Re: [PATCH] mm: use this_cpu_cmpxchg_double in put_cpu_partial

From: Wei Yang
Date: Wed Nov 21 2018 - 19:37:03 EST


On Tue, Nov 20, 2018 at 07:18:13PM -0800, Wengang Wang wrote:
>Hi Wei,
>
>I think you will receive my reply to Zhong, But I am copying my comments for
>that patch here (again):
>
>Copy starts ==>
>
>I am not sure if the patch you mentioned intended to fix the problem here.
>With that patch the negative page->pobjects would become a large positive
>value,
>it will win the compare with s->cpu_partial and go ahead to unfreeze the
>partial slabs.
>Though it may be not a perfect fix for this issue, it really fixes (or
>workarounds) the issue here.
>I'd like to skip my patch..
>
><=== Copy ends

Thanks.

I still didn't get the point. Let's see whether I would get your replay
to that thread.

>
>thanks,
>
>wengang
>
>
>On 2018/11/20 19:02, Wei Yang wrote:
>> On Tue, Nov 20, 2018 at 09:58:58AM -0800, Wengang Wang wrote:
>> > Hi Wei,
>> >
>> >
>> > On 2018/11/17 17:02, Wei Yang wrote:
>> > > On Fri, Nov 16, 2018 at 05:33:35PM -0800, 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
>> > > Well, I seems to understand your description.
>> > >
>> > > There are two slabs
>> > >
>> > > * one which put_cpu_partial() trying to free an object
>> > > * one which is the first slab in cpu_partial list
>> > >
>> > > There is some tricky case, the first slab in cpu_partial list we
>> > > reference to will change since interrupt is not disabled.
>> > Yes, two slabs involved here just as you said above.
>> > And yes, the case is really tricky, but it's there.
>> >
>> > > > interrupt handlers. Theoretically, after we have a reference to the it,
>> > > ^^^
>> > > one more word?
>> > sorry, "the" should not be there.
>> >
>> > > > stored in _oldpage_, the first slab on the partial list on this CPU can be
>> > > ^^^
>> > > One little suggestion here, mayby use cpu_partial would be more easy to
>> > > understand. I confused this with the partial list in kmem_cache_node at
>> > > the first time. :-)
>> > Right, making others understanding easily is very important. I just meant
>> > cpu_partial.
>> >
>> > > > 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
>> > > Actually, no matter what happens after the removal of the first slab in
>> > > cpu_partial, it would leads to problem.
>> > Maybe you are right, what I see is the problem on the page->pobjects.
>> >
>> > > > 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.
>> > > Maybe not an exact solution.
>> > >
>> > > I took a look into the code and change log.
>> > >
>> > > _tid_ is introduced by commit 8a5ec0ba42c4 ('Lockless (and preemptless)
>> > > fastpaths for slub'), which is used to guard cpu_freelist. While we don't
>> > > modify _tid_ when cpu_partial changes.
>> > >
>> > > May need another _tid_ for cpu_partial?
>> > Right, _tid_ changes later than cpu_partial changes.
>> >
>> > As pointed out by Zhong Jiang, the pobjects issue is fixed by commit
>> Where you discussed this issue? Any reference I could get a look?
>>
>> > e5d9998f3e09 (not sure if by side effect, see my replay there),
>> I took a look at this commit e5d9998f3e09 ('slub: make ->cpu_partial
>> unsigned int'), but not see some relationship between them.
>>
>> Would you mind show me a link or cc me in case you have further
>> discussion?
>>
>> Thanks.
>>
>> > I'd skip this patch.?? If we found other problems regarding the change of
>> > cpu_partial, let's fix them. What do you think?
>> >
>> > thanks,
>> > wengang

--
Wei Yang
Help you, Help me