Re: [PATCH] mm: check the return value of lookup_page_ext for all call sites

From: Shi, Yang
Date: Wed Jun 01 2016 - 16:53:09 EST


On 5/29/2016 11:08 PM, Minchan Kim wrote:
On Mon, May 30, 2016 at 02:39:06PM +0900, Joonsoo Kim wrote:
On Fri, May 27, 2016 at 05:11:08PM +0900, Minchan Kim wrote:
On Fri, May 27, 2016 at 03:08:39PM +0900, Joonsoo Kim wrote:
On Fri, May 27, 2016 at 02:14:32PM +0900, Minchan Kim wrote:
On Thu, May 26, 2016 at 04:15:28PM -0700, Shi, Yang wrote:
On 5/25/2016 5:37 PM, Minchan Kim wrote:
On Tue, May 24, 2016 at 11:58:11AM +0900, Minchan Kim wrote:
On Mon, May 23, 2016 at 10:16:08AM -0700, Yang Shi wrote:
Per the discussion with Joonsoo Kim [1], we need check the return value of
lookup_page_ext() for all call sites since it might return NULL in some cases,
although it is unlikely, i.e. memory hotplug.

Tested with ltp with "page_owner=0".

[1] http://lkml.kernel.org/r/20160519002809.GA10245@js1304-P5Q-DELUXE

Signed-off-by: Yang Shi <yang.shi@xxxxxxxxxx>

I didn't read code code in detail to see how page_ext memory space
allocated in boot code and memory hotplug but to me, it's not good
to check NULL whenever we calls lookup_page_ext.

More dangerous thing is now page_ext is used by optionable feature(ie, not
critical for system stability) but if we want to use page_ext as
another important tool for the system in future,
it could be a serious problem.

Hello, Minchan.

Hi Joonsoo,


I wonder how pages that isn't managed by kernel yet will cause serious
problem. Until onlining, these pages are out of our scope. Any
information about them would be useless until it is actually
activated. I guess that returning NULL for those pages will not hurt
any functionality. Do you have any possible scenario that this causes the
serious problem?

I don't have any specific usecase now. That's why I said "in future".
And I don't want to argue whether there is possible scenario or not
to make the feature useful but if you want, I should write novel.
One of example, pop up my mind, xen, hv and even memory_hotplug itself
might want to use page_ext for their functionality extension to hook
guest pages.

There is no detail so I can't guess how to use it and how it causes
the serious problem. But, we can do it when it is really needed.


My opinion is that page_ext is extension of struct page so it would
be better to allow any operation on struct page without any limitation
if we can do it. Whether it's useful or useless depend on random
usecase and we don't need to limit that way from the beginning.

If there is no drawback, it would be a better design. But, we have
trade-off that for some case that the memory is added but not
onlined, there is memory saving if we allocate page_ext later.
So, in current situation that there is no user to require such
guarantee, I don't think it's worth doing right now.

However, current design allows deferred page_ext population so any user
of page_ext should keep it in mind and should either make fallback plan
or don't use page_ext for those cases. If we decide go this way through
discussion, at least, we should make such limitation more clear to
somewhere in this chance, maybe around page_ext_operation->need comment.

Agreed.

Okay, We realized from this discussion that by design, guest of page_ext
at the meoment should know his page_ext access from the page can be failed
so every caller should prepare for it.

Shi, Yang, Please include some comment about that in your patch to
prevent further reviewer waste his time with repeating same discussion
and client of page_ext can know the limitation.


My comment's point is that we should consider that way at least. It's
worth to discuss pros and cons, what's the best and what makes that way
hesitate if we can't.

Yes, your suggestion would be good for future direction, but, for now,
I think that inserting NULL to all callers is right fix.

1) Current design that page_ext is allocated when online is design
decision of page_ext to save memory as much as possible. Fixing
possible problem within this design decision looks good to me.

Okay. Shi Yang, please include this comment in your patch, too.

There is already such comment in lookup_page_ext:

* The sanity checks the page allocator does upon freeing a
* page can reach here before the page_ext arrays are
* allocated when feeding a range of pages to the allocator
* for the first time during bootup or memory hotplug.

I will add more details, i.e. newly added but not onlined memory could reach here too.



2) Maybe, we need to backport fixes because it would crash older
kernels. In this case, fix with NULL is easy to backport.

Agreed.

Then, Shi Yang need to mark the page as stable.

Agreed.

Shi, Please resend your patch with hard testing and more better
description with marking it as stable.

Will come with an incremental patch to remove CONFIG_DEBUG #ifdef and fix the improve the comments since the comments are included by it.

Thanks,
Yang

And I know another race problem about Shi's patch.
I will reply to the thread.




And, allocation such memory space doesn't come from free. If someone
just add the memory device and don't online it, these memory will be

Here goes several questions.
Cced hotplug guys

1.
If someone just add the memory device without onlining, kernel code
can return pfn_valid == true on the offlined page?

AFAIK, yes.

2.
If so, it means memmap on offline memory is already populated somewhere.
Where is the memmap allocated? part of offlined memory space or other memory?

Other memory.

3. Could we allocate page_ext in part of offline memory space so that
it doesn't consume online memory.

wasted. I don't know if there is such a usecase but it's possible
scenario.



Can we put some hooks of page_ext into memory-hotplug so guarantee
that page_ext memory space is allocated with memmap space at the
same time? IOW, once every PFN wakers find a page is valid, page_ext
is valid, too so lookup_page_ext never returns NULL on valid page
by design.

I hope we consider this direction, too.

Yang, Could you think about this?

Thanks a lot for the suggestion. Sorry for the late reply, I was
busy on preparing patches. I do agree this is a direction we should
look into, but I haven't got time to think about it deeper. I hope
Joonsoo could chime in too since he is the original author for page
extension.


Even, your patch was broken, I think.
It doesn't work with !CONFIG_DEBUG_VM && !CONFIG_PAGE_POISONING because
lookup_page_ext doesn't return NULL in that case.

Actually, I think the #ifdef should be removed if lookup_page_ext()
is possible to return NULL. It sounds not make sense returning NULL
only when DEBUG_VM is enabled. It should return NULL no matter what
debug config is selected. If Joonsoo agrees with me I'm going to
come up with a patch to fix it.

Agreed but let's wait for Minchan's response.

If we goes this way, how to guarantee this race?

kpageflags_read
stable_page_flags
page_is_idle
lookup_page_ext
section = __pfn_to_section(pfn)
offline_pages
memory_notify(MEM_OFFLINE)
offline_page_ext
ms->page_ext = NULL
section->page_ext + pfn

I think that it is a fundamental problem of memory hotplug.
There is similar race with struct page for offlined memory.



kpageflags_read
pfn_valid
remove_memory
stable_page_flags
crash!

I already reported similar race problem to memory hotplug guys but
didn't get any answer.

lkml.kernel.org/r/20151221031501.GA32524@js1304-P5Q-DELUXE

Who's in charge of memory-hotplug? Kame, Could you nudge him?


Thanks.