Re: [RFC] can we use vmalloc to alloc thread stack if compaction failed

From: Joonsoo Kim
Date: Mon Aug 01 2016 - 01:41:14 EST


On Fri, Jul 29, 2016 at 12:47:38PM -0700, Andy Lutomirski wrote:
> ---------- Forwarded message ----------
> From: "Joonsoo Kim" <iamjoonsoo.kim@xxxxxxx>
> Date: Jul 28, 2016 7:57 PM
> Subject: Re: [RFC] can we use vmalloc to alloc thread stack if compaction failed
> To: "Andy Lutomirski" <luto@xxxxxxxxxx>
> Cc: "Xishi Qiu" <qiuxishi@xxxxxxxxxx>, "Michal Hocko"
> <mhocko@xxxxxxxxxx>, "Tejun Heo" <tj@xxxxxxxxxx>, "Ingo Molnar"
> <mingo@xxxxxxxxxx>, "Peter Zijlstra" <peterz@xxxxxxxxxxxxx>, "LKML"
> <linux-kernel@xxxxxxxxxxxxxxx>, "Linux MM" <linux-mm@xxxxxxxxx>,
> "Yisheng Xie" <xieyisheng1@xxxxxxxxxx>
>
> > On Thu, Jul 28, 2016 at 08:07:51AM -0700, Andy Lutomirski wrote:
> > > On Thu, Jul 28, 2016 at 3:51 AM, Xishi Qiu <qiuxishi@xxxxxxxxxx> wrote:
> > > > On 2016/7/28 17:43, Michal Hocko wrote:
> > > >
> > > >> On Thu 28-07-16 16:45:06, Xishi Qiu wrote:
> > > >>> On 2016/7/28 15:58, Michal Hocko wrote:
> > > >>>
> > > >>>> On Thu 28-07-16 15:41:53, Xishi Qiu wrote:
> > > >>>>> On 2016/7/28 15:20, Michal Hocko wrote:
> > > >>>>>
> > > >>>>>> On Thu 28-07-16 15:08:26, Xishi Qiu wrote:
> > > >>>>>>> Usually THREAD_SIZE_ORDER is 2, it means we need to alloc 16kb continuous
> > > >>>>>>> physical memory during fork a new process.
> > > >>>>>>>
> > > >>>>>>> If the system's memory is very small, especially the smart phone, maybe there
> > > >>>>>>> is only 1G memory. So the free memory is very small and compaction is not
> > > >>>>>>> always success in slowpath(__alloc_pages_slowpath), then alloc thread stack
> > > >>>>>>> may be failed for memory fragment.
> > > >>>>>>
> > > >>>>>> Well, with the current implementation of the page allocator those
> > > >>>>>> requests will not fail in most cases. The oom killer would be invoked in
> > > >>>>>> order to free up some memory.
> > > >>>>>>
> > > >>>>>
> > > >>>>> Hi Michal,
> > > >>>>>
> > > >>>>> Yes, it success in most cases, but I did have seen this problem in some
> > > >>>>> stress-test.
> > > >>>>>
> > > >>>>> DMA free:470628kB, but alloc 2 order block failed during fork a new process.
> > > >>>>> There are so many memory fragments and the large block may be soon taken by
> > > >>>>> others after compact because of stress-test.
> > > >>>>>
> > > >>>>> --- dmesg messages ---
> > > >>>>> 07-13 08:41:51.341 <4>[309805.658142s][pid:1361,cpu5,sManagerService]sManagerService: page allocation failure: order:2, mode:0x2000d1
> > > >>>>
> > > >>>> Yes but this is __GFP_DMA allocation. I guess you have already reported
> > > >>>> this failure and you've been told that this is quite unexpected for the
> > > >>>> kernel stack allocation. It is your out-of-tree patch which just makes
> > > >>>> things worse because DMA restricted allocations are considered "lowmem"
> > > >>>> and so they do not invoke OOM killer and do not retry like regular
> > > >>>> GFP_KERNEL allocations.
> > > >>>
> > > >>> Hi Michal,
> > > >>>
> > > >>> Yes, we add GFP_DMA, but I don't think this is the key for the problem.
> > > >>
> > > >> You are restricting the allocation request to a single zone which is
> > > >> definitely not good. Look at how many larger order pages are available
> > > >> in the Normal zone.
> > > >>
> > > >>> If we do oom-killer, maybe we will get a large block later, but there
> > > >>> is enough free memory before oom(although most of them are fragments).
> > > >>
> > > >> Killing a task is of course the last resort action. It would give you
> > > >> larger order blocks used for the victims thread.
> > > >>
> > > >>> I wonder if we can alloc success without kill any process in this situation.
> > > >>
> > > >> Sure it would be preferable to compact that memory but that might be
> > > >> hard with your restriction in place. Consider that DMA zone would tend
> > > >> to be less movable than normal zones as users would have to pin it for
> > > >> DMA. Your DMA is really large so this might turn out to just happen to
> > > >> work but note that the primary problem here is that you put a zone
> > > >> restriction for your allocations.
> > > >>
> > > >>> Maybe use vmalloc is a good way, but I don't know the influence.
> > > >>
> > > >> You can have a look at vmalloc patches posted by Andy. They are not that
> > > >> trivial.
> > > >>
> > > >
> > > > Hi Michal,
> > > >
> > > > Thank you for your comment, could you give me the link?
> > > >
> > >
> > > I've been keeping it mostly up to date in this branch:
> > >
> > > https://git.kernel.org/cgit/linux/kernel/git/luto/linux.git/log/?h=x86/vmap_stack
> > >
> > > It's currently out of sync due to a bunch of the patches being queued
> > > elsewhere for the merge window.
> >
> > Hello, Andy.
> >
> > I have some questions about it.
> >
> > IIUC, to turn on HAVE_ARCH_VMAP_STACK on different architecture, there
> > is nothing to be done in architecture side if the architecture doesn't
> > support lazily faults in top-level paging entries for the vmalloc
> > area. Is my understanding is correct?
> >
>
> There should be nothing fundamental that needs to be done. On the
> other hand, it might be good to make sure the arch code can print a
> clean stack trace on stack overflow.
>
> If it's helpful, I just pushed out anew

You mean that you can turn on HAVE_ARCH_VMAP_STACK on the other arch? It
would be helpful. :)

>
> > And, I'd like to know how you search problematic places using kernel
> > stack for DMA.
> >
>
> I did some searching for problematic sg_init_buf calls using
> Coccinelle. I'm not very good at Coccinelle, so I may have missed
> something.

I'm also not familiar with Coccinelle. Could you share your .cocci
script? I can think of following one but there would be a better way.

virtual report

@stack_var depends on report@
type T1;
expression E1, E2;
identifier I1;
@@
(
* T1 I1;
)
...
(
* sg_init_one(E1, &I1, E2)
|
* sg_set_buf(E1, &I1, E2)
)

@stack_arr depends on report@
type T1;
expression E1, E2, E3;
identifier I1;
@@
(
* T1 I1[E1];
)
...
(
* sg_init_one(E2, I1, E3)
|
* sg_set_buf(E2, I1, E3)
)


> For the most part, DMA API debugging should have found the problems
> already. The ones I found were in drivers that didn't do real DMA:
> crypto users and virtio.

Ah... using stack for DMA API is already prohibited.

Thanks.