lkml.org 
[lkml]   [2016]   [Aug]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [RFC] can we use vmalloc to alloc thread stack if compaction failed
On Sun, Jul 31, 2016 at 10:30 PM, Joonsoo Kim <iamjoonsoo.kim@lge.com> wrote:
> On Fri, Jul 29, 2016 at 12:47:38PM -0700, Andy Lutomirski wrote:
>> ---------- Forwarded message ----------
>> From: "Joonsoo Kim" <iamjoonsoo.kim@lge.com>
>> 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@kernel.org>
>> Cc: "Xishi Qiu" <qiuxishi@huawei.com>, "Michal Hocko"
>> <mhocko@kernel.org>, "Tejun Heo" <tj@kernel.org>, "Ingo Molnar"
>> <mingo@kernel.org>, "Peter Zijlstra" <peterz@infradead.org>, "LKML"
>> <linux-kernel@vger.kernel.org>, "Linux MM" <linux-mm@kvack.org>,
>> "Yisheng Xie" <xieyisheng1@huawei.com>
>>
>> > 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@huawei.com> 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)
> )
>
>

$ cat sgstack.cocci
@@
local idexpression S;
expression A, B;
@@

(
* sg_init_one(A, &S, B)
|
* virt_to_phys(&S)


not very inspiring. I barely understand Coccinelle syntax, and sadly
I find the manual nearly incomprehensible. I can read the grammar,
but that doesn't mean I know what the various declarations do.

\
 
 \ /
  Last update: 2016-08-10 21:41    [W:0.074 / U:0.036 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site