Re: [PATCH] mm: page_is_guard return false when page_ext arrays are not allocated yet

From: Shi, Yang
Date: Tue May 24 2016 - 12:37:29 EST


On 5/23/2016 10:37 PM, Joonsoo Kim wrote:
On Fri, May 20, 2016 at 10:00:06AM -0700, Shi, Yang wrote:
On 5/19/2016 7:40 PM, Joonsoo Kim wrote:
2016-05-20 2:18 GMT+09:00 Shi, Yang <yang.shi@xxxxxxxxxx>:
On 5/18/2016 5:28 PM, Joonsoo Kim wrote:

Vlastiml, thanks for ccing me on original bug report.

On Wed, May 18, 2016 at 03:23:45PM -0700, Yang Shi wrote:

When enabling the below kernel configs:

CONFIG_DEFERRED_STRUCT_PAGE_INIT
CONFIG_DEBUG_PAGEALLOC
CONFIG_PAGE_EXTENSION
CONFIG_DEBUG_VM

kernel bootup may fail due to the following oops:

BUG: unable to handle kernel NULL pointer dereference at (null)
IP: [<ffffffff8118d982>] free_pcppages_bulk+0x2d2/0x8d0
PGD 0
Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
Modules linked in:
CPU: 11 PID: 106 Comm: pgdatinit1 Not tainted 4.6.0-rc5-next-20160427 #26
Hardware name: Intel Corporation S5520HC/S5520HC, BIOS
S5500.86B.01.10.0025.030220091519 03/02/2009
task: ffff88017c080040 ti: ffff88017c084000 task.ti: ffff88017c084000
RIP: 0010:[<ffffffff8118d982>] [<ffffffff8118d982>]
free_pcppages_bulk+0x2d2/0x8d0
RSP: 0000:ffff88017c087c48 EFLAGS: 00010046
RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000001
RDX: 0000000000000980 RSI: 0000000000000080 RDI: 0000000000660401
RBP: ffff88017c087cd0 R08: 0000000000000401 R09: 0000000000000009
R10: ffff88017c080040 R11: 000000000000000a R12: 0000000000000400
R13: ffffea0019810000 R14: ffffea0019810040 R15: ffff88066cfe6080
FS: 0000000000000000(0000) GS:ffff88066cd40000(0000)
knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000000000000 CR3: 0000000002406000 CR4: 00000000000006e0
Stack:
ffff88066cd5bbd8 ffff88066cfe6640 0000000000000000 0000000000000000
0000001f0000001f ffff88066cd5bbe8 ffffea0019810000 000000008118f53e
0000000000000009 0000000000000401 ffffffff0000000a 0000000000000001
Call Trace:
[<ffffffff8118f602>] free_hot_cold_page+0x192/0x1d0
[<ffffffff8118f69c>] __free_pages+0x5c/0x90
[<ffffffff8262a676>] __free_pages_boot_core+0x11a/0x14e
[<ffffffff8262a6fa>] deferred_free_range+0x50/0x62
[<ffffffff8262aa46>] deferred_init_memmap+0x220/0x3c3
[<ffffffff8262a826>] ? setup_per_cpu_pageset+0x35/0x35
[<ffffffff8108b1f8>] kthread+0xf8/0x110
[<ffffffff81c1b732>] ret_from_fork+0x22/0x40
[<ffffffff8108b100>] ? kthread_create_on_node+0x200/0x200
Code: 49 89 d4 48 c1 e0 06 49 01 c5 e9 de fe ff ff 4c 89 f7 44 89 4d b8
4c 89 45 c0 44 89 5d c8 48 89 4d d0 e8 62 c7 07 00 48 8b 4d d0 <48> 8b 00 44
8b 5d c8 4c 8b 45 c0 44 8b 4d b8 a8 02 0f 84 05 ff
RIP [<ffffffff8118d982>] free_pcppages_bulk+0x2d2/0x8d0
RSP <ffff88017c087c48>
CR2: 0000000000000000

The problem is lookup_page_ext() returns NULL then page_is_guard() tried
to
access it in page freeing.

page_is_guard() depends on PAGE_EXT_DEBUG_GUARD bit of page extension
flag, but
freeing page might 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.


Patch itself looks find to me because I also found that this kind of
problem happens during memory hotplug. So, we need to fix more sites,
all callers of lookup_page_ext().


Yes, I agree. I will come up with a patch or a couple of patches to check
the return value of lookup_page_ext().


But, I'd like to know how your problem occurs during bootup.
debug_guardpage_enabled() is turned to 'enable' after page_ext is
initialized. Before that, page_is_guard() unconditionally returns
false so I think that the problem what you mentioned can't happen.

Could you check that when debug_guardpage_enabled() returns 'enable'
and init_section_page_ext() is called?


I think the problem is I have CONFIG_DEFERRED_STRUCT_PAGE_INIT enabled,
which will defer some struct pages initialization to "pgdatinitX" kernel
thread in page_alloc_init_late(). But, page_ext_init() is called before it.
So, it leads debug_guardpage_enabled() return true, but page extension is
not allocated yet for the struct pages initialized by "pgdatinitX".

No. After page_ext_init(), it is ensured that all page extension is initialized.

It sounds page_ext_init() should be called after page_alloc_init_late(). Or
it should be just incompatible with CONFIG_DEFERRED_STRUCT_PAGE_INIT.

I will try to move the init call around.

We need to investigate more. I guess that problem is introduced by
CONFIG_DEFERRED_STRUCT_PAGE_INIT. It makes pfn_to_nid() invalid
until page_alloc_init_late() is done. That is a big side-effect. If
there is pfn walker
and it uses pfn_to_nid() between memmap_init_zone() and page_alloc_init_late(),
it also has same problem. So, we need to think how to fix it more
carefully.

Thanks for the analysis. I think you are correct. Since pfn_to_nid()
depends on memmap which has not been fully setup yet until
page_alloc_init_late() is done.

So, for such usecase early_pfn_to_nid() should be used.


Anyway, to make sure that my assumption is true, could you confirm that
below change fix your problem?

Yes, it does.


Thanks.

----->8----------
diff --git a/mm/page_ext.c b/mm/page_ext.c
index 2d864e6..cac5dc9 100644
--- a/mm/page_ext.c
+++ b/mm/page_ext.c
@@ -391,7 +391,7 @@ void __init page_ext_init(void)
* -------------pfn-------------->
* N0 | N1 | N2 | N0 | N1 | N2|....
*/
- if (pfn_to_nid(pfn) != nid)
+ if (!early_pfn_in_nid(pfn nid))

early_pfn_in_nid() is static function in page_alloc.c. I'm supposed
early_pfn_to_nid() should be used.

Thanks for checking. Then, please revert your patch "mm: call
page_ext_init() after all struct pages are initialized" and apply this
change, because deferring page_ext_init() would make page owner which
uses page_ext miss some early page allocation callsites. Although it
already miss some early page allocation callsites, we don't need to
miss more.

Definitely agree. I will come up with an incremental patch to adopt the new solution.

Thanks,
Yang


Thanks.