linux-kernel - Re: [PATCH] mm/hmm: Fix bad subpage pointer in try_to_unmap

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Mon, 15 Jul 2019 17:38:04 -0700
From:   Ralph Campbell <rcampbell@...dia.com>
To:     John Hubbard <jhubbard@...dia.com>,
        Andrew Morton <akpm@...ux-foundation.org>
CC:     <linux-mm@...ck.org>, <linux-kernel@...r.kernel.org>,
        Jérôme Glisse <jglisse@...hat.com>,
        "Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>,
        Mike Kravetz <mike.kravetz@...cle.com>,
        Jason Gunthorpe <jgg@...lanox.com>
Subject: Re: [PATCH] mm/hmm: Fix bad subpage pointer in try_to_unmap_one


On 7/15/19 4:34 PM, John Hubbard wrote:
> On 7/15/19 3:00 PM, Andrew Morton wrote:
>> On Tue, 9 Jul 2019 18:24:57 -0700 Ralph Campbell <rcampbell@...dia.com> wrote:
>>
>>>
>>> On 7/9/19 5:28 PM, Andrew Morton wrote:
>>>> On Tue, 9 Jul 2019 15:35:56 -0700 Ralph Campbell <rcampbell@...dia.com> wrote:
>>>>
>>>>> When migrating a ZONE device private page from device memory to system
>>>>> memory, the subpage pointer is initialized from a swap pte which computes
>>>>> an invalid page pointer. A kernel panic results such as:
>>>>>
>>>>> BUG: unable to handle page fault for address: ffffea1fffffffc8
>>>>>
>>>>> Initialize subpage correctly before calling page_remove_rmap().
>>>>
>>>> I think this is
>>>>
>>>> Fixes:  a5430dda8a3a1c ("mm/migrate: support un-addressable ZONE_DEVICE page in migration")
>>>> Cc: stable
>>>>
>>>> yes?
>>>>
>>>
>>> Yes. Can you add this or should I send a v2?
>>
>> I updated the patch.  Could we please have some review input?
>>
>>
>> From: Ralph Campbell <rcampbell@...dia.com>
>> Subject: mm/hmm: fix bad subpage pointer in try_to_unmap_one
>>
>> When migrating a ZONE device private page from device memory to system
>> memory, the subpage pointer is initialized from a swap pte which computes
>> an invalid page pointer. A kernel panic results such as:
>>
>> BUG: unable to handle page fault for address: ffffea1fffffffc8
>>
>> Initialize subpage correctly before calling page_remove_rmap().
>>
>> Link: http://lkml.kernel.org/r/20190709223556.28908-1-rcampbell@nvidia.com
>> Fixes: a5430dda8a3a1c ("mm/migrate: support un-addressable ZONE_DEVICE page in migration")
>> Signed-off-by: Ralph Campbell <rcampbell@...dia.com>
>> Cc: "Jérôme Glisse" <jglisse@...hat.com>
>> Cc: "Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>
>> Cc: Mike Kravetz <mike.kravetz@...cle.com>
>> Cc: Jason Gunthorpe <jgg@...lanox.com>
>> Cc: <stable@...r.kernel.org>
>> Signed-off-by: Andrew Morton <akpm@...ux-foundation.org>
>> ---
>>
>>   mm/rmap.c |    1 +
>>   1 file changed, 1 insertion(+)
>>
>> --- a/mm/rmap.c~mm-hmm-fix-bad-subpage-pointer-in-try_to_unmap_one
>> +++ a/mm/rmap.c
>> @@ -1476,6 +1476,7 @@ static bool try_to_unmap_one(struct page
>>   			 * No need to invalidate here it will synchronize on
>>   			 * against the special swap migration pte.
>>   			 */
>> +			subpage = page;
>>   			goto discard;
>>   		}
>>   
> 
> Hi Ralph and everyone,
> 
> While the above prevents a crash, I'm concerned that it is still not
> an accurate fix. This fix leads to repeatedly removing the rmap, against the
> same struct page, which is odd, and also doesn't directly address the
> root cause, which I understand to be: this routine can't handle migrating
> the zero page properly--over and back, anyway. (We should also mention more
> about how this is triggered, in the commit description.)
> 
> I'll take a closer look at possible fixes (I have to step out for a bit) soon,
> but any more experienced help is also appreciated here.
> 
> thanks,

I'm not surprised at the confusion. It took me quite awhile to 
understand how migrate_vma() works with ZONE_DEVICE private memory.
The big point to be aware of is that when migrating a page to
device private memory, the source page's page->mapping pointer
is copied to the ZONE_DEVICE struct page and the page_mapcount()
is increased. So, the kernel sees the page as being "mapped"
but the page table entry as being is_swap_pte() so the CPU will fault
if it tries to access the mapped address.
So yes, the source anon page is unmapped, DMA'ed to the device,
and then mapped again. Then on a CPU fault, the zone device page
is unmapped, DMA'ed to system memory, and mapped again.
The rmap_walk() is used to clear the temporary migration pte so
that is another important detail of how migrate_vma() works.
At the moment, only single anon private pages can migrate to
device private memory so there are no subpages and setting it to "page"
should be correct for now. I'm looking at supporting migration of
transparent huge pages but that is a work in progress.
Let me know how much of all that you think should be in the change log.
Getting an Acked-by from Jerome would be nice too.

I see Christoph Hellwig got confused by this too [1].
I have a patch to clear page->mapping when freeing ZONE_DEVICE private
struct pages which I'll send out soon.
I'll probably also add some comments to struct page to include the
above info and maybe remove the _zd_pad_1 field.

[1] 740d6310ed4cd5c78e63 ("mm: don't clear ->mapping in hmm_devmem_free")