lkml.org 
[lkml]   [2021]   [Feb]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH] mm, kasan: don't poison boot memory
Date
On 22.02.21 16:13, George Kennedy wrote:
>
>
> On 2/22/2021 4:52 AM, David Hildenbrand wrote:
>> On 20.02.21 00:04, George Kennedy wrote:
>>>
>>>
>>> On 2/19/2021 11:45 AM, George Kennedy wrote:
>>>>
>>>>
>>>> On 2/18/2021 7:09 PM, Andrey Konovalov wrote:
>>>>> On Fri, Feb 19, 2021 at 1:06 AM George Kennedy
>>>>> <george.kennedy@oracle.com> wrote:
>>>>>>
>>>>>>
>>>>>> On 2/18/2021 3:55 AM, David Hildenbrand wrote:
>>>>>>> On 17.02.21 21:56, Andrey Konovalov wrote:
>>>>>>>> During boot, all non-reserved memblock memory is exposed to the
>>>>>>>> buddy
>>>>>>>> allocator. Poisoning all that memory with KASAN lengthens boot
>>>>>>>> time,
>>>>>>>> especially on systems with large amount of RAM. This patch makes
>>>>>>>> page_alloc to not call kasan_free_pages() on all new memory.
>>>>>>>>
>>>>>>>> __free_pages_core() is used when exposing fresh memory during
>>>>>>>> system
>>>>>>>> boot and when onlining memory during hotplug. This patch adds a new
>>>>>>>> FPI_SKIP_KASAN_POISON flag and passes it to __free_pages_ok()
>>>>>>>> through
>>>>>>>> free_pages_prepare() from __free_pages_core().
>>>>>>>>
>>>>>>>> This has little impact on KASAN memory tracking.
>>>>>>>>
>>>>>>>> Assuming that there are no references to newly exposed pages
>>>>>>>> before they
>>>>>>>> are ever allocated, there won't be any intended (but buggy)
>>>>>>>> accesses to
>>>>>>>> that memory that KASAN would normally detect.
>>>>>>>>
>>>>>>>> However, with this patch, KASAN stops detecting wild and large
>>>>>>>> out-of-bounds accesses that happen to land on a fresh memory page
>>>>>>>> that
>>>>>>>> was never allocated. This is taken as an acceptable trade-off.
>>>>>>>>
>>>>>>>> All memory allocated normally when the boot is over keeps getting
>>>>>>>> poisoned as usual.
>>>>>>>>
>>>>>>>> Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
>>>>>>>> Change-Id: Iae6b1e4bb8216955ffc14af255a7eaaa6f35324d
>>>>>>> Not sure this is the right thing to do, see
>>>>>>>
>>>>>>> https://lkml.kernel.org/r/bcf8925d-0949-3fe1-baa8-cc536c529860@oracle.com
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Reversing the order in which memory gets allocated + used during
>>>>>>> boot
>>>>>>> (in a patch by me) might have revealed an invalid memory access
>>>>>>> during
>>>>>>> boot.
>>>>>>>
>>>>>>> I suspect that that issue would no longer get detected with your
>>>>>>> patch, as the invalid memory access would simply not get detected.
>>>>>>> Now, I cannot prove that :)
>>>>>> Since David's patch we're having trouble with the iBFT ACPI table,
>>>>>> which
>>>>>> is mapped in via kmap() - see acpi_map() in "drivers/acpi/osl.c".
>>>>>> KASAN
>>>>>> detects that it is being used after free when ibft_init() accesses
>>>>>> the
>>>>>> iBFT table, but as of yet we can't find where it get's freed (we've
>>>>>> instrumented calls to kunmap()).
>>>>> Maybe it doesn't get freed, but what you see is a wild or a large
>>>>> out-of-bounds access. Since KASAN marks all memory as freed during the
>>>>> memblock->page_alloc transition, such bugs can manifest as
>>>>> use-after-frees.
>>>>
>>>> It gets freed and re-used. By the time the iBFT table is accessed by
>>>> ibft_init() the page has been over-written.
>>>>
>>>> Setting page flags like the following before the call to kmap()
>>>> prevents the iBFT table page from being freed:
>>>
>>> Cleaned up version:
>>>
>>> diff --git a/drivers/acpi/osl.c b/drivers/acpi/osl.c
>>> index 0418feb..8f0a8e7 100644
>>> --- a/drivers/acpi/osl.c
>>> +++ b/drivers/acpi/osl.c
>>> @@ -287,9 +287,12 @@ static void __iomem *acpi_map(acpi_physical_address
>>> pg_off, unsigned long pg_sz)
>>>
>>>        pfn = pg_off >> PAGE_SHIFT;
>>>        if (should_use_kmap(pfn)) {
>>> +        struct page *page = pfn_to_page(pfn);
>>> +
>>>            if (pg_sz > PAGE_SIZE)
>>>                return NULL;
>>> -        return (void __iomem __force *)kmap(pfn_to_page(pfn));
>>> +        SetPageReserved(page);
>>> +        return (void __iomem __force *)kmap(page);
>>>        } else
>>>            return acpi_os_ioremap(pg_off, pg_sz);
>>>    }
>>> @@ -299,9 +302,12 @@ static void acpi_unmap(acpi_physical_address
>>> pg_off, void __iomem *vaddr)
>>>        unsigned long pfn;
>>>
>>>        pfn = pg_off >> PAGE_SHIFT;
>>> -    if (should_use_kmap(pfn))
>>> -        kunmap(pfn_to_page(pfn));
>>> -    else
>>> +    if (should_use_kmap(pfn)) {
>>> +        struct page *page = pfn_to_page(pfn);
>>> +
>>> +        ClearPageReserved(page);
>>> +        kunmap(page);
>>> +    } else
>>>            iounmap(vaddr);
>>>    }
>>>
>>> David, the above works, but wondering why it is now necessary. kunmap()
>>> is not hit. What other ways could a page mapped via kmap() be unmapped?
>>>
>>
>> Let me look into the code ... I have little experience with ACPI
>> details, so bear with me.
>>
>> I assume that acpi_map()/acpi_unmap() map some firmware blob that is
>> provided via firmware/bios/... to us.
>>
>> should_use_kmap() tells us whether
>> a) we have a "struct page" and should kmap() that one
>> b) we don't have a "struct page" and should ioremap.
>>
>> As it is a blob, the firmware should always reserve that memory region
>> via memblock (e.g., memblock_reserve()), such that we either
>> 1) don't create a memmap ("struct page") at all (-> case b) )
>> 2) if we have to create e memmap, we mark the page PG_reserved and
>>    *never* expose it to the buddy (-> case a) )
>>
>>
>> Are you telling me that in this case we might have a memmap for the HW
>> blob that is *not* PG_reserved? In that case it most probably got
>> exposed to the buddy where it can happily get allocated/freed.
>>
>> The latent BUG would be that that blob gets exposed to the system like
>> ordinary RAM, and not reserved via memblock early during boot.
>> Assuming that blob has a low physical address, with my patch it will
>> get allocated/used a lot earlier - which would mean we trigger this
>> latent BUG now more easily.
>>
>> There have been similar latent BUGs on ARM boards that my patch
>> discovered where special RAM regions did not get marked as reserved
>> via the device tree properly.
>>
>> Now, this is just a wild guess :) Can you dump the page when mapping
>> (before PageReserved()) and when unmapping, to see what the state of
>> that memmap is?
>
> Thank you David for the explanation and your help on this,
>
> dump_page() before PageReserved and before kmap() in the above patch:
>
> [    1.116480] ACPI: Core revision 20201113
> [    1.117628] XXX acpi_map: about to call kmap()...
> [    1.118561] page:ffffea0002f914c0 refcount:0 mapcount:0
> mapping:0000000000000000 index:0x0 pfn:0xbe453
> [    1.120381] flags: 0xfffffc0000000()
> [    1.121116] raw: 000fffffc0000000 ffffea0002f914c8 ffffea0002f914c8
> 0000000000000000
> [    1.122638] raw: 0000000000000000 0000000000000000 00000000ffffffff
> 0000000000000000
> [    1.124146] page dumped because: acpi_map pre SetPageReserved
>
> I also added dump_page() before unmapping, but it is not hit. The
> following for the same pfn now shows up I believe as a result of setting
> PageReserved:
>
> [   28.098208] BUG:Bad page state in process mo dprobe  pfn:be453
> [   28.098394] page:ffffea0002f914c0 refcount:0 mapcount:0
> mapping:0000000000000000 index:0x1 pfn:0xbe453
> [   28.098394] flags: 0xfffffc0001000(reserved)
> [   28.098394] raw: 000fffffc0001000 dead000000000100 dead000000000122
> 0000000000000000
> [   28.098394] raw: 0000000000000001 0000000000000000 00000000ffffffff
> 0000000000000000
> [   28.098394] page dumped because: PAGE_FLAGS_CHECK_AT_PREP flag(s) set
> [   28.098394] page_owner info is not present (never set?)
> [   28.098394] Modules linked in:
> [   28.098394] CPU: 2 PID: 204 Comm: modprobe Not tainted 5.11.0-3dbd5e3 #66
> [   28.098394] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996),
> BIOS 0.0.0 02/06/2015
> [   28.098394] Call Trace:
> [   28.098394]  dump_stack+0xdb/0x120
> [   28.098394]  bad_page.cold.108+0xc6/0xcb
> [   28.098394]  check_new_page_bad+0x47/0xa0
> [   28.098394]  get_page_from_freelist+0x30cd/0x5730
> [   28.098394]  ? __isolate_free_page+0x4f0/0x4f0
> [   28.098394]  ? init_object+0x7e/0x90
> [   28.098394]  __alloc_pages_nodemask+0x2d8/0x650
> [   28.098394]  ? write_comp_data+0x2f/0x90
> [   28.098394]  ? __alloc_pages_slowpath.constprop.103+0x2110/0x2110
> [   28.098394]  ? __sanitizer_cov_trace_pc+0x21/0x50
> [   28.098394]  alloc_pages_vma+0xe2/0x560
> [   28.098394]  do_fault+0x194/0x12c0
> [   28.098394]  ? write_comp_data+0x2f/0x90
> [   28.098394]  __handle_mm_fault+0x1650/0x26c0
> [   28.098394]  ? copy_page_range+0x1350/0x1350
> [   28.098394]  ? write_comp_data+0x2f/0x90
> [   28.098394]  ? write_comp_data+0x2f/0x90
> [   28.098394]  handle_mm_fault+0x1f9/0x810
> [   28.098394]  ? write_comp_data+0x2f/0x90
> [   28.098394]  do_user_addr_fault+0x6f7/0xca0
> [   28.098394]  exc_page_fault+0xaf/0x1a0
> [   28.098394]  asm_exc_page_fault+0x1e/0x30
> [   28.098394] RIP: 0010:__clear_user+0x30/0x60

I think the PAGE_FLAGS_CHECK_AT_PREP check in this instance means that
someone is trying to allocate that page with the PG_reserved bit set.
This means that the page actually was exposed to the buddy.

However, when you SetPageReserved(), I don't think that PG_buddy is set
and the refcount is 0. That could indicate that the page is on the buddy
PCP list. Could be that it is getting reused a couple of times.

The PFN 0xbe453 looks a little strange, though. Do we expect ACPI tables
close to 3 GiB ? No idea. Could it be that you are trying to map a wrong
table? Just a guess.

>
> What would be  the correct way to reserve the page so that the above
> would not be hit?

I would have assumed that if this is a binary blob, that someone (which
I think would be acpi code) reserved via memblock_reserve() early during
boot.

E.g., see drivers/acpi/tables.c:acpi_table_upgrade()->memblock_reserve().

--
Thanks,

David / dhildenb

\
 
 \ /
  Last update: 2021-02-22 17:17    [W:0.079 / U:44.868 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site