[RFC] mm: Fix memory corruption caused by deferred page initialization

[Gavin Shan]

On Sun, Mar 27, 2016 at 12:37:09AM +1100, Gavin Shan wrote:
>On Sat, Mar 26, 2016 at 08:47:17PM +1100, Michael Ellerman wrote:
>>Hi Gavin,
>>
>>On Fri, 2016-25-03 at 16:05:29 UTC, Gavin Shan wrote:
>>> During deferred page initialization, the pages are moved from memblock
>>> or bootmem to buddy allocator without checking they were reserved. Those
>>> reserved pages can be reallocated to somebody else by buddy/slab allocator.
>>> It leads to memory corruption and potential kernel crash eventually.
>>
>>Can you give me a bit more detail on what the bug is?
>>
>>I haven't seen any issues on my systems, but I realise now I haven't enabled
>>DEFERRED_STRUCT_PAGE_INIT - I assumed it was enabled by default.
>>
>>How did this get tested before submission?
>>
>
>Michael, I have to reply with same context in another thread in case 
>somebody else wants to understand more: Li, who is in the cc list, is
>backporting deferred page initialization (CONFIG_DEFERRED_STRUCT_PAGE_INIT)
>from upstream kernel to RHEL 7.2 or 7.3 kernel (3.10.0-357.el7). RHEL kernel
>has (!CONFIG_NO_BOOTMEM && CONFIG_DEFERRED_STRUCT_PAGE_INIT), meaning
>bootmem is enabled. She eventually runs into kernel crash and I jumped
>in to help understanding the root cause.
>
>There're two related kernel config options: ARCH_SUPPORTS_DEFERRED_STRUCT_PAGE_INIT
>and DEFERRED_STRUCT_PAGE_INIT. The former one is enabled on PPC by default.
>The later one isn't enabled by default.
>
>There are two test cases I had:
>
>- With (!CONFIG_NO_BOOTMEM && CONFIG_DEFERRED_STRUCT_PAGE_INIT)
>on PowerNV platform, upstream kernel (4.5.rc7) and additional patch to support
>bootmem as it was removed on powerpc a while ago.
>
>- With (CONFIG_NO_BOOTMEM && CONFIG_DEFERRED_STRUCT_PAGE_INIT) on PowerNV platform,
>upstream kernel (4.5.rc7), I dumped the reserved memblock regions and added printk
>in function deferred_init_memmap() to check if memblock reserved PFN 0x1fff80 (one
>page in memblock reserved region#31, refer to the below kernel log) is released
>to buddy allocator or not when doing deferred page struct initialization. I did
>see that PFN is released to buddy allocator at that time. However, I didn't see
>kernel crash and it would be luck and the current deferred page struct initialization
>implementation: The pages in region [0, 2GB] except the memblock reserved ones are
>presented to buddy allocator at early stage. It's not deferred. So for the pages in
>[0, 2GB], we don't have consistency issue between memblock and buddy allocator.
>The pages in region [2GB ...] are all presented to buddy allocator despite they're
>reserved in memblock or not. It ensures the kernel text section isn't corrupted
>and we're lucky not seeing program interrupt because of illegal instruction.
>

After more debugging, it turns out that Michael is correct: we don't have problem
when CONFIG_NO_BOOTMEM=y. In the case, the page frames in [2G ...] is marked as
reserved in early stage (as below function calls reveal). During the deferred
initialization stage, those reserved pages won't be released to buddy allocator:

- Below function calls mark reserved pages according to memblock reserved regions:
  init/main.c::start_kernel()
  init/main.c::mm_init()
  arch/powerpc/mm/mem.c::mem_init()
  nobootmem.c::free_all_bootmem()            <-> bootmem.c::free_all_bootmem() on !CONFIG_NO_BOOTMEM
  nobootmem.c::free_low_memory_core_early()
  nobootmem.c::reserve_bootmem_region()

- In page_alloc.c::deferred_init_memmap(), the reserved pages aren't released
  to buddy allocator with below check:

                        if (page->flags) {
                                VM_BUG_ON(page_zone(page) != zone);
                                goto free_range;
                        }


So the issue is only existing when CONFIG_NO_BOOTMEM=n. The alternative fix would
be similar to what we have on !CONFIG_NO_BOOTMEM: In early stage, all page structs
for bootmem reserved pages are initialized and mark them with PG_reserved. I'm
not sure it's worthy to fix it as we won't support bootmem as Michael mentioned.

Thanks,
Gavin