[PATCH v3] arm64: mm: Fix NOMAP page initialization

[Hanjun Guo]

On 2017/1/6 16:37, Ard Biesheuvel wrote:
> On 6 January 2017 at 01:07, Hanjun Guo <hanjun.guo at linaro.org> wrote:
>> On 2017/1/5 10:03, Hanjun Guo wrote:
>>>
>>> On 2017/1/4 21:56, Ard Biesheuvel wrote:
>>>>
>>>> On 16 December 2016 at 16:54, Robert Richter <rrichter at cavium.com> wrote:
>>>>>
>>>>> On ThunderX systems with certain memory configurations we see the
>>>>> following BUG_ON():
>>>>>
>>>>>  kernel BUG at mm/page_alloc.c:1848!
>>>>>
>>>>> This happens for some configs with 64k page size enabled. The BUG_ON()
>>>>> checks if start and end page of a memmap range belongs to the same
>>>>> zone.
>>>>>
>>>>> The BUG_ON() check fails if a memory zone contains NOMAP regions. In
>>>>> this case the node information of those pages is not initialized. This
>>>>> causes an inconsistency of the page links with wrong zone and node
>>>>> information for that pages. NOMAP pages from node 1 still point to the
>>>>> mem zone from node 0 and have the wrong nid assigned.
>>>>>
>>>>> The reason for the mis-configuration is a change in pfn_valid() which
>>>>> reports pages marked NOMAP as invalid:
>>>>>
>>>>>  68709f45385a arm64: only consider memblocks with NOMAP cleared for
>>>>> linear mapping
>>>>>
>>>>> This causes pages marked as nomap being no longer reassigned to the
>>>>> new zone in memmap_init_zone() by calling __init_single_pfn().
>>>>>
>>>>> Fixing this by implementing an arm64 specific early_pfn_valid(). This
>>>>> causes all pages of sections with memory including NOMAP ranges to be
>>>>> initialized by __init_single_page() and ensures consistency of page
>>>>> links to zone, node and section.
>>>>>
>>>>
>>>> I like this solution a lot better than the first one, but I am still
>>>> somewhat uneasy about having the kernel reason about attributes of
>>>> pages it should not touch in the first place. But the fact that
>>>> early_pfn_valid() is only used a single time in the whole kernel does
>>>> give some confidence that we are not simply moving the problem
>>>> elsewhere.
>>>>
>>>> Given that you are touching arch/arm/ as well as arch/arm64, could you
>>>> explain why only arm64 needs this treatment? Is it simply because we
>>>> don't have NUMA support there?
>>>>
>>>> Considering that Hisilicon D05 suffered from the same issue, I would
>>>> like to get some coverage there as well. Hanjun, is this something you
>>>> can arrange? Thanks
>>>
>>>
>>> Sure, we will test this patch with LTP MM stress test (which triggers
>>> the bug on D05), and give the feedback.
>>
>>
>> a update here, tested on 4.9,
>>
>>  - Applied Ard's two patches only
>>  - Applied Robert's patch only
>>
>> Both of them can work fine on D05 with NUMA enabled, which means
>> boot ok and LTP MM stress test is passed.
>>
>
> Thanks a lot Hanjun.
>
> Any comments on the performance impact (including boot time) ?

Didn't collect the performance data yet, any recommended test
suite?  Is it sysbench ok? we can test it and collect the data.

Thanks
Hanjun