Re: [PATCH] sparse: Track the boundaries of memory sections for accurate checks

From: Dan Williams
Date: Wed Sep 14 2016 - 18:53:55 EST


On Wed, Sep 14, 2016 at 3:11 PM, Raslan, KarimAllah <karahmed@xxxxxxxxx> wrote:
>
> Ahmed, Karim Allah
> karahmed@xxxxxxxxx
>
>
>
>> On Sep 15, 2016, at 12:05 AM, Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
>>
>> On Wed, Sep 14, 2016 at 2:40 PM, Raslan, KarimAllah <karahmed@xxxxxxxxx> wrote:
>>>
>>>
>>> On 6/20/16, 10:23 AM, "Michal Hocko" <mhocko@xxxxxxxxxx> wrote:
>>>
>>> On Sat 18-06-16 12:11:19, KarimAllah Ahmed wrote:
>>>> When sparse memory model is used an array of memory sections is created to
>>>> track each block of contiguous physical pages. Each element of this array
>>>> contains PAGES_PER_SECTION pages. During the creation of this array the actual
>>>> boundaries of the memory block is lost, so the whole block is either considered
>>>> as present or not.
>>>>
>>>> pfn_valid() in the sparse memory configuration checks which memory sections the
>>>> pfn belongs to then checks whether it's present or not. This yields sub-optimal
>>>> results when the available memory doesn't cover the whole memory section,
>>>> because pfn_valid will return 'true' even for the unavailable pfns at the
>>>> boundaries of the memory section.
>>>
>>> Please be more verbose of _why_ the patch is needed. Why those
>>> "sub-optimal results" matter?
>>>
>>> Does this make sense to you ?
>>
>> [ channeling my inner akpm ]
>>
>> What's the user visible effect of this change? What code is getting
>> tripped up by pfn_valid() being imprecise, and why is changing
>> pfn_valid() the preferred fix?
>
> I did expand the commit message in v2 of this patch to answer these questions:
>
> https://patchwork.kernel.org/patch/9190737/
>

Ah, ok that gives more information about how it is "potentially"
problematic, so I assume you are hitting those problems in practice?
That way the patch can be marked for -stable if this is a problem
others are likely to run into in older kernels. When pfn_valid()
fails does /proc/iomem show that address "System RAM"? If not then we
could alternatively convert these problematic usages to use
region_intersects().