Re: [PATCH] powerpc: Perform a bounds check in arch_add_memory

[David Hildenbrand <david@redhat.com>]

On 02.09.19 01:54, Alastair D'Silva wrote:
> On Tue, 2019-08-27 at 09:13 +0200, David Hildenbrand wrote:
>> On 27.08.19 08:39, Alastair D'Silva wrote:
>>> On Tue, 2019-08-27 at 08:28 +0200, Michal Hocko wrote:
>>>> On Tue 27-08-19 15:20:46, Alastair D'Silva wrote:
>>>>> From: Alastair D'Silva <alastair@d-silva.org>
>>>>>
>>>>> It is possible for firmware to allocate memory ranges outside
>>>>> the range of physical memory that we support
>>>>> (MAX_PHYSMEM_BITS).
>>>>
>>>> Doesn't that count as a FW bug? Do you have any evidence of that
>>>> in
>>>> the
>>>> field? Just wondering...
>>>>
>>>
>>> Not outside our lab, but OpenCAPI attached LPC memory is assigned
>>> addresses based on the slot/NPU it is connected to. These addresses
>>> prior to:
>>> 4ffe713b7587 ("powerpc/mm: Increase the max addressable memory to
>>> 2PB")
>>> were inaccessible and resulted in bogus sections - see our
>>> discussion
>>> on 'mm: Trigger bug on if a section is not found in __section_nr'.
>>> Doing this check here was your suggestion :)
>>>
>>> It's entirely possible that a similar problem will occur in the
>>> future,
>>> and it's cheap to guard against, which is why I've added this.
>>>
>>
>> If you keep it here, I guess this should be wrapped by a
>> WARN_ON_ONCE().
>>
>> If we move it to common code (e.g., __add_pages() or add_memory()),
>> then
>> probably not. I can see that s390x allows to configure
>> MAX_PHYSMEM_BITS,
>> so the check could actually make sense.
>>
> 
> I couldn't see a nice platform indepedent way to determine the
> allowable address range, but if there is, then I'll move this to the
> generic code instead.
> 

At least on the !ZONE_DEVICE path we have

__add_memory() -> register_memory_resource() ...

return ERR_PTR(-E2BIG);


I was thinking about something like

int add_pages()
{
	if ((start + size - 1) >> MAX_PHYSMEM_BITS)
		return -E2BIG;	

	return arch_add_memory(...)
}

And switching users of arch_add_memory() to add_pages(). However, x86
already has an add_pages() function, so that would need some more thought.

Maybe simply renaming the existing add_pages() to arch_add_pages().

add_pages(): Create virtual mapping
__add_pages(): Don't create virtual mapping

arch_add_memory(): Arch backend for add_pages()
arch_add_pages(): Arch backend for __add_pages()

It would be even more consistent if we would have arch_add_pages() vs.
__arch_add_pages().

-- 

Thanks,

David / dhildenb