Re: [PATCH 2/2] x86, e820 disable ACPI Memory Hotplug if memory mapping is specified by user [v2]

[Prarit Bhargava <prarit@redhat.com>]

On 01/13/2014 07:33 PM, Toshi Kani wrote:
> On Mon, 2014-01-13 at 18:39 -0500, Prarit Bhargava wrote:
>>
>> On 01/13/2014 03:31 PM, KOSAKI Motohiro wrote:
>>> On Sun, Jan 12, 2014 at 6:46 PM, Prarit Bhargava <prarit@redhat.com> wrote:
>>>>
>>>>
>>>> On 01/11/2014 11:35 AM, 7eggert@gmx.de wrote:
>>>>>
>>>>>
>>>>> On Fri, 10 Jan 2014, Prarit Bhargava wrote:
>>>>>
>>>>>> kdump uses memmap=exactmap and mem=X values to configure the memory
>>>>>> mapping for the kdump kernel.  If memory is hotadded during the boot of
>>>>>> the kdump kernel it is possible that the page tables for the new memory
>>>>>> cause the kdump kernel to run out of memory.
>>>>>>
>>>>>> Since the user has specified a specific mapping ACPI Memory Hotplug should be
>>>>>> disabled in this case.
>>>>>
>>>>> I'll ask just in case: Is it possible to want memory hotplug in spite of
>>>>> using memmap=exactmap or mem=X?
>>>>
>>>> Good question -- I can't think of a case.  When a user specifies "memmap" or
>>>> "mem" IMO they are asking for a very specific memory configuration.  Having
>>>> extra memory added above what the user has specified seems to defeat the purpose
>>>> of "memmap" and "mem".
>>>
>>> May be yes, may be no.
>>>
>>> They are often used for a wrokaround to avoid broken firmware issue.
>>> If we have no way
>>> to explicitly enable hotplug. We will lose a workaround.
>>>
>>> Perhaps, there is no matter. Today, memory hotplug is only used on
>>> high-end machine
>>> and their firmware is carefully developped and don't have a serious
>>> issue almostly. Though.
>>
>> Oof -- sorry Kosaki :(  I didn't see this until just now (and your subsequent
>> ACK on the updated patch).
>>
>> I just remembered that we did have a processor vendor's whitebox that would not
>> boot unless we specified a specific memmap and we did specify memmap=exactmap to
>> boot the system correctly and the system had hotplug memory.
>>
>> So it means that I should not key off of "memmap=exactmap".
> 
> I do not think it makes sense.  You needed memmap=exactmap as a
> workaround because the kernel did not boot with the firmware's memory
> info.  So, it's broken, and you requested the kernel to ignore the
> firmware info.
> 
> Why do you think memory hotplug needs to be supported under such
> condition, which has to use the broken firmware info?

There was a memory address region that was not in the e820 map that had to be
reserved for a specific device's use.  It was not so we used memmap=exactmap and
other memmap entries to "rewrite" the e820 map to see if the system would boot;
then I filed a bug against the hardware vendor ;)

So, yes, in that case I did want memory hotplug & memmap=exactmap.  Admittedly
this was a rare corner case, and I certainly could have recompiled the kernel.

P.

> 
> Thanks,
> -Toshi 
> 
> 
> 

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>