Re: [RESEND PATCH] x86/mm: only allow memmap=XX!YY over existing RAM

[Dan Williams]

On Tue, Jun 28, 2016 at 10:58 AM, H. Peter Anvin <hpa@xxxxxxxxx> wrote:
> On 06/28/16 09:33, Dan Williams wrote:
>> On Tue, Jun 28, 2016 at 1:31 AM, Yigal Korman <yigal@xxxxxxxxxxxxx> wrote:
>>> Before this patch, passing a range that is beyond the physical memory
>>> range will succeed, the user will see a /dev/pmem0 and will be able to
>>> access it. Reads will always return 0 and writes will be silently
>>> ignored.
>>>
>>> I've gotten more than one bug report about mkfs.{xfs,ext4} or nvml
>>> failing that were eventually tracked down to be wrong values passed to
>>> memmap.
>>>
>>> This patch prevents the above issue by instead of adding a new memory
>>> range, only update a RAM memory range with the PRAM type. This way,
>>> passing the wrong memmap will either not give you a pmem at all or give
>>> you a smaller one that actually has RAM behind it.
>>>
>>> And if someone still needs to fake a pmem that doesn't have RAM behind
>>> it, they can simply do memmap=XX@YY,XX!YY.
>>>
>>> Signed-off-by: Yigal Korman <yigal@xxxxxxxxxxxxx>
>>> Acked-by: Dan Williams <dan.j.williams@xxxxxxxxx>
>>> Acked-by: Johannes Thumshirn <jthumshirn@xxxxxxx>
>>> Tested-by: Boaz Harrosh <boaz@xxxxxxxxxxxxx>
>>> ---
>>
>> I have some other libnvdimm fixes heading upstream shortly if x86
>> folks just want to ack this one...
>>
>
> I'm concerned about this.  This would mean that memory not marked as RAM
> because it is persistent and you don't want the OS to accidentally
> clobber persistent RAM can't be added.

Ah true.  Specifically you are worried about the case where a
platform-firmware has mis-marked pmem as reserved memory (or some
other type) and would like to correct it to be pram.

> So it seems like The Wrong
> Thing.  If all you want is simulated pram then it shouldn't care about
> addresses in the first place and instead we should just specify it by
> quantity.

Yes, agree we need an explicit "simulate pram" option independent of
memmap=, or just continue to educate users that if they try to
simulate pmem and specify an invalid range they get to keep all the
broken pieces.