Re: [pmem_attach_disk] WARNING: CPU: 46 PID: 518 at kernel/memremap.c:363 devm_memremap_pages+0x350/0x4b0

[Dan Williams]

On Mon, Oct 30, 2017 at 5:00 PM, Fengguang Wu <fengguang.wu@xxxxxxxxx> wrote:
> Hi Dan,
>
> On Mon, Oct 30, 2017 at 08:59:46AM -0700, Dan Williams wrote:
>>
>> On Mon, Oct 30, 2017 at 12:40 AM, Fengguang Wu <fengguang.wu@xxxxxxxxx>
>> wrote:
>>>
>>>
>>> CC nvdimm maintainers.
>>>
>>> On Sun, Oct 29, 2017 at 11:51:55PM +0100, Fengguang Wu wrote:
>>>>
>>>>
>>>> Hi Linus,
>>>>
>>>> Up to now we see the below boot error/warnings when testing v4.14-rc6.
>>>>
>>>> They hit the RC release mainly due to various imperfections in 0day's
>>>> auto bisection. So I manually list them here and CC the likely easy to
>>>> debug ones to the corresponding maintainers in the followup emails.
>>>>
>>>> boot_successes: 4700
>>>> boot_failures: 247
>>>
>>>
>>>
>>> [...]
>>>
>>>> WARNING:at_kernel/memremap.c:#devm_memremap_pages: 1
>>>
>>>
>>>
>>> Bisect failed, hope it's not hard to debug:
>>>
>>>         Start
>>> [   18.989316] devm_memremap_pages attempted on mixed region [mem
>>> 0x680000000-0x103dffffff flags 0x200]
>>
>>
>>
>> This appears to be a problem in the test environment. "Persistent
>> Memory" can only be specified on a minimum of a 128MB boundary if it
>> intersects "System RAM". Assuming I did my math right this appears to
>> end on 16MB boundary. Fixing this problem in the kernel would require
>> this patch set:
>>
>>    "mm: sub-section memory hotplug support":
>> https://lwn.net/Articles/707908/
>
>
> Good to know that!
>
>> ...but I have abandoned / pushed that to the back of my queue since
>> BIOS induced version of this problem does not appear to trigger in
>> practice. I assume this test is using memmap=ss!nn?
>
>
> Yes, we used
>
>        memmap=104G!26G memmap=104G!154G
> The warning showed up only once -- attached is the full dmesg.
>

Something is going wrong with memmap= because you are not getting 1G
aligned address ranges. I think you would have better luck switching
to the official nvdimm emulation in qemu-kvm rather than relying on
memmap= which is just a fragile / unreliable interface. In fact we
should look to deprecate it and point everyone to use the standard
methods. We just have a problem of legacy pre-ACPI6 platforms that
have no other way than a kernel command line to identify persistent
memory ranges.