[RFC PATCH 1/2] ARM: mm: support memory-failure

[David Hildenbrand (Red Hat)]

On 23.09.25 06:10, Xie Yuanbin wrote:
> Arnd Bergmann wrote:
>>>> It would be helpful to be more specific about what you
>>>> want to do with this.
>>>>
>>>> Are you working on a driver that would actually make use of
>>>> the exported interface?
>>>
>>> Thanks for your reply.
>>>
>>> Yes, In fact, we have developed a hardware component to detect DDR bit
>>> transitions (software does not sense the detection behavior). Once a bit
>>> transition is detected, an interrupt is reported to the CPU.
>>>
>>> On the software side, we have developed a driver module ko to register
>>> the interrupt callback to perform soft page offline to the corresponding
>>> physical pages.
>>>
>>> In fact, we will export `soft_offline_page` for ko to use (we can ensure
>>> that it is not called in the interrupt context), but I have looked at the
>>> code and found that `memory_failure_queue` and `memory_failure` can also
>>> be used, which are already exported.
>>
>> Ok
>>
>>>> I see only a very small number of
>>>> drivers that call memory_failure(), and none of them are
>>>> usable on Arm.
>>>
>>> I think that not all drivers are in the open source kernel code.
>>> As far as I know, there should be similar third-party drivers in other
>>> architectures that use memory-failure functions, like x86 or arm64.
>>> I am not a specialist in drivers, so if I have made any mistakes,
>>> please correct me.
>>
>> I'm not familiar with the memory-failure support, but this sounds
>> like something that is usually done with a drivers/edac/ driver.
>> There are many SoC specific drivers, including for 32-bit Arm
>> SoCs.
>>
>> Have you considered adding an EDAC driver first? I don't know
>> how the other platforms that have EDAC drivers handle failures,
>> but I would assume that either that subsystem already contains
>> functionality for taking pages offline,
> 
> I'm very sorry, I tried my best to do this,
> but it seems impossible to achieve.
> I am a kernel developer rathder than a driver developer. I have tried to
> communicate with driver developers, but open source is very difficult due
> to the involvement of proprietary hardware and algorithms.
> 
>> or this is something
>> that should be done in a way that works for all of them without
>> requiring an extra driver.
> 
> Yes, I think that the memory-failure feature should not be associated with
> specific architectures or drivers.
> 
> I have read the memory-failure's doc and code,
> and found the following features, which are user useable,
> are not associated with specific drivers:
> 
> 1. `/sys/devices/system/memory/soft_offline_page`:
> see https://www.kernel.org/doc/Documentation/ABI/testing/sysfs-memory-page-offline
> 
> This interface only exists when CONFIG_MEMORY_HOTPLUG is enabled, but
> ARM cannot enable it.
> However, I have read the code and believe that it should not require a
> lot of effort to decouple these two, allowing the interface to exist
> even if mem-hotplug is disabled.

It's all about the /sys/devices/system/memory/ directory, which 
traditionally only made sense for memory hotplug. Well, still does to 
most degree.

Not sure whether some user space (chmem?) senses for 
/sys/devices/system/memory/ to detect memory hotplug capabilities.

But given soft_offline_page is a pure testing mechanism, I wouldn't be 
too concerned about that for now.

> 
> 2. The syscall madvise with `MADV_SOFT_OFFLINE/MADV_HWPOISON` flags:
> 
> According to the documentation, this interface is currently only used for
> testing. However, if the user program can map the specified physical
> address, it can actually be used for memory-failure.

It's mostly a testing-only interface. It could be used for other things, 
but really detecting MCE and handling it properly is kernel responsibility.

> 
> 3. The CONFIG_HWPOISON_INJECT which depends on CONFIG_MEMORY_FAILURE:
> see https://docs.kernel.org/mm/hwpoison.html
> 
> It seems to allow input of physical addresses and trigger memory-failure,
> but according to the doc, it seems to be used only for testing.

Right, all these interfaces are testing only.

> 
> 
> Additionally, I noticed that in the memory-failure doc
> https://docs.kernel.org/mm/hwpoison.html, it mentions that
> "The main target right now is KVM guests, but it works for all kinds of
> applications." This seems to confirm my speculation that the
> memory-failure feature should not be associated with specific
> architectures or drivers.

Can you go into more details which exact functionality in 
memory-failure.c you would be interested in using?

Only soft-offlining or also the other (possibly architecture-specific) 
handling?

-- 
Cheers

David