Re: [PATCH V11 1/5] mm/hotplug: Introduce arch callback validating the hot remove range

[Anshuman Khandual]

On 01/13/2020 04:07 PM, David Hildenbrand wrote:
> On 13.01.20 10:50, Anshuman Khandual wrote:
>>
>>
>> On 01/13/2020 02:44 PM, David Hildenbrand wrote:
>>>
>>>
>>>> Am 13.01.2020 um 10:10 schrieb Anshuman Khandual <anshuman.khandual@xxxxxxx>:
>>>>
>>>> ï
>>>>
>>>>> On 01/10/2020 02:12 PM, David Hildenbrand wrote:
>>>>>> On 10.01.20 04:09, Anshuman Khandual wrote:
>>>>>> Currently there are two interfaces to initiate memory range hot removal i.e
>>>>>> remove_memory() and __remove_memory() which then calls try_remove_memory().
>>>>>> Platform gets called with arch_remove_memory() to tear down required kernel
>>>>>> page tables and other arch specific procedures. But there are platforms
>>>>>> like arm64 which might want to prevent removal of certain specific memory
>>>>>> ranges irrespective of their present usage or movability properties.
>>>>>
>>>>> Why? Is this only relevant for boot memory? I hope so, otherwise the
>>>>> arch code needs fixing IMHO.
>>>>
>>>> Right, it is relevant only for the boot memory on arm64 platform. But this
>>>> new arch callback makes it flexible to reject any given memory range.
>>>>
>>>>>
>>>>> If it's only boot memory, we should disallow offlining instead via a
>>>>> memory notifier - much cleaner.
>>>>
>>>> Dont have much detail understanding of MMU notifier mechanism but from some
>>>> initial reading, it seems like we need to have a mm_struct for a notifier
>>>> to monitor various events on the page table. Just wondering how a physical
>>>> memory range like boot memory can be monitored because it can be used both
>>>> for for kernel (init_mm) or user space process at same time. Is there some
>>>> mechanism we could do this ?
>>>>
>>>>>
>>>>>>
>>>>>> Current arch call back arch_remove_memory() is too late in the process to
>>>>>> abort memory hot removal as memory block devices and firmware memory map
>>>>>> entries would have already been removed. Platforms should be able to abort
>>>>>> the process before taking the mem_hotplug_lock with mem_hotplug_begin().
>>>>>> This essentially requires a new arch callback for memory range validation.
>>>>>
>>>>> I somewhat dislike this very much. Memory removal should never fail if
>>>>> used sanely. See e.g., __remove_memory(), it will BUG() whenever
>>>>> something like that would strike.
>>>>>
>>>>>>
>>>>>> This differentiates memory range validation between memory hot add and hot
>>>>>> remove paths before carving out a new helper check_hotremove_memory_range()
>>>>>> which incorporates a new arch callback. This call back provides platforms
>>>>>> an opportunity to refuse memory removal at the very onset. In future the
>>>>>> same principle can be extended for memory hot add path if required.
>>>>>>
>>>>>> Platforms can choose to override this callback in order to reject specific
>>>>>> memory ranges from removal or can just fallback to a default implementation
>>>>>> which allows removal of all memory ranges.
>>>>>
>>>>> I suspect we want really want to disallow offlining instead. E.g., I
>>>>
>>>> If boot memory pages can be prevented from being offlined for sure, then it
>>>> would indirectly definitely prevent hot remove process as well.
>>>>
>>>>> remember s390x does that with certain areas needed for dumping/kexec.
>>>>
>>>> Could not find any references to mmu_notifier in arch/s390 or any other arch
>>>> for that matter apart from KVM (which has an user space component), could you
>>>> please give some pointers ?
>>>
>>> Memory (hotplug) notifier, not MMU notifier :)
>>
>> They are so similarly named :)
>>
>>>
>>> Not on my notebook right now, grep for MEM_GOING_OFFLINE, that should be it.
>>>
>>
>> Got it, thanks ! But we will still need boot memory enumeration via MEMBLOCK_BOOT
>> to reject affected offline requests in the callback.
> 
> Do you really need that?
> 
> We have SECTION_IS_EARLY. You could iterate all involved sections (for
> which you are getting notified) and check if any one of these is marked
> SECTION_IS_EARLY. then, it was added during boot and not via add_memory().

Seems to be a better approach than adding a new memblock flag.

> 
>