Message-ID: <5847F0AA.5050109@huawei.com>
Date:   Wed, 7 Dec 2016 19:21:14 +0800
From:   Xishi Qiu <qiuxishi@...wei.com>
To:     Scott Branden <scott.branden@...adcom.com>
CC:     Arnd Bergmann <arnd@...db.de>,
        Russell King <linux@...linux.org.uk>,
        Catalin Marinas <catalin.marinas@....com>,
        Will Deacon <will.deacon@....com>,
        Ard Biesheuvel <ard.biesheuvel@...aro.org>,
        Mark Rutland <mark.rutland@....com>, <bielski@...tmail.net>,
        BCM Kernel Feedback <bcm-kernel-feedback-list@...adcom.com>,
        Tang Chen <tangchen@...fujitsu.com>,
        <linux-arm-kernel@...ts.infradead.org>,
        <linux-kernel@...r.kernel.org>, "Luck, Tony" <tony.luck@...el.com>
Subject: Re: [RFC PATCH 0/2] arm64: memory-hotplug: Add Memory Hotplug support

On 2016/12/7 16:43, Scott Branden wrote:

> Hi Xishi,
> 
> I followed you suggestions and found pfn_valid is always true.  Answers to your questions inline.
> 
> I could keep debugging this but hope Marcin sends out some code - I'm quite willing to test and help clean up the patchset.
> 
> On 16-12-01 07:11 PM, Xishi Qiu wrote:
>> On 2016/12/2 10:38, Scott Branden wrote:
>>
>>> Hi Xishi,
>>>
>>> Thanks for the reply - please see comments below.
>>>
>>> On 16-12-01 05:49 PM, Xishi Qiu wrote:
>>>> On 2016/12/2 8:19, Scott Branden wrote:
>>>>
>>>>> This patchset is sent for comment to add memory hotplug support for ARM64
>>>>> based platforms.  It follows hotplug code added for other architectures
>>>>> in the linux kernel.
>>>>>
>>>>> I tried testing the memory hotplug feature following documentation from
>>>>> Documentation/memory-hotplug.txt.  I don't think it is working as expected
>>>>> - see below:
>>>>>
>>>>> To add memory to the system I did the following:
>>>>> echo 0x400000000 > /sys/devices/system/memory/probe
>>>>>
>>>>> The memory is displayed as system ram:
>>>>> cat /proc/iomem:
>>>>> 74000000-77ffffff : System RAM
>>>>>   74080000-748dffff : Kernel code
>>>>>   74950000-749d2fff : Kernel data
>>>>> 400000000-43fffffff : System RAM
>>>>>
>>>>> But does not seem to be added to the kernel memory.
>>>>> /proc/meminfo did not change.
>>>>>
>>>>> What else needs to be done so the memory is added to the kernel memory
>>>>> pool for normal allocation?
>>>>>
>>>>
>>>> Hi Scott,
>>>>
>>>> Do you mean it still don't support hod-add after apply this patchset?
>>>
>>> After applying the patch it appears to partially support hot-add. Please let me know if you think it is working as expected?
>>>
>>> The memory probe functions in that the memory is registered with the system and shows up in /proc/iomem.  But, the memory is not available in /proc/meminfo.  Do you think something else needs to be adjusted for ARM64 to hotadd the memory
>>>
>>> I just found another clue:
>>> under /sys/devices/system/memory I only see one memory entry (before or after I try to hotadd additional memory).
>>>
>>> /sys/devices/system/memory # ls
>>> auto_online_blocks  memory0            uevent
>>> block_size_bytes    probe
>>>
>>> In arch/arm64/include/asm/sparsemem.h if I change SECTION_SIZE_BITS from 30 to 28 and recompile I get the following:
>>> /sys/devices/system/memory # ls
>>> auto_online_blocks  memory7            uevent
>>> block_size_bytes    probe
>>>
>>>
>>> In arch/arm64/include/asm/sparsemem.h if I change SECTION_SIZE_BITS from 30 to 27 and recompile I get the following:
>>> /sys/devices/system/memory # ls
>>> auto_online_blocks  memory14            uevent
>>> block_size_bytes    probe
>>>
>>> If looks to me like something is not working properly in the ARM64 implementation.  I should expect to see multiple memoryX entries under /sys/devices/system/memory?
>>>
>>
>> Hi Scott,
>>
>> 1. Do you enable the following configs?
>> CONFIG_SPARSEMEM
>> MEMORY_HOTPLUG
>> CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
> Yes, these configs are enabled
>>
>> 2. I find you missed create mapping in arch_add_memory(), and x86 has it.
> Could you please explain this further?  The patch I submitted hass arch_add_memory identical to the ia64 implementation.

Hi Scott,

I think we should create page table first for the new hotadd memory.
e.g. create_mapping_late(start, __phys_to_virt(start), size, PAGE_KERNEL);

I don't know why ia64 don't have this step.
CC Tony

>>
>> 3. We will add memblock first, so pfn_valid() maybe always return true(in the
>> following function), and this will lead __add_section() failed. Please check
>> it.
> You are correct - pfn_valid always returns true.  The function is in arch/arm64/mm/init.c and different than the one you indicated below:
> 

> #ifdef CONFIG_HAVE_ARCH_PFN_VALID
> int pfn_valid(unsigned long pfn)
> {
>     return memblock_is_map_memory(pfn << PAGE_SHIFT);
> }
> EXPORT_SYMBOL(pfn_valid);
> #endif
> 
>>
>> int pfn_valid(unsigned long pfn)
>> {
>>     return (pfn & PFN_MASK) == pfn && memblock_is_memory(pfn << PAGE_SHIFT);
>> }
>>
>> add_memory
>>   add_memory_resource
>>     memblock_add_node
>>       arch_add_memory
>>         __add_pages
>>           __add_section
>>             pfn_valid
>>
>> Thanks,
>> Xishi Qiu
>>
>>>
>>>
>>>>
>>>> Thanks,
>>>> Xishi Qiu
>>>>
>>>>> Scott Branden (2):
>>>>>   arm64: memory-hotplug: Add MEMORY_HOTPLUG, MEMORY_HOTREMOVE,
>>>>>     MEMORY_PROBE
>>>>>   arm64: defconfig: enable MEMORY_HOTPLUG config options
>>>>>
>>>>>  arch/arm64/Kconfig           | 10 ++++++++++
>>>>>  arch/arm64/configs/defconfig |  3 +++
>>>>>  arch/arm64/mm/init.c         | 42 ++++++++++++++++++++++++++++++++++++++++++
>>>>>  3 files changed, 55 insertions(+)
>>>>>
>>>>
>>>>
>>>>
>>>
>>> .
>>>
>>
>>
>>
> 
> .
> 

Powered by blists - more mailing lists