[PATCH v5 01/14] memory-hotplug: try to offline the memory twice to avoid dependence

[Glauber Costa]

On 02/06/2013 02:10 PM, Tang Chen wrote:
> On 02/06/2013 05:17 PM, Tang Chen wrote:
>> Hi all,
>>
>> On 02/06/2013 11:07 AM, Tang Chen wrote:
>>> Hi Glauber, all,
>>>
>>> An old thing I want to discuss with you. :)
>>>
>>> On 01/09/2013 11:09 PM, Glauber Costa wrote:
>>>>>>> memory can't be offlined when CONFIG_MEMCG is selected.
>>>>>>> For example: there is a memory device on node 1. The address range
>>>>>>> is [1G, 1.5G). You will find 4 new directories memory8, memory9,
>>>>>>> memory10,
>>>>>>> and memory11 under the directory /sys/devices/system/memory/.
>>>>>>>
>>>>>>> If CONFIG_MEMCG is selected, we will allocate memory to store page
>>>>>>> cgroup
>>>>>>> when we online pages. When we online memory8, the memory stored
>>>>>>> page cgroup
>>>>>>> is not provided by this memory device. But when we online memory9,
>>>>>>> the memory
>>>>>>> stored page cgroup may be provided by memory8. So we can't offline
>>>>>>> memory8
>>>>>>> now. We should offline the memory in the reversed order.
>>>>>>>
>>>>>>> When the memory device is hotremoved, we will auto offline memory
>>>>>>> provided
>>>>>>> by this memory device. But we don't know which memory is onlined
>>>>>>> first, so
>>>>>>> offlining memory may fail. In such case, iterate twice to offline
>>>>>>> the memory.
>>>>>>> 1st iterate: offline every non primary memory block.
>>>>>>> 2nd iterate: offline primary (i.e. first added) memory block.
>>>>>>>
>>>>>>> This idea is suggested by KOSAKI Motohiro.
>>>>>>>
>>>>>>> Signed-off-by: Wen Congyang<wency at cn.fujitsu.com>
>>>>>>
>>>>>> Maybe there is something here that I am missing - I admit that I came
>>>>>> late to this one, but this really sounds like a very ugly hack, that
>>>>>> really has no place in here.
>>>>>>
>>>>>> Retrying, of course, may make sense, if we have reasonable belief
>>>>>> that
>>>>>> we may now succeed. If this is the case, you need to document - in
>>>>>> the
>>>>>> code - while is that.
>>>>>>
>>>>>> The memcg argument, however, doesn't really cut it. Why can't we make
>>>>>> all page_cgroup allocations local to the node they are describing? If
>>>>>> memcg is the culprit here, we should fix it, and not retry. If
>>>>>> there is
>>>>>> still any benefit in retrying, then we retry being very specific
>>>>>> about why.
>>>>>
>>>>> We try to make all page_cgroup allocations local to the node they are
>>>>> describing
>>>>> now. If the memory is the first memory onlined in this node, we will
>>>>> allocate
>>>>> it from the other node.
>>>>>
>>>>> For example, node1 has 4 memory blocks: 8-11, and we online it from 8
>>>>> to 11
>>>>> 1. memory block 8, page_cgroup allocations are in the other nodes
>>>>> 2. memory block 9, page_cgroup allocations are in memory block 8
>>>>>
>>>>> So we should offline memory block 9 first. But we don't know in which
>>>>> order
>>>>> the user online the memory block.
>>>>>
>>>>> I think we can modify memcg like this:
>>>>> allocate the memory from the memory block they are describing
>>>>>
>>>>> I am not sure it is OK to do so.
>>>>
>>>> I don't see a reason why not.
>>>>
>>>> You would have to tweak a bit the lookup function for page_cgroup, but
>>>> assuming you will always have the pfns and limits, it should be easy
>>>> to do.
>>>>
>>>> I think the only tricky part is that today we have a single
>>>> node_page_cgroup, and we would of course have to have one per memory
>>>> block. My assumption is that the number of memory blocks is limited and
>>>> likely not very big. So even a static array would do.
>>>>
>>>
>>> About the idea "allocate the memory from the memory block they are
>>> describing",
>>>
>>> online_pages()
>>> |-->memory_notify(MEM_GOING_ONLINE, &arg) ----------- memory of this
>>> section is not in buddy yet.
>>> |-->page_cgroup_callback()
>>> |-->online_page_cgroup()
>>> |-->init_section_page_cgroup()
>>> |-->alloc_page_cgroup() --------- allocate page_cgroup from buddy
>>> system.
>>>
>>> When onlining pages, we allocate page_cgroup from buddy. And the being
>>> onlined pages are not in
>>> buddy yet. I think we can reserve some memory in the section for
>>> page_cgroup, and return all the
>>> rest to the buddy.
>>>
>>> But when the system is booting,
>>>
>>> start_kernel()
>>> |-->setup_arch()
>>> |-->mm_init()
>>> | |-->mem_init()
>>> | |-->numa_free_all_bootmem() -------------- all the pages are in buddy
>>> system.
>>> |-->page_cgroup_init()
>>> |-->init_section_page_cgroup()
>>> |-->alloc_page_cgroup() ------------------ I don't know how to reserve
>>> memory in each section.
>>>
>>> So any idea about how to deal with it when the system is booting please?
>>>
>>
>> How about this way.
>>
>> 1) Add a new flag PAGE_CGROUP_INFO, like SECTION_INFO and
>> MIX_SECTION_INFO.
>> 2) In sparse_init(), reserve some beginning pages of each section as
>> bootmem.
> 
> Hi all,
> 
> After digging into bootmem code, I met another problem.
> 
> memblock allocates memory from high address to low address, using
> memblock.current_limit
> to remember where the upper limit is. What I am doing will produce a lot
> of fragments,
> and the memory will be non-contiguous. So we need to modify memblock again.
> 
> I don't think it's a good idea. How do you think ?
> 
> Thanks. :)
> 
>> 3) In register_page_bootmem_info_section(), set these pages as
>> page->lru.next = PAGE_CGROUP_INFO;
>>
>> Then these pages will not go to buddy system.
>>
>> But I do worry about the fragment problem because part of each section
>> will
>> be used in the very beginning.
>>
>> Thanks. :)
>>
>>>
>>> And one more question, a memory section is 128MB in Linux. If we reserve
>>> part of the them for page_cgroup,
>>> then anyone who wants to allocate a contiguous memory larger than 128MB,
>>> it will fail, right ?
>>> Is it OK ?
No, it is not.

Another take on this: Can't we free all the page_cgroup structure before
we actually start removing the sections ? If we do this, we would be
basically left with no problem at all, since when your code starts
running we would no longer have any page_cgroup allocated.

All you have to guarantee is that it happens after the memory block is
already isolated and allocations no longer can reach it.

What do you think ?