Re: Cannot hot remove a memory device

[Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>]

(2013/08/04 9:37), Toshi Kani wrote:
> On Sat, 2013-08-03 at 03:01 +0200, Rafael J. Wysocki wrote:
>> On Friday, August 02, 2013 06:04:40 PM Toshi Kani wrote:
>>> On Sat, 2013-08-03 at 01:43 +0200, Rafael J. Wysocki wrote:
>>>> On Friday, August 02, 2013 03:46:15 PM Toshi Kani wrote:
>>>>> On Thu, 2013-08-01 at 23:43 +0200, Rafael J. Wysocki wrote:
>>>>>> Hi,
>>>>>>
>>>>>> Thanks for your report.
>>>>>>
>>>>>> On Thursday, August 01, 2013 05:37:21 PM Yasuaki Ishimatsu wrote:
>>>>>>> By following commit, I cannot hot remove a memory device.
>>>>>>>
>>>>>>> ACPI / memhotplug: Bind removable memory blocks to ACPI device nodes
>>>>>>> commit e2ff39400d81233374e780b133496a2296643d7d
>>>>>>>
>>>>>>> Details are follows:
>>>>>>> When I add a memory device, acpi_memory_enable_device() always fails
>>>>>>> as follows:
>>>>>>>
>>>>>>> ...
>>>>>>> [ 1271.114116]  [ffffea121c400000-ffffea121c7fffff] PMD -> [ffff880813c00000-ffff880813ffffff] on node 3
>>>>>>> [ 1271.128682]  [ffffea121c800000-ffffea121cbfffff] PMD -> [ffff880813800000-ffff880813bfffff] on node 3
>>>>>>> [ 1271.143298]  [ffffea121cc00000-ffffea121cffffff] PMD -> [ffff880813000000-ffff8808133fffff] on node 3
>>>>>>> [ 1271.157799]  [ffffea121d000000-ffffea121d3fffff] PMD -> [ffff880812c00000-ffff880812ffffff] on node 3
>>>>>>> [ 1271.172341]  [ffffea121d400000-ffffea121d7fffff] PMD -> [ffff880812800000-ffff880812bfffff] on node 3
>>>>>>> [ 1271.186872]  [ffffea121d800000-ffffea121dbfffff] PMD -> [ffff880812400000-ffff8808127fffff] on node 3
>>>>>>> [ 1271.201481]  [ffffea121dc00000-ffffea121dffffff] PMD -> [ffff880812000000-ffff8808123fffff] on node 3
>>>>>>> [ 1271.216041]  [ffffea121e000000-ffffea121e3fffff] PMD -> [ffff880811c00000-ffff880811ffffff] on node 3
>>>>>>> [ 1271.230623]  [ffffea121e400000-ffffea121e7fffff] PMD -> [ffff880811800000-ffff880811bfffff] on node 3
>>>>>>> [ 1271.245148]  [ffffea121e800000-ffffea121ebfffff] PMD -> [ffff880811400000-ffff8808117fffff] on node 3
>>>>>>> [ 1271.259683]  [ffffea121ec00000-ffffea121effffff] PMD -> [ffff880811000000-ffff8808113fffff] on node 3
>>>>>>> [ 1271.274194]  [ffffea121f000000-ffffea121f3fffff] PMD -> [ffff880810c00000-ffff880810ffffff] on node 3
>>>>>>> [ 1271.288764]  [ffffea121f400000-ffffea121f7fffff] PMD -> [ffff880810800000-ffff880810bfffff] on node 3
>>>>>
>>>>> It appears that each memory object only has 64MB of memory.  This is
>>>>> less than the memory block size, which is 128MB.  This means that a
>>>>> single memory block associates with two ACPI memory device objects.
>>>>
>>>> That'd be bad.
>>>>
>>>> How did that work before if that indeed is the case?
>>>
>>> Well, it looks to me that it has never worked before...
>>>
>>>>>>> ...	
>>>>>>> [ 1271.325841] acpi PNP0C80:03: acpi_memory_enable_device() error
>>>>>>
>>>>>> Well, the only new way acpi_memory_enable_device() can fail after that commit
>>>>>> is a failure in acpi_bind_memory_blocks().
>>>>>
>>>>> Agreed.
>>>>>
>>>>>> This means that either handle is NULL, which I think we can exclude, because
>>>>>> acpi_memory_enable_device() wouldn't be called at all if that were the case, or
>>>>>> there's a more subtle error in acpi_bind_one().
>>>>>>
>>>>>> One that comes to mind is that we may be calling acpi_bind_one() twice for the
>>>>>> same memory region, in which it will trigger -EINVAL from the sanity check in
>>>>>> there.
>>>>>
>>>>> I think it fails with -EINVAL at the place with dev_warn(dev, "ACPI
>>>>> handle is already set\n").  When two ACPI memory objects associate with
>>>>> a same memory block, the bind procedure of the 2nd ACPI memory object
>>>>> sees that ACPI_HANDLE(dev) is already set to the 1st ACPI memory object.
>>>>
>>>> That sound's plausible, but I wonder how we can fix that?
>>>>
>>>> There's no way for a single physical device to have two different ACPI
>>>> "companions".  It looks like the memory blocks should be 64 M each in that
>>>> case.  Or we need to create two child devices for each memory block and
>>>> associate each of them with an ACPI object.  That would lead to complications
>>>> in the user space interface, though.
>>>
>>> Right.  Even bigger issue is that I do not think __add_pages() and
>>> __remove_pages() can add / delete a memory chunk that is less than
>>> 128MB.  128MB is the granularity of them.  So, we may just have to fail
>>> this case gracefully.
>>
>> Sigh.
>>
>> BTW, why do you think they are 64 M each (it's late and I'm obviously tired)?
>
> Oops!  Sorry, I had confused the above messages with the one in
> init_memory_mapping(), which shows a memory range being added, i.e. the
> size of an ACPI memory device object.  But the above messages actually
> came from vmemmap_populate_hugepages(), which was called during boot-up.
> So, these messages have nothing to do with ACPI memory device objects.
> And even worse, I do not seem to be able to count a number of zeros...
> In the above messages, each memory range is 4MB (0x400000), not 64MB
> (0x4000000)...  My bad. :-(
>
> So, while we may still need to do something for the less-than-128MB
> issue, Yasuaki may be hitting a different one.  Let's wait for Yasuaki
> to give us more info.

acpi_bind_memory_blocks() failed with -ENOSPC.

int acpi_bind_one(struct device *dev, acpi_handle handle)
{
...
	/* allocate physical node id according to physical_node_id_bitmap */
	physical_node->node_id =
		find_first_zero_bit(acpi_dev->physical_node_id_bitmap,
		ACPI_MAX_PHYSICAL_NODE);
	if (physical_node->node_id >= ACPI_MAX_PHYSICAL_NODE) {
		retval = -ENOSPC; => here
		goto err_free;
	}

When adding memory device, acpi_bind_memroy_blocks() calls acpi_bind_one()
"memory device size / 128MiB" times. So ACPI_MAX_PHYSICAL_NODE need to
be set "memory device size / 128MiB" or more. But ACPI_MAX_PHYSICAL_NODE is 32.
So acpi_bind_memory_blocks() always failed with -ENOSPC.

I'll test again after increasing ACPI_MAX_PHYSICAL_NODE to enough size.

Thanks,
Yasuaki Ishimatsu

>
> Thanks,
> -Toshi
>
>
>
>
>
>
>
>
>