On 27.01.22 10:41, Jonghyeon Kim wrote:
> On Wed, Jan 26, 2022 at 06:04:50PM +0100, David Hildenbrand wrote:
>> On 26.01.22 18:00, Jonghyeon Kim wrote:
>>> Export shrink_zone_span() and update_pgdat_span() functions to head
>>> file. We need to update real number of spanned pages for NUMA nodes and
>>> zones when we add memory device node such as device dax memory.
>>>
>>
>> Can you elaborate a bit more what you intend to fix?
>>
>> Memory onlining/offlining is reponsible for updating the node/zone span,
>> and that's triggered when the dax/kmem mamory gets onlined/offlined.
>>
> Sure, sorry for the lack of explanation of the intended fix.
>
> Before onlining nvdimm memory using dax(devdax or fsdax), these memory belong to
> cpu NUMA nodes, which extends span pages of node/zone as a ZONE_DEVICE. So there
> is no problem because node/zone contain these additional non-visible memory
> devices to the system.
> But, if we online dax-memory, zone[ZONE_DEVICE] of CPU NUMA node is hot-plugged
> to new NUMA node(but CPU-less). I think there is no need to hold
> zone[ZONE_DEVICE] pages on the original node.
>
> Additionally, spanned pages are also used to calculate the end pfn of a node.
> Thus, it is needed to maintain accurate page stats for node/zone.
>
> My machine contains two CPU-socket consisting of DRAM and Intel DCPMM
> (DC persistent memory modules) with App-Direct mode.
>
> Below are my test results.
>
> Before memory onlining:
>
> # ndctl create-namespace --mode=devdax
> # ndctl create-namespace --mode=devdax
> # cat /proc/zoneinfo | grep -E "Node|spanned" | paste - -
> Node 0, zone DMA spanned 4095
> Node 0, zone DMA32 spanned 1044480
> Node 0, zone Normal spanned 7864320
> Node 0, zone Movable spanned 0
> Node 0, zone Device spanned 66060288
> Node 1, zone DMA spanned 0
> Node 1, zone DMA32 spanned 0
> Node 1, zone Normal spanned 8388608
> Node 1, zone Movable spanned 0
> Node 1, zone Device spanned 66060288
>
> After memory onlining:
>
> # daxctl reconfigure-device --mode=system-ram --no-online dax0.0
> # daxctl reconfigure-device --mode=system-ram --no-online dax1.0
>
> # cat /proc/zoneinfo | grep -E "Node|spanned" | paste - -
> Node 0, zone DMA spanned 4095
> Node 0, zone DMA32 spanned 1044480
> Node 0, zone Normal spanned 7864320
> Node 0, zone Movable spanned 0
> Node 0, zone Device spanned 66060288
> Node 1, zone DMA spanned 0
> Node 1, zone DMA32 spanned 0
> Node 1, zone Normal spanned 8388608
> Node 1, zone Movable spanned 0
> Node 1, zone Device spanned 66060288
> Node 2, zone DMA spanned 0
> Node 2, zone DMA32 spanned 0
> Node 2, zone Normal spanned 65011712
> Node 2, zone Movable spanned 0
> Node 2, zone Device spanned 0
> Node 3, zone DMA spanned 0
> Node 3, zone DMA32 spanned 0
> Node 3, zone Normal spanned 65011712
> Node 3, zone Movable spanned 0
> Node 3, zone Device spanned 0
>
> As we can see, Node 0 and 1 still have zone_device pages after memory onlining.
> This causes problem that Node 0 and Node 2 have same end of pfn values, also
> Node 1 and Node 3 have same problem.
Thanks for the information, that makes it clearer.
While this unfortunate, the node/zone span is something fairly
unreliable/unusable for user space. Nodes and zones can overlap just easily.
What counts are present/managed pages in the node/zone.
So at least I don't count this as something that "needs fixing",
it's more something that's nice to handle better if easily possible.
See below.
>
>>> Signed-off-by: Jonghyeon Kim <tome01@xxxxxxxxxx>
>>> ---
>>> include/linux/memory_hotplug.h | 3 +++
>>> mm/memory_hotplug.c | 6 ++++--
>>> 2 files changed, 7 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/include/linux/memory_hotplug.h b/include/linux/memory_hotplug.h
>>> index be48e003a518..25c7f60c317e 100644
>>> --- a/include/linux/memory_hotplug.h
>>> +++ b/include/linux/memory_hotplug.h
>>> @@ -337,6 +337,9 @@ extern void move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
>>> extern void remove_pfn_range_from_zone(struct zone *zone,
>>> unsigned long start_pfn,
>>> unsigned long nr_pages);
>>> +extern void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
>>> + unsigned long end_pfn);
>>> +extern void update_pgdat_span(struct pglist_data *pgdat);
>>> extern bool is_memblock_offlined(struct memory_block *mem);
>>> extern int sparse_add_section(int nid, unsigned long pfn,
>>> unsigned long nr_pages, struct vmem_altmap *altmap);
>>> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
>>> index 2a9627dc784c..38f46a9ef853 100644
>>> --- a/mm/memory_hotplug.c
>>> +++ b/mm/memory_hotplug.c
>>> @@ -389,7 +389,7 @@ static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
>>> return 0;
>>> }
>>>
>>> -static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
>>> +void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
>>> unsigned long end_pfn)
>>> {
>>> unsigned long pfn;
>>> @@ -428,8 +428,9 @@ static void shrink_zone_span(struct zone *zone, unsigned long start_pfn,
>>> }
>>> }
>>> }
>>> +EXPORT_SYMBOL_GPL(shrink_zone_span);
>>
>> Exporting both as symbols feels very wrong. This is memory
>> onlining/offlining internal stuff.
>
> I agree with you that your comment. I will find another approach to avoid
> directly using onlining/offlining internal stuff while updating node/zone span.
IIRC, to handle what you intend to handle properly want to look into teaching
remove_pfn_range_from_zone() to handle zone_is_zone_device().
There is a big fat comment:
/*
* Zone shrinking code cannot properly deal with ZONE_DEVICE. So
* we will not try to shrink the zones - which is okay as
* set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
*/
if (zone_is_zone_device(zone))
return;
Similarly, try_offline_node() spells this out:
/*
* If the node still spans pages (especially ZONE_DEVICE), don't
* offline it. A node spans memory after move_pfn_range_to_zone(),
* e.g., after the memory block was onlined.
*/
if (pgdat->node_spanned_pages)
return;
So once you handle remove_pfn_range_from_zone() cleanly, you'll cleanly handle
try_offline_node() implicitly.
Trying to update the node span manually without teaching node/zone shrinking code how to
handle ZONE_DEVICE properly is just a hack that will only sometimes work. Especially, it
won't work if the range of interest is still surrounded by other ranges.
--
Thanks,
David / dhildenb
