Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Message-ID: <df613a5e-bf32-a03e-e06f-5dcb3444c3d4@redhat.com>
Date:   Fri, 28 Jan 2022 09:10:21 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.4.0
Content-Language: en-US
To:     Jonghyeon Kim <tome01@ajou.ac.kr>
Cc:     dan.j.williams@intel.com, vishal.l.verma@intel.com,
        dave.jiang@intel.com, akpm@linux-foundation.org,
        nvdimm@lists.linux.dev, linux-kernel@vger.kernel.org,
        linux-mm@kvack.org
References: <20220126170002.19754-1-tome01@ajou.ac.kr>
 <5d02ea0e-aca6-a64b-23de-bc9307572d17@redhat.com>
 <20220127094142.GA31409@swarm08>
 <696b782f-0b50-9861-a34d-cf750d4244bd@redhat.com>
 <20220128041959.GA20345@swarm08>
From:   David Hildenbrand <david@redhat.com>
Organization: Red Hat
Subject: Re: [PATCH 1/2] mm/memory_hotplug: Export shrink span functions for
 zone and node
In-Reply-To: <20220128041959.GA20345@swarm08>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 28.01.22 05:19, Jonghyeon Kim wrote:
> On Thu, Jan 27, 2022 at 10:54:23AM +0100, David Hildenbrand wrote:
>> 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@ajou.ac.kr>
>>>>> ---
>>>>>  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 for your pointing out, I missed those comments.
> I will keep trying to handle node/zone span updating process.

The only safe thing right now for on ZONE_DEVICE in
remove_pfn_range_from_zone() would be removing the given range from the
start/end of the zone range, but we must not scan using the existing
functions.

As soon as we start actual *scanning* via find_smallest...
find_biggest... in shrink_zone_span() we would mistakenly skip other
ZONE_DEVICE ranges and mess up.

Assume you would have a ZONE_DEVICE layout like

[  DEV 0 | Hole | DEV 1 | Hole | DEV 2 ]

What we actually want to do when removing

* DEV 0 is scanning low->high until we find DEV 1
* DEV 1 is doing nothing, because we cannot shrink
* DEV 2 is scanning high -> low until we find DEV 1


I assume we'd want to call in shrink_zone_span() two new functions for
ZONE_DEVICE:
find_smallest_zone_device_pfn
find_biggest_zone_device_pfn

Which would be able to do exactly that scanning, eventually, using
get_dev_pagemap() or some similar source of information.

-- 
Thanks,

David / dhildenb