Re: [PATCH] mm: hugetlb: eliminate memory-less nodes handling

[Muchun Song]

> On Sep 1, 2022, at 17:00, Aneesh Kumar K V <aneesh.kumar@xxxxxxxxxxxxx> wrote:
> 
> On 9/1/22 2:00 PM, Muchun Song wrote:
>> The memory-notify-based approach aims to handle meory-less nodes, however, it just adds
>> the complexity of code as pointed by David in thread [1].  The handling of memory-less
>> nodes is introduced by commit 4faf8d950ec4 ("hugetlb: handle memory hot-plug events").
>> From its commit message, we cannot find any necessity of handling this case. So, we can
>> simply register/unregister sysfs entries in register_node/unregister_node to simlify the
>> code.
> 
> Isn't that hotplug callback added because in hugetlb_register_all_nodes() we register
> sysfs nodes only for N_MEMORY nodes? 

I think you might right. I have looked at the commit 9a30523066cd which introduces the sysfs
creation. I saw it create the sysfs for every possible node.

	for (nid = 0; nid < nr_node_ids; nid++)
		hugetlb_register_node(node);

And then I checked the commit 9b5e5d0fdc91, which said it was a preparation for handling
memory-less nodes via memory hotplug.

> 
> 
>> 
>> https://lore.kernel.org/linux-mm/60933ffc-b850-976c-78a0-0ee6e0ea9ef0@xxxxxxxxxx/ [1]
>> Suggested-by: David Hildenbrand <david@xxxxxxxxxx>
>> Signed-off-by: Muchun Song <songmuchun@xxxxxxxxxxxxx>
>> ---
>> drivers/base/node.c  |  7 +++++--
>> include/linux/node.h |  5 +++++
>> mm/hugetlb.c         | 37 ++++++++++---------------------------
>> 3 files changed, 20 insertions(+), 29 deletions(-)
>> 
>> diff --git a/drivers/base/node.c b/drivers/base/node.c
>> index ed391cb09999..cf115d5a9b8a 100644
>> --- a/drivers/base/node.c
>> +++ b/drivers/base/node.c
>> @@ -608,10 +608,12 @@ static int register_node(struct node *node, int num)
>> 	node->dev.groups = node_dev_groups;
>> 	error = device_register(&node->dev);
>> 
>> -	if (error)
>> +	if (error) {
>> 		put_device(&node->dev);
>> -	else
>> +	} else {
>> +		hugetlb_register_node(node);
>> 		compaction_register_node(node);
>> +	}
>> 
> 
> 
> I guess this will handle register of sysfs hugetlb files for new NUMA nodes
> after hugetlb_initialized = true;

Yes.

> 
> But what about N_CPU that can get memory added later. Do we need to update
> hugetlb_register_all_nodes() to handle N_ONLINE? 

I think we should.

> 
> 
>> 	return error;
>> }
>> @@ -625,6 +627,7 @@ static int register_node(struct node *node, int num)
>>  */
>> void unregister_node(struct node *node)
>> {
>> +	hugetlb_unregister_node(node);
>> 	compaction_unregister_node(node);
>> 	node_remove_accesses(node);
>> 	node_remove_caches(node);
>> diff --git a/include/linux/node.h b/include/linux/node.h
>> index 427a5975cf40..f5d41498c2bf 100644
>> --- a/include/linux/node.h
>> +++ b/include/linux/node.h
>> @@ -138,6 +138,11 @@ extern void unregister_memory_block_under_nodes(struct memory_block *mem_blk);
>> extern int register_memory_node_under_compute_node(unsigned int mem_nid,
>> 						   unsigned int cpu_nid,
>> 						   unsigned access);
>> +
>> +#ifdef CONFIG_HUGETLBFS
>> +void hugetlb_register_node(struct node *node);
>> +void hugetlb_unregister_node(struct node *node);
>> +#endif
>> #else
>> static inline void node_dev_init(void)
>> {
>> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
>> index d0617d64d718..722e862bb6be 100644
>> --- a/mm/hugetlb.c
>> +++ b/mm/hugetlb.c
>> @@ -3898,6 +3898,7 @@ static void __init hugetlb_sysfs_init(void)
>> }
>> 
>> #ifdef CONFIG_NUMA
>> +static bool hugetlb_initialized __ro_after_init;
>> 
>> /*
>>  * node_hstate/s - associate per node hstate attributes, via their kobjects,
>> @@ -3953,7 +3954,7 @@ static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
>>  * Unregister hstate attributes from a single node device.
>>  * No-op if no hstate attributes attached.
>>  */
>> -static void hugetlb_unregister_node(struct node *node)
>> +void hugetlb_unregister_node(struct node *node)
>> {
>> 	struct hstate *h;
>> 	struct node_hstate *nhs = &node_hstates[node->dev.id];
>> @@ -3983,19 +3984,22 @@ static void hugetlb_unregister_node(struct node *node)
>>  * Register hstate attributes for a single node device.
>>  * No-op if attributes already registered.
>>  */
>> -static int hugetlb_register_node(struct node *node)
>> +void hugetlb_register_node(struct node *node)
>> {
>> 	struct hstate *h;
>> 	struct node_hstate *nhs = &node_hstates[node->dev.id];
>> 	int err;
>> 
>> +	if (!hugetlb_initialized)
>> +		return;
>> +
>> 	if (nhs->hugepages_kobj)
>> -		return 0;		/* already allocated */
>> +		return;		/* already allocated */
>> 
>> 	nhs->hugepages_kobj = kobject_create_and_add("hugepages",
>> 							&node->dev.kobj);
>> 	if (!nhs->hugepages_kobj)
>> -		return -ENOMEM;
>> +		return;
>> 
>> 	for_each_hstate(h) {
>> 		err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
>> @@ -4005,28 +4009,9 @@ static int hugetlb_register_node(struct node *node)
>> 			pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
>> 				h->name, node->dev.id);
>> 			hugetlb_unregister_node(node);
>> -			return -ENOMEM;
>> +			break;
>> 		}
>> 	}
>> -	return 0;
>> -}
>> -
>> -static int __meminit hugetlb_memory_callback(struct notifier_block *self,
>> -					     unsigned long action, void *arg)
>> -{
>> -	int ret = 0;
>> -	struct memory_notify *mnb = arg;
>> -	int nid = mnb->status_change_nid;
>> -
>> -	if (nid == NUMA_NO_NODE)
>> -		return NOTIFY_DONE;
>> -
>> -	if (action == MEM_GOING_ONLINE)
>> -		ret = hugetlb_register_node(node_devices[nid]);
>> -	else if (action == MEM_CANCEL_ONLINE || action == MEM_OFFLINE)
>> -		hugetlb_unregister_node(node_devices[nid]);
>> -
>> -	return notifier_from_errno(ret);
>> }
>> 
>> /*
>> @@ -4038,11 +4023,9 @@ static void __init hugetlb_register_all_nodes(void)
>> {
>> 	int nid;
>> 
>> -	get_online_mems();
>> -	hotplug_memory_notifier(hugetlb_memory_callback, 0);
>> +	hugetlb_initialized = true;
>> 	for_each_node_state(nid, N_MEMORY)
> 
> 
> Should this be for_each_online_node() ?

So, yes.

Thanks for your review.

Muchun.

> 
>> 		hugetlb_register_node(node_devices[nid]);
>> -	put_online_mems();
>> }
>> #else	/* !CONFIG_NUMA */