Re: [RFC PATCH] mm: add last level page table numa info to /proc/pid/numa_pgtable

[haoxin <xhao@linux.alibaba.com>]

在 2022/7/31 上午1:29, Matthew Wilcox 写道:
> On Sun, Jul 31, 2022 at 12:35:28AM +0800, Xin Hao wrote:
>> In many data center servers, the shared memory architectures is
>> Non-Uniform Memory Access (NUMA), remote numa node data access
>> often brings a high latency problem, but what we are easy to ignore
>> is that the page table remote numa access, It can also leads to a
>> performance degradation.
>>
>> So there add a new interface in /proc, This will help developers to
>> get more info about performance issues if they are caused by cross-NUMA.
> Interesting.  The implementation seems rather more complex than
> necessary though.
>
>> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
>> index 2d04e3470d4c..a51befb47ea8 100644
>> --- a/fs/proc/task_mmu.c
>> +++ b/fs/proc/task_mmu.c
>> @@ -1999,4 +1999,133 @@ const struct file_operations proc_pid_numa_maps_operations = {
>>   	.release	= proc_map_release,
>>   };
>>
>> +struct pgtable_numa_maps {
>> +	unsigned long node[MAX_NUMNODES];
>> +};
>> +
>> +struct pgtable_numa_private {
>> +	struct proc_maps_private proc_maps;
>> +	struct pgtable_numa_maps md;
>> +};
> struct pgtable_numa_private {
> 	struct proc_maps_private proc_maps;
> 	unsigned long node[MAX_NUMNODES];
> };
>
>> +static void gather_pgtable_stats(struct page *page, struct pgtable_numa_maps *md)
>> +{
>> +	md->node[page_to_nid(page)] += 1;
>> +}
>> +
>> +static struct page *can_gather_pgtable_numa_stats(pmd_t pmd, struct vm_area_struct *vma,
>> +		unsigned long addr)
>> +{
>> +	struct page *page;
>> +	int nid;
>> +
>> +	if (!pmd_present(pmd))
>> +		return NULL;
>> +
>> +	if (pmd_huge(pmd))
>> +		return NULL;
>> +
>> +	page = pmd_page(pmd);
>> +	nid = page_to_nid(page);
>> +	if (!node_isset(nid, node_states[N_MEMORY]))
>> +		return NULL;
>> +
>> +	return page;
>> +}
>> +
>> +static int gather_pgtable_numa_stats(pmd_t *pmd, unsigned long addr,
>> +		unsigned long end, struct mm_walk *walk)
>> +{
>> +	struct pgtable_numa_maps *md = walk->private;
>> +	struct vm_area_struct *vma = walk->vma;
>> +	struct page *page;
>> +
>> +	if (pmd_huge(*pmd)) {
>> +		struct page *pmd_page;
>> +
>> +		pmd_page = virt_to_page(pmd);
>> +		if (!pmd_page)
>> +			return 0;
>> +
>> +		if (!node_isset(page_to_nid(pmd_page), node_states[N_MEMORY]))
>> +			return 0;
>> +
>> +		gather_pgtable_stats(pmd_page, md);
>> +		goto out;
>> +	}
>> +
>> +	page = can_gather_pgtable_numa_stats(*pmd, vma, addr);
>> +	if (!page)
>> +		return 0;
>> +
>> +	gather_pgtable_stats(page, md);
>> +
>> +out:
>> +	cond_resched();
>> +	return 0;
>> +}
> static int gather_pgtable_numa_stats(pmd_t *pmd, unsigned long addr,
> 		unsigned long end, struct mm_walk *walk)
> {
> 	struct pgtable_numa_private *priv = walk->private;
> 	struct vm_area_struct *vma = walk->vma;
> 	struct page *page;
> 	int nid;
>
> 	if (pmd_huge(*pmd)) {
> 		page = virt_to_page(pmd);
> 	} else {
> 		page = pmd_page(*pmd);
> 	}
>
> 	nid = page_to_nid(page);
> 	priv->node[nid]++;
>
> 	return 0;
> }
Oh,  Thank you for reviewing the code, i will fix it in the next version.