mm: add last level page table numa info to /proc/pid/numa_pgtable

RFC PATCH Superseded Xin Hao 2 participants 3 messages Jul 30, 2022 – Jul 31, 2022

Subsystems (3)

Memory management Proc filesystem Memory mapping

lore

Share & tools

lore

Xin Hao

author

2022-07-30 16:35 UTC #0 lore raw

RFC PATCH 1/1 mm: add last level page table numa info to /proc/pid/numa_pgtable

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.

Signed-off-by Xin Hao <[email protected]>

---
 fs/proc/base.c     |   6 ++-
 fs/proc/internal.h |   1 +
 fs/proc/task_mmu.c | 129 +++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 134 insertions(+), 2 deletions(-)

fs/proc/base.c +4 -2

index 8dfa36a99c74..de60351813ec 100644
@@ -2626,7 +2626,7 @@ static struct dentry *proc_pident_instantiate(struct dentry *dentry,
 	return d_splice_alias(inode, dentry);
 }
 
-static struct dentry *proc_pident_lookup(struct inode *dir,
+static struct dentry *proc_pident_lookup(struct inode *dir,
 					 struct dentry *dentry,
 					 const struct pid_entry *p,
 					 const struct pid_entry *end)

@@ -2839,7 +2839,7 @@ static const struct pid_entry attr_dir_stuff[] = {
 
 static int proc_attr_dir_readdir(struct file *file, struct dir_context *ctx)
 {
-	return proc_pident_readdir(file, ctx,
+	return proc_pident_readdir(file, ctx,
 				   attr_dir_stuff, ARRAY_SIZE(attr_dir_stuff));
 }

@@ -3224,6 +3224,7 @@ static const struct pid_entry tgid_base_stuff[] = {
 	REG("maps",       S_IRUGO, proc_pid_maps_operations),
 #ifdef CONFIG_NUMA
 	REG("numa_maps",  S_IRUGO, proc_pid_numa_maps_operations),
+	REG("numa_pgtable", S_IRUGO, proc_pid_numa_pgtable_operations),
 #endif
 	REG("mem",        S_IRUSR|S_IWUSR, proc_mem_operations),
 	LNK("cwd",        proc_cwd_link),

@@ -3571,6 +3572,7 @@ static const struct pid_entry tid_base_stuff[] = {
 #endif
 #ifdef CONFIG_NUMA
 	REG("numa_maps", S_IRUGO, proc_pid_numa_maps_operations),
+	REG("numa_pgtable", S_IRUGO, proc_pid_numa_pgtable_operations),
 #endif
 	REG("mem",       S_IRUSR|S_IWUSR, proc_mem_operations),
 	LNK("cwd",       proc_cwd_link),

fs/proc/internal.h +1 -0

index 06a80f78433d..e7ed9ef097b6 100644
@@ -296,6 +296,7 @@ struct mm_struct *proc_mem_open(struct inode *inode, unsigned int mode);
 
 extern const struct file_operations proc_pid_maps_operations;
 extern const struct file_operations proc_pid_numa_maps_operations;
+extern const struct file_operations proc_pid_numa_pgtable_operations;
 extern const struct file_operations proc_pid_smaps_operations;
 extern const struct file_operations proc_pid_smaps_rollup_operations;
 extern const struct file_operations proc_clear_refs_operations;

fs/proc/task_mmu.c +129 -0

133 lines not on this page · raw message

--
2.31.0

Matthew Wilcox

2022-07-30 17:29 UTC #1 lore raw

Quoted 8 lines · Xin Hao

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.

Quoted 6 lines

fs/proc/task_mmu.c +0 -0

index 2d04e3470d4c..a51befb47ea8 100644
@@ -1999,4 +1999,133 @@ const struct file_operations proc_pid_numa_maps_operations = {
 	.release	= proc_map_release,

fs/proc/task_mmu.c const struct file_operations proc_pid_numa_maps_operations = { · line 2009

Quoted 10 lines

fs/proc/task_mmu.c +8 -0

 };
 
+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];
};

Quoted 46 lines

fs/proc/task_mmu.c +46 -0

+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;
+	}
+

fs/proc/task_mmu.c const struct file_operations proc_pid_numa_maps_operations = { · line 2066

Quoted 10 lines

fs/proc/task_mmu.c +10 -0

+	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;
}

haoxin

author

2022-07-31 14:11 UTC #2 lore raw

Quoted 100 lines · Matthew Wilcox

Quoted 8 lines · Xin Hao
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.
Quoted 6 lines
fs/proc/task_mmu.c +0 -0
index 2d04e3470d4c..a51befb47ea8 100644
@@ -1999,4 +1999,133 @@ const struct file_operations proc_pid_numa_maps_operations = {
  	.release	= proc_map_release,
fs/proc/task_mmu.c const struct file_operations proc_pid_numa_maps_operations = { · line 2009
Quoted 10 lines
fs/proc/task_mmu.c +8 -0
  };
 
+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];
};
Quoted 46 lines
fs/proc/task_mmu.c +46 -0
+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;
+	}
+
fs/proc/task_mmu.c const struct file_operations proc_pid_numa_maps_operations = { · line 2066
Quoted 10 lines
fs/proc/task_mmu.c +10 -0
+	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 {

Quoted 8 lines

		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.