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.

V2 -> V3
1, Fix compile warning bug.

V1 -> V2
1, As Matthew Wilcox advise, Simplify the code.
2, Do some codes format fix.

V2: https://lore.kernel.org/linux-mm/[email protected]/
V1: https://lore.kernel.org/linux-mm/[email protected]/T/

Reported-by kernel test robot <[email protected]>

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

---
  fs/proc/base.c     |  2 ++
  fs/proc/internal.h |  1 +
  fs/proc/task_mmu.c | 87 ++++++++++++++++++++++++++++++++++++++++++++++
  3 files changed, 90 insertions(+)

fs/proc/base.c +2 -0

index 8dfa36a99c74..487e82dd3275 100644
@@ -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 +6 -0

index 2d04e3470d4c..77b7a49757f5 100644
@@ -1999,4 +1999,91 @@ const struct file_operations proc_pid_numa_maps_operations = {
  	.release	= proc_map_release,
  };
 
+struct pgtable_numa_private {
+	struct proc_maps_private proc_maps;
+	unsigned long node[MAX_NUMNODES];
+};
+
+static int gather_pgtable_numa_stats(pmd_t *pmd, unsigned long addr,

fs/proc/task_mmu.c +71 -1

+	}
+
+	nid = page_to_nid(page);
+	priv->node[nid]++;
+
+	return 0;
+}
+
+static const struct mm_walk_ops show_numa_pgtable_ops = {
+	.pmd_entry = gather_pgtable_numa_stats,
+};
+
+/*
+ * Display the page talbe allocated per node via /proc.
+ */
+static int show_numa_pgtable(struct seq_file *m, void *v)
+{
+	struct pgtable_numa_private *numa_priv = m->private;
+	struct vm_area_struct *vma = v;
+	struct mm_struct *mm = vma->vm_mm;
+	struct file *file = vma->vm_file;
+	int nid;
+
+	if (!mm)
+		return 0;
+
+	memset(numa_priv->node, 0, sizeof(numa_priv->node));
+
+	seq_printf(m, "%08lx ", vma->vm_start);
+
+	if (file) {
+		seq_puts(m, " file=");
+		seq_file_path(m, file, "\n\t= ");
+	} else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
+		seq_puts(m, " heap");
+	} else if (is_stack(vma)) {
+		seq_puts(m, " stack");
+	}
+
+	/* mmap_lock is held by m_start */
+	walk_page_vma(vma, &show_numa_pgtable_ops, numa_priv);
+
+	for_each_node_state(nid, N_MEMORY) {
+		if (numa_priv->node[nid])
+			seq_printf(m, " N%d=%lu", nid, numa_priv->node[nid]);
+	}
+	seq_putc(m, '\n');
+
+	return 0;
+}
+
+static const struct seq_operations proc_pid_numa_pgtable_op = {
+	.start  = m_start,
+	.next   = m_next,
+	.stop   = m_stop,
+	.show   = show_numa_pgtable,
+};
+
+static int pid_numa_pgtable_open(struct inode *inode, struct file *file)
+{
+	return proc_maps_open(inode, file, &proc_pid_numa_pgtable_op,
+			sizeof(struct pgtable_numa_private));
+}
+
+const struct file_operations proc_pid_numa_pgtable_operations = {
+	.open		= pid_numa_pgtable_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= proc_map_release,
+};
+
  #endif /* CONFIG_NUMA */
--

2.31.0

Hi Xin,

Quoted 9 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.

Quoted 6 lines

V2 -> V3
1, Fix compile warning bug.

V1 -> V2
1, As Matthew Wilcox advise, Simplify the code.
2, Do some codes format fix.

Please move the changes history under your 'Signed-off-by' with '---'.

Quoted 61 lines

V2: https://lore.kernel.org/linux-mm/[email protected]/
V1: https://lore.kernel.org/linux-mm/[email protected]/T/

Reported-by kernel test robot <[email protected]>

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

---
  fs/proc/base.c     |  2 ++
  fs/proc/internal.h |  1 +
  fs/proc/task_mmu.c | 87 ++++++++++++++++++++++++++++++++++++++++++++++

  3 files changed, 90 insertions(+)

fs/proc/base.c +1 -0

index 8dfa36a99c74..487e82dd3275 100644
@@ -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),

fs/proc/base.c +1 -0

@@ -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 +6 -0

index 2d04e3470d4c..77b7a49757f5 100644
@@ -1999,4 +1999,91 @@ const struct file_operations 
 proc_pid_numa_maps_operations = {
       .release    = proc_map_release,
   };
 
+struct pgtable_numa_private {
+    struct proc_maps_private proc_maps;
+    unsigned long node[MAX_NUMNODES];
+};
+
+static int gather_pgtable_numa_stats(pmd_t *pmd, unsigned long addr,

fs/proc/task_mmu.c const struct file_operations · line 2018

Quoted 10 lines

fs/proc/task_mmu.c +10 -0

+                     unsigned long end, struct mm_walk *walk)
+{
+    struct pgtable_numa_private *priv = walk->private;
+    struct page *page;
+    int nid;
+
+    if (pmd_huge(*pmd)) {
+        page = virt_to_page(pmd);
+    } else {
+        page = pmd_page(*pmd);

You should validate if the pmd is valid or present before getting the pagetable page.

if (pmd_none(*pmd) || !pmd_present(*pmd))

Another issue is I think you should hold the pmd lock to call pmd_page(), since after the validation of pmd_huge(), the pmd entry can be modified by other threads if you did not hold the pmd lock.

Quoted 76 lines

fs/proc/task_mmu.c +34 -0

+    }
+
+    nid = page_to_nid(page);
+    priv->node[nid]++;
+
+    return 0;
+}
+
+static const struct mm_walk_ops show_numa_pgtable_ops = {
+    .pmd_entry = gather_pgtable_numa_stats,
+};
+
+/*
+ * Display the page talbe allocated per node via /proc.
+ */
+static int show_numa_pgtable(struct seq_file *m, void *v)
+{
+    struct pgtable_numa_private *numa_priv = m->private;
+    struct vm_area_struct *vma = v;
+    struct mm_struct *mm = vma->vm_mm;
+    struct file *file = vma->vm_file;
+    int nid;
+
+    if (!mm)
+        return 0;
+
+    memset(numa_priv->node, 0, sizeof(numa_priv->node));
+
+    seq_printf(m, "%08lx ", vma->vm_start);
+
+    if (file) {
+        seq_puts(m, " file=");
+        seq_file_path(m, file, "\n\t= ");
+    } else if (vma->vm_start <= mm->brk && vma->vm_end >= 

mm->start_brk) {
+        seq_puts(m, " heap");
+    } else if (is_stack(vma)) {
+        seq_puts(m, " stack");
+    }

fs/proc/task_mmu.c +21 -0

+
+    /* mmap_lock is held by m_start */
+    walk_page_vma(vma, &show_numa_pgtable_ops, numa_priv);
+
+    for_each_node_state(nid, N_MEMORY) {
+        if (numa_priv->node[nid])
+            seq_printf(m, " N%d=%lu", nid, numa_priv->node[nid]);
+    }
+    seq_putc(m, '\n');
+
+    return 0;
+}
+
+static const struct seq_operations proc_pid_numa_pgtable_op = {
+    .start  = m_start,
+    .next   = m_next,
+    .stop   = m_stop,
+    .show   = show_numa_pgtable,
+};
+
+static int pid_numa_pgtable_open(struct inode *inode, struct file 

*file)

fs/proc/task_mmu.c +12 -0

+{
+    return proc_maps_open(inode, file, &proc_pid_numa_pgtable_op,
+            sizeof(struct pgtable_numa_private));
+}
+
+const struct file_operations proc_pid_numa_pgtable_operations = {
+    .open        = pid_numa_pgtable_open,
+    .read        = seq_read,
+    .llseek        = seq_lseek,
+    .release    = proc_map_release,
+};
+

  #endif /* CONFIG_NUMA */

Signature · 1 line

2.31.0