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.
V4 -> V3
1, Add some detects and spinlock when operating PMD page as Baolin Wang advised.
V2 -> V3
1, Fix compile warning bug.
V1 -> V2
1, As Matthew Wilcox advise, Simplify the code.
2, Do some codes format fix.
V3: https://lore.kernel.org/linux-mm/[email protected]/
V2: https://lore.kernel.org/linux-mm/[email protected]/
V1: https://lore.kernel.org/linux-mm/[email protected]/T/
Xin Hao (1):
mm: add last level page table numa info to /proc/pid/numa_pgtable
fs/proc/base.c | 2 ++
fs/proc/internal.h | 1 +
fs/proc/task_mmu.c | 87 ++++++++++++++++++++++++++++++++++++++++++++++
3 files changed, 90 insertions(+)
--
2.31.0
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.
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(+)
diff --git a/fs/proc/base.c b/fs/proc/base.c
index 8dfa36a99c74..487e82dd3275 100644
--- a/fs/proc/base.c
+++ b/fs/proc/base.c
@@ -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),
diff --git a/fs/proc/internal.h b/fs/proc/internal.h
index 06a80f78433d..e7ed9ef097b6 100644
--- a/fs/proc/internal.h
+++ b/fs/proc/internal.h
@@ -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;
diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
index 2d04e3470d4c..77b7a49757f5 100644
--- a/fs/proc/task_mmu.c
+++ b/fs/proc/task_mmu.c
@@ -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,
+ 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);
+ }
+
+ 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
On 01.08.22 14:17, 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.
Let me try rewriting:
"
Many data center servers employ Non-Uniform Memory Access (NUMA)
architectures. Remote numa memory access results in high latency. While
memory placement is one issue, sub-optimal page table placement can also
result in surprise 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.
Why do we only care about "last level page table", why not about the others?
IMHO, we could emit something like "0, 1, 3, 0" instead for a given user
space address, showing the NUMA node the page table belongs to from
highest to lowest page table level.
>
> Reported-by: kernel test robot <[email protected]>
The kernel test robot reported that we need "/proc/pid/numa_pgtable" ?! :)
Just drop that unless it's a follow-up fix.
> 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(+)
>
> diff --git a/fs/proc/base.c b/fs/proc/base.c
> index 8dfa36a99c74..487e82dd3275 100644
> --- a/fs/proc/base.c
> +++ b/fs/proc/base.c
> @@ -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),
> diff --git a/fs/proc/internal.h b/fs/proc/internal.h
> index 06a80f78433d..e7ed9ef097b6 100644
> --- a/fs/proc/internal.h
> +++ b/fs/proc/internal.h
> @@ -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;
> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
> index 2d04e3470d4c..77b7a49757f5 100644
> --- a/fs/proc/task_mmu.c
> +++ b/fs/proc/task_mmu.c
> @@ -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,
> + 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);
> + }
> +
> + 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.
s/talbe/table/
but the comment somehow doesn't make sense. We don't display a page table.
> + */
> +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
>
--
Thanks,
David / dhildenb
在 2022/8/1 下午9:28, David Hildenbrand 写道:
> On 01.08.22 14:17, 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.
> Let me try rewriting:
>
> "
> Many data center servers employ Non-Uniform Memory Access (NUMA)
> architectures. Remote numa memory access results in high latency. While
> memory placement is one issue, sub-optimal page table placement can also
> result in surprise performance degradation.
> "
Thanks, it reads more clearly.
>> 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.
>
> Why do we only care about "last level page table", why not about the others?
>
> IMHO, we could emit something like "0, 1, 3, 0" instead for a given user
> space address, showing the NUMA node the page table belongs to from
> highest to lowest page table level.
I have planned to implement the PTE page table in this version first,
and then support other page tables in the next patch later.
>
>> Reported-by: kernel test robot <[email protected]>
> The kernel test robot reported that we need "/proc/pid/numa_pgtable" ?! :)
>
> Just drop that unless it's a follow-up fix.
Get it.
>
>> 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(+)
>>
>> diff --git a/fs/proc/base.c b/fs/proc/base.c
>> index 8dfa36a99c74..487e82dd3275 100644
>> --- a/fs/proc/base.c
>> +++ b/fs/proc/base.c
>> @@ -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),
>> diff --git a/fs/proc/internal.h b/fs/proc/internal.h
>> index 06a80f78433d..e7ed9ef097b6 100644
>> --- a/fs/proc/internal.h
>> +++ b/fs/proc/internal.h
>> @@ -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;
>> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
>> index 2d04e3470d4c..77b7a49757f5 100644
>> --- a/fs/proc/task_mmu.c
>> +++ b/fs/proc/task_mmu.c
>> @@ -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,
>> + 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);
>> + }
>> +
>> + 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.
> s/talbe/table/
>
> but the comment somehow doesn't make sense. We don't display a page table.
>
>> + */
>> +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
>>
>
On 04.08.22 10:04, haoxin wrote:
>
> 在 2022/8/1 下午9:28, David Hildenbrand 写道:
>> On 01.08.22 14:17, 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.
>> Let me try rewriting:
>>
>> "
>> Many data center servers employ Non-Uniform Memory Access (NUMA)
>> architectures. Remote numa memory access results in high latency. While
>> memory placement is one issue, sub-optimal page table placement can also
>> result in surprise performance degradation.
>> "
> Thanks, it reads more clearly.
>
>>> 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.
>>
>> Why do we only care about "last level page table", why not about the others?
>>
>> IMHO, we could emit something like "0, 1, 3, 0" instead for a given user
>> space address, showing the NUMA node the page table belongs to from
>> highest to lowest page table level.
>
> I have planned to implement the PTE page table in this version first,
> and then support other page tables in the next patch later.
If there are plans, let's do it all at once, to get a good and single
interface to expose that information.
--
Thanks,
David / dhildenb
在 2022/8/4 下午4:12, David Hildenbrand 写道:
> On 04.08.22 10:04, haoxin wrote:
>> 在 2022/8/1 下午9:28, David Hildenbrand 写道:
>>> On 01.08.22 14:17, 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.
>>> Let me try rewriting:
>>>
>>> "
>>> Many data center servers employ Non-Uniform Memory Access (NUMA)
>>> architectures. Remote numa memory access results in high latency. While
>>> memory placement is one issue, sub-optimal page table placement can also
>>> result in surprise performance degradation.
>>> "
>> Thanks, it reads more clearly.
>>
>>>> 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.
>>> Why do we only care about "last level page table", why not about the others?
>>>
>>> IMHO, we could emit something like "0, 1, 3, 0" instead for a given user
>>> space address, showing the NUMA node the page table belongs to from
>>> highest to lowest page table level.
>> I have planned to implement the PTE page table in this version first,
>> and then support other page tables in the next patch later.
> If there are plans, let's do it all at once, to get a good and single
> interface to expose that information.
Ok, thanks for your suggestion, I will implement it in the next version.
>