[PATCH v3] smaps: Report PMD page size for pure PMD mappings

[PATCH v3] smaps: Report PMD page size for pure PMD mappings

[Andi Kleen]

When a smaps mapping is only PMD mappings report the
PMD page size for MMUPageSize instead of the base page size.

This is a revised version of an earlier patch that tried
to report multiple page sizes, but there were many objections
mainly centered around compatibility for mixed page size
reporting. This patch side steps all of this by
only handling the non mixed case in the simplest possible way.
It also avoids a problem introduced with v2 that page sizes
for mappings with no pages were incorrectly reported.

Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
 fs/proc/task_mmu.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
index e091931d7ca1..a5c7bc88a539 100644
--- a/fs/proc/task_mmu.c
+++ b/fs/proc/task_mmu.c
@@ -1371,6 +1371,7 @@ static int show_smap(struct seq_file *m, void *v)
 {
 	struct vm_area_struct *vma = v;
 	struct mem_size_stats mss = {};
+	unsigned ps;
 
 	smap_gather_stats(vma, &mss, 0);
 
@@ -1378,7 +1379,16 @@ static int show_smap(struct seq_file *m, void *v)
 
 	SEQ_PUT_DEC("Size:           ", vma->vm_end - vma->vm_start);
 	SEQ_PUT_DEC(" kB\nKernelPageSize: ", vma_kernel_pagesize(vma));
-	SEQ_PUT_DEC(" kB\nMMUPageSize:    ", vma_mmu_pagesize(vma));
+	ps = vma_mmu_pagesize(vma);
+	/*
+	 * When the mapping is only PMD THP report the correct page size.
+	 * When multiple pages are there the user has to figure it out
+	 * from other fields.
+	 */
+	if (mss.shmem_thp + mss.file_thp + mss.anonymous_thp == mss.resident &&
+	    mss.resident)
+		ps = HPAGE_PMD_SIZE;
+	SEQ_PUT_DEC(" kB\nMMUPageSize:    ", ps);
 	seq_puts(m, " kB\n");
 
 	__show_smap(m, &mss, false);
-- 
2.53.0

Re: [PATCH v3] smaps: Report PMD page size for pure PMD mappings

[Vlastimil Babka]

Why do you keep not CCing people from v1/v2 discussions? David said that
already on v2. It seems rather rude to me.

On 3/3/26 9:15 PM, Andi Kleen wrote:
> When a smaps mapping is only PMD mappings report the
> PMD page size for MMUPageSize instead of the base page size.
> 
> This is a revised version of an earlier patch that tried
> to report multiple page sizes, but there were many objections
> mainly centered around compatibility for mixed page size
> reporting. This patch side steps all of this by
> only handling the non mixed case in the simplest possible way.
> It also avoids a problem introduced with v2 that page sizes
> for mappings with no pages were incorrectly reported.
> 
> Signed-off-by: Andi Kleen <ak@linux.intel.com>

Will this be useful in practice or just confusing? There can be e.g.
misaligned mappings, or other reasons why there won't be 100% THP
coverage. This all or nothing value (IIUC) seems inferior to the
counters we have that say how much is pmd mapped, so I'm not really sure
it's worth changing this.

> ---
>  fs/proc/task_mmu.c | 12 +++++++++++-
>  1 file changed, 11 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
> index e091931d7ca1..a5c7bc88a539 100644
> --- a/fs/proc/task_mmu.c
> +++ b/fs/proc/task_mmu.c
> @@ -1371,6 +1371,7 @@ static int show_smap(struct seq_file *m, void *v)
>  {
>  	struct vm_area_struct *vma = v;
>  	struct mem_size_stats mss = {};
> +	unsigned ps;
>  
>  	smap_gather_stats(vma, &mss, 0);
>  
> @@ -1378,7 +1379,16 @@ static int show_smap(struct seq_file *m, void *v)
>  
>  	SEQ_PUT_DEC("Size:           ", vma->vm_end - vma->vm_start);
>  	SEQ_PUT_DEC(" kB\nKernelPageSize: ", vma_kernel_pagesize(vma));
> -	SEQ_PUT_DEC(" kB\nMMUPageSize:    ", vma_mmu_pagesize(vma));
> +	ps = vma_mmu_pagesize(vma);
> +	/*
> +	 * When the mapping is only PMD THP report the correct page size.
> +	 * When multiple pages are there the user has to figure it out
> +	 * from other fields.
> +	 */
> +	if (mss.shmem_thp + mss.file_thp + mss.anonymous_thp == mss.resident &&
> +	    mss.resident)
> +		ps = HPAGE_PMD_SIZE;
> +	SEQ_PUT_DEC(" kB\nMMUPageSize:    ", ps);
>  	seq_puts(m, " kB\n");
>  
>  	__show_smap(m, &mss, false);

Re: [PATCH v3] smaps: Report PMD page size for pure PMD mappings

[Andi Kleen]

On Wed, Mar 04, 2026 at 06:29:09PM +0100, Vlastimil Babka wrote:
> Why do you keep not CCing people from v1/v2 discussions? David said that
> already on v2. It seems rather rude to me.
> 
> On 3/3/26 9:15 PM, Andi Kleen wrote:
> > When a smaps mapping is only PMD mappings report the
> > PMD page size for MMUPageSize instead of the base page size.
> > 
> > This is a revised version of an earlier patch that tried
> > to report multiple page sizes, but there were many objections
> > mainly centered around compatibility for mixed page size
> > reporting. This patch side steps all of this by
> > only handling the non mixed case in the simplest possible way.
> > It also avoids a problem introduced with v2 that page sizes
> > for mappings with no pages were incorrectly reported.
> > 
> > Signed-off-by: Andi Kleen <ak@linux.intel.com>
> 
> Will this be useful in practice or just confusing? There can be e.g.

It might have saved me some wasted time.

I found the current situation confusing!

> misaligned mappings, or other reasons why there won't be 100% THP
> coverage. This all or nothing value (IIUC) seems inferior to the

s/nothing/4K like what's there today/

Yes probably multiple page sizes would be better, but I'm not reopening
that discussion now.

> counters we have that say how much is pmd mapped, so I'm not really sure
> it's worth changing this.

The other counters still exist of course.

I don't see how there can be confusion if we report the correct page
size if the mapping is only that page size vs a page size that is not 
used at all.

-Andi

Re: [PATCH v3] smaps: Report PMD page size for pure PMD mappings

[David Hildenbrand (Arm)]

On 3/4/26 18:29, Vlastimil Babka wrote:
> Why do you keep not CCing people from v1/v2 discussions? David said that
> already on v2. It seems rather rude to me.
> 
> On 3/3/26 9:15 PM, Andi Kleen wrote:
>> When a smaps mapping is only PMD mappings report the
>> PMD page size for MMUPageSize instead of the base page size.
>>
>> This is a revised version of an earlier patch that tried
>> to report multiple page sizes, but there were many objections
>> mainly centered around compatibility for mixed page size
>> reporting. This patch side steps all of this by
>> only handling the non mixed case in the simplest possible way.
>> It also avoids a problem introduced with v2 that page sizes
>> for mappings with no pages were incorrectly reported.
>>
>> Signed-off-by: Andi Kleen <ak@linux.intel.com>
> 
> Will this be useful in practice or just confusing? There can be e.g.
> misaligned mappings, or other reasons why there won't be 100% THP
> coverage. This all or nothing value (IIUC) seems inferior to the
> counters we have that say how much is pmd mapped, so I'm not really sure
> it's worth changing this.

My opinion on this remains unchanged.

Not CCing me once more is questionable and makes me want to recommend
Andi to work on different parts of the kernel.

I sent a doc update to clarify this. Won't make everybody happy, but we
can't turn back time.

-- 
Cheers,

David