[PATCH 0/3] mm/smaps: Fixes and optimizations on shmem swap handling

[PATCH 0/3] mm/smaps: Fixes and optimizations on shmem swap handling

[Peter Xu]

This series grows from the patch previously posted here:

  [PATCH] mm/smaps: Use vma->vm_pgoff directly when counting partial swap
  https://lore.kernel.org/lkml/20210916215839.95177-1-peterx@redhat.com/

Vlastimil reported a bug that is even more important to fix than the cleanup,
so I put it as patch 1 here.  There's a test program we can use to verify the
bug before/after the patch.  I used the same program to test patch 2/3 because
it covers walking shmem swap both in page cache and in pgtables.

Patch 2 is the original patch, though with a tiny touchup as Vlastimil
suggested.

Patch 3 is a further cleanup of the shmem swap logic, hopefully make it even
cleaner.

Please review, thanks.

Peter Xu (3):
  mm/smaps: Fix shmem pte hole swap calculation
  mm/smaps: Use vma->vm_pgoff directly when counting partial swap
  mm/smaps: Simplify shmem handling of pte holes

 fs/proc/task_mmu.c | 28 ++++++++++++++++------------
 mm/shmem.c         |  5 ++---
 2 files changed, 18 insertions(+), 15 deletions(-)

-- 
2.31.1

[PATCH 1/3] mm/smaps: Fix shmem pte hole swap calculation

[Peter Xu]

The shmem swap calculation on the privately writable mappings are using wrong
parameters as spotted by Vlastimil.  Fix them.  That's introduced in commit
48131e03ca4e, when rework shmem_swap_usage to shmem_partial_swap_usage.

Test program:

==================

void main(void)
{
    char *buffer, *p;
    int i, fd;

    fd = memfd_create("test", 0);
    assert(fd > 0);

    /* isize==2M*3, fill in pages, swap them out */
    ftruncate(fd, SIZE_2M * 3);
    buffer = mmap(NULL, SIZE_2M * 3, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
    assert(buffer);
    for (i = 0, p = buffer; i < SIZE_2M * 3 / 4096; i++) {
        *p = 1;
        p += 4096;
    }
    madvise(buffer, SIZE_2M * 3, MADV_PAGEOUT);
    munmap(buffer, SIZE_2M * 3);

    /*
     * Remap with private+writtable mappings on partial of the inode (<= 2M*3),
     * while the size must also be >= 2M*2 to make sure there's a none pmd so
     * smaps_pte_hole will be triggered.
     */
    buffer = mmap(NULL, SIZE_2M * 2, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
    printf("pid=%d, buffer=%p\n", getpid(), buffer);

    /* Check /proc/$PID/smap_rollup, should see 4MB swap */
    sleep(1000000);
}
==================

Before the patch, smaps_rollup shows <4MB swap and the number will be random
depending on the alignment of the buffer of mmap() allocated.  After this
patch, it'll show 4MB.

Fixes: 48131e03ca4e ("mm, proc: reduce cost of /proc/pid/smaps for unpopulated shmem mappings")
Reported-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Peter Xu <peterx@redhat.com>
---
 fs/proc/task_mmu.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
index cf25be3e0321..2197f669e17b 100644
--- a/fs/proc/task_mmu.c
+++ b/fs/proc/task_mmu.c
@@ -478,9 +478,11 @@ static int smaps_pte_hole(unsigned long addr, unsigned long end,
 			  __always_unused int depth, struct mm_walk *walk)
 {
 	struct mem_size_stats *mss = walk->private;
+	struct vm_area_struct *vma = walk->vma;
 
-	mss->swap += shmem_partial_swap_usage(
-			walk->vma->vm_file->f_mapping, addr, end);
+	mss->swap += shmem_partial_swap_usage(walk->vma->vm_file->f_mapping,
+					      linear_page_index(vma, addr),
+					      linear_page_index(vma, end));
 
 	return 0;
 }
-- 
2.31.1

[PATCH 2/3] mm/smaps: Use vma->vm_pgoff directly when counting partial swap

[Peter Xu]

As it's trying to cover the whole vma anyways, use direct vm_pgoff value and
vma_pages() rather than linear_page_index.

Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
---
 mm/shmem.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/mm/shmem.c b/mm/shmem.c
index 96ccf6e941aa..d2620db8c938 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -856,9 +856,8 @@ unsigned long shmem_swap_usage(struct vm_area_struct *vma)
 		return swapped << PAGE_SHIFT;
 
 	/* Here comes the more involved part */
-	return shmem_partial_swap_usage(mapping,
-			linear_page_index(vma, vma->vm_start),
-			linear_page_index(vma, vma->vm_end));
+	return shmem_partial_swap_usage(mapping, vma->vm_pgoff,
+					vma->vm_pgoff + vma_pages(vma));
 }
 
 /*
-- 
2.31.1

[PATCH 3/3] mm/smaps: Simplify shmem handling of pte holes

[Peter Xu]

Firstly, check_shmem_swap variable is actually not necessary, because it's
always set with pte_hole hook; checking each would work.

Meanwhile, the check within smaps_pte_entry is not easy to follow.  E.g.,
pte_none() check is not needed as "!pte_present && !is_swap_pte" is the same.
Since at it, use the pte_hole() helper rather than dup the page cache lookup.

Still keep the CONFIG_SHMEM part so the code can be optimized to nop for !SHMEM.

There will be a very slight functional change in smaps_pte_entry(), that for
!SHMEM we'll return early for pte_none (before checking page==NULL), but that's
even nicer.

Cc: Hugh Dickins <hughd@google.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Peter Xu <peterx@redhat.com>
---
 fs/proc/task_mmu.c | 22 ++++++++++++----------
 1 file changed, 12 insertions(+), 10 deletions(-)

diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
index 2197f669e17b..ad667dbc96f5 100644
--- a/fs/proc/task_mmu.c
+++ b/fs/proc/task_mmu.c
@@ -397,7 +397,6 @@ struct mem_size_stats {
 	u64 pss_shmem;
 	u64 pss_locked;
 	u64 swap_pss;
-	bool check_shmem_swap;
 };
 
 static void smaps_page_accumulate(struct mem_size_stats *mss,
@@ -490,6 +489,16 @@ static int smaps_pte_hole(unsigned long addr, unsigned long end,
 #define smaps_pte_hole		NULL
 #endif /* CONFIG_SHMEM */
 
+static void smaps_pte_hole_lookup(unsigned long addr, struct mm_walk *walk)
+{
+#ifdef CONFIG_SHMEM
+	if (walk->ops->pte_hole) {
+		/* depth is not used */
+		smaps_pte_hole(addr, addr + PAGE_SIZE, 0, walk);
+	}
+#endif
+}
+
 static void smaps_pte_entry(pte_t *pte, unsigned long addr,
 		struct mm_walk *walk)
 {
@@ -518,12 +527,8 @@ static void smaps_pte_entry(pte_t *pte, unsigned long addr,
 			}
 		} else if (is_pfn_swap_entry(swpent))
 			page = pfn_swap_entry_to_page(swpent);
-	} else if (unlikely(IS_ENABLED(CONFIG_SHMEM) && mss->check_shmem_swap
-							&& pte_none(*pte))) {
-		page = xa_load(&vma->vm_file->f_mapping->i_pages,
-						linear_page_index(vma, addr));
-		if (xa_is_value(page))
-			mss->swap += PAGE_SIZE;
+	} else {
+		smaps_pte_hole_lookup(addr, walk);
 		return;
 	}
 
@@ -737,8 +742,6 @@ static void smap_gather_stats(struct vm_area_struct *vma,
 		return;
 
 #ifdef CONFIG_SHMEM
-	/* In case of smaps_rollup, reset the value from previous vma */
-	mss->check_shmem_swap = false;
 	if (vma->vm_file && shmem_mapping(vma->vm_file->f_mapping)) {
 		/*
 		 * For shared or readonly shmem mappings we know that all
@@ -756,7 +759,6 @@ static void smap_gather_stats(struct vm_area_struct *vma,
 					!(vma->vm_flags & VM_WRITE))) {
 			mss->swap += shmem_swapped;
 		} else {
-			mss->check_shmem_swap = true;
 			ops = &smaps_shmem_walk_ops;
 		}
 	}
-- 
2.31.1

Re: [PATCH 1/3] mm/smaps: Fix shmem pte hole swap calculation

[Vlastimil Babka]

On 9/17/21 18:47, Peter Xu wrote:
> The shmem swap calculation on the privately writable mappings are using wrong
> parameters as spotted by Vlastimil.  Fix them.  That's introduced in commit
> 48131e03ca4e, when rework shmem_swap_usage to shmem_partial_swap_usage.
> 
> Test program:
> 
> ==================
> 
> void main(void)
> {
>     char *buffer, *p;
>     int i, fd;
> 
>     fd = memfd_create("test", 0);
>     assert(fd > 0);
> 
>     /* isize==2M*3, fill in pages, swap them out */
>     ftruncate(fd, SIZE_2M * 3);
>     buffer = mmap(NULL, SIZE_2M * 3, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
>     assert(buffer);
>     for (i = 0, p = buffer; i < SIZE_2M * 3 / 4096; i++) {
>         *p = 1;
>         p += 4096;
>     }
>     madvise(buffer, SIZE_2M * 3, MADV_PAGEOUT);
>     munmap(buffer, SIZE_2M * 3);
> 
>     /*
>      * Remap with private+writtable mappings on partial of the inode (<= 2M*3),
>      * while the size must also be >= 2M*2 to make sure there's a none pmd so
>      * smaps_pte_hole will be triggered.
>      */
>     buffer = mmap(NULL, SIZE_2M * 2, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
>     printf("pid=%d, buffer=%p\n", getpid(), buffer);
> 
>     /* Check /proc/$PID/smap_rollup, should see 4MB swap */
>     sleep(1000000);
> }
> ==================
> 
> Before the patch, smaps_rollup shows <4MB swap and the number will be random
> depending on the alignment of the buffer of mmap() allocated.  After this
> patch, it'll show 4MB.
> 
> Fixes: 48131e03ca4e ("mm, proc: reduce cost of /proc/pid/smaps for unpopulated shmem mappings")

Thanks, too bad I didn't spot it when sending that patch :)

> Reported-by: Vlastimil Babka <vbabka@suse.cz>
> Signed-off-by: Peter Xu <peterx@redhat.com>

Reviewed-by: Vlastimil Babka <vbabka@suse.cz>

> ---
>  fs/proc/task_mmu.c | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
> index cf25be3e0321..2197f669e17b 100644
> --- a/fs/proc/task_mmu.c
> +++ b/fs/proc/task_mmu.c
> @@ -478,9 +478,11 @@ static int smaps_pte_hole(unsigned long addr, unsigned long end,
>  			  __always_unused int depth, struct mm_walk *walk)
>  {
>  	struct mem_size_stats *mss = walk->private;
> +	struct vm_area_struct *vma = walk->vma;
>  
> -	mss->swap += shmem_partial_swap_usage(
> -			walk->vma->vm_file->f_mapping, addr, end);
> +	mss->swap += shmem_partial_swap_usage(walk->vma->vm_file->f_mapping,
> +					      linear_page_index(vma, addr),
> +					      linear_page_index(vma, end));
>  
>  	return 0;
>  }
> 

Re: [PATCH 2/3] mm/smaps: Use vma->vm_pgoff directly when counting partial swap

[Vlastimil Babka]

On 9/17/21 18:47, Peter Xu wrote:
> As it's trying to cover the whole vma anyways, use direct vm_pgoff value and
> vma_pages() rather than linear_page_index.
> 
> Cc: Vlastimil Babka <vbabka@suse.cz>
> Cc: Hugh Dickins <hughd@google.com>
> Signed-off-by: Peter Xu <peterx@redhat.com>

Reviewed-by: Vlastimil Babka <vbabka@suse.cz>

> ---
>  mm/shmem.c | 5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/mm/shmem.c b/mm/shmem.c
> index 96ccf6e941aa..d2620db8c938 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -856,9 +856,8 @@ unsigned long shmem_swap_usage(struct vm_area_struct *vma)
>  		return swapped << PAGE_SHIFT;
>  
>  	/* Here comes the more involved part */
> -	return shmem_partial_swap_usage(mapping,
> -			linear_page_index(vma, vma->vm_start),
> -			linear_page_index(vma, vma->vm_end));
> +	return shmem_partial_swap_usage(mapping, vma->vm_pgoff,
> +					vma->vm_pgoff + vma_pages(vma));
>  }
>  
>  /*
> 

Re: [PATCH 3/3] mm/smaps: Simplify shmem handling of pte holes

[Vlastimil Babka]

On 9/17/21 18:47, Peter Xu wrote:
> Firstly, check_shmem_swap variable is actually not necessary, because it's
> always set with pte_hole hook; checking each would work.

Right...

> Meanwhile, the check within smaps_pte_entry is not easy to follow.  E.g.,
> pte_none() check is not needed as "!pte_present && !is_swap_pte" is the same.

Seems to be true, indeed.

> Since at it, use the pte_hole() helper rather than dup the page cache lookup.

pte_hole() is for checking a range and we are calling it for single page,
isnt't that causing larger overhead in the end? There's xarray involved, so
maybe Matthew will know best.

> Still keep the CONFIG_SHMEM part so the code can be optimized to nop for !SHMEM.
> 
> There will be a very slight functional change in smaps_pte_entry(), that for
> !SHMEM we'll return early for pte_none (before checking page==NULL), but that's
> even nicer.

I don't think this is true, 'unlikely(IS_ENABLED(CONFIG_SHMEM))' will be a
compile-time constant false and shortcut the rest of the 'if' evaluation
thus there will be no page check? Or I misunderstood.

> Cc: Hugh Dickins <hughd@google.com>
> Cc: Matthew Wilcox <willy@infradead.org>
> Cc: Vlastimil Babka <vbabka@suse.cz>
> Signed-off-by: Peter Xu <peterx@redhat.com>
> ---
>  fs/proc/task_mmu.c | 22 ++++++++++++----------
>  1 file changed, 12 insertions(+), 10 deletions(-)
> 
> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
> index 2197f669e17b..ad667dbc96f5 100644
> --- a/fs/proc/task_mmu.c
> +++ b/fs/proc/task_mmu.c
> @@ -397,7 +397,6 @@ struct mem_size_stats {
>  	u64 pss_shmem;
>  	u64 pss_locked;
>  	u64 swap_pss;
> -	bool check_shmem_swap;
>  };
>  
>  static void smaps_page_accumulate(struct mem_size_stats *mss,
> @@ -490,6 +489,16 @@ static int smaps_pte_hole(unsigned long addr, unsigned long end,
>  #define smaps_pte_hole		NULL
>  #endif /* CONFIG_SHMEM */
>  
> +static void smaps_pte_hole_lookup(unsigned long addr, struct mm_walk *walk)
> +{
> +#ifdef CONFIG_SHMEM
> +	if (walk->ops->pte_hole) {
> +		/* depth is not used */
> +		smaps_pte_hole(addr, addr + PAGE_SIZE, 0, walk);
> +	}
> +#endif
> +}
> +
>  static void smaps_pte_entry(pte_t *pte, unsigned long addr,
>  		struct mm_walk *walk)
>  {
> @@ -518,12 +527,8 @@ static void smaps_pte_entry(pte_t *pte, unsigned long addr,
>  			}
>  		} else if (is_pfn_swap_entry(swpent))
>  			page = pfn_swap_entry_to_page(swpent);
> -	} else if (unlikely(IS_ENABLED(CONFIG_SHMEM) && mss->check_shmem_swap
> -							&& pte_none(*pte))) {
> -		page = xa_load(&vma->vm_file->f_mapping->i_pages,
> -						linear_page_index(vma, addr));
> -		if (xa_is_value(page))
> -			mss->swap += PAGE_SIZE;
> +	} else {
> +		smaps_pte_hole_lookup(addr, walk);
>  		return;
>  	}
>  
> @@ -737,8 +742,6 @@ static void smap_gather_stats(struct vm_area_struct *vma,
>  		return;
>  
>  #ifdef CONFIG_SHMEM
> -	/* In case of smaps_rollup, reset the value from previous vma */
> -	mss->check_shmem_swap = false;
>  	if (vma->vm_file && shmem_mapping(vma->vm_file->f_mapping)) {
>  		/*
>  		 * For shared or readonly shmem mappings we know that all
> @@ -756,7 +759,6 @@ static void smap_gather_stats(struct vm_area_struct *vma,
>  					!(vma->vm_flags & VM_WRITE))) {
>  			mss->swap += shmem_swapped;
>  		} else {
> -			mss->check_shmem_swap = true;
>  			ops = &smaps_shmem_walk_ops;
>  		}
>  	}
> 

Re: [PATCH 3/3] mm/smaps: Simplify shmem handling of pte holes

[Peter Xu]

Hi, Vlastimil,

On Tue, Oct 05, 2021 at 01:15:05PM +0200, Vlastimil Babka wrote:
> > Since at it, use the pte_hole() helper rather than dup the page cache lookup.
> 
> pte_hole() is for checking a range and we are calling it for single page,
> isnt't that causing larger overhead in the end? There's xarray involved, so
> maybe Matthew will know best.

Per my understanding, pte_hole() calls xas_load() too at last, just like the
old code; it's just that the xas_for_each() of shmem_partial_swap_usage() will
only run one iteration, iiuc.

> 
> > Still keep the CONFIG_SHMEM part so the code can be optimized to nop for !SHMEM.
> > 
> > There will be a very slight functional change in smaps_pte_entry(), that for
> > !SHMEM we'll return early for pte_none (before checking page==NULL), but that's
> > even nicer.
> 
> I don't think this is true, 'unlikely(IS_ENABLED(CONFIG_SHMEM))' will be a
> compile-time constant false and shortcut the rest of the 'if' evaluation
> thus there will be no page check? Or I misunderstood.

The page check I was referring is this one in smaps_pte_entry():

	if (!page)
		return;

After the change, with !SHMEM the "else" block will be kept there (unlike the
old code as you mentioned it'll be optimized), the smaps_pte_hole_lookup() will
be noop so it'll be a direct "return" in that "else", then it should return a
bit earlier by not checking "!page" (because in that case pte_none must have
page==NULL).

Thanks,

-- 
Peter Xu