[RFC] mm:change /proc/smaps caculation behavior

[RFC] mm:change /proc/smaps caculation behavior

[Wang, Yalin]

This patch change smaps pagetable walk behavior, to make
sure not skip VM_PFNMAP pagetables,
so that we can calculate COW pages of VM_PFNMAP as normal pages.

Signed-off-by: Yalin Wang <yalin.wang@sonymobile.com>
---
 fs/proc/task_mmu.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
index c7267e9..00a5b73 100644
--- a/fs/proc/task_mmu.c
+++ b/fs/proc/task_mmu.c
@@ -503,6 +503,15 @@ static void smaps_pte_entry(pte_t *pte, unsigned long addr,
 	smaps_account(mss, page, PAGE_SIZE, pte_young(*pte), pte_dirty(*pte));
 }
 
+static int smaps_test_walk(unsigned long addr, unsigned long next,
+		struct mm_walk *walk)
+{
+	/*
+	 * don't skip VM_PFNMAP, so that we can caculate some COW pages.
+	 */
+	return 0;
+}
+
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 static void smaps_pmd_entry(pmd_t *pmd, unsigned long addr,
 		struct mm_walk *walk)
@@ -616,6 +625,7 @@ static int show_smap(struct seq_file *m, void *v, int is_pid)
 	struct mem_size_stats mss;
 	struct mm_walk smaps_walk = {
 		.pmd_entry = smaps_pte_range,
+		.test_walk = smaps_test_walk,
 		.mm = vma->vm_mm,
 		.private = &mss,
 	};
-- 
2.2.2

[RFC V2] mm:change smaps/pagemap_read calculation behavior

[Wang, Yalin]

This patch change smaps/pagemap_read pagetable walk behavior, to make
sure not skip VM_PFNMAP pagetables,
so that we can calculate COW pages of VM_PFNMAP as normal pages.

Signed-off-by: Yalin Wang <yalin.wang@sonymobile.com>
---
 fs/proc/task_mmu.c | 2 ++
 include/linux/mm.h | 2 ++
 mm/pagewalk.c      | 5 +++++
 3 files changed, 9 insertions(+)

diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
index c7267e9..e7d7c43 100644
--- a/fs/proc/task_mmu.c
+++ b/fs/proc/task_mmu.c
@@ -616,6 +616,7 @@ static int show_smap(struct seq_file *m, void *v, int is_pid)
 	struct mem_size_stats mss;
 	struct mm_walk smaps_walk = {
 		.pmd_entry = smaps_pte_range,
+		.test_walk = generic_walk_page_test_no_skip,
 		.mm = vma->vm_mm,
 		.private = &mss,
 	};
@@ -1264,6 +1265,7 @@ static ssize_t pagemap_read(struct file *file, char __user *buf,
 
 	pagemap_walk.pmd_entry = pagemap_pte_range;
 	pagemap_walk.pte_hole = pagemap_pte_hole;
+	pagemap_walk.test_walk = generic_walk_page_test_no_skip;
 #ifdef CONFIG_HUGETLB_PAGE
 	pagemap_walk.hugetlb_entry = pagemap_hugetlb_range;
 #endif
diff --git a/include/linux/mm.h b/include/linux/mm.h
index b976d9f..07f71c5 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1191,6 +1191,8 @@ struct mm_walk {
 	void *private;
 };
 
+int generic_walk_page_test_no_skip(unsigned long start, unsigned long end,
+		struct mm_walk *walk);
 int walk_page_range(unsigned long addr, unsigned long end,
 		struct mm_walk *walk);
 int walk_page_vma(struct vm_area_struct *vma, struct mm_walk *walk);
diff --git a/mm/pagewalk.c b/mm/pagewalk.c
index 75c1f28..14f38d5 100644
--- a/mm/pagewalk.c
+++ b/mm/pagewalk.c
@@ -206,6 +206,11 @@ static int __walk_page_range(unsigned long start, unsigned long end,
 	return err;
 }
 
+int generic_walk_page_test_no_skip(unsigned long start, unsigned long end,
+			struct mm_walk *walk)
+{
+	return 0;
+}
 /**
  * walk_page_range - walk page table with caller specific callbacks
  *
-- 
2.2.2

Re: [RFC V2] mm:change smaps/pagemap_read calculation behavior

[Naoya Horiguchi]

On Fri, Jan 30, 2015 at 03:47:54PM +0800, Wang, Yalin wrote:
> This patch change smaps/pagemap_read pagetable walk behavior, to make
> sure not skip VM_PFNMAP pagetables,
> so that we can calculate COW pages of VM_PFNMAP as normal pages.
> 
> Signed-off-by: Yalin Wang <yalin.wang@sonymobile.com>

Hi Yalin,

The original motivation of the VM_PFNMAP code in pagewalk.c comes from the
following patch:

  commit a9ff785e4437c83d2179161e012f5bdfbd6381f0
  Author: Cliff Wickman <cpw@sgi.com>
  Date:   Fri May 24 15:55:36 2013 -0700
  
      mm/pagewalk.c: walk_page_range should avoid VM_PFNMAP areas

, where Cliff stated that some kind of vma(VM_PFNMAP) caused kernel panic
when walk_page_range() was called over it. So I don't think that re-enabling
to walk over every vma(VM_PFNMAP) unexceptionally is a good idea.

If you really want to get some information from a vma(VM_PFNMAP) via these
interfaces, I recommend you to implement proper judging code which returns
0 for your vma(VM_PFNMAP) and returns 1 for Cliff's vma(VM_PFNMAP).

Thanks,
Naoya Horiguchi

> ---
>  fs/proc/task_mmu.c | 2 ++
>  include/linux/mm.h | 2 ++
>  mm/pagewalk.c      | 5 +++++
>  3 files changed, 9 insertions(+)
> 
> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
> index c7267e9..e7d7c43 100644
> --- a/fs/proc/task_mmu.c
> +++ b/fs/proc/task_mmu.c
> @@ -616,6 +616,7 @@ static int show_smap(struct seq_file *m, void *v, int is_pid)
>  	struct mem_size_stats mss;
>  	struct mm_walk smaps_walk = {
>  		.pmd_entry = smaps_pte_range,
> +		.test_walk = generic_walk_page_test_no_skip,
>  		.mm = vma->vm_mm,
>  		.private = &mss,
>  	};
> @@ -1264,6 +1265,7 @@ static ssize_t pagemap_read(struct file *file, char __user *buf,
>  
>  	pagemap_walk.pmd_entry = pagemap_pte_range;
>  	pagemap_walk.pte_hole = pagemap_pte_hole;
> +	pagemap_walk.test_walk = generic_walk_page_test_no_skip;
>  #ifdef CONFIG_HUGETLB_PAGE
>  	pagemap_walk.hugetlb_entry = pagemap_hugetlb_range;
>  #endif
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index b976d9f..07f71c5 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -1191,6 +1191,8 @@ struct mm_walk {
>  	void *private;
>  };
>  
> +int generic_walk_page_test_no_skip(unsigned long start, unsigned long end,
> +		struct mm_walk *walk);
>  int walk_page_range(unsigned long addr, unsigned long end,
>  		struct mm_walk *walk);
>  int walk_page_vma(struct vm_area_struct *vma, struct mm_walk *walk);
> diff --git a/mm/pagewalk.c b/mm/pagewalk.c
> index 75c1f28..14f38d5 100644
> --- a/mm/pagewalk.c
> +++ b/mm/pagewalk.c
> @@ -206,6 +206,11 @@ static int __walk_page_range(unsigned long start, unsigned long end,
>  	return err;
>  }
>  
> +int generic_walk_page_test_no_skip(unsigned long start, unsigned long end,
> +			struct mm_walk *walk)
> +{
> +	return 0;
> +}
>  /**
>   * walk_page_range - walk page table with caller specific callbacks
>   *
> -- 
> 2.2.2
> 

RE: [RFC V2] mm:change smaps/pagemap_read calculation behavior

[Wang, Yalin]

> -----Original Message-----
> From: Naoya Horiguchi [mailto:n-horiguchi@ah.jp.nec.com]
> Sent: Friday, January 30, 2015 4:24 PM
> To: Wang, Yalin
> Cc: 'akpm@linux-foundation.org'; 'kirill.shutemov@linux.intel.com';
> 'oleg@redhat.com'; 'gorcunov@openvz.org'; 'pfeiner@google.com';
> 'aquini@redhat.com'; 'linux-kernel@vger.kernel.org'
> Subject: Re: [RFC V2] mm:change smaps/pagemap_read calculation behavior
> 
> On Fri, Jan 30, 2015 at 03:47:54PM +0800, Wang, Yalin wrote:
> > This patch change smaps/pagemap_read pagetable walk behavior, to make
> > sure not skip VM_PFNMAP pagetables,
> > so that we can calculate COW pages of VM_PFNMAP as normal pages.
> >
> > Signed-off-by: Yalin Wang <yalin.wang@sonymobile.com>
> 
> Hi Yalin,
> 
> The original motivation of the VM_PFNMAP code in pagewalk.c comes from the
> following patch:
> 
>   commit a9ff785e4437c83d2179161e012f5bdfbd6381f0
>   Author: Cliff Wickman <cpw@sgi.com>
>   Date:   Fri May 24 15:55:36 2013 -0700
> 
>       mm/pagewalk.c: walk_page_range should avoid VM_PFNMAP areas
> 
> , where Cliff stated that some kind of vma(VM_PFNMAP) caused kernel panic
> when walk_page_range() was called over it. So I don't think that re-
> enabling
> to walk over every vma(VM_PFNMAP) unexceptionally is a good idea.
> 
> If you really want to get some information from a vma(VM_PFNMAP) via these
> interfaces, I recommend you to implement proper judging code which returns
> 0 for your vma(VM_PFNMAP) and returns 1 for Cliff's vma(VM_PFNMAP).
> 
I see, but I am curious that why kernel panic when I just
access process pagetables in page_table_walk()?
Is it caused by hardware problem?

The reason that I want to enable it is to see some drivers map some COW pages 
With VM_PFNMAP, so that user space can get correct page allocation info for 
These COW special pages with VM_PFNMAP flag.

Thanks