[RFC][PATCH 0/3] Extend the mincore() report bits

[RFC][PATCH 0/3] Extend the mincore() report bits

[Pavel Emelyanov]

Hi.

We're currently working on the "Checkpoint-Restart in the Userspace" project. The
aim of the project is to dump a full (as full as possible) state of a running process
set (e.g. an application or a container) and recreate them at exactly the same state
later [1]

One of the parts of this state-dumping is dumping the memory of the processes. Since
apps usually mmap() more memory, than they actually use, dumping the whole mapping
contents is too expensive. To reduce this data we currently use the mincore() syscall
to check which pages are currently in memory and should be dumped, and which are not
and can be skipped. This simple trick reduces the dump size greatly, buy has three
problems.

1. File pages, that were not mapped by a task, but that were brought to page cache
   somehow (readahead or shared lib usage by other task) are reported as present and we 
   dump them, while we can skip them, since reading the page from file again at restore 
   time gives the correct page;
2. File pages, that are mapped by private mapping but that are not yet cow-ed are also
   reported as present, but we can skip them as well -- they can be re-read from disk;
2. Pages, that are swapped out are not reported by existing mincore(), and we skip them,
   and this is a bug :(

That said, I propose to add 2 more bits to the mincore per-page report to address these
problems. Plz, find details in the respective patches.

[1]
The project homepage is at http://criu.org
The project sources can be found here:
tools:  https://github.com/cyrillos/crtools
kernel: https://github.com/cyrillos/linux-2.6

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

[PATCH 1/3] mincore: Introduce named constant for existing bit

[Pavel Emelyanov]

Signed-off-by: Pavel Emelyanov <xemul@parallels.com>

---
 include/linux/mman.h |    2 ++
 mm/huge_memory.c     |    2 +-
 mm/mincore.c         |   10 +++++-----
 3 files changed, 8 insertions(+), 6 deletions(-)

diff --git a/include/linux/mman.h b/include/linux/mman.h
index 8b74e9b..e4fda1e 100644
--- a/include/linux/mman.h
+++ b/include/linux/mman.h
@@ -10,6 +10,8 @@
 #define OVERCOMMIT_ALWAYS		1
 #define OVERCOMMIT_NEVER		2
 
+#define MINCORE_RESIDENT	0x1
+
 #ifdef __KERNEL__
 #include <linux/mm.h>
 #include <linux/percpu_counter.h>
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 36b3d98..4f87067 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -1045,7 +1045,7 @@ int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 			 * All logical pages in the range are present
 			 * if backed by a huge page.
 			 */
-			memset(vec, 1, (end - addr) >> PAGE_SHIFT);
+			memset(vec, MINCORE_RESIDENT, (end - addr) >> PAGE_SHIFT);
 		}
 	} else
 		spin_unlock(&vma->vm_mm->page_table_lock);
diff --git a/mm/mincore.c b/mm/mincore.c
index 636a868..b719cdd 100644
--- a/mm/mincore.c
+++ b/mm/mincore.c
@@ -38,7 +38,7 @@ static void mincore_hugetlb_page_range(struct vm_area_struct *vma,
 				       addr & huge_page_mask(h));
 		present = ptep && !huge_pte_none(huge_ptep_get(ptep));
 		while (1) {
-			*vec = present;
+			*vec = (present ? MINCORE_RESIDENT : 0);
 			vec++;
 			addr += PAGE_SIZE;
 			if (addr == end)
@@ -83,7 +83,7 @@ static unsigned char mincore_page(struct address_space *mapping, pgoff_t pgoff)
 		page_cache_release(page);
 	}
 
-	return present;
+	return present ? MINCORE_RESIDENT : 0;
 }
 
 static void mincore_unmapped_range(struct vm_area_struct *vma,
@@ -122,7 +122,7 @@ static void mincore_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 		if (pte_none(pte))
 			mincore_unmapped_range(vma, addr, next, vec);
 		else if (pte_present(pte))
-			*vec = 1;
+			*vec = MINCORE_RESIDENT;
 		else if (pte_file(pte)) {
 			pgoff = pte_to_pgoff(pte);
 			*vec = mincore_page(vma->vm_file->f_mapping, pgoff);
@@ -131,14 +131,14 @@ static void mincore_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 
 			if (is_migration_entry(entry)) {
 				/* migration entries are always uptodate */
-				*vec = 1;
+				*vec = MINCORE_RESIDENT;
 			} else {
 #ifdef CONFIG_SWAP
 				pgoff = entry.val;
 				*vec = mincore_page(&swapper_space, pgoff);
 #else
 				WARN_ON(1);
-				*vec = 1;
+				*vec = MINCORE_RESIDENT;
 #endif
 			}
 		}
-- 
1.5.5.6

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

[PATCH 2/3] mincore: Introduce the MINCORE_ANON bit

[Pavel Emelyanov]

When creating a memory dump of a running application it is better to have
the smallest possible set of pages. Using the existing mincore() bit helps,
but not much -- it repotes pages from page cache, which have not necessarily
being mapped by an application, and pages from private file mappings, that
are not yet being cow-ed and thus their contents doesn't differ from file.

Introduce the 2nd bit of mincore, that reports whether a page is not backed
by a file on disk. I.e. all pages from anonymous mappings and those pages
from private file mappings, that have already being cow-ed by write.

Signed-off-by: Pavel Emelyanov <xemul@parallels.com>

---
 include/linux/mman.h |    1 +
 mm/mincore.c         |   15 +++++++++++++--
 2 files changed, 14 insertions(+), 2 deletions(-)

diff --git a/include/linux/mman.h b/include/linux/mman.h
index e4fda1e..9d1de16 100644
--- a/include/linux/mman.h
+++ b/include/linux/mman.h
@@ -11,6 +11,7 @@
 #define OVERCOMMIT_NEVER		2
 
 #define MINCORE_RESIDENT	0x1
+#define MINCORE_ANON		0x2
 
 #ifdef __KERNEL__
 #include <linux/mm.h>
diff --git a/mm/mincore.c b/mm/mincore.c
index b719cdd..3163dfb 100644
--- a/mm/mincore.c
+++ b/mm/mincore.c
@@ -38,7 +38,7 @@ static void mincore_hugetlb_page_range(struct vm_area_struct *vma,
 				       addr & huge_page_mask(h));
 		present = ptep && !huge_pte_none(huge_ptep_get(ptep));
 		while (1) {
-			*vec = (present ? MINCORE_RESIDENT : 0);
+			*vec = (present ? MINCORE_RESIDENT : 0) | MINCORE_ANON;
 			vec++;
 			addr += PAGE_SIZE;
 			if (addr == end)
@@ -86,6 +86,17 @@ static unsigned char mincore_page(struct address_space *mapping, pgoff_t pgoff)
 	return present ? MINCORE_RESIDENT : 0;
 }
 
+static unsigned char mincore_pte(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
+{
+	struct page *pg;
+
+	pg = vm_normal_page(vma, addr, pte);
+	if (!pg)
+		return 0;
+	else
+		return PageAnon(pg) ? MINCORE_ANON : 0;
+}
+
 static void mincore_unmapped_range(struct vm_area_struct *vma,
 				unsigned long addr, unsigned long end,
 				unsigned char *vec)
@@ -122,7 +133,7 @@ static void mincore_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 		if (pte_none(pte))
 			mincore_unmapped_range(vma, addr, next, vec);
 		else if (pte_present(pte))
-			*vec = MINCORE_RESIDENT;
+			*vec = MINCORE_RESIDENT | mincore_pte(vma, addr, pte);
 		else if (pte_file(pte)) {
 			pgoff = pte_to_pgoff(pte);
 			*vec = mincore_page(vma->vm_file->f_mapping, pgoff);
-- 
1.5.5.6

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/3] mincore: Introduce the MINCORE_ANON bit

[KOSAKI Motohiro]

> +static unsigned char mincore_pte(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
> +{
> +       struct page *pg;
> +
> +       pg = vm_normal_page(vma, addr, pte);
> +       if (!pg)
> +               return 0;
> +       else
> +               return PageAnon(pg) ? MINCORE_ANON : 0;
> +}
> +

How do your program handle tmpfs pages (and/or ram device pages)?

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/3] mincore: Introduce the MINCORE_ANON bit

[Pavel Emelyanov]

On 12/26/2011 04:05 AM, KOSAKI Motohiro wrote:
>> +static unsigned char mincore_pte(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
>> +{
>> +       struct page *pg;
>> +
>> +       pg = vm_normal_page(vma, addr, pte);
>> +       if (!pg)
>> +               return 0;
>> +       else
>> +               return PageAnon(pg) ? MINCORE_ANON : 0;
>> +}
>> +
> 
> How do your program handle tmpfs pages (and/or ram device pages)?
> .

Do you mean mapped files from tmpfs? Currently just any other file.
Do you see problems with this patch wrt tmpfs?

Thanks,
Pavel

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/3] mincore: Introduce the MINCORE_ANON bit

[KOSAKI Motohiro]

2011/12/26 Pavel Emelyanov <xemul@parallels.com>:
> On 12/26/2011 04:05 AM, KOSAKI Motohiro wrote:
>>> +static unsigned char mincore_pte(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
>>> +{
>>> +       struct page *pg;
>>> +
>>> +       pg = vm_normal_page(vma, addr, pte);
>>> +       if (!pg)
>>> +               return 0;
>>> +       else
>>> +               return PageAnon(pg) ? MINCORE_ANON : 0;
>>> +}
>>> +
>>
>> How do your program handle tmpfs pages (and/or ram device pages)?
>
> Do you mean mapped files from tmpfs? Currently just any other file.
> Do you see problems with this patch wrt tmpfs?

If you don't save mapped file on tmpfs or other volatile devices, the process
might not restored. The data might already destroyed. The common strategy
are two,

1) save all opened file by different ways.
2) save all mapped file even though clean file cache.

In both case, we don't reduce freezed data size. So, I'm interesting
you strategy.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/3] mincore: Introduce the MINCORE_ANON bit

[Pavel Emelyanov]

On 12/26/2011 08:35 PM, KOSAKI Motohiro wrote:
> 2011/12/26 Pavel Emelyanov <xemul@parallels.com>:
>> On 12/26/2011 04:05 AM, KOSAKI Motohiro wrote:
>>>> +static unsigned char mincore_pte(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
>>>> +{
>>>> +       struct page *pg;
>>>> +
>>>> +       pg = vm_normal_page(vma, addr, pte);
>>>> +       if (!pg)
>>>> +               return 0;
>>>> +       else
>>>> +               return PageAnon(pg) ? MINCORE_ANON : 0;
>>>> +}
>>>> +
>>>
>>> How do your program handle tmpfs pages (and/or ram device pages)?
>>
>> Do you mean mapped files from tmpfs? Currently just any other file.
>> Do you see problems with this patch wrt tmpfs?
> 
> If you don't save mapped file on tmpfs or other volatile devices, the process
> might not restored. The data might already destroyed. 

Yes I know this, thanks :\

> The common strategy are two,
> 
> 1) save all opened file by different ways.
> 2) save all mapped file even though clean file cache.
> 
> In both case, we don't reduce freezed data size. So, I'm interesting
> you strategy.

The tmpfs contents itself is supposed to be preserved, it's not a problem. The problem I'm trying
to solve here is which page from task mappings (i.e. vm_area_struct-s) to save and which not to.

Do the proposed MINCORE_RESIDENT and MINCORE_ANON bits have problems with this from
your POV?

Thanks,
Pavel

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/3] mincore: Introduce the MINCORE_ANON bit

[KOSAKI Motohiro]

> The tmpfs contents itself is supposed to be preserved, it's not a problem. The problem I'm trying
> to solve here is which page from task mappings (i.e. vm_area_struct-s) to save and which not to.
>
> Do the proposed MINCORE_RESIDENT and MINCORE_ANON bits have problems with this from
> your POV?

No problem.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

[PATCH 3/3] mincore: Introduce the MINCORE_SWAP bit

[Pavel Emelyanov]

When a page is swapped out it should be included in the memory dump,
but the existing mincore() doesn't report the set bit for such pages.

Thus add a bit for swapped out pages.

Signed-off-by: Pavel Emelyanov <xemul@parallels.com>

---
 include/linux/mman.h |    1 +
 mm/mincore.c         |    2 +-
 2 files changed, 2 insertions(+), 1 deletions(-)

diff --git a/include/linux/mman.h b/include/linux/mman.h
index 9d1de16..bfe4038 100644
--- a/include/linux/mman.h
+++ b/include/linux/mman.h
@@ -12,6 +12,7 @@
 
 #define MINCORE_RESIDENT	0x1
 #define MINCORE_ANON		0x2
+#define MINCORE_SWAP		0x4
 
 #ifdef __KERNEL__
 #include <linux/mm.h>
diff --git a/mm/mincore.c b/mm/mincore.c
index 3163dfb..82c5c3e 100644
--- a/mm/mincore.c
+++ b/mm/mincore.c
@@ -146,7 +146,7 @@ static void mincore_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 			} else {
 #ifdef CONFIG_SWAP
 				pgoff = entry.val;
-				*vec = mincore_page(&swapper_space, pgoff);
+				*vec = mincore_page(&swapper_space, pgoff) | MINCORE_SWAP;
 #else
 				WARN_ON(1);
 				*vec = MINCORE_RESIDENT;
-- 
1.5.5.6

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 3/3] mincore: Introduce the MINCORE_SWAP bit

[Konstantin Khlebnikov]

Pavel Emelyanov wrote:
> When a page is swapped out it should be included in the memory dump,
> but the existing mincore() doesn't report the set bit for such pages.
>
> Thus add a bit for swapped out pages.

We can reuse ANON bit there.
Plus you forget to handle migration-entries, they can be both file and anon too.

>
> Signed-off-by: Pavel Emelyanov<xemul@parallels.com>
>
> ---
>   include/linux/mman.h |    1 +
>   mm/mincore.c         |    2 +-
>   2 files changed, 2 insertions(+), 1 deletions(-)
>
> diff --git a/include/linux/mman.h b/include/linux/mman.h
> index 9d1de16..bfe4038 100644
> --- a/include/linux/mman.h
> +++ b/include/linux/mman.h
> @@ -12,6 +12,7 @@
>
>   #define MINCORE_RESIDENT	0x1
>   #define MINCORE_ANON		0x2
> +#define MINCORE_SWAP		0x4
>
>   #ifdef __KERNEL__
>   #include<linux/mm.h>
> diff --git a/mm/mincore.c b/mm/mincore.c
> index 3163dfb..82c5c3e 100644
> --- a/mm/mincore.c
> +++ b/mm/mincore.c
> @@ -146,7 +146,7 @@ static void mincore_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
>   			} else {
>   #ifdef CONFIG_SWAP
>   				pgoff = entry.val;
> -				*vec = mincore_page(&swapper_space, pgoff);
> +				*vec = mincore_page(&swapper_space, pgoff) | MINCORE_SWAP;
>   #else
>   				WARN_ON(1);
>   				*vec = MINCORE_RESIDENT;

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>