[RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Balbir Singh]

Found-by: Hugh Dickins <hugh@veritas.com>

mem_cgroup_charge() in unuse_pte() is called under a lock, the pte_lock. That's
clearly incorrect, since we pass GFP_KERNEL to mem_cgroup_charge() for
allocation of page_cgroup.

This patch release the lock and reacquires the lock after the call to
mem_cgroup_charge().

Tested on a powerpc box by calling swapoff in the middle of a cgroup
running a workload that pushes pages to swap.

Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com>
---

 mm/swapfile.c |   16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)

diff -puN mm/swapfile.c~memory-controller-fix-unuse-pte-charging mm/swapfile.c
--- linux-2.6.23-rc8/mm/swapfile.c~memory-controller-fix-unuse-pte-charging	2007-10-03 13:45:56.000000000 +0530
+++ linux-2.6.23-rc8-balbir/mm/swapfile.c	2007-10-05 08:49:54.000000000 +0530
@@ -507,11 +507,18 @@ unsigned int count_swap_pages(int type, 
  * just let do_wp_page work it out if a write is requested later - to
  * force COW, vm_page_prot omits write permission from any private vma.
  */
-static int unuse_pte(struct vm_area_struct *vma, pte_t *pte,
-		unsigned long addr, swp_entry_t entry, struct page *page)
+static int unuse_pte(struct vm_area_struct *vma, pte_t *pte, pmd_t *pmd,
+		unsigned long addr, swp_entry_t entry, struct page *page,
+		spinlock_t **ptl)
 {
-	if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
+	pte_unmap_unlock(pte - 1, *ptl);
+
+	if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL)) {
+		pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
 		return -ENOMEM;
+	}
+
+	pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
 
 	inc_mm_counter(vma->vm_mm, anon_rss);
 	get_page(page);
@@ -543,7 +550,8 @@ static int unuse_pte_range(struct vm_are
 		 * Test inline before going to call unuse_pte.
 		 */
 		if (unlikely(pte_same(*pte, swp_pte))) {
-			ret = unuse_pte(vma, pte++, addr, entry, page);
+			ret = unuse_pte(vma, pte++, pmd, addr, entry, page,
+					&ptl);
 			break;
 		}
 	} while (pte++, addr += PAGE_SIZE, addr != end);
_

-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Hugh Dickins]

On Fri, 5 Oct 2007, Balbir Singh wrote:
> 
> Found-by: Hugh Dickins <hugh@veritas.com>
> 
> mem_cgroup_charge() in unuse_pte() is called under a lock, the pte_lock. That's
> clearly incorrect, since we pass GFP_KERNEL to mem_cgroup_charge() for
> allocation of page_cgroup.
> 
> This patch release the lock and reacquires the lock after the call to
> mem_cgroup_charge().
> 
> Tested on a powerpc box by calling swapoff in the middle of a cgroup
> running a workload that pushes pages to swap.

Hard to test it adequately at present, while that call
to mem_cgroup_charge is never allocating anything new.

Sorry, it's a bit ugly (the intertwining of unuse_pte and its caller),
it's got a bug, and fixing that bug makes it uglier.

The bug is that you ignore the pte ptr returned by pte_offset_map_lock:
we could be preempted on to a different cpu just there, so a different
cpu's kmap_atomic area used, with a different pte pointer; which would
need to be passed back to the caller for when it unmaps.

I much prefer my patch appended further down: considering how it's safe
for you to drop the ptl there because of holding page lock, pushed me
into seeing that we can actually do our scanning without ptl, which in
many configurations has the advantage of staying preemptible (though
preemptible swapoff is not terribly high on anyone's ticklist ;).

But you may well prefer that we split it into two: with me taking
responsibility and blame for the preliminary patch which relaxes
the locking, and you then adding the mem_cgroup_charge (and the
exceptional mem_cgroup_uncharge_page) on top of that.

Hugh

> 
> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com>
> ---
> 
>  mm/swapfile.c |   16 ++++++++++++----
>  1 file changed, 12 insertions(+), 4 deletions(-)
> 
> diff -puN mm/swapfile.c~memory-controller-fix-unuse-pte-charging mm/swapfile.c
> --- linux-2.6.23-rc8/mm/swapfile.c~memory-controller-fix-unuse-pte-charging	2007-10-03 13:45:56.000000000 +0530
> +++ linux-2.6.23-rc8-balbir/mm/swapfile.c	2007-10-05 08:49:54.000000000 +0530
> @@ -507,11 +507,18 @@ unsigned int count_swap_pages(int type, 
>   * just let do_wp_page work it out if a write is requested later - to
>   * force COW, vm_page_prot omits write permission from any private vma.
>   */
> -static int unuse_pte(struct vm_area_struct *vma, pte_t *pte,
> -		unsigned long addr, swp_entry_t entry, struct page *page)
> +static int unuse_pte(struct vm_area_struct *vma, pte_t *pte, pmd_t *pmd,
> +		unsigned long addr, swp_entry_t entry, struct page *page,
> +		spinlock_t **ptl)
>  {
> -	if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
> +	pte_unmap_unlock(pte - 1, *ptl);
> +
> +	if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL)) {
> +		pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
>  		return -ENOMEM;
> +	}
> +
> +	pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
>  
>  	inc_mm_counter(vma->vm_mm, anon_rss);
>  	get_page(page);
> @@ -543,7 +550,8 @@ static int unuse_pte_range(struct vm_are
>  		 * Test inline before going to call unuse_pte.
>  		 */
>  		if (unlikely(pte_same(*pte, swp_pte))) {
> -			ret = unuse_pte(vma, pte++, addr, entry, page);
> +			ret = unuse_pte(vma, pte++, pmd, addr, entry, page,
> +					&ptl);
>  			break;
>  		}
>  	} while (pte++, addr += PAGE_SIZE, addr != end);

--- 2.6.23-rc8-mm2/mm/swapfile.c	2007-09-27 12:03:36.000000000 +0100
+++ linux/mm/swapfile.c	2007-10-07 14:33:05.000000000 +0100
@@ -507,11 +507,23 @@ unsigned int count_swap_pages(int type, 
  * just let do_wp_page work it out if a write is requested later - to
  * force COW, vm_page_prot omits write permission from any private vma.
  */
-static int unuse_pte(struct vm_area_struct *vma, pte_t *pte,
+static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
 		unsigned long addr, swp_entry_t entry, struct page *page)
 {
+	spinlock_t *ptl;
+	pte_t *pte;
+	int ret = 1;
+
 	if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
-		return -ENOMEM;
+		ret = -ENOMEM;
+
+	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
+	if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
+		if (ret > 0)
+			mem_cgroup_uncharge_page(page);
+		ret = 0;
+		goto out;
+	}
 
 	inc_mm_counter(vma->vm_mm, anon_rss);
 	get_page(page);
@@ -524,7 +536,9 @@ static int unuse_pte(struct vm_area_stru
 	 * immediately swapped out again after swapon.
 	 */
 	activate_page(page);
-	return 1;
+out:
+	pte_unmap_unlock(pte, ptl);
+	return ret;
 }
 
 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
@@ -533,21 +547,33 @@ static int unuse_pte_range(struct vm_are
 {
 	pte_t swp_pte = swp_entry_to_pte(entry);
 	pte_t *pte;
-	spinlock_t *ptl;
 	int ret = 0;
 
-	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
+	/*
+	 * We don't actually need pte lock while scanning for swp_pte:
+	 * since we hold page lock, swp_pte cannot be inserted into or
+	 * removed from a page table while we're scanning; but on some
+	 * architectures (e.g. i386 with PAE) we might catch a glimpse
+	 * of unmatched parts which look like swp_pte, so unuse_pte
+	 * must recheck under pte lock.  Scanning without the lock
+	 * is preemptible if CONFIG_PREEMPT without CONFIG_HIGHPTE.
+	 */
+	pte = pte_offset_map(pmd, addr);
 	do {
 		/*
 		 * swapoff spends a _lot_ of time in this loop!
 		 * Test inline before going to call unuse_pte.
 		 */
 		if (unlikely(pte_same(*pte, swp_pte))) {
-			ret = unuse_pte(vma, pte++, addr, entry, page);
-			break;
+			pte_unmap(pte);
+			ret = unuse_pte(vma, pmd, addr, entry, page);
+			if (ret)
+				goto out;
+			pte = pte_offset_map(pmd, addr);
 		}
 	} while (pte++, addr += PAGE_SIZE, addr != end);
-	pte_unmap_unlock(pte - 1, ptl);
+	pte_unmap(pte - 1);
+out:
 	return ret;
 }
 

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Balbir Singh]

Hugh Dickins wrote:
> On Fri, 5 Oct 2007, Balbir Singh wrote:
>> Found-by: Hugh Dickins <hugh@veritas.com>
>>
>> mem_cgroup_charge() in unuse_pte() is called under a lock, the pte_lock. That's
>> clearly incorrect, since we pass GFP_KERNEL to mem_cgroup_charge() for
>> allocation of page_cgroup.
>>
>> This patch release the lock and reacquires the lock after the call to
>> mem_cgroup_charge().
>>
>> Tested on a powerpc box by calling swapoff in the middle of a cgroup
>> running a workload that pushes pages to swap.
> 
> Hard to test it adequately at present, while that call
> to mem_cgroup_charge is never allocating anything new.
> 

Yes, your right!

> Sorry, it's a bit ugly (the intertwining of unuse_pte and its caller),
> it's got a bug, and fixing that bug makes it uglier.
> 
> The bug is that you ignore the pte ptr returned by pte_offset_map_lock:
> we could be preempted on to a different cpu just there, so a different
> cpu's kmap_atomic area used, with a different pte pointer; which would
> need to be passed back to the caller for when it unmaps.
> 

Good point! I forgot that we unmap the pte when we unlock it

> I much prefer my patch appended further down: considering how it's safe
> for you to drop the ptl there because of holding page lock, pushed me
> into seeing that we can actually do our scanning without ptl, which in
> many configurations has the advantage of staying preemptible (though
> preemptible swapoff is not terribly high on anyone's ticklist ;).
> 

I like your patch, my comments on it below

> But you may well prefer that we split it into two: with me taking
> responsibility and blame for the preliminary patch which relaxes
> the locking, and you then adding the mem_cgroup_charge (and the
> exceptional mem_cgroup_uncharge_page) on top of that.
> 

Sounds good, you could submit both parts to Andrew. I think
Andrew would like to split up the patches as well, so that
the major change of scanning without the lock and the
memory controller fix are two different patches. My changes
are pretty trivial and well covered under your patch.

> Hugh
> 
>> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com>
>> ---
>>
>>  mm/swapfile.c |   16 ++++++++++++----
>>  1 file changed, 12 insertions(+), 4 deletions(-)
>>
>> diff -puN mm/swapfile.c~memory-controller-fix-unuse-pte-charging mm/swapfile.c
>> --- linux-2.6.23-rc8/mm/swapfile.c~memory-controller-fix-unuse-pte-charging	2007-10-03 13:45:56.000000000 +0530
>> +++ linux-2.6.23-rc8-balbir/mm/swapfile.c	2007-10-05 08:49:54.000000000 +0530
>> @@ -507,11 +507,18 @@ unsigned int count_swap_pages(int type, 
>>   * just let do_wp_page work it out if a write is requested later - to
>>   * force COW, vm_page_prot omits write permission from any private vma.
>>   */
>> -static int unuse_pte(struct vm_area_struct *vma, pte_t *pte,
>> -		unsigned long addr, swp_entry_t entry, struct page *page)
>> +static int unuse_pte(struct vm_area_struct *vma, pte_t *pte, pmd_t *pmd,
>> +		unsigned long addr, swp_entry_t entry, struct page *page,
>> +		spinlock_t **ptl)
>>  {
>> -	if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
>> +	pte_unmap_unlock(pte - 1, *ptl);
>> +
>> +	if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL)) {
>> +		pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
>>  		return -ENOMEM;
>> +	}
>> +
>> +	pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
>>  
>>  	inc_mm_counter(vma->vm_mm, anon_rss);
>>  	get_page(page);
>> @@ -543,7 +550,8 @@ static int unuse_pte_range(struct vm_are
>>  		 * Test inline before going to call unuse_pte.
>>  		 */
>>  		if (unlikely(pte_same(*pte, swp_pte))) {
>> -			ret = unuse_pte(vma, pte++, addr, entry, page);
>> +			ret = unuse_pte(vma, pte++, pmd, addr, entry, page,
>> +					&ptl);
>>  			break;
>>  		}
>>  	} while (pte++, addr += PAGE_SIZE, addr != end);
> 
> --- 2.6.23-rc8-mm2/mm/swapfile.c	2007-09-27 12:03:36.000000000 +0100
> +++ linux/mm/swapfile.c	2007-10-07 14:33:05.000000000 +0100
> @@ -507,11 +507,23 @@ unsigned int count_swap_pages(int type, 
>   * just let do_wp_page work it out if a write is requested later - to
>   * force COW, vm_page_prot omits write permission from any private vma.
>   */
> -static int unuse_pte(struct vm_area_struct *vma, pte_t *pte,
> +static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
>  		unsigned long addr, swp_entry_t entry, struct page *page)
>  {
> +	spinlock_t *ptl;
> +	pte_t *pte;
> +	int ret = 1;
> +
>  	if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
> -		return -ENOMEM;
> +		ret = -ENOMEM;
> +


With this change I think, ret = mem_cgroup_charge(....) makes
more sense

> +	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);

We map the pte in unuse_pte_range() and unmap it before
calling this routing, can't we keep it mapped and acquire the lock
here? Looking through pte_offset_map_lock, it did not seem
like mapping and acquiring the lock were indivisible operations.

> +	if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
> +		if (ret > 0)
> +			mem_cgroup_uncharge_page(page);


Then we can check for ret >= 0

> +		ret = 0;
> +		goto out;
> +	}
> 
>  	inc_mm_counter(vma->vm_mm, anon_rss);
>  	get_page(page);
> @@ -524,7 +536,9 @@ static int unuse_pte(struct vm_area_stru
>  	 * immediately swapped out again after swapon.
>  	 */
>  	activate_page(page);
> -	return 1;
> +out:
> +	pte_unmap_unlock(pte, ptl);
> +	return ret;
>  }
> 
>  static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
> @@ -533,21 +547,33 @@ static int unuse_pte_range(struct vm_are
>  {
>  	pte_t swp_pte = swp_entry_to_pte(entry);
>  	pte_t *pte;
> -	spinlock_t *ptl;
>  	int ret = 0;
> 
> -	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
> +	/*
> +	 * We don't actually need pte lock while scanning for swp_pte:
> +	 * since we hold page lock, swp_pte cannot be inserted into or
> +	 * removed from a page table while we're scanning; but on some
> +	 * architectures (e.g. i386 with PAE) we might catch a glimpse
> +	 * of unmatched parts which look like swp_pte, so unuse_pte
> +	 * must recheck under pte lock.  Scanning without the lock
> +	 * is preemptible if CONFIG_PREEMPT without CONFIG_HIGHPTE.
> +	 */
> +	pte = pte_offset_map(pmd, addr);
>  	do {
>  		/*
>  		 * swapoff spends a _lot_ of time in this loop!
>  		 * Test inline before going to call unuse_pte.
>  		 */
>  		if (unlikely(pte_same(*pte, swp_pte))) {
> -			ret = unuse_pte(vma, pte++, addr, entry, page);
> -			break;
> +			pte_unmap(pte);
> +			ret = unuse_pte(vma, pmd, addr, entry, page);
> +			if (ret)
> +				goto out;
> +			pte = pte_offset_map(pmd, addr);
>  		}
>  	} while (pte++, addr += PAGE_SIZE, addr != end);
> -	pte_unmap_unlock(pte - 1, ptl);
> +	pte_unmap(pte - 1);
> +out:
>  	return ret;
>  }
> 


-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Balbir Singh]

Hugh Dickins wrote:
> 
> --- 2.6.23-rc8-mm2/mm/swapfile.c	2007-09-27 12:03:36.000000000 +0100
> +++ linux/mm/swapfile.c	2007-10-07 14:33:05.000000000 +0100
> @@ -507,11 +507,23 @@ unsigned int count_swap_pages(int type, 
>   * just let do_wp_page work it out if a write is requested later - to
>   * force COW, vm_page_prot omits write permission from any private vma.
>   */
> -static int unuse_pte(struct vm_area_struct *vma, pte_t *pte,
> +static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
>  		unsigned long addr, swp_entry_t entry, struct page *page)
>  {
> +	spinlock_t *ptl;
> +	pte_t *pte;
> +	int ret = 1;
> +
>  	if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
> -		return -ENOMEM;
> +		ret = -ENOMEM;
> +
> +	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
> +	if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
> +		if (ret > 0)
> +			mem_cgroup_uncharge_page(page);
> +		ret = 0;
> +		goto out;
> +	}
> 
>  	inc_mm_counter(vma->vm_mm, anon_rss);
>  	get_page(page);
> @@ -524,7 +536,9 @@ static int unuse_pte(struct vm_area_stru
>  	 * immediately swapped out again after swapon.
>  	 */
>  	activate_page(page);
> -	return 1;
> +out:
> +	pte_unmap_unlock(pte, ptl);
> +	return ret;
>  }
> 
>  static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
> @@ -533,21 +547,33 @@ static int unuse_pte_range(struct vm_are
>  {
>  	pte_t swp_pte = swp_entry_to_pte(entry);
>  	pte_t *pte;
> -	spinlock_t *ptl;
>  	int ret = 0;
> 
> -	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
> +	/*
> +	 * We don't actually need pte lock while scanning for swp_pte:
> +	 * since we hold page lock, swp_pte cannot be inserted into or
> +	 * removed from a page table while we're scanning; but on some
> +	 * architectures (e.g. i386 with PAE) we might catch a glimpse
> +	 * of unmatched parts which look like swp_pte, so unuse_pte
> +	 * must recheck under pte lock.  Scanning without the lock
> +	 * is preemptible if CONFIG_PREEMPT without CONFIG_HIGHPTE.
> +	 */
> +	pte = pte_offset_map(pmd, addr);
>  	do {
>  		/*
>  		 * swapoff spends a _lot_ of time in this loop!
>  		 * Test inline before going to call unuse_pte.
>  		 */
>  		if (unlikely(pte_same(*pte, swp_pte))) {
> -			ret = unuse_pte(vma, pte++, addr, entry, page);
> -			break;
> +			pte_unmap(pte);
> +			ret = unuse_pte(vma, pmd, addr, entry, page);
> +			if (ret)
> +				goto out;
> +			pte = pte_offset_map(pmd, addr);
>  		}
>  	} while (pte++, addr += PAGE_SIZE, addr != end);
> -	pte_unmap_unlock(pte - 1, ptl);
> +	pte_unmap(pte - 1);
> +out:
>  	return ret;
>  }
> 

I tested this patch and it seems to be working fine. I tried swapoff -a
in the middle of tests consuming swap. Not 100% rigorous, but a good
test nevertheless.

Tested-by: Balbir Singh <balbir@linux.vnet.ibm.com>

-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Hugh Dickins]

On Mon, 15 Oct 2007, Balbir Singh wrote:
> Hugh Dickins wrote:
> > 
> > --- 2.6.23-rc8-mm2/mm/swapfile.c	2007-09-27 12:03:36.000000000 +0100
> > +++ linux/mm/swapfile.c	2007-10-07 14:33:05.000000000 +0100
> > @@ -507,11 +507,23 @@ unsigned int count_swap_pages(int type, 
> >   * just let do_wp_page work it out if a write is requested later - to
> >   * force COW, vm_page_prot omits write permission from any private vma.
> >   */
> > -static int unuse_pte(struct vm_area_struct *vma, pte_t *pte,
> > +static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
> >  		unsigned long addr, swp_entry_t entry, struct page *page)
...
> 
> I tested this patch and it seems to be working fine. I tried swapoff -a
> in the middle of tests consuming swap. Not 100% rigorous, but a good
> test nevertheless.
> 
> Tested-by: Balbir Singh <balbir@linux.vnet.ibm.com>

Thanks, Balbir.  Sorry for the delay.  I've not forgotten our
agreement that I should be splitting it into before-and-after
mem cgroup patches.  But it's low priority for me until we're
genuinely assigning to a cgroup there.  Hope to get back to
looking into that tomorrow, but no promises.

I think you still see no problem, where I claim that simply
omitting the mem charge mods from mm/swap_state.c leads to OOMs?
Maybe our difference is because my memhog in the cgroup is using
more memory than RAM, not just more memory than allowed to the
cgroup.  I suspect that arrives at a state (when the swapcache
pages are not charged) where it cannot locate the pages it needs
to reclaim to stay within its limit.

Hugh

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Balbir Singh]

Hugh Dickins wrote:
> On Mon, 15 Oct 2007, Balbir Singh wrote:
>> Hugh Dickins wrote:
>>> --- 2.6.23-rc8-mm2/mm/swapfile.c	2007-09-27 12:03:36.000000000 +0100
>>> +++ linux/mm/swapfile.c	2007-10-07 14:33:05.000000000 +0100
>>> @@ -507,11 +507,23 @@ unsigned int count_swap_pages(int type, 
>>>   * just let do_wp_page work it out if a write is requested later - to
>>>   * force COW, vm_page_prot omits write permission from any private vma.
>>>   */
>>> -static int unuse_pte(struct vm_area_struct *vma, pte_t *pte,
>>> +static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
>>>  		unsigned long addr, swp_entry_t entry, struct page *page)
> ...
>> I tested this patch and it seems to be working fine. I tried swapoff -a
>> in the middle of tests consuming swap. Not 100% rigorous, but a good
>> test nevertheless.
>>
>> Tested-by: Balbir Singh <balbir@linux.vnet.ibm.com>
> 
> Thanks, Balbir.  Sorry for the delay.  I've not forgotten our
> agreement that I should be splitting it into before-and-after
> mem cgroup patches.  But it's low priority for me until we're
> genuinely assigning to a cgroup there.  Hope to get back to
> looking into that tomorrow, but no promises.
> 

No Problem. We have some time with this one.

> I think you still see no problem, where I claim that simply
> omitting the mem charge mods from mm/swap_state.c leads to OOMs?
> Maybe our difference is because my memhog in the cgroup is using
> more memory than RAM, not just more memory than allowed to the
> cgroup.  I suspect that arrives at a state (when the swapcache
> pages are not charged) where it cannot locate the pages it needs
> to reclaim to stay within its limit.
> 

Yes, in my case there I use memory less than RAM and more than that
is allowed by the cgroup. It's quite possible that in your case the
swapcache has grown significantly without any limit/control on it.
The memhog program is using memory at a rate much higher than the
rate of reclaim. Could you share your memhog program, please?
In the use case you've mentioned/tested, having these mods to
control swapcache is actually useful, right?

Could you share your major objections at this point with the memory
controller at this point. I hope to be able to look into/resolve them
as my first priority in my list of items to work on.


> Hugh


-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Hugh Dickins]

On Wed, 24 Oct 2007, Balbir Singh wrote:
> Hugh Dickins wrote:
> > 
> > Thanks, Balbir.  Sorry for the delay.  I've not forgotten our
> > agreement that I should be splitting it into before-and-after
> > mem cgroup patches.  But it's low priority for me until we're
> > genuinely assigning to a cgroup there.  Hope to get back to
> > looking into that tomorrow, but no promises.
> 
> No Problem. We have some time with this one.

Phew - I still haven't got there.

> > I think you still see no problem, where I claim that simply
> > omitting the mem charge mods from mm/swap_state.c leads to OOMs?
> > Maybe our difference is because my memhog in the cgroup is using
> > more memory than RAM, not just more memory than allowed to the
> > cgroup.  I suspect that arrives at a state (when the swapcache
> > pages are not charged) where it cannot locate the pages it needs
> > to reclaim to stay within its limit.
> 
> Yes, in my case there I use memory less than RAM and more than that
> is allowed by the cgroup. It's quite possible that in your case the
> swapcache has grown significantly without any limit/control on it.
> The memhog program is using memory at a rate much higher than the
> rate of reclaim. Could you share your memhog program, please?

Gosh, it's nothing special.  Appended below, but please don't shame
me by taking it too seriously.  Defaults to working on a 600M mmap
because I'm in the habit of booting mem=512M.  You probably have
something better yourself that you'd rather use.

> In the use case you've mentioned/tested, having these mods to
> control swapcache is actually useful, right?

No idea what you mean by "these mods to control swapcache"?

With your mem_cgroup mods in mm/swap_state.c, swapoff assigns
the pages read in from swap to whoever's running swapoff and your
unuse_pte mem_cgroup_charge never does anything useful: swap pages
should get assigned to the appropriate cgroups at that point.

Without your mem_cgroup mods in mm/swap_state.c, unuse_pte makes
the right assignments (I believe).  But I find that swapout (using
600M in a 512M machine) from a 200M cgroup quickly OOMs, whereas
it behaves correctly with your mm/swap_state.c.

Thought little yet about what happens to shmem swapped pages,
and swap readahead pages; but still suspect that they and the
above issue will need a "limbo" cgroup, for pages which are
expected to belong to a not-yet-identified mem cgroup.

> 
> Could you share your major objections at this point with the memory
> controller at this point. I hope to be able to look into/resolve them
> as my first priority in my list of items to work on.

The things I've noticed so far, as mentioned before and above.

But it does worry me that I only came here through finding swapoff
broken by that unuse_mm return value, and then found one issue
after another.  It feels like the mem cgroup people haven't really
thought through or tested swap at all, and that if I looked further
I'd uncover more.

That's simply FUD, and I apologize if I'm being unfair: but that
is how it feels, and I expect we all know that phase in a project
when solving one problem uncovers three - suggests it's not ready.

Hugh

/* swapout.c */
#include <unistd.h>
#include <stdlib.h>
#include <stdio.h>
#include <errno.h>
#include <sys/mman.h>

int main(int argc, char *argv[])
{
	unsigned long *base = (unsigned long *)0x08400000;
	unsigned long size;
	unsigned long limit;
	unsigned long i;
	char *ptr = NULL;

	size = argv[1]? strtoul(argv[1], &ptr, 0): 600;
	if (size >= 3*1024)
		size = 0;
	size *= 1024*1024;
	limit = size / sizeof(unsigned long);
	if (size == 0 || base + limit + 1024 > &size) {
		errno = EINVAL;
		perror("swapout");
		exit(1);
	}
	base = mmap(base, size, PROT_READ|PROT_WRITE,
		    MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
	if (base == (unsigned long *)(-1)) {
		perror("mmap");
		exit(1);
	}
	for (i = 0; i < limit; i++)
		base[i] = i;
	if (ptr && *ptr == '.') {
		printf("Type <Return> to continue ");
		fflush(stdout);
		getchar();
	}
	for (i = 0; i < limit; i++)
		base[i] = limit - i;
	return 0;
}

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Balbir Singh]

Hugh Dickins wrote:
> Gosh, it's nothing special.  Appended below, but please don't shame
> me by taking it too seriously.  Defaults to working on a 600M mmap
> because I'm in the habit of booting mem=512M.  You probably have
> something better yourself that you'd rather use.
> 

Thanks for sending it. I do have something more generic that I got
from my colleague.

>> In the use case you've mentioned/tested, having these mods to
>> control swapcache is actually useful, right?
> 
> No idea what you mean by "these mods to control swapcache"?
> 

Yes

> With your mem_cgroup mods in mm/swap_state.c, swapoff assigns
> the pages read in from swap to whoever's running swapoff and your
> unuse_pte mem_cgroup_charge never does anything useful: swap pages
> should get assigned to the appropriate cgroups at that point.
> 
> Without your mem_cgroup mods in mm/swap_state.c, unuse_pte makes
> the right assignments (I believe).  But I find that swapout (using
> 600M in a 512M machine) from a 200M cgroup quickly OOMs, whereas
> it behaves correctly with your mm/swap_state.c.
> 

I'll try this test and play with your test

> Thought little yet about what happens to shmem swapped pages,
> and swap readahead pages; but still suspect that they and the
> above issue will need a "limbo" cgroup, for pages which are
> expected to belong to a not-yet-identified mem cgroup.
> 

This is something I am yet to experiment with. I suspect this
should be easy to do if we decide to go this route.


>> Could you share your major objections at this point with the memory
>> controller at this point. I hope to be able to look into/resolve them
>> as my first priority in my list of items to work on.
> 
> The things I've noticed so far, as mentioned before and above.
> 
> But it does worry me that I only came here through finding swapoff
> broken by that unuse_mm return value, and then found one issue
> after another.  It feels like the mem cgroup people haven't really
> thought through or tested swap at all, and that if I looked further
> I'd uncover more.
> 

I thought so far that you've found a couple of bugs and one issue
with the way we account for swapcache. Other users, KAMEZAWA,
YAMAMOTO have been using and enhancing the memory controller.
I can point you to a set of links where I posted all the test
results. Swap was tested mostly through swapout/swapin when the
cgroup goes over limit. Please do help uncover as many bugs
as possible, please look more closely as you find more time.


> That's simply FUD, and I apologize if I'm being unfair: but that
> is how it feels, and I expect we all know that phase in a project
> when solving one problem uncovers three - suggests it's not ready.
> 

I disagree, all projects/code do have bugs, which we are trying to
resolve, but I don't think there are any major design drawbacks
that *cannot* be fixed. We discussed the design at VM-Summit and
everyone agreed it was the way to go forward (even though Double
LRU has its complexity).

> Hugh

[snip]

Thanks for the review and your valuable feedback!

-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Balbir Singh]

On Mon, Oct 29, 2007 at 01:57:40AM +0530, Balbir Singh wrote:
Hugh Dickins wrote:

[snip]
 
> Without your mem_cgroup mods in mm/swap_state.c, unuse_pte makes
> the right assignments (I believe).  But I find that swapout (using
> 600M in a 512M machine) from a 200M cgroup quickly OOMs, whereas
> it behaves correctly with your mm/swap_state.c.
> 

On my UML setup, I booted the UML instance with 512M of memory and
used the swapout program that you shared. I tried two things


1. Ran swapout without any changes. The program ran well without
   any OOM condition occuring, lot of reclaim occured.
2. Ran swapout with the changes to mm/swap_state.c removed (diff below)
   and I still did not see any OOM. The reclaim count was much lesser
   since swap cache did not get accounted back to the cgroup from
   which pages were being evicted.

I am not sure why I don't see the OOM that you see, still trying. May be
I missing something obvious at this late hour in the night :-)

Output of the tests
-------------------

balbir@ubuntu:/container/swapout$ cat memory.limit_in_bytes
209715200
balbir@ubuntu:/container/swapout$ cat memory.usage_in_bytes
65536
balbir@ubuntu:/container/swapout$ cat tasks
1815
1847
balbir@ubuntu:/container/swapout$ ps
  PID TTY          TIME CMD
 1815 pts/0    00:00:00 bash
 1848 pts/0    00:00:00 ps
balbir@ubuntu:/container/swapout$ ~/swapout
balbir@ubuntu:/container/swapout$ echo $?
0
balbir@ubuntu:/container/swapout$ cat memory.failcnt
18

Diff to remove mods from swap_state.c (for testing only)
--------------------------------------------------------

--- mm/swap_state.c.org	2007-10-29 01:42:14.000000000 +0530
+++ mm/swap_state.c	2007-10-29 01:52:48.000000000 +0530
@@ -79,10 +79,6 @@ static int __add_to_swap_cache(struct pa
 	BUG_ON(PageSwapCache(page));
 	BUG_ON(PagePrivate(page));
 
-	error = mem_cgroup_cache_charge(page, current->mm, gfp_mask);
-	if (error)
-		goto out;
-
 	error = radix_tree_preload(gfp_mask);
 	if (!error) {
 		write_lock_irq(&swapper_space.tree_lock);
@@ -94,14 +90,11 @@ static int __add_to_swap_cache(struct pa
 			set_page_private(page, entry.val);
 			total_swapcache_pages++;
 			__inc_zone_page_state(page, NR_FILE_PAGES);
-		} else
-			mem_cgroup_uncharge_page(page);
+		}
 
 		write_unlock_irq(&swapper_space.tree_lock);
 		radix_tree_preload_end();
-	} else
-		mem_cgroup_uncharge_page(page);
-out:
+	}
 	return error;
 }
 
@@ -141,7 +134,6 @@ void __delete_from_swap_cache(struct pag
 	BUG_ON(PageWriteback(page));
 	BUG_ON(PagePrivate(page));
 
-	mem_cgroup_uncharge_page(page);
 	radix_tree_delete(&swapper_space.page_tree, page_private(page));
 	set_page_private(page, 0);
 	ClearPageSwapCache(page);
-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Hugh Dickins]

On Mon, 29 Oct 2007, Balbir Singh wrote:
> On Mon, Oct 29, 2007 at 01:57:40AM +0530, Balbir Singh wrote:
> Hugh Dickins wrote:
> 
> [snip]
>  
> > Without your mem_cgroup mods in mm/swap_state.c, unuse_pte makes
> > the right assignments (I believe).  But I find that swapout (using
> > 600M in a 512M machine) from a 200M cgroup quickly OOMs, whereas
> > it behaves correctly with your mm/swap_state.c.
> > 
> 
> On my UML setup, I booted the UML instance with 512M of memory and
> used the swapout program that you shared. I tried two things
> 
> 
> 1. Ran swapout without any changes. The program ran well without
>    any OOM condition occuring, lot of reclaim occured.
> 2. Ran swapout with the changes to mm/swap_state.c removed (diff below)
>    and I still did not see any OOM. The reclaim count was much lesser
>    since swap cache did not get accounted back to the cgroup from
>    which pages were being evicted.
> 
> I am not sure why I don't see the OOM that you see, still trying. May be
> I missing something obvious at this late hour in the night :-)

I reconfirm that I do see those OOMs.  I'll have to try harder to
analyze how they come about: I sure don't expect you to debug a
problem you cannot reproduce.  But what happens if you try it
native rather than using UML?

Hugh

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Balbir Singh]

Hugh Dickins wrote:
> On Mon, 29 Oct 2007, Balbir Singh wrote:
>> On Mon, Oct 29, 2007 at 01:57:40AM +0530, Balbir Singh wrote:
>> Hugh Dickins wrote:
>>
>> [snip]
>>  
>>> Without your mem_cgroup mods in mm/swap_state.c, unuse_pte makes
>>> the right assignments (I believe).  But I find that swapout (using
>>> 600M in a 512M machine) from a 200M cgroup quickly OOMs, whereas
>>> it behaves correctly with your mm/swap_state.c.
>>>
>> On my UML setup, I booted the UML instance with 512M of memory and
>> used the swapout program that you shared. I tried two things
>>
>>
>> 1. Ran swapout without any changes. The program ran well without
>>    any OOM condition occuring, lot of reclaim occured.
>> 2. Ran swapout with the changes to mm/swap_state.c removed (diff below)
>>    and I still did not see any OOM. The reclaim count was much lesser
>>    since swap cache did not get accounted back to the cgroup from
>>    which pages were being evicted.
>>
>> I am not sure why I don't see the OOM that you see, still trying. May be
>> I missing something obvious at this late hour in the night :-)
> 
> I reconfirm that I do see those OOMs.  I'll have to try harder to
> analyze how they come about: I sure don't expect you to debug a
> problem you cannot reproduce.  But what happens if you try it
> native rather than using UML?
> 
> Hugh

On a real box - a powerpc machine that I have access to

1. I don't see the OOM with the mods removed (I have swap
   space at-least twice of RAM - with mem=512M, I have at-least
   1G of swap).
2. Running under the container is much much faster than running
   swapout in the root container. The machine is almost unusable
   if swapout is run under the root container

At this momemnt, I suspect one of two things

1. Our mods to swap_state.c are different
2. Our configuration is different, main-memory to swap-size ratio


-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Hugh Dickins]

On Tue, 30 Oct 2007, Balbir Singh wrote:
> 
> At this momemnt, I suspect one of two things
> 
> 1. Our mods to swap_state.c are different

I believe they're the same (just take swap_state.c back to how it
was without mem_cgroup mods) - or would be, if after finding this
effect I hadn't added a "swap_in_cg" switch to move between the
two behaviours to study it better (though I do need to remember
to swapoff and swapon between the two: sometimes I do forget).

> 2. Our configuration is different, main-memory to swap-size ratio

I doubt the swapsize is relevant: just so long as there's some (a
little more than 200M I guess); I've got 1GB-2GB on different boxes.

There may well be something about our configs that's significantly
different.  I'd failed to mention SMP (4 cpu), and that I happen
to have /proc/sys/vm/swappiness 100; but find it happens on UP
also, and when I go back to default swappiness 60.

I've reordered your mail for more dramatic effect...
> 
> On a real box - a powerpc machine that I have access to

I've tried on 3 Intel and 1 PowerPC now: the Intels show the OOMs
and the PowerPC does not.  I rather doubt it's an Intel versus
PowerPC issue as such, but interesting that we see the same.

> 
> 1. I don't see the OOM with the mods removed (I have swap
>    space at-least twice of RAM - with mem=512M, I have at-least
>    1G of swap).

mem=512M with 1G of swap, yes, I'm the same.

> 2. Running under the container is much much faster than running
>    swapout in the root container. The machine is almost unusable
>    if swapout is run under the root container

That's rather interesting, isn't it?  Probably irrelevant to the
OOM issue we're investigating, but worthy of investigation in itself.

Maybe I saw the same on the PowerPC: I simply forgot to set up the
cgroup one time, and my sequence of three swapouts (sometimes only
two out of three OOM, on those boxes that do OOM) seemed to take a
very long time (but I wasn't trying to do anything else on it at
the same time, so didn't notice if it was "unusable").

I'll probe on.

Hugh

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [RFC] [-mm PATCH] Memory controller fix swap charging context in unuse_pte()

[Balbir Singh]

Hugh Dickins wrote:
> On Tue, 30 Oct 2007, Balbir Singh wrote:
>> At this momemnt, I suspect one of two things
>>
>> 1. Our mods to swap_state.c are different
> 
> I believe they're the same (just take swap_state.c back to how it
> was without mem_cgroup mods) - or would be, if after finding this
> effect I hadn't added a "swap_in_cg" switch to move between the
> two behaviours to study it better (though I do need to remember
> to swapoff and swapon between the two: sometimes I do forget).
> 
>> 2. Our configuration is different, main-memory to swap-size ratio
> 
> I doubt the swapsize is relevant: just so long as there's some (a
> little more than 200M I guess); I've got 1GB-2GB on different boxes.
> 

I agree, just wanted to make sure that there is enough swap

> There may well be something about our configs that's significantly
> different.  I'd failed to mention SMP (4 cpu), and that I happen
> to have /proc/sys/vm/swappiness 100; but find it happens on UP
> also, and when I go back to default swappiness 60.
> 

OK.. so those are out of the equation

> I've reordered your mail for more dramatic effect...
>> On a real box - a powerpc machine that I have access to
> 
> I've tried on 3 Intel and 1 PowerPC now: the Intels show the OOMs
> and the PowerPC does not.  I rather doubt it's an Intel versus
> PowerPC issue as such, but interesting that we see the same.
> 

Very surprising, I am surprised that it's architecture dependent.
Let me try and grab an Intel box and try.

>> 1. I don't see the OOM with the mods removed (I have swap
>>    space at-least twice of RAM - with mem=512M, I have at-least
>>    1G of swap).
> 
> mem=512M with 1G of swap, yes, I'm the same.
> 
>> 2. Running under the container is much much faster than running
>>    swapout in the root container. The machine is almost unusable
>>    if swapout is run under the root container
> 
> That's rather interesting, isn't it?  Probably irrelevant to the
> OOM issue we're investigating, but worthy of investigation in itself.
> 

Yes, it irrelevant, but I find it to be a good use case for using the
memory controller :-) I found that kswapd running at prio -5, seemed
to hog quite a bit of the CPU. But it needs more independent
investigation, like you've suggested.

> Maybe I saw the same on the PowerPC: I simply forgot to set up the
> cgroup one time, and my sequence of three swapouts (sometimes only
> two out of three OOM, on those boxes that do OOM) seemed to take a
> very long time (but I wasn't trying to do anything else on it at
> the same time, so didn't notice if it was "unusable").
> 
> I'll probe on.
> 

Me too.. I'll try and acquire a good x86_64 box and test on it.

> Hugh
> 
> --
> 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/ .
> Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>


-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL

--
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/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>