ZERO_PAGE refcounting causes cache line bouncing

ZERO_PAGE refcounting causes cache line bouncing

[Christoph Lameter]

We have issues with ZERO_PAGE refcounting causing severe cacheline 
bouncing. ZERO_PAGES are mapped into multiple processes running on 
multiple nodes. Refcounter modifications therefore have to acquire a 
remote exclusive cacheline.

Could we somehow fix this? There are a couple of ways to do this:

1. No refcounting on reserved pages in the VM. ZERO_PAGEs are
   reserved and there is no point in refcounting them since they
   will not go away.

2. Having a percpu or pernode ZERO_PAGE?

   May be a simpler solution but then we still may have issues
   if the ZERO_PAGE gets "freed" from other processors/ nodes.

--
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: ZERO_PAGE refcounting causes cache line bouncing

[William Lee Irwin III]

On Fri, Mar 16, 2007 at 03:17:39PM -0700, Christoph Lameter wrote:
> We have issues with ZERO_PAGE refcounting causing severe cacheline 
> bouncing. ZERO_PAGES are mapped into multiple processes running on 
> multiple nodes. Refcounter modifications therefore have to acquire a 
> remote exclusive cacheline.
> Could we somehow fix this? There are a couple of ways to do this:
> 1. No refcounting on reserved pages in the VM. ZERO_PAGEs are
>    reserved and there is no point in refcounting them since they
>    will not go away.
> 2. Having a percpu or pernode ZERO_PAGE?
>    May be a simpler solution but then we still may have issues
>    if the ZERO_PAGE gets "freed" from other processors/ nodes.

It's dumb to refcount the zero page. Someone should've noticed this
when the PG_reserved patches went in. I can't think of an easy way
around this apart from a backout. OTOH it's a simple matter of
programming to arrange for it without a backout.

Provisions should be made for per-node zero pages in addition to this.
AFAICT the primary thing needed is to wrap checks for a page being a
zero page with some testing function instead of using a raw equality
check. This is above and beyond solving the mere zero page refcount
problem; I'm saying that both proposals should be done even though only
one is needed to resolve the bouncing issue.

I guess this is an "ack of the concept" of sorts. None of this is so
involved that I should jump on it and try to get a patch out ahead of
you.


-- wli

--
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: ZERO_PAGE refcounting causes cache line bouncing

[Nick Piggin]

William Lee Irwin III wrote:
> On Fri, Mar 16, 2007 at 03:17:39PM -0700, Christoph Lameter wrote:
> 
>>We have issues with ZERO_PAGE refcounting causing severe cacheline 
>>bouncing. ZERO_PAGES are mapped into multiple processes running on 
>>multiple nodes. Refcounter modifications therefore have to acquire a 
>>remote exclusive cacheline.
>>Could we somehow fix this? There are a couple of ways to do this:
>>1. No refcounting on reserved pages in the VM. ZERO_PAGEs are
>>   reserved and there is no point in refcounting them since they
>>   will not go away.
>>2. Having a percpu or pernode ZERO_PAGE?
>>   May be a simpler solution but then we still may have issues
>>   if the ZERO_PAGE gets "freed" from other processors/ nodes.
> 
> 
> It's dumb to refcount the zero page. Someone should've noticed this
> when the PG_reserved patches went in.

The patch author did notice ;)

http://groups.google.com/group/fa.linux.kernel/msg/48108a5faa20a667?hl=en&

But was subsequently told that special casing the ZERO_PAGE was stupid
(as if it isn't a special object in the VM), and so it was removed (and
a warning added to the changelog)

http://lwn.net/Articles/155280/

I also subsequently submitted a standalone patch to reintroduce it after
the core patch was merged, but again was rejected (can't find the link for
that one, off hand).


> I can't think of an easy way
> around this apart from a backout. OTOH it's a simple matter of
> programming to arrange for it without a backout.

Yes, I have the patch to do it quite easily. Per-node ZERO_PAGE could be
another option, but that's going to cost another page flag if we wish to
recognise the zero page in wp faults like we do now (hmm, for some reason
it is OK to special case it _there_).


> Provisions should be made for per-node zero pages in addition to this.
> AFAICT the primary thing needed is to wrap checks for a page being a
> zero page with some testing function instead of using a raw equality
> check. This is above and beyond solving the mere zero page refcount
> problem; I'm saying that both proposals should be done even though only
> one is needed to resolve the bouncing issue.

I've always thought the bouncing issue was a silly one and should be
fixed, of course. Maybe the reason my fix was vetoed was lack of numbers.
Christoph, would you oblige? I'll dig out the patch and repost.

-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com 

--
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: ZERO_PAGE refcounting causes cache line bouncing

[Nick Piggin]

[-- Attachment #1: Type: text/plain, Size: 550 bytes --]

Nick Piggin wrote:

> I've always thought the bouncing issue was a silly one and should be
> fixed, of course. Maybe the reason my fix was vetoed was lack of numbers.
> Christoph, would you oblige? I'll dig out the patch and repost.

Something like this roughly should get rid of ZERO_PAGE _count and _mapcount
manipulation for anonymous pages. (others still exist, XIP and /dev/zero, but
they should not be a large concern AFAIKS).

I haven't booted this, but it is a quick forward port + some fixes and
simplifications.

-- 
SUSE Labs, Novell Inc.

[-- Attachment #2: mm-special-case-ZERO_PAGE.patch --]
[-- Type: text/plain, Size: 1012 bytes --]

Index: linux-2.6/mm/memory.c
===================================================================
--- linux-2.6.orig/mm/memory.c
+++ linux-2.6/mm/memory.c
@@ -665,7 +665,8 @@ static unsigned long zap_pte_range(struc
 			ptent = ptep_get_and_clear_full(mm, addr, pte,
 							tlb->fullmm);
 			tlb_remove_tlb_entry(tlb, pte, addr);
-			if (unlikely(!page))
+			if (unlikely(!page ||
+				(!vma->vm_file && page == ZERO_PAGE(addr))))
 				continue;
 			if (unlikely(details) && details->nonlinear_vma
 			    && linear_page_index(details->nonlinear_vma,
@@ -2152,15 +2153,12 @@ static int do_anonymous_page(struct mm_s
 	} else {
 		/* Map the ZERO_PAGE - vm_page_prot is readonly */
 		page = ZERO_PAGE(address);
-		page_cache_get(page);
 		entry = mk_pte(page, vma->vm_page_prot);
 
 		ptl = pte_lockptr(mm, pmd);
 		spin_lock(ptl);
 		if (!pte_none(*page_table))
-			goto release;
-		inc_mm_counter(mm, file_rss);
-		page_add_file_rmap(page);
+			goto unlock;
 	}
 
 	set_pte_at(mm, address, page_table, entry);

Re: ZERO_PAGE refcounting causes cache line bouncing

[Nick Piggin]

Nick Piggin wrote:

> Something like this roughly should get rid of ZERO_PAGE _count and 
> _mapcount
> manipulation for anonymous pages. (others still exist, XIP and 
> /dev/zero, but
> they should not be a large concern AFAIKS).
> 
> I haven't booted this, but it is a quick forward port + some fixes and
> simplifications.
> 
> 
> ------------------------------------------------------------------------
> 
> Index: linux-2.6/mm/memory.c
> ===================================================================
> --- linux-2.6.orig/mm/memory.c
> +++ linux-2.6/mm/memory.c
> @@ -665,7 +665,8 @@ static unsigned long zap_pte_range(struc
>  			ptent = ptep_get_and_clear_full(mm, addr, pte,
>  							tlb->fullmm);
>  			tlb_remove_tlb_entry(tlb, pte, addr);
> -			if (unlikely(!page))
> +			if (unlikely(!page ||
> +				(!vma->vm_file && page == ZERO_PAGE(addr))))
>  				continue;

Hmm, well I suppose it would be cleaner if this check used the one in
handle_pte_fault instead of !vma->vm_file ie. (!vma->vm_ops ||
!vma->vm_ops->nopage)


>  			if (unlikely(details) && details->nonlinear_vma
>  			    && linear_page_index(details->nonlinear_vma,
> @@ -2152,15 +2153,12 @@ static int do_anonymous_page(struct mm_s
>  	} else {
>  		/* Map the ZERO_PAGE - vm_page_prot is readonly */
>  		page = ZERO_PAGE(address);
> -		page_cache_get(page);
>  		entry = mk_pte(page, vma->vm_page_prot);
>  
>  		ptl = pte_lockptr(mm, pmd);
>  		spin_lock(ptl);
>  		if (!pte_none(*page_table))
> -			goto release;
> -		inc_mm_counter(mm, file_rss);
> -		page_add_file_rmap(page);
> +			goto unlock;
>  	}
>  
>  	set_pte_at(mm, address, page_table, entry);


-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com 

--
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: ZERO_PAGE refcounting causes cache line bouncing

[Nick Piggin]

Nick Piggin wrote:
> Nick Piggin wrote:
> 
>> Something like this roughly should get rid of ZERO_PAGE _count and 
>> _mapcount
>> manipulation for anonymous pages. (others still exist, XIP and 
>> /dev/zero, but
>> they should not be a large concern AFAIKS).
>>
>> I haven't booted this, but it is a quick forward port + some fixes and
>> simplifications.
>>
>>
>> ------------------------------------------------------------------------
>>
>> Index: linux-2.6/mm/memory.c
>> ===================================================================
>> --- linux-2.6.orig/mm/memory.c
>> +++ linux-2.6/mm/memory.c
>> @@ -665,7 +665,8 @@ static unsigned long zap_pte_range(struc
>>              ptent = ptep_get_and_clear_full(mm, addr, pte,
>>                              tlb->fullmm);
>>              tlb_remove_tlb_entry(tlb, pte, addr);
>> -            if (unlikely(!page))
>> +            if (unlikely(!page ||
>> +                (!vma->vm_file && page == ZERO_PAGE(addr))))
>>                  continue;
> 
> 
> Hmm, well I suppose it would be cleaner if this check used the one in
> handle_pte_fault instead of !vma->vm_file ie. (!vma->vm_ops ||
> !vma->vm_ops->nopage)

Bah, I also missed a reject for a similar hunk required in copy_one_pte.

I don't think there is anything more required after that, though... I
will actually test it tomorrow and send an updated patch with proper
changelog.

-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com 

--
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: ZERO_PAGE refcounting causes cache line bouncing

[Robin Holt]

On Mon, Mar 19, 2007 at 04:56:47PM +1100, Nick Piggin wrote:
> Yes, I have the patch to do it quite easily. Per-node ZERO_PAGE could be
> another option, but that's going to cost another page flag if we wish to
> recognise the zero page in wp faults like we do now (hmm, for some reason
> it is OK to special case it _there_).

Could we do a per-node ZERO_PAGE as a pointer from the node structure
and then use a page_to_nid to get back to the node and compare the page
to the node's zero page instead of using another page flag which would
actually only be used on numa?

Thanks,
Robin

--
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: ZERO_PAGE refcounting causes cache line bouncing

[William Lee Irwin III]

On Mon, Mar 19, 2007 at 04:56:47PM +1100, Nick Piggin wrote:
> Yes, I have the patch to do it quite easily. Per-node ZERO_PAGE could be
> another option, but that's going to cost another page flag if we wish to
> recognise the zero page in wp faults like we do now (hmm, for some reason
> it is OK to special case it _there_).

No need for a page flag. A per-node array of struct page * can be used
to check by merely indexing into it with the nid of the page's node. e.g.

struct page *get_zero_page(int nid, unsigned long addr)
{
	return zero_pages[nid][(addr & SOME_ARCHDEP_MASK) >> PAGE_SHIFT];
}

/* any time we fish one out of a pte we have a uvaddr */
int is_zero_page_addr(struct page *page, unsigned long address)
{
	return page == get_zero_page(page_to_nid(page), address);
}


-- wli

--
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: ZERO_PAGE refcounting causes cache line bouncing

[Christoph Lameter]

On Mon, 19 Mar 2007, Nick Piggin wrote:

> I've always thought the bouncing issue was a silly one and should be
> fixed, of course. Maybe the reason my fix was vetoed was lack of numbers.
> Christoph, would you oblige? I'll dig out the patch and repost.

Well this occurs on a 1024p system that is only sporadically available. It 
gets so bad it hangs completely. Could you also provide patch against 
SLES10? We get get some bug action on this one I think.

--
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: ZERO_PAGE refcounting causes cache line bouncing

[Christoph Lameter]

On Mon, 19 Mar 2007, Nick Piggin wrote:

> I haven't booted this, but it is a quick forward port + some fixes and
> simplifications.

Eeek patch vanished.

The comparison with ZERO_PAGE may fail if we have multiple zero pages. 
Would it be possible to check for PageReserved?

--
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: ZERO_PAGE refcounting causes cache line bouncing

[Nick Piggin]

William Lee Irwin III wrote:
> On Mon, Mar 19, 2007 at 04:56:47PM +1100, Nick Piggin wrote:
> 
>>Yes, I have the patch to do it quite easily. Per-node ZERO_PAGE could be
>>another option, but that's going to cost another page flag if we wish to
>>recognise the zero page in wp faults like we do now (hmm, for some reason
>>it is OK to special case it _there_).
> 
> 
> No need for a page flag. A per-node array of struct page * can be used
> to check by merely indexing into it with the nid of the page's node. e.g.
> 
> struct page *get_zero_page(int nid, unsigned long addr)
> {
> 	return zero_pages[nid][(addr & SOME_ARCHDEP_MASK) >> PAGE_SHIFT];
> }
> 
> /* any time we fish one out of a pte we have a uvaddr */
> int is_zero_page_addr(struct page *page, unsigned long address)
> {
> 	return page == get_zero_page(page_to_nid(page), address);
> }

Ah, good point :)

-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com 

--
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: ZERO_PAGE refcounting causes cache line bouncing

[Nick Piggin]

Robin Holt wrote:
> On Mon, Mar 19, 2007 at 04:56:47PM +1100, Nick Piggin wrote:
> 
>>Yes, I have the patch to do it quite easily. Per-node ZERO_PAGE could be
>>another option, but that's going to cost another page flag if we wish to
>>recognise the zero page in wp faults like we do now (hmm, for some reason
>>it is OK to special case it _there_).
> 
> 
> Could we do a per-node ZERO_PAGE as a pointer from the node structure
> and then use a page_to_nid to get back to the node and compare the page
> to the node's zero page instead of using another page flag which would
> actually only be used on numa?

Yes, that's a nice way to do it.

-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com 

--
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: ZERO_PAGE refcounting causes cache line bouncing

[Nick Piggin]

Christoph Lameter wrote:
> On Mon, 19 Mar 2007, Nick Piggin wrote:
> 
> 
>>I haven't booted this, but it is a quick forward port + some fixes and
>>simplifications.
> 
> 
> Eeek patch vanished.
> 
> The comparison with ZERO_PAGE may fail if we have multiple zero pages. 
> Would it be possible to check for PageReserved?

That still wasn't quite right either.

I don't want to check for PageReserved, because I want to get rid of
that flag one day. The Robin/Bill approach for multiple zero pages
will work, though.

-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com 

--
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>