[PATCH] mm, x86, pat: Improve scaling of pat_pagerange_is_ram()

[PATCH] mm, x86, pat: Improve scaling of pat_pagerange_is_ram()

[John Dykstra]

Function pat_pagerange_is_ram() scales poorly to large address ranges,
because it probes the resource tree for each page.  On a 2.6 GHz
Opteron, this function consumes 34 ms. for a 1 GB range.  It is called
twice during untrack_pfn_vma(), slowing process cleanup and handicapping
the OOM killer.

This replacement based on walk_system_ram_range() consumes less than 1
ms. under the same conditions.

Signed-off-by: John Dykstra <jdykstra@cray.com> on behalf of Cray Inc. 
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
---
 arch/x86/mm/pat.c      |   55 ++++++++++++++++++++++++++++++-----------------
 include/linux/ioport.h |    2 +
 kernel/resource.c      |    2 +-
 3 files changed, 38 insertions(+), 21 deletions(-)

diff --git a/arch/x86/mm/pat.c b/arch/x86/mm/pat.c
index f6ff57b..c119afb 100644
--- a/arch/x86/mm/pat.c
+++ b/arch/x86/mm/pat.c
@@ -160,29 +160,44 @@ static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
 
 static int pat_pagerange_is_ram(resource_size_t start, resource_size_t end)
 {
-	int ram_page = 0, not_rampage = 0;
-	unsigned long page_nr;
+	struct resource res;
+	resource_size_t pg_end, after_ram;
+	int ram = 0, not_ram = 0;
 
-	for (page_nr = (start >> PAGE_SHIFT); page_nr < (end >> PAGE_SHIFT);
-	     ++page_nr) {
-		/*
-		 * For legacy reasons, physical address range in the legacy ISA
-		 * region is tracked as non-RAM. This will allow users of
-		 * /dev/mem to map portions of legacy ISA region, even when
-		 * some of those portions are listed(or not even listed) with
-		 * different e820 types(RAM/reserved/..)
-		 */
-		if (page_nr >= (ISA_END_ADDRESS >> PAGE_SHIFT) &&
-		    page_is_ram(page_nr))
-			ram_page = 1;
-		else
-			not_rampage = 1;
+	res.start = start & PHYSICAL_PAGE_MASK;
 
-		if (ram_page == not_rampage)
+	/*
+	 * For legacy reasons, physical address range in the legacy ISA
+	 * region is tracked as non-RAM. This will allow users of
+	 * /dev/mem to map portions of legacy ISA region, even when
+	 * some of those portions are listed(or not even listed) with
+	 * different e820 types(RAM/reserved/..)
+	 */
+	if (res.start < ISA_END_ADDRESS) {
+		not_ram = 1;
+		res.start = ISA_END_ADDRESS;
+	}
+
+	pg_end = (end + PAGE_SIZE - 1) & PHYSICAL_PAGE_MASK;
+	res.end = pg_end;
+	res.flags = IORESOURCE_MEM | IORESOURCE_BUSY;
+	after_ram = res.start;
+	while ((res.start < res.end) &&
+		(find_next_system_ram(&res, "System RAM") >= 0)) {
+		if (res.start > after_ram)
+			not_ram = 1;
+		if (res.end > res.start)
+			ram = 1;
+
+		if (ram && not_ram)
 			return -1;
+
+		after_ram = res.end + 1;
+		res.start = res.end + 1;
+		res.end = pg_end;
 	}
 
-	return ram_page;
+	return ram;
 }
 
 /*
diff --git a/include/linux/ioport.h b/include/linux/ioport.h
index e885ba2..273a725 100644
--- a/include/linux/ioport.h
+++ b/include/linux/ioport.h
@@ -223,5 +223,7 @@ extern int
 walk_system_ram_range(unsigned long start_pfn, unsigned long nr_pages,
 		void *arg, int (*func)(unsigned long, unsigned long, void *));
 
+extern int find_next_system_ram(struct resource *res, char *name);
+
 #endif /* __ASSEMBLY__ */
 #endif	/* _LINUX_IOPORT_H */
diff --git a/kernel/resource.c b/kernel/resource.c
index 7e8ea66..dd8f553 100644
--- a/kernel/resource.c
+++ b/kernel/resource.c
@@ -283,7 +283,7 @@ EXPORT_SYMBOL(release_resource);
  * the caller must specify res->start, res->end, res->flags and "name".
  * If found, returns 0, res is overwritten, if not found, returns -1.
  */
-static int find_next_system_ram(struct resource *res, char *name)
+int find_next_system_ram(struct resource *res, char *name)
 {
 	resource_size_t start, end;
 	struct resource *p;
-- 
1.7.0.4

[PATCH] mm, x86, pat: Improve scaling of pat_pagerange_is_ram()

[John Dykstra]

Function pat_pagerange_is_ram() scales poorly to large address ranges,
because it probes the resource tree for each page.  On a 2.6 GHz
Opteron, this function consumes 34 ms. for a 1 GB range.  It is called
twice during untrack_pfn_vma(), slowing process cleanup and handicapping
the OOM killer.

This replacement based on walk_system_ram_range() consumes less than 1
ms. under the same conditions.

Signed-off-by: John Dykstra <jdykstra@cray.com> on behalf of Cray Inc. 
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
---
 arch/x86/mm/pat.c      |   55 ++++++++++++++++++++++++++++++-----------------
 include/linux/ioport.h |    2 +
 kernel/resource.c      |    2 +-
 3 files changed, 38 insertions(+), 21 deletions(-)

diff --git a/arch/x86/mm/pat.c b/arch/x86/mm/pat.c
index f6ff57b..c119afb 100644
--- a/arch/x86/mm/pat.c
+++ b/arch/x86/mm/pat.c
@@ -160,29 +160,44 @@ static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
 
 static int pat_pagerange_is_ram(resource_size_t start, resource_size_t end)
 {
-	int ram_page = 0, not_rampage = 0;
-	unsigned long page_nr;
+	struct resource res;
+	resource_size_t pg_end, after_ram;
+	int ram = 0, not_ram = 0;
 
-	for (page_nr = (start >> PAGE_SHIFT); page_nr < (end >> PAGE_SHIFT);
-	     ++page_nr) {
-		/*
-		 * For legacy reasons, physical address range in the legacy ISA
-		 * region is tracked as non-RAM. This will allow users of
-		 * /dev/mem to map portions of legacy ISA region, even when
-		 * some of those portions are listed(or not even listed) with
-		 * different e820 types(RAM/reserved/..)
-		 */
-		if (page_nr >= (ISA_END_ADDRESS >> PAGE_SHIFT) &&
-		    page_is_ram(page_nr))
-			ram_page = 1;
-		else
-			not_rampage = 1;
+	res.start = start & PHYSICAL_PAGE_MASK;
 
-		if (ram_page == not_rampage)
+	/*
+	 * For legacy reasons, physical address range in the legacy ISA
+	 * region is tracked as non-RAM. This will allow users of
+	 * /dev/mem to map portions of legacy ISA region, even when
+	 * some of those portions are listed(or not even listed) with
+	 * different e820 types(RAM/reserved/..)
+	 */
+	if (res.start < ISA_END_ADDRESS) {
+		not_ram = 1;
+		res.start = ISA_END_ADDRESS;
+	}
+
+	pg_end = (end + PAGE_SIZE - 1) & PHYSICAL_PAGE_MASK;
+	res.end = pg_end;
+	res.flags = IORESOURCE_MEM | IORESOURCE_BUSY;
+	after_ram = res.start;
+	while ((res.start < res.end) &&
+		(find_next_system_ram(&res, "System RAM") >= 0)) {
+		if (res.start > after_ram)
+			not_ram = 1;
+		if (res.end > res.start)
+			ram = 1;
+
+		if (ram && not_ram)
 			return -1;
+
+		after_ram = res.end + 1;
+		res.start = res.end + 1;
+		res.end = pg_end;
 	}
 
-	return ram_page;
+	return ram;
 }
 
 /*
diff --git a/include/linux/ioport.h b/include/linux/ioport.h
index e885ba2..273a725 100644
--- a/include/linux/ioport.h
+++ b/include/linux/ioport.h
@@ -223,5 +223,7 @@ extern int
 walk_system_ram_range(unsigned long start_pfn, unsigned long nr_pages,
 		void *arg, int (*func)(unsigned long, unsigned long, void *));
 
+extern int find_next_system_ram(struct resource *res, char *name);
+
 #endif /* __ASSEMBLY__ */
 #endif	/* _LINUX_IOPORT_H */
diff --git a/kernel/resource.c b/kernel/resource.c
index 7e8ea66..dd8f553 100644
--- a/kernel/resource.c
+++ b/kernel/resource.c
@@ -283,7 +283,7 @@ EXPORT_SYMBOL(release_resource);
  * the caller must specify res->start, res->end, res->flags and "name".
  * If found, returns 0, res is overwritten, if not found, returns -1.
  */
-static int find_next_system_ram(struct resource *res, char *name)
+int find_next_system_ram(struct resource *res, char *name)
 {
 	resource_size_t start, end;
 	struct resource *p;
-- 
1.7.0.4

Re: [PATCH] mm, x86, pat: Improve scaling of pat_pagerange_is_ram()

[Suresh Siddha]

On Mon, 2012-05-14 at 15:26 -0500, John Dykstra wrote:
> Function pat_pagerange_is_ram() scales poorly to large address ranges,
> because it probes the resource tree for each page.  On a 2.6 GHz
> Opteron, this function consumes 34 ms. for a 1 GB range.  It is called
> twice during untrack_pfn_vma(), slowing process cleanup and handicapping
> the OOM killer.
> 
> This replacement based on walk_system_ram_range() consumes less than 1
> ms. under the same conditions.
> 
> Signed-off-by: John Dykstra <jdykstra@cray.com> on behalf of Cray Inc. 
> Cc: Suresh Siddha <suresh.b.siddha@intel.com>
> ---
>  arch/x86/mm/pat.c      |   55 ++++++++++++++++++++++++++++++-----------------
>  include/linux/ioport.h |    2 +
>  kernel/resource.c      |    2 +-
>  3 files changed, 38 insertions(+), 21 deletions(-)
> 
> diff --git a/arch/x86/mm/pat.c b/arch/x86/mm/pat.c
> index f6ff57b..c119afb 100644
> --- a/arch/x86/mm/pat.c
> +++ b/arch/x86/mm/pat.c
> @@ -160,29 +160,44 @@ static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
>  
>  static int pat_pagerange_is_ram(resource_size_t start, resource_size_t end)
>  {
> -	int ram_page = 0, not_rampage = 0;
> -	unsigned long page_nr;
> +	struct resource res;
> +	resource_size_t pg_end, after_ram;
> +	int ram = 0, not_ram = 0;
>  
> -	for (page_nr = (start >> PAGE_SHIFT); page_nr < (end >> PAGE_SHIFT);
> -	     ++page_nr) {
> -		/*
> -		 * For legacy reasons, physical address range in the legacy ISA
> -		 * region is tracked as non-RAM. This will allow users of
> -		 * /dev/mem to map portions of legacy ISA region, even when
> -		 * some of those portions are listed(or not even listed) with
> -		 * different e820 types(RAM/reserved/..)
> -		 */
> -		if (page_nr >= (ISA_END_ADDRESS >> PAGE_SHIFT) &&
> -		    page_is_ram(page_nr))
> -			ram_page = 1;
> -		else
> -			not_rampage = 1;
> +	res.start = start & PHYSICAL_PAGE_MASK;
>  
> -		if (ram_page == not_rampage)
> +	/*
> +	 * For legacy reasons, physical address range in the legacy ISA
> +	 * region is tracked as non-RAM. This will allow users of
> +	 * /dev/mem to map portions of legacy ISA region, even when
> +	 * some of those portions are listed(or not even listed) with
> +	 * different e820 types(RAM/reserved/..)
> +	 */
> +	if (res.start < ISA_END_ADDRESS) {
> +		not_ram = 1;
> +		res.start = ISA_END_ADDRESS;
> +	}
> +
> +	pg_end = (end + PAGE_SIZE - 1) & PHYSICAL_PAGE_MASK;
> +	res.end = pg_end;
> +	res.flags = IORESOURCE_MEM | IORESOURCE_BUSY;
> +	after_ram = res.start;
> +	while ((res.start < res.end) &&
> +		(find_next_system_ram(&res, "System RAM") >= 0)) {
> +		if (res.start > after_ram)
> +			not_ram = 1;
> +		if (res.end > res.start)
> +			ram = 1;
> +
> +		if (ram && not_ram)
>  			return -1;
> +
> +		after_ram = res.end + 1;
> +		res.start = res.end + 1;
> +		res.end = pg_end;
>  	}

Instead of duplicating what kernel/resource.c:walk_system_ram_range() is
already doing, can we just provide a callback that can be used with
walk_system_ram_range() and see if the expected RAM pages is what the
callback also sees.

That will greatly simplify the patch and avoid code duplication.

thanks,
suresh

Re: [PATCH] mm, x86, pat: Improve scaling of pat_pagerange_is_ram()

[Ingo Molnar]

* Suresh Siddha <suresh.b.siddha@intel.com> wrote:

> On Mon, 2012-05-14 at 15:26 -0500, John Dykstra wrote:
> > Function pat_pagerange_is_ram() scales poorly to large address ranges,
> > because it probes the resource tree for each page.  On a 2.6 GHz
> > Opteron, this function consumes 34 ms. for a 1 GB range.  It is called
> > twice during untrack_pfn_vma(), slowing process cleanup and handicapping
> > the OOM killer.
> > 
> > This replacement based on walk_system_ram_range() consumes less than 1
> > ms. under the same conditions.

Nice performance improvement!

> > 
> > Signed-off-by: John Dykstra <jdykstra@cray.com> on behalf of Cray Inc. 
> > Cc: Suresh Siddha <suresh.b.siddha@intel.com>
> > ---
> >  arch/x86/mm/pat.c      |   55 ++++++++++++++++++++++++++++++-----------------
> >  include/linux/ioport.h |    2 +
> >  kernel/resource.c      |    2 +-
> >  3 files changed, 38 insertions(+), 21 deletions(-)
> > 
> > diff --git a/arch/x86/mm/pat.c b/arch/x86/mm/pat.c
> > index f6ff57b..c119afb 100644
> > --- a/arch/x86/mm/pat.c
> > +++ b/arch/x86/mm/pat.c
> > @@ -160,29 +160,44 @@ static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
> >  
> >  static int pat_pagerange_is_ram(resource_size_t start, resource_size_t end)
> >  {
> > -	int ram_page = 0, not_rampage = 0;
> > -	unsigned long page_nr;
> > +	struct resource res;
> > +	resource_size_t pg_end, after_ram;
> > +	int ram = 0, not_ram = 0;
> >  
> > -	for (page_nr = (start >> PAGE_SHIFT); page_nr < (end >> PAGE_SHIFT);
> > -	     ++page_nr) {
> > -		/*
> > -		 * For legacy reasons, physical address range in the legacy ISA
> > -		 * region is tracked as non-RAM. This will allow users of
> > -		 * /dev/mem to map portions of legacy ISA region, even when
> > -		 * some of those portions are listed(or not even listed) with
> > -		 * different e820 types(RAM/reserved/..)
> > -		 */
> > -		if (page_nr >= (ISA_END_ADDRESS >> PAGE_SHIFT) &&
> > -		    page_is_ram(page_nr))
> > -			ram_page = 1;
> > -		else
> > -			not_rampage = 1;
> > +	res.start = start & PHYSICAL_PAGE_MASK;
> >  
> > -		if (ram_page == not_rampage)
> > +	/*
> > +	 * For legacy reasons, physical address range in the legacy ISA
> > +	 * region is tracked as non-RAM. This will allow users of
> > +	 * /dev/mem to map portions of legacy ISA region, even when
> > +	 * some of those portions are listed(or not even listed) with
> > +	 * different e820 types(RAM/reserved/..)
> > +	 */
> > +	if (res.start < ISA_END_ADDRESS) {
> > +		not_ram = 1;
> > +		res.start = ISA_END_ADDRESS;
> > +	}
> > +
> > +	pg_end = (end + PAGE_SIZE - 1) & PHYSICAL_PAGE_MASK;
> > +	res.end = pg_end;
> > +	res.flags = IORESOURCE_MEM | IORESOURCE_BUSY;
> > +	after_ram = res.start;
> > +	while ((res.start < res.end) &&
> > +		(find_next_system_ram(&res, "System RAM") >= 0)) {
> > +		if (res.start > after_ram)
> > +			not_ram = 1;
> > +		if (res.end > res.start)
> > +			ram = 1;
> > +
> > +		if (ram && not_ram)
> >  			return -1;
> > +
> > +		after_ram = res.end + 1;
> > +		res.start = res.end + 1;
> > +		res.end = pg_end;
> >  	}
> 
> Instead of duplicating what kernel/resource.c:walk_system_ram_range() is
> already doing, can we just provide a callback that can be used with
> walk_system_ram_range() and see if the expected RAM pages is what the
> callback also sees.
> 
> That will greatly simplify the patch and avoid code duplication.

Agreed.

Thanks,

	Ingo