[RFC] patch [1/1] x86_64 numa aware sparsemem add_memory functinality

[RFC] patch [1/1] x86_64 numa aware sparsemem add_memory functinality

[keith mannthey]

[-- Attachment #1: Type: text/plain, Size: 1179 bytes --]

Hello all,
  This patch is an attempt to add a numa ware add_memory functionality
to x86_64 using CONFIG_SPARSEMEM.  The add memory function today just
grabs the pgdat from node 0 and adds the memory there.  On a numa system
this is functional but not optimal/correct. 

  The SRAT can expose future memory locality.  This information is
already tracked by the nodes_add data structure (it keeps the
memory/node locality information) from the SRAT code.  The code in
srat.c is built around RESERVE_HOTADD.  This patch is a little subtle in
the way it uses the existing code for use with sparsemem.  Perhaps
acpi_numa_memory_affinity_init needs a larger refactor to fit both
RESERVE_HOTADD and sparsemem.  

  This patch still hotadd_percent as a flag to the whole srat parsing
code to disable and contain broken bios.  It's functionality is retained
and an on off switch to sparsemem hot-add.  Without changing the safety
mechanisms build into the current SRAT code I have provided a path for
the sparsemem hot-add path to get to the nodes_add data for use at
runtime. 

  This is a 1st run at the patch, it works with 2.6.17

Signed-off-by:  Keith Mannthey <kmannth@us.ibm.com>

[-- Attachment #2: patch-2.6.17-nodes-add-v1.patch --]
[-- Type: text/x-patch, Size: 3550 bytes --]

diff -urN linux-2.6.17/arch/x86_64/mm/init.c linux-2.6.17-work/arch/x86_64/mm/init.c
--- linux-2.6.17/arch/x86_64/mm/init.c	2006-06-17 21:49:35.000000000 -0400
+++ linux-2.6.17-work/arch/x86_64/mm/init.c	2006-06-20 21:41:30.000000000 -0400
@@ -553,7 +553,7 @@
  */
 int add_memory(u64 start, u64 size)
 {
-	struct pglist_data *pgdat = NODE_DATA(0);
+	struct pglist_data *pgdat = NODE_DATA(new_memory_to_node(start,start+size));
 	struct zone *zone = pgdat->node_zones + MAX_NR_ZONES-2;
 	unsigned long start_pfn = start >> PAGE_SHIFT;
 	unsigned long nr_pages = size >> PAGE_SHIFT;
diff -urN linux-2.6.17/arch/x86_64/mm/srat.c linux-2.6.17-work/arch/x86_64/mm/srat.c
--- linux-2.6.17/arch/x86_64/mm/srat.c	2006-06-20 20:25:33.000000000 -0400
+++ linux-2.6.17-work/arch/x86_64/mm/srat.c	2006-06-20 21:44:54.000000000 -0400
@@ -32,10 +32,10 @@
 static nodemask_t nodes_parsed __initdata;
 static nodemask_t nodes_found __initdata;
 static struct bootnode nodes[MAX_NUMNODES] __initdata;
-static struct bootnode nodes_add[MAX_NUMNODES] __initdata;
+static struct bootnode nodes_add[MAX_NUMNODES];
 static int found_add_area __initdata;
 int hotadd_percent __initdata = 0;
-#ifndef RESERVE_HOTADD 
+#if !defined(RESERVE_HOTADD) && !defined(CONFIG_MEMORY_HOTPLUG)
 #define hotadd_percent 0	/* Ignore all settings */
 #endif
 static u8 pxm2node[256] = { [0 ... 255] = 0xff };
@@ -219,9 +219,9 @@
 	allocated += mem;
 	return 1;
 }
-
+#endif
 /*
- * It is fine to add this area to the nodes data it will be used later
+ * It is fine to add this area to the nodes_add data it will be used later
  * This code supports one contigious hot add area per node.
  */
 static int reserve_hotadd(int node, unsigned long start, unsigned long end)
@@ -247,15 +247,14 @@
 		printk(KERN_ERR "SRAT: Hotplug area has existing memory\n");
 		return -1;
 	}
-
+#ifdef RESERVE_HOTADD
 	if (!hotadd_enough_memory(&nodes_add[node]))  {
 		printk(KERN_ERR "SRAT: Hotplug area too large\n");
 		return -1;
 	}
-
+#endif 
 	/* Looks good */
 
- 	found_add_area = 1;
 	if (nd->start == nd->end) {
  		nd->start = start;
  		nd->end = end;
@@ -273,14 +272,16 @@
 			printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
  	}
 
- 	if ((nd->end >> PAGE_SHIFT) > end_pfn)
- 		end_pfn = nd->end >> PAGE_SHIFT;
-
 	if (changed)
 	 	printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
+#ifdef RESERVE_HOTADD	
+ 	found_add_area = 1;
+	if ((nd->end >> PAGE_SHIFT) > end_pfn)
+ 		end_pfn = nd->end >> PAGE_SHIFT;
 	return 0;
+#endif 
+	return -1;
 }
-#endif
 
 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
 void __init
@@ -338,7 +339,6 @@
 	printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm,
 	       nd->start, nd->end);
 
-#ifdef RESERVE_HOTADD
  	if (ma->flags.hot_pluggable && reserve_hotadd(node, start, end) < 0) {
 		/* Ignore hotadd region. Undo damage */
 		printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
@@ -346,7 +346,6 @@
 		if ((nd->start | nd->end) == 0)
 			node_clear(node, nodes_parsed);
 	}
-#endif
 }
 
 /* Sanity check to catch more bad SRATs (they are amazingly common).
@@ -479,5 +478,15 @@
 	index = acpi_slit->localities * node_to_pxm(a);
 	return acpi_slit->entry[index + node_to_pxm(b)];
 }
-
 EXPORT_SYMBOL(__node_distance);
+
+int new_memory_to_node(unsigned long start, unsigned long end) {
+	int i,ret;
+	ret=0;
+	for_each_node(i){
+		if (nodes_add[i].start <= start && nodes_add[i].end >= end)
+			ret = i;		
+	}
+	return ret;
+}
+EXPORT_SYMBOL(new_memory_to_node);

Re: [Lhms-devel] [RFC] patch [1/1] x86_64 numa aware sparsemem add_memory functinality

[KAMEZAWA Hiroyuki]

On Tue, 20 Jun 2006 22:43:01 -0700
keith mannthey <kmannth@us.ibm.com> wrote:

> Hello all,
>   This patch is an attempt to add a numa ware add_memory functionality
> to x86_64 using CONFIG_SPARSEMEM.  The add memory function today just
> grabs the pgdat from node 0 and adds the memory there.  On a numa system
> this is functional but not optimal/correct. 
> 

At first, sorry for confusing.
reserve_hotadd()/memory-hot-add with preallocated mem_map things are 
maintained by x86_64 and Andi Kleen (maybe).
So we (lhms people) are not familiar with this.

And yes, mem_map should be allocated from local node.
I'm now preparing "dynamic local mem_map allocation" for lhms's memory hotplug,
which doesn't depend on SRAT.

Regards,
-Kame


>   The SRAT can expose future memory locality.  This information is
> already tracked by the nodes_add data structure (it keeps the
> memory/node locality information) from the SRAT code.  The code in
> srat.c is built around RESERVE_HOTADD.  This patch is a little subtle in
> the way it uses the existing code for use with sparsemem.  Perhaps
> acpi_numa_memory_affinity_init needs a larger refactor to fit both
> RESERVE_HOTADD and sparsemem.  
> 
>   This patch still hotadd_percent as a flag to the whole srat parsing
> code to disable and contain broken bios.  It's functionality is retained
> and an on off switch to sparsemem hot-add.  Without changing the safety
> mechanisms build into the current SRAT code I have provided a path for
> the sparsemem hot-add path to get to the nodes_add data for use at
> runtime. 
> 
>   This is a 1st run at the patch, it works with 2.6.17
> 
> Signed-off-by:  Keith Mannthey <kmannth@us.ibm.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: [Lhms-devel] [RFC] patch [1/1] x86_64 numa aware sparsemem add_memory functinality

[keith mannthey]

On Wed, 2006-06-21 at 15:06 +0900, KAMEZAWA Hiroyuki wrote:
> On Tue, 20 Jun 2006 22:43:01 -0700
> keith mannthey <kmannth@us.ibm.com> wrote:
> 
> > Hello all,
> >   This patch is an attempt to add a numa ware add_memory functionality
> > to x86_64 using CONFIG_SPARSEMEM.  The add memory function today just
> > grabs the pgdat from node 0 and adds the memory there.  On a numa system
> > this is functional but not optimal/correct. 
> > 
> 
> At first, sorry for confusing.
> reserve_hotadd()/memory-hot-add with preallocated mem_map things are 
> maintained by x86_64 and Andi Kleen (maybe).
> So we (lhms people) are not familiar with this.
Agreeded. I don't expect lhms to know much about reserve_hotadd(). 
  Right now SPARSEMEM adds all it's memory to node 0 in x86_64.  This is
the problem I am trying to fix.  I doesn't make sense to me to rewrite
the SRAT code when RESERVE_HOTADD has done most of the work already. 

> And yes, mem_map should be allocated from local node.
> I'm now preparing "dynamic local mem_map allocation" for lhms's memory hotplug,
> which doesn't depend on SRAT.

How do you know which node to add the memory too without something like
the SRAT that define memory locality of hot-add zones? SPARSEMEM doesn't
depend on SRAT (it just needs to use to to know what zone to add to.)

This patch isn't about mem_map allocation rather what zone to add the
memory to when doing SPASEMEM hot-add.  A numa aware mem_map allocation
would belong in generic SPARSEMEM code. 

> 
> >   The SRAT can expose future memory locality.  This information is
> > already tracked by the nodes_add data structure (it keeps the
> > memory/node locality information) from the SRAT code.  The code in
> > srat.c is built around RESERVE_HOTADD.  This patch is a little subtle in
> > the way it uses the existing code for use with sparsemem.  Perhaps
> > acpi_numa_memory_affinity_init needs a larger refactor to fit both
> > RESERVE_HOTADD and sparsemem.  
> > 
> >   This patch still hotadd_percent as a flag to the whole srat parsing
> > code to disable and contain broken bios.  It's functionality is retained
> > and an on off switch to sparsemem hot-add.  Without changing the safety
> > mechanisms build into the current SRAT code I have provided a path for
> > the sparsemem hot-add path to get to the nodes_add data for use at
> > runtime. 
> > 
> >   This is a 1st run at the patch, it works with 2.6.17
> > 
> > Signed-off-by:  Keith Mannthey <kmannth@us.ibm.com>
> > 
> 
> 
> 
> _______________________________________________
> Lhms-devel mailing list
> Lhms-devel@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/lhms-devel
-- 
keith mannthey <kmannth@us.ibm.com>
Linux Technology Center IBM

--
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: [Lhms-devel] [RFC] patch [1/1] x86_64 numa aware sparsemem add_memory functinality

[Yasunori Goto]

> On Tue, 20 Jun 2006 22:43:01 -0700
> keith mannthey <kmannth@us.ibm.com> wrote:
> 
> > Hello all,
> >   This patch is an attempt to add a numa ware add_memory functionality
> > to x86_64 using CONFIG_SPARSEMEM.  The add memory function today just
> > grabs the pgdat from node 0 and adds the memory there.  On a numa system
> > this is functional but not optimal/correct. 
> > 
> 
> At first, sorry for confusing.
> reserve_hotadd()/memory-hot-add with preallocated mem_map things are 
> maintained by x86_64 and Andi Kleen (maybe).
> So we (lhms people) are not familiar with this.
>
> And yes, mem_map should be allocated from local node.
> I'm now preparing "dynamic local mem_map allocation" for lhms's memory hotplug,
> which doesn't depend on SRAT.

I wrote patches for NUMA aware memory hotplug with sparsemem.
It is already included in current -mm.
He means he would like to make the patch for -mm. Could you check it?
But, I've not try it with RESERVE_HOT_ADD. I just try it with sparsemem.
Sorry. :-(

In my patch, if new memory is in new node, new node id is decided by PXM
in dsdt. So, it must work even if srat does not define hot-add area.

Thanks.

-- 
Yasunori Goto 


--
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: [Lhms-devel] [RFC] patch [1/1] x86_64 numa aware sparsemem add_memory functinality

[KAMEZAWA Hiroyuki]

On Tue, 20 Jun 2006 23:25:01 -0700
keith mannthey <kmannth@us.ibm.com> wrote:

> > And yes, mem_map should be allocated from local node.
> > I'm now preparing "dynamic local mem_map allocation" for lhms's memory hotplug,
> > which doesn't depend on SRAT.
> 
> How do you know which node to add the memory too without something like
> the SRAT that define memory locality of hot-add zones? SPARSEMEM doesn't
> depend on SRAT (it just needs to use to to know what zone to add to.)
> 
Now, acpi's _PXM method is supported by acpi-memory-hotadd. (See -mm.)
I'll use it. and current add_memory() -mm is this.
==
int arch_add_memory(int nid, u64 start, u64 size)
{
        struct pglist_data *pgdat = NODE_DATA(nid);
        struct zone *zone = pgdat->node_zones + MAX_NR_ZONES-2;
        unsigned long start_pfn = start >> PAGE_SHIFT;
        unsigned long nr_pages = size >> PAGE_SHIFT;
==
nid is passed by caller.


> This patch isn't about mem_map allocation rather what zone to add the
> memory to when doing SPASEMEM hot-add.  A numa aware mem_map allocation
> would belong in generic SPARSEMEM code. 
> 
I also need to do NUMA-aware 
- pgdat allocation , wait table allocation ..and so on...

I'll add memory allocater which allocates memory from newly-added-memory.

-Kame

--
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] patch [1/1] x86_64 numa aware sparsemem add_memory functinality

[Dave Hansen]

>  int add_memory(u64 start, u64 size)
>  {
> -       struct pglist_data *pgdat = NODE_DATA(0);
> +       struct pglist_data *pgdat = NODE_DATA(new_memory_to_node(start,start+size));
>         struct zone *zone = pgdat->node_zones + MAX_NR_ZONES-2;

How about just having new_memory_to_node() take the range and return the
pgdat?  Should make that line a bit shorter.

> -#ifndef RESERVE_HOTADD 
> +#if !defined(RESERVE_HOTADD) && !defined(CONFIG_MEMORY_HOTPLUG)
>  #define hotadd_percent 0       /* Ignore all settings */
>  #endif
>  static u8 pxm2node[256] = { [0 ... 255] = 0xff };
> @@ -219,9 +219,9 @@
>         allocated += mem;
>         return 1;
>  }
> -
> +#endif
>  /*

Could this use another Kconfig option which gives a name to this
condition?

> +#ifdef RESERVE_HOTADD
>         if (!hotadd_enough_memory(&nodes_add[node]))  {
>                 printk(KERN_ERR "SRAT: Hotplug area too large\n");
>                 return -1;
>         }
> -
> +#endif 

This #ifdef is probably better handled by an #ifdef in the header for
hotadd_enough_memory().

-- Dave

--
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: [Lhms-devel] [RFC] patch [1/1] x86_64 numa aware sparsemem add_memory functinality

[keith mannthey]

On Fri, 2006-06-23 at 10:13 -0700, Dave Hansen wrote: 
> >  int add_memory(u64 start, u64 size)
> >  {
> > -       struct pglist_data *pgdat = NODE_DATA(0);
> > +       struct pglist_data *pgdat = NODE_DATA(new_memory_to_node(start,start+size));
> >         struct zone *zone = pgdat->node_zones + MAX_NR_ZONES-2;
> 
> How about just having new_memory_to_node() take the range and return the
> pgdat?  Should make that line a bit shorter.

In the -mm tree things are a little different.  The acpi layer (and
something for ppc) is passing the nid down the a generic add memory
call.  

  This int add_memory(u64 start, u64 size) is going away with something
more like int add_memory(int nid, u64 start, u64 size) this changes
things some. 

  I have patches against the -mm stack but I had a little trouble with
my testbox's file-system last night so I should have them out this
afternoon.  

> > -#ifndef RESERVE_HOTADD 
> > +#if !defined(RESERVE_HOTADD) && !defined(CONFIG_MEMORY_HOTPLUG)
> >  #define hotadd_percent 0       /* Ignore all settings */
> >  #endif
> >  static u8 pxm2node[256] = { [0 ... 255] = 0xff };
> > @@ -219,9 +219,9 @@
> >         allocated += mem;
> >         return 1;
> >  }
> > -
> > +#endif
> >  /*
> 
> Could this use another Kconfig option which gives a name to this
> condition?

  This is sort of a redundant force off.  I am not sure if there is a
code path to the SRAT code without  RESERVE_HOTADD or
CONFIG_MEMORY_HOTPLUG defined.  

  hotadd_percent can only change from 0 with an explicit command-line
numa=hotadd=XXX boot so maybe taking this  
#define hotadd_percent 0
out all together might be the better way to go if the code patch is
going to be shared. 

> 
> > +#ifdef RESERVE_HOTADD
> >         if (!hotadd_enough_memory(&nodes_add[node]))  {
> >                 printk(KERN_ERR "SRAT: Hotplug area too large\n");
> >                 return -1;
> >         }
> > -
> > +#endif 
> 
> This #ifdef is probably better handled by an #ifdef in the header for
> hotadd_enough_memory().

  hotadd_enough_memory is static there is no header entry for it. 

Thanks for the feedback,
  Keith 

--
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>

[RFC] Patch [1/4] x86_64 sparsmem add- save nodes_add data for later

[keith mannthey]

[-- Attachment #1: Type: text/plain, Size: 833 bytes --]

Hello All,
  The following patch reserves the nodes_add data for use later.  It
retains the bulk of the hotadd_precent defense that is built into the
SRAT code. It is a little subtle but it gets the job done.  

  The code saves off the hot-add area ranges without extending the end
of memory.  It then creates arch_find_node which will be use in the next
patch.  

 arch_find_node is passed a memory range and it looks for a nodes_add
area it fits into.   If no area is found it returns 0.

 With this and the other 3 patches I can do SPARSEMEM x86_64 hot-add on
my hardware. (the first 2 patches are one I consider real the other 2
are more to point out issues)

 It is built against 2.6.17-mm1 x86_64 and the current memory_hotplug
work but should apply on any current srat.c

Signed-off-by:  Keith Mannthey <kmannth@us.ibm.com>


[-- Attachment #2: patch-2.6.17-mm1-save_nodes_add --]
[-- Type: text/x-patch, Size: 2650 bytes --]

diff -urN linux-2.6.17-mm1-orig/arch/x86_64/mm/srat.c linux-2.6.17-mm1/arch/x86_64/mm/srat.c
--- linux-2.6.17-mm1-orig/arch/x86_64/mm/srat.c	2006-06-23 16:12:00.000000000 -0400
+++ linux-2.6.17-mm1/arch/x86_64/mm/srat.c	2006-06-23 18:43:03.000000000 -0400
@@ -34,9 +34,6 @@
 static struct bootnode nodes_add[MAX_NUMNODES] __initdata;
 static int found_add_area __initdata;
 int hotadd_percent __initdata = 0;
-#ifndef RESERVE_HOTADD
-#define hotadd_percent 0	/* Ignore all settings */
-#endif
 
 /* Too small nodes confuse the VM badly. Usually they result
    from BIOS bugs. */
@@ -199,9 +196,9 @@
 	allocated += mem;
 	return 1;
 }
-
+#endif
 /*
- * It is fine to add this area to the nodes data it will be used later
+ * It is fine to add this area to the nodes_add data it will be used later
  * This code supports one contigious hot add area per node.
  */
 static int reserve_hotadd(int node, unsigned long start, unsigned long end)
@@ -227,15 +224,14 @@
 		printk(KERN_ERR "SRAT: Hotplug area has existing memory\n");
 		return -1;
 	}
-
+#ifdef RESERVE_HOTADD
 	if (!hotadd_enough_memory(&nodes_add[node]))  {
 		printk(KERN_ERR "SRAT: Hotplug area too large\n");
 		return -1;
 	}
-
+#endif 
 	/* Looks good */
 
- 	found_add_area = 1;
 	if (nd->start == nd->end) {
  		nd->start = start;
  		nd->end = end;
@@ -253,14 +249,16 @@
 			printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
  	}
 
- 	if ((nd->end >> PAGE_SHIFT) > end_pfn)
- 		end_pfn = nd->end >> PAGE_SHIFT;
-
 	if (changed)
 	 	printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
+#ifdef RESERVE_HOTADD	
+ 	found_add_area = 1;
+	if ((nd->end >> PAGE_SHIFT) > end_pfn)
+ 		end_pfn = nd->end >> PAGE_SHIFT;
 	return 0;
+#endif 
+	return -1;
 }
-#endif
 
 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
 void __init
@@ -318,7 +316,6 @@
 	printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm,
 	       nd->start, nd->end);
 
-#ifdef RESERVE_HOTADD
  	if (ma->flags.hot_pluggable && reserve_hotadd(node, start, end) < 0) {
 		/* Ignore hotadd region. Undo damage */
 		printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
@@ -326,7 +323,6 @@
 		if ((nd->start | nd->end) == 0)
 			node_clear(node, nodes_parsed);
 	}
-#endif
 }
 
 /* Sanity check to catch more bad SRATs (they are amazingly common).
@@ -450,3 +446,14 @@
 }
 
 EXPORT_SYMBOL(__node_distance);
+
+int arch_find_node(unsigned long start, unsigned long size) {
+	int i, ret = 0;
+	unsigned long end = start+size;
+
+	for_each_node(i){
+		if (nodes_add[i].start <= start && nodes_add[i].end >= end)
+			ret = i;
+	}
+	return ret;
+}