Skip to contentSkip to conversation

add dev_to_node()

PATCH In review Christoph Hellwig 7 participants 17 messages Oct 30, 2006 – Nov 11, 2006

Subsystems (7)
lore
Share & tools
lore
#0 lore raw

PATCH 2/3 add dev_to_node()

Davem suggested to get the node-affinity information directly from struct device instead of having the caller extreact it from the pci_dev. This patch adds dev_to_node() to the topology API for that. The implementation is rather ugly as we need to compare the bus operations which we can't do inline in a header without pulling all kinds of mess in.

Thus provide an out of line dev_to_node for ppc and let everyone else use the dummy variant in asm-generic.h for now.

include/asm-generic/topology.h +4 -1
Index: linux-2.6/include/asm-generic/topology.h
===================================================================
@@ -45,11 +45,14 @@
 #define pcibus_to_node(node)	(-1)
 #endif
 
+#ifndef dev_to_node
+#define dev_to_node(dev)	(-1)
+#endif
+
 #ifndef pcibus_to_cpumask
 #define pcibus_to_cpumask(bus)	(pcibus_to_node(bus) == -1 ? \
 					CPU_MASK_ALL : \
 					node_to_cpumask(pcibus_to_node(bus)) \
 				)
 #endif
-
 #endif /* _ASM_GENERIC_TOPOLOGY_H */
include/asm-powerpc/topology.h +2 -0
Index: linux-2.6/include/asm-powerpc/topology.h
===================================================================
@@ -5,6 +5,7 @@
 
 struct sys_device;
 struct device_node;
+struct device;
 
 #ifdef CONFIG_NUMA
 
@@ -33,6 +34,7 @@
 
 struct pci_bus;
 extern int pcibus_to_node(struct pci_bus *bus);
+int dev_to_node(struct device *dev);
 
 #define pcibus_to_cpumask(bus)	(pcibus_to_node(bus) == -1 ? \
 					CPU_MASK_ALL : \
arch/powerpc/kernel/pci_64.c +8 -0
Index: linux-2.6/arch/powerpc/kernel/pci_64.c
===================================================================
@@ -1424,4 +1424,12 @@
 	return phb->node;
 }
 EXPORT_SYMBOL(pcibus_to_node);
+
+int dev_to_node(struct device *dev)
+{
+	if (dev->bus == &pci_bus_type)
+		return pcibus_to_node(to_pci_dev(dev)->bus);
+	return -1;
+}
+EXPORT_SYMBOL(dev_to_node);
 #endif

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

maintainer
#1 lore raw
Quoted 11 lines

Davem suggested to get the node-affinity information directly from struct device instead of having the caller extreact it from the pci_dev. This patch adds dev_to_node() to the topology API for that. The implementation is rather ugly as we need to compare the bus operations which we can't do inline in a header without pulling all kinds of mess in.

Thus provide an out of line dev_to_node for ppc and let everyone else use the dummy variant in asm-generic.h for now.

It may be a bit much to be calling all the way through up to the PCI layer just to pluck out a simple integer, don't you think? The PCI bus pointer comparison is just a symptom of how silly this is.

Especially since this will be used for every packet allocation a device makes.

So, please add some sanity to this situation and just put the node into the generic struct device. :-)

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

maintainer
#3 lore raw
Quoted 7 lines · David Miller
Quoted 3 lines

So, please add some sanity to this situation and just put the node into the generic struct device. :-)

Good. Then we can remove the node from the pci structure and get rid of pcibus_to_node?

Yes, that's possible, because the idea is that the arch specific bus layer code would initialize the node value. Therefore, there would be no need for things like pcibus_to_node() any longer.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#4 lore raw
Quoted 3 lines · David Miller

Yes, that's possible, because the idea is that the arch specific bus layer code would initialize the node value. Therefore, there would be no need for things like pcibus_to_node() any longer.

Then lets rename pcibus_to_node to dev_to_node() throughout the kernel. Provide a -1 default. Then other device layers that are not based on pci will also be able to exploit NUMA locality.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#5 lore raw
In reply to David Miller
Quoted 9 lines · David Miller

It may be a bit much to be calling all the way through up to the PCI layer just to pluck out a simple integer, don't you think? The PCI bus pointer comparison is just a symptom of how silly this is.

Especially since this will be used for every packet allocation a device makes.

So, please add some sanity to this situation and just put the node into the generic struct device. :-)

I was concerned about growing struct device, on smaller system it already eats up a lot of memory. But we can make the node member conditional on CONFIG_NUMA, as I did in the patch below.

This directly replaces PATCH 2/2 (the one we're replying to), all others remain unmodified.

include/linux/device.h +9 -0
Index: linux-2.6/include/linux/device.h
===================================================================
@@ -347,6 +347,9 @@
 					   BIOS data),reserved for device core*/
 	struct dev_pm_info	power;
 
+#ifdef CONFIG_NUMA
+	int		numa_node;	/* NUMA node this device is close to */
+#endif
 	u64		*dma_mask;	/* dma mask (if dma'able device) */
 	u64		coherent_dma_mask;/* Like dma_mask, but for
 					     alloc_coherent mappings as
@@ -368,6 +371,12 @@
 	void	(*release)(struct device * dev);
 };
 
+#ifdef CONFIG_NUMA
+#define dev_to_node(dev)	((dev)->numa_node)
+#else
+#define dev_to_node(dev)	(-1)
+#endif
+
 static inline void *
 dev_get_drvdata (struct device *dev)
 {
drivers/base/core.c +1 -0
Index: linux-2.6/drivers/base/core.c
===================================================================
@@ -381,6 +381,7 @@
 	INIT_LIST_HEAD(&dev->node);
 	init_MUTEX(&dev->sem);
 	device_init_wakeup(dev, 0);
+	dev->numa_node = -1;
 }
 
 /**
drivers/pci/probe.c +1 -0
Index: linux-2.6/drivers/pci/probe.c
===================================================================
@@ -846,6 +846,7 @@
 	dev->dev.release = pci_release_dev;
 	pci_dev_get(dev);
 
+	dev->dev.numa_node = pcibus_to_node(bus);
 	dev->dev.dma_mask = &dev->dma_mask;
 	dev->dev.coherent_dma_mask = 0xffffffffull;
 

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#6 lore raw

On Sat, Nov 04, 2006 at 11:56:29PM +0100, Christoph Hellwig wrote:

This will break the compile for !NUMA if someone ends up doing a bisect and lands here as a bisect point.

You introduce this nice wrapper..

include/linux/device.h line 382

Quoted 9 lines
include/linux/device.h +6 -0
+#ifdef CONFIG_NUMA
+#define dev_to_node(dev)	((dev)->numa_node)
+#else
+#define dev_to_node(dev)	(-1)
+#endif
+
 static inline void *
 dev_get_drvdata (struct device *dev)
 {

And then don't use it here..

drivers/base/core.c line 387

Quoted 12 lines

Index: linux-2.6/drivers/base/core.c
===================================================================

drivers/base/core.c +1 -0
@@ -381,6 +381,7 @@
 	INIT_LIST_HEAD(&dev->node);
 	init_MUTEX(&dev->sem);
 	device_init_wakeup(dev, 0);
+	dev->numa_node = -1;
 }
 
 /**

and here.

drivers/pci/probe.c line 851

Quoted 11 lines

Index: linux-2.6/drivers/pci/probe.c
===================================================================

drivers/pci/probe.c +1 -0
@@ -846,6 +846,7 @@
 	dev->dev.release = pci_release_dev;
 	pci_dev_get(dev);
 
+	dev->dev.numa_node = pcibus_to_node(bus);
 	dev->dev.dma_mask = &dev->dma_mask;
 	dev->dev.coherent_dma_mask = 0xffffffffull;

Dave

Signature · 7 lines
http://www.codemonkey.org.uk

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
#7 lore raw
Quoted 6 lines · Dave Jones

On Sat, Nov 04, 2006 at 11:56:29PM +0100, Christoph Hellwig wrote:

This will break the compile for !NUMA if someone ends up doing a bisect and lands here as a bisect point.

You introduce this nice wrapper..

Yes, I'm stupid :) Updated version will follow ASAP.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#8 lore raw
In reply to Dave Jones
Quoted 6 lines · Dave Jones

On Sat, Nov 04, 2006 at 11:56:29PM +0100, Christoph Hellwig wrote:

This will break the compile for !NUMA if someone ends up doing a bisect and lands here as a bisect point.

You introduce this nice wrapper..

The dev_to_node wrapper is not enough as we can't assign to (-1) for the non-NUMA case. So I added a second macro, set_dev_node for that.

The patch below compiles and works on numa and non-NUMA platforms.

include/linux/device.h +11 -0
Index: linux-2.6/include/linux/device.h
===================================================================
@@ -347,6 +347,9 @@
 					   BIOS data),reserved for device core*/
 	struct dev_pm_info	power;
 
+#ifdef CONFIG_NUMA
+	int		numa_node;	/* NUMA node this device is close to */
+#endif
 	u64		*dma_mask;	/* dma mask (if dma'able device) */
 	u64		coherent_dma_mask;/* Like dma_mask, but for
 					     alloc_coherent mappings as
@@ -368,6 +371,14 @@
 	void	(*release)(struct device * dev);
 };
 
+#ifdef CONFIG_NUMA
+#define dev_to_node(dev)	((dev)->numa_node)
+#define set_dev_node(dev, node)	((dev)->numa_node = node)
+#else
+#define dev_to_node(dev)	(-1)
+#define set_dev_node(dev, node)	do { } while (0)
+#endif
+
 static inline void *
 dev_get_drvdata (struct device *dev)
 {
drivers/base/core.c +1 -0
Index: linux-2.6/drivers/base/core.c
===================================================================
@@ -381,6 +381,7 @@
 	INIT_LIST_HEAD(&dev->node);
 	init_MUTEX(&dev->sem);
 	device_init_wakeup(dev, 0);
+	set_dev_node(dev, -1);
 }
 
 /**
drivers/pci/probe.c +1 -0
Index: linux-2.6/drivers/pci/probe.c
===================================================================
@@ -846,6 +846,7 @@
 	dev->dev.release = pci_release_dev;
 	pci_dev_get(dev);
 
+	set_dev_node(&dev->dev, pcibus_to_node(bus));
 	dev->dev.dma_mask = &dev->dma_mask;
 	dev->dev.coherent_dma_mask = 0xffffffffull;
 

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

maintainer
#9 lore raw

From: Christoph Hellwig <[email protected]>
Date: Sun, 5 Nov 2006 00:53:23 +0100

Quoted 6 lines · Dave Jones
Quoted 5 lines · Christoph Hellwig

This will break the compile for !NUMA if someone ends up doing a bisect and lands here as a bisect point.

Quoted 8 lines
Quoted 1 line

You introduce this nice wrapper..

The dev_to_node wrapper is not enough as we can't assign to (-1) for the non-NUMA case. So I added a second macro, set_dev_node for that.

The patch below compiles and works on numa and non-NUMA platforms.

Looks good to me.

#10 lore raw
Quoted 1 line · David Miller

Looks good to me.

So what's the right path to get this in? There's one patch touching MM code, one adding something to the driver core and then finally a networking patch depending on the previous two. Do you want to take them all and send them in through the networking tree? Or should we put the burden on Andrew?

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#11 lore raw
In reply to Christoph Hellwig
Quoted 6 lines · Christoph Hellwig
Quoted 5 lines · Dave Jones

On Sat, Nov 04, 2006 at 11:56:29PM +0100, Christoph Hellwig wrote:

This will break the compile for !NUMA if someone ends up doing a bisect and lands here as a bisect point.

Quoted 8 lines
Quoted 1 line

You introduce this nice wrapper..

The dev_to_node wrapper is not enough as we can't assign to (-1) for the non-NUMA case. So I added a second macro, set_dev_node for that.

The patch below compiles and works on numa and non-NUMA platforms.

Hi Christoph,
dev_to_node does not work as expected on x86_64 (and i386). This is because node value returned by pcibus_to_node is initialized after a struct device is created with current x86_64 code.

We need the node value initialized before the call to pci_scan_bus_parented, as the generic devices are allocated and initialized
off pci_scan_child_bus, which gets called from pci_scan_bus_parented The following patch does that using "pci_sysdata" introduced by the PCI domain patches in -mm.

arch/i386/pci/acpi.c +9 -2
Index: linux-2.6.19-rc4mm2/arch/i386/pci/acpi.c
===================================================================
@@ -9,6 +9,7 @@ struct pci_bus * __devinit pci_acpi_scan
 {
 	struct pci_bus *bus;
 	struct pci_sysdata *sd;
+	int pxm;
 
 	/* Allocate per-root-bus (not per bus) arch-specific data.
 	 * TODO: leak; this memory is never freed.
@@ -30,15 +31,21 @@ struct pci_bus * __devinit pci_acpi_scan
 	}
 #endif /* CONFIG_PCI_DOMAINS */
 
+	sd->node = -1;
+
+	pxm = acpi_get_pxm(device->handle);
+#ifdef CONFIG_ACPI_NUMA
+	if (pxm >= 0)
+		sd->node = pxm_to_node(pxm);
+#endif
+
 	bus = pci_scan_bus_parented(NULL, busnum, &pci_root_ops, sd);
 	if (!bus)
 		kfree(sd);
 
 #ifdef CONFIG_ACPI_NUMA
 	if (bus != NULL) {
-		int pxm = acpi_get_pxm(device->handle);
 		if (pxm >= 0) {
-			sd->node = pxm_to_node(pxm);
 			printk("bus %d -> pxm %d -> node %d\n",
 				busnum, pxm, sd->node);
 		}

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#12 lore raw
Quoted 21 lines · Ravikiran G Thirumalai
Quoted 6 lines · Christoph Hellwig
Quoted 5 lines · Dave Jones

On Sat, Nov 04, 2006 at 11:56:29PM +0100, Christoph Hellwig wrote:

This will break the compile for !NUMA if someone ends up doing a bisect and lands here as a bisect point.

Quoted 8 lines
Quoted 1 line

You introduce this nice wrapper..

The dev_to_node wrapper is not enough as we can't assign to (-1) for the non-NUMA case. So I added a second macro, set_dev_node for that.

The patch below compiles and works on numa and non-NUMA platforms.

Hi Christoph,
dev_to_node does not work as expected on x86_64 (and i386). This is because node value returned by pcibus_to_node is initialized after a struct device is created with current x86_64 code.

Quoted 5 lines

We need the node value initialized before the call to pci_scan_bus_parented, as the generic devices are allocated and initialized
off pci_scan_child_bus, which gets called from pci_scan_bus_parented The following patch does that using "pci_sysdata" introduced by the PCI domain patches in -mm.

A nice, that some non-cell folks actually care for this patch. As far as my x86_64 pci code knowledge is concerned that patch look fine to me.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#13 lore raw
In reply to Christoph Hellwig

Hi, I have a question.

include/linux/device.h line 354

Quoted 13 lines · On Sat, 4 Nov 2006 23:56:29 +0100 Christoph Hellwig <[email protected]> wrote:

Index: linux-2.6/include/linux/device.h
===================================================================

include/linux/device.h +4 -0
@@ -347,6 +347,9 @@
 					   BIOS data),reserved for device core*/
 	struct dev_pm_info	power;
 
+#ifdef CONFIG_NUMA
+	int		numa_node;	/* NUMA node this device is close to */
+#endif
 
+	dev->dev.numa_node = pcibus_to_node(bus);

Does this "node" is guaranteed to be online ?

if node is not online, NODE_DATA(node) is NULL or not initialized. Then, alloc_pages_node() at el. will panic.

I wonder there are no code for creating NODE_DATA() for device-only-node.

-Kame

#15 lore raw
Quoted 6 lines · Christoph Lameter
Quoted 1 line · KAMEZAWA Hiroyuki

I wonder there are no code for creating NODE_DATA() for device-only-node.

On IA64 we remap nodes with no memory / cpus to the nearest node with memory. I think that is sufficient.

I don't think this happens anymore. Back in the ~2.6.5 days, when we would configure our numa platforms with 100% of memory interleaved [in hardware at cache line granularity], the cpus would move to the interleaved "pseudo-node" and the memoryless nodes would be removed. numactl --hardware would show something like this:

# uname -r
2.6.5-7.244-default
# numactl --hardware
available: 1 nodes (0-0)
node 0 size: 65443 MB
node 0 free: 64506 MB

I started seeing different behavior about the time SPARSEMEM went in. Now, with a 2.6.16 base kernel [same platform, hardware interleaved memory], I see:

# uname -r# numactl --hardware
available: 5 nodes (0-4)
node 0 size: 0 MB
node 0 free: 0 MB
node 1 size: 0 MB
node 1 free: 0 MB
node 2 size: 0 MB
node 2 free: 0 MB
node 3 size: 0 MB
node 3 free: 0 MB
node 4 size: 65439 MB
node 4 free: 64492 MB
node distances:
node   0   1   2   3   4
  0:  10  17  17  17  14
  1:  17  10  17  17  14
  2:  17  17  10  17  14
  3:  17  17  17  10  14
  4:  14  14  14  14  10
2.6.16.21-0.8-default

[Aside: The firmware/SLIT says that the interleaved memory is closer to all nodes that other nodes' memory. This has interesting implications for the "overflow" zone lists...]

Lee

#16 lore raw
Quoted 9 lines · On Fri, 10 Nov 2006 13:28:25 -0500 Lee Schermerhorn <[email protected]>…
Quoted 6 lines · Christoph Lameter
Quoted 1 line · KAMEZAWA Hiroyuki

I wonder there are no code for creating NODE_DATA() for device-only-node.

On IA64 we remap nodes with no memory / cpus to the nearest node with memory. I think that is sufficient.

I don't think this happens anymore.

In my understanding , from drivers/acpi/numa.c,
a node is created by a pxm found in SRAT table at boot time.

the node-number for the pxm which was not found in SRAT at boot time is "-1". please check how acpi_map_pxm_to_node() is used.

If pci's node-id is based on pxm, checking return vaule of pxm_to_node() will be good.

-Kame

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>