[PATCH 0/4] NUMA support for spufs

[PATCH 0/4] NUMA support for spufs

[Arnd Bergmann]

The current version of spufs breaks upon boot when NUMA support is enabled.
I'd like to fix that before 2.6.17 so we can use the same kernel image
on Cell that we use on other powerpc systems using NUMA.

This series has four patches, the first two are required to make it work
at all, the other two are there for a somewhat cleaner solution.

The second patch (sparsemem-interaction-with-memory-add-bug-fixes.patch)
actually comes from 2.6.17-rc2-mm1, but in order to make this work,
we need it included in mainline along with the others.

	Arnd <><
--

[PATCH 1/4] spufs: fix for CONFIG_NUMA

[Arnd Bergmann]

From: Joel H Schopp <jschopp@us.ibm.com>

Based on an older patch from  Mike Kravetz <kravetz@us.ibm.com>

We need to have a mem_map for high addresses in order to make
fops->no_page work on spufs mem and register files. So far, we
have used the memory_present() function during early bootup,
but that did not work when CONFIG_NUMA was enabled.

We now use the __add_pages() function to add the mem_map
when loading the spufs module, which is a lot nicer.

Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
---
Index: linus-2.6/arch/powerpc/platforms/cell/setup.c
===================================================================
--- linus-2.6.orig/arch/powerpc/platforms/cell/setup.c	2006-04-29 02:51:29.000000000 +0200
+++ linus-2.6/arch/powerpc/platforms/cell/setup.c	2006-04-29 02:51:33.000000000 +0200
@@ -29,6 +29,8 @@
 #include <linux/seq_file.h>
 #include <linux/root_dev.h>
 #include <linux/console.h>
+#include <linux/mutex.h>
+#include <linux/memory_hotplug.h>
 
 #include <asm/mmu.h>
 #include <asm/processor.h>
@@ -46,6 +48,7 @@
 #include <asm/cputable.h>
 #include <asm/ppc-pci.h>
 #include <asm/irq.h>
+#include <asm/spu.h>
 
 #include "interrupt.h"
 #include "iommu.h"
@@ -69,77 +72,6 @@
 	of_node_put(root);
 }
 
-#ifdef CONFIG_SPARSEMEM
-static int __init find_spu_node_id(struct device_node *spe)
-{
-	unsigned int *id;
-#ifdef CONFIG_NUMA
-	struct device_node *cpu;
-	cpu = spe->parent->parent;
-	id = (unsigned int *)get_property(cpu, "node-id", NULL);
-#else
-	id = NULL;
-#endif
-	return id ? *id : 0;
-}
-
-static void __init cell_spuprop_present(struct device_node *spe,
-				       const char *prop, int early)
-{
-	struct address_prop {
-		unsigned long address;
-		unsigned int len;
-	} __attribute__((packed)) *p;
-	int proplen;
-
-	unsigned long start_pfn, end_pfn, pfn;
-	int node_id;
-
-	p = (void*)get_property(spe, prop, &proplen);
-	WARN_ON(proplen != sizeof (*p));
-
-	node_id = find_spu_node_id(spe);
-
-	start_pfn = p->address >> PAGE_SHIFT;
-	end_pfn = (p->address + p->len + PAGE_SIZE - 1) >> PAGE_SHIFT;
-
-	/* We need to call memory_present *before* the call to sparse_init,
-	   but we can initialize the page structs only *after* that call.
-	   Thus, we're being called twice. */
-	if (early)
-		memory_present(node_id, start_pfn, end_pfn);
-	else {
-		/* As the pages backing SPU LS and I/O are outside the range
-		   of regular memory, their page structs were not initialized
-		   by free_area_init. Do it here instead. */
-		for (pfn = start_pfn; pfn < end_pfn; pfn++) {
-			struct page *page = pfn_to_page(pfn);
-			set_page_links(page, ZONE_DMA, node_id, pfn);
-			init_page_count(page);
-			reset_page_mapcount(page);
-			SetPageReserved(page);
-			INIT_LIST_HEAD(&page->lru);
-		}
-	}
-}
-
-static void __init cell_spumem_init(int early)
-{
-	struct device_node *node;
-	for (node = of_find_node_by_type(NULL, "spe");
-			node; node = of_find_node_by_type(node, "spe")) {
-		cell_spuprop_present(node, "local-store", early);
-		cell_spuprop_present(node, "problem", early);
-		cell_spuprop_present(node, "priv1", early);
-		cell_spuprop_present(node, "priv2", early);
-	}
-}
-#else
-static void __init cell_spumem_init(int early)
-{
-}
-#endif
-
 static void cell_progress(char *s, unsigned short hex)
 {
 	printk("*** %04x : %s\n", hex, s ? s : "");
@@ -172,8 +104,6 @@
 #endif
 
 	mmio_nvram_init();
-
-	cell_spumem_init(0);
 }
 
 /*
@@ -189,8 +119,6 @@
 
 	ppc64_interrupt_controller = IC_CELL_PIC;
 
-	cell_spumem_init(1);
-
 	DBG(" <- cell_init_early()\n");
 }
 
Index: linus-2.6/arch/powerpc/platforms/cell/Kconfig
===================================================================
--- linus-2.6.orig/arch/powerpc/platforms/cell/Kconfig	2006-04-29 02:51:29.000000000 +0200
+++ linus-2.6/arch/powerpc/platforms/cell/Kconfig	2006-04-29 02:51:33.000000000 +0200
@@ -12,7 +12,8 @@
 
 config SPUFS_MMAP
 	bool
-	depends on SPU_FS && SPARSEMEM && !PPC_64K_PAGES
+	depends on SPU_FS && SPARSEMEM
+	select MEMORY_HOTPLUG
 	default y
 
 endmenu
Index: linus-2.6/arch/powerpc/platforms/cell/spu_base.c
===================================================================
--- linus-2.6.orig/arch/powerpc/platforms/cell/spu_base.c	2006-04-29 02:51:29.000000000 +0200
+++ linus-2.6/arch/powerpc/platforms/cell/spu_base.c	2006-04-29 02:51:33.000000000 +0200
@@ -516,6 +516,56 @@
 }
 EXPORT_SYMBOL_GPL(spu_irq_setaffinity);
 
+static int __init find_spu_node_id(struct device_node *spe)
+{
+	unsigned int *id;
+	struct device_node *cpu;
+	cpu = spe->parent->parent;
+	id = (unsigned int *)get_property(cpu, "node-id", NULL);
+	return id ? *id : 0;
+}
+
+static int __init cell_spuprop_present(struct device_node *spe,
+					const char *prop)
+{
+	static DEFINE_MUTEX(add_spumem_mutex);
+
+	struct address_prop {
+		unsigned long address;
+		unsigned int len;
+	} __attribute__((packed)) *p;
+	int proplen;
+
+	unsigned long start_pfn, nr_pages;
+	int node_id;
+	struct pglist_data *pgdata;
+	struct zone *zone;
+	int ret;
+
+	p = (void*)get_property(spe, prop, &proplen);
+	WARN_ON(proplen != sizeof (*p));
+
+	start_pfn = p->address >> PAGE_SHIFT;
+	nr_pages = ((unsigned long)p->len + PAGE_SIZE - 1) >> PAGE_SHIFT;
+
+	/*
+	 * XXX need to get the correct NUMA node in here. This may
+	 * be different from the spe::node_id property, e.g. when
+	 * the host firmware is not NUMA aware.
+	 */
+	node_id = 0;
+
+	pgdata = NODE_DATA(node_id);
+	zone = pgdata->node_zones;
+
+	/* XXX rethink locking here */
+	mutex_lock(&add_spumem_mutex);
+	ret = __add_pages(zone, start_pfn, nr_pages);
+	mutex_unlock(&add_spumem_mutex);
+
+	return ret;
+}
+
 static void __iomem * __init map_spe_prop(struct device_node *n,
 						 const char *name)
 {
@@ -526,6 +576,8 @@
 
 	void *p;
 	int proplen;
+	void* ret = NULL;
+	int err = 0;
 
 	p = get_property(n, name, &proplen);
 	if (proplen != sizeof (struct address_prop))
@@ -533,7 +585,14 @@
 
 	prop = p;
 
-	return ioremap(prop->address, prop->len);
+	err = cell_spuprop_present(n, name);
+	if (err && (err != -EEXIST))
+		goto out;
+
+	ret = ioremap(prop->address, prop->len);
+
+ out:
+	return ret;
 }
 
 static void spu_unmap(struct spu *spu)
@@ -593,17 +652,6 @@
 	return ret;
 }
 
-static int __init find_spu_node_id(struct device_node *spe)
-{
-	unsigned int *id;
-	struct device_node *cpu;
-
-	cpu = spe->parent->parent;
-	id = (unsigned int *)get_property(cpu, "node-id", NULL);
-
-	return id ? *id : 0;
-}
-
 static int __init create_spu(struct device_node *spe)
 {
 	struct spu *spu;
Index: linus-2.6/mm/memory_hotplug.c
===================================================================
--- linus-2.6.orig/mm/memory_hotplug.c	2006-04-29 02:51:29.000000000 +0200
+++ linus-2.6/mm/memory_hotplug.c	2006-04-29 02:51:33.000000000 +0200
@@ -69,12 +69,16 @@
 	for (i = 0; i < nr_pages; i += PAGES_PER_SECTION) {
 		err = __add_section(zone, phys_start_pfn + i);
 
-		if (err)
+		/* We want to keep adding the rest of the
+		 * sections if the first ones already exist
+		 */
+		if (err && (err != -EEXIST))
 			break;
 	}
 
 	return err;
 }
+EXPORT_SYMBOL_GPL(__add_pages);
 
 static void grow_zone_span(struct zone *zone,
 		unsigned long start_pfn, unsigned long end_pfn)

--

[PATCH 2/4] sparsemem interaction with memory add bug fixes

[Arnd Bergmann]

From: Mike Kravetz <mjkravetz@verizon.net>

This patch fixes two bugs with the way sparsemem interacts with memory add.
They are:

- memory leak if memmap for section already exists

- calling alloc_bootmem_node() after boot

These bugs were discovered and a first cut at the fixes were provided by
Arnd Bergmann <arnd@arndb.de> and Joel Schopp <jschopp@us.ibm.com>.

Signed-off-by: Mike Kravetz <kravetz@us.ibm.com>
Signed-off-by: Joel Schopp <jschopp@austin.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
---

 mm/sparse.c |    9 ++++++---
 1 files changed, 6 insertions(+), 3 deletions(-)

diff -puN mm/sparse.c~sparsemem-interaction-with-memory-add-bug-fixes mm/sparse.c
--- 25/mm/sparse.c~sparsemem-interaction-with-memory-add-bug-fixes	Wed Apr 12 14:38:04 2006
+++ 25-akpm/mm/sparse.c	Wed Apr 12 14:38:04 2006
@@ -32,7 +32,10 @@ static struct mem_section *sparse_index_
 	unsigned long array_size = SECTIONS_PER_ROOT *
 				   sizeof(struct mem_section);
 
-	section = alloc_bootmem_node(NODE_DATA(nid), array_size);
+	if (system_state == SYSTEM_RUNNING)
+		section = kmalloc_node(array_size, GFP_KERNEL, nid);
+	else
+		section = alloc_bootmem_node(NODE_DATA(nid), array_size);
 
 	if (section)
 		memset(section, 0, array_size);
@@ -281,9 +284,9 @@ int sparse_add_one_section(struct zone *
 
 	ret = sparse_init_one_section(ms, section_nr, memmap);
 
-	if (ret <= 0)
-		__kfree_section_memmap(memmap, nr_pages);
 out:
 	pgdat_resize_unlock(pgdat, &flags);
+	if (ret <= 0)
+		__kfree_section_memmap(memmap, nr_pages);
 	return ret;
 }
_

--

[PATCH 3/4] powerpc: Allow devices to register with numa topology

[Arnd Bergmann]

From: Jeremy Kerr <jk@ozlabs.org>

Change of_node_to_nid() to traverse the device tree, looking for a numa
id. Cell uses this to assign ids to SPUs, which are children of the CPU
node. Existing users of of_node_to_nid() are altered to use
of_node_to_nid_single(), which doesn't do the traversal.

Export an attach_sysdev_to_node() function, allowing system devices (eg.
SPUs) to link themselves into the numa topology in sysfs.

Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
---
Index: linus-2.6/arch/powerpc/mm/numa.c
===================================================================
--- linus-2.6.orig/arch/powerpc/mm/numa.c	2006-04-29 01:50:25.000000000 +0200
+++ linus-2.6/arch/powerpc/mm/numa.c	2006-04-29 02:02:28.000000000 +0200
@@ -194,7 +194,7 @@
 /* Returns nid in the range [0..MAX_NUMNODES-1], or -1 if no useful numa
  * info is found.
  */
-static int of_node_to_nid(struct device_node *device)
+static int of_node_to_nid_single(struct device_node *device)
 {
 	int nid = -1;
 	unsigned int *tmp;
@@ -216,6 +216,28 @@
 	return nid;
 }
 
+/* Walk the device tree upwards, looking for an associativity id */
+int of_node_to_nid(struct device_node *device)
+{
+	struct device_node *tmp;
+	int nid = -1;
+
+	of_node_get(device);
+	while (device) {
+		nid = of_node_to_nid_single(device);
+		if (nid != -1)
+			break;
+
+	        tmp = device;
+		device = of_get_parent(tmp);
+		of_node_put(tmp);
+	}
+	of_node_put(device);
+
+	return nid;
+}
+EXPORT_SYMBOL_GPL(of_node_to_nid);
+
 /*
  * In theory, the "ibm,associativity" property may contain multiple
  * associativity lists because a resource may be multiply connected
@@ -300,7 +322,7 @@
 		goto out;
 	}
 
-	nid = of_node_to_nid(cpu);
+	nid = of_node_to_nid_single(cpu);
 
 	if (nid < 0 || !node_online(nid))
 		nid = any_online_node(NODE_MASK_ALL);
@@ -393,7 +415,7 @@
 
 		cpu = find_cpu_node(i);
 		BUG_ON(!cpu);
-		nid = of_node_to_nid(cpu);
+		nid = of_node_to_nid_single(cpu);
 		of_node_put(cpu);
 
 		/*
@@ -437,7 +459,7 @@
 		 * have associativity properties.  If none, then
 		 * everything goes to default_nid.
 		 */
-		nid = of_node_to_nid(memory);
+		nid = of_node_to_nid_single(memory);
 		if (nid < 0)
 			nid = default_nid;
 		node_set_online(nid);
@@ -776,7 +798,7 @@
 ha_new_range:
 		start = read_n_cells(n_mem_addr_cells, &memcell_buf);
 		size = read_n_cells(n_mem_size_cells, &memcell_buf);
-		nid = of_node_to_nid(memory);
+		nid = of_node_to_nid_single(memory);
 
 		/* Domains not present at boot default to 0 */
 		if (nid < 0 || !node_online(nid))
Index: linus-2.6/include/asm-powerpc/topology.h
===================================================================
--- linus-2.6.orig/include/asm-powerpc/topology.h	2006-04-29 01:50:25.000000000 +0200
+++ linus-2.6/include/asm-powerpc/topology.h	2006-04-29 02:02:28.000000000 +0200
@@ -4,6 +4,9 @@
 
 #include <linux/config.h>
 
+struct sys_device;
+struct device_node;
+
 #ifdef CONFIG_NUMA
 
 #include <asm/mmzone.h>
@@ -27,6 +30,8 @@
 	return first_cpu(tmp);
 }
 
+int of_node_to_nid(struct device_node *device);
+
 #define pcibus_to_node(node)    (-1)
 #define pcibus_to_cpumask(bus)	(cpu_online_map)
 
@@ -57,10 +62,29 @@
 
 extern void __init dump_numa_cpu_topology(void);
 
+extern int sysfs_add_device_to_node(struct sys_device *dev, int nid);
+extern void sysfs_remove_device_from_node(struct sys_device *dev, int nid);
+
 #else
 
+static inline int of_node_to_nid(struct device_node *device)
+{
+	return 0;
+}
+
 static inline void dump_numa_cpu_topology(void) {}
 
+static inline int sysfs_add_device_to_node(struct sys_device *dev, int nid)
+{
+	return 0;
+}
+
+static inline void sysfs_remove_device_from_node(struct sys_device *dev,
+						int nid)
+{
+}
+
+
 #include <asm-generic/topology.h>
 
 #endif /* CONFIG_NUMA */
Index: linus-2.6/arch/powerpc/kernel/sysfs.c
===================================================================
--- linus-2.6.orig/arch/powerpc/kernel/sysfs.c	2006-04-29 01:50:25.000000000 +0200
+++ linus-2.6/arch/powerpc/kernel/sysfs.c	2006-04-29 02:02:28.000000000 +0200
@@ -322,13 +322,31 @@
 		}
 	}
 }
+
+int sysfs_add_device_to_node(struct sys_device *dev, int nid)
+{
+	struct node *node = &node_devices[nid];
+	return sysfs_create_link(&node->sysdev.kobj, &dev->kobj,
+			kobject_name(&dev->kobj));
+}
+
+void sysfs_remove_device_from_node(struct sys_device *dev, int nid)
+{
+	struct node *node = &node_devices[nid];
+	sysfs_remove_link(&node->sysdev.kobj, kobject_name(&dev->kobj));
+}
+
 #else
 static void register_nodes(void)
 {
 	return;
 }
+
 #endif
 
+EXPORT_SYMBOL_GPL(sysfs_add_device_to_node);
+EXPORT_SYMBOL_GPL(sysfs_remove_device_from_node);
+
 /* Only valid if CPU is present. */
 static ssize_t show_physical_id(struct sys_device *dev, char *buf)
 {

--

[PATCH 4/4] powerpc: cell: Add numa id to struct spu

[Arnd Bergmann]

From: Jeremy Kerr <jk@ozlabs.org>

Add an nid member to the spu structure, and store the numa id of the spu
there on creation.

Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
---
Index: linus-2.6/arch/powerpc/platforms/cell/spu_base.c
===================================================================
--- linus-2.6.orig/arch/powerpc/platforms/cell/spu_base.c	2006-04-29 02:51:33.000000000 +0200
+++ linus-2.6/arch/powerpc/platforms/cell/spu_base.c	2006-04-29 02:56:58.000000000 +0200
@@ -525,8 +525,8 @@
 	return id ? *id : 0;
 }
 
-static int __init cell_spuprop_present(struct device_node *spe,
-					const char *prop)
+static int __init cell_spuprop_present(struct spu *spu, struct device_node *spe,
+		const char *prop)
 {
 	static DEFINE_MUTEX(add_spumem_mutex);
 
@@ -537,7 +537,6 @@
 	int proplen;
 
 	unsigned long start_pfn, nr_pages;
-	int node_id;
 	struct pglist_data *pgdata;
 	struct zone *zone;
 	int ret;
@@ -548,14 +547,7 @@
 	start_pfn = p->address >> PAGE_SHIFT;
 	nr_pages = ((unsigned long)p->len + PAGE_SIZE - 1) >> PAGE_SHIFT;
 
-	/*
-	 * XXX need to get the correct NUMA node in here. This may
-	 * be different from the spe::node_id property, e.g. when
-	 * the host firmware is not NUMA aware.
-	 */
-	node_id = 0;
-
-	pgdata = NODE_DATA(node_id);
+	pgdata = NODE_DATA(spu->nid);
 	zone = pgdata->node_zones;
 
 	/* XXX rethink locking here */
@@ -566,8 +558,8 @@
 	return ret;
 }
 
-static void __iomem * __init map_spe_prop(struct device_node *n,
-						 const char *name)
+static void __iomem * __init map_spe_prop(struct spu *spu,
+		struct device_node *n, const char *name)
 {
 	struct address_prop {
 		unsigned long address;
@@ -585,7 +577,7 @@
 
 	prop = p;
 
-	err = cell_spuprop_present(n, name);
+	err = cell_spuprop_present(spu, n, name);
 	if (err && (err != -EEXIST))
 		goto out;
 
@@ -603,44 +595,45 @@
 	iounmap((u8 __iomem *)spu->local_store);
 }
 
-static int __init spu_map_device(struct spu *spu, struct device_node *spe)
+static int __init spu_map_device(struct spu *spu, struct device_node *node)
 {
 	char *prop;
 	int ret;
 
 	ret = -ENODEV;
-	prop = get_property(spe, "isrc", NULL);
+	prop = get_property(node, "isrc", NULL);
 	if (!prop)
 		goto out;
 	spu->isrc = *(unsigned int *)prop;
 
-	spu->name = get_property(spe, "name", NULL);
+	spu->name = get_property(node, "name", NULL);
 	if (!spu->name)
 		goto out;
 
-	prop = get_property(spe, "local-store", NULL);
+	prop = get_property(node, "local-store", NULL);
 	if (!prop)
 		goto out;
 	spu->local_store_phys = *(unsigned long *)prop;
 
 	/* we use local store as ram, not io memory */
-	spu->local_store = (void __force *)map_spe_prop(spe, "local-store");
+	spu->local_store = (void __force *)
+		map_spe_prop(spu, node, "local-store");
 	if (!spu->local_store)
 		goto out;
 
-	prop = get_property(spe, "problem", NULL);
+	prop = get_property(node, "problem", NULL);
 	if (!prop)
 		goto out_unmap;
 	spu->problem_phys = *(unsigned long *)prop;
 
-	spu->problem= map_spe_prop(spe, "problem");
+	spu->problem= map_spe_prop(spu, node, "problem");
 	if (!spu->problem)
 		goto out_unmap;
 
-	spu->priv1= map_spe_prop(spe, "priv1");
+	spu->priv1= map_spe_prop(spu, node, "priv1");
 	/* priv1 is not available on a hypervisor */
 
-	spu->priv2= map_spe_prop(spe, "priv2");
+	spu->priv2= map_spe_prop(spu, node, "priv2");
 	if (!spu->priv2)
 		goto out_unmap;
 	ret = 0;
@@ -668,6 +661,10 @@
 		goto out_free;
 
 	spu->node = find_spu_node_id(spe);
+	spu->nid = of_node_to_nid(spe);
+	if (spu->nid == -1)
+		spu->nid = 0;
+
 	spu->stop_code = 0;
 	spu->slb_replace = 0;
 	spu->mm = NULL;
Index: linus-2.6/include/asm-powerpc/spu.h
===================================================================
--- linus-2.6.orig/include/asm-powerpc/spu.h	2006-04-29 02:51:29.000000000 +0200
+++ linus-2.6/include/asm-powerpc/spu.h	2006-04-29 02:56:58.000000000 +0200
@@ -117,6 +117,7 @@
 	struct list_head list;
 	struct list_head sched_list;
 	int number;
+	int nid;
 	u32 isrc;
 	u32 node;
 	u64 flags;

--

Re: [PATCH 0/4] NUMA support for spufs

[Paul Mackerras]

Arnd Bergmann writes:

> The current version of spufs breaks upon boot when NUMA support is enabled.
> I'd like to fix that before 2.6.17 so we can use the same kernel image
> on Cell that we use on other powerpc systems using NUMA.

Andrew,

If 2/4 of this series looks OK to you to go into 2.6.17, let me know
and I'll push the others to Linus (or you can if you prefer).

Thanks,
Paul.

Re: [PATCH 0/4] NUMA support for spufs

[Andrew Morton]

Paul Mackerras <paulus@samba.org> wrote:
>
> Arnd Bergmann writes:
> 
> > The current version of spufs breaks upon boot when NUMA support is enabled.
> > I'd like to fix that before 2.6.17 so we can use the same kernel image
> > on Cell that we use on other powerpc systems using NUMA.
> 
> Andrew,
> 
> If 2/4 of this series looks OK to you to go into 2.6.17,

OK by me.  1/4 touches __add_pages(), which it undergoing quite a bit of
churn at present, but we'll work it out.

> let me know
> and I'll push the others to Linus (or you can if you prefer).

I'll queue all four for the next batch (within the next 48 hours).  If you
merge them first then no probs.