[SPARC64]: Add NUMA support.
authorDavid S. Miller <davem@davemloft.net>
Wed, 23 Apr 2008 12:40:25 +0000 (05:40 -0700)
committerDavid S. Miller <davem@davemloft.net>
Thu, 24 Apr 2008 06:32:17 +0000 (23:32 -0700)
Currently there is only code to parse NUMA attributes on
sun4v/niagara systems, but later on we will add such parsing
for older systems.

Signed-off-by: David S. Miller <davem@davemloft.net>
Makefile
arch/sparc64/Kconfig
arch/sparc64/defconfig
arch/sparc64/kernel/sysfs.c
arch/sparc64/mm/init.c
include/asm-sparc64/mmzone.h [new file with mode: 0644]
include/asm-sparc64/topology.h

index 3dbc826bb8e6d73c175d772a497f17accd61c23e..d35c5246fce5b338f1781ba2b2ae4c0224ad902e 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -1,7 +1,7 @@
 VERSION = 2
 PATCHLEVEL = 6
 SUBLEVEL = 25
-EXTRAVERSION =
+EXTRAVERSION = -numa
 NAME = Funky Weasel is Jiggy wit it
 
 # *DOCUMENTATION*
index df3eacb5ca15dcaf1f716148d1dcdcc1e7c33195..8acc5cc386212c4cba07ff12b93cd2fabf0a57b6 100644 (file)
@@ -250,6 +250,26 @@ endchoice
 
 endmenu
 
+config NUMA
+       bool "NUMA support"
+
+config NODES_SHIFT
+       int
+       default "4"
+       depends on NEED_MULTIPLE_NODES
+
+# Some NUMA nodes have memory ranges that span
+# other nodes.  Even though a pfn is valid and
+# between a node's start and end pfns, it may not
+# reside on that node.  See memmap_init_zone()
+# for details.
+config NODES_SPAN_OTHER_NODES
+       def_bool y
+       depends on NEED_MULTIPLE_NODES
+
+config ARCH_POPULATES_NODE_MAP
+       def_bool y
+
 config ARCH_SELECT_MEMORY_MODEL
        def_bool y
 
index e1835868ad36f5e30b383a31c5d0f26193dc1785..92f79680f70d92da247c2e80130db38af86363ad 100644 (file)
@@ -1,7 +1,7 @@
 #
 # Automatically generated make config: don't edit
-# Linux kernel version: 2.6.25
-# Sun Apr 20 01:33:21 2008
+# Linux kernel version: 2.6.25-numa
+# Wed Apr 23 04:49:08 2008
 #
 CONFIG_SPARC=y
 CONFIG_SPARC64=y
@@ -152,6 +152,8 @@ CONFIG_GENERIC_CALIBRATE_DELAY=y
 CONFIG_HUGETLB_PAGE_SIZE_4MB=y
 # CONFIG_HUGETLB_PAGE_SIZE_512K is not set
 # CONFIG_HUGETLB_PAGE_SIZE_64K is not set
+# CONFIG_NUMA is not set
+CONFIG_ARCH_POPULATES_NODE_MAP=y
 CONFIG_ARCH_SELECT_MEMORY_MODEL=y
 CONFIG_ARCH_SPARSEMEM_ENABLE=y
 CONFIG_ARCH_SPARSEMEM_DEFAULT=y
@@ -787,7 +789,6 @@ CONFIG_I2C_ALGOBIT=y
 # CONFIG_SENSORS_PCF8574 is not set
 # CONFIG_PCF8575 is not set
 # CONFIG_SENSORS_PCF8591 is not set
-# CONFIG_TPS65010 is not set
 # CONFIG_SENSORS_MAX6875 is not set
 # CONFIG_SENSORS_TSL2550 is not set
 # CONFIG_I2C_DEBUG_CORE is not set
@@ -869,6 +870,7 @@ CONFIG_SSB_POSSIBLE=y
 # Multifunction device drivers
 #
 # CONFIG_MFD_SM501 is not set
+# CONFIG_HTC_PASIC3 is not set
 
 #
 # Multimedia devices
@@ -1219,10 +1221,6 @@ CONFIG_USB_STORAGE=m
 # CONFIG_NEW_LEDS is not set
 # CONFIG_INFINIBAND is not set
 # CONFIG_RTC_CLASS is not set
-
-#
-# Userspace I/O
-#
 # CONFIG_UIO is not set
 
 #
@@ -1399,6 +1397,7 @@ CONFIG_SCHEDSTATS=y
 CONFIG_DEBUG_BUGVERBOSE=y
 # CONFIG_DEBUG_INFO is not set
 # CONFIG_DEBUG_VM is not set
+# CONFIG_DEBUG_WRITECOUNT is not set
 # CONFIG_DEBUG_LIST is not set
 # CONFIG_DEBUG_SG is not set
 # CONFIG_BOOT_PRINTK_DELAY is not set
@@ -1425,53 +1424,82 @@ CONFIG_ASYNC_CORE=m
 CONFIG_ASYNC_MEMCPY=m
 CONFIG_ASYNC_XOR=m
 CONFIG_CRYPTO=y
+
+#
+# Crypto core or helper
+#
 CONFIG_CRYPTO_ALGAPI=y
 CONFIG_CRYPTO_AEAD=y
 CONFIG_CRYPTO_BLKCIPHER=y
-# CONFIG_CRYPTO_SEQIV is not set
 CONFIG_CRYPTO_HASH=y
 CONFIG_CRYPTO_MANAGER=y
+CONFIG_CRYPTO_GF128MUL=m
+CONFIG_CRYPTO_NULL=m
+# CONFIG_CRYPTO_CRYPTD is not set
+CONFIG_CRYPTO_AUTHENC=y
+CONFIG_CRYPTO_TEST=m
+
+#
+# Authenticated Encryption with Associated Data
+#
+# CONFIG_CRYPTO_CCM is not set
+# CONFIG_CRYPTO_GCM is not set
+# CONFIG_CRYPTO_SEQIV is not set
+
+#
+# Block modes
+#
+CONFIG_CRYPTO_CBC=y
+# CONFIG_CRYPTO_CTR is not set
+# CONFIG_CRYPTO_CTS is not set
+CONFIG_CRYPTO_ECB=m
+CONFIG_CRYPTO_LRW=m
+CONFIG_CRYPTO_PCBC=m
+CONFIG_CRYPTO_XTS=m
+
+#
+# Hash modes
+#
 CONFIG_CRYPTO_HMAC=y
 CONFIG_CRYPTO_XCBC=y
-CONFIG_CRYPTO_NULL=m
+
+#
+# Digest
+#
+CONFIG_CRYPTO_CRC32C=m
 CONFIG_CRYPTO_MD4=y
 CONFIG_CRYPTO_MD5=y
+CONFIG_CRYPTO_MICHAEL_MIC=m
 CONFIG_CRYPTO_SHA1=y
 CONFIG_CRYPTO_SHA256=m
 CONFIG_CRYPTO_SHA512=m
-CONFIG_CRYPTO_WP512=m
 CONFIG_CRYPTO_TGR192=m
-CONFIG_CRYPTO_GF128MUL=m
-CONFIG_CRYPTO_ECB=m
-CONFIG_CRYPTO_CBC=y
-CONFIG_CRYPTO_PCBC=m
-CONFIG_CRYPTO_LRW=m
-CONFIG_CRYPTO_XTS=m
-# CONFIG_CRYPTO_CTR is not set
-# CONFIG_CRYPTO_GCM is not set
-# CONFIG_CRYPTO_CCM is not set
-# CONFIG_CRYPTO_CRYPTD is not set
-CONFIG_CRYPTO_DES=y
-CONFIG_CRYPTO_FCRYPT=m
-CONFIG_CRYPTO_BLOWFISH=m
-CONFIG_CRYPTO_TWOFISH=m
-CONFIG_CRYPTO_TWOFISH_COMMON=m
-CONFIG_CRYPTO_SERPENT=m
+CONFIG_CRYPTO_WP512=m
+
+#
+# Ciphers
+#
 CONFIG_CRYPTO_AES=m
+CONFIG_CRYPTO_ANUBIS=m
+CONFIG_CRYPTO_ARC4=m
+CONFIG_CRYPTO_BLOWFISH=m
+CONFIG_CRYPTO_CAMELLIA=m
 CONFIG_CRYPTO_CAST5=m
 CONFIG_CRYPTO_CAST6=m
-CONFIG_CRYPTO_TEA=m
-CONFIG_CRYPTO_ARC4=m
+CONFIG_CRYPTO_DES=y
+CONFIG_CRYPTO_FCRYPT=m
 CONFIG_CRYPTO_KHAZAD=m
-CONFIG_CRYPTO_ANUBIS=m
-CONFIG_CRYPTO_SEED=m
 # CONFIG_CRYPTO_SALSA20 is not set
+CONFIG_CRYPTO_SEED=m
+CONFIG_CRYPTO_SERPENT=m
+CONFIG_CRYPTO_TEA=m
+CONFIG_CRYPTO_TWOFISH=m
+CONFIG_CRYPTO_TWOFISH_COMMON=m
+
+#
+# Compression
+#
 CONFIG_CRYPTO_DEFLATE=y
-CONFIG_CRYPTO_MICHAEL_MIC=m
-CONFIG_CRYPTO_CRC32C=m
-CONFIG_CRYPTO_CAMELLIA=m
-CONFIG_CRYPTO_TEST=m
-CONFIG_CRYPTO_AUTHENC=y
 # CONFIG_CRYPTO_LZO is not set
 CONFIG_CRYPTO_HW=y
 # CONFIG_CRYPTO_DEV_HIFN_795X is not set
@@ -1492,3 +1520,4 @@ CONFIG_PLIST=y
 CONFIG_HAS_IOMEM=y
 CONFIG_HAS_IOPORT=y
 CONFIG_HAS_DMA=y
+CONFIG_HAVE_LMB=y
index 52816c7be0b96f8015c997c268dd5a866269979e..e885034a6b73a865c049fa5ca62ca5dab3a88fc1 100644 (file)
@@ -273,10 +273,22 @@ static void __init check_mmu_stats(void)
                mmu_stats_supported = 1;
 }
 
+static void register_nodes(void)
+{
+#ifdef CONFIG_NUMA
+       int i;
+
+       for (i = 0; i < MAX_NUMNODES; i++)
+               register_one_node(i);
+#endif
+}
+
 static int __init topology_init(void)
 {
        int cpu;
 
+       register_nodes();
+
        check_mmu_stats();
 
        register_cpu_notifier(&sysfs_cpu_nb);
index 8e0e86787127f2f5ec4566ba5a29f0b7fe5ba615..177d8aaeec42e2d147b92617dd308afa9c2baec1 100644 (file)
@@ -25,6 +25,7 @@
 #include <linux/sort.h>
 #include <linux/percpu.h>
 #include <linux/lmb.h>
+#include <linux/mmzone.h>
 
 #include <asm/head.h>
 #include <asm/system.h>
@@ -73,9 +74,7 @@ extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
 #define MAX_BANKS      32
 
 static struct linux_prom64_registers pavail[MAX_BANKS] __initdata;
-static struct linux_prom64_registers pavail_rescan[MAX_BANKS] __initdata;
 static int pavail_ents __initdata;
-static int pavail_rescan_ents __initdata;
 
 static int cmp_p64(const void *a, const void *b)
 {
@@ -716,19 +715,28 @@ out:
                smp_new_mmu_context_version();
 }
 
-/* Find a free area for the bootmem map, avoiding the kernel image
- * and the initial ramdisk.
- */
-static unsigned long __init choose_bootmap_pfn(unsigned long start_pfn,
-                                              unsigned long end_pfn)
+static int numa_enabled = 1;
+static int numa_debug;
+
+static int __init early_numa(char *p)
 {
-       unsigned long bootmap_size;
+       if (!p)
+               return 0;
+
+       if (strstr(p, "off"))
+               numa_enabled = 0;
 
-       bootmap_size = bootmem_bootmap_pages(end_pfn - start_pfn);
-       bootmap_size <<= PAGE_SHIFT;
+       if (strstr(p, "debug"))
+               numa_debug = 1;
 
-       return lmb_alloc(bootmap_size, PAGE_SIZE) >> PAGE_SHIFT;
+       return 0;
 }
+early_param("numa", early_numa);
+
+#define numadbg(f, a...) \
+do {   if (numa_debug) \
+               printk(KERN_INFO f, ## a); \
+} while (0)
 
 static void __init find_ramdisk(unsigned long phys_base)
 {
@@ -755,6 +763,9 @@ static void __init find_ramdisk(unsigned long phys_base)
                ramdisk_image -= KERNBASE;
                ramdisk_image += phys_base;
 
+               numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
+                       ramdisk_image, sparc_ramdisk_size);
+
                initrd_start = ramdisk_image;
                initrd_end = ramdisk_image + sparc_ramdisk_size;
 
@@ -763,60 +774,625 @@ static void __init find_ramdisk(unsigned long phys_base)
 #endif
 }
 
-/* About pages_avail, this is the value we will use to calculate
- * the zholes_size[] argument given to free_area_init_node().  The
- * page allocator uses this to calculate nr_kernel_pages,
- * nr_all_pages and zone->present_pages.  On NUMA it is used
- * to calculate zone->min_unmapped_pages and zone->min_slab_pages.
- *
- * So this number should really be set to what the page allocator
- * actually ends up with.  This means:
- * 1) It should include bootmem map pages, we'll release those.
- * 2) It should not include the kernel image, except for the
- *    __init sections which we will also release.
- * 3) It should include the initrd image, since we'll release
- *    that too.
+struct node_mem_mask {
+       unsigned long mask;
+       unsigned long val;
+       unsigned long bootmem_paddr;
+};
+static struct node_mem_mask node_masks[MAX_NUMNODES];
+static int num_node_masks;
+
+int numa_cpu_lookup_table[NR_CPUS];
+cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
+
+#ifdef CONFIG_NEED_MULTIPLE_NODES
+static bootmem_data_t plat_node_bdata[MAX_NUMNODES];
+
+struct mdesc_mblock {
+       u64     base;
+       u64     size;
+       u64     offset; /* RA-to-PA */
+};
+static struct mdesc_mblock *mblocks;
+static int num_mblocks;
+
+static unsigned long ra_to_pa(unsigned long addr)
+{
+       int i;
+
+       for (i = 0; i < num_mblocks; i++) {
+               struct mdesc_mblock *m = &mblocks[i];
+
+               if (addr >= m->base &&
+                   addr < (m->base + m->size)) {
+                       addr += m->offset;
+                       break;
+               }
+       }
+       return addr;
+}
+
+static int find_node(unsigned long addr)
+{
+       int i;
+
+       addr = ra_to_pa(addr);
+       for (i = 0; i < num_node_masks; i++) {
+               struct node_mem_mask *p = &node_masks[i];
+
+               if ((addr & p->mask) == p->val)
+                       return i;
+       }
+       return -1;
+}
+
+static unsigned long nid_range(unsigned long start, unsigned long end,
+                              int *nid)
+{
+       *nid = find_node(start);
+       start += PAGE_SIZE;
+       while (start < end) {
+               int n = find_node(start);
+
+               if (n != *nid)
+                       break;
+               start += PAGE_SIZE;
+       }
+
+       return start;
+}
+#else
+static unsigned long nid_range(unsigned long start, unsigned long end,
+                              int *nid)
+{
+       *nid = 0;
+       return end;
+}
+#endif
+
+/* This must be invoked after performing all of the necessary
+ * add_active_range() calls for 'nid'.  We need to be able to get
+ * correct data from get_pfn_range_for_nid().
  */
-static unsigned long __init bootmem_init(unsigned long *pages_avail,
-                                        unsigned long phys_base)
+static void __init allocate_node_data(int nid)
+{
+       unsigned long paddr, num_pages, start_pfn, end_pfn;
+       struct pglist_data *p;
+
+#ifdef CONFIG_NEED_MULTIPLE_NODES
+       paddr = lmb_alloc_nid(sizeof(struct pglist_data),
+                             SMP_CACHE_BYTES, nid, nid_range);
+       if (!paddr) {
+               prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
+               prom_halt();
+       }
+       NODE_DATA(nid) = __va(paddr);
+       memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
+
+       NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
+#endif
+
+       p = NODE_DATA(nid);
+
+       get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
+       p->node_start_pfn = start_pfn;
+       p->node_spanned_pages = end_pfn - start_pfn;
+
+       if (p->node_spanned_pages) {
+               num_pages = bootmem_bootmap_pages(p->node_spanned_pages);
+
+               paddr = lmb_alloc_nid(num_pages << PAGE_SHIFT, PAGE_SIZE, nid,
+                                     nid_range);
+               if (!paddr) {
+                       prom_printf("Cannot allocate bootmap for nid[%d]\n",
+                                 nid);
+                       prom_halt();
+               }
+               node_masks[nid].bootmem_paddr = paddr;
+       }
+}
+
+static void init_node_masks_nonnuma(void)
 {
-       unsigned long end_pfn;
        int i;
 
-       *pages_avail = lmb_phys_mem_size() >> PAGE_SHIFT;
-       end_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
+       numadbg("Initializing tables for non-numa.\n");
 
-       /* Initialize the boot-time allocator. */
-       max_pfn = max_low_pfn = end_pfn;
-       min_low_pfn = (phys_base >> PAGE_SHIFT);
+       node_masks[0].mask = node_masks[0].val = 0;
+       num_node_masks = 1;
 
-       init_bootmem_node(NODE_DATA(0),
-                         choose_bootmap_pfn(min_low_pfn, end_pfn),
-                         min_low_pfn, end_pfn);
+       for (i = 0; i < NR_CPUS; i++)
+               numa_cpu_lookup_table[i] = 0;
 
-       /* Now register the available physical memory with the
-        * allocator.
-        */
-       for (i = 0; i < lmb.memory.cnt; i++)
-               free_bootmem(lmb.memory.region[i].base,
-                            lmb_size_bytes(&lmb.memory, i));
+       numa_cpumask_lookup_table[0] = CPU_MASK_ALL;
+}
+
+#ifdef CONFIG_NEED_MULTIPLE_NODES
+struct pglist_data *node_data[MAX_NUMNODES];
+
+EXPORT_SYMBOL(numa_cpu_lookup_table);
+EXPORT_SYMBOL(numa_cpumask_lookup_table);
+EXPORT_SYMBOL(node_data);
+
+struct mdesc_mlgroup {
+       u64     node;
+       u64     latency;
+       u64     match;
+       u64     mask;
+};
+static struct mdesc_mlgroup *mlgroups;
+static int num_mlgroups;
+
+static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
+                                  u32 cfg_handle)
+{
+       u64 arc;
+
+       mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
+               u64 target = mdesc_arc_target(md, arc);
+               const u64 *val;
+
+               val = mdesc_get_property(md, target,
+                                        "cfg-handle", NULL);
+               if (val && *val == cfg_handle)
+                       return 0;
+       }
+       return -ENODEV;
+}
+
+static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
+                                   u32 cfg_handle)
+{
+       u64 arc, candidate, best_latency = ~(u64)0;
+
+       candidate = MDESC_NODE_NULL;
+       mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
+               u64 target = mdesc_arc_target(md, arc);
+               const char *name = mdesc_node_name(md, target);
+               const u64 *val;
+
+               if (strcmp(name, "pio-latency-group"))
+                       continue;
+
+               val = mdesc_get_property(md, target, "latency", NULL);
+               if (!val)
+                       continue;
+
+               if (*val < best_latency) {
+                       candidate = target;
+                       best_latency = *val;
+               }
+       }
+
+       if (candidate == MDESC_NODE_NULL)
+               return -ENODEV;
+
+       return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
+}
+
+int of_node_to_nid(struct device_node *dp)
+{
+       const struct linux_prom64_registers *regs;
+       struct mdesc_handle *md;
+       u32 cfg_handle;
+       int count, nid;
+       u64 grp;
+
+       if (!mlgroups)
+               return -1;
+
+       regs = of_get_property(dp, "reg", NULL);
+       if (!regs)
+               return -1;
+
+       cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
+
+       md = mdesc_grab();
+
+       count = 0;
+       nid = -1;
+       mdesc_for_each_node_by_name(md, grp, "group") {
+               if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
+                       nid = count;
+                       break;
+               }
+               count++;
+       }
+
+       mdesc_release(md);
+
+       return nid;
+}
+
+static void add_node_ranges(void)
+{
+       int i;
+
+       for (i = 0; i < lmb.memory.cnt; i++) {
+               unsigned long size = lmb_size_bytes(&lmb.memory, i);
+               unsigned long start, end;
+
+               start = lmb.memory.region[i].base;
+               end = start + size;
+               while (start < end) {
+                       unsigned long this_end;
+                       int nid;
+
+                       this_end = nid_range(start, end, &nid);
+
+                       numadbg("Adding active range nid[%d] "
+                               "start[%lx] end[%lx]\n",
+                               nid, start, this_end);
+
+                       add_active_range(nid,
+                                        start >> PAGE_SHIFT,
+                                        this_end >> PAGE_SHIFT);
+
+                       start = this_end;
+               }
+       }
+}
 
-       for (i = 0; i < lmb.reserved.cnt; i++)
-               reserve_bootmem(lmb.reserved.region[i].base,
-                               lmb_size_bytes(&lmb.reserved, i),
-                               BOOTMEM_DEFAULT);
+static int __init grab_mlgroups(struct mdesc_handle *md)
+{
+       unsigned long paddr;
+       int count = 0;
+       u64 node;
+
+       mdesc_for_each_node_by_name(md, node, "memory-latency-group")
+               count++;
+       if (!count)
+               return -ENOENT;
+
+       paddr = lmb_alloc(count * sizeof(struct mdesc_mlgroup),
+                         SMP_CACHE_BYTES);
+       if (!paddr)
+               return -ENOMEM;
+
+       mlgroups = __va(paddr);
+       num_mlgroups = count;
+
+       count = 0;
+       mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
+               struct mdesc_mlgroup *m = &mlgroups[count++];
+               const u64 *val;
+
+               m->node = node;
+
+               val = mdesc_get_property(md, node, "latency", NULL);
+               m->latency = *val;
+               val = mdesc_get_property(md, node, "address-match", NULL);
+               m->match = *val;
+               val = mdesc_get_property(md, node, "address-mask", NULL);
+               m->mask = *val;
+
+               numadbg("MLGROUP[%d]: node[%lx] latency[%lx] "
+                       "match[%lx] mask[%lx]\n",
+                       count - 1, m->node, m->latency, m->match, m->mask);
+       }
 
-       *pages_avail -= PAGE_ALIGN(kern_size) >> PAGE_SHIFT;
+       return 0;
+}
 
-       for (i = 0; i < lmb.memory.cnt; ++i) {
-               unsigned long start_pfn, end_pfn, pages;
+static int __init grab_mblocks(struct mdesc_handle *md)
+{
+       unsigned long paddr;
+       int count = 0;
+       u64 node;
+
+       mdesc_for_each_node_by_name(md, node, "mblock")
+               count++;
+       if (!count)
+               return -ENOENT;
+
+       paddr = lmb_alloc(count * sizeof(struct mdesc_mblock),
+                         SMP_CACHE_BYTES);
+       if (!paddr)
+               return -ENOMEM;
+
+       mblocks = __va(paddr);
+       num_mblocks = count;
+
+       count = 0;
+       mdesc_for_each_node_by_name(md, node, "mblock") {
+               struct mdesc_mblock *m = &mblocks[count++];
+               const u64 *val;
+
+               val = mdesc_get_property(md, node, "base", NULL);
+               m->base = *val;
+               val = mdesc_get_property(md, node, "size", NULL);
+               m->size = *val;
+               val = mdesc_get_property(md, node,
+                                        "address-congruence-offset", NULL);
+               m->offset = *val;
+
+               numadbg("MBLOCK[%d]: base[%lx] size[%lx] offset[%lx]\n",
+                       count - 1, m->base, m->size, m->offset);
+       }
+
+       return 0;
+}
+
+static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
+                                              u64 grp, cpumask_t *mask)
+{
+       u64 arc;
+
+       cpus_clear(*mask);
+
+       mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
+               u64 target = mdesc_arc_target(md, arc);
+               const char *name = mdesc_node_name(md, target);
+               const u64 *id;
+
+               if (strcmp(name, "cpu"))
+                       continue;
+               id = mdesc_get_property(md, target, "id", NULL);
+               if (*id < NR_CPUS)
+                       cpu_set(*id, *mask);
+       }
+}
+
+static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
+{
+       int i;
+
+       for (i = 0; i < num_mlgroups; i++) {
+               struct mdesc_mlgroup *m = &mlgroups[i];
+               if (m->node == node)
+                       return m;
+       }
+       return NULL;
+}
+
+static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
+                                     int index)
+{
+       struct mdesc_mlgroup *candidate = NULL;
+       u64 arc, best_latency = ~(u64)0;
+       struct node_mem_mask *n;
+
+       mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
+               u64 target = mdesc_arc_target(md, arc);
+               struct mdesc_mlgroup *m = find_mlgroup(target);
+               if (!m)
+                       continue;
+               if (m->latency < best_latency) {
+                       candidate = m;
+                       best_latency = m->latency;
+               }
+       }
+       if (!candidate)
+               return -ENOENT;
+
+       if (num_node_masks != index) {
+               printk(KERN_ERR "Inconsistent NUMA state, "
+                      "index[%d] != num_node_masks[%d]\n",
+                      index, num_node_masks);
+               return -EINVAL;
+       }
+
+       n = &node_masks[num_node_masks++];
+
+       n->mask = candidate->mask;
+       n->val = candidate->match;
+
+       numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%lx])\n",
+               index, n->mask, n->val, candidate->latency);
+
+       return 0;
+}
+
+static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
+                                        int index)
+{
+       cpumask_t mask;
+       int cpu;
+
+       numa_parse_mdesc_group_cpus(md, grp, &mask);
+
+       for_each_cpu_mask(cpu, mask)
+               numa_cpu_lookup_table[cpu] = index;
+       numa_cpumask_lookup_table[index] = mask;
+
+       if (numa_debug) {
+               printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
+               for_each_cpu_mask(cpu, mask)
+                       printk("%d ", cpu);
+               printk("]\n");
+       }
+
+       return numa_attach_mlgroup(md, grp, index);
+}
+
+static int __init numa_parse_mdesc(void)
+{
+       struct mdesc_handle *md = mdesc_grab();
+       int i, err, count;
+       u64 node;
+
+       node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
+       if (node == MDESC_NODE_NULL) {
+               mdesc_release(md);
+               return -ENOENT;
+       }
+
+       err = grab_mblocks(md);
+       if (err < 0)
+               goto out;
+
+       err = grab_mlgroups(md);
+       if (err < 0)
+               goto out;
+
+       count = 0;
+       mdesc_for_each_node_by_name(md, node, "group") {
+               err = numa_parse_mdesc_group(md, node, count);
+               if (err < 0)
+                       break;
+               count++;
+       }
+
+       add_node_ranges();
+
+       for (i = 0; i < num_node_masks; i++) {
+               allocate_node_data(i);
+               node_set_online(i);
+       }
+
+       err = 0;
+out:
+       mdesc_release(md);
+       return err;
+}
+
+static int __init numa_parse_sun4u(void)
+{
+       return -1;
+}
+
+static int __init bootmem_init_numa(void)
+{
+       int err = -1;
+
+       numadbg("bootmem_init_numa()\n");
+
+       if (numa_enabled) {
+               if (tlb_type == hypervisor)
+                       err = numa_parse_mdesc();
+               else
+                       err = numa_parse_sun4u();
+       }
+       return err;
+}
+
+#else
+
+static int bootmem_init_numa(void)
+{
+       return -1;
+}
+
+#endif
+
+static void __init bootmem_init_nonnuma(void)
+{
+       unsigned long top_of_ram = lmb_end_of_DRAM();
+       unsigned long total_ram = lmb_phys_mem_size();
+       unsigned int i;
+
+       numadbg("bootmem_init_nonnuma()\n");
+
+       printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
+              top_of_ram, total_ram);
+       printk(KERN_INFO "Memory hole size: %ldMB\n",
+              (top_of_ram - total_ram) >> 20);
+
+       init_node_masks_nonnuma();
+
+       for (i = 0; i < lmb.memory.cnt; i++) {
+               unsigned long size = lmb_size_bytes(&lmb.memory, i);
+               unsigned long start_pfn, end_pfn;
+
+               if (!size)
+                       continue;
 
-               pages = lmb_size_pages(&lmb.memory, i);
                start_pfn = lmb.memory.region[i].base >> PAGE_SHIFT;
-               end_pfn = start_pfn + pages;
+               end_pfn = start_pfn + lmb_size_pages(&lmb.memory, i);
+               add_active_range(0, start_pfn, end_pfn);
+       }
 
-               memory_present(0, start_pfn, end_pfn);
+       allocate_node_data(0);
+
+       node_set_online(0);
+}
+
+static void __init reserve_range_in_node(int nid, unsigned long start,
+                                        unsigned long end)
+{
+       numadbg("    reserve_range_in_node(nid[%d],start[%lx],end[%lx]\n",
+               nid, start, end);
+       while (start < end) {
+               unsigned long this_end;
+               int n;
+
+               this_end = nid_range(start, end, &n);
+               if (n == nid) {
+                       numadbg("      MATCH reserving range [%lx:%lx]\n",
+                               start, this_end);
+                       reserve_bootmem_node(NODE_DATA(nid), start,
+                                            (this_end - start), BOOTMEM_DEFAULT);
+               } else
+                       numadbg("      NO MATCH, advancing start to %lx\n",
+                               this_end);
+
+               start = this_end;
        }
+}
+
+static void __init trim_reserved_in_node(int nid)
+{
+       int i;
+
+       numadbg("  trim_reserved_in_node(%d)\n", nid);
+
+       for (i = 0; i < lmb.reserved.cnt; i++) {
+               unsigned long start = lmb.reserved.region[i].base;
+               unsigned long size = lmb_size_bytes(&lmb.reserved, i);
+               unsigned long end = start + size;
+
+               reserve_range_in_node(nid, start, end);
+       }
+}
+
+static void __init bootmem_init_one_node(int nid)
+{
+       struct pglist_data *p;
+
+       numadbg("bootmem_init_one_node(%d)\n", nid);
+
+       p = NODE_DATA(nid);
+
+       if (p->node_spanned_pages) {
+               unsigned long paddr = node_masks[nid].bootmem_paddr;
+               unsigned long end_pfn;
+
+               end_pfn = p->node_start_pfn + p->node_spanned_pages;
+
+               numadbg("  init_bootmem_node(%d, %lx, %lx, %lx)\n",
+                       nid, paddr >> PAGE_SHIFT, p->node_start_pfn, end_pfn);
+
+               init_bootmem_node(p, paddr >> PAGE_SHIFT,
+                                 p->node_start_pfn, end_pfn);
+
+               numadbg("  free_bootmem_with_active_regions(%d, %lx)\n",
+                       nid, end_pfn);
+               free_bootmem_with_active_regions(nid, end_pfn);
+
+               trim_reserved_in_node(nid);
+
+               numadbg("  sparse_memory_present_with_active_regions(%d)\n",
+                       nid);
+               sparse_memory_present_with_active_regions(nid);
+       }
+}
+
+static unsigned long __init bootmem_init(unsigned long phys_base)
+{
+       unsigned long end_pfn;
+       int nid;
+
+       end_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
+       max_pfn = max_low_pfn = end_pfn;
+       min_low_pfn = (phys_base >> PAGE_SHIFT);
+
+       if (bootmem_init_numa() < 0)
+               bootmem_init_nonnuma();
+
+       /* XXX cpu notifier XXX */
+
+       for_each_online_node(nid)
+               bootmem_init_one_node(nid);
 
        sparse_init();
 
@@ -1112,7 +1688,7 @@ void __init setup_per_cpu_areas(void)
 
 void __init paging_init(void)
 {
-       unsigned long end_pfn, pages_avail, shift, phys_base;
+       unsigned long end_pfn, shift, phys_base;
        unsigned long real_end, i;
 
        /* These build time checkes make sure that the dcache_dirty_cpu()
@@ -1220,27 +1796,21 @@ void __init paging_init(void)
                sun4v_mdesc_init();
 
        /* Setup bootmem... */
-       pages_avail = 0;
-       last_valid_pfn = end_pfn = bootmem_init(&pages_avail, phys_base);
+       last_valid_pfn = end_pfn = bootmem_init(phys_base);
 
+#ifndef CONFIG_NEED_MULTIPLE_NODES
        max_mapnr = last_valid_pfn;
-
+#endif
        kernel_physical_mapping_init();
 
        {
-               unsigned long zones_size[MAX_NR_ZONES];
-               unsigned long zholes_size[MAX_NR_ZONES];
-               int znum;
+               unsigned long max_zone_pfns[MAX_NR_ZONES];
 
-               for (znum = 0; znum < MAX_NR_ZONES; znum++)
-                       zones_size[znum] = zholes_size[znum] = 0;
+               memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 
-               zones_size[ZONE_NORMAL] = end_pfn;
-               zholes_size[ZONE_NORMAL] = end_pfn - pages_avail;
+               max_zone_pfns[ZONE_NORMAL] = end_pfn;
 
-               free_area_init_node(0, &contig_page_data, zones_size,
-                                   __pa(PAGE_OFFSET) >> PAGE_SHIFT,
-                                   zholes_size);
+               free_area_init_nodes(max_zone_pfns);
        }
 
        printk("Booting Linux...\n");
@@ -1249,21 +1819,52 @@ void __init paging_init(void)
        cpu_probe();
 }
 
-static void __init taint_real_pages(void)
+int __init page_in_phys_avail(unsigned long paddr)
+{
+       int i;
+
+       paddr &= PAGE_MASK;
+
+       for (i = 0; i < pavail_ents; i++) {
+               unsigned long start, end;
+
+               start = pavail[i].phys_addr;
+               end = start + pavail[i].reg_size;
+
+               if (paddr >= start && paddr < end)
+                       return 1;
+       }
+       if (paddr >= kern_base && paddr < (kern_base + kern_size))
+               return 1;
+#ifdef CONFIG_BLK_DEV_INITRD
+       if (paddr >= __pa(initrd_start) &&
+           paddr < __pa(PAGE_ALIGN(initrd_end)))
+               return 1;
+#endif
+
+       return 0;
+}
+
+static struct linux_prom64_registers pavail_rescan[MAX_BANKS] __initdata;
+static int pavail_rescan_ents __initdata;
+
+/* Certain OBP calls, such as fetching "available" properties, can
+ * claim physical memory.  So, along with initializing the valid
+ * address bitmap, what we do here is refetch the physical available
+ * memory list again, and make sure it provides at least as much
+ * memory as 'pavail' does.
+ */
+static void setup_valid_addr_bitmap_from_pavail(void)
 {
        int i;
 
        read_obp_memory("available", &pavail_rescan[0], &pavail_rescan_ents);
 
-       /* Find changes discovered in the physmem available rescan and
-        * reserve the lost portions in the bootmem maps.
-        */
        for (i = 0; i < pavail_ents; i++) {
                unsigned long old_start, old_end;
 
                old_start = pavail[i].phys_addr;
-               old_end = old_start +
-                       pavail[i].reg_size;
+               old_end = old_start + pavail[i].reg_size;
                while (old_start < old_end) {
                        int n;
 
@@ -1281,7 +1882,16 @@ static void __init taint_real_pages(void)
                                        goto do_next_page;
                                }
                        }
-                       reserve_bootmem(old_start, PAGE_SIZE, BOOTMEM_DEFAULT);
+
+                       prom_printf("mem_init: Lost memory in pavail\n");
+                       prom_printf("mem_init: OLD start[%lx] size[%lx]\n",
+                                   pavail[i].phys_addr,
+                                   pavail[i].reg_size);
+                       prom_printf("mem_init: NEW start[%lx] size[%lx]\n",
+                                   pavail_rescan[i].phys_addr,
+                                   pavail_rescan[i].reg_size);
+                       prom_printf("mem_init: Cannot continue, aborting.\n");
+                       prom_halt();
 
                do_next_page:
                        old_start += PAGE_SIZE;
@@ -1289,32 +1899,6 @@ static void __init taint_real_pages(void)
        }
 }
 
-int __init page_in_phys_avail(unsigned long paddr)
-{
-       int i;
-
-       paddr &= PAGE_MASK;
-
-       for (i = 0; i < pavail_rescan_ents; i++) {
-               unsigned long start, end;
-
-               start = pavail_rescan[i].phys_addr;
-               end = start + pavail_rescan[i].reg_size;
-
-               if (paddr >= start && paddr < end)
-                       return 1;
-       }
-       if (paddr >= kern_base && paddr < (kern_base + kern_size))
-               return 1;
-#ifdef CONFIG_BLK_DEV_INITRD
-       if (paddr >= __pa(initrd_start) &&
-           paddr < __pa(PAGE_ALIGN(initrd_end)))
-               return 1;
-#endif
-
-       return 0;
-}
-
 void __init mem_init(void)
 {
        unsigned long codepages, datapages, initpages;
@@ -1337,14 +1921,26 @@ void __init mem_init(void)
                addr += PAGE_SIZE;
        }
 
-       taint_real_pages();
+       setup_valid_addr_bitmap_from_pavail();
 
        high_memory = __va(last_valid_pfn << PAGE_SHIFT);
 
+#ifdef CONFIG_NEED_MULTIPLE_NODES
+       for_each_online_node(i) {
+               if (NODE_DATA(i)->node_spanned_pages != 0) {
+                       totalram_pages +=
+                               free_all_bootmem_node(NODE_DATA(i));
+               }
+       }
+#else
+       totalram_pages = free_all_bootmem();
+#endif
+
        /* We subtract one to account for the mem_map_zero page
         * allocated below.
         */
-       totalram_pages = num_physpages = free_all_bootmem() - 1;
+       totalram_pages -= 1;
+       num_physpages = totalram_pages;
 
        /*
         * Set up the zero page, mark it reserved, so that page count
diff --git a/include/asm-sparc64/mmzone.h b/include/asm-sparc64/mmzone.h
new file mode 100644 (file)
index 0000000..ebf5986
--- /dev/null
@@ -0,0 +1,17 @@
+#ifndef _SPARC64_MMZONE_H
+#define _SPARC64_MMZONE_H
+
+#ifdef CONFIG_NEED_MULTIPLE_NODES
+
+extern struct pglist_data *node_data[];
+
+#define NODE_DATA(nid)         (node_data[nid])
+#define node_start_pfn(nid)    (NODE_DATA(nid)->node_start_pfn)
+#define node_end_pfn(nid)      (NODE_DATA(nid)->node_end_pfn)
+
+extern int numa_cpu_lookup_table[];
+extern cpumask_t numa_cpumask_lookup_table[];
+
+#endif /* CONFIG_NEED_MULTIPLE_NODES */
+
+#endif /* _SPARC64_MMZONE_H */
index c6b557034f686623ab59a6a109b92302b5e4a3d2..001c04027c82b3f774a8c4aaeed12f5eff8c8835 100644 (file)
@@ -1,6 +1,77 @@
 #ifndef _ASM_SPARC64_TOPOLOGY_H
 #define _ASM_SPARC64_TOPOLOGY_H
 
+#ifdef CONFIG_NUMA
+
+#include <asm/mmzone.h>
+
+static inline int cpu_to_node(int cpu)
+{
+       return numa_cpu_lookup_table[cpu];
+}
+
+#define parent_node(node)      (node)
+
+static inline cpumask_t node_to_cpumask(int node)
+{
+       return numa_cpumask_lookup_table[node];
+}
+
+/* Returns a pointer to the cpumask of CPUs on Node 'node'. */
+#define node_to_cpumask_ptr(v, node)           \
+               cpumask_t *v = &(numa_cpumask_lookup_table[node])
+
+#define node_to_cpumask_ptr_next(v, node)      \
+                          v = &(numa_cpumask_lookup_table[node])
+
+static inline int node_to_first_cpu(int node)
+{
+       cpumask_t tmp;
+       tmp = node_to_cpumask(node);
+       return first_cpu(tmp);
+}
+
+struct pci_bus;
+#ifdef CONFIG_PCI
+extern int pcibus_to_node(struct pci_bus *pbus);
+#else
+static inline int pcibus_to_node(struct pci_bus *pbus)
+{
+       return -1;
+}
+#endif
+
+#define pcibus_to_cpumask(bus) \
+       (pcibus_to_node(bus) == -1 ? \
+        CPU_MASK_ALL : \
+        node_to_cpumask(pcibus_to_node(bus)))
+
+#define SD_NODE_INIT (struct sched_domain) {           \
+       .min_interval           = 8,                    \
+       .max_interval           = 32,                   \
+       .busy_factor            = 32,                   \
+       .imbalance_pct          = 125,                  \
+       .cache_nice_tries       = 2,                    \
+       .busy_idx               = 3,                    \
+       .idle_idx               = 2,                    \
+       .newidle_idx            = 0,                    \
+       .wake_idx               = 1,                    \
+       .forkexec_idx           = 1,                    \
+       .flags                  = SD_LOAD_BALANCE       \
+                               | SD_BALANCE_FORK       \
+                               | SD_BALANCE_EXEC       \
+                               | SD_SERIALIZE          \
+                               | SD_WAKE_BALANCE,      \
+       .last_balance           = jiffies,              \
+       .balance_interval       = 1,                    \
+}
+
+#else /* CONFIG_NUMA */
+
+#include <asm-generic/topology.h>
+
+#endif /* !(CONFIG_NUMA) */
+
 #ifdef CONFIG_SMP
 #define topology_physical_package_id(cpu)      (cpu_data(cpu).proc_id)
 #define topology_core_id(cpu)                  (cpu_data(cpu).core_id)
@@ -10,8 +81,6 @@
 #define smt_capable()                          (sparc64_multi_core)
 #endif /* CONFIG_SMP */
 
-#include <asm-generic/topology.h>
-
 #define cpu_coregroup_map(cpu)                 (cpu_core_map[cpu])
 
 #endif /* _ASM_SPARC64_TOPOLOGY_H */