x86_64: prepare shared mm/srat.c
authorThomas Gleixner <tglx@linutronix.de>
Thu, 11 Oct 2007 09:16:08 +0000 (11:16 +0200)
committerThomas Gleixner <tglx@linutronix.de>
Thu, 11 Oct 2007 09:16:08 +0000 (11:16 +0200)
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/x86_64/mm/Makefile
arch/x86_64/mm/srat.c [deleted file]
arch/x86_64/mm/srat_64.c [new file with mode: 0644]

index d25ac86fe27a93de7cf98eee014a5e90c1f5bfbc..f5fbb8053acda9d754ecb82a9050b3803ed9d6ec 100644 (file)
@@ -6,6 +6,6 @@ obj-y    := init.o fault.o ioremap.o extable.o pageattr.o mmap.o
 obj-$(CONFIG_HUGETLB_PAGE) += hugetlbpage.o
 obj-$(CONFIG_NUMA) += numa.o
 obj-$(CONFIG_K8_NUMA) += k8topology.o
-obj-$(CONFIG_ACPI_NUMA) += srat.o
+obj-$(CONFIG_ACPI_NUMA) += srat_64.o
 
 hugetlbpage-y = ../../i386/mm/hugetlbpage.o
diff --git a/arch/x86_64/mm/srat.c b/arch/x86_64/mm/srat.c
deleted file mode 100644 (file)
index acdf03e..0000000
+++ /dev/null
@@ -1,566 +0,0 @@
-/*
- * ACPI 3.0 based NUMA setup
- * Copyright 2004 Andi Kleen, SuSE Labs.
- *
- * Reads the ACPI SRAT table to figure out what memory belongs to which CPUs.
- *
- * Called from acpi_numa_init while reading the SRAT and SLIT tables.
- * Assumes all memory regions belonging to a single proximity domain
- * are in one chunk. Holes between them will be included in the node.
- */
-
-#include <linux/kernel.h>
-#include <linux/acpi.h>
-#include <linux/mmzone.h>
-#include <linux/bitmap.h>
-#include <linux/module.h>
-#include <linux/topology.h>
-#include <linux/bootmem.h>
-#include <linux/mm.h>
-#include <asm/proto.h>
-#include <asm/numa.h>
-#include <asm/e820.h>
-
-int acpi_numa __initdata;
-
-static struct acpi_table_slit *acpi_slit;
-
-static nodemask_t nodes_parsed __initdata;
-static struct bootnode nodes[MAX_NUMNODES] __initdata;
-static struct bootnode nodes_add[MAX_NUMNODES];
-static int found_add_area __initdata;
-int hotadd_percent __initdata = 0;
-
-/* Too small nodes confuse the VM badly. Usually they result
-   from BIOS bugs. */
-#define NODE_MIN_SIZE (4*1024*1024)
-
-static __init int setup_node(int pxm)
-{
-       return acpi_map_pxm_to_node(pxm);
-}
-
-static __init int conflicting_nodes(unsigned long start, unsigned long end)
-{
-       int i;
-       for_each_node_mask(i, nodes_parsed) {
-               struct bootnode *nd = &nodes[i];
-               if (nd->start == nd->end)
-                       continue;
-               if (nd->end > start && nd->start < end)
-                       return i;
-               if (nd->end == end && nd->start == start)
-                       return i;
-       }
-       return -1;
-}
-
-static __init void cutoff_node(int i, unsigned long start, unsigned long end)
-{
-       struct bootnode *nd = &nodes[i];
-
-       if (found_add_area)
-               return;
-
-       if (nd->start < start) {
-               nd->start = start;
-               if (nd->end < nd->start)
-                       nd->start = nd->end;
-       }
-       if (nd->end > end) {
-               nd->end = end;
-               if (nd->start > nd->end)
-                       nd->start = nd->end;
-       }
-}
-
-static __init void bad_srat(void)
-{
-       int i;
-       printk(KERN_ERR "SRAT: SRAT not used.\n");
-       acpi_numa = -1;
-       found_add_area = 0;
-       for (i = 0; i < MAX_LOCAL_APIC; i++)
-               apicid_to_node[i] = NUMA_NO_NODE;
-       for (i = 0; i < MAX_NUMNODES; i++)
-               nodes_add[i].start = nodes[i].end = 0;
-       remove_all_active_ranges();
-}
-
-static __init inline int srat_disabled(void)
-{
-       return numa_off || acpi_numa < 0;
-}
-
-/*
- * A lot of BIOS fill in 10 (= no distance) everywhere. This messes
- * up the NUMA heuristics which wants the local node to have a smaller
- * distance than the others.
- * Do some quick checks here and only use the SLIT if it passes.
- */
-static __init int slit_valid(struct acpi_table_slit *slit)
-{
-       int i, j;
-       int d = slit->locality_count;
-       for (i = 0; i < d; i++) {
-               for (j = 0; j < d; j++)  {
-                       u8 val = slit->entry[d*i + j];
-                       if (i == j) {
-                               if (val != LOCAL_DISTANCE)
-                                       return 0;
-                       } else if (val <= LOCAL_DISTANCE)
-                               return 0;
-               }
-       }
-       return 1;
-}
-
-/* Callback for SLIT parsing */
-void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
-{
-       if (!slit_valid(slit)) {
-               printk(KERN_INFO "ACPI: SLIT table looks invalid. Not used.\n");
-               return;
-       }
-       acpi_slit = slit;
-}
-
-/* Callback for Proximity Domain -> LAPIC mapping */
-void __init
-acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
-{
-       int pxm, node;
-       if (srat_disabled())
-               return;
-       if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
-               bad_srat();
-               return;
-       }
-       if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
-               return;
-       pxm = pa->proximity_domain_lo;
-       node = setup_node(pxm);
-       if (node < 0) {
-               printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
-               bad_srat();
-               return;
-       }
-       apicid_to_node[pa->apic_id] = node;
-       acpi_numa = 1;
-       printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
-              pxm, pa->apic_id, node);
-}
-
-#ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
-/*
- * Protect against too large hotadd areas that would fill up memory.
- */
-static int hotadd_enough_memory(struct bootnode *nd)
-{
-       static unsigned long allocated;
-       static unsigned long last_area_end;
-       unsigned long pages = (nd->end - nd->start) >> PAGE_SHIFT;
-       long mem = pages * sizeof(struct page);
-       unsigned long addr;
-       unsigned long allowed;
-       unsigned long oldpages = pages;
-
-       if (mem < 0)
-               return 0;
-       allowed = (end_pfn - absent_pages_in_range(0, end_pfn)) * PAGE_SIZE;
-       allowed = (allowed / 100) * hotadd_percent;
-       if (allocated + mem > allowed) {
-               unsigned long range;
-               /* Give them at least part of their hotadd memory upto hotadd_percent
-                  It would be better to spread the limit out
-                  over multiple hotplug areas, but that is too complicated
-                  right now */
-               if (allocated >= allowed)
-                       return 0;
-               range = allowed - allocated;
-               pages = (range / PAGE_SIZE);
-               mem = pages * sizeof(struct page);
-               nd->end = nd->start + range;
-       }
-       /* Not completely fool proof, but a good sanity check */
-       addr = find_e820_area(last_area_end, end_pfn<<PAGE_SHIFT, mem);
-       if (addr == -1UL)
-               return 0;
-       if (pages != oldpages)
-               printk(KERN_NOTICE "SRAT: Hotadd area limited to %lu bytes\n",
-                       pages << PAGE_SHIFT);
-       last_area_end = addr + mem;
-       allocated += mem;
-       return 1;
-}
-
-static int update_end_of_memory(unsigned long end)
-{
-       found_add_area = 1;
-       if ((end >> PAGE_SHIFT) > end_pfn)
-               end_pfn = end >> PAGE_SHIFT;
-       return 1;
-}
-
-static inline int save_add_info(void)
-{
-       return hotadd_percent > 0;
-}
-#else
-int update_end_of_memory(unsigned long end) {return -1;}
-static int hotadd_enough_memory(struct bootnode *nd) {return 1;}
-#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
-static inline int save_add_info(void) {return 1;}
-#else
-static inline int save_add_info(void) {return 0;}
-#endif
-#endif
-/*
- * Update nodes_add and decide if to include add are in the zone.
- * Both SPARSE and RESERVE need nodes_add infomation.
- * This code supports one contigious hot add area per node.
- */
-static int reserve_hotadd(int node, unsigned long start, unsigned long end)
-{
-       unsigned long s_pfn = start >> PAGE_SHIFT;
-       unsigned long e_pfn = end >> PAGE_SHIFT;
-       int ret = 0, changed = 0;
-       struct bootnode *nd = &nodes_add[node];
-
-       /* I had some trouble with strange memory hotadd regions breaking
-          the boot. Be very strict here and reject anything unexpected.
-          If you want working memory hotadd write correct SRATs.
-
-          The node size check is a basic sanity check to guard against
-          mistakes */
-       if ((signed long)(end - start) < NODE_MIN_SIZE) {
-               printk(KERN_ERR "SRAT: Hotplug area too small\n");
-               return -1;
-       }
-
-       /* This check might be a bit too strict, but I'm keeping it for now. */
-       if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
-               printk(KERN_ERR
-                       "SRAT: Hotplug area %lu -> %lu has existing memory\n",
-                       s_pfn, e_pfn);
-               return -1;
-       }
-
-       if (!hotadd_enough_memory(&nodes_add[node]))  {
-               printk(KERN_ERR "SRAT: Hotplug area too large\n");
-               return -1;
-       }
-
-       /* Looks good */
-
-       if (nd->start == nd->end) {
-               nd->start = start;
-               nd->end = end;
-               changed = 1;
-       } else {
-               if (nd->start == end) {
-                       nd->start = start;
-                       changed = 1;
-               }
-               if (nd->end == start) {
-                       nd->end = end;
-                       changed = 1;
-               }
-               if (!changed)
-                       printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
-       }
-
-       ret = update_end_of_memory(nd->end);
-
-       if (changed)
-               printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
-       return ret;
-}
-
-/* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
-void __init
-acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
-{
-       struct bootnode *nd, oldnode;
-       unsigned long start, end;
-       int node, pxm;
-       int i;
-
-       if (srat_disabled())
-               return;
-       if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
-               bad_srat();
-               return;
-       }
-       if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
-               return;
-
-       if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
-               return;
-       start = ma->base_address;
-       end = start + ma->length;
-       pxm = ma->proximity_domain;
-       node = setup_node(pxm);
-       if (node < 0) {
-               printk(KERN_ERR "SRAT: Too many proximity domains.\n");
-               bad_srat();
-               return;
-       }
-       i = conflicting_nodes(start, end);
-       if (i == node) {
-               printk(KERN_WARNING
-               "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
-                       pxm, start, end, nodes[i].start, nodes[i].end);
-       } else if (i >= 0) {
-               printk(KERN_ERR
-                      "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
-                      pxm, start, end, node_to_pxm(i),
-                       nodes[i].start, nodes[i].end);
-               bad_srat();
-               return;
-       }
-       nd = &nodes[node];
-       oldnode = *nd;
-       if (!node_test_and_set(node, nodes_parsed)) {
-               nd->start = start;
-               nd->end = end;
-       } else {
-               if (start < nd->start)
-                       nd->start = start;
-               if (nd->end < end)
-                       nd->end = end;
-       }
-
-       printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm,
-              nd->start, nd->end);
-       e820_register_active_regions(node, nd->start >> PAGE_SHIFT,
-                                               nd->end >> PAGE_SHIFT);
-       push_node_boundaries(node, nd->start >> PAGE_SHIFT,
-                                               nd->end >> PAGE_SHIFT);
-
-       if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) &&
-           (reserve_hotadd(node, start, end) < 0)) {
-               /* Ignore hotadd region. Undo damage */
-               printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
-               *nd = oldnode;
-               if ((nd->start | nd->end) == 0)
-                       node_clear(node, nodes_parsed);
-       }
-}
-
-/* Sanity check to catch more bad SRATs (they are amazingly common).
-   Make sure the PXMs cover all memory. */
-static int __init nodes_cover_memory(const struct bootnode *nodes)
-{
-       int i;
-       unsigned long pxmram, e820ram;
-
-       pxmram = 0;
-       for_each_node_mask(i, nodes_parsed) {
-               unsigned long s = nodes[i].start >> PAGE_SHIFT;
-               unsigned long e = nodes[i].end >> PAGE_SHIFT;
-               pxmram += e - s;
-               pxmram -= absent_pages_in_range(s, e);
-               if ((long)pxmram < 0)
-                       pxmram = 0;
-       }
-
-       e820ram = end_pfn - absent_pages_in_range(0, end_pfn);
-       /* We seem to lose 3 pages somewhere. Allow a bit of slack. */
-       if ((long)(e820ram - pxmram) >= 1*1024*1024) {
-               printk(KERN_ERR
-       "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
-                       (pxmram << PAGE_SHIFT) >> 20,
-                       (e820ram << PAGE_SHIFT) >> 20);
-               return 0;
-       }
-       return 1;
-}
-
-static void unparse_node(int node)
-{
-       int i;
-       node_clear(node, nodes_parsed);
-       for (i = 0; i < MAX_LOCAL_APIC; i++) {
-               if (apicid_to_node[i] == node)
-                       apicid_to_node[i] = NUMA_NO_NODE;
-       }
-}
-
-void __init acpi_numa_arch_fixup(void) {}
-
-/* Use the information discovered above to actually set up the nodes. */
-int __init acpi_scan_nodes(unsigned long start, unsigned long end)
-{
-       int i;
-
-       if (acpi_numa <= 0)
-               return -1;
-
-       /* First clean up the node list */
-       for (i = 0; i < MAX_NUMNODES; i++) {
-               cutoff_node(i, start, end);
-               if ((nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) {
-                       unparse_node(i);
-                       node_set_offline(i);
-               }
-       }
-
-       if (!nodes_cover_memory(nodes)) {
-               bad_srat();
-               return -1;
-       }
-
-       memnode_shift = compute_hash_shift(nodes, MAX_NUMNODES);
-       if (memnode_shift < 0) {
-               printk(KERN_ERR
-                    "SRAT: No NUMA node hash function found. Contact maintainer\n");
-               bad_srat();
-               return -1;
-       }
-
-       node_possible_map = nodes_parsed;
-
-       /* Finally register nodes */
-       for_each_node_mask(i, node_possible_map)
-               setup_node_bootmem(i, nodes[i].start, nodes[i].end);
-       /* Try again in case setup_node_bootmem missed one due
-          to missing bootmem */
-       for_each_node_mask(i, node_possible_map)
-               if (!node_online(i))
-                       setup_node_bootmem(i, nodes[i].start, nodes[i].end);
-
-       for (i = 0; i < NR_CPUS; i++) {
-               if (cpu_to_node[i] == NUMA_NO_NODE)
-                       continue;
-               if (!node_isset(cpu_to_node[i], node_possible_map))
-                       numa_set_node(i, NUMA_NO_NODE);
-       }
-       numa_init_array();
-       return 0;
-}
-
-#ifdef CONFIG_NUMA_EMU
-static int __init find_node_by_addr(unsigned long addr)
-{
-       int ret = NUMA_NO_NODE;
-       int i;
-
-       for_each_node_mask(i, nodes_parsed) {
-               /*
-                * Find the real node that this emulated node appears on.  For
-                * the sake of simplicity, we only use a real node's starting
-                * address to determine which emulated node it appears on.
-                */
-               if (addr >= nodes[i].start && addr < nodes[i].end) {
-                       ret = i;
-                       break;
-               }
-       }
-       return i;
-}
-
-/*
- * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
- * mappings that respect the real ACPI topology but reflect our emulated
- * environment.  For each emulated node, we find which real node it appears on
- * and create PXM to NID mappings for those fake nodes which mirror that
- * locality.  SLIT will now represent the correct distances between emulated
- * nodes as a result of the real topology.
- */
-void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
-{
-       int i, j;
-       int fake_node_to_pxm_map[MAX_NUMNODES] = {
-               [0 ... MAX_NUMNODES-1] = PXM_INVAL
-       };
-       unsigned char fake_apicid_to_node[MAX_LOCAL_APIC] = {
-               [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
-       };
-
-       printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
-                        "topology.\n");
-       for (i = 0; i < num_nodes; i++) {
-               int nid, pxm;
-
-               nid = find_node_by_addr(fake_nodes[i].start);
-               if (nid == NUMA_NO_NODE)
-                       continue;
-               pxm = node_to_pxm(nid);
-               if (pxm == PXM_INVAL)
-                       continue;
-               fake_node_to_pxm_map[i] = pxm;
-               /*
-                * For each apicid_to_node mapping that exists for this real
-                * node, it must now point to the fake node ID.
-                */
-               for (j = 0; j < MAX_LOCAL_APIC; j++)
-                       if (apicid_to_node[j] == nid)
-                               fake_apicid_to_node[j] = i;
-       }
-       for (i = 0; i < num_nodes; i++)
-               __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
-       memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
-
-       nodes_clear(nodes_parsed);
-       for (i = 0; i < num_nodes; i++)
-               if (fake_nodes[i].start != fake_nodes[i].end)
-                       node_set(i, nodes_parsed);
-       WARN_ON(!nodes_cover_memory(fake_nodes));
-}
-
-static int null_slit_node_compare(int a, int b)
-{
-       return node_to_pxm(a) == node_to_pxm(b);
-}
-#else
-static int null_slit_node_compare(int a, int b)
-{
-       return a == b;
-}
-#endif /* CONFIG_NUMA_EMU */
-
-void __init srat_reserve_add_area(int nodeid)
-{
-       if (found_add_area && nodes_add[nodeid].end) {
-               u64 total_mb;
-
-               printk(KERN_INFO "SRAT: Reserving hot-add memory space "
-                               "for node %d at %Lx-%Lx\n",
-                       nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end);
-               total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start)
-                                       >> PAGE_SHIFT;
-               total_mb *= sizeof(struct page);
-               total_mb >>= 20;
-               printk(KERN_INFO "SRAT: This will cost you %Lu MB of "
-                               "pre-allocated memory.\n", (unsigned long long)total_mb);
-               reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start,
-                              nodes_add[nodeid].end - nodes_add[nodeid].start);
-       }
-}
-
-int __node_distance(int a, int b)
-{
-       int index;
-
-       if (!acpi_slit)
-               return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
-                                                     REMOTE_DISTANCE;
-       index = acpi_slit->locality_count * node_to_pxm(a);
-       return acpi_slit->entry[index + node_to_pxm(b)];
-}
-
-EXPORT_SYMBOL(__node_distance);
-
-int memory_add_physaddr_to_nid(u64 start)
-{
-       int i, ret = 0;
-
-       for_each_node(i)
-               if (nodes_add[i].start <= start && nodes_add[i].end > start)
-                       ret = i;
-
-       return ret;
-}
-EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
-
diff --git a/arch/x86_64/mm/srat_64.c b/arch/x86_64/mm/srat_64.c
new file mode 100644 (file)
index 0000000..acdf03e
--- /dev/null
@@ -0,0 +1,566 @@
+/*
+ * ACPI 3.0 based NUMA setup
+ * Copyright 2004 Andi Kleen, SuSE Labs.
+ *
+ * Reads the ACPI SRAT table to figure out what memory belongs to which CPUs.
+ *
+ * Called from acpi_numa_init while reading the SRAT and SLIT tables.
+ * Assumes all memory regions belonging to a single proximity domain
+ * are in one chunk. Holes between them will be included in the node.
+ */
+
+#include <linux/kernel.h>
+#include <linux/acpi.h>
+#include <linux/mmzone.h>
+#include <linux/bitmap.h>
+#include <linux/module.h>
+#include <linux/topology.h>
+#include <linux/bootmem.h>
+#include <linux/mm.h>
+#include <asm/proto.h>
+#include <asm/numa.h>
+#include <asm/e820.h>
+
+int acpi_numa __initdata;
+
+static struct acpi_table_slit *acpi_slit;
+
+static nodemask_t nodes_parsed __initdata;
+static struct bootnode nodes[MAX_NUMNODES] __initdata;
+static struct bootnode nodes_add[MAX_NUMNODES];
+static int found_add_area __initdata;
+int hotadd_percent __initdata = 0;
+
+/* Too small nodes confuse the VM badly. Usually they result
+   from BIOS bugs. */
+#define NODE_MIN_SIZE (4*1024*1024)
+
+static __init int setup_node(int pxm)
+{
+       return acpi_map_pxm_to_node(pxm);
+}
+
+static __init int conflicting_nodes(unsigned long start, unsigned long end)
+{
+       int i;
+       for_each_node_mask(i, nodes_parsed) {
+               struct bootnode *nd = &nodes[i];
+               if (nd->start == nd->end)
+                       continue;
+               if (nd->end > start && nd->start < end)
+                       return i;
+               if (nd->end == end && nd->start == start)
+                       return i;
+       }
+       return -1;
+}
+
+static __init void cutoff_node(int i, unsigned long start, unsigned long end)
+{
+       struct bootnode *nd = &nodes[i];
+
+       if (found_add_area)
+               return;
+
+       if (nd->start < start) {
+               nd->start = start;
+               if (nd->end < nd->start)
+                       nd->start = nd->end;
+       }
+       if (nd->end > end) {
+               nd->end = end;
+               if (nd->start > nd->end)
+                       nd->start = nd->end;
+       }
+}
+
+static __init void bad_srat(void)
+{
+       int i;
+       printk(KERN_ERR "SRAT: SRAT not used.\n");
+       acpi_numa = -1;
+       found_add_area = 0;
+       for (i = 0; i < MAX_LOCAL_APIC; i++)
+               apicid_to_node[i] = NUMA_NO_NODE;
+       for (i = 0; i < MAX_NUMNODES; i++)
+               nodes_add[i].start = nodes[i].end = 0;
+       remove_all_active_ranges();
+}
+
+static __init inline int srat_disabled(void)
+{
+       return numa_off || acpi_numa < 0;
+}
+
+/*
+ * A lot of BIOS fill in 10 (= no distance) everywhere. This messes
+ * up the NUMA heuristics which wants the local node to have a smaller
+ * distance than the others.
+ * Do some quick checks here and only use the SLIT if it passes.
+ */
+static __init int slit_valid(struct acpi_table_slit *slit)
+{
+       int i, j;
+       int d = slit->locality_count;
+       for (i = 0; i < d; i++) {
+               for (j = 0; j < d; j++)  {
+                       u8 val = slit->entry[d*i + j];
+                       if (i == j) {
+                               if (val != LOCAL_DISTANCE)
+                                       return 0;
+                       } else if (val <= LOCAL_DISTANCE)
+                               return 0;
+               }
+       }
+       return 1;
+}
+
+/* Callback for SLIT parsing */
+void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
+{
+       if (!slit_valid(slit)) {
+               printk(KERN_INFO "ACPI: SLIT table looks invalid. Not used.\n");
+               return;
+       }
+       acpi_slit = slit;
+}
+
+/* Callback for Proximity Domain -> LAPIC mapping */
+void __init
+acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
+{
+       int pxm, node;
+       if (srat_disabled())
+               return;
+       if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
+               bad_srat();
+               return;
+       }
+       if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
+               return;
+       pxm = pa->proximity_domain_lo;
+       node = setup_node(pxm);
+       if (node < 0) {
+               printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
+               bad_srat();
+               return;
+       }
+       apicid_to_node[pa->apic_id] = node;
+       acpi_numa = 1;
+       printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
+              pxm, pa->apic_id, node);
+}
+
+#ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
+/*
+ * Protect against too large hotadd areas that would fill up memory.
+ */
+static int hotadd_enough_memory(struct bootnode *nd)
+{
+       static unsigned long allocated;
+       static unsigned long last_area_end;
+       unsigned long pages = (nd->end - nd->start) >> PAGE_SHIFT;
+       long mem = pages * sizeof(struct page);
+       unsigned long addr;
+       unsigned long allowed;
+       unsigned long oldpages = pages;
+
+       if (mem < 0)
+               return 0;
+       allowed = (end_pfn - absent_pages_in_range(0, end_pfn)) * PAGE_SIZE;
+       allowed = (allowed / 100) * hotadd_percent;
+       if (allocated + mem > allowed) {
+               unsigned long range;
+               /* Give them at least part of their hotadd memory upto hotadd_percent
+                  It would be better to spread the limit out
+                  over multiple hotplug areas, but that is too complicated
+                  right now */
+               if (allocated >= allowed)
+                       return 0;
+               range = allowed - allocated;
+               pages = (range / PAGE_SIZE);
+               mem = pages * sizeof(struct page);
+               nd->end = nd->start + range;
+       }
+       /* Not completely fool proof, but a good sanity check */
+       addr = find_e820_area(last_area_end, end_pfn<<PAGE_SHIFT, mem);
+       if (addr == -1UL)
+               return 0;
+       if (pages != oldpages)
+               printk(KERN_NOTICE "SRAT: Hotadd area limited to %lu bytes\n",
+                       pages << PAGE_SHIFT);
+       last_area_end = addr + mem;
+       allocated += mem;
+       return 1;
+}
+
+static int update_end_of_memory(unsigned long end)
+{
+       found_add_area = 1;
+       if ((end >> PAGE_SHIFT) > end_pfn)
+               end_pfn = end >> PAGE_SHIFT;
+       return 1;
+}
+
+static inline int save_add_info(void)
+{
+       return hotadd_percent > 0;
+}
+#else
+int update_end_of_memory(unsigned long end) {return -1;}
+static int hotadd_enough_memory(struct bootnode *nd) {return 1;}
+#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
+static inline int save_add_info(void) {return 1;}
+#else
+static inline int save_add_info(void) {return 0;}
+#endif
+#endif
+/*
+ * Update nodes_add and decide if to include add are in the zone.
+ * Both SPARSE and RESERVE need nodes_add infomation.
+ * This code supports one contigious hot add area per node.
+ */
+static int reserve_hotadd(int node, unsigned long start, unsigned long end)
+{
+       unsigned long s_pfn = start >> PAGE_SHIFT;
+       unsigned long e_pfn = end >> PAGE_SHIFT;
+       int ret = 0, changed = 0;
+       struct bootnode *nd = &nodes_add[node];
+
+       /* I had some trouble with strange memory hotadd regions breaking
+          the boot. Be very strict here and reject anything unexpected.
+          If you want working memory hotadd write correct SRATs.
+
+          The node size check is a basic sanity check to guard against
+          mistakes */
+       if ((signed long)(end - start) < NODE_MIN_SIZE) {
+               printk(KERN_ERR "SRAT: Hotplug area too small\n");
+               return -1;
+       }
+
+       /* This check might be a bit too strict, but I'm keeping it for now. */
+       if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
+               printk(KERN_ERR
+                       "SRAT: Hotplug area %lu -> %lu has existing memory\n",
+                       s_pfn, e_pfn);
+               return -1;
+       }
+
+       if (!hotadd_enough_memory(&nodes_add[node]))  {
+               printk(KERN_ERR "SRAT: Hotplug area too large\n");
+               return -1;
+       }
+
+       /* Looks good */
+
+       if (nd->start == nd->end) {
+               nd->start = start;
+               nd->end = end;
+               changed = 1;
+       } else {
+               if (nd->start == end) {
+                       nd->start = start;
+                       changed = 1;
+               }
+               if (nd->end == start) {
+                       nd->end = end;
+                       changed = 1;
+               }
+               if (!changed)
+                       printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
+       }
+
+       ret = update_end_of_memory(nd->end);
+
+       if (changed)
+               printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
+       return ret;
+}
+
+/* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
+void __init
+acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
+{
+       struct bootnode *nd, oldnode;
+       unsigned long start, end;
+       int node, pxm;
+       int i;
+
+       if (srat_disabled())
+               return;
+       if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
+               bad_srat();
+               return;
+       }
+       if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
+               return;
+
+       if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
+               return;
+       start = ma->base_address;
+       end = start + ma->length;
+       pxm = ma->proximity_domain;
+       node = setup_node(pxm);
+       if (node < 0) {
+               printk(KERN_ERR "SRAT: Too many proximity domains.\n");
+               bad_srat();
+               return;
+       }
+       i = conflicting_nodes(start, end);
+       if (i == node) {
+               printk(KERN_WARNING
+               "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
+                       pxm, start, end, nodes[i].start, nodes[i].end);
+       } else if (i >= 0) {
+               printk(KERN_ERR
+                      "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
+                      pxm, start, end, node_to_pxm(i),
+                       nodes[i].start, nodes[i].end);
+               bad_srat();
+               return;
+       }
+       nd = &nodes[node];
+       oldnode = *nd;
+       if (!node_test_and_set(node, nodes_parsed)) {
+               nd->start = start;
+               nd->end = end;
+       } else {
+               if (start < nd->start)
+                       nd->start = start;
+               if (nd->end < end)
+                       nd->end = end;
+       }
+
+       printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm,
+              nd->start, nd->end);
+       e820_register_active_regions(node, nd->start >> PAGE_SHIFT,
+                                               nd->end >> PAGE_SHIFT);
+       push_node_boundaries(node, nd->start >> PAGE_SHIFT,
+                                               nd->end >> PAGE_SHIFT);
+
+       if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) &&
+           (reserve_hotadd(node, start, end) < 0)) {
+               /* Ignore hotadd region. Undo damage */
+               printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
+               *nd = oldnode;
+               if ((nd->start | nd->end) == 0)
+                       node_clear(node, nodes_parsed);
+       }
+}
+
+/* Sanity check to catch more bad SRATs (they are amazingly common).
+   Make sure the PXMs cover all memory. */
+static int __init nodes_cover_memory(const struct bootnode *nodes)
+{
+       int i;
+       unsigned long pxmram, e820ram;
+
+       pxmram = 0;
+       for_each_node_mask(i, nodes_parsed) {
+               unsigned long s = nodes[i].start >> PAGE_SHIFT;
+               unsigned long e = nodes[i].end >> PAGE_SHIFT;
+               pxmram += e - s;
+               pxmram -= absent_pages_in_range(s, e);
+               if ((long)pxmram < 0)
+                       pxmram = 0;
+       }
+
+       e820ram = end_pfn - absent_pages_in_range(0, end_pfn);
+       /* We seem to lose 3 pages somewhere. Allow a bit of slack. */
+       if ((long)(e820ram - pxmram) >= 1*1024*1024) {
+               printk(KERN_ERR
+       "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
+                       (pxmram << PAGE_SHIFT) >> 20,
+                       (e820ram << PAGE_SHIFT) >> 20);
+               return 0;
+       }
+       return 1;
+}
+
+static void unparse_node(int node)
+{
+       int i;
+       node_clear(node, nodes_parsed);
+       for (i = 0; i < MAX_LOCAL_APIC; i++) {
+               if (apicid_to_node[i] == node)
+                       apicid_to_node[i] = NUMA_NO_NODE;
+       }
+}
+
+void __init acpi_numa_arch_fixup(void) {}
+
+/* Use the information discovered above to actually set up the nodes. */
+int __init acpi_scan_nodes(unsigned long start, unsigned long end)
+{
+       int i;
+
+       if (acpi_numa <= 0)
+               return -1;
+
+       /* First clean up the node list */
+       for (i = 0; i < MAX_NUMNODES; i++) {
+               cutoff_node(i, start, end);
+               if ((nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) {
+                       unparse_node(i);
+                       node_set_offline(i);
+               }
+       }
+
+       if (!nodes_cover_memory(nodes)) {
+               bad_srat();
+               return -1;
+       }
+
+       memnode_shift = compute_hash_shift(nodes, MAX_NUMNODES);
+       if (memnode_shift < 0) {
+               printk(KERN_ERR
+                    "SRAT: No NUMA node hash function found. Contact maintainer\n");
+               bad_srat();
+               return -1;
+       }
+
+       node_possible_map = nodes_parsed;
+
+       /* Finally register nodes */
+       for_each_node_mask(i, node_possible_map)
+               setup_node_bootmem(i, nodes[i].start, nodes[i].end);
+       /* Try again in case setup_node_bootmem missed one due
+          to missing bootmem */
+       for_each_node_mask(i, node_possible_map)
+               if (!node_online(i))
+                       setup_node_bootmem(i, nodes[i].start, nodes[i].end);
+
+       for (i = 0; i < NR_CPUS; i++) {
+               if (cpu_to_node[i] == NUMA_NO_NODE)
+                       continue;
+               if (!node_isset(cpu_to_node[i], node_possible_map))
+                       numa_set_node(i, NUMA_NO_NODE);
+       }
+       numa_init_array();
+       return 0;
+}
+
+#ifdef CONFIG_NUMA_EMU
+static int __init find_node_by_addr(unsigned long addr)
+{
+       int ret = NUMA_NO_NODE;
+       int i;
+
+       for_each_node_mask(i, nodes_parsed) {
+               /*
+                * Find the real node that this emulated node appears on.  For
+                * the sake of simplicity, we only use a real node's starting
+                * address to determine which emulated node it appears on.
+                */
+               if (addr >= nodes[i].start && addr < nodes[i].end) {
+                       ret = i;
+                       break;
+               }
+       }
+       return i;
+}
+
+/*
+ * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
+ * mappings that respect the real ACPI topology but reflect our emulated
+ * environment.  For each emulated node, we find which real node it appears on
+ * and create PXM to NID mappings for those fake nodes which mirror that
+ * locality.  SLIT will now represent the correct distances between emulated
+ * nodes as a result of the real topology.
+ */
+void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
+{
+       int i, j;
+       int fake_node_to_pxm_map[MAX_NUMNODES] = {
+               [0 ... MAX_NUMNODES-1] = PXM_INVAL
+       };
+       unsigned char fake_apicid_to_node[MAX_LOCAL_APIC] = {
+               [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
+       };
+
+       printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
+                        "topology.\n");
+       for (i = 0; i < num_nodes; i++) {
+               int nid, pxm;
+
+               nid = find_node_by_addr(fake_nodes[i].start);
+               if (nid == NUMA_NO_NODE)
+                       continue;
+               pxm = node_to_pxm(nid);
+               if (pxm == PXM_INVAL)
+                       continue;
+               fake_node_to_pxm_map[i] = pxm;
+               /*
+                * For each apicid_to_node mapping that exists for this real
+                * node, it must now point to the fake node ID.
+                */
+               for (j = 0; j < MAX_LOCAL_APIC; j++)
+                       if (apicid_to_node[j] == nid)
+                               fake_apicid_to_node[j] = i;
+       }
+       for (i = 0; i < num_nodes; i++)
+               __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
+       memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
+
+       nodes_clear(nodes_parsed);
+       for (i = 0; i < num_nodes; i++)
+               if (fake_nodes[i].start != fake_nodes[i].end)
+                       node_set(i, nodes_parsed);
+       WARN_ON(!nodes_cover_memory(fake_nodes));
+}
+
+static int null_slit_node_compare(int a, int b)
+{
+       return node_to_pxm(a) == node_to_pxm(b);
+}
+#else
+static int null_slit_node_compare(int a, int b)
+{
+       return a == b;
+}
+#endif /* CONFIG_NUMA_EMU */
+
+void __init srat_reserve_add_area(int nodeid)
+{
+       if (found_add_area && nodes_add[nodeid].end) {
+               u64 total_mb;
+
+               printk(KERN_INFO "SRAT: Reserving hot-add memory space "
+                               "for node %d at %Lx-%Lx\n",
+                       nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end);
+               total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start)
+                                       >> PAGE_SHIFT;
+               total_mb *= sizeof(struct page);
+               total_mb >>= 20;
+               printk(KERN_INFO "SRAT: This will cost you %Lu MB of "
+                               "pre-allocated memory.\n", (unsigned long long)total_mb);
+               reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start,
+                              nodes_add[nodeid].end - nodes_add[nodeid].start);
+       }
+}
+
+int __node_distance(int a, int b)
+{
+       int index;
+
+       if (!acpi_slit)
+               return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
+                                                     REMOTE_DISTANCE;
+       index = acpi_slit->locality_count * node_to_pxm(a);
+       return acpi_slit->entry[index + node_to_pxm(b)];
+}
+
+EXPORT_SYMBOL(__node_distance);
+
+int memory_add_physaddr_to_nid(u64 start)
+{
+       int i, ret = 0;
+
+       for_each_node(i)
+               if (nodes_add[i].start <= start && nodes_add[i].end > start)
+                       ret = i;
+
+       return ret;
+}
+EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
+