arch/s390/numa/numa.c

   1 /*
   2  * NUMA support for s390
   3  *
   4  * Implement NUMA core code.
   5  *
   6  * Copyright IBM Corp. 2015
   7  */
   8
   9 #define KMSG_COMPONENT "numa"
  10 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  11
  12 #include <linux/kernel.h>
  13 #include <linux/mmzone.h>
  14 #include <linux/cpumask.h>
  15 #include <linux/bootmem.h>
  16 #include <linux/memblock.h>
  17 #include <linux/slab.h>
  18 #include <linux/node.h>
  19
  20 #include <asm/numa.h>
  21 #include "numa_mode.h"
  22
  23 pg_data_t *node_data[MAX_NUMNODES];
  24 EXPORT_SYMBOL(node_data);
  25
  26 cpumask_t node_to_cpumask_map[MAX_NUMNODES];
  27 EXPORT_SYMBOL(node_to_cpumask_map);
  28
  29 const struct numa_mode numa_mode_plain = {
  30         .name = "plain",
  31 };
  32
  33 static const struct numa_mode *mode = &numa_mode_plain;
  34
  35 int numa_pfn_to_nid(unsigned long pfn)
  36 {
  37         return mode->__pfn_to_nid ? mode->__pfn_to_nid(pfn) : 0;
  38 }
  39
  40 void numa_update_cpu_topology(void)
  41 {
  42         if (mode->update_cpu_topology)
  43                 mode->update_cpu_topology();
  44 }
  45
  46 int __node_distance(int a, int b)
  47 {
  48         return mode->distance ? mode->distance(a, b) : 0;
  49 }
  50
  51 int numa_debug_enabled;
  52
  53 /*
  54  * alloc_node_data() - Allocate node data
  55  */
  56 static __init pg_data_t *alloc_node_data(void)
  57 {
  58         pg_data_t *res;
  59
  60         res = (pg_data_t *) memblock_alloc(sizeof(pg_data_t), 1);
  61         if (!res)
  62                 panic("Could not allocate memory for node data!\n");
  63         memset(res, 0, sizeof(pg_data_t));
  64         return res;
  65 }
  66
  67 /*
  68  * numa_setup_memory() - Assign bootmem to nodes
  69  *
  70  * The memory is first added to memblock without any respect to nodes.
  71  * This is fixed before remaining memblock memory is handed over to the
  72  * buddy allocator.
  73  * An important side effect is that large bootmem allocations might easily
  74  * cross node boundaries, which can be needed for large allocations with
  75  * smaller memory stripes in each node (i.e. when using NUMA emulation).
  76  *
  77  * Memory defines nodes:
  78  * Therefore this routine also sets the nodes online with memory.
  79  */
  80 static void __init numa_setup_memory(void)
  81 {
  82         unsigned long cur_base, align, end_of_dram;
  83         int nid = 0;
  84
  85         end_of_dram = memblock_end_of_DRAM();
  86         align = mode->align ? mode->align() : ULONG_MAX;
  87
  88         /*
  89          * Step through all available memory and assign it to the nodes
  90          * indicated by the mode implementation.
  91          * All nodes which are seen here will be set online.
  92          */
  93         cur_base = 0;
  94         do {
  95                 nid = numa_pfn_to_nid(PFN_DOWN(cur_base));
  96                 node_set_online(nid);
  97                 memblock_set_node(cur_base, align, &memblock.memory, nid);
  98                 cur_base += align;
  99         } while (cur_base < end_of_dram);
 100
 101         /* Allocate and fill out node_data */
 102         for (nid = 0; nid < MAX_NUMNODES; nid++)
 103                 NODE_DATA(nid) = alloc_node_data();
 104
 105         for_each_online_node(nid) {
 106                 unsigned long start_pfn, end_pfn;
 107                 unsigned long t_start, t_end;
 108                 int i;
 109
 110                 start_pfn = ULONG_MAX;
 111                 end_pfn = 0;
 112                 for_each_mem_pfn_range(i, nid, &t_start, &t_end, NULL) {
 113                         if (t_start < start_pfn)
 114                                 start_pfn = t_start;
 115                         if (t_end > end_pfn)
 116                                 end_pfn = t_end;
 117                 }
 118                 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
 119                 NODE_DATA(nid)->node_id = nid;
 120         }
 121 }
 122
 123 /*
 124  * numa_setup() - Earliest initialization
 125  *
 126  * Assign the mode and call the mode's setup routine.
 127  */
 128 void __init numa_setup(void)
 129 {
 130         pr_info("NUMA mode: %s\n", mode->name);
 131         if (mode->setup)
 132                 mode->setup();
 133         numa_setup_memory();
 134         memblock_dump_all();
 135 }
 136
 137
 138 /*
 139  * numa_init_early() - Initialization initcall
 140  *
 141  * This runs when only one CPU is online and before the first
 142  * topology update is called for by the scheduler.
 143  */
 144 static int __init numa_init_early(void)
 145 {
 146         /* Attach all possible CPUs to node 0 for now. */
 147         cpumask_copy(&node_to_cpumask_map[0], cpu_possible_mask);
 148         return 0;
 149 }
 150 early_initcall(numa_init_early);
 151
 152 /*
 153  * numa_init_late() - Initialization initcall
 154  *
 155  * Register NUMA nodes.
 156  */
 157 static int __init numa_init_late(void)
 158 {
 159         int nid;
 160
 161         for_each_online_node(nid)
 162                 register_one_node(nid);
 163         return 0;
 164 }
 165 device_initcall(numa_init_late);
 166
 167 static int __init parse_debug(char *parm)
 168 {
 169         numa_debug_enabled = 1;
 170         return 0;
 171 }
 172 early_param("numa_debug", parse_debug);
 173
 174 static int __init parse_numa(char *parm)
 175 {
 176         if (strcmp(parm, numa_mode_plain.name) == 0)
 177                 mode = &numa_mode_plain;
 178 #ifdef CONFIG_NUMA_EMU
 179         if (strcmp(parm, numa_mode_emu.name) == 0)
 180                 mode = &numa_mode_emu;
 181 #endif
 182         return 0;
 183 }
 184 early_param("numa", parse_numa);