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 static void plain_setup(void)
  30 {
  31         node_set(0, node_possible_map);
  32 }
  33
  34 const struct numa_mode numa_mode_plain = {
  35         .name = "plain",
  36         .setup = plain_setup,
  37 };
  38
  39 static const struct numa_mode *mode = &numa_mode_plain;
  40
  41 int numa_pfn_to_nid(unsigned long pfn)
  42 {
  43         return mode->__pfn_to_nid ? mode->__pfn_to_nid(pfn) : 0;
  44 }
  45
  46 void numa_update_cpu_topology(void)
  47 {
  48         if (mode->update_cpu_topology)
  49                 mode->update_cpu_topology();
  50 }
  51
  52 int __node_distance(int a, int b)
  53 {
  54         return mode->distance ? mode->distance(a, b) : 0;
  55 }
  56
  57 int numa_debug_enabled;
  58
  59 /*
  60  * alloc_node_data() - Allocate node data
  61  */
  62 static __init pg_data_t *alloc_node_data(void)
  63 {
  64         pg_data_t *res;
  65
  66         res = (pg_data_t *) memblock_alloc(sizeof(pg_data_t), 8);
  67         memset(res, 0, sizeof(pg_data_t));
  68         return res;
  69 }
  70
  71 /*
  72  * numa_setup_memory() - Assign bootmem to nodes
  73  *
  74  * The memory is first added to memblock without any respect to nodes.
  75  * This is fixed before remaining memblock memory is handed over to the
  76  * buddy allocator.
  77  * An important side effect is that large bootmem allocations might easily
  78  * cross node boundaries, which can be needed for large allocations with
  79  * smaller memory stripes in each node (i.e. when using NUMA emulation).
  80  *
  81  * Memory defines nodes:
  82  * Therefore this routine also sets the nodes online with memory.
  83  */
  84 static void __init numa_setup_memory(void)
  85 {
  86         unsigned long cur_base, align, end_of_dram;
  87         int nid = 0;
  88
  89         end_of_dram = memblock_end_of_DRAM();
  90         align = mode->align ? mode->align() : ULONG_MAX;
  91
  92         /*
  93          * Step through all available memory and assign it to the nodes
  94          * indicated by the mode implementation.
  95          * All nodes which are seen here will be set online.
  96          */
  97         cur_base = 0;
  98         do {
  99                 nid = numa_pfn_to_nid(PFN_DOWN(cur_base));
 100                 node_set_online(nid);
 101                 memblock_set_node(cur_base, align, &memblock.memory, nid);
 102                 cur_base += align;
 103         } while (cur_base < end_of_dram);
 104
 105         /* Allocate and fill out node_data */
 106         for (nid = 0; nid < MAX_NUMNODES; nid++)
 107                 NODE_DATA(nid) = alloc_node_data();
 108
 109         for_each_online_node(nid) {
 110                 unsigned long start_pfn, end_pfn;
 111                 unsigned long t_start, t_end;
 112                 int i;
 113
 114                 start_pfn = ULONG_MAX;
 115                 end_pfn = 0;
 116                 for_each_mem_pfn_range(i, nid, &t_start, &t_end, NULL) {
 117                         if (t_start < start_pfn)
 118                                 start_pfn = t_start;
 119                         if (t_end > end_pfn)
 120                                 end_pfn = t_end;
 121                 }
 122                 NODE_DATA(nid)->node_spanned_pages = end_pfn - start_pfn;
 123                 NODE_DATA(nid)->node_id = nid;
 124         }
 125 }
 126
 127 /*
 128  * numa_setup() - Earliest initialization
 129  *
 130  * Assign the mode and call the mode's setup routine.
 131  */
 132 void __init numa_setup(void)
 133 {
 134         pr_info("NUMA mode: %s\n", mode->name);
 135         nodes_clear(node_possible_map);
 136         if (mode->setup)
 137                 mode->setup();
 138         numa_setup_memory();
 139         memblock_dump_all();
 140 }
 141
 142 /*
 143  * numa_init_early() - Initialization initcall
 144  *
 145  * This runs when only one CPU is online and before the first
 146  * topology update is called for by the scheduler.
 147  */
 148 static int __init numa_init_early(void)
 149 {
 150         /* Attach all possible CPUs to node 0 for now. */
 151         cpumask_copy(&node_to_cpumask_map[0], cpu_possible_mask);
 152         return 0;
 153 }
 154 early_initcall(numa_init_early);
 155
 156 /*
 157  * numa_init_late() - Initialization initcall
 158  *
 159  * Register NUMA nodes.
 160  */
 161 static int __init numa_init_late(void)
 162 {
 163         int nid;
 164
 165         for_each_online_node(nid)
 166                 register_one_node(nid);
 167         return 0;
 168 }
 169 arch_initcall(numa_init_late);
 170
 171 static int __init parse_debug(char *parm)
 172 {
 173         numa_debug_enabled = 1;
 174         return 0;
 175 }
 176 early_param("numa_debug", parse_debug);
 177
 178 static int __init parse_numa(char *parm)
 179 {
 180         if (strcmp(parm, numa_mode_plain.name) == 0)
 181                 mode = &numa_mode_plain;
 182 #ifdef CONFIG_NUMA_EMU
 183         if (strcmp(parm, numa_mode_emu.name) == 0)
 184                 mode = &numa_mode_emu;
 185 #endif
 186         return 0;
 187 }
 188 early_param("numa", parse_numa);