arch/x86/kernel/setup_percpu.c

   1 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   2
   3 #include <linux/kernel.h>
   4 #include <linux/module.h>
   5 #include <linux/init.h>
   6 #include <linux/bootmem.h>
   7 #include <linux/percpu.h>
   8 #include <linux/kexec.h>
   9 #include <linux/crash_dump.h>
  10 #include <linux/smp.h>
  11 #include <linux/topology.h>
  12 #include <linux/pfn.h>
  13 #include <asm/sections.h>
  14 #include <asm/processor.h>
  15 #include <asm/setup.h>
  16 #include <asm/mpspec.h>
  17 #include <asm/apicdef.h>
  18 #include <asm/highmem.h>
  19 #include <asm/proto.h>
  20 #include <asm/cpumask.h>
  21 #include <asm/cpu.h>
  22 #include <asm/stackprotector.h>
  23
  24 DEFINE_PER_CPU_READ_MOSTLY(int, cpu_number);
  25 EXPORT_PER_CPU_SYMBOL(cpu_number);
  26
  27 #ifdef CONFIG_X86_64
  28 #define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
  29 #else
  30 #define BOOT_PERCPU_OFFSET 0
  31 #endif
  32
  33 DEFINE_PER_CPU(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
  34 EXPORT_PER_CPU_SYMBOL(this_cpu_off);
  35
  36 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly = {
  37         [0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
  38 };
  39 EXPORT_SYMBOL(__per_cpu_offset);
  40
  41 /*
  42  * On x86_64 symbols referenced from code should be reachable using
  43  * 32bit relocations.  Reserve space for static percpu variables in
  44  * modules so that they are always served from the first chunk which
  45  * is located at the percpu segment base.  On x86_32, anything can
  46  * address anywhere.  No need to reserve space in the first chunk.
  47  */
  48 #ifdef CONFIG_X86_64
  49 #define PERCPU_FIRST_CHUNK_RESERVE      PERCPU_MODULE_RESERVE
  50 #else
  51 #define PERCPU_FIRST_CHUNK_RESERVE      0
  52 #endif
  53
  54 #ifdef CONFIG_X86_32
  55 /**
  56  * pcpu_need_numa - determine percpu allocation needs to consider NUMA
  57  *
  58  * If NUMA is not configured or there is only one NUMA node available,
  59  * there is no reason to consider NUMA.  This function determines
  60  * whether percpu allocation should consider NUMA or not.
  61  *
  62  * RETURNS:
  63  * true if NUMA should be considered; otherwise, false.
  64  */
  65 static bool __init pcpu_need_numa(void)
  66 {
  67 #ifdef CONFIG_NEED_MULTIPLE_NODES
  68         pg_data_t *last = NULL;
  69         unsigned int cpu;
  70
  71         for_each_possible_cpu(cpu) {
  72                 int node = early_cpu_to_node(cpu);
  73
  74                 if (node_online(node) && NODE_DATA(node) &&
  75                     last && last != NODE_DATA(node))
  76                         return true;
  77
  78                 last = NODE_DATA(node);
  79         }
  80 #endif
  81         return false;
  82 }
  83 #endif
  84
  85 /**
  86  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
  87  * @cpu: cpu to allocate for
  88  * @size: size allocation in bytes
  89  * @align: alignment
  90  *
  91  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
  92  * does the right thing for NUMA regardless of the current
  93  * configuration.
  94  *
  95  * RETURNS:
  96  * Pointer to the allocated area on success, NULL on failure.
  97  */
  98 static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size,
  99                                         unsigned long align)
 100 {
 101         const unsigned long goal = __pa(MAX_DMA_ADDRESS);
 102 #ifdef CONFIG_NEED_MULTIPLE_NODES
 103         int node = early_cpu_to_node(cpu);
 104         void *ptr;
 105
 106         if (!node_online(node) || !NODE_DATA(node)) {
 107                 ptr = __alloc_bootmem_nopanic(size, align, goal);
 108                 pr_info("cpu %d has no node %d or node-local memory\n",
 109                         cpu, node);
 110                 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
 111                          cpu, size, __pa(ptr));
 112         } else {
 113                 ptr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
 114                                                    size, align, goal);
 115                 pr_debug("per cpu data for cpu%d %lu bytes on node%d at %016lx\n",
 116                          cpu, size, node, __pa(ptr));
 117         }
 118         return ptr;
 119 #else
 120         return __alloc_bootmem_nopanic(size, align, goal);
 121 #endif
 122 }
 123
 124 /*
 125  * Helpers for first chunk memory allocation
 126  */
 127 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
 128 {
 129         return pcpu_alloc_bootmem(cpu, size, align);
 130 }
 131
 132 static void __init pcpu_fc_free(void *ptr, size_t size)
 133 {
 134         free_bootmem(__pa(ptr), size);
 135 }
 136
 137 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
 138 {
 139 #ifdef CONFIG_NEED_MULTIPLE_NODES
 140         if (early_cpu_to_node(from) == early_cpu_to_node(to))
 141                 return LOCAL_DISTANCE;
 142         else
 143                 return REMOTE_DISTANCE;
 144 #else
 145         return LOCAL_DISTANCE;
 146 #endif
 147 }
 148
 149 static void __init pcpup_populate_pte(unsigned long addr)
 150 {
 151         populate_extra_pte(addr);
 152 }
 153
 154 static inline void setup_percpu_segment(int cpu)
 155 {
 156 #ifdef CONFIG_X86_32
 157         struct desc_struct gdt;
 158
 159         pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF,
 160                         0x2 | DESCTYPE_S, 0x8);
 161         gdt.s = 1;
 162         write_gdt_entry(get_cpu_gdt_table(cpu),
 163                         GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S);
 164 #endif
 165 }
 166
 167 void __init setup_per_cpu_areas(void)
 168 {
 169         unsigned int cpu;
 170         unsigned long delta;
 171         int rc;
 172
 173         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
 174                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 175
 176         /*
 177          * Allocate percpu area.  Embedding allocator is our favorite;
 178          * however, on NUMA configurations, it can result in very
 179          * sparse unit mapping and vmalloc area isn't spacious enough
 180          * on 32bit.  Use page in that case.
 181          */
 182 #ifdef CONFIG_X86_32
 183         if (pcpu_chosen_fc == PCPU_FC_AUTO && pcpu_need_numa())
 184                 pcpu_chosen_fc = PCPU_FC_PAGE;
 185 #endif
 186         rc = -EINVAL;
 187         if (pcpu_chosen_fc != PCPU_FC_PAGE) {
 188                 const size_t dyn_size = PERCPU_MODULE_RESERVE +
 189                         PERCPU_DYNAMIC_RESERVE - PERCPU_FIRST_CHUNK_RESERVE;
 190                 size_t atom_size;
 191
 192                 /*
 193                  * On 64bit, use PMD_SIZE for atom_size so that embedded
 194                  * percpu areas are aligned to PMD.  This, in the future,
 195                  * can also allow using PMD mappings in vmalloc area.  Use
 196                  * PAGE_SIZE on 32bit as vmalloc space is highly contended
 197                  * and large vmalloc area allocs can easily fail.
 198                  */
 199 #ifdef CONFIG_X86_64
 200                 atom_size = PMD_SIZE;
 201 #else
 202                 atom_size = PAGE_SIZE;
 203 #endif
 204                 rc = pcpu_embed_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
 205                                             dyn_size, atom_size,
 206                                             pcpu_cpu_distance,
 207                                             pcpu_fc_alloc, pcpu_fc_free);
 208                 if (rc < 0)
 209                         pr_warning("%s allocator failed (%d), falling back to page size\n",
 210                                    pcpu_fc_names[pcpu_chosen_fc], rc);
 211         }
 212         if (rc < 0)
 213                 rc = pcpu_page_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
 214                                            pcpu_fc_alloc, pcpu_fc_free,
 215                                            pcpup_populate_pte);
 216         if (rc < 0)
 217                 panic("cannot initialize percpu area (err=%d)", rc);
 218
 219         /* alrighty, percpu areas up and running */
 220         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 221         for_each_possible_cpu(cpu) {
 222                 per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu];
 223                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 224                 per_cpu(cpu_number, cpu) = cpu;
 225                 setup_percpu_segment(cpu);
 226                 setup_stack_canary_segment(cpu);
 227                 /*
 228                  * Copy data used in early init routines from the
 229                  * initial arrays to the per cpu data areas.  These
 230                  * arrays then become expendable and the *_early_ptr's
 231                  * are zeroed indicating that the static arrays are
 232                  * gone.
 233                  */
 234 #ifdef CONFIG_X86_LOCAL_APIC
 235                 per_cpu(x86_cpu_to_apicid, cpu) =
 236                         early_per_cpu_map(x86_cpu_to_apicid, cpu);
 237                 per_cpu(x86_bios_cpu_apicid, cpu) =
 238                         early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 239 #endif
 240 #ifdef CONFIG_X86_32
 241                 per_cpu(x86_cpu_to_logical_apicid, cpu) =
 242                         early_per_cpu_map(x86_cpu_to_logical_apicid, cpu);
 243 #endif
 244 #ifdef CONFIG_X86_64
 245                 per_cpu(irq_stack_ptr, cpu) =
 246                         per_cpu(irq_stack_union.irq_stack, cpu) +
 247                         IRQ_STACK_SIZE - 64;
 248 #endif
 249 #ifdef CONFIG_NUMA
 250                 per_cpu(x86_cpu_to_node_map, cpu) =
 251                         early_per_cpu_map(x86_cpu_to_node_map, cpu);
 252                 /*
 253                  * Ensure that the boot cpu numa_node is correct when the boot
 254                  * cpu is on a node that doesn't have memory installed.
 255                  * Also cpu_up() will call cpu_to_node() for APs when
 256                  * MEMORY_HOTPLUG is defined, before per_cpu(numa_node) is set
 257                  * up later with c_init aka intel_init/amd_init.
 258                  * So set them all (boot cpu and all APs).
 259                  */
 260                 set_cpu_numa_node(cpu, early_cpu_to_node(cpu));
 261 #endif
 262                 /*
 263                  * Up to this point, the boot CPU has been using .init.data
 264                  * area.  Reload any changed state for the boot CPU.
 265                  */
 266                 if (!cpu)
 267                         switch_to_new_gdt(cpu);
 268         }
 269
 270         /* indicate the early static arrays will soon be gone */
 271 #ifdef CONFIG_X86_LOCAL_APIC
 272         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
 273         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
 274 #endif
 275 #ifdef CONFIG_X86_32
 276         early_per_cpu_ptr(x86_cpu_to_logical_apicid) = NULL;
 277 #endif
 278 #ifdef CONFIG_NUMA
 279         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 280 #endif
 281
 282         /* Setup node to cpumask map */
 283         setup_node_to_cpumask_map();
 284
 285         /* Setup cpu initialized, callin, callout masks */
 286         setup_cpu_local_masks();
 287 }