arch/x86/kernel/setup_percpu.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   3
   4 #include <linux/kernel.h>
   5 #include <linux/export.h>
   6 #include <linux/init.h>
   7 #include <linux/memblock.h>
   8 #include <linux/percpu.h>
   9 #include <linux/kexec.h>
  10 #include <linux/crash_dump.h>
  11 #include <linux/smp.h>
  12 #include <linux/topology.h>
  13 #include <linux/pfn.h>
  14 #include <asm/sections.h>
  15 #include <asm/processor.h>
  16 #include <asm/desc.h>
  17 #include <asm/setup.h>
  18 #include <asm/mpspec.h>
  19 #include <asm/apicdef.h>
  20 #include <asm/highmem.h>
  21 #include <asm/proto.h>
  22 #include <asm/cpumask.h>
  23 #include <asm/cpu.h>
  24 #include <asm/stackprotector.h>
  25
  26 DEFINE_PER_CPU_READ_MOSTLY(int, cpu_number);
  27 EXPORT_PER_CPU_SYMBOL(cpu_number);
  28
  29 #ifdef CONFIG_X86_64
  30 #define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
  31 #else
  32 #define BOOT_PERCPU_OFFSET 0
  33 #endif
  34
  35 DEFINE_PER_CPU_READ_MOSTLY(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
  36 EXPORT_PER_CPU_SYMBOL(this_cpu_off);
  37
  38 unsigned long __per_cpu_offset[NR_CPUS] __ro_after_init = {
  39         [0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
  40 };
  41 EXPORT_SYMBOL(__per_cpu_offset);
  42
  43 /*
  44  * On x86_64 symbols referenced from code should be reachable using
  45  * 32bit relocations.  Reserve space for static percpu variables in
  46  * modules so that they are always served from the first chunk which
  47  * is located at the percpu segment base.  On x86_32, anything can
  48  * address anywhere.  No need to reserve space in the first chunk.
  49  */
  50 #ifdef CONFIG_X86_64
  51 #define PERCPU_FIRST_CHUNK_RESERVE      PERCPU_MODULE_RESERVE
  52 #else
  53 #define PERCPU_FIRST_CHUNK_RESERVE      0
  54 #endif
  55
  56 #ifdef CONFIG_X86_32
  57 /**
  58  * pcpu_need_numa - determine percpu allocation needs to consider NUMA
  59  *
  60  * If NUMA is not configured or there is only one NUMA node available,
  61  * there is no reason to consider NUMA.  This function determines
  62  * whether percpu allocation should consider NUMA or not.
  63  *
  64  * RETURNS:
  65  * true if NUMA should be considered; otherwise, false.
  66  */
  67 static bool __init pcpu_need_numa(void)
  68 {
  69 #ifdef CONFIG_NUMA
  70         pg_data_t *last = NULL;
  71         unsigned int cpu;
  72
  73         for_each_possible_cpu(cpu) {
  74                 int node = early_cpu_to_node(cpu);
  75
  76                 if (node_online(node) && NODE_DATA(node) &&
  77                     last && last != NODE_DATA(node))
  78                         return true;
  79
  80                 last = NODE_DATA(node);
  81         }
  82 #endif
  83         return false;
  84 }
  85 #endif
  86
  87 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
  88 {
  89 #ifdef CONFIG_NUMA
  90         if (early_cpu_to_node(from) == early_cpu_to_node(to))
  91                 return LOCAL_DISTANCE;
  92         else
  93                 return REMOTE_DISTANCE;
  94 #else
  95         return LOCAL_DISTANCE;
  96 #endif
  97 }
  98
  99 static int __init pcpu_cpu_to_node(int cpu)
 100 {
 101         return early_cpu_to_node(cpu);
 102 }
 103
 104 void __init pcpu_populate_pte(unsigned long addr)
 105 {
 106         populate_extra_pte(addr);
 107 }
 108
 109 static inline void setup_percpu_segment(int cpu)
 110 {
 111 #ifdef CONFIG_X86_32
 112         struct desc_struct d = GDT_ENTRY_INIT(0x8092, per_cpu_offset(cpu),
 113                                               0xFFFFF);
 114
 115         write_gdt_entry(get_cpu_gdt_rw(cpu), GDT_ENTRY_PERCPU, &d, DESCTYPE_S);
 116 #endif
 117 }
 118
 119 void __init setup_per_cpu_areas(void)
 120 {
 121         unsigned int cpu;
 122         unsigned long delta;
 123         int rc;
 124
 125         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%u nr_node_ids:%u\n",
 126                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 127
 128         /*
 129          * Allocate percpu area.  Embedding allocator is our favorite;
 130          * however, on NUMA configurations, it can result in very
 131          * sparse unit mapping and vmalloc area isn't spacious enough
 132          * on 32bit.  Use page in that case.
 133          */
 134 #ifdef CONFIG_X86_32
 135         if (pcpu_chosen_fc == PCPU_FC_AUTO && pcpu_need_numa())
 136                 pcpu_chosen_fc = PCPU_FC_PAGE;
 137 #endif
 138         rc = -EINVAL;
 139         if (pcpu_chosen_fc != PCPU_FC_PAGE) {
 140                 const size_t dyn_size = PERCPU_MODULE_RESERVE +
 141                         PERCPU_DYNAMIC_RESERVE - PERCPU_FIRST_CHUNK_RESERVE;
 142                 size_t atom_size;
 143
 144                 /*
 145                  * On 64bit, use PMD_SIZE for atom_size so that embedded
 146                  * percpu areas are aligned to PMD.  This, in the future,
 147                  * can also allow using PMD mappings in vmalloc area.  Use
 148                  * PAGE_SIZE on 32bit as vmalloc space is highly contended
 149                  * and large vmalloc area allocs can easily fail.
 150                  */
 151 #ifdef CONFIG_X86_64
 152                 atom_size = PMD_SIZE;
 153 #else
 154                 atom_size = PAGE_SIZE;
 155 #endif
 156                 rc = pcpu_embed_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
 157                                             dyn_size, atom_size,
 158                                             pcpu_cpu_distance,
 159                                             pcpu_cpu_to_node);
 160                 if (rc < 0)
 161                         pr_warn("%s allocator failed (%d), falling back to page size\n",
 162                                 pcpu_fc_names[pcpu_chosen_fc], rc);
 163         }
 164         if (rc < 0)
 165                 rc = pcpu_page_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
 166                                            pcpu_cpu_to_node);
 167         if (rc < 0)
 168                 panic("cannot initialize percpu area (err=%d)", rc);
 169
 170         /* alrighty, percpu areas up and running */
 171         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 172         for_each_possible_cpu(cpu) {
 173                 per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu];
 174                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 175                 per_cpu(cpu_number, cpu) = cpu;
 176                 setup_percpu_segment(cpu);
 177                 /*
 178                  * Copy data used in early init routines from the
 179                  * initial arrays to the per cpu data areas.  These
 180                  * arrays then become expendable and the *_early_ptr's
 181                  * are zeroed indicating that the static arrays are
 182                  * gone.
 183                  */
 184 #ifdef CONFIG_X86_LOCAL_APIC
 185                 per_cpu(x86_cpu_to_apicid, cpu) =
 186                         early_per_cpu_map(x86_cpu_to_apicid, cpu);
 187                 per_cpu(x86_bios_cpu_apicid, cpu) =
 188                         early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 189                 per_cpu(x86_cpu_to_acpiid, cpu) =
 190                         early_per_cpu_map(x86_cpu_to_acpiid, cpu);
 191 #endif
 192 #ifdef CONFIG_X86_32
 193                 per_cpu(x86_cpu_to_logical_apicid, cpu) =
 194                         early_per_cpu_map(x86_cpu_to_logical_apicid, cpu);
 195 #endif
 196 #ifdef CONFIG_NUMA
 197                 per_cpu(x86_cpu_to_node_map, cpu) =
 198                         early_per_cpu_map(x86_cpu_to_node_map, cpu);
 199                 /*
 200                  * Ensure that the boot cpu numa_node is correct when the boot
 201                  * cpu is on a node that doesn't have memory installed.
 202                  * Also cpu_up() will call cpu_to_node() for APs when
 203                  * MEMORY_HOTPLUG is defined, before per_cpu(numa_node) is set
 204                  * up later with c_init aka intel_init/amd_init.
 205                  * So set them all (boot cpu and all APs).
 206                  */
 207                 set_cpu_numa_node(cpu, early_cpu_to_node(cpu));
 208 #endif
 209                 /*
 210                  * Up to this point, the boot CPU has been using .init.data
 211                  * area.  Reload any changed state for the boot CPU.
 212                  */
 213                 if (!cpu)
 214                         switch_to_new_gdt(cpu);
 215         }
 216
 217         /* indicate the early static arrays will soon be gone */
 218 #ifdef CONFIG_X86_LOCAL_APIC
 219         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
 220         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
 221         early_per_cpu_ptr(x86_cpu_to_acpiid) = NULL;
 222 #endif
 223 #ifdef CONFIG_X86_32
 224         early_per_cpu_ptr(x86_cpu_to_logical_apicid) = NULL;
 225 #endif
 226 #ifdef CONFIG_NUMA
 227         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 228 #endif
 229
 230         /* Setup node to cpumask map */
 231         setup_node_to_cpumask_map();
 232
 233         /* Setup cpu initialized, callin, callout masks */
 234         setup_cpu_local_masks();
 235
 236         /*
 237          * Sync back kernel address range again.  We already did this in
 238          * setup_arch(), but percpu data also needs to be available in
 239          * the smpboot asm and arch_sync_kernel_mappings() doesn't sync to
 240          * swapper_pg_dir on 32-bit. The per-cpu mappings need to be available
 241          * there too.
 242          *
 243          * FIXME: Can the later sync in setup_cpu_entry_areas() replace
 244          * this call?
 245          */
 246         sync_initial_page_table();
 247 }