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/bootmem.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_NEED_MULTIPLE_NODES
  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 /**
  88  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
  89  * @cpu: cpu to allocate for
  90  * @size: size allocation in bytes
  91  * @align: alignment
  92  *
  93  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
  94  * does the right thing for NUMA regardless of the current
  95  * configuration.
  96  *
  97  * RETURNS:
  98  * Pointer to the allocated area on success, NULL on failure.
  99  */
 100 static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size,
 101                                         unsigned long align)
 102 {
 103         const unsigned long goal = __pa(MAX_DMA_ADDRESS);
 104 #ifdef CONFIG_NEED_MULTIPLE_NODES
 105         int node = early_cpu_to_node(cpu);
 106         void *ptr;
 107
 108         if (!node_online(node) || !NODE_DATA(node)) {
 109                 ptr = __alloc_bootmem_nopanic(size, align, goal);
 110                 pr_info("cpu %d has no node %d or node-local memory\n",
 111                         cpu, node);
 112                 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
 113                          cpu, size, __pa(ptr));
 114         } else {
 115                 ptr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
 116                                                    size, align, goal);
 117                 pr_debug("per cpu data for cpu%d %lu bytes on node%d at %016lx\n",
 118                          cpu, size, node, __pa(ptr));
 119         }
 120         return ptr;
 121 #else
 122         return __alloc_bootmem_nopanic(size, align, goal);
 123 #endif
 124 }
 125
 126 /*
 127  * Helpers for first chunk memory allocation
 128  */
 129 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
 130 {
 131         return pcpu_alloc_bootmem(cpu, size, align);
 132 }
 133
 134 static void __init pcpu_fc_free(void *ptr, size_t size)
 135 {
 136         free_bootmem(__pa(ptr), size);
 137 }
 138
 139 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
 140 {
 141 #ifdef CONFIG_NEED_MULTIPLE_NODES
 142         if (early_cpu_to_node(from) == early_cpu_to_node(to))
 143                 return LOCAL_DISTANCE;
 144         else
 145                 return REMOTE_DISTANCE;
 146 #else
 147         return LOCAL_DISTANCE;
 148 #endif
 149 }
 150
 151 static void __init pcpup_populate_pte(unsigned long addr)
 152 {
 153         populate_extra_pte(addr);
 154 }
 155
 156 static inline void setup_percpu_segment(int cpu)
 157 {
 158 #ifdef CONFIG_X86_32
 159         struct desc_struct d = GDT_ENTRY_INIT(0x8092, per_cpu_offset(cpu),
 160                                               0xFFFFF);
 161
 162         write_gdt_entry(get_cpu_gdt_rw(cpu), GDT_ENTRY_PERCPU, &d, DESCTYPE_S);
 163 #endif
 164 }
 165
 166 void __init setup_per_cpu_areas(void)
 167 {
 168         unsigned int cpu;
 169         unsigned long delta;
 170         int rc;
 171
 172         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%u nr_node_ids:%d\n",
 173                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 174
 175         /*
 176          * Allocate percpu area.  Embedding allocator is our favorite;
 177          * however, on NUMA configurations, it can result in very
 178          * sparse unit mapping and vmalloc area isn't spacious enough
 179          * on 32bit.  Use page in that case.
 180          */
 181 #ifdef CONFIG_X86_32
 182         if (pcpu_chosen_fc == PCPU_FC_AUTO && pcpu_need_numa())
 183                 pcpu_chosen_fc = PCPU_FC_PAGE;
 184 #endif
 185         rc = -EINVAL;
 186         if (pcpu_chosen_fc != PCPU_FC_PAGE) {
 187                 const size_t dyn_size = PERCPU_MODULE_RESERVE +
 188                         PERCPU_DYNAMIC_RESERVE - PERCPU_FIRST_CHUNK_RESERVE;
 189                 size_t atom_size;
 190
 191                 /*
 192                  * On 64bit, use PMD_SIZE for atom_size so that embedded
 193                  * percpu areas are aligned to PMD.  This, in the future,
 194                  * can also allow using PMD mappings in vmalloc area.  Use
 195                  * PAGE_SIZE on 32bit as vmalloc space is highly contended
 196                  * and large vmalloc area allocs can easily fail.
 197                  */
 198 #ifdef CONFIG_X86_64
 199                 atom_size = PMD_SIZE;
 200 #else
 201                 atom_size = PAGE_SIZE;
 202 #endif
 203                 rc = pcpu_embed_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
 204                                             dyn_size, atom_size,
 205                                             pcpu_cpu_distance,
 206                                             pcpu_fc_alloc, pcpu_fc_free);
 207                 if (rc < 0)
 208                         pr_warning("%s allocator failed (%d), falling back to page size\n",
 209                                    pcpu_fc_names[pcpu_chosen_fc], rc);
 210         }
 211         if (rc < 0)
 212                 rc = pcpu_page_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
 213                                            pcpu_fc_alloc, pcpu_fc_free,
 214                                            pcpup_populate_pte);
 215         if (rc < 0)
 216                 panic("cannot initialize percpu area (err=%d)", rc);
 217
 218         /* alrighty, percpu areas up and running */
 219         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 220         for_each_possible_cpu(cpu) {
 221                 per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu];
 222                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 223                 per_cpu(cpu_number, cpu) = cpu;
 224                 setup_percpu_segment(cpu);
 225                 setup_stack_canary_segment(cpu);
 226                 /*
 227                  * Copy data used in early init routines from the
 228                  * initial arrays to the per cpu data areas.  These
 229                  * arrays then become expendable and the *_early_ptr's
 230                  * are zeroed indicating that the static arrays are
 231                  * gone.
 232                  */
 233 #ifdef CONFIG_X86_LOCAL_APIC
 234                 per_cpu(x86_cpu_to_apicid, cpu) =
 235                         early_per_cpu_map(x86_cpu_to_apicid, cpu);
 236                 per_cpu(x86_bios_cpu_apicid, cpu) =
 237                         early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 238                 per_cpu(x86_cpu_to_acpiid, cpu) =
 239                         early_per_cpu_map(x86_cpu_to_acpiid, cpu);
 240 #endif
 241 #ifdef CONFIG_X86_32
 242                 per_cpu(x86_cpu_to_logical_apicid, cpu) =
 243                         early_per_cpu_map(x86_cpu_to_logical_apicid, cpu);
 244 #endif
 245 #ifdef CONFIG_X86_64
 246                 per_cpu(irq_stack_ptr, cpu) =
 247                         per_cpu(irq_stack_union.irq_stack, cpu) +
 248                         IRQ_STACK_SIZE;
 249 #endif
 250 #ifdef CONFIG_NUMA
 251                 per_cpu(x86_cpu_to_node_map, cpu) =
 252                         early_per_cpu_map(x86_cpu_to_node_map, cpu);
 253                 /*
 254                  * Ensure that the boot cpu numa_node is correct when the boot
 255                  * cpu is on a node that doesn't have memory installed.
 256                  * Also cpu_up() will call cpu_to_node() for APs when
 257                  * MEMORY_HOTPLUG is defined, before per_cpu(numa_node) is set
 258                  * up later with c_init aka intel_init/amd_init.
 259                  * So set them all (boot cpu and all APs).
 260                  */
 261                 set_cpu_numa_node(cpu, early_cpu_to_node(cpu));
 262 #endif
 263                 /*
 264                  * Up to this point, the boot CPU has been using .init.data
 265                  * area.  Reload any changed state for the boot CPU.
 266                  */
 267                 if (!cpu)
 268                         switch_to_new_gdt(cpu);
 269         }
 270
 271         /* indicate the early static arrays will soon be gone */
 272 #ifdef CONFIG_X86_LOCAL_APIC
 273         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
 274         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
 275         early_per_cpu_ptr(x86_cpu_to_acpiid) = NULL;
 276 #endif
 277 #ifdef CONFIG_X86_32
 278         early_per_cpu_ptr(x86_cpu_to_logical_apicid) = NULL;
 279 #endif
 280 #ifdef CONFIG_NUMA
 281         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 282 #endif
 283
 284         /* Setup node to cpumask map */
 285         setup_node_to_cpumask_map();
 286
 287         /* Setup cpu initialized, callin, callout masks */
 288         setup_cpu_local_masks();
 289
 290         /*
 291          * Sync back kernel address range again.  We already did this in
 292          * setup_arch(), but percpu data also needs to be available in
 293          * the smpboot asm.  We can't reliably pick up percpu mappings
 294          * using vmalloc_fault(), because exception dispatch needs
 295          * percpu data.
 296          *
 297          * FIXME: Can the later sync in setup_cpu_entry_areas() replace
 298          * this call?
 299          */
 300         sync_initial_page_table();
 301 }