arch/x86/kernel/setup_percpu.c

   1 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   2
   3 #include <linux/kernel.h>
   4 #include <linux/export.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/desc.h>
  16 #include <asm/setup.h>
  17 #include <asm/mpspec.h>
  18 #include <asm/apicdef.h>
  19 #include <asm/highmem.h>
  20 #include <asm/proto.h>
  21 #include <asm/cpumask.h>
  22 #include <asm/cpu.h>
  23 #include <asm/stackprotector.h>
  24
  25 DEFINE_PER_CPU_READ_MOSTLY(int, cpu_number);
  26 EXPORT_PER_CPU_SYMBOL(cpu_number);
  27
  28 #ifdef CONFIG_X86_64
  29 #define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
  30 #else
  31 #define BOOT_PERCPU_OFFSET 0
  32 #endif
  33
  34 DEFINE_PER_CPU_READ_MOSTLY(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
  35 EXPORT_PER_CPU_SYMBOL(this_cpu_off);
  36
  37 unsigned long __per_cpu_offset[NR_CPUS] __ro_after_init = {
  38         [0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
  39 };
  40 EXPORT_SYMBOL(__per_cpu_offset);
  41
  42 /*
  43  * On x86_64 symbols referenced from code should be reachable using
  44  * 32bit relocations.  Reserve space for static percpu variables in
  45  * modules so that they are always served from the first chunk which
  46  * is located at the percpu segment base.  On x86_32, anything can
  47  * address anywhere.  No need to reserve space in the first chunk.
  48  */
  49 #ifdef CONFIG_X86_64
  50 #define PERCPU_FIRST_CHUNK_RESERVE      PERCPU_MODULE_RESERVE
  51 #else
  52 #define PERCPU_FIRST_CHUNK_RESERVE      0
  53 #endif
  54
  55 #ifdef CONFIG_X86_32
  56 /**
  57  * pcpu_need_numa - determine percpu allocation needs to consider NUMA
  58  *
  59  * If NUMA is not configured or there is only one NUMA node available,
  60  * there is no reason to consider NUMA.  This function determines
  61  * whether percpu allocation should consider NUMA or not.
  62  *
  63  * RETURNS:
  64  * true if NUMA should be considered; otherwise, false.
  65  */
  66 static bool __init pcpu_need_numa(void)
  67 {
  68 #ifdef CONFIG_NEED_MULTIPLE_NODES
  69         pg_data_t *last = NULL;
  70         unsigned int cpu;
  71
  72         for_each_possible_cpu(cpu) {
  73                 int node = early_cpu_to_node(cpu);
  74
  75                 if (node_online(node) && NODE_DATA(node) &&
  76                     last && last != NODE_DATA(node))
  77                         return true;
  78
  79                 last = NODE_DATA(node);
  80         }
  81 #endif
  82         return false;
  83 }
  84 #endif
  85
  86 /**
  87  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
  88  * @cpu: cpu to allocate for
  89  * @size: size allocation in bytes
  90  * @align: alignment
  91  *
  92  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
  93  * does the right thing for NUMA regardless of the current
  94  * configuration.
  95  *
  96  * RETURNS:
  97  * Pointer to the allocated area on success, NULL on failure.
  98  */
  99 static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size,
 100                                         unsigned long align)
 101 {
 102         const unsigned long goal = __pa(MAX_DMA_ADDRESS);
 103 #ifdef CONFIG_NEED_MULTIPLE_NODES
 104         int node = early_cpu_to_node(cpu);
 105         void *ptr;
 106
 107         if (!node_online(node) || !NODE_DATA(node)) {
 108                 ptr = __alloc_bootmem_nopanic(size, align, goal);
 109                 pr_info("cpu %d has no node %d or node-local memory\n",
 110                         cpu, node);
 111                 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
 112                          cpu, size, __pa(ptr));
 113         } else {
 114                 ptr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
 115                                                    size, align, goal);
 116                 pr_debug("per cpu data for cpu%d %lu bytes on node%d at %016lx\n",
 117                          cpu, size, node, __pa(ptr));
 118         }
 119         return ptr;
 120 #else
 121         return __alloc_bootmem_nopanic(size, align, goal);
 122 #endif
 123 }
 124
 125 /*
 126  * Helpers for first chunk memory allocation
 127  */
 128 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
 129 {
 130         return pcpu_alloc_bootmem(cpu, size, align);
 131 }
 132
 133 static void __init pcpu_fc_free(void *ptr, size_t size)
 134 {
 135         free_bootmem(__pa(ptr), size);
 136 }
 137
 138 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
 139 {
 140 #ifdef CONFIG_NEED_MULTIPLE_NODES
 141         if (early_cpu_to_node(from) == early_cpu_to_node(to))
 142                 return LOCAL_DISTANCE;
 143         else
 144                 return REMOTE_DISTANCE;
 145 #else
 146         return LOCAL_DISTANCE;
 147 #endif
 148 }
 149
 150 static void __init pcpup_populate_pte(unsigned long addr)
 151 {
 152         populate_extra_pte(addr);
 153 }
 154
 155 static inline void setup_percpu_segment(int cpu)
 156 {
 157 #ifdef CONFIG_X86_32
 158         struct desc_struct d = GDT_ENTRY_INIT(0x8092, per_cpu_offset(cpu),
 159                                               0xFFFFF);
 160
 161         write_gdt_entry(get_cpu_gdt_rw(cpu), GDT_ENTRY_PERCPU, &d, DESCTYPE_S);
 162 #endif
 163 }
 164
 165 void __init setup_per_cpu_areas(void)
 166 {
 167         unsigned int cpu;
 168         unsigned long delta;
 169         int rc;
 170
 171         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%u nr_node_ids:%d\n",
 172                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 173
 174         /*
 175          * Allocate percpu area.  Embedding allocator is our favorite;
 176          * however, on NUMA configurations, it can result in very
 177          * sparse unit mapping and vmalloc area isn't spacious enough
 178          * on 32bit.  Use page in that case.
 179          */
 180 #ifdef CONFIG_X86_32
 181         if (pcpu_chosen_fc == PCPU_FC_AUTO && pcpu_need_numa())
 182                 pcpu_chosen_fc = PCPU_FC_PAGE;
 183 #endif
 184         rc = -EINVAL;
 185         if (pcpu_chosen_fc != PCPU_FC_PAGE) {
 186                 const size_t dyn_size = PERCPU_MODULE_RESERVE +
 187                         PERCPU_DYNAMIC_RESERVE - PERCPU_FIRST_CHUNK_RESERVE;
 188                 size_t atom_size;
 189
 190                 /*
 191                  * On 64bit, use PMD_SIZE for atom_size so that embedded
 192                  * percpu areas are aligned to PMD.  This, in the future,
 193                  * can also allow using PMD mappings in vmalloc area.  Use
 194                  * PAGE_SIZE on 32bit as vmalloc space is highly contended
 195                  * and large vmalloc area allocs can easily fail.
 196                  */
 197 #ifdef CONFIG_X86_64
 198                 atom_size = PMD_SIZE;
 199 #else
 200                 atom_size = PAGE_SIZE;
 201 #endif
 202                 rc = pcpu_embed_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
 203                                             dyn_size, atom_size,
 204                                             pcpu_cpu_distance,
 205                                             pcpu_fc_alloc, pcpu_fc_free);
 206                 if (rc < 0)
 207                         pr_warning("%s allocator failed (%d), falling back to page size\n",
 208                                    pcpu_fc_names[pcpu_chosen_fc], rc);
 209         }
 210         if (rc < 0)
 211                 rc = pcpu_page_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
 212                                            pcpu_fc_alloc, pcpu_fc_free,
 213                                            pcpup_populate_pte);
 214         if (rc < 0)
 215                 panic("cannot initialize percpu area (err=%d)", rc);
 216
 217         /* alrighty, percpu areas up and running */
 218         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 219         for_each_possible_cpu(cpu) {
 220                 per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu];
 221                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 222                 per_cpu(cpu_number, cpu) = cpu;
 223                 setup_percpu_segment(cpu);
 224                 setup_stack_canary_segment(cpu);
 225                 /*
 226                  * Copy data used in early init routines from the
 227                  * initial arrays to the per cpu data areas.  These
 228                  * arrays then become expendable and the *_early_ptr's
 229                  * are zeroed indicating that the static arrays are
 230                  * gone.
 231                  */
 232 #ifdef CONFIG_X86_LOCAL_APIC
 233                 per_cpu(x86_cpu_to_apicid, cpu) =
 234                         early_per_cpu_map(x86_cpu_to_apicid, cpu);
 235                 per_cpu(x86_bios_cpu_apicid, cpu) =
 236                         early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 237                 per_cpu(x86_cpu_to_acpiid, cpu) =
 238                         early_per_cpu_map(x86_cpu_to_acpiid, 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;
 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         early_per_cpu_ptr(x86_cpu_to_acpiid) = NULL;
 275 #endif
 276 #ifdef CONFIG_X86_32
 277         early_per_cpu_ptr(x86_cpu_to_logical_apicid) = NULL;
 278 #endif
 279 #ifdef CONFIG_NUMA
 280         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 281 #endif
 282
 283         /* Setup node to cpumask map */
 284         setup_node_to_cpumask_map();
 285
 286         /* Setup cpu initialized, callin, callout masks */
 287         setup_cpu_local_masks();
 288
 289 #ifdef CONFIG_X86_32
 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         clone_pgd_range(initial_page_table + KERNEL_PGD_BOUNDARY,
 298                         swapper_pg_dir     + KERNEL_PGD_BOUNDARY,
 299                         KERNEL_PGD_PTRS);
 300
 301         /*
 302          * sync back low identity map too.  It is used for example
 303          * in the 32-bit EFI stub.
 304          */
 305         clone_pgd_range(initial_page_table,
 306                         swapper_pg_dir     + KERNEL_PGD_BOUNDARY,
 307                         min(KERNEL_PGD_PTRS, KERNEL_PGD_BOUNDARY));
 308 #endif
 309 }