arch/x86/kernel/setup_percpu.c

   1 #include <linux/kernel.h>
   2 #include <linux/module.h>
   3 #include <linux/init.h>
   4 #include <linux/bootmem.h>
   5 #include <linux/percpu.h>
   6 #include <linux/kexec.h>
   7 #include <linux/crash_dump.h>
   8 #include <linux/smp.h>
   9 #include <linux/topology.h>
  10 #include <linux/pfn.h>
  11 #include <asm/sections.h>
  12 #include <asm/processor.h>
  13 #include <asm/setup.h>
  14 #include <asm/mpspec.h>
  15 #include <asm/apicdef.h>
  16 #include <asm/highmem.h>
  17 #include <asm/proto.h>
  18 #include <asm/cpumask.h>
  19 #include <asm/cpu.h>
  20 #include <asm/stackprotector.h>
  21
  22 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
  23 # define DBG(x...) printk(KERN_DEBUG x)
  24 #else
  25 # define DBG(x...)
  26 #endif
  27
  28 DEFINE_PER_CPU(int, cpu_number);
  29 EXPORT_PER_CPU_SYMBOL(cpu_number);
  30
  31 #ifdef CONFIG_X86_64
  32 #define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
  33 #else
  34 #define BOOT_PERCPU_OFFSET 0
  35 #endif
  36
  37 DEFINE_PER_CPU(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
  38 EXPORT_PER_CPU_SYMBOL(this_cpu_off);
  39
  40 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly = {
  41         [0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
  42 };
  43 EXPORT_SYMBOL(__per_cpu_offset);
  44
  45 /*
  46  * On x86_64 symbols referenced from code should be reachable using
  47  * 32bit relocations.  Reserve space for static percpu variables in
  48  * modules so that they are always served from the first chunk which
  49  * is located at the percpu segment base.  On x86_32, anything can
  50  * address anywhere.  No need to reserve space in the first chunk.
  51  */
  52 #ifdef CONFIG_X86_64
  53 #define PERCPU_FIRST_CHUNK_RESERVE      PERCPU_MODULE_RESERVE
  54 #else
  55 #define PERCPU_FIRST_CHUNK_RESERVE      0
  56 #endif
  57
  58 /**
  59  * pcpu_need_numa - determine percpu allocation needs to consider NUMA
  60  *
  61  * If NUMA is not configured or there is only one NUMA node available,
  62  * there is no reason to consider NUMA.  This function determines
  63  * whether percpu allocation should consider NUMA or not.
  64  *
  65  * RETURNS:
  66  * true if NUMA should be considered; otherwise, false.
  67  */
  68 static bool __init pcpu_need_numa(void)
  69 {
  70 #ifdef CONFIG_NEED_MULTIPLE_NODES
  71         pg_data_t *last = NULL;
  72         unsigned int cpu;
  73
  74         for_each_possible_cpu(cpu) {
  75                 int node = early_cpu_to_node(cpu);
  76
  77                 if (node_online(node) && NODE_DATA(node) &&
  78                     last && last != NODE_DATA(node))
  79                         return true;
  80
  81                 last = NODE_DATA(node);
  82         }
  83 #endif
  84         return false;
  85 }
  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 "
 118                          "%016lx\n", 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 /*
 140  * Large page remapping allocator
 141  */
 142 #ifdef CONFIG_NEED_MULTIPLE_NODES
 143 static void __init pcpul_map(void *ptr, size_t size, void *addr)
 144 {
 145         pmd_t *pmd, pmd_v;
 146
 147         pmd = populate_extra_pmd((unsigned long)addr);
 148         pmd_v = pfn_pmd(page_to_pfn(virt_to_page(ptr)), PAGE_KERNEL_LARGE);
 149         set_pmd(pmd, pmd_v);
 150 }
 151
 152 static int pcpu_lpage_cpu_distance(unsigned int from, unsigned int to)
 153 {
 154         if (early_cpu_to_node(from) == early_cpu_to_node(to))
 155                 return LOCAL_DISTANCE;
 156         else
 157                 return REMOTE_DISTANCE;
 158 }
 159
 160 static ssize_t __init setup_pcpu_lpage(bool chosen)
 161 {
 162         size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
 163         size_t dyn_size = reserve - PERCPU_FIRST_CHUNK_RESERVE;
 164         struct pcpu_alloc_info *ai;
 165         ssize_t ret;
 166
 167         /* on non-NUMA, embedding is better */
 168         if (!chosen && !pcpu_need_numa())
 169                 return -EINVAL;
 170
 171         /* need PSE */
 172         if (!cpu_has_pse) {
 173                 pr_warning("PERCPU: lpage allocator requires PSE\n");
 174                 return -EINVAL;
 175         }
 176
 177         /* allocate and build unit_map */
 178         ai = pcpu_build_alloc_info(PERCPU_FIRST_CHUNK_RESERVE, dyn_size,
 179                                    PMD_SIZE, pcpu_lpage_cpu_distance);
 180         if (IS_ERR(ai)) {
 181                 pr_warning("PERCPU: failed to build unit_map (%ld)\n",
 182                            PTR_ERR(ai));
 183                 return PTR_ERR(ai);
 184         }
 185
 186         /* do the parameters look okay? */
 187         if (!chosen) {
 188                 size_t vm_size = VMALLOC_END - VMALLOC_START;
 189                 size_t tot_size = 0;
 190                 int group;
 191
 192                 for (group = 0; group < ai->nr_groups; group++)
 193                         tot_size += ai->unit_size * ai->groups[group].nr_units;
 194
 195                 /* don't consume more than 20% of vmalloc area */
 196                 if (tot_size > vm_size / 5) {
 197                         pr_info("PERCPU: too large chunk size %zuMB for "
 198                                 "large page remap\n", tot_size >> 20);
 199                         ret = -EINVAL;
 200                         goto out_free;
 201                 }
 202         }
 203
 204         ret = pcpu_lpage_first_chunk(ai, pcpu_fc_alloc, pcpu_fc_free,
 205                                      pcpul_map);
 206 out_free:
 207         pcpu_free_alloc_info(ai);
 208         return ret;
 209 }
 210 #else
 211 static ssize_t __init setup_pcpu_lpage(bool chosen)
 212 {
 213         return -EINVAL;
 214 }
 215 #endif
 216
 217 /*
 218  * Embedding allocator
 219  *
 220  * The first chunk is sized to just contain the static area plus
 221  * module and dynamic reserves and embedded into linear physical
 222  * mapping so that it can use PMD mapping without additional TLB
 223  * pressure.
 224  */
 225 static ssize_t __init setup_pcpu_embed(bool chosen)
 226 {
 227         size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
 228
 229         /*
 230          * If large page isn't supported, there's no benefit in doing
 231          * this.  Also, embedding allocation doesn't play well with
 232          * NUMA.
 233          */
 234         if (!chosen && (!cpu_has_pse || pcpu_need_numa()))
 235                 return -EINVAL;
 236
 237         return pcpu_embed_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
 238                                       reserve - PERCPU_FIRST_CHUNK_RESERVE);
 239 }
 240
 241 /*
 242  * Page allocator
 243  *
 244  * Boring fallback 4k page allocator.  This allocator puts more
 245  * pressure on PTE TLBs but other than that behaves nicely on both UMA
 246  * and NUMA.
 247  */
 248 static void __init pcpup_populate_pte(unsigned long addr)
 249 {
 250         populate_extra_pte(addr);
 251 }
 252
 253 static ssize_t __init setup_pcpu_page(void)
 254 {
 255         return pcpu_page_first_chunk(PERCPU_FIRST_CHUNK_RESERVE,
 256                                      pcpu_fc_alloc, pcpu_fc_free,
 257                                      pcpup_populate_pte);
 258 }
 259
 260 static inline void setup_percpu_segment(int cpu)
 261 {
 262 #ifdef CONFIG_X86_32
 263         struct desc_struct gdt;
 264
 265         pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF,
 266                         0x2 | DESCTYPE_S, 0x8);
 267         gdt.s = 1;
 268         write_gdt_entry(get_cpu_gdt_table(cpu),
 269                         GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S);
 270 #endif
 271 }
 272
 273 void __init setup_per_cpu_areas(void)
 274 {
 275         unsigned int cpu;
 276         unsigned long delta;
 277         size_t pcpu_unit_size;
 278         ssize_t ret;
 279
 280         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
 281                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 282
 283         /*
 284          * Allocate percpu area.  If PSE is supported, try to make use
 285          * of large page mappings.  Please read comments on top of
 286          * each allocator for details.
 287          */
 288         ret = -EINVAL;
 289         if (pcpu_chosen_fc != PCPU_FC_AUTO) {
 290                 if (pcpu_chosen_fc != PCPU_FC_PAGE) {
 291                         if (pcpu_chosen_fc == PCPU_FC_LPAGE)
 292                                 ret = setup_pcpu_lpage(true);
 293                         else
 294                                 ret = setup_pcpu_embed(true);
 295
 296                         if (ret < 0)
 297                                 pr_warning("PERCPU: %s allocator failed (%zd), "
 298                                            "falling back to page size\n",
 299                                            pcpu_fc_names[pcpu_chosen_fc], ret);
 300                 }
 301         } else {
 302                 ret = setup_pcpu_lpage(false);
 303                 if (ret < 0)
 304                         ret = setup_pcpu_embed(false);
 305         }
 306         if (ret < 0)
 307                 ret = setup_pcpu_page();
 308         if (ret < 0)
 309                 panic("cannot initialize percpu area (err=%zd)", ret);
 310
 311         pcpu_unit_size = ret;
 312
 313         /* alrighty, percpu areas up and running */
 314         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 315         for_each_possible_cpu(cpu) {
 316                 per_cpu_offset(cpu) =
 317                         delta + pcpu_unit_map[cpu] * pcpu_unit_size;
 318                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 319                 per_cpu(cpu_number, cpu) = cpu;
 320                 setup_percpu_segment(cpu);
 321                 setup_stack_canary_segment(cpu);
 322                 /*
 323                  * Copy data used in early init routines from the
 324                  * initial arrays to the per cpu data areas.  These
 325                  * arrays then become expendable and the *_early_ptr's
 326                  * are zeroed indicating that the static arrays are
 327                  * gone.
 328                  */
 329 #ifdef CONFIG_X86_LOCAL_APIC
 330                 per_cpu(x86_cpu_to_apicid, cpu) =
 331                         early_per_cpu_map(x86_cpu_to_apicid, cpu);
 332                 per_cpu(x86_bios_cpu_apicid, cpu) =
 333                         early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 334 #endif
 335 #ifdef CONFIG_X86_64
 336                 per_cpu(irq_stack_ptr, cpu) =
 337                         per_cpu(irq_stack_union.irq_stack, cpu) +
 338                         IRQ_STACK_SIZE - 64;
 339 #ifdef CONFIG_NUMA
 340                 per_cpu(x86_cpu_to_node_map, cpu) =
 341                         early_per_cpu_map(x86_cpu_to_node_map, cpu);
 342 #endif
 343 #endif
 344                 /*
 345                  * Up to this point, the boot CPU has been using .data.init
 346                  * area.  Reload any changed state for the boot CPU.
 347                  */
 348                 if (cpu == boot_cpu_id)
 349                         switch_to_new_gdt(cpu);
 350         }
 351
 352         /* indicate the early static arrays will soon be gone */
 353 #ifdef CONFIG_X86_LOCAL_APIC
 354         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
 355         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
 356 #endif
 357 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
 358         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 359 #endif
 360
 361 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
 362         /*
 363          * make sure boot cpu node_number is right, when boot cpu is on the
 364          * node that doesn't have mem installed
 365          */
 366         per_cpu(node_number, boot_cpu_id) = cpu_to_node(boot_cpu_id);
 367 #endif
 368
 369         /* Setup node to cpumask map */
 370         setup_node_to_cpumask_map();
 371
 372         /* Setup cpu initialized, callin, callout masks */
 373         setup_cpu_local_masks();
 374 }