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)
 130 {
 131         return pcpu_alloc_bootmem(cpu, size, size);
 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 ssize_t __init setup_pcpu_lpage(size_t static_size, bool chosen)
 153 {
 154         size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
 155
 156         if (!chosen) {
 157                 size_t vm_size = VMALLOC_END - VMALLOC_START;
 158                 size_t tot_size = num_possible_cpus() * PMD_SIZE;
 159
 160                 /* on non-NUMA, embedding is better */
 161                 if (!pcpu_need_numa())
 162                         return -EINVAL;
 163
 164                 /* don't consume more than 20% of vmalloc area */
 165                 if (tot_size > vm_size / 5) {
 166                         pr_info("PERCPU: too large chunk size %zuMB for "
 167                                 "large page remap\n", tot_size >> 20);
 168                         return -EINVAL;
 169                 }
 170         }
 171
 172         /* need PSE */
 173         if (!cpu_has_pse) {
 174                 pr_warning("PERCPU: lpage allocator requires PSE\n");
 175                 return -EINVAL;
 176         }
 177
 178         return pcpu_lpage_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE,
 179                                       reserve - PERCPU_FIRST_CHUNK_RESERVE,
 180                                       PMD_SIZE,
 181                                       pcpu_fc_alloc, pcpu_fc_free, pcpul_map);
 182 }
 183 #else
 184 static ssize_t __init setup_pcpu_lpage(size_t static_size, bool chosen)
 185 {
 186         return -EINVAL;
 187 }
 188 #endif
 189
 190 /*
 191  * Embedding allocator
 192  *
 193  * The first chunk is sized to just contain the static area plus
 194  * module and dynamic reserves and embedded into linear physical
 195  * mapping so that it can use PMD mapping without additional TLB
 196  * pressure.
 197  */
 198 static ssize_t __init setup_pcpu_embed(size_t static_size, bool chosen)
 199 {
 200         size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
 201
 202         /*
 203          * If large page isn't supported, there's no benefit in doing
 204          * this.  Also, embedding allocation doesn't play well with
 205          * NUMA.
 206          */
 207         if (!chosen && (!cpu_has_pse || pcpu_need_numa()))
 208                 return -EINVAL;
 209
 210         return pcpu_embed_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE,
 211                                       reserve - PERCPU_FIRST_CHUNK_RESERVE);
 212 }
 213
 214 /*
 215  * 4k allocator
 216  *
 217  * Boring fallback 4k allocator.  This allocator puts more pressure on
 218  * PTE TLBs but other than that behaves nicely on both UMA and NUMA.
 219  */
 220 static void __init pcpu4k_populate_pte(unsigned long addr)
 221 {
 222         populate_extra_pte(addr);
 223 }
 224
 225 static ssize_t __init setup_pcpu_4k(size_t static_size)
 226 {
 227         return pcpu_4k_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE,
 228                                    pcpu_fc_alloc, pcpu_fc_free,
 229                                    pcpu4k_populate_pte);
 230 }
 231
 232 /* for explicit first chunk allocator selection */
 233 static char pcpu_chosen_alloc[16] __initdata;
 234
 235 static int __init percpu_alloc_setup(char *str)
 236 {
 237         strncpy(pcpu_chosen_alloc, str, sizeof(pcpu_chosen_alloc) - 1);
 238         return 0;
 239 }
 240 early_param("percpu_alloc", percpu_alloc_setup);
 241
 242 static inline void setup_percpu_segment(int cpu)
 243 {
 244 #ifdef CONFIG_X86_32
 245         struct desc_struct gdt;
 246
 247         pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF,
 248                         0x2 | DESCTYPE_S, 0x8);
 249         gdt.s = 1;
 250         write_gdt_entry(get_cpu_gdt_table(cpu),
 251                         GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S);
 252 #endif
 253 }
 254
 255 void __init setup_per_cpu_areas(void)
 256 {
 257         size_t static_size = __per_cpu_end - __per_cpu_start;
 258         unsigned int cpu;
 259         unsigned long delta;
 260         size_t pcpu_unit_size;
 261         ssize_t ret;
 262
 263         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
 264                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 265
 266         /*
 267          * Allocate percpu area.  If PSE is supported, try to make use
 268          * of large page mappings.  Please read comments on top of
 269          * each allocator for details.
 270          */
 271         ret = -EINVAL;
 272         if (strlen(pcpu_chosen_alloc)) {
 273                 if (strcmp(pcpu_chosen_alloc, "4k")) {
 274                         if (!strcmp(pcpu_chosen_alloc, "lpage"))
 275                                 ret = setup_pcpu_lpage(static_size, true);
 276                         else if (!strcmp(pcpu_chosen_alloc, "embed"))
 277                                 ret = setup_pcpu_embed(static_size, true);
 278                         else
 279                                 pr_warning("PERCPU: unknown allocator %s "
 280                                            "specified\n", pcpu_chosen_alloc);
 281                         if (ret < 0)
 282                                 pr_warning("PERCPU: %s allocator failed (%zd), "
 283                                            "falling back to 4k\n",
 284                                            pcpu_chosen_alloc, ret);
 285                 }
 286         } else {
 287                 ret = setup_pcpu_lpage(static_size, false);
 288                 if (ret < 0)
 289                         ret = setup_pcpu_embed(static_size, false);
 290         }
 291         if (ret < 0)
 292                 ret = setup_pcpu_4k(static_size);
 293         if (ret < 0)
 294                 panic("cannot allocate static percpu area (%zu bytes, err=%zd)",
 295                       static_size, ret);
 296
 297         pcpu_unit_size = ret;
 298
 299         /* alrighty, percpu areas up and running */
 300         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 301         for_each_possible_cpu(cpu) {
 302                 per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size;
 303                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 304                 per_cpu(cpu_number, cpu) = cpu;
 305                 setup_percpu_segment(cpu);
 306                 setup_stack_canary_segment(cpu);
 307                 /*
 308                  * Copy data used in early init routines from the
 309                  * initial arrays to the per cpu data areas.  These
 310                  * arrays then become expendable and the *_early_ptr's
 311                  * are zeroed indicating that the static arrays are
 312                  * gone.
 313                  */
 314 #ifdef CONFIG_X86_LOCAL_APIC
 315                 per_cpu(x86_cpu_to_apicid, cpu) =
 316                         early_per_cpu_map(x86_cpu_to_apicid, cpu);
 317                 per_cpu(x86_bios_cpu_apicid, cpu) =
 318                         early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 319 #endif
 320 #ifdef CONFIG_X86_64
 321                 per_cpu(irq_stack_ptr, cpu) =
 322                         per_cpu(irq_stack_union.irq_stack, cpu) +
 323                         IRQ_STACK_SIZE - 64;
 324 #ifdef CONFIG_NUMA
 325                 per_cpu(x86_cpu_to_node_map, cpu) =
 326                         early_per_cpu_map(x86_cpu_to_node_map, cpu);
 327 #endif
 328 #endif
 329                 /*
 330                  * Up to this point, the boot CPU has been using .data.init
 331                  * area.  Reload any changed state for the boot CPU.
 332                  */
 333                 if (cpu == boot_cpu_id)
 334                         switch_to_new_gdt(cpu);
 335         }
 336
 337         /* indicate the early static arrays will soon be gone */
 338 #ifdef CONFIG_X86_LOCAL_APIC
 339         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
 340         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
 341 #endif
 342 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
 343         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 344 #endif
 345
 346 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
 347         /*
 348          * make sure boot cpu node_number is right, when boot cpu is on the
 349          * node that doesn't have mem installed
 350          */
 351         per_cpu(node_number, boot_cpu_id) = cpu_to_node(boot_cpu_id);
 352 #endif
 353
 354         /* Setup node to cpumask map */
 355         setup_node_to_cpumask_map();
 356
 357         /* Setup cpu initialized, callin, callout masks */
 358         setup_cpu_local_masks();
 359 }