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  * Large page remap allocator
 128  *
 129  * This allocator uses PMD page as unit.  A PMD page is allocated for
 130  * each cpu and each is remapped into vmalloc area using PMD mapping.
 131  * As PMD page is quite large, only part of it is used for the first
 132  * chunk.  Unused part is returned to the bootmem allocator.
 133  *
 134  * So, the PMD pages are mapped twice - once to the physical mapping
 135  * and to the vmalloc area for the first percpu chunk.  The double
 136  * mapping does add one more PMD TLB entry pressure but still is much
 137  * better than only using 4k mappings while still being NUMA friendly.
 138  */
 139 #ifdef CONFIG_NEED_MULTIPLE_NODES
 140 struct pcpul_ent {
 141         unsigned int    cpu;
 142         void            *ptr;
 143 };
 144
 145 static size_t pcpul_size __initdata;
 146 static struct pcpul_ent *pcpul_map __initdata;
 147 static struct vm_struct pcpul_vm;
 148
 149 static struct page * __init pcpul_get_page(unsigned int cpu, int pageno)
 150 {
 151         size_t off = (size_t)pageno << PAGE_SHIFT;
 152
 153         if (off >= pcpul_size)
 154                 return NULL;
 155
 156         return virt_to_page(pcpul_map[cpu].ptr + off);
 157 }
 158
 159 static ssize_t __init setup_pcpu_lpage(size_t static_size)
 160 {
 161         size_t map_size, dyn_size;
 162         unsigned int cpu;
 163         ssize_t ret;
 164
 165         /*
 166          * If large page isn't supported, there's no benefit in doing
 167          * this.  Also, on non-NUMA, embedding is better.
 168          *
 169          * NOTE: disabled for now.
 170          */
 171         if (true || !cpu_has_pse || !pcpu_need_numa())
 172                 return -EINVAL;
 173
 174         /*
 175          * Currently supports only single page.  Supporting multiple
 176          * pages won't be too difficult if it ever becomes necessary.
 177          */
 178         pcpul_size = PFN_ALIGN(static_size + PERCPU_MODULE_RESERVE +
 179                                PERCPU_DYNAMIC_RESERVE);
 180         if (pcpul_size > PMD_SIZE) {
 181                 pr_warning("PERCPU: static data is larger than large page, "
 182                            "can't use large page\n");
 183                 return -EINVAL;
 184         }
 185         dyn_size = pcpul_size - static_size - PERCPU_FIRST_CHUNK_RESERVE;
 186
 187         /* allocate pointer array and alloc large pages */
 188         map_size = PFN_ALIGN(num_possible_cpus() * sizeof(pcpul_map[0]));
 189         pcpul_map = alloc_bootmem(map_size);
 190
 191         for_each_possible_cpu(cpu) {
 192                 pcpul_map[cpu].cpu = cpu;
 193                 pcpul_map[cpu].ptr = pcpu_alloc_bootmem(cpu, PMD_SIZE,
 194                                                         PMD_SIZE);
 195                 if (!pcpul_map[cpu].ptr)
 196                         goto enomem;
 197
 198                 /*
 199                  * Only use pcpul_size bytes and give back the rest.
 200                  *
 201                  * Ingo: The 2MB up-rounding bootmem is needed to make
 202                  * sure the partial 2MB page is still fully RAM - it's
 203                  * not well-specified to have a PAT-incompatible area
 204                  * (unmapped RAM, device memory, etc.) in that hole.
 205                  */
 206                 free_bootmem(__pa(pcpul_map[cpu].ptr + pcpul_size),
 207                              PMD_SIZE - pcpul_size);
 208
 209                 memcpy(pcpul_map[cpu].ptr, __per_cpu_load, static_size);
 210         }
 211
 212         /* allocate address and map */
 213         pcpul_vm.flags = VM_ALLOC;
 214         pcpul_vm.size = num_possible_cpus() * PMD_SIZE;
 215         vm_area_register_early(&pcpul_vm, PMD_SIZE);
 216
 217         for_each_possible_cpu(cpu) {
 218                 pmd_t *pmd, pmd_v;
 219
 220                 pmd = populate_extra_pmd((unsigned long)pcpul_vm.addr +
 221                                          cpu * PMD_SIZE);
 222                 pmd_v = pfn_pmd(page_to_pfn(virt_to_page(pcpul_map[cpu].ptr)),
 223                                 PAGE_KERNEL_LARGE);
 224                 set_pmd(pmd, pmd_v);
 225         }
 226
 227         /* we're ready, commit */
 228         pr_info("PERCPU: Remapped at %p with large pages, static data "
 229                 "%zu bytes\n", pcpul_vm.addr, static_size);
 230
 231         ret = pcpu_setup_first_chunk(pcpul_get_page, static_size,
 232                                      PERCPU_FIRST_CHUNK_RESERVE, dyn_size,
 233                                      PMD_SIZE, pcpul_vm.addr, NULL);
 234         goto out_free_map;
 235
 236 enomem:
 237         for_each_possible_cpu(cpu)
 238                 if (pcpul_map[cpu].ptr)
 239                         free_bootmem(__pa(pcpul_map[cpu].ptr), pcpul_size);
 240         ret = -ENOMEM;
 241 out_free_map:
 242         free_bootmem(__pa(pcpul_map), map_size);
 243         return ret;
 244 }
 245 #else
 246 static ssize_t __init setup_pcpu_lpage(size_t static_size)
 247 {
 248         return -EINVAL;
 249 }
 250 #endif
 251
 252 /*
 253  * Embedding allocator
 254  *
 255  * The first chunk is sized to just contain the static area plus
 256  * module and dynamic reserves and embedded into linear physical
 257  * mapping so that it can use PMD mapping without additional TLB
 258  * pressure.
 259  */
 260 static ssize_t __init setup_pcpu_embed(size_t static_size)
 261 {
 262         size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
 263
 264         /*
 265          * If large page isn't supported, there's no benefit in doing
 266          * this.  Also, embedding allocation doesn't play well with
 267          * NUMA.
 268          */
 269         if (!cpu_has_pse || pcpu_need_numa())
 270                 return -EINVAL;
 271
 272         return pcpu_embed_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE,
 273                                       reserve - PERCPU_FIRST_CHUNK_RESERVE, -1);
 274 }
 275
 276 /*
 277  * 4k page allocator
 278  *
 279  * This is the basic allocator.  Static percpu area is allocated
 280  * page-by-page and most of initialization is done by the generic
 281  * setup function.
 282  */
 283 static struct page **pcpu4k_pages __initdata;
 284 static int pcpu4k_nr_static_pages __initdata;
 285
 286 static struct page * __init pcpu4k_get_page(unsigned int cpu, int pageno)
 287 {
 288         if (pageno < pcpu4k_nr_static_pages)
 289                 return pcpu4k_pages[cpu * pcpu4k_nr_static_pages + pageno];
 290         return NULL;
 291 }
 292
 293 static void __init pcpu4k_populate_pte(unsigned long addr)
 294 {
 295         populate_extra_pte(addr);
 296 }
 297
 298 static ssize_t __init setup_pcpu_4k(size_t static_size)
 299 {
 300         size_t pages_size;
 301         unsigned int cpu;
 302         int i, j;
 303         ssize_t ret;
 304
 305         pcpu4k_nr_static_pages = PFN_UP(static_size);
 306
 307         /* unaligned allocations can't be freed, round up to page size */
 308         pages_size = PFN_ALIGN(pcpu4k_nr_static_pages * num_possible_cpus()
 309                                * sizeof(pcpu4k_pages[0]));
 310         pcpu4k_pages = alloc_bootmem(pages_size);
 311
 312         /* allocate and copy */
 313         j = 0;
 314         for_each_possible_cpu(cpu)
 315                 for (i = 0; i < pcpu4k_nr_static_pages; i++) {
 316                         void *ptr;
 317
 318                         ptr = pcpu_alloc_bootmem(cpu, PAGE_SIZE, PAGE_SIZE);
 319                         if (!ptr)
 320                                 goto enomem;
 321
 322                         memcpy(ptr, __per_cpu_load + i * PAGE_SIZE, PAGE_SIZE);
 323                         pcpu4k_pages[j++] = virt_to_page(ptr);
 324                 }
 325
 326         /* we're ready, commit */
 327         pr_info("PERCPU: Allocated %d 4k pages, static data %zu bytes\n",
 328                 pcpu4k_nr_static_pages, static_size);
 329
 330         ret = pcpu_setup_first_chunk(pcpu4k_get_page, static_size,
 331                                      PERCPU_FIRST_CHUNK_RESERVE, -1,
 332                                      -1, NULL, pcpu4k_populate_pte);
 333         goto out_free_ar;
 334
 335 enomem:
 336         while (--j >= 0)
 337                 free_bootmem(__pa(page_address(pcpu4k_pages[j])), PAGE_SIZE);
 338         ret = -ENOMEM;
 339 out_free_ar:
 340         free_bootmem(__pa(pcpu4k_pages), pages_size);
 341         return ret;
 342 }
 343
 344 static inline void setup_percpu_segment(int cpu)
 345 {
 346 #ifdef CONFIG_X86_32
 347         struct desc_struct gdt;
 348
 349         pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF,
 350                         0x2 | DESCTYPE_S, 0x8);
 351         gdt.s = 1;
 352         write_gdt_entry(get_cpu_gdt_table(cpu),
 353                         GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S);
 354 #endif
 355 }
 356
 357 /*
 358  * Great future plan:
 359  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
 360  * Always point %gs to its beginning
 361  */
 362 void __init setup_per_cpu_areas(void)
 363 {
 364         size_t static_size = __per_cpu_end - __per_cpu_start;
 365         unsigned int cpu;
 366         unsigned long delta;
 367         size_t pcpu_unit_size;
 368         ssize_t ret;
 369
 370         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
 371                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 372
 373         /*
 374          * Allocate percpu area.  If PSE is supported, try to make use
 375          * of large page mappings.  Please read comments on top of
 376          * each allocator for details.
 377          */
 378         ret = setup_pcpu_lpage(static_size);
 379         if (ret < 0)
 380                 ret = setup_pcpu_embed(static_size);
 381         if (ret < 0)
 382                 ret = setup_pcpu_4k(static_size);
 383         if (ret < 0)
 384                 panic("cannot allocate static percpu area (%zu bytes, err=%zd)",
 385                       static_size, ret);
 386
 387         pcpu_unit_size = ret;
 388
 389         /* alrighty, percpu areas up and running */
 390         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 391         for_each_possible_cpu(cpu) {
 392                 per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size;
 393                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 394                 per_cpu(cpu_number, cpu) = cpu;
 395                 setup_percpu_segment(cpu);
 396                 setup_stack_canary_segment(cpu);
 397                 /*
 398                  * Copy data used in early init routines from the
 399                  * initial arrays to the per cpu data areas.  These
 400                  * arrays then become expendable and the *_early_ptr's
 401                  * are zeroed indicating that the static arrays are
 402                  * gone.
 403                  */
 404 #ifdef CONFIG_X86_LOCAL_APIC
 405                 per_cpu(x86_cpu_to_apicid, cpu) =
 406                         early_per_cpu_map(x86_cpu_to_apicid, cpu);
 407                 per_cpu(x86_bios_cpu_apicid, cpu) =
 408                         early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 409 #endif
 410 #ifdef CONFIG_X86_64
 411                 per_cpu(irq_stack_ptr, cpu) =
 412                         per_cpu(irq_stack_union.irq_stack, cpu) +
 413                         IRQ_STACK_SIZE - 64;
 414 #ifdef CONFIG_NUMA
 415                 per_cpu(x86_cpu_to_node_map, cpu) =
 416                         early_per_cpu_map(x86_cpu_to_node_map, cpu);
 417 #endif
 418 #endif
 419                 /*
 420                  * Up to this point, the boot CPU has been using .data.init
 421                  * area.  Reload any changed state for the boot CPU.
 422                  */
 423                 if (cpu == boot_cpu_id)
 424                         switch_to_new_gdt(cpu);
 425         }
 426
 427         /* indicate the early static arrays will soon be gone */
 428 #ifdef CONFIG_X86_LOCAL_APIC
 429         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
 430         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
 431 #endif
 432 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
 433         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 434 #endif
 435
 436 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
 437         /*
 438          * make sure boot cpu node_number is right, when boot cpu is on the
 439          * node that doesn't have mem installed
 440          */
 441         per_cpu(node_number, boot_cpu_id) = cpu_to_node(boot_cpu_id);
 442 #endif
 443
 444         /* Setup node to cpumask map */
 445         setup_node_to_cpumask_map();
 446
 447         /* Setup cpu initialized, callin, callout masks */
 448         setup_cpu_local_masks();
 449 }