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  * pcpu_need_numa - determine percpu allocation needs to consider NUMA
  47  *
  48  * If NUMA is not configured or there is only one NUMA node available,
  49  * there is no reason to consider NUMA.  This function determines
  50  * whether percpu allocation should consider NUMA or not.
  51  *
  52  * RETURNS:
  53  * true if NUMA should be considered; otherwise, false.
  54  */
  55 static bool __init pcpu_need_numa(void)
  56 {
  57 #ifdef CONFIG_NEED_MULTIPLE_NODES
  58         pg_data_t *last = NULL;
  59         unsigned int cpu;
  60
  61         for_each_possible_cpu(cpu) {
  62                 int node = early_cpu_to_node(cpu);
  63
  64                 if (node_online(node) && NODE_DATA(node) &&
  65                     last && last != NODE_DATA(node))
  66                         return true;
  67
  68                 last = NODE_DATA(node);
  69         }
  70 #endif
  71         return false;
  72 }
  73
  74 /**
  75  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
  76  * @cpu: cpu to allocate for
  77  * @size: size allocation in bytes
  78  * @align: alignment
  79  *
  80  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
  81  * does the right thing for NUMA regardless of the current
  82  * configuration.
  83  *
  84  * RETURNS:
  85  * Pointer to the allocated area on success, NULL on failure.
  86  */
  87 static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size,
  88                                         unsigned long align)
  89 {
  90         const unsigned long goal = __pa(MAX_DMA_ADDRESS);
  91 #ifdef CONFIG_NEED_MULTIPLE_NODES
  92         int node = early_cpu_to_node(cpu);
  93         void *ptr;
  94
  95         if (!node_online(node) || !NODE_DATA(node)) {
  96                 ptr = __alloc_bootmem_nopanic(size, align, goal);
  97                 pr_info("cpu %d has no node %d or node-local memory\n",
  98                         cpu, node);
  99                 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
 100                          cpu, size, __pa(ptr));
 101         } else {
 102                 ptr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
 103                                                    size, align, goal);
 104                 pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
 105                          "%016lx\n", cpu, size, node, __pa(ptr));
 106         }
 107         return ptr;
 108 #else
 109         return __alloc_bootmem_nopanic(size, align, goal);
 110 #endif
 111 }
 112
 113 /*
 114  * Remap allocator
 115  *
 116  * This allocator uses PMD page as unit.  A PMD page is allocated for
 117  * each cpu and each is remapped into vmalloc area using PMD mapping.
 118  * As PMD page is quite large, only part of it is used for the first
 119  * chunk.  Unused part is returned to the bootmem allocator.
 120  *
 121  * So, the PMD pages are mapped twice - once to the physical mapping
 122  * and to the vmalloc area for the first percpu chunk.  The double
 123  * mapping does add one more PMD TLB entry pressure but still is much
 124  * better than only using 4k mappings while still being NUMA friendly.
 125  */
 126 #ifdef CONFIG_NEED_MULTIPLE_NODES
 127 static size_t pcpur_size __initdata;
 128 static void **pcpur_ptrs __initdata;
 129
 130 static struct page * __init pcpur_get_page(unsigned int cpu, int pageno)
 131 {
 132         size_t off = (size_t)pageno << PAGE_SHIFT;
 133
 134         if (off >= pcpur_size)
 135                 return NULL;
 136
 137         return virt_to_page(pcpur_ptrs[cpu] + off);
 138 }
 139
 140 static ssize_t __init setup_pcpu_remap(size_t static_size)
 141 {
 142         static struct vm_struct vm;
 143         pg_data_t *last;
 144         size_t ptrs_size;
 145         unsigned int cpu;
 146         ssize_t ret;
 147
 148         /*
 149          * If large page isn't supported, there's no benefit in doing
 150          * this.  Also, on non-NUMA, embedding is better.
 151          */
 152         if (!cpu_has_pse || pcpu_need_numa())
 153                 return -EINVAL;
 154
 155         last = NULL;
 156         for_each_possible_cpu(cpu) {
 157                 int node = early_cpu_to_node(cpu);
 158
 159                 if (node_online(node) && NODE_DATA(node) &&
 160                     last && last != NODE_DATA(node))
 161                         goto proceed;
 162
 163                 last = NODE_DATA(node);
 164         }
 165         return -EINVAL;
 166
 167 proceed:
 168         /*
 169          * Currently supports only single page.  Supporting multiple
 170          * pages won't be too difficult if it ever becomes necessary.
 171          */
 172         pcpur_size = PFN_ALIGN(static_size + PERCPU_DYNAMIC_RESERVE);
 173         if (pcpur_size > PMD_SIZE) {
 174                 pr_warning("PERCPU: static data is larger than large page, "
 175                            "can't use large page\n");
 176                 return -EINVAL;
 177         }
 178
 179         /* allocate pointer array and alloc large pages */
 180         ptrs_size = PFN_ALIGN(num_possible_cpus() * sizeof(pcpur_ptrs[0]));
 181         pcpur_ptrs = alloc_bootmem(ptrs_size);
 182
 183         for_each_possible_cpu(cpu) {
 184                 pcpur_ptrs[cpu] = pcpu_alloc_bootmem(cpu, PMD_SIZE, PMD_SIZE);
 185                 if (!pcpur_ptrs[cpu])
 186                         goto enomem;
 187
 188                 /*
 189                  * Only use pcpur_size bytes and give back the rest.
 190                  *
 191                  * Ingo: The 2MB up-rounding bootmem is needed to make
 192                  * sure the partial 2MB page is still fully RAM - it's
 193                  * not well-specified to have a PAT-incompatible area
 194                  * (unmapped RAM, device memory, etc.) in that hole.
 195                  */
 196                 free_bootmem(__pa(pcpur_ptrs[cpu] + pcpur_size),
 197                              PMD_SIZE - pcpur_size);
 198
 199                 memcpy(pcpur_ptrs[cpu], __per_cpu_load, static_size);
 200         }
 201
 202         /* allocate address and map */
 203         vm.flags = VM_ALLOC;
 204         vm.size = num_possible_cpus() * PMD_SIZE;
 205         vm_area_register_early(&vm, PMD_SIZE);
 206
 207         for_each_possible_cpu(cpu) {
 208                 pmd_t *pmd;
 209
 210                 pmd = populate_extra_pmd((unsigned long)vm.addr
 211                                          + cpu * PMD_SIZE);
 212                 set_pmd(pmd, pfn_pmd(page_to_pfn(virt_to_page(pcpur_ptrs[cpu])),
 213                                      PAGE_KERNEL_LARGE));
 214         }
 215
 216         /* we're ready, commit */
 217         pr_info("PERCPU: Remapped at %p with large pages, static data "
 218                 "%zu bytes\n", vm.addr, static_size);
 219
 220         ret = pcpu_setup_first_chunk(pcpur_get_page, static_size, PMD_SIZE,
 221                                      pcpur_size - static_size, vm.addr, NULL);
 222         goto out_free_ar;
 223
 224 enomem:
 225         for_each_possible_cpu(cpu)
 226                 if (pcpur_ptrs[cpu])
 227                         free_bootmem(__pa(pcpur_ptrs[cpu]), PMD_SIZE);
 228         ret = -ENOMEM;
 229 out_free_ar:
 230         free_bootmem(__pa(pcpur_ptrs), ptrs_size);
 231         return ret;
 232 }
 233 #else
 234 static ssize_t __init setup_pcpu_remap(size_t static_size)
 235 {
 236         return -EINVAL;
 237 }
 238 #endif
 239
 240 /*
 241  * Embedding allocator
 242  *
 243  * The first chunk is sized to just contain the static area plus
 244  * PERCPU_DYNAMIC_RESERVE and allocated as a contiguous area using
 245  * bootmem allocator and used as-is without being mapped into vmalloc
 246  * area.  This enables the first chunk to piggy back on the linear
 247  * physical PMD mapping and doesn't add any additional pressure to
 248  * TLB.
 249  */
 250 static void *pcpue_ptr __initdata;
 251 static size_t pcpue_unit_size __initdata;
 252
 253 static struct page * __init pcpue_get_page(unsigned int cpu, int pageno)
 254 {
 255         return virt_to_page(pcpue_ptr + cpu * pcpue_unit_size
 256                             + ((size_t)pageno << PAGE_SHIFT));
 257 }
 258
 259 static ssize_t __init setup_pcpu_embed(size_t static_size)
 260 {
 261         unsigned int cpu;
 262
 263         /*
 264          * If large page isn't supported, there's no benefit in doing
 265          * this.  Also, embedding allocation doesn't play well with
 266          * NUMA.
 267          */
 268         if (!cpu_has_pse || pcpu_need_numa())
 269                 return -EINVAL;
 270
 271         /* allocate and copy */
 272         pcpue_unit_size = PFN_ALIGN(static_size + PERCPU_DYNAMIC_RESERVE);
 273         pcpue_unit_size = max_t(size_t, pcpue_unit_size, PCPU_MIN_UNIT_SIZE);
 274         pcpue_ptr = pcpu_alloc_bootmem(0, num_possible_cpus() * pcpue_unit_size,
 275                                        PAGE_SIZE);
 276         if (!pcpue_ptr)
 277                 return -ENOMEM;
 278
 279         for_each_possible_cpu(cpu)
 280                 memcpy(pcpue_ptr + cpu * pcpue_unit_size, __per_cpu_load,
 281                        static_size);
 282
 283         /* we're ready, commit */
 284         pr_info("PERCPU: Embedded %zu pages at %p, static data %zu bytes\n",
 285                 pcpue_unit_size >> PAGE_SHIFT, pcpue_ptr, static_size);
 286
 287         return pcpu_setup_first_chunk(pcpue_get_page, static_size,
 288                                       pcpue_unit_size,
 289                                       pcpue_unit_size - static_size, pcpue_ptr,
 290                                       NULL);
 291 }
 292
 293 /*
 294  * 4k page allocator
 295  *
 296  * This is the basic allocator.  Static percpu area is allocated
 297  * page-by-page and most of initialization is done by the generic
 298  * setup function.
 299  */
 300 static struct page **pcpu4k_pages __initdata;
 301 static int pcpu4k_nr_static_pages __initdata;
 302
 303 static struct page * __init pcpu4k_get_page(unsigned int cpu, int pageno)
 304 {
 305         if (pageno < pcpu4k_nr_static_pages)
 306                 return pcpu4k_pages[cpu * pcpu4k_nr_static_pages + pageno];
 307         return NULL;
 308 }
 309
 310 static void __init pcpu4k_populate_pte(unsigned long addr)
 311 {
 312         populate_extra_pte(addr);
 313 }
 314
 315 static ssize_t __init setup_pcpu_4k(size_t static_size)
 316 {
 317         size_t pages_size;
 318         unsigned int cpu;
 319         int i, j;
 320         ssize_t ret;
 321
 322         pcpu4k_nr_static_pages = PFN_UP(static_size);
 323
 324         /* unaligned allocations can't be freed, round up to page size */
 325         pages_size = PFN_ALIGN(pcpu4k_nr_static_pages * num_possible_cpus()
 326                                * sizeof(pcpu4k_pages[0]));
 327         pcpu4k_pages = alloc_bootmem(pages_size);
 328
 329         /* allocate and copy */
 330         j = 0;
 331         for_each_possible_cpu(cpu)
 332                 for (i = 0; i < pcpu4k_nr_static_pages; i++) {
 333                         void *ptr;
 334
 335                         ptr = pcpu_alloc_bootmem(cpu, PAGE_SIZE, PAGE_SIZE);
 336                         if (!ptr)
 337                                 goto enomem;
 338
 339                         memcpy(ptr, __per_cpu_load + i * PAGE_SIZE, PAGE_SIZE);
 340                         pcpu4k_pages[j++] = virt_to_page(ptr);
 341                 }
 342
 343         /* we're ready, commit */
 344         pr_info("PERCPU: Allocated %d 4k pages, static data %zu bytes\n",
 345                 pcpu4k_nr_static_pages, static_size);
 346
 347         ret = pcpu_setup_first_chunk(pcpu4k_get_page, static_size, -1, -1, NULL,
 348                                      pcpu4k_populate_pte);
 349         goto out_free_ar;
 350
 351 enomem:
 352         while (--j >= 0)
 353                 free_bootmem(__pa(page_address(pcpu4k_pages[j])), PAGE_SIZE);
 354         ret = -ENOMEM;
 355 out_free_ar:
 356         free_bootmem(__pa(pcpu4k_pages), pages_size);
 357         return ret;
 358 }
 359
 360 static inline void setup_percpu_segment(int cpu)
 361 {
 362 #ifdef CONFIG_X86_32
 363         struct desc_struct gdt;
 364
 365         pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF,
 366                         0x2 | DESCTYPE_S, 0x8);
 367         gdt.s = 1;
 368         write_gdt_entry(get_cpu_gdt_table(cpu),
 369                         GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S);
 370 #endif
 371 }
 372
 373 /*
 374  * Great future plan:
 375  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
 376  * Always point %gs to its beginning
 377  */
 378 void __init setup_per_cpu_areas(void)
 379 {
 380         size_t static_size = __per_cpu_end - __per_cpu_start;
 381         unsigned int cpu;
 382         unsigned long delta;
 383         size_t pcpu_unit_size;
 384         ssize_t ret;
 385
 386         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
 387                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 388
 389         /*
 390          * Allocate percpu area.  If PSE is supported, try to make use
 391          * of large page mappings.  Please read comments on top of
 392          * each allocator for details.
 393          */
 394         ret = setup_pcpu_remap(static_size);
 395         if (ret < 0)
 396                 ret = setup_pcpu_embed(static_size);
 397         if (ret < 0)
 398                 ret = setup_pcpu_4k(static_size);
 399         if (ret < 0)
 400                 panic("cannot allocate static percpu area (%zu bytes, err=%zd)",
 401                       static_size, ret);
 402
 403         pcpu_unit_size = ret;
 404
 405         /* alrighty, percpu areas up and running */
 406         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 407         for_each_possible_cpu(cpu) {
 408                 per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size;
 409                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 410                 per_cpu(cpu_number, cpu) = cpu;
 411                 setup_percpu_segment(cpu);
 412                 setup_stack_canary_segment(cpu);
 413                 /*
 414                  * Copy data used in early init routines from the
 415                  * initial arrays to the per cpu data areas.  These
 416                  * arrays then become expendable and the *_early_ptr's
 417                  * are zeroed indicating that the static arrays are
 418                  * gone.
 419                  */
 420 #ifdef CONFIG_X86_LOCAL_APIC
 421                 per_cpu(x86_cpu_to_apicid, cpu) =
 422                         early_per_cpu_map(x86_cpu_to_apicid, cpu);
 423                 per_cpu(x86_bios_cpu_apicid, cpu) =
 424                         early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 425 #endif
 426 #ifdef CONFIG_X86_64
 427                 per_cpu(irq_stack_ptr, cpu) =
 428                         per_cpu(irq_stack_union.irq_stack, cpu) +
 429                         IRQ_STACK_SIZE - 64;
 430 #ifdef CONFIG_NUMA
 431                 per_cpu(x86_cpu_to_node_map, cpu) =
 432                         early_per_cpu_map(x86_cpu_to_node_map, cpu);
 433 #endif
 434 #endif
 435                 /*
 436                  * Up to this point, the boot CPU has been using .data.init
 437                  * area.  Reload any changed state for the boot CPU.
 438                  */
 439                 if (cpu == boot_cpu_id)
 440                         switch_to_new_gdt(cpu);
 441         }
 442
 443         /* indicate the early static arrays will soon be gone */
 444 #ifdef CONFIG_X86_LOCAL_APIC
 445         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
 446         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
 447 #endif
 448 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
 449         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 450 #endif
 451
 452         /* Setup node to cpumask map */
 453         setup_node_to_cpumask_map();
 454
 455         /* Setup cpu initialized, callin, callout masks */
 456         setup_cpu_local_masks();
 457 }