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;
 146 static struct pcpul_ent *pcpul_map;
 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, bool chosen)
 160 {
 161         size_t map_size, dyn_size;
 162         unsigned int cpu;
 163         int i, j;
 164         ssize_t ret;
 165
 166         if (!chosen) {
 167                 size_t vm_size = VMALLOC_END - VMALLOC_START;
 168                 size_t tot_size = nr_cpu_ids * PMD_SIZE;
 169
 170                 /* on non-NUMA, embedding is better */
 171                 if (!pcpu_need_numa())
 172                         return -EINVAL;
 173
 174                 /* don't consume more than 20% of vmalloc area */
 175                 if (tot_size > vm_size / 5) {
 176                         pr_info("PERCPU: too large chunk size %zuMB for "
 177                                 "large page remap\n", tot_size >> 20);
 178                         return -EINVAL;
 179                 }
 180         }
 181
 182         /* need PSE */
 183         if (!cpu_has_pse) {
 184                 pr_warning("PERCPU: lpage allocator requires PSE\n");
 185                 return -EINVAL;
 186         }
 187
 188         /*
 189          * Currently supports only single page.  Supporting multiple
 190          * pages won't be too difficult if it ever becomes necessary.
 191          */
 192         pcpul_size = PFN_ALIGN(static_size + PERCPU_MODULE_RESERVE +
 193                                PERCPU_DYNAMIC_RESERVE);
 194         if (pcpul_size > PMD_SIZE) {
 195                 pr_warning("PERCPU: static data is larger than large page, "
 196                            "can't use large page\n");
 197                 return -EINVAL;
 198         }
 199         dyn_size = pcpul_size - static_size - PERCPU_FIRST_CHUNK_RESERVE;
 200
 201         /* allocate pointer array and alloc large pages */
 202         map_size = PFN_ALIGN(nr_cpu_ids * sizeof(pcpul_map[0]));
 203         pcpul_map = alloc_bootmem(map_size);
 204
 205         for_each_possible_cpu(cpu) {
 206                 pcpul_map[cpu].cpu = cpu;
 207                 pcpul_map[cpu].ptr = pcpu_alloc_bootmem(cpu, PMD_SIZE,
 208                                                         PMD_SIZE);
 209                 if (!pcpul_map[cpu].ptr) {
 210                         pr_warning("PERCPU: failed to allocate large page "
 211                                    "for cpu%u\n", cpu);
 212                         goto enomem;
 213                 }
 214
 215                 /*
 216                  * Only use pcpul_size bytes and give back the rest.
 217                  *
 218                  * Ingo: The 2MB up-rounding bootmem is needed to make
 219                  * sure the partial 2MB page is still fully RAM - it's
 220                  * not well-specified to have a PAT-incompatible area
 221                  * (unmapped RAM, device memory, etc.) in that hole.
 222                  */
 223                 free_bootmem(__pa(pcpul_map[cpu].ptr + pcpul_size),
 224                              PMD_SIZE - pcpul_size);
 225
 226                 memcpy(pcpul_map[cpu].ptr, __per_cpu_load, static_size);
 227         }
 228
 229         /* allocate address and map */
 230         pcpul_vm.flags = VM_ALLOC;
 231         pcpul_vm.size = nr_cpu_ids * PMD_SIZE;
 232         vm_area_register_early(&pcpul_vm, PMD_SIZE);
 233
 234         for_each_possible_cpu(cpu) {
 235                 pmd_t *pmd, pmd_v;
 236
 237                 pmd = populate_extra_pmd((unsigned long)pcpul_vm.addr +
 238                                          cpu * PMD_SIZE);
 239                 pmd_v = pfn_pmd(page_to_pfn(virt_to_page(pcpul_map[cpu].ptr)),
 240                                 PAGE_KERNEL_LARGE);
 241                 set_pmd(pmd, pmd_v);
 242         }
 243
 244         /* we're ready, commit */
 245         pr_info("PERCPU: Remapped at %p with large pages, static data "
 246                 "%zu bytes\n", pcpul_vm.addr, static_size);
 247
 248         ret = pcpu_setup_first_chunk(pcpul_get_page, static_size,
 249                                      PERCPU_FIRST_CHUNK_RESERVE, dyn_size,
 250                                      PMD_SIZE, pcpul_vm.addr, NULL);
 251
 252         /* sort pcpul_map array for pcpu_lpage_remapped() */
 253         for (i = 0; i < nr_cpu_ids - 1; i++)
 254                 for (j = i + 1; j < nr_cpu_ids; j++)
 255                         if (pcpul_map[i].ptr > pcpul_map[j].ptr) {
 256                                 struct pcpul_ent tmp = pcpul_map[i];
 257                                 pcpul_map[i] = pcpul_map[j];
 258                                 pcpul_map[j] = tmp;
 259                         }
 260
 261         return ret;
 262
 263 enomem:
 264         for_each_possible_cpu(cpu)
 265                 if (pcpul_map[cpu].ptr)
 266                         free_bootmem(__pa(pcpul_map[cpu].ptr), pcpul_size);
 267         free_bootmem(__pa(pcpul_map), map_size);
 268         return -ENOMEM;
 269 }
 270
 271 /**
 272  * pcpu_lpage_remapped - determine whether a kaddr is in pcpul recycled area
 273  * @kaddr: the kernel address in question
 274  *
 275  * Determine whether @kaddr falls in the pcpul recycled area.  This is
 276  * used by pageattr to detect VM aliases and break up the pcpu PMD
 277  * mapping such that the same physical page is not mapped under
 278  * different attributes.
 279  *
 280  * The recycled area is always at the tail of a partially used PMD
 281  * page.
 282  *
 283  * RETURNS:
 284  * Address of corresponding remapped pcpu address if match is found;
 285  * otherwise, NULL.
 286  */
 287 void *pcpu_lpage_remapped(void *kaddr)
 288 {
 289         void *pmd_addr = (void *)((unsigned long)kaddr & PMD_MASK);
 290         unsigned long offset = (unsigned long)kaddr & ~PMD_MASK;
 291         int left = 0, right = nr_cpu_ids - 1;
 292         int pos;
 293
 294         /* pcpul in use at all? */
 295         if (!pcpul_map)
 296                 return NULL;
 297
 298         /* okay, perform binary search */
 299         while (left <= right) {
 300                 pos = (left + right) / 2;
 301
 302                 if (pcpul_map[pos].ptr < pmd_addr)
 303                         left = pos + 1;
 304                 else if (pcpul_map[pos].ptr > pmd_addr)
 305                         right = pos - 1;
 306                 else {
 307                         /* it shouldn't be in the area for the first chunk */
 308                         WARN_ON(offset < pcpul_size);
 309
 310                         return pcpul_vm.addr +
 311                                 pcpul_map[pos].cpu * PMD_SIZE + offset;
 312                 }
 313         }
 314
 315         return NULL;
 316 }
 317 #else
 318 static ssize_t __init setup_pcpu_lpage(size_t static_size, bool chosen)
 319 {
 320         return -EINVAL;
 321 }
 322 #endif
 323
 324 /*
 325  * Embedding allocator
 326  *
 327  * The first chunk is sized to just contain the static area plus
 328  * module and dynamic reserves and embedded into linear physical
 329  * mapping so that it can use PMD mapping without additional TLB
 330  * pressure.
 331  */
 332 static ssize_t __init setup_pcpu_embed(size_t static_size, bool chosen)
 333 {
 334         size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
 335
 336         /*
 337          * If large page isn't supported, there's no benefit in doing
 338          * this.  Also, embedding allocation doesn't play well with
 339          * NUMA.
 340          */
 341         if (!chosen && (!cpu_has_pse || pcpu_need_numa()))
 342                 return -EINVAL;
 343
 344         return pcpu_embed_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE,
 345                                       reserve - PERCPU_FIRST_CHUNK_RESERVE, -1);
 346 }
 347
 348 /*
 349  * 4k page allocator
 350  *
 351  * This is the basic allocator.  Static percpu area is allocated
 352  * page-by-page and most of initialization is done by the generic
 353  * setup function.
 354  */
 355 static struct page **pcpu4k_pages __initdata;
 356 static int pcpu4k_nr_static_pages __initdata;
 357
 358 static struct page * __init pcpu4k_get_page(unsigned int cpu, int pageno)
 359 {
 360         if (pageno < pcpu4k_nr_static_pages)
 361                 return pcpu4k_pages[cpu * pcpu4k_nr_static_pages + pageno];
 362         return NULL;
 363 }
 364
 365 static void __init pcpu4k_populate_pte(unsigned long addr)
 366 {
 367         populate_extra_pte(addr);
 368 }
 369
 370 static ssize_t __init setup_pcpu_4k(size_t static_size)
 371 {
 372         size_t pages_size;
 373         unsigned int cpu;
 374         int i, j;
 375         ssize_t ret;
 376
 377         pcpu4k_nr_static_pages = PFN_UP(static_size);
 378
 379         /* unaligned allocations can't be freed, round up to page size */
 380         pages_size = PFN_ALIGN(pcpu4k_nr_static_pages * nr_cpu_ids
 381                                * sizeof(pcpu4k_pages[0]));
 382         pcpu4k_pages = alloc_bootmem(pages_size);
 383
 384         /* allocate and copy */
 385         j = 0;
 386         for_each_possible_cpu(cpu)
 387                 for (i = 0; i < pcpu4k_nr_static_pages; i++) {
 388                         void *ptr;
 389
 390                         ptr = pcpu_alloc_bootmem(cpu, PAGE_SIZE, PAGE_SIZE);
 391                         if (!ptr) {
 392                                 pr_warning("PERCPU: failed to allocate "
 393                                            "4k page for cpu%u\n", cpu);
 394                                 goto enomem;
 395                         }
 396
 397                         memcpy(ptr, __per_cpu_load + i * PAGE_SIZE, PAGE_SIZE);
 398                         pcpu4k_pages[j++] = virt_to_page(ptr);
 399                 }
 400
 401         /* we're ready, commit */
 402         pr_info("PERCPU: Allocated %d 4k pages, static data %zu bytes\n",
 403                 pcpu4k_nr_static_pages, static_size);
 404
 405         ret = pcpu_setup_first_chunk(pcpu4k_get_page, static_size,
 406                                      PERCPU_FIRST_CHUNK_RESERVE, -1,
 407                                      -1, NULL, pcpu4k_populate_pte);
 408         goto out_free_ar;
 409
 410 enomem:
 411         while (--j >= 0)
 412                 free_bootmem(__pa(page_address(pcpu4k_pages[j])), PAGE_SIZE);
 413         ret = -ENOMEM;
 414 out_free_ar:
 415         free_bootmem(__pa(pcpu4k_pages), pages_size);
 416         return ret;
 417 }
 418
 419 /* for explicit first chunk allocator selection */
 420 static char pcpu_chosen_alloc[16] __initdata;
 421
 422 static int __init percpu_alloc_setup(char *str)
 423 {
 424         strncpy(pcpu_chosen_alloc, str, sizeof(pcpu_chosen_alloc) - 1);
 425         return 0;
 426 }
 427 early_param("percpu_alloc", percpu_alloc_setup);
 428
 429 static inline void setup_percpu_segment(int cpu)
 430 {
 431 #ifdef CONFIG_X86_32
 432         struct desc_struct gdt;
 433
 434         pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF,
 435                         0x2 | DESCTYPE_S, 0x8);
 436         gdt.s = 1;
 437         write_gdt_entry(get_cpu_gdt_table(cpu),
 438                         GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S);
 439 #endif
 440 }
 441
 442 void __init setup_per_cpu_areas(void)
 443 {
 444         size_t static_size = __per_cpu_end - __per_cpu_start;
 445         unsigned int cpu;
 446         unsigned long delta;
 447         size_t pcpu_unit_size;
 448         ssize_t ret;
 449
 450         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
 451                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 452
 453         /*
 454          * Allocate percpu area.  If PSE is supported, try to make use
 455          * of large page mappings.  Please read comments on top of
 456          * each allocator for details.
 457          */
 458         ret = -EINVAL;
 459         if (strlen(pcpu_chosen_alloc)) {
 460                 if (strcmp(pcpu_chosen_alloc, "4k")) {
 461                         if (!strcmp(pcpu_chosen_alloc, "lpage"))
 462                                 ret = setup_pcpu_lpage(static_size, true);
 463                         else if (!strcmp(pcpu_chosen_alloc, "embed"))
 464                                 ret = setup_pcpu_embed(static_size, true);
 465                         else
 466                                 pr_warning("PERCPU: unknown allocator %s "
 467                                            "specified\n", pcpu_chosen_alloc);
 468                         if (ret < 0)
 469                                 pr_warning("PERCPU: %s allocator failed (%zd), "
 470                                            "falling back to 4k\n",
 471                                            pcpu_chosen_alloc, ret);
 472                 }
 473         } else {
 474                 ret = setup_pcpu_lpage(static_size, false);
 475                 if (ret < 0)
 476                         ret = setup_pcpu_embed(static_size, false);
 477         }
 478         if (ret < 0)
 479                 ret = setup_pcpu_4k(static_size);
 480         if (ret < 0)
 481                 panic("cannot allocate static percpu area (%zu bytes, err=%zd)",
 482                       static_size, ret);
 483
 484         pcpu_unit_size = ret;
 485
 486         /* alrighty, percpu areas up and running */
 487         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 488         for_each_possible_cpu(cpu) {
 489                 per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size;
 490                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 491                 per_cpu(cpu_number, cpu) = cpu;
 492                 setup_percpu_segment(cpu);
 493                 setup_stack_canary_segment(cpu);
 494                 /*
 495                  * Copy data used in early init routines from the
 496                  * initial arrays to the per cpu data areas.  These
 497                  * arrays then become expendable and the *_early_ptr's
 498                  * are zeroed indicating that the static arrays are
 499                  * gone.
 500                  */
 501 #ifdef CONFIG_X86_LOCAL_APIC
 502                 per_cpu(x86_cpu_to_apicid, cpu) =
 503                         early_per_cpu_map(x86_cpu_to_apicid, cpu);
 504                 per_cpu(x86_bios_cpu_apicid, cpu) =
 505                         early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 506 #endif
 507 #ifdef CONFIG_X86_64
 508                 per_cpu(irq_stack_ptr, cpu) =
 509                         per_cpu(irq_stack_union.irq_stack, cpu) +
 510                         IRQ_STACK_SIZE - 64;
 511 #ifdef CONFIG_NUMA
 512                 per_cpu(x86_cpu_to_node_map, cpu) =
 513                         early_per_cpu_map(x86_cpu_to_node_map, cpu);
 514 #endif
 515 #endif
 516                 /*
 517                  * Up to this point, the boot CPU has been using .data.init
 518                  * area.  Reload any changed state for the boot CPU.
 519                  */
 520                 if (cpu == boot_cpu_id)
 521                         switch_to_new_gdt(cpu);
 522         }
 523
 524         /* indicate the early static arrays will soon be gone */
 525 #ifdef CONFIG_X86_LOCAL_APIC
 526         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
 527         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
 528 #endif
 529 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
 530         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 531 #endif
 532
 533 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
 534         /*
 535          * make sure boot cpu node_number is right, when boot cpu is on the
 536          * node that doesn't have mem installed
 537          */
 538         per_cpu(node_number, boot_cpu_id) = cpu_to_node(boot_cpu_id);
 539 #endif
 540
 541         /* Setup node to cpumask map */
 542         setup_node_to_cpumask_map();
 543
 544         /* Setup cpu initialized, callin, callout masks */
 545         setup_cpu_local_masks();
 546 }