arch/x86/mm/init_64.c

   1 /*
   2  *  linux/arch/x86_64/mm/init.c
   3  *
   4  *  Copyright (C) 1995  Linus Torvalds
   5  *  Copyright (C) 2000  Pavel Machek <pavel@ucw.cz>
   6  *  Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
   7  */
   8
   9 #include <linux/signal.h>
  10 #include <linux/sched.h>
  11 #include <linux/kernel.h>
  12 #include <linux/errno.h>
  13 #include <linux/string.h>
  14 #include <linux/types.h>
  15 #include <linux/ptrace.h>
  16 #include <linux/mman.h>
  17 #include <linux/mm.h>
  18 #include <linux/swap.h>
  19 #include <linux/smp.h>
  20 #include <linux/init.h>
  21 #include <linux/initrd.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/bootmem.h>
  24 #include <linux/memblock.h>
  25 #include <linux/proc_fs.h>
  26 #include <linux/pci.h>
  27 #include <linux/pfn.h>
  28 #include <linux/poison.h>
  29 #include <linux/dma-mapping.h>
  30 #include <linux/memory.h>
  31 #include <linux/memory_hotplug.h>
  32 #include <linux/memremap.h>
  33 #include <linux/nmi.h>
  34 #include <linux/gfp.h>
  35 #include <linux/kcore.h>
  36
  37 #include <asm/processor.h>
  38 #include <asm/bios_ebda.h>
  39 #include <linux/uaccess.h>
  40 #include <asm/pgtable.h>
  41 #include <asm/pgalloc.h>
  42 #include <asm/dma.h>
  43 #include <asm/fixmap.h>
  44 #include <asm/e820/api.h>
  45 #include <asm/apic.h>
  46 #include <asm/tlb.h>
  47 #include <asm/mmu_context.h>
  48 #include <asm/proto.h>
  49 #include <asm/smp.h>
  50 #include <asm/sections.h>
  51 #include <asm/kdebug.h>
  52 #include <asm/numa.h>
  53 #include <asm/set_memory.h>
  54 #include <asm/init.h>
  55 #include <asm/uv/uv.h>
  56 #include <asm/setup.h>
  57
  58 #include "mm_internal.h"
  59
  60 #include "ident_map.c"
  61
  62 /*
  63  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  64  * physical space so we can cache the place of the first one and move
  65  * around without checking the pgd every time.
  66  */
  67
  68 pteval_t __supported_pte_mask __read_mostly = ~0;
  69 EXPORT_SYMBOL_GPL(__supported_pte_mask);
  70
  71 int force_personality32;
  72
  73 /*
  74  * noexec32=on|off
  75  * Control non executable heap for 32bit processes.
  76  * To control the stack too use noexec=off
  77  *
  78  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
  79  * off  PROT_READ implies PROT_EXEC
  80  */
  81 static int __init nonx32_setup(char *str)
  82 {
  83         if (!strcmp(str, "on"))
  84                 force_personality32 &= ~READ_IMPLIES_EXEC;
  85         else if (!strcmp(str, "off"))
  86                 force_personality32 |= READ_IMPLIES_EXEC;
  87         return 1;
  88 }
  89 __setup("noexec32=", nonx32_setup);
  90
  91 /*
  92  * When memory was added make sure all the processes MM have
  93  * suitable PGD entries in the local PGD level page.
  94  */
  95 void sync_global_pgds(unsigned long start, unsigned long end)
  96 {
  97         unsigned long addr;
  98
  99         for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
 100                 pgd_t *pgd_ref = pgd_offset_k(addr);
 101                 const p4d_t *p4d_ref;
 102                 struct page *page;
 103
 104                 /*
 105                  * With folded p4d, pgd_none() is always false, we need to
 106                  * handle synchonization on p4d level.
 107                  */
 108                 BUILD_BUG_ON(pgd_none(*pgd_ref));
 109                 p4d_ref = p4d_offset(pgd_ref, addr);
 110
 111                 if (p4d_none(*p4d_ref))
 112                         continue;
 113
 114                 spin_lock(&pgd_lock);
 115                 list_for_each_entry(page, &pgd_list, lru) {
 116                         pgd_t *pgd;
 117                         p4d_t *p4d;
 118                         spinlock_t *pgt_lock;
 119
 120                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
 121                         p4d = p4d_offset(pgd, addr);
 122                         /* the pgt_lock only for Xen */
 123                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 124                         spin_lock(pgt_lock);
 125
 126                         if (!p4d_none(*p4d_ref) && !p4d_none(*p4d))
 127                                 BUG_ON(p4d_page_vaddr(*p4d)
 128                                        != p4d_page_vaddr(*p4d_ref));
 129
 130                         if (p4d_none(*p4d))
 131                                 set_p4d(p4d, *p4d_ref);
 132
 133                         spin_unlock(pgt_lock);
 134                 }
 135                 spin_unlock(&pgd_lock);
 136         }
 137 }
 138
 139 /*
 140  * NOTE: This function is marked __ref because it calls __init function
 141  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 142  */
 143 static __ref void *spp_getpage(void)
 144 {
 145         void *ptr;
 146
 147         if (after_bootmem)
 148                 ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
 149         else
 150                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 151
 152         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 153                 panic("set_pte_phys: cannot allocate page data %s\n",
 154                         after_bootmem ? "after bootmem" : "");
 155         }
 156
 157         pr_debug("spp_getpage %p\n", ptr);
 158
 159         return ptr;
 160 }
 161
 162 static p4d_t *fill_p4d(pgd_t *pgd, unsigned long vaddr)
 163 {
 164         if (pgd_none(*pgd)) {
 165                 p4d_t *p4d = (p4d_t *)spp_getpage();
 166                 pgd_populate(&init_mm, pgd, p4d);
 167                 if (p4d != p4d_offset(pgd, 0))
 168                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 169                                p4d, p4d_offset(pgd, 0));
 170         }
 171         return p4d_offset(pgd, vaddr);
 172 }
 173
 174 static pud_t *fill_pud(p4d_t *p4d, unsigned long vaddr)
 175 {
 176         if (p4d_none(*p4d)) {
 177                 pud_t *pud = (pud_t *)spp_getpage();
 178                 p4d_populate(&init_mm, p4d, pud);
 179                 if (pud != pud_offset(p4d, 0))
 180                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 181                                pud, pud_offset(p4d, 0));
 182         }
 183         return pud_offset(p4d, vaddr);
 184 }
 185
 186 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 187 {
 188         if (pud_none(*pud)) {
 189                 pmd_t *pmd = (pmd_t *) spp_getpage();
 190                 pud_populate(&init_mm, pud, pmd);
 191                 if (pmd != pmd_offset(pud, 0))
 192                         printk(KERN_ERR "PAGETABLE BUG #02! %p <-> %p\n",
 193                                pmd, pmd_offset(pud, 0));
 194         }
 195         return pmd_offset(pud, vaddr);
 196 }
 197
 198 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 199 {
 200         if (pmd_none(*pmd)) {
 201                 pte_t *pte = (pte_t *) spp_getpage();
 202                 pmd_populate_kernel(&init_mm, pmd, pte);
 203                 if (pte != pte_offset_kernel(pmd, 0))
 204                         printk(KERN_ERR "PAGETABLE BUG #03!\n");
 205         }
 206         return pte_offset_kernel(pmd, vaddr);
 207 }
 208
 209 static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
 210 {
 211         pmd_t *pmd = fill_pmd(pud, vaddr);
 212         pte_t *pte = fill_pte(pmd, vaddr);
 213
 214         set_pte(pte, new_pte);
 215
 216         /*
 217          * It's enough to flush this one mapping.
 218          * (PGE mappings get flushed as well)
 219          */
 220         __flush_tlb_one(vaddr);
 221 }
 222
 223 void set_pte_vaddr_p4d(p4d_t *p4d_page, unsigned long vaddr, pte_t new_pte)
 224 {
 225         p4d_t *p4d = p4d_page + p4d_index(vaddr);
 226         pud_t *pud = fill_pud(p4d, vaddr);
 227
 228         __set_pte_vaddr(pud, vaddr, new_pte);
 229 }
 230
 231 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 232 {
 233         pud_t *pud = pud_page + pud_index(vaddr);
 234
 235         __set_pte_vaddr(pud, vaddr, new_pte);
 236 }
 237
 238 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 239 {
 240         pgd_t *pgd;
 241         p4d_t *p4d_page;
 242
 243         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 244
 245         pgd = pgd_offset_k(vaddr);
 246         if (pgd_none(*pgd)) {
 247                 printk(KERN_ERR
 248                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 249                 return;
 250         }
 251
 252         p4d_page = p4d_offset(pgd, 0);
 253         set_pte_vaddr_p4d(p4d_page, vaddr, pteval);
 254 }
 255
 256 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 257 {
 258         pgd_t *pgd;
 259         p4d_t *p4d;
 260         pud_t *pud;
 261
 262         pgd = pgd_offset_k(vaddr);
 263         p4d = fill_p4d(pgd, vaddr);
 264         pud = fill_pud(p4d, vaddr);
 265         return fill_pmd(pud, vaddr);
 266 }
 267
 268 pte_t * __init populate_extra_pte(unsigned long vaddr)
 269 {
 270         pmd_t *pmd;
 271
 272         pmd = populate_extra_pmd(vaddr);
 273         return fill_pte(pmd, vaddr);
 274 }
 275
 276 /*
 277  * Create large page table mappings for a range of physical addresses.
 278  */
 279 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 280                                         enum page_cache_mode cache)
 281 {
 282         pgd_t *pgd;
 283         p4d_t *p4d;
 284         pud_t *pud;
 285         pmd_t *pmd;
 286         pgprot_t prot;
 287
 288         pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
 289                 pgprot_val(pgprot_4k_2_large(cachemode2pgprot(cache)));
 290         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 291         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 292                 pgd = pgd_offset_k((unsigned long)__va(phys));
 293                 if (pgd_none(*pgd)) {
 294                         p4d = (p4d_t *) spp_getpage();
 295                         set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE |
 296                                                 _PAGE_USER));
 297                 }
 298                 p4d = p4d_offset(pgd, (unsigned long)__va(phys));
 299                 if (p4d_none(*p4d)) {
 300                         pud = (pud_t *) spp_getpage();
 301                         set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE |
 302                                                 _PAGE_USER));
 303                 }
 304                 pud = pud_offset(p4d, (unsigned long)__va(phys));
 305                 if (pud_none(*pud)) {
 306                         pmd = (pmd_t *) spp_getpage();
 307                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 308                                                 _PAGE_USER));
 309                 }
 310                 pmd = pmd_offset(pud, phys);
 311                 BUG_ON(!pmd_none(*pmd));
 312                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 313         }
 314 }
 315
 316 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 317 {
 318         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
 319 }
 320
 321 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 322 {
 323         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
 324 }
 325
 326 /*
 327  * The head.S code sets up the kernel high mapping:
 328  *
 329  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 330  *
 331  * phys_base holds the negative offset to the kernel, which is added
 332  * to the compile time generated pmds. This results in invalid pmds up
 333  * to the point where we hit the physaddr 0 mapping.
 334  *
 335  * We limit the mappings to the region from _text to _brk_end.  _brk_end
 336  * is rounded up to the 2MB boundary. This catches the invalid pmds as
 337  * well, as they are located before _text:
 338  */
 339 void __init cleanup_highmap(void)
 340 {
 341         unsigned long vaddr = __START_KERNEL_map;
 342         unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
 343         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 344         pmd_t *pmd = level2_kernel_pgt;
 345
 346         /*
 347          * Native path, max_pfn_mapped is not set yet.
 348          * Xen has valid max_pfn_mapped set in
 349          *      arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
 350          */
 351         if (max_pfn_mapped)
 352                 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 353
 354         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
 355                 if (pmd_none(*pmd))
 356                         continue;
 357                 if (vaddr < (unsigned long) _text || vaddr > end)
 358                         set_pmd(pmd, __pmd(0));
 359         }
 360 }
 361
 362 /*
 363  * Create PTE level page table mapping for physical addresses.
 364  * It returns the last physical address mapped.
 365  */
 366 static unsigned long __meminit
 367 phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
 368               pgprot_t prot)
 369 {
 370         unsigned long pages = 0, paddr_next;
 371         unsigned long paddr_last = paddr_end;
 372         pte_t *pte;
 373         int i;
 374
 375         pte = pte_page + pte_index(paddr);
 376         i = pte_index(paddr);
 377
 378         for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
 379                 paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
 380                 if (paddr >= paddr_end) {
 381                         if (!after_bootmem &&
 382                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 383                                              E820_TYPE_RAM) &&
 384                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 385                                              E820_TYPE_RESERVED_KERN))
 386                                 set_pte(pte, __pte(0));
 387                         continue;
 388                 }
 389
 390                 /*
 391                  * We will re-use the existing mapping.
 392                  * Xen for example has some special requirements, like mapping
 393                  * pagetable pages as RO. So assume someone who pre-setup
 394                  * these mappings are more intelligent.
 395                  */
 396                 if (!pte_none(*pte)) {
 397                         if (!after_bootmem)
 398                                 pages++;
 399                         continue;
 400                 }
 401
 402                 if (0)
 403                         pr_info("   pte=%p addr=%lx pte=%016lx\n", pte, paddr,
 404                                 pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 405                 pages++;
 406                 set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, prot));
 407                 paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
 408         }
 409
 410         update_page_count(PG_LEVEL_4K, pages);
 411
 412         return paddr_last;
 413 }
 414
 415 /*
 416  * Create PMD level page table mapping for physical addresses. The virtual
 417  * and physical address have to be aligned at this level.
 418  * It returns the last physical address mapped.
 419  */
 420 static unsigned long __meminit
 421 phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
 422               unsigned long page_size_mask, pgprot_t prot)
 423 {
 424         unsigned long pages = 0, paddr_next;
 425         unsigned long paddr_last = paddr_end;
 426
 427         int i = pmd_index(paddr);
 428
 429         for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
 430                 pmd_t *pmd = pmd_page + pmd_index(paddr);
 431                 pte_t *pte;
 432                 pgprot_t new_prot = prot;
 433
 434                 paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
 435                 if (paddr >= paddr_end) {
 436                         if (!after_bootmem &&
 437                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 438                                              E820_TYPE_RAM) &&
 439                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 440                                              E820_TYPE_RESERVED_KERN))
 441                                 set_pmd(pmd, __pmd(0));
 442                         continue;
 443                 }
 444
 445                 if (!pmd_none(*pmd)) {
 446                         if (!pmd_large(*pmd)) {
 447                                 spin_lock(&init_mm.page_table_lock);
 448                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
 449                                 paddr_last = phys_pte_init(pte, paddr,
 450                                                            paddr_end, prot);
 451                                 spin_unlock(&init_mm.page_table_lock);
 452                                 continue;
 453                         }
 454                         /*
 455                          * If we are ok with PG_LEVEL_2M mapping, then we will
 456                          * use the existing mapping,
 457                          *
 458                          * Otherwise, we will split the large page mapping but
 459                          * use the same existing protection bits except for
 460                          * large page, so that we don't violate Intel's TLB
 461                          * Application note (317080) which says, while changing
 462                          * the page sizes, new and old translations should
 463                          * not differ with respect to page frame and
 464                          * attributes.
 465                          */
 466                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 467                                 if (!after_bootmem)
 468                                         pages++;
 469                                 paddr_last = paddr_next;
 470                                 continue;
 471                         }
 472                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 473                 }
 474
 475                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 476                         pages++;
 477                         spin_lock(&init_mm.page_table_lock);
 478                         set_pte((pte_t *)pmd,
 479                                 pfn_pte((paddr & PMD_MASK) >> PAGE_SHIFT,
 480                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 481                         spin_unlock(&init_mm.page_table_lock);
 482                         paddr_last = paddr_next;
 483                         continue;
 484                 }
 485
 486                 pte = alloc_low_page();
 487                 paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot);
 488
 489                 spin_lock(&init_mm.page_table_lock);
 490                 pmd_populate_kernel(&init_mm, pmd, pte);
 491                 spin_unlock(&init_mm.page_table_lock);
 492         }
 493         update_page_count(PG_LEVEL_2M, pages);
 494         return paddr_last;
 495 }
 496
 497 /*
 498  * Create PUD level page table mapping for physical addresses. The virtual
 499  * and physical address do not have to be aligned at this level. KASLR can
 500  * randomize virtual addresses up to this level.
 501  * It returns the last physical address mapped.
 502  */
 503 static unsigned long __meminit
 504 phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
 505               unsigned long page_size_mask)
 506 {
 507         unsigned long pages = 0, paddr_next;
 508         unsigned long paddr_last = paddr_end;
 509         unsigned long vaddr = (unsigned long)__va(paddr);
 510         int i = pud_index(vaddr);
 511
 512         for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
 513                 pud_t *pud;
 514                 pmd_t *pmd;
 515                 pgprot_t prot = PAGE_KERNEL;
 516
 517                 vaddr = (unsigned long)__va(paddr);
 518                 pud = pud_page + pud_index(vaddr);
 519                 paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
 520
 521                 if (paddr >= paddr_end) {
 522                         if (!after_bootmem &&
 523                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 524                                              E820_TYPE_RAM) &&
 525                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 526                                              E820_TYPE_RESERVED_KERN))
 527                                 set_pud(pud, __pud(0));
 528                         continue;
 529                 }
 530
 531                 if (!pud_none(*pud)) {
 532                         if (!pud_large(*pud)) {
 533                                 pmd = pmd_offset(pud, 0);
 534                                 paddr_last = phys_pmd_init(pmd, paddr,
 535                                                            paddr_end,
 536                                                            page_size_mask,
 537                                                            prot);
 538                                 __flush_tlb_all();
 539                                 continue;
 540                         }
 541                         /*
 542                          * If we are ok with PG_LEVEL_1G mapping, then we will
 543                          * use the existing mapping.
 544                          *
 545                          * Otherwise, we will split the gbpage mapping but use
 546                          * the same existing protection  bits except for large
 547                          * page, so that we don't violate Intel's TLB
 548                          * Application note (317080) which says, while changing
 549                          * the page sizes, new and old translations should
 550                          * not differ with respect to page frame and
 551                          * attributes.
 552                          */
 553                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 554                                 if (!after_bootmem)
 555                                         pages++;
 556                                 paddr_last = paddr_next;
 557                                 continue;
 558                         }
 559                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 560                 }
 561
 562                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 563                         pages++;
 564                         spin_lock(&init_mm.page_table_lock);
 565                         set_pte((pte_t *)pud,
 566                                 pfn_pte((paddr & PUD_MASK) >> PAGE_SHIFT,
 567                                         PAGE_KERNEL_LARGE));
 568                         spin_unlock(&init_mm.page_table_lock);
 569                         paddr_last = paddr_next;
 570                         continue;
 571                 }
 572
 573                 pmd = alloc_low_page();
 574                 paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
 575                                            page_size_mask, prot);
 576
 577                 spin_lock(&init_mm.page_table_lock);
 578                 pud_populate(&init_mm, pud, pmd);
 579                 spin_unlock(&init_mm.page_table_lock);
 580         }
 581         __flush_tlb_all();
 582
 583         update_page_count(PG_LEVEL_1G, pages);
 584
 585         return paddr_last;
 586 }
 587
 588 /*
 589  * Create page table mapping for the physical memory for specific physical
 590  * addresses. The virtual and physical addresses have to be aligned on PMD level
 591  * down. It returns the last physical address mapped.
 592  */
 593 unsigned long __meminit
 594 kernel_physical_mapping_init(unsigned long paddr_start,
 595                              unsigned long paddr_end,
 596                              unsigned long page_size_mask)
 597 {
 598         bool pgd_changed = false;
 599         unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
 600
 601         paddr_last = paddr_end;
 602         vaddr = (unsigned long)__va(paddr_start);
 603         vaddr_end = (unsigned long)__va(paddr_end);
 604         vaddr_start = vaddr;
 605
 606         for (; vaddr < vaddr_end; vaddr = vaddr_next) {
 607                 pgd_t *pgd = pgd_offset_k(vaddr);
 608                 p4d_t *p4d;
 609                 pud_t *pud;
 610
 611                 vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
 612
 613                 BUILD_BUG_ON(pgd_none(*pgd));
 614                 p4d = p4d_offset(pgd, vaddr);
 615                 if (p4d_val(*p4d)) {
 616                         pud = (pud_t *)p4d_page_vaddr(*p4d);
 617                         paddr_last = phys_pud_init(pud, __pa(vaddr),
 618                                                    __pa(vaddr_end),
 619                                                    page_size_mask);
 620                         continue;
 621                 }
 622
 623                 pud = alloc_low_page();
 624                 paddr_last = phys_pud_init(pud, __pa(vaddr), __pa(vaddr_end),
 625                                            page_size_mask);
 626
 627                 spin_lock(&init_mm.page_table_lock);
 628                 p4d_populate(&init_mm, p4d, pud);
 629                 spin_unlock(&init_mm.page_table_lock);
 630                 pgd_changed = true;
 631         }
 632
 633         if (pgd_changed)
 634                 sync_global_pgds(vaddr_start, vaddr_end - 1);
 635
 636         __flush_tlb_all();
 637
 638         return paddr_last;
 639 }
 640
 641 #ifndef CONFIG_NUMA
 642 void __init initmem_init(void)
 643 {
 644         memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
 645 }
 646 #endif
 647
 648 void __init paging_init(void)
 649 {
 650         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 651         sparse_init();
 652
 653         /*
 654          * clear the default setting with node 0
 655          * note: don't use nodes_clear here, that is really clearing when
 656          *       numa support is not compiled in, and later node_set_state
 657          *       will not set it back.
 658          */
 659         node_clear_state(0, N_MEMORY);
 660         if (N_MEMORY != N_NORMAL_MEMORY)
 661                 node_clear_state(0, N_NORMAL_MEMORY);
 662
 663         zone_sizes_init();
 664 }
 665
 666 /*
 667  * Memory hotplug specific functions
 668  */
 669 #ifdef CONFIG_MEMORY_HOTPLUG
 670 /*
 671  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 672  * updating.
 673  */
 674 static void  update_end_of_memory_vars(u64 start, u64 size)
 675 {
 676         unsigned long end_pfn = PFN_UP(start + size);
 677
 678         if (end_pfn > max_pfn) {
 679                 max_pfn = end_pfn;
 680                 max_low_pfn = end_pfn;
 681                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 682         }
 683 }
 684
 685 /*
 686  * Memory is added always to NORMAL zone. This means you will never get
 687  * additional DMA/DMA32 memory.
 688  */
 689 int arch_add_memory(int nid, u64 start, u64 size, bool for_device)
 690 {
 691         struct pglist_data *pgdat = NODE_DATA(nid);
 692         struct zone *zone = pgdat->node_zones +
 693                 zone_for_memory(nid, start, size, ZONE_NORMAL, for_device);
 694         unsigned long start_pfn = start >> PAGE_SHIFT;
 695         unsigned long nr_pages = size >> PAGE_SHIFT;
 696         int ret;
 697
 698         init_memory_mapping(start, start + size);
 699
 700         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 701         WARN_ON_ONCE(ret);
 702
 703         /* update max_pfn, max_low_pfn and high_memory */
 704         update_end_of_memory_vars(start, size);
 705
 706         return ret;
 707 }
 708 EXPORT_SYMBOL_GPL(arch_add_memory);
 709
 710 #define PAGE_INUSE 0xFD
 711
 712 static void __meminit free_pagetable(struct page *page, int order)
 713 {
 714         unsigned long magic;
 715         unsigned int nr_pages = 1 << order;
 716         struct vmem_altmap *altmap = to_vmem_altmap((unsigned long) page);
 717
 718         if (altmap) {
 719                 vmem_altmap_free(altmap, nr_pages);
 720                 return;
 721         }
 722
 723         /* bootmem page has reserved flag */
 724         if (PageReserved(page)) {
 725                 __ClearPageReserved(page);
 726
 727                 magic = (unsigned long)page->freelist;
 728                 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
 729                         while (nr_pages--)
 730                                 put_page_bootmem(page++);
 731                 } else
 732                         while (nr_pages--)
 733                                 free_reserved_page(page++);
 734         } else
 735                 free_pages((unsigned long)page_address(page), order);
 736 }
 737
 738 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
 739 {
 740         pte_t *pte;
 741         int i;
 742
 743         for (i = 0; i < PTRS_PER_PTE; i++) {
 744                 pte = pte_start + i;
 745                 if (!pte_none(*pte))
 746                         return;
 747         }
 748
 749         /* free a pte talbe */
 750         free_pagetable(pmd_page(*pmd), 0);
 751         spin_lock(&init_mm.page_table_lock);
 752         pmd_clear(pmd);
 753         spin_unlock(&init_mm.page_table_lock);
 754 }
 755
 756 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
 757 {
 758         pmd_t *pmd;
 759         int i;
 760
 761         for (i = 0; i < PTRS_PER_PMD; i++) {
 762                 pmd = pmd_start + i;
 763                 if (!pmd_none(*pmd))
 764                         return;
 765         }
 766
 767         /* free a pmd talbe */
 768         free_pagetable(pud_page(*pud), 0);
 769         spin_lock(&init_mm.page_table_lock);
 770         pud_clear(pud);
 771         spin_unlock(&init_mm.page_table_lock);
 772 }
 773
 774 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
 775 {
 776         pud_t *pud;
 777         int i;
 778
 779         for (i = 0; i < PTRS_PER_PUD; i++) {
 780                 pud = pud_start + i;
 781                 if (!pud_none(*pud))
 782                         return;
 783         }
 784
 785         /* free a pud talbe */
 786         free_pagetable(p4d_page(*p4d), 0);
 787         spin_lock(&init_mm.page_table_lock);
 788         p4d_clear(p4d);
 789         spin_unlock(&init_mm.page_table_lock);
 790 }
 791
 792 static void __meminit
 793 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
 794                  bool direct)
 795 {
 796         unsigned long next, pages = 0;
 797         pte_t *pte;
 798         void *page_addr;
 799         phys_addr_t phys_addr;
 800
 801         pte = pte_start + pte_index(addr);
 802         for (; addr < end; addr = next, pte++) {
 803                 next = (addr + PAGE_SIZE) & PAGE_MASK;
 804                 if (next > end)
 805                         next = end;
 806
 807                 if (!pte_present(*pte))
 808                         continue;
 809
 810                 /*
 811                  * We mapped [0,1G) memory as identity mapping when
 812                  * initializing, in arch/x86/kernel/head_64.S. These
 813                  * pagetables cannot be removed.
 814                  */
 815                 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
 816                 if (phys_addr < (phys_addr_t)0x40000000)
 817                         return;
 818
 819                 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
 820                         /*
 821                          * Do not free direct mapping pages since they were
 822                          * freed when offlining, or simplely not in use.
 823                          */
 824                         if (!direct)
 825                                 free_pagetable(pte_page(*pte), 0);
 826
 827                         spin_lock(&init_mm.page_table_lock);
 828                         pte_clear(&init_mm, addr, pte);
 829                         spin_unlock(&init_mm.page_table_lock);
 830
 831                         /* For non-direct mapping, pages means nothing. */
 832                         pages++;
 833                 } else {
 834                         /*
 835                          * If we are here, we are freeing vmemmap pages since
 836                          * direct mapped memory ranges to be freed are aligned.
 837                          *
 838                          * If we are not removing the whole page, it means
 839                          * other page structs in this page are being used and
 840                          * we canot remove them. So fill the unused page_structs
 841                          * with 0xFD, and remove the page when it is wholly
 842                          * filled with 0xFD.
 843                          */
 844                         memset((void *)addr, PAGE_INUSE, next - addr);
 845
 846                         page_addr = page_address(pte_page(*pte));
 847                         if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
 848                                 free_pagetable(pte_page(*pte), 0);
 849
 850                                 spin_lock(&init_mm.page_table_lock);
 851                                 pte_clear(&init_mm, addr, pte);
 852                                 spin_unlock(&init_mm.page_table_lock);
 853                         }
 854                 }
 855         }
 856
 857         /* Call free_pte_table() in remove_pmd_table(). */
 858         flush_tlb_all();
 859         if (direct)
 860                 update_page_count(PG_LEVEL_4K, -pages);
 861 }
 862
 863 static void __meminit
 864 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
 865                  bool direct)
 866 {
 867         unsigned long next, pages = 0;
 868         pte_t *pte_base;
 869         pmd_t *pmd;
 870         void *page_addr;
 871
 872         pmd = pmd_start + pmd_index(addr);
 873         for (; addr < end; addr = next, pmd++) {
 874                 next = pmd_addr_end(addr, end);
 875
 876                 if (!pmd_present(*pmd))
 877                         continue;
 878
 879                 if (pmd_large(*pmd)) {
 880                         if (IS_ALIGNED(addr, PMD_SIZE) &&
 881                             IS_ALIGNED(next, PMD_SIZE)) {
 882                                 if (!direct)
 883                                         free_pagetable(pmd_page(*pmd),
 884                                                        get_order(PMD_SIZE));
 885
 886                                 spin_lock(&init_mm.page_table_lock);
 887                                 pmd_clear(pmd);
 888                                 spin_unlock(&init_mm.page_table_lock);
 889                                 pages++;
 890                         } else {
 891                                 /* If here, we are freeing vmemmap pages. */
 892                                 memset((void *)addr, PAGE_INUSE, next - addr);
 893
 894                                 page_addr = page_address(pmd_page(*pmd));
 895                                 if (!memchr_inv(page_addr, PAGE_INUSE,
 896                                                 PMD_SIZE)) {
 897                                         free_pagetable(pmd_page(*pmd),
 898                                                        get_order(PMD_SIZE));
 899
 900                                         spin_lock(&init_mm.page_table_lock);
 901                                         pmd_clear(pmd);
 902                                         spin_unlock(&init_mm.page_table_lock);
 903                                 }
 904                         }
 905
 906                         continue;
 907                 }
 908
 909                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
 910                 remove_pte_table(pte_base, addr, next, direct);
 911                 free_pte_table(pte_base, pmd);
 912         }
 913
 914         /* Call free_pmd_table() in remove_pud_table(). */
 915         if (direct)
 916                 update_page_count(PG_LEVEL_2M, -pages);
 917 }
 918
 919 static void __meminit
 920 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
 921                  bool direct)
 922 {
 923         unsigned long next, pages = 0;
 924         pmd_t *pmd_base;
 925         pud_t *pud;
 926         void *page_addr;
 927
 928         pud = pud_start + pud_index(addr);
 929         for (; addr < end; addr = next, pud++) {
 930                 next = pud_addr_end(addr, end);
 931
 932                 if (!pud_present(*pud))
 933                         continue;
 934
 935                 if (pud_large(*pud)) {
 936                         if (IS_ALIGNED(addr, PUD_SIZE) &&
 937                             IS_ALIGNED(next, PUD_SIZE)) {
 938                                 if (!direct)
 939                                         free_pagetable(pud_page(*pud),
 940                                                        get_order(PUD_SIZE));
 941
 942                                 spin_lock(&init_mm.page_table_lock);
 943                                 pud_clear(pud);
 944                                 spin_unlock(&init_mm.page_table_lock);
 945                                 pages++;
 946                         } else {
 947                                 /* If here, we are freeing vmemmap pages. */
 948                                 memset((void *)addr, PAGE_INUSE, next - addr);
 949
 950                                 page_addr = page_address(pud_page(*pud));
 951                                 if (!memchr_inv(page_addr, PAGE_INUSE,
 952                                                 PUD_SIZE)) {
 953                                         free_pagetable(pud_page(*pud),
 954                                                        get_order(PUD_SIZE));
 955
 956                                         spin_lock(&init_mm.page_table_lock);
 957                                         pud_clear(pud);
 958                                         spin_unlock(&init_mm.page_table_lock);
 959                                 }
 960                         }
 961
 962                         continue;
 963                 }
 964
 965                 pmd_base = pmd_offset(pud, 0);
 966                 remove_pmd_table(pmd_base, addr, next, direct);
 967                 free_pmd_table(pmd_base, pud);
 968         }
 969
 970         if (direct)
 971                 update_page_count(PG_LEVEL_1G, -pages);
 972 }
 973
 974 static void __meminit
 975 remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
 976                  bool direct)
 977 {
 978         unsigned long next, pages = 0;
 979         pud_t *pud_base;
 980         p4d_t *p4d;
 981
 982         p4d = p4d_start + p4d_index(addr);
 983         for (; addr < end; addr = next, p4d++) {
 984                 next = p4d_addr_end(addr, end);
 985
 986                 if (!p4d_present(*p4d))
 987                         continue;
 988
 989                 BUILD_BUG_ON(p4d_large(*p4d));
 990
 991                 pud_base = pud_offset(p4d, 0);
 992                 remove_pud_table(pud_base, addr, next, direct);
 993                 free_pud_table(pud_base, p4d);
 994         }
 995
 996         if (direct)
 997                 update_page_count(PG_LEVEL_512G, -pages);
 998 }
 999
1000 /* start and end are both virtual address. */
1001 static void __meminit
1002 remove_pagetable(unsigned long start, unsigned long end, bool direct)
1003 {
1004         unsigned long next;
1005         unsigned long addr;
1006         pgd_t *pgd;
1007         p4d_t *p4d;
1008
1009         for (addr = start; addr < end; addr = next) {
1010                 next = pgd_addr_end(addr, end);
1011
1012                 pgd = pgd_offset_k(addr);
1013                 if (!pgd_present(*pgd))
1014                         continue;
1015
1016                 p4d = p4d_offset(pgd, 0);
1017                 remove_p4d_table(p4d, addr, next, direct);
1018         }
1019
1020         flush_tlb_all();
1021 }
1022
1023 void __ref vmemmap_free(unsigned long start, unsigned long end)
1024 {
1025         remove_pagetable(start, end, false);
1026 }
1027
1028 #ifdef CONFIG_MEMORY_HOTREMOVE
1029 static void __meminit
1030 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1031 {
1032         start = (unsigned long)__va(start);
1033         end = (unsigned long)__va(end);
1034
1035         remove_pagetable(start, end, true);
1036 }
1037
1038 int __ref arch_remove_memory(u64 start, u64 size)
1039 {
1040         unsigned long start_pfn = start >> PAGE_SHIFT;
1041         unsigned long nr_pages = size >> PAGE_SHIFT;
1042         struct page *page = pfn_to_page(start_pfn);
1043         struct vmem_altmap *altmap;
1044         struct zone *zone;
1045         int ret;
1046
1047         /* With altmap the first mapped page is offset from @start */
1048         altmap = to_vmem_altmap((unsigned long) page);
1049         if (altmap)
1050                 page += vmem_altmap_offset(altmap);
1051         zone = page_zone(page);
1052         ret = __remove_pages(zone, start_pfn, nr_pages);
1053         WARN_ON_ONCE(ret);
1054         kernel_physical_mapping_remove(start, start + size);
1055
1056         return ret;
1057 }
1058 #endif
1059 #endif /* CONFIG_MEMORY_HOTPLUG */
1060
1061 static struct kcore_list kcore_vsyscall;
1062
1063 static void __init register_page_bootmem_info(void)
1064 {
1065 #ifdef CONFIG_NUMA
1066         int i;
1067
1068         for_each_online_node(i)
1069                 register_page_bootmem_info_node(NODE_DATA(i));
1070 #endif
1071 }
1072
1073 void __init mem_init(void)
1074 {
1075         pci_iommu_alloc();
1076
1077         /* clear_bss() already clear the empty_zero_page */
1078
1079         register_page_bootmem_info();
1080
1081         /* this will put all memory onto the freelists */
1082         free_all_bootmem();
1083         after_bootmem = 1;
1084
1085         /* Register memory areas for /proc/kcore */
1086         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR,
1087                          PAGE_SIZE, KCORE_OTHER);
1088
1089         mem_init_print_info(NULL);
1090 }
1091
1092 int kernel_set_to_readonly;
1093
1094 void set_kernel_text_rw(void)
1095 {
1096         unsigned long start = PFN_ALIGN(_text);
1097         unsigned long end = PFN_ALIGN(__stop___ex_table);
1098
1099         if (!kernel_set_to_readonly)
1100                 return;
1101
1102         pr_debug("Set kernel text: %lx - %lx for read write\n",
1103                  start, end);
1104
1105         /*
1106          * Make the kernel identity mapping for text RW. Kernel text
1107          * mapping will always be RO. Refer to the comment in
1108          * static_protections() in pageattr.c
1109          */
1110         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1111 }
1112
1113 void set_kernel_text_ro(void)
1114 {
1115         unsigned long start = PFN_ALIGN(_text);
1116         unsigned long end = PFN_ALIGN(__stop___ex_table);
1117
1118         if (!kernel_set_to_readonly)
1119                 return;
1120
1121         pr_debug("Set kernel text: %lx - %lx for read only\n",
1122                  start, end);
1123
1124         /*
1125          * Set the kernel identity mapping for text RO.
1126          */
1127         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1128 }
1129
1130 void mark_rodata_ro(void)
1131 {
1132         unsigned long start = PFN_ALIGN(_text);
1133         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1134         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1135         unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1136         unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1137         unsigned long all_end;
1138
1139         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1140                (end - start) >> 10);
1141         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1142
1143         kernel_set_to_readonly = 1;
1144
1145         /*
1146          * The rodata/data/bss/brk section (but not the kernel text!)
1147          * should also be not-executable.
1148          *
1149          * We align all_end to PMD_SIZE because the existing mapping
1150          * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1151          * split the PMD and the reminder between _brk_end and the end
1152          * of the PMD will remain mapped executable.
1153          *
1154          * Any PMD which was setup after the one which covers _brk_end
1155          * has been zapped already via cleanup_highmem().
1156          */
1157         all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1158         set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1159
1160 #ifdef CONFIG_CPA_DEBUG
1161         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1162         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1163
1164         printk(KERN_INFO "Testing CPA: again\n");
1165         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1166 #endif
1167
1168         free_init_pages("unused kernel",
1169                         (unsigned long) __va(__pa_symbol(text_end)),
1170                         (unsigned long) __va(__pa_symbol(rodata_start)));
1171         free_init_pages("unused kernel",
1172                         (unsigned long) __va(__pa_symbol(rodata_end)),
1173                         (unsigned long) __va(__pa_symbol(_sdata)));
1174
1175         debug_checkwx();
1176 }
1177
1178 int kern_addr_valid(unsigned long addr)
1179 {
1180         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1181         pgd_t *pgd;
1182         p4d_t *p4d;
1183         pud_t *pud;
1184         pmd_t *pmd;
1185         pte_t *pte;
1186
1187         if (above != 0 && above != -1UL)
1188                 return 0;
1189
1190         pgd = pgd_offset_k(addr);
1191         if (pgd_none(*pgd))
1192                 return 0;
1193
1194         p4d = p4d_offset(pgd, addr);
1195         if (p4d_none(*p4d))
1196                 return 0;
1197
1198         pud = pud_offset(p4d, addr);
1199         if (pud_none(*pud))
1200                 return 0;
1201
1202         if (pud_large(*pud))
1203                 return pfn_valid(pud_pfn(*pud));
1204
1205         pmd = pmd_offset(pud, addr);
1206         if (pmd_none(*pmd))
1207                 return 0;
1208
1209         if (pmd_large(*pmd))
1210                 return pfn_valid(pmd_pfn(*pmd));
1211
1212         pte = pte_offset_kernel(pmd, addr);
1213         if (pte_none(*pte))
1214                 return 0;
1215
1216         return pfn_valid(pte_pfn(*pte));
1217 }
1218
1219 static unsigned long probe_memory_block_size(void)
1220 {
1221         unsigned long bz = MIN_MEMORY_BLOCK_SIZE;
1222
1223         /* if system is UV or has 64GB of RAM or more, use large blocks */
1224         if (is_uv_system() || ((max_pfn << PAGE_SHIFT) >= (64UL << 30)))
1225                 bz = 2UL << 30; /* 2GB */
1226
1227         pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1228
1229         return bz;
1230 }
1231
1232 static unsigned long memory_block_size_probed;
1233 unsigned long memory_block_size_bytes(void)
1234 {
1235         if (!memory_block_size_probed)
1236                 memory_block_size_probed = probe_memory_block_size();
1237
1238         return memory_block_size_probed;
1239 }
1240
1241 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1242 /*
1243  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1244  */
1245 static long __meminitdata addr_start, addr_end;
1246 static void __meminitdata *p_start, *p_end;
1247 static int __meminitdata node_start;
1248
1249 static int __meminit vmemmap_populate_hugepages(unsigned long start,
1250                 unsigned long end, int node, struct vmem_altmap *altmap)
1251 {
1252         unsigned long addr;
1253         unsigned long next;
1254         pgd_t *pgd;
1255         p4d_t *p4d;
1256         pud_t *pud;
1257         pmd_t *pmd;
1258
1259         for (addr = start; addr < end; addr = next) {
1260                 next = pmd_addr_end(addr, end);
1261
1262                 pgd = vmemmap_pgd_populate(addr, node);
1263                 if (!pgd)
1264                         return -ENOMEM;
1265
1266                 p4d = vmemmap_p4d_populate(pgd, addr, node);
1267                 if (!p4d)
1268                         return -ENOMEM;
1269
1270                 pud = vmemmap_pud_populate(p4d, addr, node);
1271                 if (!pud)
1272                         return -ENOMEM;
1273
1274                 pmd = pmd_offset(pud, addr);
1275                 if (pmd_none(*pmd)) {
1276                         void *p;
1277
1278                         p = __vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
1279                         if (p) {
1280                                 pte_t entry;
1281
1282                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1283                                                 PAGE_KERNEL_LARGE);
1284                                 set_pmd(pmd, __pmd(pte_val(entry)));
1285
1286                                 /* check to see if we have contiguous blocks */
1287                                 if (p_end != p || node_start != node) {
1288                                         if (p_start)
1289                                                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1290                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1291                                         addr_start = addr;
1292                                         node_start = node;
1293                                         p_start = p;
1294                                 }
1295
1296                                 addr_end = addr + PMD_SIZE;
1297                                 p_end = p + PMD_SIZE;
1298                                 continue;
1299                         } else if (altmap)
1300                                 return -ENOMEM; /* no fallback */
1301                 } else if (pmd_large(*pmd)) {
1302                         vmemmap_verify((pte_t *)pmd, node, addr, next);
1303                         continue;
1304                 }
1305                 pr_warn_once("vmemmap: falling back to regular page backing\n");
1306                 if (vmemmap_populate_basepages(addr, next, node))
1307                         return -ENOMEM;
1308         }
1309         return 0;
1310 }
1311
1312 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
1313 {
1314         struct vmem_altmap *altmap = to_vmem_altmap(start);
1315         int err;
1316
1317         if (boot_cpu_has(X86_FEATURE_PSE))
1318                 err = vmemmap_populate_hugepages(start, end, node, altmap);
1319         else if (altmap) {
1320                 pr_err_once("%s: no cpu support for altmap allocations\n",
1321                                 __func__);
1322                 err = -ENOMEM;
1323         } else
1324                 err = vmemmap_populate_basepages(start, end, node);
1325         if (!err)
1326                 sync_global_pgds(start, end - 1);
1327         return err;
1328 }
1329
1330 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1331 void register_page_bootmem_memmap(unsigned long section_nr,
1332                                   struct page *start_page, unsigned long size)
1333 {
1334         unsigned long addr = (unsigned long)start_page;
1335         unsigned long end = (unsigned long)(start_page + size);
1336         unsigned long next;
1337         pgd_t *pgd;
1338         p4d_t *p4d;
1339         pud_t *pud;
1340         pmd_t *pmd;
1341         unsigned int nr_pages;
1342         struct page *page;
1343
1344         for (; addr < end; addr = next) {
1345                 pte_t *pte = NULL;
1346
1347                 pgd = pgd_offset_k(addr);
1348                 if (pgd_none(*pgd)) {
1349                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1350                         continue;
1351                 }
1352                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1353
1354                 p4d = p4d_offset(pgd, addr);
1355                 if (p4d_none(*p4d)) {
1356                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1357                         continue;
1358                 }
1359                 get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1360
1361                 pud = pud_offset(p4d, addr);
1362                 if (pud_none(*pud)) {
1363                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1364                         continue;
1365                 }
1366                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1367
1368                 if (!boot_cpu_has(X86_FEATURE_PSE)) {
1369                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1370                         pmd = pmd_offset(pud, addr);
1371                         if (pmd_none(*pmd))
1372                                 continue;
1373                         get_page_bootmem(section_nr, pmd_page(*pmd),
1374                                          MIX_SECTION_INFO);
1375
1376                         pte = pte_offset_kernel(pmd, addr);
1377                         if (pte_none(*pte))
1378                                 continue;
1379                         get_page_bootmem(section_nr, pte_page(*pte),
1380                                          SECTION_INFO);
1381                 } else {
1382                         next = pmd_addr_end(addr, end);
1383
1384                         pmd = pmd_offset(pud, addr);
1385                         if (pmd_none(*pmd))
1386                                 continue;
1387
1388                         nr_pages = 1 << (get_order(PMD_SIZE));
1389                         page = pmd_page(*pmd);
1390                         while (nr_pages--)
1391                                 get_page_bootmem(section_nr, page++,
1392                                                  SECTION_INFO);
1393                 }
1394         }
1395 }
1396 #endif
1397
1398 void __meminit vmemmap_populate_print_last(void)
1399 {
1400         if (p_start) {
1401                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1402                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1403                 p_start = NULL;
1404                 p_end = NULL;
1405                 node_start = 0;
1406         }
1407 }
1408 #endif