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 <[email protected]>
   6  *  Copyright (C) 2002,2003 Andi Kleen <[email protected]>
   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/memblock.h>
  24 #include <linux/proc_fs.h>
  25 #include <linux/pci.h>
  26 #include <linux/pfn.h>
  27 #include <linux/poison.h>
  28 #include <linux/dma-mapping.h>
  29 #include <linux/memory.h>
  30 #include <linux/memory_hotplug.h>
  31 #include <linux/memremap.h>
  32 #include <linux/nmi.h>
  33 #include <linux/gfp.h>
  34 #include <linux/kcore.h>
  35
  36 #include <asm/processor.h>
  37 #include <asm/bios_ebda.h>
  38 #include <linux/uaccess.h>
  39 #include <asm/pgtable.h>
  40 #include <asm/pgalloc.h>
  41 #include <asm/dma.h>
  42 #include <asm/fixmap.h>
  43 #include <asm/e820/api.h>
  44 #include <asm/apic.h>
  45 #include <asm/tlb.h>
  46 #include <asm/mmu_context.h>
  47 #include <asm/proto.h>
  48 #include <asm/smp.h>
  49 #include <asm/sections.h>
  50 #include <asm/kdebug.h>
  51 #include <asm/numa.h>
  52 #include <asm/set_memory.h>
  53 #include <asm/init.h>
  54 #include <asm/uv/uv.h>
  55 #include <asm/setup.h>
  56
  57 #include "mm_internal.h"
  58
  59 #include "ident_map.c"
  60
  61 /*
  62  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  63  * physical space so we can cache the place of the first one and move
  64  * around without checking the pgd every time.
  65  */
  66
  67 /* Bits supported by the hardware: */
  68 pteval_t __supported_pte_mask __read_mostly = ~0;
  69 /* Bits allowed in normal kernel mappings: */
  70 pteval_t __default_kernel_pte_mask __read_mostly = ~0;
  71 EXPORT_SYMBOL_GPL(__supported_pte_mask);
  72 /* Used in PAGE_KERNEL_* macros which are reasonably used out-of-tree: */
  73 EXPORT_SYMBOL(__default_kernel_pte_mask);
  74
  75 int force_personality32;
  76
  77 /*
  78  * noexec32=on|off
  79  * Control non executable heap for 32bit processes.
  80  * To control the stack too use noexec=off
  81  *
  82  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
  83  * off  PROT_READ implies PROT_EXEC
  84  */
  85 static int __init nonx32_setup(char *str)
  86 {
  87         if (!strcmp(str, "on"))
  88                 force_personality32 &= ~READ_IMPLIES_EXEC;
  89         else if (!strcmp(str, "off"))
  90                 force_personality32 |= READ_IMPLIES_EXEC;
  91         return 1;
  92 }
  93 __setup("noexec32=", nonx32_setup);
  94
  95 static void sync_global_pgds_l5(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                 const pgd_t *pgd_ref = pgd_offset_k(addr);
 101                 struct page *page;
 102
 103                 /* Check for overflow */
 104                 if (addr < start)
 105                         break;
 106
 107                 if (pgd_none(*pgd_ref))
 108                         continue;
 109
 110                 spin_lock(&pgd_lock);
 111                 list_for_each_entry(page, &pgd_list, lru) {
 112                         pgd_t *pgd;
 113                         spinlock_t *pgt_lock;
 114
 115                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
 116                         /* the pgt_lock only for Xen */
 117                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 118                         spin_lock(pgt_lock);
 119
 120                         if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
 121                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 122
 123                         if (pgd_none(*pgd))
 124                                 set_pgd(pgd, *pgd_ref);
 125
 126                         spin_unlock(pgt_lock);
 127                 }
 128                 spin_unlock(&pgd_lock);
 129         }
 130 }
 131
 132 static void sync_global_pgds_l4(unsigned long start, unsigned long end)
 133 {
 134         unsigned long addr;
 135
 136         for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
 137                 pgd_t *pgd_ref = pgd_offset_k(addr);
 138                 const p4d_t *p4d_ref;
 139                 struct page *page;
 140
 141                 /*
 142                  * With folded p4d, pgd_none() is always false, we need to
 143                  * handle synchonization on p4d level.
 144                  */
 145                 MAYBE_BUILD_BUG_ON(pgd_none(*pgd_ref));
 146                 p4d_ref = p4d_offset(pgd_ref, addr);
 147
 148                 if (p4d_none(*p4d_ref))
 149                         continue;
 150
 151                 spin_lock(&pgd_lock);
 152                 list_for_each_entry(page, &pgd_list, lru) {
 153                         pgd_t *pgd;
 154                         p4d_t *p4d;
 155                         spinlock_t *pgt_lock;
 156
 157                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
 158                         p4d = p4d_offset(pgd, addr);
 159                         /* the pgt_lock only for Xen */
 160                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 161                         spin_lock(pgt_lock);
 162
 163                         if (!p4d_none(*p4d_ref) && !p4d_none(*p4d))
 164                                 BUG_ON(p4d_page_vaddr(*p4d)
 165                                        != p4d_page_vaddr(*p4d_ref));
 166
 167                         if (p4d_none(*p4d))
 168                                 set_p4d(p4d, *p4d_ref);
 169
 170                         spin_unlock(pgt_lock);
 171                 }
 172                 spin_unlock(&pgd_lock);
 173         }
 174 }
 175
 176 /*
 177  * When memory was added make sure all the processes MM have
 178  * suitable PGD entries in the local PGD level page.
 179  */
 180 void sync_global_pgds(unsigned long start, unsigned long end)
 181 {
 182         if (pgtable_l5_enabled())
 183                 sync_global_pgds_l5(start, end);
 184         else
 185                 sync_global_pgds_l4(start, end);
 186 }
 187
 188 /*
 189  * NOTE: This function is marked __ref because it calls __init function
 190  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 191  */
 192 static __ref void *spp_getpage(void)
 193 {
 194         void *ptr;
 195
 196         if (after_bootmem)
 197                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 198         else
 199                 ptr = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
 200
 201         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 202                 panic("set_pte_phys: cannot allocate page data %s\n",
 203                         after_bootmem ? "after bootmem" : "");
 204         }
 205
 206         pr_debug("spp_getpage %p\n", ptr);
 207
 208         return ptr;
 209 }
 210
 211 static p4d_t *fill_p4d(pgd_t *pgd, unsigned long vaddr)
 212 {
 213         if (pgd_none(*pgd)) {
 214                 p4d_t *p4d = (p4d_t *)spp_getpage();
 215                 pgd_populate(&init_mm, pgd, p4d);
 216                 if (p4d != p4d_offset(pgd, 0))
 217                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 218                                p4d, p4d_offset(pgd, 0));
 219         }
 220         return p4d_offset(pgd, vaddr);
 221 }
 222
 223 static pud_t *fill_pud(p4d_t *p4d, unsigned long vaddr)
 224 {
 225         if (p4d_none(*p4d)) {
 226                 pud_t *pud = (pud_t *)spp_getpage();
 227                 p4d_populate(&init_mm, p4d, pud);
 228                 if (pud != pud_offset(p4d, 0))
 229                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 230                                pud, pud_offset(p4d, 0));
 231         }
 232         return pud_offset(p4d, vaddr);
 233 }
 234
 235 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 236 {
 237         if (pud_none(*pud)) {
 238                 pmd_t *pmd = (pmd_t *) spp_getpage();
 239                 pud_populate(&init_mm, pud, pmd);
 240                 if (pmd != pmd_offset(pud, 0))
 241                         printk(KERN_ERR "PAGETABLE BUG #02! %p <-> %p\n",
 242                                pmd, pmd_offset(pud, 0));
 243         }
 244         return pmd_offset(pud, vaddr);
 245 }
 246
 247 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 248 {
 249         if (pmd_none(*pmd)) {
 250                 pte_t *pte = (pte_t *) spp_getpage();
 251                 pmd_populate_kernel(&init_mm, pmd, pte);
 252                 if (pte != pte_offset_kernel(pmd, 0))
 253                         printk(KERN_ERR "PAGETABLE BUG #03!\n");
 254         }
 255         return pte_offset_kernel(pmd, vaddr);
 256 }
 257
 258 static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
 259 {
 260         pmd_t *pmd = fill_pmd(pud, vaddr);
 261         pte_t *pte = fill_pte(pmd, vaddr);
 262
 263         set_pte(pte, new_pte);
 264
 265         /*
 266          * It's enough to flush this one mapping.
 267          * (PGE mappings get flushed as well)
 268          */
 269         __flush_tlb_one_kernel(vaddr);
 270 }
 271
 272 void set_pte_vaddr_p4d(p4d_t *p4d_page, unsigned long vaddr, pte_t new_pte)
 273 {
 274         p4d_t *p4d = p4d_page + p4d_index(vaddr);
 275         pud_t *pud = fill_pud(p4d, vaddr);
 276
 277         __set_pte_vaddr(pud, vaddr, new_pte);
 278 }
 279
 280 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 281 {
 282         pud_t *pud = pud_page + pud_index(vaddr);
 283
 284         __set_pte_vaddr(pud, vaddr, new_pte);
 285 }
 286
 287 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 288 {
 289         pgd_t *pgd;
 290         p4d_t *p4d_page;
 291
 292         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 293
 294         pgd = pgd_offset_k(vaddr);
 295         if (pgd_none(*pgd)) {
 296                 printk(KERN_ERR
 297                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 298                 return;
 299         }
 300
 301         p4d_page = p4d_offset(pgd, 0);
 302         set_pte_vaddr_p4d(p4d_page, vaddr, pteval);
 303 }
 304
 305 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 306 {
 307         pgd_t *pgd;
 308         p4d_t *p4d;
 309         pud_t *pud;
 310
 311         pgd = pgd_offset_k(vaddr);
 312         p4d = fill_p4d(pgd, vaddr);
 313         pud = fill_pud(p4d, vaddr);
 314         return fill_pmd(pud, vaddr);
 315 }
 316
 317 pte_t * __init populate_extra_pte(unsigned long vaddr)
 318 {
 319         pmd_t *pmd;
 320
 321         pmd = populate_extra_pmd(vaddr);
 322         return fill_pte(pmd, vaddr);
 323 }
 324
 325 /*
 326  * Create large page table mappings for a range of physical addresses.
 327  */
 328 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 329                                         enum page_cache_mode cache)
 330 {
 331         pgd_t *pgd;
 332         p4d_t *p4d;
 333         pud_t *pud;
 334         pmd_t *pmd;
 335         pgprot_t prot;
 336
 337         pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
 338                 pgprot_val(pgprot_4k_2_large(cachemode2pgprot(cache)));
 339         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 340         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 341                 pgd = pgd_offset_k((unsigned long)__va(phys));
 342                 if (pgd_none(*pgd)) {
 343                         p4d = (p4d_t *) spp_getpage();
 344                         set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE |
 345                                                 _PAGE_USER));
 346                 }
 347                 p4d = p4d_offset(pgd, (unsigned long)__va(phys));
 348                 if (p4d_none(*p4d)) {
 349                         pud = (pud_t *) spp_getpage();
 350                         set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE |
 351                                                 _PAGE_USER));
 352                 }
 353                 pud = pud_offset(p4d, (unsigned long)__va(phys));
 354                 if (pud_none(*pud)) {
 355                         pmd = (pmd_t *) spp_getpage();
 356                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 357                                                 _PAGE_USER));
 358                 }
 359                 pmd = pmd_offset(pud, phys);
 360                 BUG_ON(!pmd_none(*pmd));
 361                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 362         }
 363 }
 364
 365 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 366 {
 367         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
 368 }
 369
 370 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 371 {
 372         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
 373 }
 374
 375 /*
 376  * The head.S code sets up the kernel high mapping:
 377  *
 378  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 379  *
 380  * phys_base holds the negative offset to the kernel, which is added
 381  * to the compile time generated pmds. This results in invalid pmds up
 382  * to the point where we hit the physaddr 0 mapping.
 383  *
 384  * We limit the mappings to the region from _text to _brk_end.  _brk_end
 385  * is rounded up to the 2MB boundary. This catches the invalid pmds as
 386  * well, as they are located before _text:
 387  */
 388 void __init cleanup_highmap(void)
 389 {
 390         unsigned long vaddr = __START_KERNEL_map;
 391         unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
 392         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 393         pmd_t *pmd = level2_kernel_pgt;
 394
 395         /*
 396          * Native path, max_pfn_mapped is not set yet.
 397          * Xen has valid max_pfn_mapped set in
 398          *      arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
 399          */
 400         if (max_pfn_mapped)
 401                 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 402
 403         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
 404                 if (pmd_none(*pmd))
 405                         continue;
 406                 if (vaddr < (unsigned long) _text || vaddr > end)
 407                         set_pmd(pmd, __pmd(0));
 408         }
 409 }
 410
 411 /*
 412  * Create PTE level page table mapping for physical addresses.
 413  * It returns the last physical address mapped.
 414  */
 415 static unsigned long __meminit
 416 phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
 417               pgprot_t prot)
 418 {
 419         unsigned long pages = 0, paddr_next;
 420         unsigned long paddr_last = paddr_end;
 421         pte_t *pte;
 422         int i;
 423
 424         pte = pte_page + pte_index(paddr);
 425         i = pte_index(paddr);
 426
 427         for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
 428                 paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
 429                 if (paddr >= paddr_end) {
 430                         if (!after_bootmem &&
 431                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 432                                              E820_TYPE_RAM) &&
 433                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 434                                              E820_TYPE_RESERVED_KERN))
 435                                 set_pte(pte, __pte(0));
 436                         continue;
 437                 }
 438
 439                 /*
 440                  * We will re-use the existing mapping.
 441                  * Xen for example has some special requirements, like mapping
 442                  * pagetable pages as RO. So assume someone who pre-setup
 443                  * these mappings are more intelligent.
 444                  */
 445                 if (!pte_none(*pte)) {
 446                         if (!after_bootmem)
 447                                 pages++;
 448                         continue;
 449                 }
 450
 451                 if (0)
 452                         pr_info("   pte=%p addr=%lx pte=%016lx\n", pte, paddr,
 453                                 pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 454                 pages++;
 455                 set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, prot));
 456                 paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
 457         }
 458
 459         update_page_count(PG_LEVEL_4K, pages);
 460
 461         return paddr_last;
 462 }
 463
 464 /*
 465  * Create PMD level page table mapping for physical addresses. The virtual
 466  * and physical address have to be aligned at this level.
 467  * It returns the last physical address mapped.
 468  */
 469 static unsigned long __meminit
 470 phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
 471               unsigned long page_size_mask, pgprot_t prot)
 472 {
 473         unsigned long pages = 0, paddr_next;
 474         unsigned long paddr_last = paddr_end;
 475
 476         int i = pmd_index(paddr);
 477
 478         for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
 479                 pmd_t *pmd = pmd_page + pmd_index(paddr);
 480                 pte_t *pte;
 481                 pgprot_t new_prot = prot;
 482
 483                 paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
 484                 if (paddr >= paddr_end) {
 485                         if (!after_bootmem &&
 486                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 487                                              E820_TYPE_RAM) &&
 488                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 489                                              E820_TYPE_RESERVED_KERN))
 490                                 set_pmd(pmd, __pmd(0));
 491                         continue;
 492                 }
 493
 494                 if (!pmd_none(*pmd)) {
 495                         if (!pmd_large(*pmd)) {
 496                                 spin_lock(&init_mm.page_table_lock);
 497                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
 498                                 paddr_last = phys_pte_init(pte, paddr,
 499                                                            paddr_end, prot);
 500                                 spin_unlock(&init_mm.page_table_lock);
 501                                 continue;
 502                         }
 503                         /*
 504                          * If we are ok with PG_LEVEL_2M mapping, then we will
 505                          * use the existing mapping,
 506                          *
 507                          * Otherwise, we will split the large page mapping but
 508                          * use the same existing protection bits except for
 509                          * large page, so that we don't violate Intel's TLB
 510                          * Application note (317080) which says, while changing
 511                          * the page sizes, new and old translations should
 512                          * not differ with respect to page frame and
 513                          * attributes.
 514                          */
 515                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 516                                 if (!after_bootmem)
 517                                         pages++;
 518                                 paddr_last = paddr_next;
 519                                 continue;
 520                         }
 521                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 522                 }
 523
 524                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 525                         pages++;
 526                         spin_lock(&init_mm.page_table_lock);
 527                         set_pte((pte_t *)pmd,
 528                                 pfn_pte((paddr & PMD_MASK) >> PAGE_SHIFT,
 529                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 530                         spin_unlock(&init_mm.page_table_lock);
 531                         paddr_last = paddr_next;
 532                         continue;
 533                 }
 534
 535                 pte = alloc_low_page();
 536                 paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot);
 537
 538                 spin_lock(&init_mm.page_table_lock);
 539                 pmd_populate_kernel(&init_mm, pmd, pte);
 540                 spin_unlock(&init_mm.page_table_lock);
 541         }
 542         update_page_count(PG_LEVEL_2M, pages);
 543         return paddr_last;
 544 }
 545
 546 /*
 547  * Create PUD level page table mapping for physical addresses. The virtual
 548  * and physical address do not have to be aligned at this level. KASLR can
 549  * randomize virtual addresses up to this level.
 550  * It returns the last physical address mapped.
 551  */
 552 static unsigned long __meminit
 553 phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
 554               unsigned long page_size_mask)
 555 {
 556         unsigned long pages = 0, paddr_next;
 557         unsigned long paddr_last = paddr_end;
 558         unsigned long vaddr = (unsigned long)__va(paddr);
 559         int i = pud_index(vaddr);
 560
 561         for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
 562                 pud_t *pud;
 563                 pmd_t *pmd;
 564                 pgprot_t prot = PAGE_KERNEL;
 565
 566                 vaddr = (unsigned long)__va(paddr);
 567                 pud = pud_page + pud_index(vaddr);
 568                 paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
 569
 570                 if (paddr >= paddr_end) {
 571                         if (!after_bootmem &&
 572                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 573                                              E820_TYPE_RAM) &&
 574                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 575                                              E820_TYPE_RESERVED_KERN))
 576                                 set_pud(pud, __pud(0));
 577                         continue;
 578                 }
 579
 580                 if (!pud_none(*pud)) {
 581                         if (!pud_large(*pud)) {
 582                                 pmd = pmd_offset(pud, 0);
 583                                 paddr_last = phys_pmd_init(pmd, paddr,
 584                                                            paddr_end,
 585                                                            page_size_mask,
 586                                                            prot);
 587                                 __flush_tlb_all();
 588                                 continue;
 589                         }
 590                         /*
 591                          * If we are ok with PG_LEVEL_1G mapping, then we will
 592                          * use the existing mapping.
 593                          *
 594                          * Otherwise, we will split the gbpage mapping but use
 595                          * the same existing protection  bits except for large
 596                          * page, so that we don't violate Intel's TLB
 597                          * Application note (317080) which says, while changing
 598                          * the page sizes, new and old translations should
 599                          * not differ with respect to page frame and
 600                          * attributes.
 601                          */
 602                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 603                                 if (!after_bootmem)
 604                                         pages++;
 605                                 paddr_last = paddr_next;
 606                                 continue;
 607                         }
 608                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 609                 }
 610
 611                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 612                         pages++;
 613                         spin_lock(&init_mm.page_table_lock);
 614                         set_pte((pte_t *)pud,
 615                                 pfn_pte((paddr & PUD_MASK) >> PAGE_SHIFT,
 616                                         PAGE_KERNEL_LARGE));
 617                         spin_unlock(&init_mm.page_table_lock);
 618                         paddr_last = paddr_next;
 619                         continue;
 620                 }
 621
 622                 pmd = alloc_low_page();
 623                 paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
 624                                            page_size_mask, prot);
 625
 626                 spin_lock(&init_mm.page_table_lock);
 627                 pud_populate(&init_mm, pud, pmd);
 628                 spin_unlock(&init_mm.page_table_lock);
 629         }
 630         __flush_tlb_all();
 631
 632         update_page_count(PG_LEVEL_1G, pages);
 633
 634         return paddr_last;
 635 }
 636
 637 static unsigned long __meminit
 638 phys_p4d_init(p4d_t *p4d_page, unsigned long paddr, unsigned long paddr_end,
 639               unsigned long page_size_mask)
 640 {
 641         unsigned long paddr_next, paddr_last = paddr_end;
 642         unsigned long vaddr = (unsigned long)__va(paddr);
 643         int i = p4d_index(vaddr);
 644
 645         if (!pgtable_l5_enabled())
 646                 return phys_pud_init((pud_t *) p4d_page, paddr, paddr_end, page_size_mask);
 647
 648         for (; i < PTRS_PER_P4D; i++, paddr = paddr_next) {
 649                 p4d_t *p4d;
 650                 pud_t *pud;
 651
 652                 vaddr = (unsigned long)__va(paddr);
 653                 p4d = p4d_page + p4d_index(vaddr);
 654                 paddr_next = (paddr & P4D_MASK) + P4D_SIZE;
 655
 656                 if (paddr >= paddr_end) {
 657                         if (!after_bootmem &&
 658                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 659                                              E820_TYPE_RAM) &&
 660                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 661                                              E820_TYPE_RESERVED_KERN))
 662                                 set_p4d(p4d, __p4d(0));
 663                         continue;
 664                 }
 665
 666                 if (!p4d_none(*p4d)) {
 667                         pud = pud_offset(p4d, 0);
 668                         paddr_last = phys_pud_init(pud, paddr,
 669                                         paddr_end,
 670                                         page_size_mask);
 671                         __flush_tlb_all();
 672                         continue;
 673                 }
 674
 675                 pud = alloc_low_page();
 676                 paddr_last = phys_pud_init(pud, paddr, paddr_end,
 677                                            page_size_mask);
 678
 679                 spin_lock(&init_mm.page_table_lock);
 680                 p4d_populate(&init_mm, p4d, pud);
 681                 spin_unlock(&init_mm.page_table_lock);
 682         }
 683         __flush_tlb_all();
 684
 685         return paddr_last;
 686 }
 687
 688 /*
 689  * Create page table mapping for the physical memory for specific physical
 690  * addresses. The virtual and physical addresses have to be aligned on PMD level
 691  * down. It returns the last physical address mapped.
 692  */
 693 unsigned long __meminit
 694 kernel_physical_mapping_init(unsigned long paddr_start,
 695                              unsigned long paddr_end,
 696                              unsigned long page_size_mask)
 697 {
 698         bool pgd_changed = false;
 699         unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
 700
 701         paddr_last = paddr_end;
 702         vaddr = (unsigned long)__va(paddr_start);
 703         vaddr_end = (unsigned long)__va(paddr_end);
 704         vaddr_start = vaddr;
 705
 706         for (; vaddr < vaddr_end; vaddr = vaddr_next) {
 707                 pgd_t *pgd = pgd_offset_k(vaddr);
 708                 p4d_t *p4d;
 709
 710                 vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
 711
 712                 if (pgd_val(*pgd)) {
 713                         p4d = (p4d_t *)pgd_page_vaddr(*pgd);
 714                         paddr_last = phys_p4d_init(p4d, __pa(vaddr),
 715                                                    __pa(vaddr_end),
 716                                                    page_size_mask);
 717                         continue;
 718                 }
 719
 720                 p4d = alloc_low_page();
 721                 paddr_last = phys_p4d_init(p4d, __pa(vaddr), __pa(vaddr_end),
 722                                            page_size_mask);
 723
 724                 spin_lock(&init_mm.page_table_lock);
 725                 if (pgtable_l5_enabled())
 726                         pgd_populate(&init_mm, pgd, p4d);
 727                 else
 728                         p4d_populate(&init_mm, p4d_offset(pgd, vaddr), (pud_t *) p4d);
 729                 spin_unlock(&init_mm.page_table_lock);
 730                 pgd_changed = true;
 731         }
 732
 733         if (pgd_changed)
 734                 sync_global_pgds(vaddr_start, vaddr_end - 1);
 735
 736         __flush_tlb_all();
 737
 738         return paddr_last;
 739 }
 740
 741 #ifndef CONFIG_NUMA
 742 void __init initmem_init(void)
 743 {
 744         memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
 745 }
 746 #endif
 747
 748 void __init paging_init(void)
 749 {
 750         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 751         sparse_init();
 752
 753         /*
 754          * clear the default setting with node 0
 755          * note: don't use nodes_clear here, that is really clearing when
 756          *       numa support is not compiled in, and later node_set_state
 757          *       will not set it back.
 758          */
 759         node_clear_state(0, N_MEMORY);
 760         if (N_MEMORY != N_NORMAL_MEMORY)
 761                 node_clear_state(0, N_NORMAL_MEMORY);
 762
 763         zone_sizes_init();
 764 }
 765
 766 /*
 767  * Memory hotplug specific functions
 768  */
 769 #ifdef CONFIG_MEMORY_HOTPLUG
 770 /*
 771  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 772  * updating.
 773  */
 774 static void update_end_of_memory_vars(u64 start, u64 size)
 775 {
 776         unsigned long end_pfn = PFN_UP(start + size);
 777
 778         if (end_pfn > max_pfn) {
 779                 max_pfn = end_pfn;
 780                 max_low_pfn = end_pfn;
 781                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 782         }
 783 }
 784
 785 int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
 786                 struct vmem_altmap *altmap, bool want_memblock)
 787 {
 788         int ret;
 789
 790         ret = __add_pages(nid, start_pfn, nr_pages, altmap, want_memblock);
 791         WARN_ON_ONCE(ret);
 792
 793         /* update max_pfn, max_low_pfn and high_memory */
 794         update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
 795                                   nr_pages << PAGE_SHIFT);
 796
 797         return ret;
 798 }
 799
 800 int arch_add_memory(int nid, u64 start, u64 size, struct vmem_altmap *altmap,
 801                 bool want_memblock)
 802 {
 803         unsigned long start_pfn = start >> PAGE_SHIFT;
 804         unsigned long nr_pages = size >> PAGE_SHIFT;
 805
 806         init_memory_mapping(start, start + size);
 807
 808         return add_pages(nid, start_pfn, nr_pages, altmap, want_memblock);
 809 }
 810
 811 #define PAGE_INUSE 0xFD
 812
 813 static void __meminit free_pagetable(struct page *page, int order)
 814 {
 815         unsigned long magic;
 816         unsigned int nr_pages = 1 << order;
 817
 818         /* bootmem page has reserved flag */
 819         if (PageReserved(page)) {
 820                 __ClearPageReserved(page);
 821
 822                 magic = (unsigned long)page->freelist;
 823                 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
 824                         while (nr_pages--)
 825                                 put_page_bootmem(page++);
 826                 } else
 827                         while (nr_pages--)
 828                                 free_reserved_page(page++);
 829         } else
 830                 free_pages((unsigned long)page_address(page), order);
 831 }
 832
 833 static void __meminit free_hugepage_table(struct page *page,
 834                 struct vmem_altmap *altmap)
 835 {
 836         if (altmap)
 837                 vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
 838         else
 839                 free_pagetable(page, get_order(PMD_SIZE));
 840 }
 841
 842 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
 843 {
 844         pte_t *pte;
 845         int i;
 846
 847         for (i = 0; i < PTRS_PER_PTE; i++) {
 848                 pte = pte_start + i;
 849                 if (!pte_none(*pte))
 850                         return;
 851         }
 852
 853         /* free a pte talbe */
 854         free_pagetable(pmd_page(*pmd), 0);
 855         spin_lock(&init_mm.page_table_lock);
 856         pmd_clear(pmd);
 857         spin_unlock(&init_mm.page_table_lock);
 858 }
 859
 860 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
 861 {
 862         pmd_t *pmd;
 863         int i;
 864
 865         for (i = 0; i < PTRS_PER_PMD; i++) {
 866                 pmd = pmd_start + i;
 867                 if (!pmd_none(*pmd))
 868                         return;
 869         }
 870
 871         /* free a pmd talbe */
 872         free_pagetable(pud_page(*pud), 0);
 873         spin_lock(&init_mm.page_table_lock);
 874         pud_clear(pud);
 875         spin_unlock(&init_mm.page_table_lock);
 876 }
 877
 878 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
 879 {
 880         pud_t *pud;
 881         int i;
 882
 883         for (i = 0; i < PTRS_PER_PUD; i++) {
 884                 pud = pud_start + i;
 885                 if (!pud_none(*pud))
 886                         return;
 887         }
 888
 889         /* free a pud talbe */
 890         free_pagetable(p4d_page(*p4d), 0);
 891         spin_lock(&init_mm.page_table_lock);
 892         p4d_clear(p4d);
 893         spin_unlock(&init_mm.page_table_lock);
 894 }
 895
 896 static void __meminit
 897 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
 898                  bool direct)
 899 {
 900         unsigned long next, pages = 0;
 901         pte_t *pte;
 902         void *page_addr;
 903         phys_addr_t phys_addr;
 904
 905         pte = pte_start + pte_index(addr);
 906         for (; addr < end; addr = next, pte++) {
 907                 next = (addr + PAGE_SIZE) & PAGE_MASK;
 908                 if (next > end)
 909                         next = end;
 910
 911                 if (!pte_present(*pte))
 912                         continue;
 913
 914                 /*
 915                  * We mapped [0,1G) memory as identity mapping when
 916                  * initializing, in arch/x86/kernel/head_64.S. These
 917                  * pagetables cannot be removed.
 918                  */
 919                 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
 920                 if (phys_addr < (phys_addr_t)0x40000000)
 921                         return;
 922
 923                 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
 924                         /*
 925                          * Do not free direct mapping pages since they were
 926                          * freed when offlining, or simplely not in use.
 927                          */
 928                         if (!direct)
 929                                 free_pagetable(pte_page(*pte), 0);
 930
 931                         spin_lock(&init_mm.page_table_lock);
 932                         pte_clear(&init_mm, addr, pte);
 933                         spin_unlock(&init_mm.page_table_lock);
 934
 935                         /* For non-direct mapping, pages means nothing. */
 936                         pages++;
 937                 } else {
 938                         /*
 939                          * If we are here, we are freeing vmemmap pages since
 940                          * direct mapped memory ranges to be freed are aligned.
 941                          *
 942                          * If we are not removing the whole page, it means
 943                          * other page structs in this page are being used and
 944                          * we canot remove them. So fill the unused page_structs
 945                          * with 0xFD, and remove the page when it is wholly
 946                          * filled with 0xFD.
 947                          */
 948                         memset((void *)addr, PAGE_INUSE, next - addr);
 949
 950                         page_addr = page_address(pte_page(*pte));
 951                         if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
 952                                 free_pagetable(pte_page(*pte), 0);
 953
 954                                 spin_lock(&init_mm.page_table_lock);
 955                                 pte_clear(&init_mm, addr, pte);
 956                                 spin_unlock(&init_mm.page_table_lock);
 957                         }
 958                 }
 959         }
 960
 961         /* Call free_pte_table() in remove_pmd_table(). */
 962         flush_tlb_all();
 963         if (direct)
 964                 update_page_count(PG_LEVEL_4K, -pages);
 965 }
 966
 967 static void __meminit
 968 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
 969                  bool direct, struct vmem_altmap *altmap)
 970 {
 971         unsigned long next, pages = 0;
 972         pte_t *pte_base;
 973         pmd_t *pmd;
 974         void *page_addr;
 975
 976         pmd = pmd_start + pmd_index(addr);
 977         for (; addr < end; addr = next, pmd++) {
 978                 next = pmd_addr_end(addr, end);
 979
 980                 if (!pmd_present(*pmd))
 981                         continue;
 982
 983                 if (pmd_large(*pmd)) {
 984                         if (IS_ALIGNED(addr, PMD_SIZE) &&
 985                             IS_ALIGNED(next, PMD_SIZE)) {
 986                                 if (!direct)
 987                                         free_hugepage_table(pmd_page(*pmd),
 988                                                             altmap);
 989
 990                                 spin_lock(&init_mm.page_table_lock);
 991                                 pmd_clear(pmd);
 992                                 spin_unlock(&init_mm.page_table_lock);
 993                                 pages++;
 994                         } else {
 995                                 /* If here, we are freeing vmemmap pages. */
 996                                 memset((void *)addr, PAGE_INUSE, next - addr);
 997
 998                                 page_addr = page_address(pmd_page(*pmd));
 999                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1000                                                 PMD_SIZE)) {
1001                                         free_hugepage_table(pmd_page(*pmd),
1002                                                             altmap);
1003
1004                                         spin_lock(&init_mm.page_table_lock);
1005                                         pmd_clear(pmd);
1006                                         spin_unlock(&init_mm.page_table_lock);
1007                                 }
1008                         }
1009
1010                         continue;
1011                 }
1012
1013                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1014                 remove_pte_table(pte_base, addr, next, direct);
1015                 free_pte_table(pte_base, pmd);
1016         }
1017
1018         /* Call free_pmd_table() in remove_pud_table(). */
1019         if (direct)
1020                 update_page_count(PG_LEVEL_2M, -pages);
1021 }
1022
1023 static void __meminit
1024 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1025                  struct vmem_altmap *altmap, bool direct)
1026 {
1027         unsigned long next, pages = 0;
1028         pmd_t *pmd_base;
1029         pud_t *pud;
1030         void *page_addr;
1031
1032         pud = pud_start + pud_index(addr);
1033         for (; addr < end; addr = next, pud++) {
1034                 next = pud_addr_end(addr, end);
1035
1036                 if (!pud_present(*pud))
1037                         continue;
1038
1039                 if (pud_large(*pud)) {
1040                         if (IS_ALIGNED(addr, PUD_SIZE) &&
1041                             IS_ALIGNED(next, PUD_SIZE)) {
1042                                 if (!direct)
1043                                         free_pagetable(pud_page(*pud),
1044                                                        get_order(PUD_SIZE));
1045
1046                                 spin_lock(&init_mm.page_table_lock);
1047                                 pud_clear(pud);
1048                                 spin_unlock(&init_mm.page_table_lock);
1049                                 pages++;
1050                         } else {
1051                                 /* If here, we are freeing vmemmap pages. */
1052                                 memset((void *)addr, PAGE_INUSE, next - addr);
1053
1054                                 page_addr = page_address(pud_page(*pud));
1055                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1056                                                 PUD_SIZE)) {
1057                                         free_pagetable(pud_page(*pud),
1058                                                        get_order(PUD_SIZE));
1059
1060                                         spin_lock(&init_mm.page_table_lock);
1061                                         pud_clear(pud);
1062                                         spin_unlock(&init_mm.page_table_lock);
1063                                 }
1064                         }
1065
1066                         continue;
1067                 }
1068
1069                 pmd_base = pmd_offset(pud, 0);
1070                 remove_pmd_table(pmd_base, addr, next, direct, altmap);
1071                 free_pmd_table(pmd_base, pud);
1072         }
1073
1074         if (direct)
1075                 update_page_count(PG_LEVEL_1G, -pages);
1076 }
1077
1078 static void __meminit
1079 remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1080                  struct vmem_altmap *altmap, bool direct)
1081 {
1082         unsigned long next, pages = 0;
1083         pud_t *pud_base;
1084         p4d_t *p4d;
1085
1086         p4d = p4d_start + p4d_index(addr);
1087         for (; addr < end; addr = next, p4d++) {
1088                 next = p4d_addr_end(addr, end);
1089
1090                 if (!p4d_present(*p4d))
1091                         continue;
1092
1093                 BUILD_BUG_ON(p4d_large(*p4d));
1094
1095                 pud_base = pud_offset(p4d, 0);
1096                 remove_pud_table(pud_base, addr, next, altmap, direct);
1097                 /*
1098                  * For 4-level page tables we do not want to free PUDs, but in the
1099                  * 5-level case we should free them. This code will have to change
1100                  * to adapt for boot-time switching between 4 and 5 level page tables.
1101                  */
1102                 if (pgtable_l5_enabled())
1103                         free_pud_table(pud_base, p4d);
1104         }
1105
1106         if (direct)
1107                 update_page_count(PG_LEVEL_512G, -pages);
1108 }
1109
1110 /* start and end are both virtual address. */
1111 static void __meminit
1112 remove_pagetable(unsigned long start, unsigned long end, bool direct,
1113                 struct vmem_altmap *altmap)
1114 {
1115         unsigned long next;
1116         unsigned long addr;
1117         pgd_t *pgd;
1118         p4d_t *p4d;
1119
1120         for (addr = start; addr < end; addr = next) {
1121                 next = pgd_addr_end(addr, end);
1122
1123                 pgd = pgd_offset_k(addr);
1124                 if (!pgd_present(*pgd))
1125                         continue;
1126
1127                 p4d = p4d_offset(pgd, 0);
1128                 remove_p4d_table(p4d, addr, next, altmap, direct);
1129         }
1130
1131         flush_tlb_all();
1132 }
1133
1134 void __ref vmemmap_free(unsigned long start, unsigned long end,
1135                 struct vmem_altmap *altmap)
1136 {
1137         remove_pagetable(start, end, false, altmap);
1138 }
1139
1140 #ifdef CONFIG_MEMORY_HOTREMOVE
1141 static void __meminit
1142 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1143 {
1144         start = (unsigned long)__va(start);
1145         end = (unsigned long)__va(end);
1146
1147         remove_pagetable(start, end, true, NULL);
1148 }
1149
1150 int __ref arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
1151 {
1152         unsigned long start_pfn = start >> PAGE_SHIFT;
1153         unsigned long nr_pages = size >> PAGE_SHIFT;
1154         struct page *page = pfn_to_page(start_pfn);
1155         struct zone *zone;
1156         int ret;
1157
1158         /* With altmap the first mapped page is offset from @start */
1159         if (altmap)
1160                 page += vmem_altmap_offset(altmap);
1161         zone = page_zone(page);
1162         ret = __remove_pages(zone, start_pfn, nr_pages, altmap);
1163         WARN_ON_ONCE(ret);
1164         kernel_physical_mapping_remove(start, start + size);
1165
1166         return ret;
1167 }
1168 #endif
1169 #endif /* CONFIG_MEMORY_HOTPLUG */
1170
1171 static struct kcore_list kcore_vsyscall;
1172
1173 static void __init register_page_bootmem_info(void)
1174 {
1175 #ifdef CONFIG_NUMA
1176         int i;
1177
1178         for_each_online_node(i)
1179                 register_page_bootmem_info_node(NODE_DATA(i));
1180 #endif
1181 }
1182
1183 void __init mem_init(void)
1184 {
1185         pci_iommu_alloc();
1186
1187         /* clear_bss() already clear the empty_zero_page */
1188
1189         /* this will put all memory onto the freelists */
1190         memblock_free_all();
1191         after_bootmem = 1;
1192         x86_init.hyper.init_after_bootmem();
1193
1194         /*
1195          * Must be done after boot memory is put on freelist, because here we
1196          * might set fields in deferred struct pages that have not yet been
1197          * initialized, and memblock_free_all() initializes all the reserved
1198          * deferred pages for us.
1199          */
1200         register_page_bootmem_info();
1201
1202         /* Register memory areas for /proc/kcore */
1203         if (get_gate_vma(&init_mm))
1204                 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1205
1206         mem_init_print_info(NULL);
1207 }
1208
1209 int kernel_set_to_readonly;
1210
1211 void set_kernel_text_rw(void)
1212 {
1213         unsigned long start = PFN_ALIGN(_text);
1214         unsigned long end = PFN_ALIGN(__stop___ex_table);
1215
1216         if (!kernel_set_to_readonly)
1217                 return;
1218
1219         pr_debug("Set kernel text: %lx - %lx for read write\n",
1220                  start, end);
1221
1222         /*
1223          * Make the kernel identity mapping for text RW. Kernel text
1224          * mapping will always be RO. Refer to the comment in
1225          * static_protections() in pageattr.c
1226          */
1227         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1228 }
1229
1230 void set_kernel_text_ro(void)
1231 {
1232         unsigned long start = PFN_ALIGN(_text);
1233         unsigned long end = PFN_ALIGN(__stop___ex_table);
1234
1235         if (!kernel_set_to_readonly)
1236                 return;
1237
1238         pr_debug("Set kernel text: %lx - %lx for read only\n",
1239                  start, end);
1240
1241         /*
1242          * Set the kernel identity mapping for text RO.
1243          */
1244         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1245 }
1246
1247 void mark_rodata_ro(void)
1248 {
1249         unsigned long start = PFN_ALIGN(_text);
1250         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1251         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1252         unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1253         unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1254         unsigned long all_end;
1255
1256         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1257                (end - start) >> 10);
1258         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1259
1260         kernel_set_to_readonly = 1;
1261
1262         /*
1263          * The rodata/data/bss/brk section (but not the kernel text!)
1264          * should also be not-executable.
1265          *
1266          * We align all_end to PMD_SIZE because the existing mapping
1267          * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1268          * split the PMD and the reminder between _brk_end and the end
1269          * of the PMD will remain mapped executable.
1270          *
1271          * Any PMD which was setup after the one which covers _brk_end
1272          * has been zapped already via cleanup_highmem().
1273          */
1274         all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1275         set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1276
1277 #ifdef CONFIG_CPA_DEBUG
1278         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1279         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1280
1281         printk(KERN_INFO "Testing CPA: again\n");
1282         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1283 #endif
1284
1285         free_kernel_image_pages((void *)text_end, (void *)rodata_start);
1286         free_kernel_image_pages((void *)rodata_end, (void *)_sdata);
1287
1288         debug_checkwx();
1289 }
1290
1291 int kern_addr_valid(unsigned long addr)
1292 {
1293         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1294         pgd_t *pgd;
1295         p4d_t *p4d;
1296         pud_t *pud;
1297         pmd_t *pmd;
1298         pte_t *pte;
1299
1300         if (above != 0 && above != -1UL)
1301                 return 0;
1302
1303         pgd = pgd_offset_k(addr);
1304         if (pgd_none(*pgd))
1305                 return 0;
1306
1307         p4d = p4d_offset(pgd, addr);
1308         if (p4d_none(*p4d))
1309                 return 0;
1310
1311         pud = pud_offset(p4d, addr);
1312         if (pud_none(*pud))
1313                 return 0;
1314
1315         if (pud_large(*pud))
1316                 return pfn_valid(pud_pfn(*pud));
1317
1318         pmd = pmd_offset(pud, addr);
1319         if (pmd_none(*pmd))
1320                 return 0;
1321
1322         if (pmd_large(*pmd))
1323                 return pfn_valid(pmd_pfn(*pmd));
1324
1325         pte = pte_offset_kernel(pmd, addr);
1326         if (pte_none(*pte))
1327                 return 0;
1328
1329         return pfn_valid(pte_pfn(*pte));
1330 }
1331
1332 /*
1333  * Block size is the minimum amount of memory which can be hotplugged or
1334  * hotremoved. It must be power of two and must be equal or larger than
1335  * MIN_MEMORY_BLOCK_SIZE.
1336  */
1337 #define MAX_BLOCK_SIZE (2UL << 30)
1338
1339 /* Amount of ram needed to start using large blocks */
1340 #define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1341
1342 /* Adjustable memory block size */
1343 static unsigned long set_memory_block_size;
1344 int __init set_memory_block_size_order(unsigned int order)
1345 {
1346         unsigned long size = 1UL << order;
1347
1348         if (size > MEM_SIZE_FOR_LARGE_BLOCK || size < MIN_MEMORY_BLOCK_SIZE)
1349                 return -EINVAL;
1350
1351         set_memory_block_size = size;
1352         return 0;
1353 }
1354
1355 static unsigned long probe_memory_block_size(void)
1356 {
1357         unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1358         unsigned long bz;
1359
1360         /* If memory block size has been set, then use it */
1361         bz = set_memory_block_size;
1362         if (bz)
1363                 goto done;
1364
1365         /* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1366         if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1367                 bz = MIN_MEMORY_BLOCK_SIZE;
1368                 goto done;
1369         }
1370
1371         /* Find the largest allowed block size that aligns to memory end */
1372         for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1373                 if (IS_ALIGNED(boot_mem_end, bz))
1374                         break;
1375         }
1376 done:
1377         pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1378
1379         return bz;
1380 }
1381
1382 static unsigned long memory_block_size_probed;
1383 unsigned long memory_block_size_bytes(void)
1384 {
1385         if (!memory_block_size_probed)
1386                 memory_block_size_probed = probe_memory_block_size();
1387
1388         return memory_block_size_probed;
1389 }
1390
1391 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1392 /*
1393  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1394  */
1395 static long __meminitdata addr_start, addr_end;
1396 static void __meminitdata *p_start, *p_end;
1397 static int __meminitdata node_start;
1398
1399 static int __meminit vmemmap_populate_hugepages(unsigned long start,
1400                 unsigned long end, int node, struct vmem_altmap *altmap)
1401 {
1402         unsigned long addr;
1403         unsigned long next;
1404         pgd_t *pgd;
1405         p4d_t *p4d;
1406         pud_t *pud;
1407         pmd_t *pmd;
1408
1409         for (addr = start; addr < end; addr = next) {
1410                 next = pmd_addr_end(addr, end);
1411
1412                 pgd = vmemmap_pgd_populate(addr, node);
1413                 if (!pgd)
1414                         return -ENOMEM;
1415
1416                 p4d = vmemmap_p4d_populate(pgd, addr, node);
1417                 if (!p4d)
1418                         return -ENOMEM;
1419
1420                 pud = vmemmap_pud_populate(p4d, addr, node);
1421                 if (!pud)
1422                         return -ENOMEM;
1423
1424                 pmd = pmd_offset(pud, addr);
1425                 if (pmd_none(*pmd)) {
1426                         void *p;
1427
1428                         if (altmap)
1429                                 p = altmap_alloc_block_buf(PMD_SIZE, altmap);
1430                         else
1431                                 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1432                         if (p) {
1433                                 pte_t entry;
1434
1435                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1436                                                 PAGE_KERNEL_LARGE);
1437                                 set_pmd(pmd, __pmd(pte_val(entry)));
1438
1439                                 /* check to see if we have contiguous blocks */
1440                                 if (p_end != p || node_start != node) {
1441                                         if (p_start)
1442                                                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1443                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1444                                         addr_start = addr;
1445                                         node_start = node;
1446                                         p_start = p;
1447                                 }
1448
1449                                 addr_end = addr + PMD_SIZE;
1450                                 p_end = p + PMD_SIZE;
1451                                 continue;
1452                         } else if (altmap)
1453                                 return -ENOMEM; /* no fallback */
1454                 } else if (pmd_large(*pmd)) {
1455                         vmemmap_verify((pte_t *)pmd, node, addr, next);
1456                         continue;
1457                 }
1458                 if (vmemmap_populate_basepages(addr, next, node))
1459                         return -ENOMEM;
1460         }
1461         return 0;
1462 }
1463
1464 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1465                 struct vmem_altmap *altmap)
1466 {
1467         int err;
1468
1469         if (boot_cpu_has(X86_FEATURE_PSE))
1470                 err = vmemmap_populate_hugepages(start, end, node, altmap);
1471         else if (altmap) {
1472                 pr_err_once("%s: no cpu support for altmap allocations\n",
1473                                 __func__);
1474                 err = -ENOMEM;
1475         } else
1476                 err = vmemmap_populate_basepages(start, end, node);
1477         if (!err)
1478                 sync_global_pgds(start, end - 1);
1479         return err;
1480 }
1481
1482 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1483 void register_page_bootmem_memmap(unsigned long section_nr,
1484                                   struct page *start_page, unsigned long nr_pages)
1485 {
1486         unsigned long addr = (unsigned long)start_page;
1487         unsigned long end = (unsigned long)(start_page + nr_pages);
1488         unsigned long next;
1489         pgd_t *pgd;
1490         p4d_t *p4d;
1491         pud_t *pud;
1492         pmd_t *pmd;
1493         unsigned int nr_pmd_pages;
1494         struct page *page;
1495
1496         for (; addr < end; addr = next) {
1497                 pte_t *pte = NULL;
1498
1499                 pgd = pgd_offset_k(addr);
1500                 if (pgd_none(*pgd)) {
1501                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1502                         continue;
1503                 }
1504                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1505
1506                 p4d = p4d_offset(pgd, addr);
1507                 if (p4d_none(*p4d)) {
1508                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1509                         continue;
1510                 }
1511                 get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1512
1513                 pud = pud_offset(p4d, addr);
1514                 if (pud_none(*pud)) {
1515                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1516                         continue;
1517                 }
1518                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1519
1520                 if (!boot_cpu_has(X86_FEATURE_PSE)) {
1521                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1522                         pmd = pmd_offset(pud, addr);
1523                         if (pmd_none(*pmd))
1524                                 continue;
1525                         get_page_bootmem(section_nr, pmd_page(*pmd),
1526                                          MIX_SECTION_INFO);
1527
1528                         pte = pte_offset_kernel(pmd, addr);
1529                         if (pte_none(*pte))
1530                                 continue;
1531                         get_page_bootmem(section_nr, pte_page(*pte),
1532                                          SECTION_INFO);
1533                 } else {
1534                         next = pmd_addr_end(addr, end);
1535
1536                         pmd = pmd_offset(pud, addr);
1537                         if (pmd_none(*pmd))
1538                                 continue;
1539
1540                         nr_pmd_pages = 1 << get_order(PMD_SIZE);
1541                         page = pmd_page(*pmd);
1542                         while (nr_pmd_pages--)
1543                                 get_page_bootmem(section_nr, page++,
1544                                                  SECTION_INFO);
1545                 }
1546         }
1547 }
1548 #endif
1549
1550 void __meminit vmemmap_populate_print_last(void)
1551 {
1552         if (p_start) {
1553                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1554                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1555                 p_start = NULL;
1556                 p_end = NULL;
1557                 node_start = 0;
1558         }
1559 }
1560 #endif