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 /* Bits supported by the hardware: */
  69 pteval_t __supported_pte_mask __read_mostly = ~0;
  70 /* Bits allowed in normal kernel mappings: */
  71 pteval_t __default_kernel_pte_mask __read_mostly = ~0;
  72 EXPORT_SYMBOL_GPL(__supported_pte_mask);
  73 /* Used in PAGE_KERNEL_* macros which are reasonably used out-of-tree: */
  74 EXPORT_SYMBOL(__default_kernel_pte_mask);
  75
  76 int force_personality32;
  77
  78 /*
  79  * noexec32=on|off
  80  * Control non executable heap for 32bit processes.
  81  * To control the stack too use noexec=off
  82  *
  83  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
  84  * off  PROT_READ implies PROT_EXEC
  85  */
  86 static int __init nonx32_setup(char *str)
  87 {
  88         if (!strcmp(str, "on"))
  89                 force_personality32 &= ~READ_IMPLIES_EXEC;
  90         else if (!strcmp(str, "off"))
  91                 force_personality32 |= READ_IMPLIES_EXEC;
  92         return 1;
  93 }
  94 __setup("noexec32=", nonx32_setup);
  95
  96 static void sync_global_pgds_l5(unsigned long start, unsigned long end)
  97 {
  98         unsigned long addr;
  99
 100         for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
 101                 const pgd_t *pgd_ref = pgd_offset_k(addr);
 102                 struct page *page;
 103
 104                 /* Check for overflow */
 105                 if (addr < start)
 106                         break;
 107
 108                 if (pgd_none(*pgd_ref))
 109                         continue;
 110
 111                 spin_lock(&pgd_lock);
 112                 list_for_each_entry(page, &pgd_list, lru) {
 113                         pgd_t *pgd;
 114                         spinlock_t *pgt_lock;
 115
 116                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
 117                         /* the pgt_lock only for Xen */
 118                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 119                         spin_lock(pgt_lock);
 120
 121                         if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
 122                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 123
 124                         if (pgd_none(*pgd))
 125                                 set_pgd(pgd, *pgd_ref);
 126
 127                         spin_unlock(pgt_lock);
 128                 }
 129                 spin_unlock(&pgd_lock);
 130         }
 131 }
 132
 133 static void sync_global_pgds_l4(unsigned long start, unsigned long end)
 134 {
 135         unsigned long addr;
 136
 137         for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
 138                 pgd_t *pgd_ref = pgd_offset_k(addr);
 139                 const p4d_t *p4d_ref;
 140                 struct page *page;
 141
 142                 /*
 143                  * With folded p4d, pgd_none() is always false, we need to
 144                  * handle synchonization on p4d level.
 145                  */
 146                 MAYBE_BUILD_BUG_ON(pgd_none(*pgd_ref));
 147                 p4d_ref = p4d_offset(pgd_ref, addr);
 148
 149                 if (p4d_none(*p4d_ref))
 150                         continue;
 151
 152                 spin_lock(&pgd_lock);
 153                 list_for_each_entry(page, &pgd_list, lru) {
 154                         pgd_t *pgd;
 155                         p4d_t *p4d;
 156                         spinlock_t *pgt_lock;
 157
 158                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
 159                         p4d = p4d_offset(pgd, addr);
 160                         /* the pgt_lock only for Xen */
 161                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 162                         spin_lock(pgt_lock);
 163
 164                         if (!p4d_none(*p4d_ref) && !p4d_none(*p4d))
 165                                 BUG_ON(p4d_page_vaddr(*p4d)
 166                                        != p4d_page_vaddr(*p4d_ref));
 167
 168                         if (p4d_none(*p4d))
 169                                 set_p4d(p4d, *p4d_ref);
 170
 171                         spin_unlock(pgt_lock);
 172                 }
 173                 spin_unlock(&pgd_lock);
 174         }
 175 }
 176
 177 /*
 178  * When memory was added make sure all the processes MM have
 179  * suitable PGD entries in the local PGD level page.
 180  */
 181 void sync_global_pgds(unsigned long start, unsigned long end)
 182 {
 183         if (pgtable_l5_enabled)
 184                 sync_global_pgds_l5(start, end);
 185         else
 186                 sync_global_pgds_l4(start, end);
 187 }
 188
 189 /*
 190  * NOTE: This function is marked __ref because it calls __init function
 191  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 192  */
 193 static __ref void *spp_getpage(void)
 194 {
 195         void *ptr;
 196
 197         if (after_bootmem)
 198                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 199         else
 200                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 201
 202         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 203                 panic("set_pte_phys: cannot allocate page data %s\n",
 204                         after_bootmem ? "after bootmem" : "");
 205         }
 206
 207         pr_debug("spp_getpage %p\n", ptr);
 208
 209         return ptr;
 210 }
 211
 212 static p4d_t *fill_p4d(pgd_t *pgd, unsigned long vaddr)
 213 {
 214         if (pgd_none(*pgd)) {
 215                 p4d_t *p4d = (p4d_t *)spp_getpage();
 216                 pgd_populate(&init_mm, pgd, p4d);
 217                 if (p4d != p4d_offset(pgd, 0))
 218                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 219                                p4d, p4d_offset(pgd, 0));
 220         }
 221         return p4d_offset(pgd, vaddr);
 222 }
 223
 224 static pud_t *fill_pud(p4d_t *p4d, unsigned long vaddr)
 225 {
 226         if (p4d_none(*p4d)) {
 227                 pud_t *pud = (pud_t *)spp_getpage();
 228                 p4d_populate(&init_mm, p4d, pud);
 229                 if (pud != pud_offset(p4d, 0))
 230                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 231                                pud, pud_offset(p4d, 0));
 232         }
 233         return pud_offset(p4d, vaddr);
 234 }
 235
 236 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 237 {
 238         if (pud_none(*pud)) {
 239                 pmd_t *pmd = (pmd_t *) spp_getpage();
 240                 pud_populate(&init_mm, pud, pmd);
 241                 if (pmd != pmd_offset(pud, 0))
 242                         printk(KERN_ERR "PAGETABLE BUG #02! %p <-> %p\n",
 243                                pmd, pmd_offset(pud, 0));
 244         }
 245         return pmd_offset(pud, vaddr);
 246 }
 247
 248 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 249 {
 250         if (pmd_none(*pmd)) {
 251                 pte_t *pte = (pte_t *) spp_getpage();
 252                 pmd_populate_kernel(&init_mm, pmd, pte);
 253                 if (pte != pte_offset_kernel(pmd, 0))
 254                         printk(KERN_ERR "PAGETABLE BUG #03!\n");
 255         }
 256         return pte_offset_kernel(pmd, vaddr);
 257 }
 258
 259 static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
 260 {
 261         pmd_t *pmd = fill_pmd(pud, vaddr);
 262         pte_t *pte = fill_pte(pmd, vaddr);
 263
 264         set_pte(pte, new_pte);
 265
 266         /*
 267          * It's enough to flush this one mapping.
 268          * (PGE mappings get flushed as well)
 269          */
 270         __flush_tlb_one_kernel(vaddr);
 271 }
 272
 273 void set_pte_vaddr_p4d(p4d_t *p4d_page, unsigned long vaddr, pte_t new_pte)
 274 {
 275         p4d_t *p4d = p4d_page + p4d_index(vaddr);
 276         pud_t *pud = fill_pud(p4d, vaddr);
 277
 278         __set_pte_vaddr(pud, vaddr, new_pte);
 279 }
 280
 281 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 282 {
 283         pud_t *pud = pud_page + pud_index(vaddr);
 284
 285         __set_pte_vaddr(pud, vaddr, new_pte);
 286 }
 287
 288 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 289 {
 290         pgd_t *pgd;
 291         p4d_t *p4d_page;
 292
 293         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 294
 295         pgd = pgd_offset_k(vaddr);
 296         if (pgd_none(*pgd)) {
 297                 printk(KERN_ERR
 298                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 299                 return;
 300         }
 301
 302         p4d_page = p4d_offset(pgd, 0);
 303         set_pte_vaddr_p4d(p4d_page, vaddr, pteval);
 304 }
 305
 306 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 307 {
 308         pgd_t *pgd;
 309         p4d_t *p4d;
 310         pud_t *pud;
 311
 312         pgd = pgd_offset_k(vaddr);
 313         p4d = fill_p4d(pgd, vaddr);
 314         pud = fill_pud(p4d, vaddr);
 315         return fill_pmd(pud, vaddr);
 316 }
 317
 318 pte_t * __init populate_extra_pte(unsigned long vaddr)
 319 {
 320         pmd_t *pmd;
 321
 322         pmd = populate_extra_pmd(vaddr);
 323         return fill_pte(pmd, vaddr);
 324 }
 325
 326 /*
 327  * Create large page table mappings for a range of physical addresses.
 328  */
 329 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 330                                         enum page_cache_mode cache)
 331 {
 332         pgd_t *pgd;
 333         p4d_t *p4d;
 334         pud_t *pud;
 335         pmd_t *pmd;
 336         pgprot_t prot;
 337
 338         pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
 339                 pgprot_val(pgprot_4k_2_large(cachemode2pgprot(cache)));
 340         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 341         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 342                 pgd = pgd_offset_k((unsigned long)__va(phys));
 343                 if (pgd_none(*pgd)) {
 344                         p4d = (p4d_t *) spp_getpage();
 345                         set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE |
 346                                                 _PAGE_USER));
 347                 }
 348                 p4d = p4d_offset(pgd, (unsigned long)__va(phys));
 349                 if (p4d_none(*p4d)) {
 350                         pud = (pud_t *) spp_getpage();
 351                         set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE |
 352                                                 _PAGE_USER));
 353                 }
 354                 pud = pud_offset(p4d, (unsigned long)__va(phys));
 355                 if (pud_none(*pud)) {
 356                         pmd = (pmd_t *) spp_getpage();
 357                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 358                                                 _PAGE_USER));
 359                 }
 360                 pmd = pmd_offset(pud, phys);
 361                 BUG_ON(!pmd_none(*pmd));
 362                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 363         }
 364 }
 365
 366 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 367 {
 368         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
 369 }
 370
 371 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 372 {
 373         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
 374 }
 375
 376 /*
 377  * The head.S code sets up the kernel high mapping:
 378  *
 379  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 380  *
 381  * phys_base holds the negative offset to the kernel, which is added
 382  * to the compile time generated pmds. This results in invalid pmds up
 383  * to the point where we hit the physaddr 0 mapping.
 384  *
 385  * We limit the mappings to the region from _text to _brk_end.  _brk_end
 386  * is rounded up to the 2MB boundary. This catches the invalid pmds as
 387  * well, as they are located before _text:
 388  */
 389 void __init cleanup_highmap(void)
 390 {
 391         unsigned long vaddr = __START_KERNEL_map;
 392         unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
 393         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 394         pmd_t *pmd = level2_kernel_pgt;
 395
 396         /*
 397          * Native path, max_pfn_mapped is not set yet.
 398          * Xen has valid max_pfn_mapped set in
 399          *      arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
 400          */
 401         if (max_pfn_mapped)
 402                 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 403
 404         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
 405                 if (pmd_none(*pmd))
 406                         continue;
 407                 if (vaddr < (unsigned long) _text || vaddr > end)
 408                         set_pmd(pmd, __pmd(0));
 409         }
 410 }
 411
 412 /*
 413  * Create PTE level page table mapping for physical addresses.
 414  * It returns the last physical address mapped.
 415  */
 416 static unsigned long __meminit
 417 phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
 418               pgprot_t prot)
 419 {
 420         unsigned long pages = 0, paddr_next;
 421         unsigned long paddr_last = paddr_end;
 422         pte_t *pte;
 423         int i;
 424
 425         pte = pte_page + pte_index(paddr);
 426         i = pte_index(paddr);
 427
 428         for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
 429                 paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
 430                 if (paddr >= paddr_end) {
 431                         if (!after_bootmem &&
 432                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 433                                              E820_TYPE_RAM) &&
 434                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 435                                              E820_TYPE_RESERVED_KERN))
 436                                 set_pte(pte, __pte(0));
 437                         continue;
 438                 }
 439
 440                 /*
 441                  * We will re-use the existing mapping.
 442                  * Xen for example has some special requirements, like mapping
 443                  * pagetable pages as RO. So assume someone who pre-setup
 444                  * these mappings are more intelligent.
 445                  */
 446                 if (!pte_none(*pte)) {
 447                         if (!after_bootmem)
 448                                 pages++;
 449                         continue;
 450                 }
 451
 452                 if (0)
 453                         pr_info("   pte=%p addr=%lx pte=%016lx\n", pte, paddr,
 454                                 pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 455                 pages++;
 456                 set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, prot));
 457                 paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
 458         }
 459
 460         update_page_count(PG_LEVEL_4K, pages);
 461
 462         return paddr_last;
 463 }
 464
 465 /*
 466  * Create PMD level page table mapping for physical addresses. The virtual
 467  * and physical address have to be aligned at this level.
 468  * It returns the last physical address mapped.
 469  */
 470 static unsigned long __meminit
 471 phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
 472               unsigned long page_size_mask, pgprot_t prot)
 473 {
 474         unsigned long pages = 0, paddr_next;
 475         unsigned long paddr_last = paddr_end;
 476
 477         int i = pmd_index(paddr);
 478
 479         for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
 480                 pmd_t *pmd = pmd_page + pmd_index(paddr);
 481                 pte_t *pte;
 482                 pgprot_t new_prot = prot;
 483
 484                 paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
 485                 if (paddr >= paddr_end) {
 486                         if (!after_bootmem &&
 487                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 488                                              E820_TYPE_RAM) &&
 489                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 490                                              E820_TYPE_RESERVED_KERN))
 491                                 set_pmd(pmd, __pmd(0));
 492                         continue;
 493                 }
 494
 495                 if (!pmd_none(*pmd)) {
 496                         if (!pmd_large(*pmd)) {
 497                                 spin_lock(&init_mm.page_table_lock);
 498                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
 499                                 paddr_last = phys_pte_init(pte, paddr,
 500                                                            paddr_end, prot);
 501                                 spin_unlock(&init_mm.page_table_lock);
 502                                 continue;
 503                         }
 504                         /*
 505                          * If we are ok with PG_LEVEL_2M mapping, then we will
 506                          * use the existing mapping,
 507                          *
 508                          * Otherwise, we will split the large page mapping but
 509                          * use the same existing protection bits except for
 510                          * large page, so that we don't violate Intel's TLB
 511                          * Application note (317080) which says, while changing
 512                          * the page sizes, new and old translations should
 513                          * not differ with respect to page frame and
 514                          * attributes.
 515                          */
 516                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 517                                 if (!after_bootmem)
 518                                         pages++;
 519                                 paddr_last = paddr_next;
 520                                 continue;
 521                         }
 522                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 523                 }
 524
 525                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 526                         pages++;
 527                         spin_lock(&init_mm.page_table_lock);
 528                         set_pte((pte_t *)pmd,
 529                                 pfn_pte((paddr & PMD_MASK) >> PAGE_SHIFT,
 530                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 531                         spin_unlock(&init_mm.page_table_lock);
 532                         paddr_last = paddr_next;
 533                         continue;
 534                 }
 535
 536                 pte = alloc_low_page();
 537                 paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot);
 538
 539                 spin_lock(&init_mm.page_table_lock);
 540                 pmd_populate_kernel(&init_mm, pmd, pte);
 541                 spin_unlock(&init_mm.page_table_lock);
 542         }
 543         update_page_count(PG_LEVEL_2M, pages);
 544         return paddr_last;
 545 }
 546
 547 /*
 548  * Create PUD level page table mapping for physical addresses. The virtual
 549  * and physical address do not have to be aligned at this level. KASLR can
 550  * randomize virtual addresses up to this level.
 551  * It returns the last physical address mapped.
 552  */
 553 static unsigned long __meminit
 554 phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
 555               unsigned long page_size_mask)
 556 {
 557         unsigned long pages = 0, paddr_next;
 558         unsigned long paddr_last = paddr_end;
 559         unsigned long vaddr = (unsigned long)__va(paddr);
 560         int i = pud_index(vaddr);
 561
 562         for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
 563                 pud_t *pud;
 564                 pmd_t *pmd;
 565                 pgprot_t prot = PAGE_KERNEL;
 566
 567                 vaddr = (unsigned long)__va(paddr);
 568                 pud = pud_page + pud_index(vaddr);
 569                 paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
 570
 571                 if (paddr >= paddr_end) {
 572                         if (!after_bootmem &&
 573                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 574                                              E820_TYPE_RAM) &&
 575                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 576                                              E820_TYPE_RESERVED_KERN))
 577                                 set_pud(pud, __pud(0));
 578                         continue;
 579                 }
 580
 581                 if (!pud_none(*pud)) {
 582                         if (!pud_large(*pud)) {
 583                                 pmd = pmd_offset(pud, 0);
 584                                 paddr_last = phys_pmd_init(pmd, paddr,
 585                                                            paddr_end,
 586                                                            page_size_mask,
 587                                                            prot);
 588                                 __flush_tlb_all();
 589                                 continue;
 590                         }
 591                         /*
 592                          * If we are ok with PG_LEVEL_1G mapping, then we will
 593                          * use the existing mapping.
 594                          *
 595                          * Otherwise, we will split the gbpage mapping but use
 596                          * the same existing protection  bits except for large
 597                          * page, so that we don't violate Intel's TLB
 598                          * Application note (317080) which says, while changing
 599                          * the page sizes, new and old translations should
 600                          * not differ with respect to page frame and
 601                          * attributes.
 602                          */
 603                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 604                                 if (!after_bootmem)
 605                                         pages++;
 606                                 paddr_last = paddr_next;
 607                                 continue;
 608                         }
 609                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 610                 }
 611
 612                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 613                         pages++;
 614                         spin_lock(&init_mm.page_table_lock);
 615                         set_pte((pte_t *)pud,
 616                                 pfn_pte((paddr & PUD_MASK) >> PAGE_SHIFT,
 617                                         PAGE_KERNEL_LARGE));
 618                         spin_unlock(&init_mm.page_table_lock);
 619                         paddr_last = paddr_next;
 620                         continue;
 621                 }
 622
 623                 pmd = alloc_low_page();
 624                 paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
 625                                            page_size_mask, prot);
 626
 627                 spin_lock(&init_mm.page_table_lock);
 628                 pud_populate(&init_mm, pud, pmd);
 629                 spin_unlock(&init_mm.page_table_lock);
 630         }
 631         __flush_tlb_all();
 632
 633         update_page_count(PG_LEVEL_1G, pages);
 634
 635         return paddr_last;
 636 }
 637
 638 static unsigned long __meminit
 639 phys_p4d_init(p4d_t *p4d_page, unsigned long paddr, unsigned long paddr_end,
 640               unsigned long page_size_mask)
 641 {
 642         unsigned long paddr_next, paddr_last = paddr_end;
 643         unsigned long vaddr = (unsigned long)__va(paddr);
 644         int i = p4d_index(vaddr);
 645
 646         if (!pgtable_l5_enabled)
 647                 return phys_pud_init((pud_t *) p4d_page, paddr, paddr_end, page_size_mask);
 648
 649         for (; i < PTRS_PER_P4D; i++, paddr = paddr_next) {
 650                 p4d_t *p4d;
 651                 pud_t *pud;
 652
 653                 vaddr = (unsigned long)__va(paddr);
 654                 p4d = p4d_page + p4d_index(vaddr);
 655                 paddr_next = (paddr & P4D_MASK) + P4D_SIZE;
 656
 657                 if (paddr >= paddr_end) {
 658                         if (!after_bootmem &&
 659                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 660                                              E820_TYPE_RAM) &&
 661                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 662                                              E820_TYPE_RESERVED_KERN))
 663                                 set_p4d(p4d, __p4d(0));
 664                         continue;
 665                 }
 666
 667                 if (!p4d_none(*p4d)) {
 668                         pud = pud_offset(p4d, 0);
 669                         paddr_last = phys_pud_init(pud, paddr,
 670                                         paddr_end,
 671                                         page_size_mask);
 672                         __flush_tlb_all();
 673                         continue;
 674                 }
 675
 676                 pud = alloc_low_page();
 677                 paddr_last = phys_pud_init(pud, paddr, paddr_end,
 678                                            page_size_mask);
 679
 680                 spin_lock(&init_mm.page_table_lock);
 681                 p4d_populate(&init_mm, p4d, pud);
 682                 spin_unlock(&init_mm.page_table_lock);
 683         }
 684         __flush_tlb_all();
 685
 686         return paddr_last;
 687 }
 688
 689 /*
 690  * Create page table mapping for the physical memory for specific physical
 691  * addresses. The virtual and physical addresses have to be aligned on PMD level
 692  * down. It returns the last physical address mapped.
 693  */
 694 unsigned long __meminit
 695 kernel_physical_mapping_init(unsigned long paddr_start,
 696                              unsigned long paddr_end,
 697                              unsigned long page_size_mask)
 698 {
 699         bool pgd_changed = false;
 700         unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
 701
 702         paddr_last = paddr_end;
 703         vaddr = (unsigned long)__va(paddr_start);
 704         vaddr_end = (unsigned long)__va(paddr_end);
 705         vaddr_start = vaddr;
 706
 707         for (; vaddr < vaddr_end; vaddr = vaddr_next) {
 708                 pgd_t *pgd = pgd_offset_k(vaddr);
 709                 p4d_t *p4d;
 710
 711                 vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
 712
 713                 if (pgd_val(*pgd)) {
 714                         p4d = (p4d_t *)pgd_page_vaddr(*pgd);
 715                         paddr_last = phys_p4d_init(p4d, __pa(vaddr),
 716                                                    __pa(vaddr_end),
 717                                                    page_size_mask);
 718                         continue;
 719                 }
 720
 721                 p4d = alloc_low_page();
 722                 paddr_last = phys_p4d_init(p4d, __pa(vaddr), __pa(vaddr_end),
 723                                            page_size_mask);
 724
 725                 spin_lock(&init_mm.page_table_lock);
 726                 if (pgtable_l5_enabled)
 727                         pgd_populate(&init_mm, pgd, p4d);
 728                 else
 729                         p4d_populate(&init_mm, p4d_offset(pgd, vaddr), (pud_t *) p4d);
 730                 spin_unlock(&init_mm.page_table_lock);
 731                 pgd_changed = true;
 732         }
 733
 734         if (pgd_changed)
 735                 sync_global_pgds(vaddr_start, vaddr_end - 1);
 736
 737         __flush_tlb_all();
 738
 739         return paddr_last;
 740 }
 741
 742 #ifndef CONFIG_NUMA
 743 void __init initmem_init(void)
 744 {
 745         memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
 746 }
 747 #endif
 748
 749 void __init paging_init(void)
 750 {
 751         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 752         sparse_init();
 753
 754         /*
 755          * clear the default setting with node 0
 756          * note: don't use nodes_clear here, that is really clearing when
 757          *       numa support is not compiled in, and later node_set_state
 758          *       will not set it back.
 759          */
 760         node_clear_state(0, N_MEMORY);
 761         if (N_MEMORY != N_NORMAL_MEMORY)
 762                 node_clear_state(0, N_NORMAL_MEMORY);
 763
 764         zone_sizes_init();
 765 }
 766
 767 /*
 768  * Memory hotplug specific functions
 769  */
 770 #ifdef CONFIG_MEMORY_HOTPLUG
 771 /*
 772  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 773  * updating.
 774  */
 775 static void update_end_of_memory_vars(u64 start, u64 size)
 776 {
 777         unsigned long end_pfn = PFN_UP(start + size);
 778
 779         if (end_pfn > max_pfn) {
 780                 max_pfn = end_pfn;
 781                 max_low_pfn = end_pfn;
 782                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 783         }
 784 }
 785
 786 int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
 787                 struct vmem_altmap *altmap, bool want_memblock)
 788 {
 789         int ret;
 790
 791         ret = __add_pages(nid, start_pfn, nr_pages, altmap, want_memblock);
 792         WARN_ON_ONCE(ret);
 793
 794         /* update max_pfn, max_low_pfn and high_memory */
 795         update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
 796                                   nr_pages << PAGE_SHIFT);
 797
 798         return ret;
 799 }
 800
 801 int arch_add_memory(int nid, u64 start, u64 size, struct vmem_altmap *altmap,
 802                 bool want_memblock)
 803 {
 804         unsigned long start_pfn = start >> PAGE_SHIFT;
 805         unsigned long nr_pages = size >> PAGE_SHIFT;
 806
 807         init_memory_mapping(start, start + size);
 808
 809         return add_pages(nid, start_pfn, nr_pages, altmap, want_memblock);
 810 }
 811
 812 #define PAGE_INUSE 0xFD
 813
 814 static void __meminit free_pagetable(struct page *page, int order)
 815 {
 816         unsigned long magic;
 817         unsigned int nr_pages = 1 << order;
 818
 819         /* bootmem page has reserved flag */
 820         if (PageReserved(page)) {
 821                 __ClearPageReserved(page);
 822
 823                 magic = (unsigned long)page->freelist;
 824                 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
 825                         while (nr_pages--)
 826                                 put_page_bootmem(page++);
 827                 } else
 828                         while (nr_pages--)
 829                                 free_reserved_page(page++);
 830         } else
 831                 free_pages((unsigned long)page_address(page), order);
 832 }
 833
 834 static void __meminit free_hugepage_table(struct page *page,
 835                 struct vmem_altmap *altmap)
 836 {
 837         if (altmap)
 838                 vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
 839         else
 840                 free_pagetable(page, get_order(PMD_SIZE));
 841 }
 842
 843 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
 844 {
 845         pte_t *pte;
 846         int i;
 847
 848         for (i = 0; i < PTRS_PER_PTE; i++) {
 849                 pte = pte_start + i;
 850                 if (!pte_none(*pte))
 851                         return;
 852         }
 853
 854         /* free a pte talbe */
 855         free_pagetable(pmd_page(*pmd), 0);
 856         spin_lock(&init_mm.page_table_lock);
 857         pmd_clear(pmd);
 858         spin_unlock(&init_mm.page_table_lock);
 859 }
 860
 861 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
 862 {
 863         pmd_t *pmd;
 864         int i;
 865
 866         for (i = 0; i < PTRS_PER_PMD; i++) {
 867                 pmd = pmd_start + i;
 868                 if (!pmd_none(*pmd))
 869                         return;
 870         }
 871
 872         /* free a pmd talbe */
 873         free_pagetable(pud_page(*pud), 0);
 874         spin_lock(&init_mm.page_table_lock);
 875         pud_clear(pud);
 876         spin_unlock(&init_mm.page_table_lock);
 877 }
 878
 879 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
 880 {
 881         pud_t *pud;
 882         int i;
 883
 884         for (i = 0; i < PTRS_PER_PUD; i++) {
 885                 pud = pud_start + i;
 886                 if (!pud_none(*pud))
 887                         return;
 888         }
 889
 890         /* free a pud talbe */
 891         free_pagetable(p4d_page(*p4d), 0);
 892         spin_lock(&init_mm.page_table_lock);
 893         p4d_clear(p4d);
 894         spin_unlock(&init_mm.page_table_lock);
 895 }
 896
 897 static void __meminit
 898 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
 899                  bool direct)
 900 {
 901         unsigned long next, pages = 0;
 902         pte_t *pte;
 903         void *page_addr;
 904         phys_addr_t phys_addr;
 905
 906         pte = pte_start + pte_index(addr);
 907         for (; addr < end; addr = next, pte++) {
 908                 next = (addr + PAGE_SIZE) & PAGE_MASK;
 909                 if (next > end)
 910                         next = end;
 911
 912                 if (!pte_present(*pte))
 913                         continue;
 914
 915                 /*
 916                  * We mapped [0,1G) memory as identity mapping when
 917                  * initializing, in arch/x86/kernel/head_64.S. These
 918                  * pagetables cannot be removed.
 919                  */
 920                 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
 921                 if (phys_addr < (phys_addr_t)0x40000000)
 922                         return;
 923
 924                 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
 925                         /*
 926                          * Do not free direct mapping pages since they were
 927                          * freed when offlining, or simplely not in use.
 928                          */
 929                         if (!direct)
 930                                 free_pagetable(pte_page(*pte), 0);
 931
 932                         spin_lock(&init_mm.page_table_lock);
 933                         pte_clear(&init_mm, addr, pte);
 934                         spin_unlock(&init_mm.page_table_lock);
 935
 936                         /* For non-direct mapping, pages means nothing. */
 937                         pages++;
 938                 } else {
 939                         /*
 940                          * If we are here, we are freeing vmemmap pages since
 941                          * direct mapped memory ranges to be freed are aligned.
 942                          *
 943                          * If we are not removing the whole page, it means
 944                          * other page structs in this page are being used and
 945                          * we canot remove them. So fill the unused page_structs
 946                          * with 0xFD, and remove the page when it is wholly
 947                          * filled with 0xFD.
 948                          */
 949                         memset((void *)addr, PAGE_INUSE, next - addr);
 950
 951                         page_addr = page_address(pte_page(*pte));
 952                         if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
 953                                 free_pagetable(pte_page(*pte), 0);
 954
 955                                 spin_lock(&init_mm.page_table_lock);
 956                                 pte_clear(&init_mm, addr, pte);
 957                                 spin_unlock(&init_mm.page_table_lock);
 958                         }
 959                 }
 960         }
 961
 962         /* Call free_pte_table() in remove_pmd_table(). */
 963         flush_tlb_all();
 964         if (direct)
 965                 update_page_count(PG_LEVEL_4K, -pages);
 966 }
 967
 968 static void __meminit
 969 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
 970                  bool direct, struct vmem_altmap *altmap)
 971 {
 972         unsigned long next, pages = 0;
 973         pte_t *pte_base;
 974         pmd_t *pmd;
 975         void *page_addr;
 976
 977         pmd = pmd_start + pmd_index(addr);
 978         for (; addr < end; addr = next, pmd++) {
 979                 next = pmd_addr_end(addr, end);
 980
 981                 if (!pmd_present(*pmd))
 982                         continue;
 983
 984                 if (pmd_large(*pmd)) {
 985                         if (IS_ALIGNED(addr, PMD_SIZE) &&
 986                             IS_ALIGNED(next, PMD_SIZE)) {
 987                                 if (!direct)
 988                                         free_hugepage_table(pmd_page(*pmd),
 989                                                             altmap);
 990
 991                                 spin_lock(&init_mm.page_table_lock);
 992                                 pmd_clear(pmd);
 993                                 spin_unlock(&init_mm.page_table_lock);
 994                                 pages++;
 995                         } else {
 996                                 /* If here, we are freeing vmemmap pages. */
 997                                 memset((void *)addr, PAGE_INUSE, next - addr);
 998
 999                                 page_addr = page_address(pmd_page(*pmd));
1000                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1001                                                 PMD_SIZE)) {
1002                                         free_hugepage_table(pmd_page(*pmd),
1003                                                             altmap);
1004
1005                                         spin_lock(&init_mm.page_table_lock);
1006                                         pmd_clear(pmd);
1007                                         spin_unlock(&init_mm.page_table_lock);
1008                                 }
1009                         }
1010
1011                         continue;
1012                 }
1013
1014                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1015                 remove_pte_table(pte_base, addr, next, direct);
1016                 free_pte_table(pte_base, pmd);
1017         }
1018
1019         /* Call free_pmd_table() in remove_pud_table(). */
1020         if (direct)
1021                 update_page_count(PG_LEVEL_2M, -pages);
1022 }
1023
1024 static void __meminit
1025 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1026                  struct vmem_altmap *altmap, bool direct)
1027 {
1028         unsigned long next, pages = 0;
1029         pmd_t *pmd_base;
1030         pud_t *pud;
1031         void *page_addr;
1032
1033         pud = pud_start + pud_index(addr);
1034         for (; addr < end; addr = next, pud++) {
1035                 next = pud_addr_end(addr, end);
1036
1037                 if (!pud_present(*pud))
1038                         continue;
1039
1040                 if (pud_large(*pud)) {
1041                         if (IS_ALIGNED(addr, PUD_SIZE) &&
1042                             IS_ALIGNED(next, PUD_SIZE)) {
1043                                 if (!direct)
1044                                         free_pagetable(pud_page(*pud),
1045                                                        get_order(PUD_SIZE));
1046
1047                                 spin_lock(&init_mm.page_table_lock);
1048                                 pud_clear(pud);
1049                                 spin_unlock(&init_mm.page_table_lock);
1050                                 pages++;
1051                         } else {
1052                                 /* If here, we are freeing vmemmap pages. */
1053                                 memset((void *)addr, PAGE_INUSE, next - addr);
1054
1055                                 page_addr = page_address(pud_page(*pud));
1056                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1057                                                 PUD_SIZE)) {
1058                                         free_pagetable(pud_page(*pud),
1059                                                        get_order(PUD_SIZE));
1060
1061                                         spin_lock(&init_mm.page_table_lock);
1062                                         pud_clear(pud);
1063                                         spin_unlock(&init_mm.page_table_lock);
1064                                 }
1065                         }
1066
1067                         continue;
1068                 }
1069
1070                 pmd_base = pmd_offset(pud, 0);
1071                 remove_pmd_table(pmd_base, addr, next, direct, altmap);
1072                 free_pmd_table(pmd_base, pud);
1073         }
1074
1075         if (direct)
1076                 update_page_count(PG_LEVEL_1G, -pages);
1077 }
1078
1079 static void __meminit
1080 remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1081                  struct vmem_altmap *altmap, bool direct)
1082 {
1083         unsigned long next, pages = 0;
1084         pud_t *pud_base;
1085         p4d_t *p4d;
1086
1087         p4d = p4d_start + p4d_index(addr);
1088         for (; addr < end; addr = next, p4d++) {
1089                 next = p4d_addr_end(addr, end);
1090
1091                 if (!p4d_present(*p4d))
1092                         continue;
1093
1094                 BUILD_BUG_ON(p4d_large(*p4d));
1095
1096                 pud_base = pud_offset(p4d, 0);
1097                 remove_pud_table(pud_base, addr, next, altmap, direct);
1098                 /*
1099                  * For 4-level page tables we do not want to free PUDs, but in the
1100                  * 5-level case we should free them. This code will have to change
1101                  * to adapt for boot-time switching between 4 and 5 level page tables.
1102                  */
1103                 if (pgtable_l5_enabled)
1104                         free_pud_table(pud_base, p4d);
1105         }
1106
1107         if (direct)
1108                 update_page_count(PG_LEVEL_512G, -pages);
1109 }
1110
1111 /* start and end are both virtual address. */
1112 static void __meminit
1113 remove_pagetable(unsigned long start, unsigned long end, bool direct,
1114                 struct vmem_altmap *altmap)
1115 {
1116         unsigned long next;
1117         unsigned long addr;
1118         pgd_t *pgd;
1119         p4d_t *p4d;
1120
1121         for (addr = start; addr < end; addr = next) {
1122                 next = pgd_addr_end(addr, end);
1123
1124                 pgd = pgd_offset_k(addr);
1125                 if (!pgd_present(*pgd))
1126                         continue;
1127
1128                 p4d = p4d_offset(pgd, 0);
1129                 remove_p4d_table(p4d, addr, next, altmap, direct);
1130         }
1131
1132         flush_tlb_all();
1133 }
1134
1135 void __ref vmemmap_free(unsigned long start, unsigned long end,
1136                 struct vmem_altmap *altmap)
1137 {
1138         remove_pagetable(start, end, false, altmap);
1139 }
1140
1141 #ifdef CONFIG_MEMORY_HOTREMOVE
1142 static void __meminit
1143 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1144 {
1145         start = (unsigned long)__va(start);
1146         end = (unsigned long)__va(end);
1147
1148         remove_pagetable(start, end, true, NULL);
1149 }
1150
1151 int __ref arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
1152 {
1153         unsigned long start_pfn = start >> PAGE_SHIFT;
1154         unsigned long nr_pages = size >> PAGE_SHIFT;
1155         struct page *page = pfn_to_page(start_pfn);
1156         struct zone *zone;
1157         int ret;
1158
1159         /* With altmap the first mapped page is offset from @start */
1160         if (altmap)
1161                 page += vmem_altmap_offset(altmap);
1162         zone = page_zone(page);
1163         ret = __remove_pages(zone, start_pfn, nr_pages, altmap);
1164         WARN_ON_ONCE(ret);
1165         kernel_physical_mapping_remove(start, start + size);
1166
1167         return ret;
1168 }
1169 #endif
1170 #endif /* CONFIG_MEMORY_HOTPLUG */
1171
1172 static struct kcore_list kcore_vsyscall;
1173
1174 static void __init register_page_bootmem_info(void)
1175 {
1176 #ifdef CONFIG_NUMA
1177         int i;
1178
1179         for_each_online_node(i)
1180                 register_page_bootmem_info_node(NODE_DATA(i));
1181 #endif
1182 }
1183
1184 void __init mem_init(void)
1185 {
1186         pci_iommu_alloc();
1187
1188         /* clear_bss() already clear the empty_zero_page */
1189
1190         /* this will put all memory onto the freelists */
1191         free_all_bootmem();
1192         after_bootmem = 1;
1193         x86_init.hyper.init_after_bootmem();
1194
1195         /*
1196          * Must be done after boot memory is put on freelist, because here we
1197          * might set fields in deferred struct pages that have not yet been
1198          * initialized, and free_all_bootmem() initializes all the reserved
1199          * deferred pages for us.
1200          */
1201         register_page_bootmem_info();
1202
1203         /* Register memory areas for /proc/kcore */
1204         if (get_gate_vma(&init_mm))
1205                 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1206
1207         mem_init_print_info(NULL);
1208 }
1209
1210 int kernel_set_to_readonly;
1211
1212 void set_kernel_text_rw(void)
1213 {
1214         unsigned long start = PFN_ALIGN(_text);
1215         unsigned long end = PFN_ALIGN(__stop___ex_table);
1216
1217         if (!kernel_set_to_readonly)
1218                 return;
1219
1220         pr_debug("Set kernel text: %lx - %lx for read write\n",
1221                  start, end);
1222
1223         /*
1224          * Make the kernel identity mapping for text RW. Kernel text
1225          * mapping will always be RO. Refer to the comment in
1226          * static_protections() in pageattr.c
1227          */
1228         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1229 }
1230
1231 void set_kernel_text_ro(void)
1232 {
1233         unsigned long start = PFN_ALIGN(_text);
1234         unsigned long end = PFN_ALIGN(__stop___ex_table);
1235
1236         if (!kernel_set_to_readonly)
1237                 return;
1238
1239         pr_debug("Set kernel text: %lx - %lx for read only\n",
1240                  start, end);
1241
1242         /*
1243          * Set the kernel identity mapping for text RO.
1244          */
1245         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1246 }
1247
1248 void mark_rodata_ro(void)
1249 {
1250         unsigned long start = PFN_ALIGN(_text);
1251         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1252         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1253         unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1254         unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1255         unsigned long all_end;
1256
1257         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1258                (end - start) >> 10);
1259         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1260
1261         kernel_set_to_readonly = 1;
1262
1263         /*
1264          * The rodata/data/bss/brk section (but not the kernel text!)
1265          * should also be not-executable.
1266          *
1267          * We align all_end to PMD_SIZE because the existing mapping
1268          * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1269          * split the PMD and the reminder between _brk_end and the end
1270          * of the PMD will remain mapped executable.
1271          *
1272          * Any PMD which was setup after the one which covers _brk_end
1273          * has been zapped already via cleanup_highmem().
1274          */
1275         all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1276         set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1277
1278 #ifdef CONFIG_CPA_DEBUG
1279         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1280         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1281
1282         printk(KERN_INFO "Testing CPA: again\n");
1283         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1284 #endif
1285
1286         free_init_pages("unused kernel",
1287                         (unsigned long) __va(__pa_symbol(text_end)),
1288                         (unsigned long) __va(__pa_symbol(rodata_start)));
1289         free_init_pages("unused kernel",
1290                         (unsigned long) __va(__pa_symbol(rodata_end)),
1291                         (unsigned long) __va(__pa_symbol(_sdata)));
1292
1293         debug_checkwx();
1294
1295         /*
1296          * Do this after all of the manipulation of the
1297          * kernel text page tables are complete.
1298          */
1299         pti_clone_kernel_text();
1300 }
1301
1302 int kern_addr_valid(unsigned long addr)
1303 {
1304         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1305         pgd_t *pgd;
1306         p4d_t *p4d;
1307         pud_t *pud;
1308         pmd_t *pmd;
1309         pte_t *pte;
1310
1311         if (above != 0 && above != -1UL)
1312                 return 0;
1313
1314         pgd = pgd_offset_k(addr);
1315         if (pgd_none(*pgd))
1316                 return 0;
1317
1318         p4d = p4d_offset(pgd, addr);
1319         if (p4d_none(*p4d))
1320                 return 0;
1321
1322         pud = pud_offset(p4d, addr);
1323         if (pud_none(*pud))
1324                 return 0;
1325
1326         if (pud_large(*pud))
1327                 return pfn_valid(pud_pfn(*pud));
1328
1329         pmd = pmd_offset(pud, addr);
1330         if (pmd_none(*pmd))
1331                 return 0;
1332
1333         if (pmd_large(*pmd))
1334                 return pfn_valid(pmd_pfn(*pmd));
1335
1336         pte = pte_offset_kernel(pmd, addr);
1337         if (pte_none(*pte))
1338                 return 0;
1339
1340         return pfn_valid(pte_pfn(*pte));
1341 }
1342
1343 /*
1344  * Block size is the minimum amount of memory which can be hotplugged or
1345  * hotremoved. It must be power of two and must be equal or larger than
1346  * MIN_MEMORY_BLOCK_SIZE.
1347  */
1348 #define MAX_BLOCK_SIZE (2UL << 30)
1349
1350 /* Amount of ram needed to start using large blocks */
1351 #define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1352
1353 static unsigned long probe_memory_block_size(void)
1354 {
1355         unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1356         unsigned long bz;
1357
1358         /* If this is UV system, always set 2G block size */
1359         if (is_uv_system()) {
1360                 bz = MAX_BLOCK_SIZE;
1361                 goto done;
1362         }
1363
1364         /* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1365         if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1366                 bz = MIN_MEMORY_BLOCK_SIZE;
1367                 goto done;
1368         }
1369
1370         /* Find the largest allowed block size that aligns to memory end */
1371         for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1372                 if (IS_ALIGNED(boot_mem_end, bz))
1373                         break;
1374         }
1375 done:
1376         pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1377
1378         return bz;
1379 }
1380
1381 static unsigned long memory_block_size_probed;
1382 unsigned long memory_block_size_bytes(void)
1383 {
1384         if (!memory_block_size_probed)
1385                 memory_block_size_probed = probe_memory_block_size();
1386
1387         return memory_block_size_probed;
1388 }
1389
1390 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1391 /*
1392  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1393  */
1394 static long __meminitdata addr_start, addr_end;
1395 static void __meminitdata *p_start, *p_end;
1396 static int __meminitdata node_start;
1397
1398 static int __meminit vmemmap_populate_hugepages(unsigned long start,
1399                 unsigned long end, int node, struct vmem_altmap *altmap)
1400 {
1401         unsigned long addr;
1402         unsigned long next;
1403         pgd_t *pgd;
1404         p4d_t *p4d;
1405         pud_t *pud;
1406         pmd_t *pmd;
1407
1408         for (addr = start; addr < end; addr = next) {
1409                 next = pmd_addr_end(addr, end);
1410
1411                 pgd = vmemmap_pgd_populate(addr, node);
1412                 if (!pgd)
1413                         return -ENOMEM;
1414
1415                 p4d = vmemmap_p4d_populate(pgd, addr, node);
1416                 if (!p4d)
1417                         return -ENOMEM;
1418
1419                 pud = vmemmap_pud_populate(p4d, addr, node);
1420                 if (!pud)
1421                         return -ENOMEM;
1422
1423                 pmd = pmd_offset(pud, addr);
1424                 if (pmd_none(*pmd)) {
1425                         void *p;
1426
1427                         if (altmap)
1428                                 p = altmap_alloc_block_buf(PMD_SIZE, altmap);
1429                         else
1430                                 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1431                         if (p) {
1432                                 pte_t entry;
1433
1434                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1435                                                 PAGE_KERNEL_LARGE);
1436                                 set_pmd(pmd, __pmd(pte_val(entry)));
1437
1438                                 /* check to see if we have contiguous blocks */
1439                                 if (p_end != p || node_start != node) {
1440                                         if (p_start)
1441                                                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1442                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1443                                         addr_start = addr;
1444                                         node_start = node;
1445                                         p_start = p;
1446                                 }
1447
1448                                 addr_end = addr + PMD_SIZE;
1449                                 p_end = p + PMD_SIZE;
1450                                 continue;
1451                         } else if (altmap)
1452                                 return -ENOMEM; /* no fallback */
1453                 } else if (pmd_large(*pmd)) {
1454                         vmemmap_verify((pte_t *)pmd, node, addr, next);
1455                         continue;
1456                 }
1457                 if (vmemmap_populate_basepages(addr, next, node))
1458                         return -ENOMEM;
1459         }
1460         return 0;
1461 }
1462
1463 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1464                 struct vmem_altmap *altmap)
1465 {
1466         int err;
1467
1468         if (boot_cpu_has(X86_FEATURE_PSE))
1469                 err = vmemmap_populate_hugepages(start, end, node, altmap);
1470         else if (altmap) {
1471                 pr_err_once("%s: no cpu support for altmap allocations\n",
1472                                 __func__);
1473                 err = -ENOMEM;
1474         } else
1475                 err = vmemmap_populate_basepages(start, end, node);
1476         if (!err)
1477                 sync_global_pgds(start, end - 1);
1478         return err;
1479 }
1480
1481 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1482 void register_page_bootmem_memmap(unsigned long section_nr,
1483                                   struct page *start_page, unsigned long nr_pages)
1484 {
1485         unsigned long addr = (unsigned long)start_page;
1486         unsigned long end = (unsigned long)(start_page + nr_pages);
1487         unsigned long next;
1488         pgd_t *pgd;
1489         p4d_t *p4d;
1490         pud_t *pud;
1491         pmd_t *pmd;
1492         unsigned int nr_pmd_pages;
1493         struct page *page;
1494
1495         for (; addr < end; addr = next) {
1496                 pte_t *pte = NULL;
1497
1498                 pgd = pgd_offset_k(addr);
1499                 if (pgd_none(*pgd)) {
1500                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1501                         continue;
1502                 }
1503                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1504
1505                 p4d = p4d_offset(pgd, addr);
1506                 if (p4d_none(*p4d)) {
1507                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1508                         continue;
1509                 }
1510                 get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1511
1512                 pud = pud_offset(p4d, addr);
1513                 if (pud_none(*pud)) {
1514                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1515                         continue;
1516                 }
1517                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1518
1519                 if (!boot_cpu_has(X86_FEATURE_PSE)) {
1520                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1521                         pmd = pmd_offset(pud, addr);
1522                         if (pmd_none(*pmd))
1523                                 continue;
1524                         get_page_bootmem(section_nr, pmd_page(*pmd),
1525                                          MIX_SECTION_INFO);
1526
1527                         pte = pte_offset_kernel(pmd, addr);
1528                         if (pte_none(*pte))
1529                                 continue;
1530                         get_page_bootmem(section_nr, pte_page(*pte),
1531                                          SECTION_INFO);
1532                 } else {
1533                         next = pmd_addr_end(addr, end);
1534
1535                         pmd = pmd_offset(pud, addr);
1536                         if (pmd_none(*pmd))
1537                                 continue;
1538
1539                         nr_pmd_pages = 1 << get_order(PMD_SIZE);
1540                         page = pmd_page(*pmd);
1541                         while (nr_pmd_pages--)
1542                                 get_page_bootmem(section_nr, page++,
1543                                                  SECTION_INFO);
1544                 }
1545         }
1546 }
1547 #endif
1548
1549 void __meminit vmemmap_populate_print_last(void)
1550 {
1551         if (p_start) {
1552                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1553                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1554                 p_start = NULL;
1555                 p_end = NULL;
1556                 node_start = 0;
1557         }
1558 }
1559 #endif