arch/x86/mm/init_64.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  linux/arch/x86_64/mm/init.c
   4  *
   5  *  Copyright (C) 1995  Linus Torvalds
   6  *  Copyright (C) 2000  Pavel Machek <[email protected]>
   7  *  Copyright (C) 2002,2003 Andi Kleen <[email protected]>
   8  */
   9
  10 #include <linux/signal.h>
  11 #include <linux/sched.h>
  12 #include <linux/kernel.h>
  13 #include <linux/errno.h>
  14 #include <linux/string.h>
  15 #include <linux/types.h>
  16 #include <linux/ptrace.h>
  17 #include <linux/mman.h>
  18 #include <linux/mm.h>
  19 #include <linux/swap.h>
  20 #include <linux/smp.h>
  21 #include <linux/init.h>
  22 #include <linux/initrd.h>
  23 #include <linux/pagemap.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 #include <linux/bootmem_info.h>
  37
  38 #include <asm/processor.h>
  39 #include <asm/bios_ebda.h>
  40 #include <linux/uaccess.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 #include <asm/ftrace.h>
  58
  59 #include "mm_internal.h"
  60
  61 #include "ident_map.c"
  62
  63 #define DEFINE_POPULATE(fname, type1, type2, init)              \
  64 static inline void fname##_init(struct mm_struct *mm,           \
  65                 type1##_t *arg1, type2##_t *arg2, bool init)    \
  66 {                                                               \
  67         if (init)                                               \
  68                 fname##_safe(mm, arg1, arg2);                   \
  69         else                                                    \
  70                 fname(mm, arg1, arg2);                          \
  71 }
  72
  73 DEFINE_POPULATE(p4d_populate, p4d, pud, init)
  74 DEFINE_POPULATE(pgd_populate, pgd, p4d, init)
  75 DEFINE_POPULATE(pud_populate, pud, pmd, init)
  76 DEFINE_POPULATE(pmd_populate_kernel, pmd, pte, init)
  77
  78 #define DEFINE_ENTRY(type1, type2, init)                        \
  79 static inline void set_##type1##_init(type1##_t *arg1,          \
  80                         type2##_t arg2, bool init)              \
  81 {                                                               \
  82         if (init)                                               \
  83                 set_##type1##_safe(arg1, arg2);                 \
  84         else                                                    \
  85                 set_##type1(arg1, arg2);                        \
  86 }
  87
  88 DEFINE_ENTRY(p4d, p4d, init)
  89 DEFINE_ENTRY(pud, pud, init)
  90 DEFINE_ENTRY(pmd, pmd, init)
  91 DEFINE_ENTRY(pte, pte, init)
  92
  93 static inline pgprot_t prot_sethuge(pgprot_t prot)
  94 {
  95         WARN_ON_ONCE(pgprot_val(prot) & _PAGE_PAT);
  96
  97         return __pgprot(pgprot_val(prot) | _PAGE_PSE);
  98 }
  99
 100 /*
 101  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
 102  * physical space so we can cache the place of the first one and move
 103  * around without checking the pgd every time.
 104  */
 105
 106 /* Bits supported by the hardware: */
 107 pteval_t __supported_pte_mask __read_mostly = ~0;
 108 /* Bits allowed in normal kernel mappings: */
 109 pteval_t __default_kernel_pte_mask __read_mostly = ~0;
 110 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 111 /* Used in PAGE_KERNEL_* macros which are reasonably used out-of-tree: */
 112 EXPORT_SYMBOL(__default_kernel_pte_mask);
 113
 114 int force_personality32;
 115
 116 /*
 117  * noexec32=on|off
 118  * Control non executable heap for 32bit processes.
 119  *
 120  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
 121  * off  PROT_READ implies PROT_EXEC
 122  */
 123 static int __init nonx32_setup(char *str)
 124 {
 125         if (!strcmp(str, "on"))
 126                 force_personality32 &= ~READ_IMPLIES_EXEC;
 127         else if (!strcmp(str, "off"))
 128                 force_personality32 |= READ_IMPLIES_EXEC;
 129         return 1;
 130 }
 131 __setup("noexec32=", nonx32_setup);
 132
 133 static void sync_global_pgds_l5(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                 const pgd_t *pgd_ref = pgd_offset_k(addr);
 139                 struct page *page;
 140
 141                 /* Check for overflow */
 142                 if (addr < start)
 143                         break;
 144
 145                 if (pgd_none(*pgd_ref))
 146                         continue;
 147
 148                 spin_lock(&pgd_lock);
 149                 list_for_each_entry(page, &pgd_list, lru) {
 150                         pgd_t *pgd;
 151                         spinlock_t *pgt_lock;
 152
 153                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
 154                         /* the pgt_lock only for Xen */
 155                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 156                         spin_lock(pgt_lock);
 157
 158                         if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
 159                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 160
 161                         if (pgd_none(*pgd))
 162                                 set_pgd(pgd, *pgd_ref);
 163
 164                         spin_unlock(pgt_lock);
 165                 }
 166                 spin_unlock(&pgd_lock);
 167         }
 168 }
 169
 170 static void sync_global_pgds_l4(unsigned long start, unsigned long end)
 171 {
 172         unsigned long addr;
 173
 174         for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
 175                 pgd_t *pgd_ref = pgd_offset_k(addr);
 176                 const p4d_t *p4d_ref;
 177                 struct page *page;
 178
 179                 /*
 180                  * With folded p4d, pgd_none() is always false, we need to
 181                  * handle synchronization on p4d level.
 182                  */
 183                 MAYBE_BUILD_BUG_ON(pgd_none(*pgd_ref));
 184                 p4d_ref = p4d_offset(pgd_ref, addr);
 185
 186                 if (p4d_none(*p4d_ref))
 187                         continue;
 188
 189                 spin_lock(&pgd_lock);
 190                 list_for_each_entry(page, &pgd_list, lru) {
 191                         pgd_t *pgd;
 192                         p4d_t *p4d;
 193                         spinlock_t *pgt_lock;
 194
 195                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
 196                         p4d = p4d_offset(pgd, addr);
 197                         /* the pgt_lock only for Xen */
 198                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 199                         spin_lock(pgt_lock);
 200
 201                         if (!p4d_none(*p4d_ref) && !p4d_none(*p4d))
 202                                 BUG_ON(p4d_pgtable(*p4d)
 203                                        != p4d_pgtable(*p4d_ref));
 204
 205                         if (p4d_none(*p4d))
 206                                 set_p4d(p4d, *p4d_ref);
 207
 208                         spin_unlock(pgt_lock);
 209                 }
 210                 spin_unlock(&pgd_lock);
 211         }
 212 }
 213
 214 /*
 215  * When memory was added make sure all the processes MM have
 216  * suitable PGD entries in the local PGD level page.
 217  */
 218 static void sync_global_pgds(unsigned long start, unsigned long end)
 219 {
 220         if (pgtable_l5_enabled())
 221                 sync_global_pgds_l5(start, end);
 222         else
 223                 sync_global_pgds_l4(start, end);
 224 }
 225
 226 /*
 227  * NOTE: This function is marked __ref because it calls __init function
 228  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 229  */
 230 static __ref void *spp_getpage(void)
 231 {
 232         void *ptr;
 233
 234         if (after_bootmem)
 235                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 236         else
 237                 ptr = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
 238
 239         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 240                 panic("set_pte_phys: cannot allocate page data %s\n",
 241                         after_bootmem ? "after bootmem" : "");
 242         }
 243
 244         pr_debug("spp_getpage %p\n", ptr);
 245
 246         return ptr;
 247 }
 248
 249 static p4d_t *fill_p4d(pgd_t *pgd, unsigned long vaddr)
 250 {
 251         if (pgd_none(*pgd)) {
 252                 p4d_t *p4d = (p4d_t *)spp_getpage();
 253                 pgd_populate(&init_mm, pgd, p4d);
 254                 if (p4d != p4d_offset(pgd, 0))
 255                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 256                                p4d, p4d_offset(pgd, 0));
 257         }
 258         return p4d_offset(pgd, vaddr);
 259 }
 260
 261 static pud_t *fill_pud(p4d_t *p4d, unsigned long vaddr)
 262 {
 263         if (p4d_none(*p4d)) {
 264                 pud_t *pud = (pud_t *)spp_getpage();
 265                 p4d_populate(&init_mm, p4d, pud);
 266                 if (pud != pud_offset(p4d, 0))
 267                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 268                                pud, pud_offset(p4d, 0));
 269         }
 270         return pud_offset(p4d, vaddr);
 271 }
 272
 273 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 274 {
 275         if (pud_none(*pud)) {
 276                 pmd_t *pmd = (pmd_t *) spp_getpage();
 277                 pud_populate(&init_mm, pud, pmd);
 278                 if (pmd != pmd_offset(pud, 0))
 279                         printk(KERN_ERR "PAGETABLE BUG #02! %p <-> %p\n",
 280                                pmd, pmd_offset(pud, 0));
 281         }
 282         return pmd_offset(pud, vaddr);
 283 }
 284
 285 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 286 {
 287         if (pmd_none(*pmd)) {
 288                 pte_t *pte = (pte_t *) spp_getpage();
 289                 pmd_populate_kernel(&init_mm, pmd, pte);
 290                 if (pte != pte_offset_kernel(pmd, 0))
 291                         printk(KERN_ERR "PAGETABLE BUG #03!\n");
 292         }
 293         return pte_offset_kernel(pmd, vaddr);
 294 }
 295
 296 static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
 297 {
 298         pmd_t *pmd = fill_pmd(pud, vaddr);
 299         pte_t *pte = fill_pte(pmd, vaddr);
 300
 301         set_pte(pte, new_pte);
 302
 303         /*
 304          * It's enough to flush this one mapping.
 305          * (PGE mappings get flushed as well)
 306          */
 307         flush_tlb_one_kernel(vaddr);
 308 }
 309
 310 void set_pte_vaddr_p4d(p4d_t *p4d_page, unsigned long vaddr, pte_t new_pte)
 311 {
 312         p4d_t *p4d = p4d_page + p4d_index(vaddr);
 313         pud_t *pud = fill_pud(p4d, vaddr);
 314
 315         __set_pte_vaddr(pud, vaddr, new_pte);
 316 }
 317
 318 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 319 {
 320         pud_t *pud = pud_page + pud_index(vaddr);
 321
 322         __set_pte_vaddr(pud, vaddr, new_pte);
 323 }
 324
 325 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 326 {
 327         pgd_t *pgd;
 328         p4d_t *p4d_page;
 329
 330         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 331
 332         pgd = pgd_offset_k(vaddr);
 333         if (pgd_none(*pgd)) {
 334                 printk(KERN_ERR
 335                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 336                 return;
 337         }
 338
 339         p4d_page = p4d_offset(pgd, 0);
 340         set_pte_vaddr_p4d(p4d_page, vaddr, pteval);
 341 }
 342
 343 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 344 {
 345         pgd_t *pgd;
 346         p4d_t *p4d;
 347         pud_t *pud;
 348
 349         pgd = pgd_offset_k(vaddr);
 350         p4d = fill_p4d(pgd, vaddr);
 351         pud = fill_pud(p4d, vaddr);
 352         return fill_pmd(pud, vaddr);
 353 }
 354
 355 pte_t * __init populate_extra_pte(unsigned long vaddr)
 356 {
 357         pmd_t *pmd;
 358
 359         pmd = populate_extra_pmd(vaddr);
 360         return fill_pte(pmd, vaddr);
 361 }
 362
 363 /*
 364  * Create large page table mappings for a range of physical addresses.
 365  */
 366 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 367                                         enum page_cache_mode cache)
 368 {
 369         pgd_t *pgd;
 370         p4d_t *p4d;
 371         pud_t *pud;
 372         pmd_t *pmd;
 373         pgprot_t prot;
 374
 375         pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
 376                 protval_4k_2_large(cachemode2protval(cache));
 377         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 378         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 379                 pgd = pgd_offset_k((unsigned long)__va(phys));
 380                 if (pgd_none(*pgd)) {
 381                         p4d = (p4d_t *) spp_getpage();
 382                         set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE |
 383                                                 _PAGE_USER));
 384                 }
 385                 p4d = p4d_offset(pgd, (unsigned long)__va(phys));
 386                 if (p4d_none(*p4d)) {
 387                         pud = (pud_t *) spp_getpage();
 388                         set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE |
 389                                                 _PAGE_USER));
 390                 }
 391                 pud = pud_offset(p4d, (unsigned long)__va(phys));
 392                 if (pud_none(*pud)) {
 393                         pmd = (pmd_t *) spp_getpage();
 394                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 395                                                 _PAGE_USER));
 396                 }
 397                 pmd = pmd_offset(pud, phys);
 398                 BUG_ON(!pmd_none(*pmd));
 399                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 400         }
 401 }
 402
 403 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 404 {
 405         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
 406 }
 407
 408 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 409 {
 410         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
 411 }
 412
 413 /*
 414  * The head.S code sets up the kernel high mapping:
 415  *
 416  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 417  *
 418  * phys_base holds the negative offset to the kernel, which is added
 419  * to the compile time generated pmds. This results in invalid pmds up
 420  * to the point where we hit the physaddr 0 mapping.
 421  *
 422  * We limit the mappings to the region from _text to _brk_end.  _brk_end
 423  * is rounded up to the 2MB boundary. This catches the invalid pmds as
 424  * well, as they are located before _text:
 425  */
 426 void __init cleanup_highmap(void)
 427 {
 428         unsigned long vaddr = __START_KERNEL_map;
 429         unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
 430         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 431         pmd_t *pmd = level2_kernel_pgt;
 432
 433         /*
 434          * Native path, max_pfn_mapped is not set yet.
 435          * Xen has valid max_pfn_mapped set in
 436          *      arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
 437          */
 438         if (max_pfn_mapped)
 439                 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 440
 441         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
 442                 if (pmd_none(*pmd))
 443                         continue;
 444                 if (vaddr < (unsigned long) _text || vaddr > end)
 445                         set_pmd(pmd, __pmd(0));
 446         }
 447 }
 448
 449 /*
 450  * Create PTE level page table mapping for physical addresses.
 451  * It returns the last physical address mapped.
 452  */
 453 static unsigned long __meminit
 454 phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
 455               pgprot_t prot, bool init)
 456 {
 457         unsigned long pages = 0, paddr_next;
 458         unsigned long paddr_last = paddr_end;
 459         pte_t *pte;
 460         int i;
 461
 462         pte = pte_page + pte_index(paddr);
 463         i = pte_index(paddr);
 464
 465         for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
 466                 paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
 467                 if (paddr >= paddr_end) {
 468                         if (!after_bootmem &&
 469                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 470                                              E820_TYPE_RAM) &&
 471                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 472                                              E820_TYPE_RESERVED_KERN) &&
 473                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 474                                              E820_TYPE_ACPI))
 475                                 set_pte_init(pte, __pte(0), init);
 476                         continue;
 477                 }
 478
 479                 /*
 480                  * We will re-use the existing mapping.
 481                  * Xen for example has some special requirements, like mapping
 482                  * pagetable pages as RO. So assume someone who pre-setup
 483                  * these mappings are more intelligent.
 484                  */
 485                 if (!pte_none(*pte)) {
 486                         if (!after_bootmem)
 487                                 pages++;
 488                         continue;
 489                 }
 490
 491                 if (0)
 492                         pr_info("   pte=%p addr=%lx pte=%016lx\n", pte, paddr,
 493                                 pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 494                 pages++;
 495                 set_pte_init(pte, pfn_pte(paddr >> PAGE_SHIFT, prot), init);
 496                 paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
 497         }
 498
 499         update_page_count(PG_LEVEL_4K, pages);
 500
 501         return paddr_last;
 502 }
 503
 504 /*
 505  * Create PMD level page table mapping for physical addresses. The virtual
 506  * and physical address have to be aligned at this level.
 507  * It returns the last physical address mapped.
 508  */
 509 static unsigned long __meminit
 510 phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
 511               unsigned long page_size_mask, pgprot_t prot, bool init)
 512 {
 513         unsigned long pages = 0, paddr_next;
 514         unsigned long paddr_last = paddr_end;
 515
 516         int i = pmd_index(paddr);
 517
 518         for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
 519                 pmd_t *pmd = pmd_page + pmd_index(paddr);
 520                 pte_t *pte;
 521                 pgprot_t new_prot = prot;
 522
 523                 paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
 524                 if (paddr >= paddr_end) {
 525                         if (!after_bootmem &&
 526                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 527                                              E820_TYPE_RAM) &&
 528                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 529                                              E820_TYPE_RESERVED_KERN) &&
 530                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 531                                              E820_TYPE_ACPI))
 532                                 set_pmd_init(pmd, __pmd(0), init);
 533                         continue;
 534                 }
 535
 536                 if (!pmd_none(*pmd)) {
 537                         if (!pmd_leaf(*pmd)) {
 538                                 spin_lock(&init_mm.page_table_lock);
 539                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
 540                                 paddr_last = phys_pte_init(pte, paddr,
 541                                                            paddr_end, prot,
 542                                                            init);
 543                                 spin_unlock(&init_mm.page_table_lock);
 544                                 continue;
 545                         }
 546                         /*
 547                          * If we are ok with PG_LEVEL_2M mapping, then we will
 548                          * use the existing mapping,
 549                          *
 550                          * Otherwise, we will split the large page mapping but
 551                          * use the same existing protection bits except for
 552                          * large page, so that we don't violate Intel's TLB
 553                          * Application note (317080) which says, while changing
 554                          * the page sizes, new and old translations should
 555                          * not differ with respect to page frame and
 556                          * attributes.
 557                          */
 558                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 559                                 if (!after_bootmem)
 560                                         pages++;
 561                                 paddr_last = paddr_next;
 562                                 continue;
 563                         }
 564                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 565                 }
 566
 567                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 568                         pages++;
 569                         spin_lock(&init_mm.page_table_lock);
 570                         set_pmd_init(pmd,
 571                                      pfn_pmd(paddr >> PAGE_SHIFT, prot_sethuge(prot)),
 572                                      init);
 573                         spin_unlock(&init_mm.page_table_lock);
 574                         paddr_last = paddr_next;
 575                         continue;
 576                 }
 577
 578                 pte = alloc_low_page();
 579                 paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot, init);
 580
 581                 spin_lock(&init_mm.page_table_lock);
 582                 pmd_populate_kernel_init(&init_mm, pmd, pte, init);
 583                 spin_unlock(&init_mm.page_table_lock);
 584         }
 585         update_page_count(PG_LEVEL_2M, pages);
 586         return paddr_last;
 587 }
 588
 589 /*
 590  * Create PUD level page table mapping for physical addresses. The virtual
 591  * and physical address do not have to be aligned at this level. KASLR can
 592  * randomize virtual addresses up to this level.
 593  * It returns the last physical address mapped.
 594  */
 595 static unsigned long __meminit
 596 phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
 597               unsigned long page_size_mask, pgprot_t _prot, bool init)
 598 {
 599         unsigned long pages = 0, paddr_next;
 600         unsigned long paddr_last = paddr_end;
 601         unsigned long vaddr = (unsigned long)__va(paddr);
 602         int i = pud_index(vaddr);
 603
 604         for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
 605                 pud_t *pud;
 606                 pmd_t *pmd;
 607                 pgprot_t prot = _prot;
 608
 609                 vaddr = (unsigned long)__va(paddr);
 610                 pud = pud_page + pud_index(vaddr);
 611                 paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
 612
 613                 if (paddr >= paddr_end) {
 614                         if (!after_bootmem &&
 615                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 616                                              E820_TYPE_RAM) &&
 617                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 618                                              E820_TYPE_RESERVED_KERN) &&
 619                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 620                                              E820_TYPE_ACPI))
 621                                 set_pud_init(pud, __pud(0), init);
 622                         continue;
 623                 }
 624
 625                 if (!pud_none(*pud)) {
 626                         if (!pud_leaf(*pud)) {
 627                                 pmd = pmd_offset(pud, 0);
 628                                 paddr_last = phys_pmd_init(pmd, paddr,
 629                                                            paddr_end,
 630                                                            page_size_mask,
 631                                                            prot, init);
 632                                 continue;
 633                         }
 634                         /*
 635                          * If we are ok with PG_LEVEL_1G mapping, then we will
 636                          * use the existing mapping.
 637                          *
 638                          * Otherwise, we will split the gbpage mapping but use
 639                          * the same existing protection  bits except for large
 640                          * page, so that we don't violate Intel's TLB
 641                          * Application note (317080) which says, while changing
 642                          * the page sizes, new and old translations should
 643                          * not differ with respect to page frame and
 644                          * attributes.
 645                          */
 646                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 647                                 if (!after_bootmem)
 648                                         pages++;
 649                                 paddr_last = paddr_next;
 650                                 continue;
 651                         }
 652                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 653                 }
 654
 655                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 656                         pages++;
 657                         spin_lock(&init_mm.page_table_lock);
 658                         set_pud_init(pud,
 659                                      pfn_pud(paddr >> PAGE_SHIFT, prot_sethuge(prot)),
 660                                      init);
 661                         spin_unlock(&init_mm.page_table_lock);
 662                         paddr_last = paddr_next;
 663                         continue;
 664                 }
 665
 666                 pmd = alloc_low_page();
 667                 paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
 668                                            page_size_mask, prot, init);
 669
 670                 spin_lock(&init_mm.page_table_lock);
 671                 pud_populate_init(&init_mm, pud, pmd, init);
 672                 spin_unlock(&init_mm.page_table_lock);
 673         }
 674
 675         update_page_count(PG_LEVEL_1G, pages);
 676
 677         return paddr_last;
 678 }
 679
 680 static unsigned long __meminit
 681 phys_p4d_init(p4d_t *p4d_page, unsigned long paddr, unsigned long paddr_end,
 682               unsigned long page_size_mask, pgprot_t prot, bool init)
 683 {
 684         unsigned long vaddr, vaddr_end, vaddr_next, paddr_next, paddr_last;
 685
 686         paddr_last = paddr_end;
 687         vaddr = (unsigned long)__va(paddr);
 688         vaddr_end = (unsigned long)__va(paddr_end);
 689
 690         if (!pgtable_l5_enabled())
 691                 return phys_pud_init((pud_t *) p4d_page, paddr, paddr_end,
 692                                      page_size_mask, prot, init);
 693
 694         for (; vaddr < vaddr_end; vaddr = vaddr_next) {
 695                 p4d_t *p4d = p4d_page + p4d_index(vaddr);
 696                 pud_t *pud;
 697
 698                 vaddr_next = (vaddr & P4D_MASK) + P4D_SIZE;
 699                 paddr = __pa(vaddr);
 700
 701                 if (paddr >= paddr_end) {
 702                         paddr_next = __pa(vaddr_next);
 703                         if (!after_bootmem &&
 704                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 705                                              E820_TYPE_RAM) &&
 706                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 707                                              E820_TYPE_RESERVED_KERN) &&
 708                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 709                                              E820_TYPE_ACPI))
 710                                 set_p4d_init(p4d, __p4d(0), init);
 711                         continue;
 712                 }
 713
 714                 if (!p4d_none(*p4d)) {
 715                         pud = pud_offset(p4d, 0);
 716                         paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
 717                                         page_size_mask, prot, init);
 718                         continue;
 719                 }
 720
 721                 pud = alloc_low_page();
 722                 paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
 723                                            page_size_mask, prot, init);
 724
 725                 spin_lock(&init_mm.page_table_lock);
 726                 p4d_populate_init(&init_mm, p4d, pud, init);
 727                 spin_unlock(&init_mm.page_table_lock);
 728         }
 729
 730         return paddr_last;
 731 }
 732
 733 static unsigned long __meminit
 734 __kernel_physical_mapping_init(unsigned long paddr_start,
 735                                unsigned long paddr_end,
 736                                unsigned long page_size_mask,
 737                                pgprot_t prot, bool init)
 738 {
 739         bool pgd_changed = false;
 740         unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
 741
 742         paddr_last = paddr_end;
 743         vaddr = (unsigned long)__va(paddr_start);
 744         vaddr_end = (unsigned long)__va(paddr_end);
 745         vaddr_start = vaddr;
 746
 747         for (; vaddr < vaddr_end; vaddr = vaddr_next) {
 748                 pgd_t *pgd = pgd_offset_k(vaddr);
 749                 p4d_t *p4d;
 750
 751                 vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
 752
 753                 if (pgd_val(*pgd)) {
 754                         p4d = (p4d_t *)pgd_page_vaddr(*pgd);
 755                         paddr_last = phys_p4d_init(p4d, __pa(vaddr),
 756                                                    __pa(vaddr_end),
 757                                                    page_size_mask,
 758                                                    prot, init);
 759                         continue;
 760                 }
 761
 762                 p4d = alloc_low_page();
 763                 paddr_last = phys_p4d_init(p4d, __pa(vaddr), __pa(vaddr_end),
 764                                            page_size_mask, prot, init);
 765
 766                 spin_lock(&init_mm.page_table_lock);
 767                 if (pgtable_l5_enabled())
 768                         pgd_populate_init(&init_mm, pgd, p4d, init);
 769                 else
 770                         p4d_populate_init(&init_mm, p4d_offset(pgd, vaddr),
 771                                           (pud_t *) p4d, init);
 772
 773                 spin_unlock(&init_mm.page_table_lock);
 774                 pgd_changed = true;
 775         }
 776
 777         if (pgd_changed)
 778                 sync_global_pgds(vaddr_start, vaddr_end - 1);
 779
 780         return paddr_last;
 781 }
 782
 783
 784 /*
 785  * Create page table mapping for the physical memory for specific physical
 786  * addresses. Note that it can only be used to populate non-present entries.
 787  * The virtual and physical addresses have to be aligned on PMD level
 788  * down. It returns the last physical address mapped.
 789  */
 790 unsigned long __meminit
 791 kernel_physical_mapping_init(unsigned long paddr_start,
 792                              unsigned long paddr_end,
 793                              unsigned long page_size_mask, pgprot_t prot)
 794 {
 795         return __kernel_physical_mapping_init(paddr_start, paddr_end,
 796                                               page_size_mask, prot, true);
 797 }
 798
 799 /*
 800  * This function is similar to kernel_physical_mapping_init() above with the
 801  * exception that it uses set_{pud,pmd}() instead of the set_{pud,pte}_safe()
 802  * when updating the mapping. The caller is responsible to flush the TLBs after
 803  * the function returns.
 804  */
 805 unsigned long __meminit
 806 kernel_physical_mapping_change(unsigned long paddr_start,
 807                                unsigned long paddr_end,
 808                                unsigned long page_size_mask)
 809 {
 810         return __kernel_physical_mapping_init(paddr_start, paddr_end,
 811                                               page_size_mask, PAGE_KERNEL,
 812                                               false);
 813 }
 814
 815 #ifndef CONFIG_NUMA
 816 void __init initmem_init(void)
 817 {
 818         memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
 819 }
 820 #endif
 821
 822 void __init paging_init(void)
 823 {
 824         sparse_init();
 825
 826         /*
 827          * clear the default setting with node 0
 828          * note: don't use nodes_clear here, that is really clearing when
 829          *       numa support is not compiled in, and later node_set_state
 830          *       will not set it back.
 831          */
 832         node_clear_state(0, N_MEMORY);
 833         node_clear_state(0, N_NORMAL_MEMORY);
 834
 835         zone_sizes_init();
 836 }
 837
 838 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 839 #define PAGE_UNUSED 0xFD
 840
 841 /*
 842  * The unused vmemmap range, which was not yet memset(PAGE_UNUSED), ranges
 843  * from unused_pmd_start to next PMD_SIZE boundary.
 844  */
 845 static unsigned long unused_pmd_start __meminitdata;
 846
 847 static void __meminit vmemmap_flush_unused_pmd(void)
 848 {
 849         if (!unused_pmd_start)
 850                 return;
 851         /*
 852          * Clears (unused_pmd_start, PMD_END]
 853          */
 854         memset((void *)unused_pmd_start, PAGE_UNUSED,
 855                ALIGN(unused_pmd_start, PMD_SIZE) - unused_pmd_start);
 856         unused_pmd_start = 0;
 857 }
 858
 859 #ifdef CONFIG_MEMORY_HOTPLUG
 860 /* Returns true if the PMD is completely unused and thus it can be freed */
 861 static bool __meminit vmemmap_pmd_is_unused(unsigned long addr, unsigned long end)
 862 {
 863         unsigned long start = ALIGN_DOWN(addr, PMD_SIZE);
 864
 865         /*
 866          * Flush the unused range cache to ensure that memchr_inv() will work
 867          * for the whole range.
 868          */
 869         vmemmap_flush_unused_pmd();
 870         memset((void *)addr, PAGE_UNUSED, end - addr);
 871
 872         return !memchr_inv((void *)start, PAGE_UNUSED, PMD_SIZE);
 873 }
 874 #endif
 875
 876 static void __meminit __vmemmap_use_sub_pmd(unsigned long start)
 877 {
 878         /*
 879          * As we expect to add in the same granularity as we remove, it's
 880          * sufficient to mark only some piece used to block the memmap page from
 881          * getting removed when removing some other adjacent memmap (just in
 882          * case the first memmap never gets initialized e.g., because the memory
 883          * block never gets onlined).
 884          */
 885         memset((void *)start, 0, sizeof(struct page));
 886 }
 887
 888 static void __meminit vmemmap_use_sub_pmd(unsigned long start, unsigned long end)
 889 {
 890         /*
 891          * We only optimize if the new used range directly follows the
 892          * previously unused range (esp., when populating consecutive sections).
 893          */
 894         if (unused_pmd_start == start) {
 895                 if (likely(IS_ALIGNED(end, PMD_SIZE)))
 896                         unused_pmd_start = 0;
 897                 else
 898                         unused_pmd_start = end;
 899                 return;
 900         }
 901
 902         /*
 903          * If the range does not contiguously follows previous one, make sure
 904          * to mark the unused range of the previous one so it can be removed.
 905          */
 906         vmemmap_flush_unused_pmd();
 907         __vmemmap_use_sub_pmd(start);
 908 }
 909
 910
 911 static void __meminit vmemmap_use_new_sub_pmd(unsigned long start, unsigned long end)
 912 {
 913         const unsigned long page = ALIGN_DOWN(start, PMD_SIZE);
 914
 915         vmemmap_flush_unused_pmd();
 916
 917         /*
 918          * Could be our memmap page is filled with PAGE_UNUSED already from a
 919          * previous remove. Make sure to reset it.
 920          */
 921         __vmemmap_use_sub_pmd(start);
 922
 923         /*
 924          * Mark with PAGE_UNUSED the unused parts of the new memmap range
 925          */
 926         if (!IS_ALIGNED(start, PMD_SIZE))
 927                 memset((void *)page, PAGE_UNUSED, start - page);
 928
 929         /*
 930          * We want to avoid memset(PAGE_UNUSED) when populating the vmemmap of
 931          * consecutive sections. Remember for the last added PMD where the
 932          * unused range begins.
 933          */
 934         if (!IS_ALIGNED(end, PMD_SIZE))
 935                 unused_pmd_start = end;
 936 }
 937 #endif
 938
 939 /*
 940  * Memory hotplug specific functions
 941  */
 942 #ifdef CONFIG_MEMORY_HOTPLUG
 943 /*
 944  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 945  * updating.
 946  */
 947 static void update_end_of_memory_vars(u64 start, u64 size)
 948 {
 949         unsigned long end_pfn = PFN_UP(start + size);
 950
 951         if (end_pfn > max_pfn) {
 952                 max_pfn = end_pfn;
 953                 max_low_pfn = end_pfn;
 954                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 955         }
 956 }
 957
 958 int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
 959               struct mhp_params *params)
 960 {
 961         unsigned long end = ((start_pfn + nr_pages) << PAGE_SHIFT) - 1;
 962         int ret;
 963
 964         if (WARN_ON_ONCE(end > DIRECT_MAP_PHYSMEM_END))
 965                 return -ERANGE;
 966
 967         ret = __add_pages(nid, start_pfn, nr_pages, params);
 968         WARN_ON_ONCE(ret);
 969
 970         /* update max_pfn, max_low_pfn and high_memory */
 971         update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
 972                                   nr_pages << PAGE_SHIFT);
 973
 974         return ret;
 975 }
 976
 977 int arch_add_memory(int nid, u64 start, u64 size,
 978                     struct mhp_params *params)
 979 {
 980         unsigned long start_pfn = start >> PAGE_SHIFT;
 981         unsigned long nr_pages = size >> PAGE_SHIFT;
 982
 983         init_memory_mapping(start, start + size, params->pgprot);
 984
 985         return add_pages(nid, start_pfn, nr_pages, params);
 986 }
 987
 988 static void free_reserved_pages(struct page *page, unsigned long nr_pages)
 989 {
 990         while (nr_pages--)
 991                 free_reserved_page(page++);
 992 }
 993
 994 static void __meminit free_pagetable(struct page *page, int order)
 995 {
 996         /* bootmem page has reserved flag */
 997         if (PageReserved(page)) {
 998                 unsigned long nr_pages = 1 << order;
 999 #ifdef CONFIG_HAVE_BOOTMEM_INFO_NODE
1000                 enum bootmem_type type = bootmem_type(page);
1001
1002                 if (type == SECTION_INFO || type == MIX_SECTION_INFO) {
1003                         while (nr_pages--)
1004                                 put_page_bootmem(page++);
1005                 } else {
1006                         free_reserved_pages(page, nr_pages);
1007                 }
1008 #else
1009                 free_reserved_pages(page, nr_pages);
1010 #endif
1011         } else {
1012                 free_pages((unsigned long)page_address(page), order);
1013         }
1014 }
1015
1016 static void __meminit free_hugepage_table(struct page *page,
1017                 struct vmem_altmap *altmap)
1018 {
1019         if (altmap)
1020                 vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
1021         else
1022                 free_pagetable(page, get_order(PMD_SIZE));
1023 }
1024
1025 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
1026 {
1027         pte_t *pte;
1028         int i;
1029
1030         for (i = 0; i < PTRS_PER_PTE; i++) {
1031                 pte = pte_start + i;
1032                 if (!pte_none(*pte))
1033                         return;
1034         }
1035
1036         /* free a pte table */
1037         free_pagetable(pmd_page(*pmd), 0);
1038         spin_lock(&init_mm.page_table_lock);
1039         pmd_clear(pmd);
1040         spin_unlock(&init_mm.page_table_lock);
1041 }
1042
1043 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
1044 {
1045         pmd_t *pmd;
1046         int i;
1047
1048         for (i = 0; i < PTRS_PER_PMD; i++) {
1049                 pmd = pmd_start + i;
1050                 if (!pmd_none(*pmd))
1051                         return;
1052         }
1053
1054         /* free a pmd table */
1055         free_pagetable(pud_page(*pud), 0);
1056         spin_lock(&init_mm.page_table_lock);
1057         pud_clear(pud);
1058         spin_unlock(&init_mm.page_table_lock);
1059 }
1060
1061 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
1062 {
1063         pud_t *pud;
1064         int i;
1065
1066         for (i = 0; i < PTRS_PER_PUD; i++) {
1067                 pud = pud_start + i;
1068                 if (!pud_none(*pud))
1069                         return;
1070         }
1071
1072         /* free a pud table */
1073         free_pagetable(p4d_page(*p4d), 0);
1074         spin_lock(&init_mm.page_table_lock);
1075         p4d_clear(p4d);
1076         spin_unlock(&init_mm.page_table_lock);
1077 }
1078
1079 static void __meminit
1080 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
1081                  bool direct)
1082 {
1083         unsigned long next, pages = 0;
1084         pte_t *pte;
1085         phys_addr_t phys_addr;
1086
1087         pte = pte_start + pte_index(addr);
1088         for (; addr < end; addr = next, pte++) {
1089                 next = (addr + PAGE_SIZE) & PAGE_MASK;
1090                 if (next > end)
1091                         next = end;
1092
1093                 if (!pte_present(*pte))
1094                         continue;
1095
1096                 /*
1097                  * We mapped [0,1G) memory as identity mapping when
1098                  * initializing, in arch/x86/kernel/head_64.S. These
1099                  * pagetables cannot be removed.
1100                  */
1101                 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
1102                 if (phys_addr < (phys_addr_t)0x40000000)
1103                         return;
1104
1105                 if (!direct)
1106                         free_pagetable(pte_page(*pte), 0);
1107
1108                 spin_lock(&init_mm.page_table_lock);
1109                 pte_clear(&init_mm, addr, pte);
1110                 spin_unlock(&init_mm.page_table_lock);
1111
1112                 /* For non-direct mapping, pages means nothing. */
1113                 pages++;
1114         }
1115
1116         /* Call free_pte_table() in remove_pmd_table(). */
1117         flush_tlb_all();
1118         if (direct)
1119                 update_page_count(PG_LEVEL_4K, -pages);
1120 }
1121
1122 static void __meminit
1123 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
1124                  bool direct, struct vmem_altmap *altmap)
1125 {
1126         unsigned long next, pages = 0;
1127         pte_t *pte_base;
1128         pmd_t *pmd;
1129
1130         pmd = pmd_start + pmd_index(addr);
1131         for (; addr < end; addr = next, pmd++) {
1132                 next = pmd_addr_end(addr, end);
1133
1134                 if (!pmd_present(*pmd))
1135                         continue;
1136
1137                 if (pmd_leaf(*pmd)) {
1138                         if (IS_ALIGNED(addr, PMD_SIZE) &&
1139                             IS_ALIGNED(next, PMD_SIZE)) {
1140                                 if (!direct)
1141                                         free_hugepage_table(pmd_page(*pmd),
1142                                                             altmap);
1143
1144                                 spin_lock(&init_mm.page_table_lock);
1145                                 pmd_clear(pmd);
1146                                 spin_unlock(&init_mm.page_table_lock);
1147                                 pages++;
1148                         }
1149 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1150                         else if (vmemmap_pmd_is_unused(addr, next)) {
1151                                         free_hugepage_table(pmd_page(*pmd),
1152                                                             altmap);
1153                                         spin_lock(&init_mm.page_table_lock);
1154                                         pmd_clear(pmd);
1155                                         spin_unlock(&init_mm.page_table_lock);
1156                         }
1157 #endif
1158                         continue;
1159                 }
1160
1161                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1162                 remove_pte_table(pte_base, addr, next, direct);
1163                 free_pte_table(pte_base, pmd);
1164         }
1165
1166         /* Call free_pmd_table() in remove_pud_table(). */
1167         if (direct)
1168                 update_page_count(PG_LEVEL_2M, -pages);
1169 }
1170
1171 static void __meminit
1172 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1173                  struct vmem_altmap *altmap, bool direct)
1174 {
1175         unsigned long next, pages = 0;
1176         pmd_t *pmd_base;
1177         pud_t *pud;
1178
1179         pud = pud_start + pud_index(addr);
1180         for (; addr < end; addr = next, pud++) {
1181                 next = pud_addr_end(addr, end);
1182
1183                 if (!pud_present(*pud))
1184                         continue;
1185
1186                 if (pud_leaf(*pud) &&
1187                     IS_ALIGNED(addr, PUD_SIZE) &&
1188                     IS_ALIGNED(next, PUD_SIZE)) {
1189                         spin_lock(&init_mm.page_table_lock);
1190                         pud_clear(pud);
1191                         spin_unlock(&init_mm.page_table_lock);
1192                         pages++;
1193                         continue;
1194                 }
1195
1196                 pmd_base = pmd_offset(pud, 0);
1197                 remove_pmd_table(pmd_base, addr, next, direct, altmap);
1198                 free_pmd_table(pmd_base, pud);
1199         }
1200
1201         if (direct)
1202                 update_page_count(PG_LEVEL_1G, -pages);
1203 }
1204
1205 static void __meminit
1206 remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1207                  struct vmem_altmap *altmap, bool direct)
1208 {
1209         unsigned long next, pages = 0;
1210         pud_t *pud_base;
1211         p4d_t *p4d;
1212
1213         p4d = p4d_start + p4d_index(addr);
1214         for (; addr < end; addr = next, p4d++) {
1215                 next = p4d_addr_end(addr, end);
1216
1217                 if (!p4d_present(*p4d))
1218                         continue;
1219
1220                 BUILD_BUG_ON(p4d_leaf(*p4d));
1221
1222                 pud_base = pud_offset(p4d, 0);
1223                 remove_pud_table(pud_base, addr, next, altmap, direct);
1224                 /*
1225                  * For 4-level page tables we do not want to free PUDs, but in the
1226                  * 5-level case we should free them. This code will have to change
1227                  * to adapt for boot-time switching between 4 and 5 level page tables.
1228                  */
1229                 if (pgtable_l5_enabled())
1230                         free_pud_table(pud_base, p4d);
1231         }
1232
1233         if (direct)
1234                 update_page_count(PG_LEVEL_512G, -pages);
1235 }
1236
1237 /* start and end are both virtual address. */
1238 static void __meminit
1239 remove_pagetable(unsigned long start, unsigned long end, bool direct,
1240                 struct vmem_altmap *altmap)
1241 {
1242         unsigned long next;
1243         unsigned long addr;
1244         pgd_t *pgd;
1245         p4d_t *p4d;
1246
1247         for (addr = start; addr < end; addr = next) {
1248                 next = pgd_addr_end(addr, end);
1249
1250                 pgd = pgd_offset_k(addr);
1251                 if (!pgd_present(*pgd))
1252                         continue;
1253
1254                 p4d = p4d_offset(pgd, 0);
1255                 remove_p4d_table(p4d, addr, next, altmap, direct);
1256         }
1257
1258         flush_tlb_all();
1259 }
1260
1261 void __ref vmemmap_free(unsigned long start, unsigned long end,
1262                 struct vmem_altmap *altmap)
1263 {
1264         VM_BUG_ON(!PAGE_ALIGNED(start));
1265         VM_BUG_ON(!PAGE_ALIGNED(end));
1266
1267         remove_pagetable(start, end, false, altmap);
1268 }
1269
1270 static void __meminit
1271 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1272 {
1273         start = (unsigned long)__va(start);
1274         end = (unsigned long)__va(end);
1275
1276         remove_pagetable(start, end, true, NULL);
1277 }
1278
1279 void __ref arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
1280 {
1281         unsigned long start_pfn = start >> PAGE_SHIFT;
1282         unsigned long nr_pages = size >> PAGE_SHIFT;
1283
1284         __remove_pages(start_pfn, nr_pages, altmap);
1285         kernel_physical_mapping_remove(start, start + size);
1286 }
1287 #endif /* CONFIG_MEMORY_HOTPLUG */
1288
1289 static struct kcore_list kcore_vsyscall;
1290
1291 static void __init register_page_bootmem_info(void)
1292 {
1293 #if defined(CONFIG_NUMA) || defined(CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP)
1294         int i;
1295
1296         for_each_online_node(i)
1297                 register_page_bootmem_info_node(NODE_DATA(i));
1298 #endif
1299 }
1300
1301 /*
1302  * Pre-allocates page-table pages for the vmalloc area in the kernel page-table.
1303  * Only the level which needs to be synchronized between all page-tables is
1304  * allocated because the synchronization can be expensive.
1305  */
1306 static void __init preallocate_vmalloc_pages(void)
1307 {
1308         unsigned long addr;
1309         const char *lvl;
1310
1311         for (addr = VMALLOC_START; addr <= VMEMORY_END; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
1312                 pgd_t *pgd = pgd_offset_k(addr);
1313                 p4d_t *p4d;
1314                 pud_t *pud;
1315
1316                 lvl = "p4d";
1317                 p4d = p4d_alloc(&init_mm, pgd, addr);
1318                 if (!p4d)
1319                         goto failed;
1320
1321                 if (pgtable_l5_enabled())
1322                         continue;
1323
1324                 /*
1325                  * The goal here is to allocate all possibly required
1326                  * hardware page tables pointed to by the top hardware
1327                  * level.
1328                  *
1329                  * On 4-level systems, the P4D layer is folded away and
1330                  * the above code does no preallocation.  Below, go down
1331                  * to the pud _software_ level to ensure the second
1332                  * hardware level is allocated on 4-level systems too.
1333                  */
1334                 lvl = "pud";
1335                 pud = pud_alloc(&init_mm, p4d, addr);
1336                 if (!pud)
1337                         goto failed;
1338         }
1339
1340         return;
1341
1342 failed:
1343
1344         /*
1345          * The pages have to be there now or they will be missing in
1346          * process page-tables later.
1347          */
1348         panic("Failed to pre-allocate %s pages for vmalloc area\n", lvl);
1349 }
1350
1351 void __init mem_init(void)
1352 {
1353         pci_iommu_alloc();
1354
1355         /* clear_bss() already clear the empty_zero_page */
1356
1357         /* this will put all memory onto the freelists */
1358         memblock_free_all();
1359         after_bootmem = 1;
1360         x86_init.hyper.init_after_bootmem();
1361
1362         /*
1363          * Must be done after boot memory is put on freelist, because here we
1364          * might set fields in deferred struct pages that have not yet been
1365          * initialized, and memblock_free_all() initializes all the reserved
1366          * deferred pages for us.
1367          */
1368         register_page_bootmem_info();
1369
1370         /* Register memory areas for /proc/kcore */
1371         if (get_gate_vma(&init_mm))
1372                 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1373
1374         preallocate_vmalloc_pages();
1375 }
1376
1377 int kernel_set_to_readonly;
1378
1379 void mark_rodata_ro(void)
1380 {
1381         unsigned long start = PFN_ALIGN(_text);
1382         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1383         unsigned long end = (unsigned long)__end_rodata_hpage_align;
1384         unsigned long text_end = PFN_ALIGN(_etext);
1385         unsigned long rodata_end = PFN_ALIGN(__end_rodata);
1386         unsigned long all_end;
1387
1388         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1389                (end - start) >> 10);
1390         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1391
1392         kernel_set_to_readonly = 1;
1393
1394         /*
1395          * The rodata/data/bss/brk section (but not the kernel text!)
1396          * should also be not-executable.
1397          *
1398          * We align all_end to PMD_SIZE because the existing mapping
1399          * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1400          * split the PMD and the reminder between _brk_end and the end
1401          * of the PMD will remain mapped executable.
1402          *
1403          * Any PMD which was setup after the one which covers _brk_end
1404          * has been zapped already via cleanup_highmem().
1405          */
1406         all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1407         set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1408
1409         set_ftrace_ops_ro();
1410
1411 #ifdef CONFIG_CPA_DEBUG
1412         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1413         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1414
1415         printk(KERN_INFO "Testing CPA: again\n");
1416         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1417 #endif
1418
1419         free_kernel_image_pages("unused kernel image (text/rodata gap)",
1420                                 (void *)text_end, (void *)rodata_start);
1421         free_kernel_image_pages("unused kernel image (rodata/data gap)",
1422                                 (void *)rodata_end, (void *)_sdata);
1423 }
1424
1425 /*
1426  * Block size is the minimum amount of memory which can be hotplugged or
1427  * hotremoved. It must be power of two and must be equal or larger than
1428  * MIN_MEMORY_BLOCK_SIZE.
1429  */
1430 #define MAX_BLOCK_SIZE (2UL << 30)
1431
1432 /* Amount of ram needed to start using large blocks */
1433 #define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1434
1435 /* Adjustable memory block size */
1436 static unsigned long set_memory_block_size;
1437 int __init set_memory_block_size_order(unsigned int order)
1438 {
1439         unsigned long size = 1UL << order;
1440
1441         if (size > MEM_SIZE_FOR_LARGE_BLOCK || size < MIN_MEMORY_BLOCK_SIZE)
1442                 return -EINVAL;
1443
1444         set_memory_block_size = size;
1445         return 0;
1446 }
1447
1448 static unsigned long probe_memory_block_size(void)
1449 {
1450         unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1451         unsigned long bz;
1452
1453         /* If memory block size has been set, then use it */
1454         bz = set_memory_block_size;
1455         if (bz)
1456                 goto done;
1457
1458         /* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1459         if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1460                 bz = MIN_MEMORY_BLOCK_SIZE;
1461                 goto done;
1462         }
1463
1464         /*
1465          * Use max block size to minimize overhead on bare metal, where
1466          * alignment for memory hotplug isn't a concern.
1467          */
1468         if (!boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
1469                 bz = MAX_BLOCK_SIZE;
1470                 goto done;
1471         }
1472
1473         /* Find the largest allowed block size that aligns to memory end */
1474         for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1475                 if (IS_ALIGNED(boot_mem_end, bz))
1476                         break;
1477         }
1478 done:
1479         pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1480
1481         return bz;
1482 }
1483
1484 static unsigned long memory_block_size_probed;
1485 unsigned long memory_block_size_bytes(void)
1486 {
1487         if (!memory_block_size_probed)
1488                 memory_block_size_probed = probe_memory_block_size();
1489
1490         return memory_block_size_probed;
1491 }
1492
1493 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1494 /*
1495  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1496  */
1497 static long __meminitdata addr_start, addr_end;
1498 static void __meminitdata *p_start, *p_end;
1499 static int __meminitdata node_start;
1500
1501 void __meminit vmemmap_set_pmd(pmd_t *pmd, void *p, int node,
1502                                unsigned long addr, unsigned long next)
1503 {
1504         pte_t entry;
1505
1506         entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1507                         PAGE_KERNEL_LARGE);
1508         set_pmd(pmd, __pmd(pte_val(entry)));
1509
1510         /* check to see if we have contiguous blocks */
1511         if (p_end != p || node_start != node) {
1512                 if (p_start)
1513                         pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1514                                 addr_start, addr_end-1, p_start, p_end-1, node_start);
1515                 addr_start = addr;
1516                 node_start = node;
1517                 p_start = p;
1518         }
1519
1520         addr_end = addr + PMD_SIZE;
1521         p_end = p + PMD_SIZE;
1522
1523         if (!IS_ALIGNED(addr, PMD_SIZE) ||
1524                 !IS_ALIGNED(next, PMD_SIZE))
1525                 vmemmap_use_new_sub_pmd(addr, next);
1526 }
1527
1528 int __meminit vmemmap_check_pmd(pmd_t *pmd, int node,
1529                                 unsigned long addr, unsigned long next)
1530 {
1531         int large = pmd_leaf(*pmd);
1532
1533         if (pmd_leaf(*pmd)) {
1534                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1535                 vmemmap_use_sub_pmd(addr, next);
1536         }
1537
1538         return large;
1539 }
1540
1541 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1542                 struct vmem_altmap *altmap)
1543 {
1544         int err;
1545
1546         VM_BUG_ON(!PAGE_ALIGNED(start));
1547         VM_BUG_ON(!PAGE_ALIGNED(end));
1548
1549         if (end - start < PAGES_PER_SECTION * sizeof(struct page))
1550                 err = vmemmap_populate_basepages(start, end, node, NULL);
1551         else if (boot_cpu_has(X86_FEATURE_PSE))
1552                 err = vmemmap_populate_hugepages(start, end, node, altmap);
1553         else if (altmap) {
1554                 pr_err_once("%s: no cpu support for altmap allocations\n",
1555                                 __func__);
1556                 err = -ENOMEM;
1557         } else
1558                 err = vmemmap_populate_basepages(start, end, node, NULL);
1559         if (!err)
1560                 sync_global_pgds(start, end - 1);
1561         return err;
1562 }
1563
1564 #ifdef CONFIG_HAVE_BOOTMEM_INFO_NODE
1565 void register_page_bootmem_memmap(unsigned long section_nr,
1566                                   struct page *start_page, unsigned long nr_pages)
1567 {
1568         unsigned long addr = (unsigned long)start_page;
1569         unsigned long end = (unsigned long)(start_page + nr_pages);
1570         unsigned long next;
1571         pgd_t *pgd;
1572         p4d_t *p4d;
1573         pud_t *pud;
1574         pmd_t *pmd;
1575         unsigned int nr_pmd_pages;
1576         struct page *page;
1577
1578         for (; addr < end; addr = next) {
1579                 pte_t *pte = NULL;
1580
1581                 pgd = pgd_offset_k(addr);
1582                 if (pgd_none(*pgd)) {
1583                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1584                         continue;
1585                 }
1586                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1587
1588                 p4d = p4d_offset(pgd, addr);
1589                 if (p4d_none(*p4d)) {
1590                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1591                         continue;
1592                 }
1593                 get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1594
1595                 pud = pud_offset(p4d, addr);
1596                 if (pud_none(*pud)) {
1597                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1598                         continue;
1599                 }
1600                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1601
1602                 if (!boot_cpu_has(X86_FEATURE_PSE)) {
1603                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1604                         pmd = pmd_offset(pud, addr);
1605                         if (pmd_none(*pmd))
1606                                 continue;
1607                         get_page_bootmem(section_nr, pmd_page(*pmd),
1608                                          MIX_SECTION_INFO);
1609
1610                         pte = pte_offset_kernel(pmd, addr);
1611                         if (pte_none(*pte))
1612                                 continue;
1613                         get_page_bootmem(section_nr, pte_page(*pte),
1614                                          SECTION_INFO);
1615                 } else {
1616                         next = pmd_addr_end(addr, end);
1617
1618                         pmd = pmd_offset(pud, addr);
1619                         if (pmd_none(*pmd))
1620                                 continue;
1621
1622                         nr_pmd_pages = 1 << get_order(PMD_SIZE);
1623                         page = pmd_page(*pmd);
1624                         while (nr_pmd_pages--)
1625                                 get_page_bootmem(section_nr, page++,
1626                                                  SECTION_INFO);
1627                 }
1628         }
1629 }
1630 #endif
1631
1632 void __meminit vmemmap_populate_print_last(void)
1633 {
1634         if (p_start) {
1635                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1636                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1637                 p_start = NULL;
1638                 p_end = NULL;
1639                 node_start = 0;
1640         }
1641 }
1642 #endif