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 <pavel@ucw.cz>
   7  *  Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
   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
  37 #include <asm/processor.h>
  38 #include <asm/bios_ebda.h>
  39 #include <linux/uaccess.h>
  40 #include <asm/pgalloc.h>
  41 #include <asm/dma.h>
  42 #include <asm/fixmap.h>
  43 #include <asm/e820/api.h>
  44 #include <asm/apic.h>
  45 #include <asm/tlb.h>
  46 #include <asm/mmu_context.h>
  47 #include <asm/proto.h>
  48 #include <asm/smp.h>
  49 #include <asm/sections.h>
  50 #include <asm/kdebug.h>
  51 #include <asm/numa.h>
  52 #include <asm/set_memory.h>
  53 #include <asm/init.h>
  54 #include <asm/uv/uv.h>
  55 #include <asm/setup.h>
  56 #include <asm/ftrace.h>
  57
  58 #include "mm_internal.h"
  59
  60 #include "ident_map.c"
  61
  62 #define DEFINE_POPULATE(fname, type1, type2, init)              \
  63 static inline void fname##_init(struct mm_struct *mm,           \
  64                 type1##_t *arg1, type2##_t *arg2, bool init)    \
  65 {                                                               \
  66         if (init)                                               \
  67                 fname##_safe(mm, arg1, arg2);                   \
  68         else                                                    \
  69                 fname(mm, arg1, arg2);                          \
  70 }
  71
  72 DEFINE_POPULATE(p4d_populate, p4d, pud, init)
  73 DEFINE_POPULATE(pgd_populate, pgd, p4d, init)
  74 DEFINE_POPULATE(pud_populate, pud, pmd, init)
  75 DEFINE_POPULATE(pmd_populate_kernel, pmd, pte, init)
  76
  77 #define DEFINE_ENTRY(type1, type2, init)                        \
  78 static inline void set_##type1##_init(type1##_t *arg1,          \
  79                         type2##_t arg2, bool init)              \
  80 {                                                               \
  81         if (init)                                               \
  82                 set_##type1##_safe(arg1, arg2);                 \
  83         else                                                    \
  84                 set_##type1(arg1, arg2);                        \
  85 }
  86
  87 DEFINE_ENTRY(p4d, p4d, init)
  88 DEFINE_ENTRY(pud, pud, init)
  89 DEFINE_ENTRY(pmd, pmd, init)
  90 DEFINE_ENTRY(pte, pte, init)
  91
  92
  93 /*
  94  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  95  * physical space so we can cache the place of the first one and move
  96  * around without checking the pgd every time.
  97  */
  98
  99 /* Bits supported by the hardware: */
 100 pteval_t __supported_pte_mask __read_mostly = ~0;
 101 /* Bits allowed in normal kernel mappings: */
 102 pteval_t __default_kernel_pte_mask __read_mostly = ~0;
 103 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 104 /* Used in PAGE_KERNEL_* macros which are reasonably used out-of-tree: */
 105 EXPORT_SYMBOL(__default_kernel_pte_mask);
 106
 107 int force_personality32;
 108
 109 /*
 110  * noexec32=on|off
 111  * Control non executable heap for 32bit processes.
 112  * To control the stack too use noexec=off
 113  *
 114  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
 115  * off  PROT_READ implies PROT_EXEC
 116  */
 117 static int __init nonx32_setup(char *str)
 118 {
 119         if (!strcmp(str, "on"))
 120                 force_personality32 &= ~READ_IMPLIES_EXEC;
 121         else if (!strcmp(str, "off"))
 122                 force_personality32 |= READ_IMPLIES_EXEC;
 123         return 1;
 124 }
 125 __setup("noexec32=", nonx32_setup);
 126
 127 static void sync_global_pgds_l5(unsigned long start, unsigned long end)
 128 {
 129         unsigned long addr;
 130
 131         for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
 132                 const pgd_t *pgd_ref = pgd_offset_k(addr);
 133                 struct page *page;
 134
 135                 /* Check for overflow */
 136                 if (addr < start)
 137                         break;
 138
 139                 if (pgd_none(*pgd_ref))
 140                         continue;
 141
 142                 spin_lock(&pgd_lock);
 143                 list_for_each_entry(page, &pgd_list, lru) {
 144                         pgd_t *pgd;
 145                         spinlock_t *pgt_lock;
 146
 147                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
 148                         /* the pgt_lock only for Xen */
 149                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 150                         spin_lock(pgt_lock);
 151
 152                         if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
 153                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 154
 155                         if (pgd_none(*pgd))
 156                                 set_pgd(pgd, *pgd_ref);
 157
 158                         spin_unlock(pgt_lock);
 159                 }
 160                 spin_unlock(&pgd_lock);
 161         }
 162 }
 163
 164 static void sync_global_pgds_l4(unsigned long start, unsigned long end)
 165 {
 166         unsigned long addr;
 167
 168         for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
 169                 pgd_t *pgd_ref = pgd_offset_k(addr);
 170                 const p4d_t *p4d_ref;
 171                 struct page *page;
 172
 173                 /*
 174                  * With folded p4d, pgd_none() is always false, we need to
 175                  * handle synchonization on p4d level.
 176                  */
 177                 MAYBE_BUILD_BUG_ON(pgd_none(*pgd_ref));
 178                 p4d_ref = p4d_offset(pgd_ref, addr);
 179
 180                 if (p4d_none(*p4d_ref))
 181                         continue;
 182
 183                 spin_lock(&pgd_lock);
 184                 list_for_each_entry(page, &pgd_list, lru) {
 185                         pgd_t *pgd;
 186                         p4d_t *p4d;
 187                         spinlock_t *pgt_lock;
 188
 189                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
 190                         p4d = p4d_offset(pgd, addr);
 191                         /* the pgt_lock only for Xen */
 192                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 193                         spin_lock(pgt_lock);
 194
 195                         if (!p4d_none(*p4d_ref) && !p4d_none(*p4d))
 196                                 BUG_ON(p4d_page_vaddr(*p4d)
 197                                        != p4d_page_vaddr(*p4d_ref));
 198
 199                         if (p4d_none(*p4d))
 200                                 set_p4d(p4d, *p4d_ref);
 201
 202                         spin_unlock(pgt_lock);
 203                 }
 204                 spin_unlock(&pgd_lock);
 205         }
 206 }
 207
 208 /*
 209  * When memory was added make sure all the processes MM have
 210  * suitable PGD entries in the local PGD level page.
 211  */
 212 static void sync_global_pgds(unsigned long start, unsigned long end)
 213 {
 214         if (pgtable_l5_enabled())
 215                 sync_global_pgds_l5(start, end);
 216         else
 217                 sync_global_pgds_l4(start, end);
 218 }
 219
 220 void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
 221 {
 222         sync_global_pgds(start, end);
 223 }
 224
 225 /*
 226  * NOTE: This function is marked __ref because it calls __init function
 227  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 228  */
 229 static __ref void *spp_getpage(void)
 230 {
 231         void *ptr;
 232
 233         if (after_bootmem)
 234                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 235         else
 236                 ptr = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
 237
 238         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 239                 panic("set_pte_phys: cannot allocate page data %s\n",
 240                         after_bootmem ? "after bootmem" : "");
 241         }
 242
 243         pr_debug("spp_getpage %p\n", ptr);
 244
 245         return ptr;
 246 }
 247
 248 static p4d_t *fill_p4d(pgd_t *pgd, unsigned long vaddr)
 249 {
 250         if (pgd_none(*pgd)) {
 251                 p4d_t *p4d = (p4d_t *)spp_getpage();
 252                 pgd_populate(&init_mm, pgd, p4d);
 253                 if (p4d != p4d_offset(pgd, 0))
 254                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 255                                p4d, p4d_offset(pgd, 0));
 256         }
 257         return p4d_offset(pgd, vaddr);
 258 }
 259
 260 static pud_t *fill_pud(p4d_t *p4d, unsigned long vaddr)
 261 {
 262         if (p4d_none(*p4d)) {
 263                 pud_t *pud = (pud_t *)spp_getpage();
 264                 p4d_populate(&init_mm, p4d, pud);
 265                 if (pud != pud_offset(p4d, 0))
 266                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 267                                pud, pud_offset(p4d, 0));
 268         }
 269         return pud_offset(p4d, vaddr);
 270 }
 271
 272 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 273 {
 274         if (pud_none(*pud)) {
 275                 pmd_t *pmd = (pmd_t *) spp_getpage();
 276                 pud_populate(&init_mm, pud, pmd);
 277                 if (pmd != pmd_offset(pud, 0))
 278                         printk(KERN_ERR "PAGETABLE BUG #02! %p <-> %p\n",
 279                                pmd, pmd_offset(pud, 0));
 280         }
 281         return pmd_offset(pud, vaddr);
 282 }
 283
 284 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 285 {
 286         if (pmd_none(*pmd)) {
 287                 pte_t *pte = (pte_t *) spp_getpage();
 288                 pmd_populate_kernel(&init_mm, pmd, pte);
 289                 if (pte != pte_offset_kernel(pmd, 0))
 290                         printk(KERN_ERR "PAGETABLE BUG #03!\n");
 291         }
 292         return pte_offset_kernel(pmd, vaddr);
 293 }
 294
 295 static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
 296 {
 297         pmd_t *pmd = fill_pmd(pud, vaddr);
 298         pte_t *pte = fill_pte(pmd, vaddr);
 299
 300         set_pte(pte, new_pte);
 301
 302         /*
 303          * It's enough to flush this one mapping.
 304          * (PGE mappings get flushed as well)
 305          */
 306         flush_tlb_one_kernel(vaddr);
 307 }
 308
 309 void set_pte_vaddr_p4d(p4d_t *p4d_page, unsigned long vaddr, pte_t new_pte)
 310 {
 311         p4d_t *p4d = p4d_page + p4d_index(vaddr);
 312         pud_t *pud = fill_pud(p4d, vaddr);
 313
 314         __set_pte_vaddr(pud, vaddr, new_pte);
 315 }
 316
 317 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 318 {
 319         pud_t *pud = pud_page + pud_index(vaddr);
 320
 321         __set_pte_vaddr(pud, vaddr, new_pte);
 322 }
 323
 324 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 325 {
 326         pgd_t *pgd;
 327         p4d_t *p4d_page;
 328
 329         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 330
 331         pgd = pgd_offset_k(vaddr);
 332         if (pgd_none(*pgd)) {
 333                 printk(KERN_ERR
 334                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 335                 return;
 336         }
 337
 338         p4d_page = p4d_offset(pgd, 0);
 339         set_pte_vaddr_p4d(p4d_page, vaddr, pteval);
 340 }
 341
 342 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 343 {
 344         pgd_t *pgd;
 345         p4d_t *p4d;
 346         pud_t *pud;
 347
 348         pgd = pgd_offset_k(vaddr);
 349         p4d = fill_p4d(pgd, vaddr);
 350         pud = fill_pud(p4d, vaddr);
 351         return fill_pmd(pud, vaddr);
 352 }
 353
 354 pte_t * __init populate_extra_pte(unsigned long vaddr)
 355 {
 356         pmd_t *pmd;
 357
 358         pmd = populate_extra_pmd(vaddr);
 359         return fill_pte(pmd, vaddr);
 360 }
 361
 362 /*
 363  * Create large page table mappings for a range of physical addresses.
 364  */
 365 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 366                                         enum page_cache_mode cache)
 367 {
 368         pgd_t *pgd;
 369         p4d_t *p4d;
 370         pud_t *pud;
 371         pmd_t *pmd;
 372         pgprot_t prot;
 373
 374         pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
 375                 protval_4k_2_large(cachemode2protval(cache));
 376         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 377         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 378                 pgd = pgd_offset_k((unsigned long)__va(phys));
 379                 if (pgd_none(*pgd)) {
 380                         p4d = (p4d_t *) spp_getpage();
 381                         set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE |
 382                                                 _PAGE_USER));
 383                 }
 384                 p4d = p4d_offset(pgd, (unsigned long)__va(phys));
 385                 if (p4d_none(*p4d)) {
 386                         pud = (pud_t *) spp_getpage();
 387                         set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE |
 388                                                 _PAGE_USER));
 389                 }
 390                 pud = pud_offset(p4d, (unsigned long)__va(phys));
 391                 if (pud_none(*pud)) {
 392                         pmd = (pmd_t *) spp_getpage();
 393                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 394                                                 _PAGE_USER));
 395                 }
 396                 pmd = pmd_offset(pud, phys);
 397                 BUG_ON(!pmd_none(*pmd));
 398                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 399         }
 400 }
 401
 402 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 403 {
 404         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
 405 }
 406
 407 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 408 {
 409         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
 410 }
 411
 412 /*
 413  * The head.S code sets up the kernel high mapping:
 414  *
 415  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 416  *
 417  * phys_base holds the negative offset to the kernel, which is added
 418  * to the compile time generated pmds. This results in invalid pmds up
 419  * to the point where we hit the physaddr 0 mapping.
 420  *
 421  * We limit the mappings to the region from _text to _brk_end.  _brk_end
 422  * is rounded up to the 2MB boundary. This catches the invalid pmds as
 423  * well, as they are located before _text:
 424  */
 425 void __init cleanup_highmap(void)
 426 {
 427         unsigned long vaddr = __START_KERNEL_map;
 428         unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
 429         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 430         pmd_t *pmd = level2_kernel_pgt;
 431
 432         /*
 433          * Native path, max_pfn_mapped is not set yet.
 434          * Xen has valid max_pfn_mapped set in
 435          *      arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
 436          */
 437         if (max_pfn_mapped)
 438                 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 439
 440         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
 441                 if (pmd_none(*pmd))
 442                         continue;
 443                 if (vaddr < (unsigned long) _text || vaddr > end)
 444                         set_pmd(pmd, __pmd(0));
 445         }
 446 }
 447
 448 /*
 449  * Create PTE level page table mapping for physical addresses.
 450  * It returns the last physical address mapped.
 451  */
 452 static unsigned long __meminit
 453 phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
 454               pgprot_t prot, bool init)
 455 {
 456         unsigned long pages = 0, paddr_next;
 457         unsigned long paddr_last = paddr_end;
 458         pte_t *pte;
 459         int i;
 460
 461         pte = pte_page + pte_index(paddr);
 462         i = pte_index(paddr);
 463
 464         for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
 465                 paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
 466                 if (paddr >= paddr_end) {
 467                         if (!after_bootmem &&
 468                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 469                                              E820_TYPE_RAM) &&
 470                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 471                                              E820_TYPE_RESERVED_KERN))
 472                                 set_pte_init(pte, __pte(0), init);
 473                         continue;
 474                 }
 475
 476                 /*
 477                  * We will re-use the existing mapping.
 478                  * Xen for example has some special requirements, like mapping
 479                  * pagetable pages as RO. So assume someone who pre-setup
 480                  * these mappings are more intelligent.
 481                  */
 482                 if (!pte_none(*pte)) {
 483                         if (!after_bootmem)
 484                                 pages++;
 485                         continue;
 486                 }
 487
 488                 if (0)
 489                         pr_info("   pte=%p addr=%lx pte=%016lx\n", pte, paddr,
 490                                 pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 491                 pages++;
 492                 set_pte_init(pte, pfn_pte(paddr >> PAGE_SHIFT, prot), init);
 493                 paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
 494         }
 495
 496         update_page_count(PG_LEVEL_4K, pages);
 497
 498         return paddr_last;
 499 }
 500
 501 /*
 502  * Create PMD level page table mapping for physical addresses. The virtual
 503  * and physical address have to be aligned at this level.
 504  * It returns the last physical address mapped.
 505  */
 506 static unsigned long __meminit
 507 phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
 508               unsigned long page_size_mask, pgprot_t prot, bool init)
 509 {
 510         unsigned long pages = 0, paddr_next;
 511         unsigned long paddr_last = paddr_end;
 512
 513         int i = pmd_index(paddr);
 514
 515         for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
 516                 pmd_t *pmd = pmd_page + pmd_index(paddr);
 517                 pte_t *pte;
 518                 pgprot_t new_prot = prot;
 519
 520                 paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
 521                 if (paddr >= paddr_end) {
 522                         if (!after_bootmem &&
 523                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 524                                              E820_TYPE_RAM) &&
 525                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 526                                              E820_TYPE_RESERVED_KERN))
 527                                 set_pmd_init(pmd, __pmd(0), init);
 528                         continue;
 529                 }
 530
 531                 if (!pmd_none(*pmd)) {
 532                         if (!pmd_large(*pmd)) {
 533                                 spin_lock(&init_mm.page_table_lock);
 534                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
 535                                 paddr_last = phys_pte_init(pte, paddr,
 536                                                            paddr_end, prot,
 537                                                            init);
 538                                 spin_unlock(&init_mm.page_table_lock);
 539                                 continue;
 540                         }
 541                         /*
 542                          * If we are ok with PG_LEVEL_2M mapping, then we will
 543                          * use the existing mapping,
 544                          *
 545                          * Otherwise, we will split the large page mapping but
 546                          * use the same existing protection bits except for
 547                          * large page, so that we don't violate Intel's TLB
 548                          * Application note (317080) which says, while changing
 549                          * the page sizes, new and old translations should
 550                          * not differ with respect to page frame and
 551                          * attributes.
 552                          */
 553                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 554                                 if (!after_bootmem)
 555                                         pages++;
 556                                 paddr_last = paddr_next;
 557                                 continue;
 558                         }
 559                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 560                 }
 561
 562                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 563                         pages++;
 564                         spin_lock(&init_mm.page_table_lock);
 565                         set_pte_init((pte_t *)pmd,
 566                                      pfn_pte((paddr & PMD_MASK) >> PAGE_SHIFT,
 567                                              __pgprot(pgprot_val(prot) | _PAGE_PSE)),
 568                                      init);
 569                         spin_unlock(&init_mm.page_table_lock);
 570                         paddr_last = paddr_next;
 571                         continue;
 572                 }
 573
 574                 pte = alloc_low_page();
 575                 paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot, init);
 576
 577                 spin_lock(&init_mm.page_table_lock);
 578                 pmd_populate_kernel_init(&init_mm, pmd, pte, init);
 579                 spin_unlock(&init_mm.page_table_lock);
 580         }
 581         update_page_count(PG_LEVEL_2M, pages);
 582         return paddr_last;
 583 }
 584
 585 /*
 586  * Create PUD level page table mapping for physical addresses. The virtual
 587  * and physical address do not have to be aligned at this level. KASLR can
 588  * randomize virtual addresses up to this level.
 589  * It returns the last physical address mapped.
 590  */
 591 static unsigned long __meminit
 592 phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
 593               unsigned long page_size_mask, pgprot_t _prot, bool init)
 594 {
 595         unsigned long pages = 0, paddr_next;
 596         unsigned long paddr_last = paddr_end;
 597         unsigned long vaddr = (unsigned long)__va(paddr);
 598         int i = pud_index(vaddr);
 599
 600         for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
 601                 pud_t *pud;
 602                 pmd_t *pmd;
 603                 pgprot_t prot = _prot;
 604
 605                 vaddr = (unsigned long)__va(paddr);
 606                 pud = pud_page + pud_index(vaddr);
 607                 paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
 608
 609                 if (paddr >= paddr_end) {
 610                         if (!after_bootmem &&
 611                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 612                                              E820_TYPE_RAM) &&
 613                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 614                                              E820_TYPE_RESERVED_KERN))
 615                                 set_pud_init(pud, __pud(0), init);
 616                         continue;
 617                 }
 618
 619                 if (!pud_none(*pud)) {
 620                         if (!pud_large(*pud)) {
 621                                 pmd = pmd_offset(pud, 0);
 622                                 paddr_last = phys_pmd_init(pmd, paddr,
 623                                                            paddr_end,
 624                                                            page_size_mask,
 625                                                            prot, init);
 626                                 continue;
 627                         }
 628                         /*
 629                          * If we are ok with PG_LEVEL_1G mapping, then we will
 630                          * use the existing mapping.
 631                          *
 632                          * Otherwise, we will split the gbpage mapping but use
 633                          * the same existing protection  bits except for large
 634                          * page, so that we don't violate Intel's TLB
 635                          * Application note (317080) which says, while changing
 636                          * the page sizes, new and old translations should
 637                          * not differ with respect to page frame and
 638                          * attributes.
 639                          */
 640                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 641                                 if (!after_bootmem)
 642                                         pages++;
 643                                 paddr_last = paddr_next;
 644                                 continue;
 645                         }
 646                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 647                 }
 648
 649                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 650                         pages++;
 651                         spin_lock(&init_mm.page_table_lock);
 652
 653                         prot = __pgprot(pgprot_val(prot) | __PAGE_KERNEL_LARGE);
 654
 655                         set_pte_init((pte_t *)pud,
 656                                      pfn_pte((paddr & PUD_MASK) >> PAGE_SHIFT,
 657                                              prot),
 658                                      init);
 659                         spin_unlock(&init_mm.page_table_lock);
 660                         paddr_last = paddr_next;
 661                         continue;
 662                 }
 663
 664                 pmd = alloc_low_page();
 665                 paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
 666                                            page_size_mask, prot, init);
 667
 668                 spin_lock(&init_mm.page_table_lock);
 669                 pud_populate_init(&init_mm, pud, pmd, init);
 670                 spin_unlock(&init_mm.page_table_lock);
 671         }
 672
 673         update_page_count(PG_LEVEL_1G, pages);
 674
 675         return paddr_last;
 676 }
 677
 678 static unsigned long __meminit
 679 phys_p4d_init(p4d_t *p4d_page, unsigned long paddr, unsigned long paddr_end,
 680               unsigned long page_size_mask, pgprot_t prot, bool init)
 681 {
 682         unsigned long vaddr, vaddr_end, vaddr_next, paddr_next, paddr_last;
 683
 684         paddr_last = paddr_end;
 685         vaddr = (unsigned long)__va(paddr);
 686         vaddr_end = (unsigned long)__va(paddr_end);
 687
 688         if (!pgtable_l5_enabled())
 689                 return phys_pud_init((pud_t *) p4d_page, paddr, paddr_end,
 690                                      page_size_mask, prot, init);
 691
 692         for (; vaddr < vaddr_end; vaddr = vaddr_next) {
 693                 p4d_t *p4d = p4d_page + p4d_index(vaddr);
 694                 pud_t *pud;
 695
 696                 vaddr_next = (vaddr & P4D_MASK) + P4D_SIZE;
 697                 paddr = __pa(vaddr);
 698
 699                 if (paddr >= paddr_end) {
 700                         paddr_next = __pa(vaddr_next);
 701                         if (!after_bootmem &&
 702                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 703                                              E820_TYPE_RAM) &&
 704                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 705                                              E820_TYPE_RESERVED_KERN))
 706                                 set_p4d_init(p4d, __p4d(0), init);
 707                         continue;
 708                 }
 709
 710                 if (!p4d_none(*p4d)) {
 711                         pud = pud_offset(p4d, 0);
 712                         paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
 713                                         page_size_mask, prot, init);
 714                         continue;
 715                 }
 716
 717                 pud = alloc_low_page();
 718                 paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
 719                                            page_size_mask, prot, init);
 720
 721                 spin_lock(&init_mm.page_table_lock);
 722                 p4d_populate_init(&init_mm, p4d, pud, init);
 723                 spin_unlock(&init_mm.page_table_lock);
 724         }
 725
 726         return paddr_last;
 727 }
 728
 729 static unsigned long __meminit
 730 __kernel_physical_mapping_init(unsigned long paddr_start,
 731                                unsigned long paddr_end,
 732                                unsigned long page_size_mask,
 733                                pgprot_t prot, bool init)
 734 {
 735         bool pgd_changed = false;
 736         unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
 737
 738         paddr_last = paddr_end;
 739         vaddr = (unsigned long)__va(paddr_start);
 740         vaddr_end = (unsigned long)__va(paddr_end);
 741         vaddr_start = vaddr;
 742
 743         for (; vaddr < vaddr_end; vaddr = vaddr_next) {
 744                 pgd_t *pgd = pgd_offset_k(vaddr);
 745                 p4d_t *p4d;
 746
 747                 vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
 748
 749                 if (pgd_val(*pgd)) {
 750                         p4d = (p4d_t *)pgd_page_vaddr(*pgd);
 751                         paddr_last = phys_p4d_init(p4d, __pa(vaddr),
 752                                                    __pa(vaddr_end),
 753                                                    page_size_mask,
 754                                                    prot, init);
 755                         continue;
 756                 }
 757
 758                 p4d = alloc_low_page();
 759                 paddr_last = phys_p4d_init(p4d, __pa(vaddr), __pa(vaddr_end),
 760                                            page_size_mask, prot, init);
 761
 762                 spin_lock(&init_mm.page_table_lock);
 763                 if (pgtable_l5_enabled())
 764                         pgd_populate_init(&init_mm, pgd, p4d, init);
 765                 else
 766                         p4d_populate_init(&init_mm, p4d_offset(pgd, vaddr),
 767                                           (pud_t *) p4d, init);
 768
 769                 spin_unlock(&init_mm.page_table_lock);
 770                 pgd_changed = true;
 771         }
 772
 773         if (pgd_changed)
 774                 sync_global_pgds(vaddr_start, vaddr_end - 1);
 775
 776         return paddr_last;
 777 }
 778
 779
 780 /*
 781  * Create page table mapping for the physical memory for specific physical
 782  * addresses. Note that it can only be used to populate non-present entries.
 783  * The virtual and physical addresses have to be aligned on PMD level
 784  * down. It returns the last physical address mapped.
 785  */
 786 unsigned long __meminit
 787 kernel_physical_mapping_init(unsigned long paddr_start,
 788                              unsigned long paddr_end,
 789                              unsigned long page_size_mask, pgprot_t prot)
 790 {
 791         return __kernel_physical_mapping_init(paddr_start, paddr_end,
 792                                               page_size_mask, prot, true);
 793 }
 794
 795 /*
 796  * This function is similar to kernel_physical_mapping_init() above with the
 797  * exception that it uses set_{pud,pmd}() instead of the set_{pud,pte}_safe()
 798  * when updating the mapping. The caller is responsible to flush the TLBs after
 799  * the function returns.
 800  */
 801 unsigned long __meminit
 802 kernel_physical_mapping_change(unsigned long paddr_start,
 803                                unsigned long paddr_end,
 804                                unsigned long page_size_mask)
 805 {
 806         return __kernel_physical_mapping_init(paddr_start, paddr_end,
 807                                               page_size_mask, PAGE_KERNEL,
 808                                               false);
 809 }
 810
 811 #ifndef CONFIG_NUMA
 812 void __init initmem_init(void)
 813 {
 814         memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
 815 }
 816 #endif
 817
 818 void __init paging_init(void)
 819 {
 820         sparse_init();
 821
 822         /*
 823          * clear the default setting with node 0
 824          * note: don't use nodes_clear here, that is really clearing when
 825          *       numa support is not compiled in, and later node_set_state
 826          *       will not set it back.
 827          */
 828         node_clear_state(0, N_MEMORY);
 829         node_clear_state(0, N_NORMAL_MEMORY);
 830
 831         zone_sizes_init();
 832 }
 833
 834 /*
 835  * Memory hotplug specific functions
 836  */
 837 #ifdef CONFIG_MEMORY_HOTPLUG
 838 /*
 839  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 840  * updating.
 841  */
 842 static void update_end_of_memory_vars(u64 start, u64 size)
 843 {
 844         unsigned long end_pfn = PFN_UP(start + size);
 845
 846         if (end_pfn > max_pfn) {
 847                 max_pfn = end_pfn;
 848                 max_low_pfn = end_pfn;
 849                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 850         }
 851 }
 852
 853 int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
 854               struct mhp_params *params)
 855 {
 856         int ret;
 857
 858         ret = __add_pages(nid, start_pfn, nr_pages, params);
 859         WARN_ON_ONCE(ret);
 860
 861         /* update max_pfn, max_low_pfn and high_memory */
 862         update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
 863                                   nr_pages << PAGE_SHIFT);
 864
 865         return ret;
 866 }
 867
 868 int arch_add_memory(int nid, u64 start, u64 size,
 869                     struct mhp_params *params)
 870 {
 871         unsigned long start_pfn = start >> PAGE_SHIFT;
 872         unsigned long nr_pages = size >> PAGE_SHIFT;
 873
 874         init_memory_mapping(start, start + size, params->pgprot);
 875
 876         return add_pages(nid, start_pfn, nr_pages, params);
 877 }
 878
 879 #define PAGE_INUSE 0xFD
 880
 881 static void __meminit free_pagetable(struct page *page, int order)
 882 {
 883         unsigned long magic;
 884         unsigned int nr_pages = 1 << order;
 885
 886         /* bootmem page has reserved flag */
 887         if (PageReserved(page)) {
 888                 __ClearPageReserved(page);
 889
 890                 magic = (unsigned long)page->freelist;
 891                 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
 892                         while (nr_pages--)
 893                                 put_page_bootmem(page++);
 894                 } else
 895                         while (nr_pages--)
 896                                 free_reserved_page(page++);
 897         } else
 898                 free_pages((unsigned long)page_address(page), order);
 899 }
 900
 901 static void __meminit free_hugepage_table(struct page *page,
 902                 struct vmem_altmap *altmap)
 903 {
 904         if (altmap)
 905                 vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
 906         else
 907                 free_pagetable(page, get_order(PMD_SIZE));
 908 }
 909
 910 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
 911 {
 912         pte_t *pte;
 913         int i;
 914
 915         for (i = 0; i < PTRS_PER_PTE; i++) {
 916                 pte = pte_start + i;
 917                 if (!pte_none(*pte))
 918                         return;
 919         }
 920
 921         /* free a pte talbe */
 922         free_pagetable(pmd_page(*pmd), 0);
 923         spin_lock(&init_mm.page_table_lock);
 924         pmd_clear(pmd);
 925         spin_unlock(&init_mm.page_table_lock);
 926 }
 927
 928 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
 929 {
 930         pmd_t *pmd;
 931         int i;
 932
 933         for (i = 0; i < PTRS_PER_PMD; i++) {
 934                 pmd = pmd_start + i;
 935                 if (!pmd_none(*pmd))
 936                         return;
 937         }
 938
 939         /* free a pmd talbe */
 940         free_pagetable(pud_page(*pud), 0);
 941         spin_lock(&init_mm.page_table_lock);
 942         pud_clear(pud);
 943         spin_unlock(&init_mm.page_table_lock);
 944 }
 945
 946 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
 947 {
 948         pud_t *pud;
 949         int i;
 950
 951         for (i = 0; i < PTRS_PER_PUD; i++) {
 952                 pud = pud_start + i;
 953                 if (!pud_none(*pud))
 954                         return;
 955         }
 956
 957         /* free a pud talbe */
 958         free_pagetable(p4d_page(*p4d), 0);
 959         spin_lock(&init_mm.page_table_lock);
 960         p4d_clear(p4d);
 961         spin_unlock(&init_mm.page_table_lock);
 962 }
 963
 964 static void __meminit
 965 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
 966                  bool direct)
 967 {
 968         unsigned long next, pages = 0;
 969         pte_t *pte;
 970         void *page_addr;
 971         phys_addr_t phys_addr;
 972
 973         pte = pte_start + pte_index(addr);
 974         for (; addr < end; addr = next, pte++) {
 975                 next = (addr + PAGE_SIZE) & PAGE_MASK;
 976                 if (next > end)
 977                         next = end;
 978
 979                 if (!pte_present(*pte))
 980                         continue;
 981
 982                 /*
 983                  * We mapped [0,1G) memory as identity mapping when
 984                  * initializing, in arch/x86/kernel/head_64.S. These
 985                  * pagetables cannot be removed.
 986                  */
 987                 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
 988                 if (phys_addr < (phys_addr_t)0x40000000)
 989                         return;
 990
 991                 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
 992                         /*
 993                          * Do not free direct mapping pages since they were
 994                          * freed when offlining, or simplely not in use.
 995                          */
 996                         if (!direct)
 997                                 free_pagetable(pte_page(*pte), 0);
 998
 999                         spin_lock(&init_mm.page_table_lock);
1000                         pte_clear(&init_mm, addr, pte);
1001                         spin_unlock(&init_mm.page_table_lock);
1002
1003                         /* For non-direct mapping, pages means nothing. */
1004                         pages++;
1005                 } else {
1006                         /*
1007                          * If we are here, we are freeing vmemmap pages since
1008                          * direct mapped memory ranges to be freed are aligned.
1009                          *
1010                          * If we are not removing the whole page, it means
1011                          * other page structs in this page are being used and
1012                          * we canot remove them. So fill the unused page_structs
1013                          * with 0xFD, and remove the page when it is wholly
1014                          * filled with 0xFD.
1015                          */
1016                         memset((void *)addr, PAGE_INUSE, next - addr);
1017
1018                         page_addr = page_address(pte_page(*pte));
1019                         if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
1020                                 free_pagetable(pte_page(*pte), 0);
1021
1022                                 spin_lock(&init_mm.page_table_lock);
1023                                 pte_clear(&init_mm, addr, pte);
1024                                 spin_unlock(&init_mm.page_table_lock);
1025                         }
1026                 }
1027         }
1028
1029         /* Call free_pte_table() in remove_pmd_table(). */
1030         flush_tlb_all();
1031         if (direct)
1032                 update_page_count(PG_LEVEL_4K, -pages);
1033 }
1034
1035 static void __meminit
1036 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
1037                  bool direct, struct vmem_altmap *altmap)
1038 {
1039         unsigned long next, pages = 0;
1040         pte_t *pte_base;
1041         pmd_t *pmd;
1042         void *page_addr;
1043
1044         pmd = pmd_start + pmd_index(addr);
1045         for (; addr < end; addr = next, pmd++) {
1046                 next = pmd_addr_end(addr, end);
1047
1048                 if (!pmd_present(*pmd))
1049                         continue;
1050
1051                 if (pmd_large(*pmd)) {
1052                         if (IS_ALIGNED(addr, PMD_SIZE) &&
1053                             IS_ALIGNED(next, PMD_SIZE)) {
1054                                 if (!direct)
1055                                         free_hugepage_table(pmd_page(*pmd),
1056                                                             altmap);
1057
1058                                 spin_lock(&init_mm.page_table_lock);
1059                                 pmd_clear(pmd);
1060                                 spin_unlock(&init_mm.page_table_lock);
1061                                 pages++;
1062                         } else {
1063                                 /* If here, we are freeing vmemmap pages. */
1064                                 memset((void *)addr, PAGE_INUSE, next - addr);
1065
1066                                 page_addr = page_address(pmd_page(*pmd));
1067                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1068                                                 PMD_SIZE)) {
1069                                         free_hugepage_table(pmd_page(*pmd),
1070                                                             altmap);
1071
1072                                         spin_lock(&init_mm.page_table_lock);
1073                                         pmd_clear(pmd);
1074                                         spin_unlock(&init_mm.page_table_lock);
1075                                 }
1076                         }
1077
1078                         continue;
1079                 }
1080
1081                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1082                 remove_pte_table(pte_base, addr, next, direct);
1083                 free_pte_table(pte_base, pmd);
1084         }
1085
1086         /* Call free_pmd_table() in remove_pud_table(). */
1087         if (direct)
1088                 update_page_count(PG_LEVEL_2M, -pages);
1089 }
1090
1091 static void __meminit
1092 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1093                  struct vmem_altmap *altmap, bool direct)
1094 {
1095         unsigned long next, pages = 0;
1096         pmd_t *pmd_base;
1097         pud_t *pud;
1098         void *page_addr;
1099
1100         pud = pud_start + pud_index(addr);
1101         for (; addr < end; addr = next, pud++) {
1102                 next = pud_addr_end(addr, end);
1103
1104                 if (!pud_present(*pud))
1105                         continue;
1106
1107                 if (pud_large(*pud)) {
1108                         if (IS_ALIGNED(addr, PUD_SIZE) &&
1109                             IS_ALIGNED(next, PUD_SIZE)) {
1110                                 if (!direct)
1111                                         free_pagetable(pud_page(*pud),
1112                                                        get_order(PUD_SIZE));
1113
1114                                 spin_lock(&init_mm.page_table_lock);
1115                                 pud_clear(pud);
1116                                 spin_unlock(&init_mm.page_table_lock);
1117                                 pages++;
1118                         } else {
1119                                 /* If here, we are freeing vmemmap pages. */
1120                                 memset((void *)addr, PAGE_INUSE, next - addr);
1121
1122                                 page_addr = page_address(pud_page(*pud));
1123                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1124                                                 PUD_SIZE)) {
1125                                         free_pagetable(pud_page(*pud),
1126                                                        get_order(PUD_SIZE));
1127
1128                                         spin_lock(&init_mm.page_table_lock);
1129                                         pud_clear(pud);
1130                                         spin_unlock(&init_mm.page_table_lock);
1131                                 }
1132                         }
1133
1134                         continue;
1135                 }
1136
1137                 pmd_base = pmd_offset(pud, 0);
1138                 remove_pmd_table(pmd_base, addr, next, direct, altmap);
1139                 free_pmd_table(pmd_base, pud);
1140         }
1141
1142         if (direct)
1143                 update_page_count(PG_LEVEL_1G, -pages);
1144 }
1145
1146 static void __meminit
1147 remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1148                  struct vmem_altmap *altmap, bool direct)
1149 {
1150         unsigned long next, pages = 0;
1151         pud_t *pud_base;
1152         p4d_t *p4d;
1153
1154         p4d = p4d_start + p4d_index(addr);
1155         for (; addr < end; addr = next, p4d++) {
1156                 next = p4d_addr_end(addr, end);
1157
1158                 if (!p4d_present(*p4d))
1159                         continue;
1160
1161                 BUILD_BUG_ON(p4d_large(*p4d));
1162
1163                 pud_base = pud_offset(p4d, 0);
1164                 remove_pud_table(pud_base, addr, next, altmap, direct);
1165                 /*
1166                  * For 4-level page tables we do not want to free PUDs, but in the
1167                  * 5-level case we should free them. This code will have to change
1168                  * to adapt for boot-time switching between 4 and 5 level page tables.
1169                  */
1170                 if (pgtable_l5_enabled())
1171                         free_pud_table(pud_base, p4d);
1172         }
1173
1174         if (direct)
1175                 update_page_count(PG_LEVEL_512G, -pages);
1176 }
1177
1178 /* start and end are both virtual address. */
1179 static void __meminit
1180 remove_pagetable(unsigned long start, unsigned long end, bool direct,
1181                 struct vmem_altmap *altmap)
1182 {
1183         unsigned long next;
1184         unsigned long addr;
1185         pgd_t *pgd;
1186         p4d_t *p4d;
1187
1188         for (addr = start; addr < end; addr = next) {
1189                 next = pgd_addr_end(addr, end);
1190
1191                 pgd = pgd_offset_k(addr);
1192                 if (!pgd_present(*pgd))
1193                         continue;
1194
1195                 p4d = p4d_offset(pgd, 0);
1196                 remove_p4d_table(p4d, addr, next, altmap, direct);
1197         }
1198
1199         flush_tlb_all();
1200 }
1201
1202 void __ref vmemmap_free(unsigned long start, unsigned long end,
1203                 struct vmem_altmap *altmap)
1204 {
1205         remove_pagetable(start, end, false, altmap);
1206 }
1207
1208 static void __meminit
1209 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1210 {
1211         start = (unsigned long)__va(start);
1212         end = (unsigned long)__va(end);
1213
1214         remove_pagetable(start, end, true, NULL);
1215 }
1216
1217 void __ref arch_remove_memory(int nid, u64 start, u64 size,
1218                               struct vmem_altmap *altmap)
1219 {
1220         unsigned long start_pfn = start >> PAGE_SHIFT;
1221         unsigned long nr_pages = size >> PAGE_SHIFT;
1222
1223         __remove_pages(start_pfn, nr_pages, altmap);
1224         kernel_physical_mapping_remove(start, start + size);
1225 }
1226 #endif /* CONFIG_MEMORY_HOTPLUG */
1227
1228 static struct kcore_list kcore_vsyscall;
1229
1230 static void __init register_page_bootmem_info(void)
1231 {
1232 #ifdef CONFIG_NUMA
1233         int i;
1234
1235         for_each_online_node(i)
1236                 register_page_bootmem_info_node(NODE_DATA(i));
1237 #endif
1238 }
1239
1240 /*
1241  * Pre-allocates page-table pages for the vmalloc area in the kernel page-table.
1242  * Only the level which needs to be synchronized between all page-tables is
1243  * allocated because the synchronization can be expensive.
1244  */
1245 static void __init preallocate_vmalloc_pages(void)
1246 {
1247         unsigned long addr;
1248         const char *lvl;
1249
1250         for (addr = VMALLOC_START; addr <= VMALLOC_END; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
1251                 pgd_t *pgd = pgd_offset_k(addr);
1252                 p4d_t *p4d;
1253                 pud_t *pud;
1254
1255                 lvl = "p4d";
1256                 p4d = p4d_alloc(&init_mm, pgd, addr);
1257                 if (!p4d)
1258                         goto failed;
1259
1260                 /*
1261                  * With 5-level paging the P4D level is not folded. So the PGDs
1262                  * are now populated and there is no need to walk down to the
1263                  * PUD level.
1264                  */
1265                 if (pgtable_l5_enabled())
1266                         continue;
1267
1268                 lvl = "pud";
1269                 pud = pud_alloc(&init_mm, p4d, addr);
1270                 if (!pud)
1271                         goto failed;
1272         }
1273
1274         return;
1275
1276 failed:
1277
1278         /*
1279          * The pages have to be there now or they will be missing in
1280          * process page-tables later.
1281          */
1282         panic("Failed to pre-allocate %s pages for vmalloc area\n", lvl);
1283 }
1284
1285 void __init mem_init(void)
1286 {
1287         pci_iommu_alloc();
1288
1289         /* clear_bss() already clear the empty_zero_page */
1290
1291         /* this will put all memory onto the freelists */
1292         memblock_free_all();
1293         after_bootmem = 1;
1294         x86_init.hyper.init_after_bootmem();
1295
1296         /*
1297          * Must be done after boot memory is put on freelist, because here we
1298          * might set fields in deferred struct pages that have not yet been
1299          * initialized, and memblock_free_all() initializes all the reserved
1300          * deferred pages for us.
1301          */
1302         register_page_bootmem_info();
1303
1304         /* Register memory areas for /proc/kcore */
1305         if (get_gate_vma(&init_mm))
1306                 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1307
1308         preallocate_vmalloc_pages();
1309
1310         mem_init_print_info(NULL);
1311 }
1312
1313 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1314 int __init deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1315 {
1316         /*
1317          * More CPUs always led to greater speedups on tested systems, up to
1318          * all the nodes' CPUs.  Use all since the system is otherwise idle
1319          * now.
1320          */
1321         return max_t(int, cpumask_weight(node_cpumask), 1);
1322 }
1323 #endif
1324
1325 int kernel_set_to_readonly;
1326
1327 void mark_rodata_ro(void)
1328 {
1329         unsigned long start = PFN_ALIGN(_text);
1330         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1331         unsigned long end = (unsigned long)__end_rodata_hpage_align;
1332         unsigned long text_end = PFN_ALIGN(_etext);
1333         unsigned long rodata_end = PFN_ALIGN(__end_rodata);
1334         unsigned long all_end;
1335
1336         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1337                (end - start) >> 10);
1338         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1339
1340         kernel_set_to_readonly = 1;
1341
1342         /*
1343          * The rodata/data/bss/brk section (but not the kernel text!)
1344          * should also be not-executable.
1345          *
1346          * We align all_end to PMD_SIZE because the existing mapping
1347          * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1348          * split the PMD and the reminder between _brk_end and the end
1349          * of the PMD will remain mapped executable.
1350          *
1351          * Any PMD which was setup after the one which covers _brk_end
1352          * has been zapped already via cleanup_highmem().
1353          */
1354         all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1355         set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1356
1357         set_ftrace_ops_ro();
1358
1359 #ifdef CONFIG_CPA_DEBUG
1360         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1361         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1362
1363         printk(KERN_INFO "Testing CPA: again\n");
1364         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1365 #endif
1366
1367         free_kernel_image_pages("unused kernel image (text/rodata gap)",
1368                                 (void *)text_end, (void *)rodata_start);
1369         free_kernel_image_pages("unused kernel image (rodata/data gap)",
1370                                 (void *)rodata_end, (void *)_sdata);
1371
1372         debug_checkwx();
1373 }
1374
1375 int kern_addr_valid(unsigned long addr)
1376 {
1377         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1378         pgd_t *pgd;
1379         p4d_t *p4d;
1380         pud_t *pud;
1381         pmd_t *pmd;
1382         pte_t *pte;
1383
1384         if (above != 0 && above != -1UL)
1385                 return 0;
1386
1387         pgd = pgd_offset_k(addr);
1388         if (pgd_none(*pgd))
1389                 return 0;
1390
1391         p4d = p4d_offset(pgd, addr);
1392         if (p4d_none(*p4d))
1393                 return 0;
1394
1395         pud = pud_offset(p4d, addr);
1396         if (pud_none(*pud))
1397                 return 0;
1398
1399         if (pud_large(*pud))
1400                 return pfn_valid(pud_pfn(*pud));
1401
1402         pmd = pmd_offset(pud, addr);
1403         if (pmd_none(*pmd))
1404                 return 0;
1405
1406         if (pmd_large(*pmd))
1407                 return pfn_valid(pmd_pfn(*pmd));
1408
1409         pte = pte_offset_kernel(pmd, addr);
1410         if (pte_none(*pte))
1411                 return 0;
1412
1413         return pfn_valid(pte_pfn(*pte));
1414 }
1415
1416 /*
1417  * Block size is the minimum amount of memory which can be hotplugged or
1418  * hotremoved. It must be power of two and must be equal or larger than
1419  * MIN_MEMORY_BLOCK_SIZE.
1420  */
1421 #define MAX_BLOCK_SIZE (2UL << 30)
1422
1423 /* Amount of ram needed to start using large blocks */
1424 #define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1425
1426 /* Adjustable memory block size */
1427 static unsigned long set_memory_block_size;
1428 int __init set_memory_block_size_order(unsigned int order)
1429 {
1430         unsigned long size = 1UL << order;
1431
1432         if (size > MEM_SIZE_FOR_LARGE_BLOCK || size < MIN_MEMORY_BLOCK_SIZE)
1433                 return -EINVAL;
1434
1435         set_memory_block_size = size;
1436         return 0;
1437 }
1438
1439 static unsigned long probe_memory_block_size(void)
1440 {
1441         unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1442         unsigned long bz;
1443
1444         /* If memory block size has been set, then use it */
1445         bz = set_memory_block_size;
1446         if (bz)
1447                 goto done;
1448
1449         /* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1450         if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1451                 bz = MIN_MEMORY_BLOCK_SIZE;
1452                 goto done;
1453         }
1454
1455         /* Find the largest allowed block size that aligns to memory end */
1456         for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1457                 if (IS_ALIGNED(boot_mem_end, bz))
1458                         break;
1459         }
1460 done:
1461         pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1462
1463         return bz;
1464 }
1465
1466 static unsigned long memory_block_size_probed;
1467 unsigned long memory_block_size_bytes(void)
1468 {
1469         if (!memory_block_size_probed)
1470                 memory_block_size_probed = probe_memory_block_size();
1471
1472         return memory_block_size_probed;
1473 }
1474
1475 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1476 /*
1477  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1478  */
1479 static long __meminitdata addr_start, addr_end;
1480 static void __meminitdata *p_start, *p_end;
1481 static int __meminitdata node_start;
1482
1483 static int __meminit vmemmap_populate_hugepages(unsigned long start,
1484                 unsigned long end, int node, struct vmem_altmap *altmap)
1485 {
1486         unsigned long addr;
1487         unsigned long next;
1488         pgd_t *pgd;
1489         p4d_t *p4d;
1490         pud_t *pud;
1491         pmd_t *pmd;
1492
1493         for (addr = start; addr < end; addr = next) {
1494                 next = pmd_addr_end(addr, end);
1495
1496                 pgd = vmemmap_pgd_populate(addr, node);
1497                 if (!pgd)
1498                         return -ENOMEM;
1499
1500                 p4d = vmemmap_p4d_populate(pgd, addr, node);
1501                 if (!p4d)
1502                         return -ENOMEM;
1503
1504                 pud = vmemmap_pud_populate(p4d, addr, node);
1505                 if (!pud)
1506                         return -ENOMEM;
1507
1508                 pmd = pmd_offset(pud, addr);
1509                 if (pmd_none(*pmd)) {
1510                         void *p;
1511
1512                         p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
1513                         if (p) {
1514                                 pte_t entry;
1515
1516                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1517                                                 PAGE_KERNEL_LARGE);
1518                                 set_pmd(pmd, __pmd(pte_val(entry)));
1519
1520                                 /* check to see if we have contiguous blocks */
1521                                 if (p_end != p || node_start != node) {
1522                                         if (p_start)
1523                                                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1524                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1525                                         addr_start = addr;
1526                                         node_start = node;
1527                                         p_start = p;
1528                                 }
1529
1530                                 addr_end = addr + PMD_SIZE;
1531                                 p_end = p + PMD_SIZE;
1532                                 continue;
1533                         } else if (altmap)
1534                                 return -ENOMEM; /* no fallback */
1535                 } else if (pmd_large(*pmd)) {
1536                         vmemmap_verify((pte_t *)pmd, node, addr, next);
1537                         continue;
1538                 }
1539                 if (vmemmap_populate_basepages(addr, next, node, NULL))
1540                         return -ENOMEM;
1541         }
1542         return 0;
1543 }
1544
1545 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1546                 struct vmem_altmap *altmap)
1547 {
1548         int err;
1549
1550         if (end - start < PAGES_PER_SECTION * sizeof(struct page))
1551                 err = vmemmap_populate_basepages(start, end, node, NULL);
1552         else if (boot_cpu_has(X86_FEATURE_PSE))
1553                 err = vmemmap_populate_hugepages(start, end, node, altmap);
1554         else if (altmap) {
1555                 pr_err_once("%s: no cpu support for altmap allocations\n",
1556                                 __func__);
1557                 err = -ENOMEM;
1558         } else
1559                 err = vmemmap_populate_basepages(start, end, node, NULL);
1560         if (!err)
1561                 sync_global_pgds(start, end - 1);
1562         return err;
1563 }
1564
1565 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1566 void register_page_bootmem_memmap(unsigned long section_nr,
1567                                   struct page *start_page, unsigned long nr_pages)
1568 {
1569         unsigned long addr = (unsigned long)start_page;
1570         unsigned long end = (unsigned long)(start_page + nr_pages);
1571         unsigned long next;
1572         pgd_t *pgd;
1573         p4d_t *p4d;
1574         pud_t *pud;
1575         pmd_t *pmd;
1576         unsigned int nr_pmd_pages;
1577         struct page *page;
1578
1579         for (; addr < end; addr = next) {
1580                 pte_t *pte = NULL;
1581
1582                 pgd = pgd_offset_k(addr);
1583                 if (pgd_none(*pgd)) {
1584                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1585                         continue;
1586                 }
1587                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1588
1589                 p4d = p4d_offset(pgd, addr);
1590                 if (p4d_none(*p4d)) {
1591                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1592                         continue;
1593                 }
1594                 get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1595
1596                 pud = pud_offset(p4d, addr);
1597                 if (pud_none(*pud)) {
1598                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1599                         continue;
1600                 }
1601                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1602
1603                 if (!boot_cpu_has(X86_FEATURE_PSE)) {
1604                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1605                         pmd = pmd_offset(pud, addr);
1606                         if (pmd_none(*pmd))
1607                                 continue;
1608                         get_page_bootmem(section_nr, pmd_page(*pmd),
1609                                          MIX_SECTION_INFO);
1610
1611                         pte = pte_offset_kernel(pmd, addr);
1612                         if (pte_none(*pte))
1613                                 continue;
1614                         get_page_bootmem(section_nr, pte_page(*pte),
1615                                          SECTION_INFO);
1616                 } else {
1617                         next = pmd_addr_end(addr, end);
1618
1619                         pmd = pmd_offset(pud, addr);
1620                         if (pmd_none(*pmd))
1621                                 continue;
1622
1623                         nr_pmd_pages = 1 << get_order(PMD_SIZE);
1624                         page = pmd_page(*pmd);
1625                         while (nr_pmd_pages--)
1626                                 get_page_bootmem(section_nr, page++,
1627                                                  SECTION_INFO);
1628                 }
1629         }
1630 }
1631 #endif
1632
1633 void __meminit vmemmap_populate_print_last(void)
1634 {
1635         if (p_start) {
1636                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1637                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1638                 p_start = NULL;
1639                 p_end = NULL;
1640                 node_start = 0;
1641         }
1642 }
1643 #endif