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 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_memory_present_with_active_regions(MAX_NUMNODES);
 821         sparse_init();
 822
 823         /*
 824          * clear the default setting with node 0
 825          * note: don't use nodes_clear here, that is really clearing when
 826          *       numa support is not compiled in, and later node_set_state
 827          *       will not set it back.
 828          */
 829         node_clear_state(0, N_MEMORY);
 830         node_clear_state(0, N_NORMAL_MEMORY);
 831
 832         zone_sizes_init();
 833 }
 834
 835 /*
 836  * Memory hotplug specific functions
 837  */
 838 #ifdef CONFIG_MEMORY_HOTPLUG
 839 /*
 840  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 841  * updating.
 842  */
 843 static void update_end_of_memory_vars(u64 start, u64 size)
 844 {
 845         unsigned long end_pfn = PFN_UP(start + size);
 846
 847         if (end_pfn > max_pfn) {
 848                 max_pfn = end_pfn;
 849                 max_low_pfn = end_pfn;
 850                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 851         }
 852 }
 853
 854 int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
 855               struct mhp_params *params)
 856 {
 857         int ret;
 858
 859         ret = __add_pages(nid, start_pfn, nr_pages, params);
 860         WARN_ON_ONCE(ret);
 861
 862         /* update max_pfn, max_low_pfn and high_memory */
 863         update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
 864                                   nr_pages << PAGE_SHIFT);
 865
 866         return ret;
 867 }
 868
 869 int arch_add_memory(int nid, u64 start, u64 size,
 870                     struct mhp_params *params)
 871 {
 872         unsigned long start_pfn = start >> PAGE_SHIFT;
 873         unsigned long nr_pages = size >> PAGE_SHIFT;
 874
 875         init_memory_mapping(start, start + size, params->pgprot);
 876
 877         return add_pages(nid, start_pfn, nr_pages, params);
 878 }
 879
 880 #define PAGE_INUSE 0xFD
 881
 882 static void __meminit free_pagetable(struct page *page, int order)
 883 {
 884         unsigned long magic;
 885         unsigned int nr_pages = 1 << order;
 886
 887         /* bootmem page has reserved flag */
 888         if (PageReserved(page)) {
 889                 __ClearPageReserved(page);
 890
 891                 magic = (unsigned long)page->freelist;
 892                 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
 893                         while (nr_pages--)
 894                                 put_page_bootmem(page++);
 895                 } else
 896                         while (nr_pages--)
 897                                 free_reserved_page(page++);
 898         } else
 899                 free_pages((unsigned long)page_address(page), order);
 900 }
 901
 902 static void __meminit free_hugepage_table(struct page *page,
 903                 struct vmem_altmap *altmap)
 904 {
 905         if (altmap)
 906                 vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
 907         else
 908                 free_pagetable(page, get_order(PMD_SIZE));
 909 }
 910
 911 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
 912 {
 913         pte_t *pte;
 914         int i;
 915
 916         for (i = 0; i < PTRS_PER_PTE; i++) {
 917                 pte = pte_start + i;
 918                 if (!pte_none(*pte))
 919                         return;
 920         }
 921
 922         /* free a pte talbe */
 923         free_pagetable(pmd_page(*pmd), 0);
 924         spin_lock(&init_mm.page_table_lock);
 925         pmd_clear(pmd);
 926         spin_unlock(&init_mm.page_table_lock);
 927 }
 928
 929 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
 930 {
 931         pmd_t *pmd;
 932         int i;
 933
 934         for (i = 0; i < PTRS_PER_PMD; i++) {
 935                 pmd = pmd_start + i;
 936                 if (!pmd_none(*pmd))
 937                         return;
 938         }
 939
 940         /* free a pmd talbe */
 941         free_pagetable(pud_page(*pud), 0);
 942         spin_lock(&init_mm.page_table_lock);
 943         pud_clear(pud);
 944         spin_unlock(&init_mm.page_table_lock);
 945 }
 946
 947 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
 948 {
 949         pud_t *pud;
 950         int i;
 951
 952         for (i = 0; i < PTRS_PER_PUD; i++) {
 953                 pud = pud_start + i;
 954                 if (!pud_none(*pud))
 955                         return;
 956         }
 957
 958         /* free a pud talbe */
 959         free_pagetable(p4d_page(*p4d), 0);
 960         spin_lock(&init_mm.page_table_lock);
 961         p4d_clear(p4d);
 962         spin_unlock(&init_mm.page_table_lock);
 963 }
 964
 965 static void __meminit
 966 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
 967                  bool direct)
 968 {
 969         unsigned long next, pages = 0;
 970         pte_t *pte;
 971         void *page_addr;
 972         phys_addr_t phys_addr;
 973
 974         pte = pte_start + pte_index(addr);
 975         for (; addr < end; addr = next, pte++) {
 976                 next = (addr + PAGE_SIZE) & PAGE_MASK;
 977                 if (next > end)
 978                         next = end;
 979
 980                 if (!pte_present(*pte))
 981                         continue;
 982
 983                 /*
 984                  * We mapped [0,1G) memory as identity mapping when
 985                  * initializing, in arch/x86/kernel/head_64.S. These
 986                  * pagetables cannot be removed.
 987                  */
 988                 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
 989                 if (phys_addr < (phys_addr_t)0x40000000)
 990                         return;
 991
 992                 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
 993                         /*
 994                          * Do not free direct mapping pages since they were
 995                          * freed when offlining, or simplely not in use.
 996                          */
 997                         if (!direct)
 998                                 free_pagetable(pte_page(*pte), 0);
 999
1000                         spin_lock(&init_mm.page_table_lock);
1001                         pte_clear(&init_mm, addr, pte);
1002                         spin_unlock(&init_mm.page_table_lock);
1003
1004                         /* For non-direct mapping, pages means nothing. */
1005                         pages++;
1006                 } else {
1007                         /*
1008                          * If we are here, we are freeing vmemmap pages since
1009                          * direct mapped memory ranges to be freed are aligned.
1010                          *
1011                          * If we are not removing the whole page, it means
1012                          * other page structs in this page are being used and
1013                          * we canot remove them. So fill the unused page_structs
1014                          * with 0xFD, and remove the page when it is wholly
1015                          * filled with 0xFD.
1016                          */
1017                         memset((void *)addr, PAGE_INUSE, next - addr);
1018
1019                         page_addr = page_address(pte_page(*pte));
1020                         if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
1021                                 free_pagetable(pte_page(*pte), 0);
1022
1023                                 spin_lock(&init_mm.page_table_lock);
1024                                 pte_clear(&init_mm, addr, pte);
1025                                 spin_unlock(&init_mm.page_table_lock);
1026                         }
1027                 }
1028         }
1029
1030         /* Call free_pte_table() in remove_pmd_table(). */
1031         flush_tlb_all();
1032         if (direct)
1033                 update_page_count(PG_LEVEL_4K, -pages);
1034 }
1035
1036 static void __meminit
1037 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
1038                  bool direct, struct vmem_altmap *altmap)
1039 {
1040         unsigned long next, pages = 0;
1041         pte_t *pte_base;
1042         pmd_t *pmd;
1043         void *page_addr;
1044
1045         pmd = pmd_start + pmd_index(addr);
1046         for (; addr < end; addr = next, pmd++) {
1047                 next = pmd_addr_end(addr, end);
1048
1049                 if (!pmd_present(*pmd))
1050                         continue;
1051
1052                 if (pmd_large(*pmd)) {
1053                         if (IS_ALIGNED(addr, PMD_SIZE) &&
1054                             IS_ALIGNED(next, PMD_SIZE)) {
1055                                 if (!direct)
1056                                         free_hugepage_table(pmd_page(*pmd),
1057                                                             altmap);
1058
1059                                 spin_lock(&init_mm.page_table_lock);
1060                                 pmd_clear(pmd);
1061                                 spin_unlock(&init_mm.page_table_lock);
1062                                 pages++;
1063                         } else {
1064                                 /* If here, we are freeing vmemmap pages. */
1065                                 memset((void *)addr, PAGE_INUSE, next - addr);
1066
1067                                 page_addr = page_address(pmd_page(*pmd));
1068                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1069                                                 PMD_SIZE)) {
1070                                         free_hugepage_table(pmd_page(*pmd),
1071                                                             altmap);
1072
1073                                         spin_lock(&init_mm.page_table_lock);
1074                                         pmd_clear(pmd);
1075                                         spin_unlock(&init_mm.page_table_lock);
1076                                 }
1077                         }
1078
1079                         continue;
1080                 }
1081
1082                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1083                 remove_pte_table(pte_base, addr, next, direct);
1084                 free_pte_table(pte_base, pmd);
1085         }
1086
1087         /* Call free_pmd_table() in remove_pud_table(). */
1088         if (direct)
1089                 update_page_count(PG_LEVEL_2M, -pages);
1090 }
1091
1092 static void __meminit
1093 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1094                  struct vmem_altmap *altmap, bool direct)
1095 {
1096         unsigned long next, pages = 0;
1097         pmd_t *pmd_base;
1098         pud_t *pud;
1099         void *page_addr;
1100
1101         pud = pud_start + pud_index(addr);
1102         for (; addr < end; addr = next, pud++) {
1103                 next = pud_addr_end(addr, end);
1104
1105                 if (!pud_present(*pud))
1106                         continue;
1107
1108                 if (pud_large(*pud)) {
1109                         if (IS_ALIGNED(addr, PUD_SIZE) &&
1110                             IS_ALIGNED(next, PUD_SIZE)) {
1111                                 if (!direct)
1112                                         free_pagetable(pud_page(*pud),
1113                                                        get_order(PUD_SIZE));
1114
1115                                 spin_lock(&init_mm.page_table_lock);
1116                                 pud_clear(pud);
1117                                 spin_unlock(&init_mm.page_table_lock);
1118                                 pages++;
1119                         } else {
1120                                 /* If here, we are freeing vmemmap pages. */
1121                                 memset((void *)addr, PAGE_INUSE, next - addr);
1122
1123                                 page_addr = page_address(pud_page(*pud));
1124                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1125                                                 PUD_SIZE)) {
1126                                         free_pagetable(pud_page(*pud),
1127                                                        get_order(PUD_SIZE));
1128
1129                                         spin_lock(&init_mm.page_table_lock);
1130                                         pud_clear(pud);
1131                                         spin_unlock(&init_mm.page_table_lock);
1132                                 }
1133                         }
1134
1135                         continue;
1136                 }
1137
1138                 pmd_base = pmd_offset(pud, 0);
1139                 remove_pmd_table(pmd_base, addr, next, direct, altmap);
1140                 free_pmd_table(pmd_base, pud);
1141         }
1142
1143         if (direct)
1144                 update_page_count(PG_LEVEL_1G, -pages);
1145 }
1146
1147 static void __meminit
1148 remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1149                  struct vmem_altmap *altmap, bool direct)
1150 {
1151         unsigned long next, pages = 0;
1152         pud_t *pud_base;
1153         p4d_t *p4d;
1154
1155         p4d = p4d_start + p4d_index(addr);
1156         for (; addr < end; addr = next, p4d++) {
1157                 next = p4d_addr_end(addr, end);
1158
1159                 if (!p4d_present(*p4d))
1160                         continue;
1161
1162                 BUILD_BUG_ON(p4d_large(*p4d));
1163
1164                 pud_base = pud_offset(p4d, 0);
1165                 remove_pud_table(pud_base, addr, next, altmap, direct);
1166                 /*
1167                  * For 4-level page tables we do not want to free PUDs, but in the
1168                  * 5-level case we should free them. This code will have to change
1169                  * to adapt for boot-time switching between 4 and 5 level page tables.
1170                  */
1171                 if (pgtable_l5_enabled())
1172                         free_pud_table(pud_base, p4d);
1173         }
1174
1175         if (direct)
1176                 update_page_count(PG_LEVEL_512G, -pages);
1177 }
1178
1179 /* start and end are both virtual address. */
1180 static void __meminit
1181 remove_pagetable(unsigned long start, unsigned long end, bool direct,
1182                 struct vmem_altmap *altmap)
1183 {
1184         unsigned long next;
1185         unsigned long addr;
1186         pgd_t *pgd;
1187         p4d_t *p4d;
1188
1189         for (addr = start; addr < end; addr = next) {
1190                 next = pgd_addr_end(addr, end);
1191
1192                 pgd = pgd_offset_k(addr);
1193                 if (!pgd_present(*pgd))
1194                         continue;
1195
1196                 p4d = p4d_offset(pgd, 0);
1197                 remove_p4d_table(p4d, addr, next, altmap, direct);
1198         }
1199
1200         flush_tlb_all();
1201 }
1202
1203 void __ref vmemmap_free(unsigned long start, unsigned long end,
1204                 struct vmem_altmap *altmap)
1205 {
1206         remove_pagetable(start, end, false, altmap);
1207 }
1208
1209 static void __meminit
1210 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1211 {
1212         start = (unsigned long)__va(start);
1213         end = (unsigned long)__va(end);
1214
1215         remove_pagetable(start, end, true, NULL);
1216 }
1217
1218 void __ref arch_remove_memory(int nid, u64 start, u64 size,
1219                               struct vmem_altmap *altmap)
1220 {
1221         unsigned long start_pfn = start >> PAGE_SHIFT;
1222         unsigned long nr_pages = size >> PAGE_SHIFT;
1223
1224         __remove_pages(start_pfn, nr_pages, altmap);
1225         kernel_physical_mapping_remove(start, start + size);
1226 }
1227 #endif /* CONFIG_MEMORY_HOTPLUG */
1228
1229 static struct kcore_list kcore_vsyscall;
1230
1231 static void __init register_page_bootmem_info(void)
1232 {
1233 #ifdef CONFIG_NUMA
1234         int i;
1235
1236         for_each_online_node(i)
1237                 register_page_bootmem_info_node(NODE_DATA(i));
1238 #endif
1239 }
1240
1241 void __init mem_init(void)
1242 {
1243         pci_iommu_alloc();
1244
1245         /* clear_bss() already clear the empty_zero_page */
1246
1247         /* this will put all memory onto the freelists */
1248         memblock_free_all();
1249         after_bootmem = 1;
1250         x86_init.hyper.init_after_bootmem();
1251
1252         /*
1253          * Must be done after boot memory is put on freelist, because here we
1254          * might set fields in deferred struct pages that have not yet been
1255          * initialized, and memblock_free_all() initializes all the reserved
1256          * deferred pages for us.
1257          */
1258         register_page_bootmem_info();
1259
1260         /* Register memory areas for /proc/kcore */
1261         if (get_gate_vma(&init_mm))
1262                 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1263
1264         mem_init_print_info(NULL);
1265 }
1266
1267 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1268 int __init deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1269 {
1270         /*
1271          * More CPUs always led to greater speedups on tested systems, up to
1272          * all the nodes' CPUs.  Use all since the system is otherwise idle
1273          * now.
1274          */
1275         return max_t(int, cpumask_weight(node_cpumask), 1);
1276 }
1277 #endif
1278
1279 int kernel_set_to_readonly;
1280
1281 void mark_rodata_ro(void)
1282 {
1283         unsigned long start = PFN_ALIGN(_text);
1284         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1285         unsigned long end = (unsigned long)__end_rodata_hpage_align;
1286         unsigned long text_end = PFN_ALIGN(_etext);
1287         unsigned long rodata_end = PFN_ALIGN(__end_rodata);
1288         unsigned long all_end;
1289
1290         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1291                (end - start) >> 10);
1292         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1293
1294         kernel_set_to_readonly = 1;
1295
1296         /*
1297          * The rodata/data/bss/brk section (but not the kernel text!)
1298          * should also be not-executable.
1299          *
1300          * We align all_end to PMD_SIZE because the existing mapping
1301          * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1302          * split the PMD and the reminder between _brk_end and the end
1303          * of the PMD will remain mapped executable.
1304          *
1305          * Any PMD which was setup after the one which covers _brk_end
1306          * has been zapped already via cleanup_highmem().
1307          */
1308         all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1309         set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1310
1311         set_ftrace_ops_ro();
1312
1313 #ifdef CONFIG_CPA_DEBUG
1314         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1315         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1316
1317         printk(KERN_INFO "Testing CPA: again\n");
1318         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1319 #endif
1320
1321         free_kernel_image_pages("unused kernel image (text/rodata gap)",
1322                                 (void *)text_end, (void *)rodata_start);
1323         free_kernel_image_pages("unused kernel image (rodata/data gap)",
1324                                 (void *)rodata_end, (void *)_sdata);
1325
1326         debug_checkwx();
1327 }
1328
1329 int kern_addr_valid(unsigned long addr)
1330 {
1331         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1332         pgd_t *pgd;
1333         p4d_t *p4d;
1334         pud_t *pud;
1335         pmd_t *pmd;
1336         pte_t *pte;
1337
1338         if (above != 0 && above != -1UL)
1339                 return 0;
1340
1341         pgd = pgd_offset_k(addr);
1342         if (pgd_none(*pgd))
1343                 return 0;
1344
1345         p4d = p4d_offset(pgd, addr);
1346         if (p4d_none(*p4d))
1347                 return 0;
1348
1349         pud = pud_offset(p4d, addr);
1350         if (pud_none(*pud))
1351                 return 0;
1352
1353         if (pud_large(*pud))
1354                 return pfn_valid(pud_pfn(*pud));
1355
1356         pmd = pmd_offset(pud, addr);
1357         if (pmd_none(*pmd))
1358                 return 0;
1359
1360         if (pmd_large(*pmd))
1361                 return pfn_valid(pmd_pfn(*pmd));
1362
1363         pte = pte_offset_kernel(pmd, addr);
1364         if (pte_none(*pte))
1365                 return 0;
1366
1367         return pfn_valid(pte_pfn(*pte));
1368 }
1369
1370 /*
1371  * Block size is the minimum amount of memory which can be hotplugged or
1372  * hotremoved. It must be power of two and must be equal or larger than
1373  * MIN_MEMORY_BLOCK_SIZE.
1374  */
1375 #define MAX_BLOCK_SIZE (2UL << 30)
1376
1377 /* Amount of ram needed to start using large blocks */
1378 #define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1379
1380 /* Adjustable memory block size */
1381 static unsigned long set_memory_block_size;
1382 int __init set_memory_block_size_order(unsigned int order)
1383 {
1384         unsigned long size = 1UL << order;
1385
1386         if (size > MEM_SIZE_FOR_LARGE_BLOCK || size < MIN_MEMORY_BLOCK_SIZE)
1387                 return -EINVAL;
1388
1389         set_memory_block_size = size;
1390         return 0;
1391 }
1392
1393 static unsigned long probe_memory_block_size(void)
1394 {
1395         unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1396         unsigned long bz;
1397
1398         /* If memory block size has been set, then use it */
1399         bz = set_memory_block_size;
1400         if (bz)
1401                 goto done;
1402
1403         /* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1404         if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1405                 bz = MIN_MEMORY_BLOCK_SIZE;
1406                 goto done;
1407         }
1408
1409         /* Find the largest allowed block size that aligns to memory end */
1410         for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1411                 if (IS_ALIGNED(boot_mem_end, bz))
1412                         break;
1413         }
1414 done:
1415         pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1416
1417         return bz;
1418 }
1419
1420 static unsigned long memory_block_size_probed;
1421 unsigned long memory_block_size_bytes(void)
1422 {
1423         if (!memory_block_size_probed)
1424                 memory_block_size_probed = probe_memory_block_size();
1425
1426         return memory_block_size_probed;
1427 }
1428
1429 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1430 /*
1431  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1432  */
1433 static long __meminitdata addr_start, addr_end;
1434 static void __meminitdata *p_start, *p_end;
1435 static int __meminitdata node_start;
1436
1437 static int __meminit vmemmap_populate_hugepages(unsigned long start,
1438                 unsigned long end, int node, struct vmem_altmap *altmap)
1439 {
1440         unsigned long addr;
1441         unsigned long next;
1442         pgd_t *pgd;
1443         p4d_t *p4d;
1444         pud_t *pud;
1445         pmd_t *pmd;
1446
1447         for (addr = start; addr < end; addr = next) {
1448                 next = pmd_addr_end(addr, end);
1449
1450                 pgd = vmemmap_pgd_populate(addr, node);
1451                 if (!pgd)
1452                         return -ENOMEM;
1453
1454                 p4d = vmemmap_p4d_populate(pgd, addr, node);
1455                 if (!p4d)
1456                         return -ENOMEM;
1457
1458                 pud = vmemmap_pud_populate(p4d, addr, node);
1459                 if (!pud)
1460                         return -ENOMEM;
1461
1462                 pmd = pmd_offset(pud, addr);
1463                 if (pmd_none(*pmd)) {
1464                         void *p;
1465
1466                         if (altmap)
1467                                 p = altmap_alloc_block_buf(PMD_SIZE, altmap);
1468                         else
1469                                 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1470                         if (p) {
1471                                 pte_t entry;
1472
1473                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1474                                                 PAGE_KERNEL_LARGE);
1475                                 set_pmd(pmd, __pmd(pte_val(entry)));
1476
1477                                 /* check to see if we have contiguous blocks */
1478                                 if (p_end != p || node_start != node) {
1479                                         if (p_start)
1480                                                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1481                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1482                                         addr_start = addr;
1483                                         node_start = node;
1484                                         p_start = p;
1485                                 }
1486
1487                                 addr_end = addr + PMD_SIZE;
1488                                 p_end = p + PMD_SIZE;
1489                                 continue;
1490                         } else if (altmap)
1491                                 return -ENOMEM; /* no fallback */
1492                 } else if (pmd_large(*pmd)) {
1493                         vmemmap_verify((pte_t *)pmd, node, addr, next);
1494                         continue;
1495                 }
1496                 if (vmemmap_populate_basepages(addr, next, node))
1497                         return -ENOMEM;
1498         }
1499         return 0;
1500 }
1501
1502 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1503                 struct vmem_altmap *altmap)
1504 {
1505         int err;
1506
1507         if (end - start < PAGES_PER_SECTION * sizeof(struct page))
1508                 err = vmemmap_populate_basepages(start, end, node);
1509         else if (boot_cpu_has(X86_FEATURE_PSE))
1510                 err = vmemmap_populate_hugepages(start, end, node, altmap);
1511         else if (altmap) {
1512                 pr_err_once("%s: no cpu support for altmap allocations\n",
1513                                 __func__);
1514                 err = -ENOMEM;
1515         } else
1516                 err = vmemmap_populate_basepages(start, end, node);
1517         if (!err)
1518                 sync_global_pgds(start, end - 1);
1519         return err;
1520 }
1521
1522 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1523 void register_page_bootmem_memmap(unsigned long section_nr,
1524                                   struct page *start_page, unsigned long nr_pages)
1525 {
1526         unsigned long addr = (unsigned long)start_page;
1527         unsigned long end = (unsigned long)(start_page + nr_pages);
1528         unsigned long next;
1529         pgd_t *pgd;
1530         p4d_t *p4d;
1531         pud_t *pud;
1532         pmd_t *pmd;
1533         unsigned int nr_pmd_pages;
1534         struct page *page;
1535
1536         for (; addr < end; addr = next) {
1537                 pte_t *pte = NULL;
1538
1539                 pgd = pgd_offset_k(addr);
1540                 if (pgd_none(*pgd)) {
1541                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1542                         continue;
1543                 }
1544                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1545
1546                 p4d = p4d_offset(pgd, addr);
1547                 if (p4d_none(*p4d)) {
1548                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1549                         continue;
1550                 }
1551                 get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1552
1553                 pud = pud_offset(p4d, addr);
1554                 if (pud_none(*pud)) {
1555                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1556                         continue;
1557                 }
1558                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1559
1560                 if (!boot_cpu_has(X86_FEATURE_PSE)) {
1561                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1562                         pmd = pmd_offset(pud, addr);
1563                         if (pmd_none(*pmd))
1564                                 continue;
1565                         get_page_bootmem(section_nr, pmd_page(*pmd),
1566                                          MIX_SECTION_INFO);
1567
1568                         pte = pte_offset_kernel(pmd, addr);
1569                         if (pte_none(*pte))
1570                                 continue;
1571                         get_page_bootmem(section_nr, pte_page(*pte),
1572                                          SECTION_INFO);
1573                 } else {
1574                         next = pmd_addr_end(addr, end);
1575
1576                         pmd = pmd_offset(pud, addr);
1577                         if (pmd_none(*pmd))
1578                                 continue;
1579
1580                         nr_pmd_pages = 1 << get_order(PMD_SIZE);
1581                         page = pmd_page(*pmd);
1582                         while (nr_pmd_pages--)
1583                                 get_page_bootmem(section_nr, page++,
1584                                                  SECTION_INFO);
1585                 }
1586         }
1587 }
1588 #endif
1589
1590 void __meminit vmemmap_populate_print_last(void)
1591 {
1592         if (p_start) {
1593                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1594                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1595                 p_start = NULL;
1596                 p_end = NULL;
1597                 node_start = 0;
1598         }
1599 }
1600 #endif