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/pgtable.h>
  41 #include <asm/pgalloc.h>
  42 #include <asm/dma.h>
  43 #include <asm/fixmap.h>
  44 #include <asm/e820/api.h>
  45 #include <asm/apic.h>
  46 #include <asm/tlb.h>
  47 #include <asm/mmu_context.h>
  48 #include <asm/proto.h>
  49 #include <asm/smp.h>
  50 #include <asm/sections.h>
  51 #include <asm/kdebug.h>
  52 #include <asm/numa.h>
  53 #include <asm/set_memory.h>
  54 #include <asm/init.h>
  55 #include <asm/uv/uv.h>
  56 #include <asm/setup.h>
  57 #include <asm/ftrace.h>
  58
  59 #include "mm_internal.h"
  60
  61 #include "ident_map.c"
  62
  63 #define DEFINE_POPULATE(fname, type1, type2, init)              \
  64 static inline void fname##_init(struct mm_struct *mm,           \
  65                 type1##_t *arg1, type2##_t *arg2, bool init)    \
  66 {                                                               \
  67         if (init)                                               \
  68                 fname##_safe(mm, arg1, arg2);                   \
  69         else                                                    \
  70                 fname(mm, arg1, arg2);                          \
  71 }
  72
  73 DEFINE_POPULATE(p4d_populate, p4d, pud, init)
  74 DEFINE_POPULATE(pgd_populate, pgd, p4d, init)
  75 DEFINE_POPULATE(pud_populate, pud, pmd, init)
  76 DEFINE_POPULATE(pmd_populate_kernel, pmd, pte, init)
  77
  78 #define DEFINE_ENTRY(type1, type2, init)                        \
  79 static inline void set_##type1##_init(type1##_t *arg1,          \
  80                         type2##_t arg2, bool init)              \
  81 {                                                               \
  82         if (init)                                               \
  83                 set_##type1##_safe(arg1, arg2);                 \
  84         else                                                    \
  85                 set_##type1(arg1, arg2);                        \
  86 }
  87
  88 DEFINE_ENTRY(p4d, p4d, init)
  89 DEFINE_ENTRY(pud, pud, init)
  90 DEFINE_ENTRY(pmd, pmd, init)
  91 DEFINE_ENTRY(pte, pte, init)
  92
  93
  94 /*
  95  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  96  * physical space so we can cache the place of the first one and move
  97  * around without checking the pgd every time.
  98  */
  99
 100 /* Bits supported by the hardware: */
 101 pteval_t __supported_pte_mask __read_mostly = ~0;
 102 /* Bits allowed in normal kernel mappings: */
 103 pteval_t __default_kernel_pte_mask __read_mostly = ~0;
 104 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 105 /* Used in PAGE_KERNEL_* macros which are reasonably used out-of-tree: */
 106 EXPORT_SYMBOL(__default_kernel_pte_mask);
 107
 108 int force_personality32;
 109
 110 /*
 111  * noexec32=on|off
 112  * Control non executable heap for 32bit processes.
 113  * To control the stack too use noexec=off
 114  *
 115  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
 116  * off  PROT_READ implies PROT_EXEC
 117  */
 118 static int __init nonx32_setup(char *str)
 119 {
 120         if (!strcmp(str, "on"))
 121                 force_personality32 &= ~READ_IMPLIES_EXEC;
 122         else if (!strcmp(str, "off"))
 123                 force_personality32 |= READ_IMPLIES_EXEC;
 124         return 1;
 125 }
 126 __setup("noexec32=", nonx32_setup);
 127
 128 static void sync_global_pgds_l5(unsigned long start, unsigned long end)
 129 {
 130         unsigned long addr;
 131
 132         for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
 133                 const pgd_t *pgd_ref = pgd_offset_k(addr);
 134                 struct page *page;
 135
 136                 /* Check for overflow */
 137                 if (addr < start)
 138                         break;
 139
 140                 if (pgd_none(*pgd_ref))
 141                         continue;
 142
 143                 spin_lock(&pgd_lock);
 144                 list_for_each_entry(page, &pgd_list, lru) {
 145                         pgd_t *pgd;
 146                         spinlock_t *pgt_lock;
 147
 148                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
 149                         /* the pgt_lock only for Xen */
 150                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 151                         spin_lock(pgt_lock);
 152
 153                         if (!pgd_none(*pgd_ref) && !pgd_none(*pgd))
 154                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 155
 156                         if (pgd_none(*pgd))
 157                                 set_pgd(pgd, *pgd_ref);
 158
 159                         spin_unlock(pgt_lock);
 160                 }
 161                 spin_unlock(&pgd_lock);
 162         }
 163 }
 164
 165 static void sync_global_pgds_l4(unsigned long start, unsigned long end)
 166 {
 167         unsigned long addr;
 168
 169         for (addr = start; addr <= end; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
 170                 pgd_t *pgd_ref = pgd_offset_k(addr);
 171                 const p4d_t *p4d_ref;
 172                 struct page *page;
 173
 174                 /*
 175                  * With folded p4d, pgd_none() is always false, we need to
 176                  * handle synchonization on p4d level.
 177                  */
 178                 MAYBE_BUILD_BUG_ON(pgd_none(*pgd_ref));
 179                 p4d_ref = p4d_offset(pgd_ref, addr);
 180
 181                 if (p4d_none(*p4d_ref))
 182                         continue;
 183
 184                 spin_lock(&pgd_lock);
 185                 list_for_each_entry(page, &pgd_list, lru) {
 186                         pgd_t *pgd;
 187                         p4d_t *p4d;
 188                         spinlock_t *pgt_lock;
 189
 190                         pgd = (pgd_t *)page_address(page) + pgd_index(addr);
 191                         p4d = p4d_offset(pgd, addr);
 192                         /* the pgt_lock only for Xen */
 193                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 194                         spin_lock(pgt_lock);
 195
 196                         if (!p4d_none(*p4d_ref) && !p4d_none(*p4d))
 197                                 BUG_ON(p4d_page_vaddr(*p4d)
 198                                        != p4d_page_vaddr(*p4d_ref));
 199
 200                         if (p4d_none(*p4d))
 201                                 set_p4d(p4d, *p4d_ref);
 202
 203                         spin_unlock(pgt_lock);
 204                 }
 205                 spin_unlock(&pgd_lock);
 206         }
 207 }
 208
 209 /*
 210  * When memory was added make sure all the processes MM have
 211  * suitable PGD entries in the local PGD level page.
 212  */
 213 void sync_global_pgds(unsigned long start, unsigned long end)
 214 {
 215         if (pgtable_l5_enabled())
 216                 sync_global_pgds_l5(start, end);
 217         else
 218                 sync_global_pgds_l4(start, end);
 219 }
 220
 221 void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
 222 {
 223         sync_global_pgds(start, end);
 224 }
 225
 226 /*
 227  * NOTE: This function is marked __ref because it calls __init function
 228  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 229  */
 230 static __ref void *spp_getpage(void)
 231 {
 232         void *ptr;
 233
 234         if (after_bootmem)
 235                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 236         else
 237                 ptr = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
 238
 239         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 240                 panic("set_pte_phys: cannot allocate page data %s\n",
 241                         after_bootmem ? "after bootmem" : "");
 242         }
 243
 244         pr_debug("spp_getpage %p\n", ptr);
 245
 246         return ptr;
 247 }
 248
 249 static p4d_t *fill_p4d(pgd_t *pgd, unsigned long vaddr)
 250 {
 251         if (pgd_none(*pgd)) {
 252                 p4d_t *p4d = (p4d_t *)spp_getpage();
 253                 pgd_populate(&init_mm, pgd, p4d);
 254                 if (p4d != p4d_offset(pgd, 0))
 255                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 256                                p4d, p4d_offset(pgd, 0));
 257         }
 258         return p4d_offset(pgd, vaddr);
 259 }
 260
 261 static pud_t *fill_pud(p4d_t *p4d, unsigned long vaddr)
 262 {
 263         if (p4d_none(*p4d)) {
 264                 pud_t *pud = (pud_t *)spp_getpage();
 265                 p4d_populate(&init_mm, p4d, pud);
 266                 if (pud != pud_offset(p4d, 0))
 267                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 268                                pud, pud_offset(p4d, 0));
 269         }
 270         return pud_offset(p4d, vaddr);
 271 }
 272
 273 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 274 {
 275         if (pud_none(*pud)) {
 276                 pmd_t *pmd = (pmd_t *) spp_getpage();
 277                 pud_populate(&init_mm, pud, pmd);
 278                 if (pmd != pmd_offset(pud, 0))
 279                         printk(KERN_ERR "PAGETABLE BUG #02! %p <-> %p\n",
 280                                pmd, pmd_offset(pud, 0));
 281         }
 282         return pmd_offset(pud, vaddr);
 283 }
 284
 285 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 286 {
 287         if (pmd_none(*pmd)) {
 288                 pte_t *pte = (pte_t *) spp_getpage();
 289                 pmd_populate_kernel(&init_mm, pmd, pte);
 290                 if (pte != pte_offset_kernel(pmd, 0))
 291                         printk(KERN_ERR "PAGETABLE BUG #03!\n");
 292         }
 293         return pte_offset_kernel(pmd, vaddr);
 294 }
 295
 296 static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
 297 {
 298         pmd_t *pmd = fill_pmd(pud, vaddr);
 299         pte_t *pte = fill_pte(pmd, vaddr);
 300
 301         set_pte(pte, new_pte);
 302
 303         /*
 304          * It's enough to flush this one mapping.
 305          * (PGE mappings get flushed as well)
 306          */
 307         __flush_tlb_one_kernel(vaddr);
 308 }
 309
 310 void set_pte_vaddr_p4d(p4d_t *p4d_page, unsigned long vaddr, pte_t new_pte)
 311 {
 312         p4d_t *p4d = p4d_page + p4d_index(vaddr);
 313         pud_t *pud = fill_pud(p4d, vaddr);
 314
 315         __set_pte_vaddr(pud, vaddr, new_pte);
 316 }
 317
 318 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 319 {
 320         pud_t *pud = pud_page + pud_index(vaddr);
 321
 322         __set_pte_vaddr(pud, vaddr, new_pte);
 323 }
 324
 325 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 326 {
 327         pgd_t *pgd;
 328         p4d_t *p4d_page;
 329
 330         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 331
 332         pgd = pgd_offset_k(vaddr);
 333         if (pgd_none(*pgd)) {
 334                 printk(KERN_ERR
 335                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 336                 return;
 337         }
 338
 339         p4d_page = p4d_offset(pgd, 0);
 340         set_pte_vaddr_p4d(p4d_page, vaddr, pteval);
 341 }
 342
 343 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 344 {
 345         pgd_t *pgd;
 346         p4d_t *p4d;
 347         pud_t *pud;
 348
 349         pgd = pgd_offset_k(vaddr);
 350         p4d = fill_p4d(pgd, vaddr);
 351         pud = fill_pud(p4d, vaddr);
 352         return fill_pmd(pud, vaddr);
 353 }
 354
 355 pte_t * __init populate_extra_pte(unsigned long vaddr)
 356 {
 357         pmd_t *pmd;
 358
 359         pmd = populate_extra_pmd(vaddr);
 360         return fill_pte(pmd, vaddr);
 361 }
 362
 363 /*
 364  * Create large page table mappings for a range of physical addresses.
 365  */
 366 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 367                                         enum page_cache_mode cache)
 368 {
 369         pgd_t *pgd;
 370         p4d_t *p4d;
 371         pud_t *pud;
 372         pmd_t *pmd;
 373         pgprot_t prot;
 374
 375         pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
 376                 pgprot_val(pgprot_4k_2_large(cachemode2pgprot(cache)));
 377         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 378         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 379                 pgd = pgd_offset_k((unsigned long)__va(phys));
 380                 if (pgd_none(*pgd)) {
 381                         p4d = (p4d_t *) spp_getpage();
 382                         set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE |
 383                                                 _PAGE_USER));
 384                 }
 385                 p4d = p4d_offset(pgd, (unsigned long)__va(phys));
 386                 if (p4d_none(*p4d)) {
 387                         pud = (pud_t *) spp_getpage();
 388                         set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE |
 389                                                 _PAGE_USER));
 390                 }
 391                 pud = pud_offset(p4d, (unsigned long)__va(phys));
 392                 if (pud_none(*pud)) {
 393                         pmd = (pmd_t *) spp_getpage();
 394                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 395                                                 _PAGE_USER));
 396                 }
 397                 pmd = pmd_offset(pud, phys);
 398                 BUG_ON(!pmd_none(*pmd));
 399                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 400         }
 401 }
 402
 403 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 404 {
 405         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
 406 }
 407
 408 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 409 {
 410         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
 411 }
 412
 413 /*
 414  * The head.S code sets up the kernel high mapping:
 415  *
 416  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 417  *
 418  * phys_base holds the negative offset to the kernel, which is added
 419  * to the compile time generated pmds. This results in invalid pmds up
 420  * to the point where we hit the physaddr 0 mapping.
 421  *
 422  * We limit the mappings to the region from _text to _brk_end.  _brk_end
 423  * is rounded up to the 2MB boundary. This catches the invalid pmds as
 424  * well, as they are located before _text:
 425  */
 426 void __init cleanup_highmap(void)
 427 {
 428         unsigned long vaddr = __START_KERNEL_map;
 429         unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
 430         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 431         pmd_t *pmd = level2_kernel_pgt;
 432
 433         /*
 434          * Native path, max_pfn_mapped is not set yet.
 435          * Xen has valid max_pfn_mapped set in
 436          *      arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
 437          */
 438         if (max_pfn_mapped)
 439                 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 440
 441         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
 442                 if (pmd_none(*pmd))
 443                         continue;
 444                 if (vaddr < (unsigned long) _text || vaddr > end)
 445                         set_pmd(pmd, __pmd(0));
 446         }
 447 }
 448
 449 /*
 450  * Create PTE level page table mapping for physical addresses.
 451  * It returns the last physical address mapped.
 452  */
 453 static unsigned long __meminit
 454 phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
 455               pgprot_t prot, bool init)
 456 {
 457         unsigned long pages = 0, paddr_next;
 458         unsigned long paddr_last = paddr_end;
 459         pte_t *pte;
 460         int i;
 461
 462         pte = pte_page + pte_index(paddr);
 463         i = pte_index(paddr);
 464
 465         for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
 466                 paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
 467                 if (paddr >= paddr_end) {
 468                         if (!after_bootmem &&
 469                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 470                                              E820_TYPE_RAM) &&
 471                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 472                                              E820_TYPE_RESERVED_KERN))
 473                                 set_pte_init(pte, __pte(0), init);
 474                         continue;
 475                 }
 476
 477                 /*
 478                  * We will re-use the existing mapping.
 479                  * Xen for example has some special requirements, like mapping
 480                  * pagetable pages as RO. So assume someone who pre-setup
 481                  * these mappings are more intelligent.
 482                  */
 483                 if (!pte_none(*pte)) {
 484                         if (!after_bootmem)
 485                                 pages++;
 486                         continue;
 487                 }
 488
 489                 if (0)
 490                         pr_info("   pte=%p addr=%lx pte=%016lx\n", pte, paddr,
 491                                 pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 492                 pages++;
 493                 set_pte_init(pte, pfn_pte(paddr >> PAGE_SHIFT, prot), init);
 494                 paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
 495         }
 496
 497         update_page_count(PG_LEVEL_4K, pages);
 498
 499         return paddr_last;
 500 }
 501
 502 /*
 503  * Create PMD level page table mapping for physical addresses. The virtual
 504  * and physical address have to be aligned at this level.
 505  * It returns the last physical address mapped.
 506  */
 507 static unsigned long __meminit
 508 phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
 509               unsigned long page_size_mask, pgprot_t prot, bool init)
 510 {
 511         unsigned long pages = 0, paddr_next;
 512         unsigned long paddr_last = paddr_end;
 513
 514         int i = pmd_index(paddr);
 515
 516         for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
 517                 pmd_t *pmd = pmd_page + pmd_index(paddr);
 518                 pte_t *pte;
 519                 pgprot_t new_prot = prot;
 520
 521                 paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
 522                 if (paddr >= paddr_end) {
 523                         if (!after_bootmem &&
 524                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 525                                              E820_TYPE_RAM) &&
 526                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 527                                              E820_TYPE_RESERVED_KERN))
 528                                 set_pmd_init(pmd, __pmd(0), init);
 529                         continue;
 530                 }
 531
 532                 if (!pmd_none(*pmd)) {
 533                         if (!pmd_large(*pmd)) {
 534                                 spin_lock(&init_mm.page_table_lock);
 535                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
 536                                 paddr_last = phys_pte_init(pte, paddr,
 537                                                            paddr_end, prot,
 538                                                            init);
 539                                 spin_unlock(&init_mm.page_table_lock);
 540                                 continue;
 541                         }
 542                         /*
 543                          * If we are ok with PG_LEVEL_2M mapping, then we will
 544                          * use the existing mapping,
 545                          *
 546                          * Otherwise, we will split the large page mapping but
 547                          * use the same existing protection bits except for
 548                          * large page, so that we don't violate Intel's TLB
 549                          * Application note (317080) which says, while changing
 550                          * the page sizes, new and old translations should
 551                          * not differ with respect to page frame and
 552                          * attributes.
 553                          */
 554                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 555                                 if (!after_bootmem)
 556                                         pages++;
 557                                 paddr_last = paddr_next;
 558                                 continue;
 559                         }
 560                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 561                 }
 562
 563                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 564                         pages++;
 565                         spin_lock(&init_mm.page_table_lock);
 566                         set_pte_init((pte_t *)pmd,
 567                                      pfn_pte((paddr & PMD_MASK) >> PAGE_SHIFT,
 568                                              __pgprot(pgprot_val(prot) | _PAGE_PSE)),
 569                                      init);
 570                         spin_unlock(&init_mm.page_table_lock);
 571                         paddr_last = paddr_next;
 572                         continue;
 573                 }
 574
 575                 pte = alloc_low_page();
 576                 paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot, init);
 577
 578                 spin_lock(&init_mm.page_table_lock);
 579                 pmd_populate_kernel_init(&init_mm, pmd, pte, init);
 580                 spin_unlock(&init_mm.page_table_lock);
 581         }
 582         update_page_count(PG_LEVEL_2M, pages);
 583         return paddr_last;
 584 }
 585
 586 /*
 587  * Create PUD level page table mapping for physical addresses. The virtual
 588  * and physical address do not have to be aligned at this level. KASLR can
 589  * randomize virtual addresses up to this level.
 590  * It returns the last physical address mapped.
 591  */
 592 static unsigned long __meminit
 593 phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
 594               unsigned long page_size_mask, pgprot_t _prot, bool init)
 595 {
 596         unsigned long pages = 0, paddr_next;
 597         unsigned long paddr_last = paddr_end;
 598         unsigned long vaddr = (unsigned long)__va(paddr);
 599         int i = pud_index(vaddr);
 600
 601         for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
 602                 pud_t *pud;
 603                 pmd_t *pmd;
 604                 pgprot_t prot = _prot;
 605
 606                 vaddr = (unsigned long)__va(paddr);
 607                 pud = pud_page + pud_index(vaddr);
 608                 paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
 609
 610                 if (paddr >= paddr_end) {
 611                         if (!after_bootmem &&
 612                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 613                                              E820_TYPE_RAM) &&
 614                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 615                                              E820_TYPE_RESERVED_KERN))
 616                                 set_pud_init(pud, __pud(0), init);
 617                         continue;
 618                 }
 619
 620                 if (!pud_none(*pud)) {
 621                         if (!pud_large(*pud)) {
 622                                 pmd = pmd_offset(pud, 0);
 623                                 paddr_last = phys_pmd_init(pmd, paddr,
 624                                                            paddr_end,
 625                                                            page_size_mask,
 626                                                            prot, init);
 627                                 continue;
 628                         }
 629                         /*
 630                          * If we are ok with PG_LEVEL_1G mapping, then we will
 631                          * use the existing mapping.
 632                          *
 633                          * Otherwise, we will split the gbpage mapping but use
 634                          * the same existing protection  bits except for large
 635                          * page, so that we don't violate Intel's TLB
 636                          * Application note (317080) which says, while changing
 637                          * the page sizes, new and old translations should
 638                          * not differ with respect to page frame and
 639                          * attributes.
 640                          */
 641                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 642                                 if (!after_bootmem)
 643                                         pages++;
 644                                 paddr_last = paddr_next;
 645                                 continue;
 646                         }
 647                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 648                 }
 649
 650                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 651                         pages++;
 652                         spin_lock(&init_mm.page_table_lock);
 653
 654                         prot = __pgprot(pgprot_val(prot) | __PAGE_KERNEL_LARGE);
 655
 656                         set_pte_init((pte_t *)pud,
 657                                      pfn_pte((paddr & PUD_MASK) >> PAGE_SHIFT,
 658                                              prot),
 659                                      init);
 660                         spin_unlock(&init_mm.page_table_lock);
 661                         paddr_last = paddr_next;
 662                         continue;
 663                 }
 664
 665                 pmd = alloc_low_page();
 666                 paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
 667                                            page_size_mask, prot, init);
 668
 669                 spin_lock(&init_mm.page_table_lock);
 670                 pud_populate_init(&init_mm, pud, pmd, init);
 671                 spin_unlock(&init_mm.page_table_lock);
 672         }
 673
 674         update_page_count(PG_LEVEL_1G, pages);
 675
 676         return paddr_last;
 677 }
 678
 679 static unsigned long __meminit
 680 phys_p4d_init(p4d_t *p4d_page, unsigned long paddr, unsigned long paddr_end,
 681               unsigned long page_size_mask, pgprot_t prot, bool init)
 682 {
 683         unsigned long vaddr, vaddr_end, vaddr_next, paddr_next, paddr_last;
 684
 685         paddr_last = paddr_end;
 686         vaddr = (unsigned long)__va(paddr);
 687         vaddr_end = (unsigned long)__va(paddr_end);
 688
 689         if (!pgtable_l5_enabled())
 690                 return phys_pud_init((pud_t *) p4d_page, paddr, paddr_end,
 691                                      page_size_mask, prot, init);
 692
 693         for (; vaddr < vaddr_end; vaddr = vaddr_next) {
 694                 p4d_t *p4d = p4d_page + p4d_index(vaddr);
 695                 pud_t *pud;
 696
 697                 vaddr_next = (vaddr & P4D_MASK) + P4D_SIZE;
 698                 paddr = __pa(vaddr);
 699
 700                 if (paddr >= paddr_end) {
 701                         paddr_next = __pa(vaddr_next);
 702                         if (!after_bootmem &&
 703                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 704                                              E820_TYPE_RAM) &&
 705                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 706                                              E820_TYPE_RESERVED_KERN))
 707                                 set_p4d_init(p4d, __p4d(0), init);
 708                         continue;
 709                 }
 710
 711                 if (!p4d_none(*p4d)) {
 712                         pud = pud_offset(p4d, 0);
 713                         paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
 714                                         page_size_mask, prot, init);
 715                         continue;
 716                 }
 717
 718                 pud = alloc_low_page();
 719                 paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
 720                                            page_size_mask, prot, init);
 721
 722                 spin_lock(&init_mm.page_table_lock);
 723                 p4d_populate_init(&init_mm, p4d, pud, init);
 724                 spin_unlock(&init_mm.page_table_lock);
 725         }
 726
 727         return paddr_last;
 728 }
 729
 730 static unsigned long __meminit
 731 __kernel_physical_mapping_init(unsigned long paddr_start,
 732                                unsigned long paddr_end,
 733                                unsigned long page_size_mask,
 734                                pgprot_t prot, bool init)
 735 {
 736         bool pgd_changed = false;
 737         unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
 738
 739         paddr_last = paddr_end;
 740         vaddr = (unsigned long)__va(paddr_start);
 741         vaddr_end = (unsigned long)__va(paddr_end);
 742         vaddr_start = vaddr;
 743
 744         for (; vaddr < vaddr_end; vaddr = vaddr_next) {
 745                 pgd_t *pgd = pgd_offset_k(vaddr);
 746                 p4d_t *p4d;
 747
 748                 vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
 749
 750                 if (pgd_val(*pgd)) {
 751                         p4d = (p4d_t *)pgd_page_vaddr(*pgd);
 752                         paddr_last = phys_p4d_init(p4d, __pa(vaddr),
 753                                                    __pa(vaddr_end),
 754                                                    page_size_mask,
 755                                                    prot, init);
 756                         continue;
 757                 }
 758
 759                 p4d = alloc_low_page();
 760                 paddr_last = phys_p4d_init(p4d, __pa(vaddr), __pa(vaddr_end),
 761                                            page_size_mask, prot, init);
 762
 763                 spin_lock(&init_mm.page_table_lock);
 764                 if (pgtable_l5_enabled())
 765                         pgd_populate_init(&init_mm, pgd, p4d, init);
 766                 else
 767                         p4d_populate_init(&init_mm, p4d_offset(pgd, vaddr),
 768                                           (pud_t *) p4d, init);
 769
 770                 spin_unlock(&init_mm.page_table_lock);
 771                 pgd_changed = true;
 772         }
 773
 774         if (pgd_changed)
 775                 sync_global_pgds(vaddr_start, vaddr_end - 1);
 776
 777         return paddr_last;
 778 }
 779
 780
 781 /*
 782  * Create page table mapping for the physical memory for specific physical
 783  * addresses. Note that it can only be used to populate non-present entries.
 784  * The virtual and physical addresses have to be aligned on PMD level
 785  * down. It returns the last physical address mapped.
 786  */
 787 unsigned long __meminit
 788 kernel_physical_mapping_init(unsigned long paddr_start,
 789                              unsigned long paddr_end,
 790                              unsigned long page_size_mask, pgprot_t prot)
 791 {
 792         return __kernel_physical_mapping_init(paddr_start, paddr_end,
 793                                               page_size_mask, prot, true);
 794 }
 795
 796 /*
 797  * This function is similar to kernel_physical_mapping_init() above with the
 798  * exception that it uses set_{pud,pmd}() instead of the set_{pud,pte}_safe()
 799  * when updating the mapping. The caller is responsible to flush the TLBs after
 800  * the function returns.
 801  */
 802 unsigned long __meminit
 803 kernel_physical_mapping_change(unsigned long paddr_start,
 804                                unsigned long paddr_end,
 805                                unsigned long page_size_mask)
 806 {
 807         return __kernel_physical_mapping_init(paddr_start, paddr_end,
 808                                               page_size_mask, PAGE_KERNEL,
 809                                               false);
 810 }
 811
 812 #ifndef CONFIG_NUMA
 813 void __init initmem_init(void)
 814 {
 815         memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
 816 }
 817 #endif
 818
 819 void __init paging_init(void)
 820 {
 821         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 822         sparse_init();
 823
 824         /*
 825          * clear the default setting with node 0
 826          * note: don't use nodes_clear here, that is really clearing when
 827          *       numa support is not compiled in, and later node_set_state
 828          *       will not set it back.
 829          */
 830         node_clear_state(0, N_MEMORY);
 831         node_clear_state(0, N_NORMAL_MEMORY);
 832
 833         zone_sizes_init();
 834 }
 835
 836 /*
 837  * Memory hotplug specific functions
 838  */
 839 #ifdef CONFIG_MEMORY_HOTPLUG
 840 /*
 841  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 842  * updating.
 843  */
 844 static void update_end_of_memory_vars(u64 start, u64 size)
 845 {
 846         unsigned long end_pfn = PFN_UP(start + size);
 847
 848         if (end_pfn > max_pfn) {
 849                 max_pfn = end_pfn;
 850                 max_low_pfn = end_pfn;
 851                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 852         }
 853 }
 854
 855 int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
 856               struct mhp_params *params)
 857 {
 858         int ret;
 859
 860         ret = __add_pages(nid, start_pfn, nr_pages, params);
 861         WARN_ON_ONCE(ret);
 862
 863         /* update max_pfn, max_low_pfn and high_memory */
 864         update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
 865                                   nr_pages << PAGE_SHIFT);
 866
 867         return ret;
 868 }
 869
 870 int arch_add_memory(int nid, u64 start, u64 size,
 871                     struct mhp_params *params)
 872 {
 873         unsigned long start_pfn = start >> PAGE_SHIFT;
 874         unsigned long nr_pages = size >> PAGE_SHIFT;
 875
 876         init_memory_mapping(start, start + size, params->pgprot);
 877
 878         return add_pages(nid, start_pfn, nr_pages, params);
 879 }
 880
 881 #define PAGE_INUSE 0xFD
 882
 883 static void __meminit free_pagetable(struct page *page, int order)
 884 {
 885         unsigned long magic;
 886         unsigned int nr_pages = 1 << order;
 887
 888         /* bootmem page has reserved flag */
 889         if (PageReserved(page)) {
 890                 __ClearPageReserved(page);
 891
 892                 magic = (unsigned long)page->freelist;
 893                 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
 894                         while (nr_pages--)
 895                                 put_page_bootmem(page++);
 896                 } else
 897                         while (nr_pages--)
 898                                 free_reserved_page(page++);
 899         } else
 900                 free_pages((unsigned long)page_address(page), order);
 901 }
 902
 903 static void __meminit free_hugepage_table(struct page *page,
 904                 struct vmem_altmap *altmap)
 905 {
 906         if (altmap)
 907                 vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
 908         else
 909                 free_pagetable(page, get_order(PMD_SIZE));
 910 }
 911
 912 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
 913 {
 914         pte_t *pte;
 915         int i;
 916
 917         for (i = 0; i < PTRS_PER_PTE; i++) {
 918                 pte = pte_start + i;
 919                 if (!pte_none(*pte))
 920                         return;
 921         }
 922
 923         /* free a pte talbe */
 924         free_pagetable(pmd_page(*pmd), 0);
 925         spin_lock(&init_mm.page_table_lock);
 926         pmd_clear(pmd);
 927         spin_unlock(&init_mm.page_table_lock);
 928 }
 929
 930 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
 931 {
 932         pmd_t *pmd;
 933         int i;
 934
 935         for (i = 0; i < PTRS_PER_PMD; i++) {
 936                 pmd = pmd_start + i;
 937                 if (!pmd_none(*pmd))
 938                         return;
 939         }
 940
 941         /* free a pmd talbe */
 942         free_pagetable(pud_page(*pud), 0);
 943         spin_lock(&init_mm.page_table_lock);
 944         pud_clear(pud);
 945         spin_unlock(&init_mm.page_table_lock);
 946 }
 947
 948 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
 949 {
 950         pud_t *pud;
 951         int i;
 952
 953         for (i = 0; i < PTRS_PER_PUD; i++) {
 954                 pud = pud_start + i;
 955                 if (!pud_none(*pud))
 956                         return;
 957         }
 958
 959         /* free a pud talbe */
 960         free_pagetable(p4d_page(*p4d), 0);
 961         spin_lock(&init_mm.page_table_lock);
 962         p4d_clear(p4d);
 963         spin_unlock(&init_mm.page_table_lock);
 964 }
 965
 966 static void __meminit
 967 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
 968                  bool direct)
 969 {
 970         unsigned long next, pages = 0;
 971         pte_t *pte;
 972         void *page_addr;
 973         phys_addr_t phys_addr;
 974
 975         pte = pte_start + pte_index(addr);
 976         for (; addr < end; addr = next, pte++) {
 977                 next = (addr + PAGE_SIZE) & PAGE_MASK;
 978                 if (next > end)
 979                         next = end;
 980
 981                 if (!pte_present(*pte))
 982                         continue;
 983
 984                 /*
 985                  * We mapped [0,1G) memory as identity mapping when
 986                  * initializing, in arch/x86/kernel/head_64.S. These
 987                  * pagetables cannot be removed.
 988                  */
 989                 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
 990                 if (phys_addr < (phys_addr_t)0x40000000)
 991                         return;
 992
 993                 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
 994                         /*
 995                          * Do not free direct mapping pages since they were
 996                          * freed when offlining, or simplely not in use.
 997                          */
 998                         if (!direct)
 999                                 free_pagetable(pte_page(*pte), 0);
1000
1001                         spin_lock(&init_mm.page_table_lock);
1002                         pte_clear(&init_mm, addr, pte);
1003                         spin_unlock(&init_mm.page_table_lock);
1004
1005                         /* For non-direct mapping, pages means nothing. */
1006                         pages++;
1007                 } else {
1008                         /*
1009                          * If we are here, we are freeing vmemmap pages since
1010                          * direct mapped memory ranges to be freed are aligned.
1011                          *
1012                          * If we are not removing the whole page, it means
1013                          * other page structs in this page are being used and
1014                          * we canot remove them. So fill the unused page_structs
1015                          * with 0xFD, and remove the page when it is wholly
1016                          * filled with 0xFD.
1017                          */
1018                         memset((void *)addr, PAGE_INUSE, next - addr);
1019
1020                         page_addr = page_address(pte_page(*pte));
1021                         if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
1022                                 free_pagetable(pte_page(*pte), 0);
1023
1024                                 spin_lock(&init_mm.page_table_lock);
1025                                 pte_clear(&init_mm, addr, pte);
1026                                 spin_unlock(&init_mm.page_table_lock);
1027                         }
1028                 }
1029         }
1030
1031         /* Call free_pte_table() in remove_pmd_table(). */
1032         flush_tlb_all();
1033         if (direct)
1034                 update_page_count(PG_LEVEL_4K, -pages);
1035 }
1036
1037 static void __meminit
1038 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
1039                  bool direct, struct vmem_altmap *altmap)
1040 {
1041         unsigned long next, pages = 0;
1042         pte_t *pte_base;
1043         pmd_t *pmd;
1044         void *page_addr;
1045
1046         pmd = pmd_start + pmd_index(addr);
1047         for (; addr < end; addr = next, pmd++) {
1048                 next = pmd_addr_end(addr, end);
1049
1050                 if (!pmd_present(*pmd))
1051                         continue;
1052
1053                 if (pmd_large(*pmd)) {
1054                         if (IS_ALIGNED(addr, PMD_SIZE) &&
1055                             IS_ALIGNED(next, PMD_SIZE)) {
1056                                 if (!direct)
1057                                         free_hugepage_table(pmd_page(*pmd),
1058                                                             altmap);
1059
1060                                 spin_lock(&init_mm.page_table_lock);
1061                                 pmd_clear(pmd);
1062                                 spin_unlock(&init_mm.page_table_lock);
1063                                 pages++;
1064                         } else {
1065                                 /* If here, we are freeing vmemmap pages. */
1066                                 memset((void *)addr, PAGE_INUSE, next - addr);
1067
1068                                 page_addr = page_address(pmd_page(*pmd));
1069                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1070                                                 PMD_SIZE)) {
1071                                         free_hugepage_table(pmd_page(*pmd),
1072                                                             altmap);
1073
1074                                         spin_lock(&init_mm.page_table_lock);
1075                                         pmd_clear(pmd);
1076                                         spin_unlock(&init_mm.page_table_lock);
1077                                 }
1078                         }
1079
1080                         continue;
1081                 }
1082
1083                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1084                 remove_pte_table(pte_base, addr, next, direct);
1085                 free_pte_table(pte_base, pmd);
1086         }
1087
1088         /* Call free_pmd_table() in remove_pud_table(). */
1089         if (direct)
1090                 update_page_count(PG_LEVEL_2M, -pages);
1091 }
1092
1093 static void __meminit
1094 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1095                  struct vmem_altmap *altmap, bool direct)
1096 {
1097         unsigned long next, pages = 0;
1098         pmd_t *pmd_base;
1099         pud_t *pud;
1100         void *page_addr;
1101
1102         pud = pud_start + pud_index(addr);
1103         for (; addr < end; addr = next, pud++) {
1104                 next = pud_addr_end(addr, end);
1105
1106                 if (!pud_present(*pud))
1107                         continue;
1108
1109                 if (pud_large(*pud)) {
1110                         if (IS_ALIGNED(addr, PUD_SIZE) &&
1111                             IS_ALIGNED(next, PUD_SIZE)) {
1112                                 if (!direct)
1113                                         free_pagetable(pud_page(*pud),
1114                                                        get_order(PUD_SIZE));
1115
1116                                 spin_lock(&init_mm.page_table_lock);
1117                                 pud_clear(pud);
1118                                 spin_unlock(&init_mm.page_table_lock);
1119                                 pages++;
1120                         } else {
1121                                 /* If here, we are freeing vmemmap pages. */
1122                                 memset((void *)addr, PAGE_INUSE, next - addr);
1123
1124                                 page_addr = page_address(pud_page(*pud));
1125                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1126                                                 PUD_SIZE)) {
1127                                         free_pagetable(pud_page(*pud),
1128                                                        get_order(PUD_SIZE));
1129
1130                                         spin_lock(&init_mm.page_table_lock);
1131                                         pud_clear(pud);
1132                                         spin_unlock(&init_mm.page_table_lock);
1133                                 }
1134                         }
1135
1136                         continue;
1137                 }
1138
1139                 pmd_base = pmd_offset(pud, 0);
1140                 remove_pmd_table(pmd_base, addr, next, direct, altmap);
1141                 free_pmd_table(pmd_base, pud);
1142         }
1143
1144         if (direct)
1145                 update_page_count(PG_LEVEL_1G, -pages);
1146 }
1147
1148 static void __meminit
1149 remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1150                  struct vmem_altmap *altmap, bool direct)
1151 {
1152         unsigned long next, pages = 0;
1153         pud_t *pud_base;
1154         p4d_t *p4d;
1155
1156         p4d = p4d_start + p4d_index(addr);
1157         for (; addr < end; addr = next, p4d++) {
1158                 next = p4d_addr_end(addr, end);
1159
1160                 if (!p4d_present(*p4d))
1161                         continue;
1162
1163                 BUILD_BUG_ON(p4d_large(*p4d));
1164
1165                 pud_base = pud_offset(p4d, 0);
1166                 remove_pud_table(pud_base, addr, next, altmap, direct);
1167                 /*
1168                  * For 4-level page tables we do not want to free PUDs, but in the
1169                  * 5-level case we should free them. This code will have to change
1170                  * to adapt for boot-time switching between 4 and 5 level page tables.
1171                  */
1172                 if (pgtable_l5_enabled())
1173                         free_pud_table(pud_base, p4d);
1174         }
1175
1176         if (direct)
1177                 update_page_count(PG_LEVEL_512G, -pages);
1178 }
1179
1180 /* start and end are both virtual address. */
1181 static void __meminit
1182 remove_pagetable(unsigned long start, unsigned long end, bool direct,
1183                 struct vmem_altmap *altmap)
1184 {
1185         unsigned long next;
1186         unsigned long addr;
1187         pgd_t *pgd;
1188         p4d_t *p4d;
1189
1190         for (addr = start; addr < end; addr = next) {
1191                 next = pgd_addr_end(addr, end);
1192
1193                 pgd = pgd_offset_k(addr);
1194                 if (!pgd_present(*pgd))
1195                         continue;
1196
1197                 p4d = p4d_offset(pgd, 0);
1198                 remove_p4d_table(p4d, addr, next, altmap, direct);
1199         }
1200
1201         flush_tlb_all();
1202 }
1203
1204 void __ref vmemmap_free(unsigned long start, unsigned long end,
1205                 struct vmem_altmap *altmap)
1206 {
1207         remove_pagetable(start, end, false, altmap);
1208 }
1209
1210 static void __meminit
1211 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1212 {
1213         start = (unsigned long)__va(start);
1214         end = (unsigned long)__va(end);
1215
1216         remove_pagetable(start, end, true, NULL);
1217 }
1218
1219 void __ref arch_remove_memory(int nid, u64 start, u64 size,
1220                               struct vmem_altmap *altmap)
1221 {
1222         unsigned long start_pfn = start >> PAGE_SHIFT;
1223         unsigned long nr_pages = size >> PAGE_SHIFT;
1224
1225         __remove_pages(start_pfn, nr_pages, altmap);
1226         kernel_physical_mapping_remove(start, start + size);
1227 }
1228 #endif /* CONFIG_MEMORY_HOTPLUG */
1229
1230 static struct kcore_list kcore_vsyscall;
1231
1232 static void __init register_page_bootmem_info(void)
1233 {
1234 #ifdef CONFIG_NUMA
1235         int i;
1236
1237         for_each_online_node(i)
1238                 register_page_bootmem_info_node(NODE_DATA(i));
1239 #endif
1240 }
1241
1242 void __init mem_init(void)
1243 {
1244         pci_iommu_alloc();
1245
1246         /* clear_bss() already clear the empty_zero_page */
1247
1248         /* this will put all memory onto the freelists */
1249         memblock_free_all();
1250         after_bootmem = 1;
1251         x86_init.hyper.init_after_bootmem();
1252
1253         /*
1254          * Must be done after boot memory is put on freelist, because here we
1255          * might set fields in deferred struct pages that have not yet been
1256          * initialized, and memblock_free_all() initializes all the reserved
1257          * deferred pages for us.
1258          */
1259         register_page_bootmem_info();
1260
1261         /* Register memory areas for /proc/kcore */
1262         if (get_gate_vma(&init_mm))
1263                 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1264
1265         mem_init_print_info(NULL);
1266 }
1267
1268 int kernel_set_to_readonly;
1269
1270 void mark_rodata_ro(void)
1271 {
1272         unsigned long start = PFN_ALIGN(_text);
1273         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1274         unsigned long end = (unsigned long)__end_rodata_hpage_align;
1275         unsigned long text_end = PFN_ALIGN(_etext);
1276         unsigned long rodata_end = PFN_ALIGN(__end_rodata);
1277         unsigned long all_end;
1278
1279         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1280                (end - start) >> 10);
1281         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1282
1283         kernel_set_to_readonly = 1;
1284
1285         /*
1286          * The rodata/data/bss/brk section (but not the kernel text!)
1287          * should also be not-executable.
1288          *
1289          * We align all_end to PMD_SIZE because the existing mapping
1290          * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1291          * split the PMD and the reminder between _brk_end and the end
1292          * of the PMD will remain mapped executable.
1293          *
1294          * Any PMD which was setup after the one which covers _brk_end
1295          * has been zapped already via cleanup_highmem().
1296          */
1297         all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1298         set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1299
1300         set_ftrace_ops_ro();
1301
1302 #ifdef CONFIG_CPA_DEBUG
1303         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1304         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1305
1306         printk(KERN_INFO "Testing CPA: again\n");
1307         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1308 #endif
1309
1310         free_kernel_image_pages("unused kernel image (text/rodata gap)",
1311                                 (void *)text_end, (void *)rodata_start);
1312         free_kernel_image_pages("unused kernel image (rodata/data gap)",
1313                                 (void *)rodata_end, (void *)_sdata);
1314
1315         debug_checkwx();
1316 }
1317
1318 int kern_addr_valid(unsigned long addr)
1319 {
1320         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1321         pgd_t *pgd;
1322         p4d_t *p4d;
1323         pud_t *pud;
1324         pmd_t *pmd;
1325         pte_t *pte;
1326
1327         if (above != 0 && above != -1UL)
1328                 return 0;
1329
1330         pgd = pgd_offset_k(addr);
1331         if (pgd_none(*pgd))
1332                 return 0;
1333
1334         p4d = p4d_offset(pgd, addr);
1335         if (p4d_none(*p4d))
1336                 return 0;
1337
1338         pud = pud_offset(p4d, addr);
1339         if (pud_none(*pud))
1340                 return 0;
1341
1342         if (pud_large(*pud))
1343                 return pfn_valid(pud_pfn(*pud));
1344
1345         pmd = pmd_offset(pud, addr);
1346         if (pmd_none(*pmd))
1347                 return 0;
1348
1349         if (pmd_large(*pmd))
1350                 return pfn_valid(pmd_pfn(*pmd));
1351
1352         pte = pte_offset_kernel(pmd, addr);
1353         if (pte_none(*pte))
1354                 return 0;
1355
1356         return pfn_valid(pte_pfn(*pte));
1357 }
1358
1359 /*
1360  * Block size is the minimum amount of memory which can be hotplugged or
1361  * hotremoved. It must be power of two and must be equal or larger than
1362  * MIN_MEMORY_BLOCK_SIZE.
1363  */
1364 #define MAX_BLOCK_SIZE (2UL << 30)
1365
1366 /* Amount of ram needed to start using large blocks */
1367 #define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1368
1369 /* Adjustable memory block size */
1370 static unsigned long set_memory_block_size;
1371 int __init set_memory_block_size_order(unsigned int order)
1372 {
1373         unsigned long size = 1UL << order;
1374
1375         if (size > MEM_SIZE_FOR_LARGE_BLOCK || size < MIN_MEMORY_BLOCK_SIZE)
1376                 return -EINVAL;
1377
1378         set_memory_block_size = size;
1379         return 0;
1380 }
1381
1382 static unsigned long probe_memory_block_size(void)
1383 {
1384         unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1385         unsigned long bz;
1386
1387         /* If memory block size has been set, then use it */
1388         bz = set_memory_block_size;
1389         if (bz)
1390                 goto done;
1391
1392         /* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1393         if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1394                 bz = MIN_MEMORY_BLOCK_SIZE;
1395                 goto done;
1396         }
1397
1398         /* Find the largest allowed block size that aligns to memory end */
1399         for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1400                 if (IS_ALIGNED(boot_mem_end, bz))
1401                         break;
1402         }
1403 done:
1404         pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1405
1406         return bz;
1407 }
1408
1409 static unsigned long memory_block_size_probed;
1410 unsigned long memory_block_size_bytes(void)
1411 {
1412         if (!memory_block_size_probed)
1413                 memory_block_size_probed = probe_memory_block_size();
1414
1415         return memory_block_size_probed;
1416 }
1417
1418 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1419 /*
1420  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1421  */
1422 static long __meminitdata addr_start, addr_end;
1423 static void __meminitdata *p_start, *p_end;
1424 static int __meminitdata node_start;
1425
1426 static int __meminit vmemmap_populate_hugepages(unsigned long start,
1427                 unsigned long end, int node, struct vmem_altmap *altmap)
1428 {
1429         unsigned long addr;
1430         unsigned long next;
1431         pgd_t *pgd;
1432         p4d_t *p4d;
1433         pud_t *pud;
1434         pmd_t *pmd;
1435
1436         for (addr = start; addr < end; addr = next) {
1437                 next = pmd_addr_end(addr, end);
1438
1439                 pgd = vmemmap_pgd_populate(addr, node);
1440                 if (!pgd)
1441                         return -ENOMEM;
1442
1443                 p4d = vmemmap_p4d_populate(pgd, addr, node);
1444                 if (!p4d)
1445                         return -ENOMEM;
1446
1447                 pud = vmemmap_pud_populate(p4d, addr, node);
1448                 if (!pud)
1449                         return -ENOMEM;
1450
1451                 pmd = pmd_offset(pud, addr);
1452                 if (pmd_none(*pmd)) {
1453                         void *p;
1454
1455                         if (altmap)
1456                                 p = altmap_alloc_block_buf(PMD_SIZE, altmap);
1457                         else
1458                                 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1459                         if (p) {
1460                                 pte_t entry;
1461
1462                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1463                                                 PAGE_KERNEL_LARGE);
1464                                 set_pmd(pmd, __pmd(pte_val(entry)));
1465
1466                                 /* check to see if we have contiguous blocks */
1467                                 if (p_end != p || node_start != node) {
1468                                         if (p_start)
1469                                                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1470                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1471                                         addr_start = addr;
1472                                         node_start = node;
1473                                         p_start = p;
1474                                 }
1475
1476                                 addr_end = addr + PMD_SIZE;
1477                                 p_end = p + PMD_SIZE;
1478                                 continue;
1479                         } else if (altmap)
1480                                 return -ENOMEM; /* no fallback */
1481                 } else if (pmd_large(*pmd)) {
1482                         vmemmap_verify((pte_t *)pmd, node, addr, next);
1483                         continue;
1484                 }
1485                 if (vmemmap_populate_basepages(addr, next, node))
1486                         return -ENOMEM;
1487         }
1488         return 0;
1489 }
1490
1491 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1492                 struct vmem_altmap *altmap)
1493 {
1494         int err;
1495
1496         if (end - start < PAGES_PER_SECTION * sizeof(struct page))
1497                 err = vmemmap_populate_basepages(start, end, node);
1498         else if (boot_cpu_has(X86_FEATURE_PSE))
1499                 err = vmemmap_populate_hugepages(start, end, node, altmap);
1500         else if (altmap) {
1501                 pr_err_once("%s: no cpu support for altmap allocations\n",
1502                                 __func__);
1503                 err = -ENOMEM;
1504         } else
1505                 err = vmemmap_populate_basepages(start, end, node);
1506         if (!err)
1507                 sync_global_pgds(start, end - 1);
1508         return err;
1509 }
1510
1511 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1512 void register_page_bootmem_memmap(unsigned long section_nr,
1513                                   struct page *start_page, unsigned long nr_pages)
1514 {
1515         unsigned long addr = (unsigned long)start_page;
1516         unsigned long end = (unsigned long)(start_page + nr_pages);
1517         unsigned long next;
1518         pgd_t *pgd;
1519         p4d_t *p4d;
1520         pud_t *pud;
1521         pmd_t *pmd;
1522         unsigned int nr_pmd_pages;
1523         struct page *page;
1524
1525         for (; addr < end; addr = next) {
1526                 pte_t *pte = NULL;
1527
1528                 pgd = pgd_offset_k(addr);
1529                 if (pgd_none(*pgd)) {
1530                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1531                         continue;
1532                 }
1533                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1534
1535                 p4d = p4d_offset(pgd, addr);
1536                 if (p4d_none(*p4d)) {
1537                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1538                         continue;
1539                 }
1540                 get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1541
1542                 pud = pud_offset(p4d, addr);
1543                 if (pud_none(*pud)) {
1544                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1545                         continue;
1546                 }
1547                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1548
1549                 if (!boot_cpu_has(X86_FEATURE_PSE)) {
1550                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1551                         pmd = pmd_offset(pud, addr);
1552                         if (pmd_none(*pmd))
1553                                 continue;
1554                         get_page_bootmem(section_nr, pmd_page(*pmd),
1555                                          MIX_SECTION_INFO);
1556
1557                         pte = pte_offset_kernel(pmd, addr);
1558                         if (pte_none(*pte))
1559                                 continue;
1560                         get_page_bootmem(section_nr, pte_page(*pte),
1561                                          SECTION_INFO);
1562                 } else {
1563                         next = pmd_addr_end(addr, end);
1564
1565                         pmd = pmd_offset(pud, addr);
1566                         if (pmd_none(*pmd))
1567                                 continue;
1568
1569                         nr_pmd_pages = 1 << get_order(PMD_SIZE);
1570                         page = pmd_page(*pmd);
1571                         while (nr_pmd_pages--)
1572                                 get_page_bootmem(section_nr, page++,
1573                                                  SECTION_INFO);
1574                 }
1575         }
1576 }
1577 #endif
1578
1579 void __meminit vmemmap_populate_print_last(void)
1580 {
1581         if (p_start) {
1582                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1583                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1584                 p_start = NULL;
1585                 p_end = NULL;
1586                 node_start = 0;
1587         }
1588 }
1589 #endif