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
  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 /*
 221  * NOTE: This function is marked __ref because it calls __init function
 222  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 223  */
 224 static __ref void *spp_getpage(void)
 225 {
 226         void *ptr;
 227
 228         if (after_bootmem)
 229                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 230         else
 231                 ptr = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
 232
 233         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 234                 panic("set_pte_phys: cannot allocate page data %s\n",
 235                         after_bootmem ? "after bootmem" : "");
 236         }
 237
 238         pr_debug("spp_getpage %p\n", ptr);
 239
 240         return ptr;
 241 }
 242
 243 static p4d_t *fill_p4d(pgd_t *pgd, unsigned long vaddr)
 244 {
 245         if (pgd_none(*pgd)) {
 246                 p4d_t *p4d = (p4d_t *)spp_getpage();
 247                 pgd_populate(&init_mm, pgd, p4d);
 248                 if (p4d != p4d_offset(pgd, 0))
 249                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 250                                p4d, p4d_offset(pgd, 0));
 251         }
 252         return p4d_offset(pgd, vaddr);
 253 }
 254
 255 static pud_t *fill_pud(p4d_t *p4d, unsigned long vaddr)
 256 {
 257         if (p4d_none(*p4d)) {
 258                 pud_t *pud = (pud_t *)spp_getpage();
 259                 p4d_populate(&init_mm, p4d, pud);
 260                 if (pud != pud_offset(p4d, 0))
 261                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 262                                pud, pud_offset(p4d, 0));
 263         }
 264         return pud_offset(p4d, vaddr);
 265 }
 266
 267 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 268 {
 269         if (pud_none(*pud)) {
 270                 pmd_t *pmd = (pmd_t *) spp_getpage();
 271                 pud_populate(&init_mm, pud, pmd);
 272                 if (pmd != pmd_offset(pud, 0))
 273                         printk(KERN_ERR "PAGETABLE BUG #02! %p <-> %p\n",
 274                                pmd, pmd_offset(pud, 0));
 275         }
 276         return pmd_offset(pud, vaddr);
 277 }
 278
 279 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 280 {
 281         if (pmd_none(*pmd)) {
 282                 pte_t *pte = (pte_t *) spp_getpage();
 283                 pmd_populate_kernel(&init_mm, pmd, pte);
 284                 if (pte != pte_offset_kernel(pmd, 0))
 285                         printk(KERN_ERR "PAGETABLE BUG #03!\n");
 286         }
 287         return pte_offset_kernel(pmd, vaddr);
 288 }
 289
 290 static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
 291 {
 292         pmd_t *pmd = fill_pmd(pud, vaddr);
 293         pte_t *pte = fill_pte(pmd, vaddr);
 294
 295         set_pte(pte, new_pte);
 296
 297         /*
 298          * It's enough to flush this one mapping.
 299          * (PGE mappings get flushed as well)
 300          */
 301         __flush_tlb_one_kernel(vaddr);
 302 }
 303
 304 void set_pte_vaddr_p4d(p4d_t *p4d_page, unsigned long vaddr, pte_t new_pte)
 305 {
 306         p4d_t *p4d = p4d_page + p4d_index(vaddr);
 307         pud_t *pud = fill_pud(p4d, vaddr);
 308
 309         __set_pte_vaddr(pud, vaddr, new_pte);
 310 }
 311
 312 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 313 {
 314         pud_t *pud = pud_page + pud_index(vaddr);
 315
 316         __set_pte_vaddr(pud, vaddr, new_pte);
 317 }
 318
 319 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 320 {
 321         pgd_t *pgd;
 322         p4d_t *p4d_page;
 323
 324         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 325
 326         pgd = pgd_offset_k(vaddr);
 327         if (pgd_none(*pgd)) {
 328                 printk(KERN_ERR
 329                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 330                 return;
 331         }
 332
 333         p4d_page = p4d_offset(pgd, 0);
 334         set_pte_vaddr_p4d(p4d_page, vaddr, pteval);
 335 }
 336
 337 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 338 {
 339         pgd_t *pgd;
 340         p4d_t *p4d;
 341         pud_t *pud;
 342
 343         pgd = pgd_offset_k(vaddr);
 344         p4d = fill_p4d(pgd, vaddr);
 345         pud = fill_pud(p4d, vaddr);
 346         return fill_pmd(pud, vaddr);
 347 }
 348
 349 pte_t * __init populate_extra_pte(unsigned long vaddr)
 350 {
 351         pmd_t *pmd;
 352
 353         pmd = populate_extra_pmd(vaddr);
 354         return fill_pte(pmd, vaddr);
 355 }
 356
 357 /*
 358  * Create large page table mappings for a range of physical addresses.
 359  */
 360 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 361                                         enum page_cache_mode cache)
 362 {
 363         pgd_t *pgd;
 364         p4d_t *p4d;
 365         pud_t *pud;
 366         pmd_t *pmd;
 367         pgprot_t prot;
 368
 369         pgprot_val(prot) = pgprot_val(PAGE_KERNEL_LARGE) |
 370                 pgprot_val(pgprot_4k_2_large(cachemode2pgprot(cache)));
 371         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 372         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 373                 pgd = pgd_offset_k((unsigned long)__va(phys));
 374                 if (pgd_none(*pgd)) {
 375                         p4d = (p4d_t *) spp_getpage();
 376                         set_pgd(pgd, __pgd(__pa(p4d) | _KERNPG_TABLE |
 377                                                 _PAGE_USER));
 378                 }
 379                 p4d = p4d_offset(pgd, (unsigned long)__va(phys));
 380                 if (p4d_none(*p4d)) {
 381                         pud = (pud_t *) spp_getpage();
 382                         set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE |
 383                                                 _PAGE_USER));
 384                 }
 385                 pud = pud_offset(p4d, (unsigned long)__va(phys));
 386                 if (pud_none(*pud)) {
 387                         pmd = (pmd_t *) spp_getpage();
 388                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 389                                                 _PAGE_USER));
 390                 }
 391                 pmd = pmd_offset(pud, phys);
 392                 BUG_ON(!pmd_none(*pmd));
 393                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 394         }
 395 }
 396
 397 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 398 {
 399         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_WB);
 400 }
 401
 402 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 403 {
 404         __init_extra_mapping(phys, size, _PAGE_CACHE_MODE_UC);
 405 }
 406
 407 /*
 408  * The head.S code sets up the kernel high mapping:
 409  *
 410  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 411  *
 412  * phys_base holds the negative offset to the kernel, which is added
 413  * to the compile time generated pmds. This results in invalid pmds up
 414  * to the point where we hit the physaddr 0 mapping.
 415  *
 416  * We limit the mappings to the region from _text to _brk_end.  _brk_end
 417  * is rounded up to the 2MB boundary. This catches the invalid pmds as
 418  * well, as they are located before _text:
 419  */
 420 void __init cleanup_highmap(void)
 421 {
 422         unsigned long vaddr = __START_KERNEL_map;
 423         unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
 424         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 425         pmd_t *pmd = level2_kernel_pgt;
 426
 427         /*
 428          * Native path, max_pfn_mapped is not set yet.
 429          * Xen has valid max_pfn_mapped set in
 430          *      arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
 431          */
 432         if (max_pfn_mapped)
 433                 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 434
 435         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
 436                 if (pmd_none(*pmd))
 437                         continue;
 438                 if (vaddr < (unsigned long) _text || vaddr > end)
 439                         set_pmd(pmd, __pmd(0));
 440         }
 441 }
 442
 443 /*
 444  * Create PTE level page table mapping for physical addresses.
 445  * It returns the last physical address mapped.
 446  */
 447 static unsigned long __meminit
 448 phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
 449               pgprot_t prot, bool init)
 450 {
 451         unsigned long pages = 0, paddr_next;
 452         unsigned long paddr_last = paddr_end;
 453         pte_t *pte;
 454         int i;
 455
 456         pte = pte_page + pte_index(paddr);
 457         i = pte_index(paddr);
 458
 459         for (; i < PTRS_PER_PTE; i++, paddr = paddr_next, pte++) {
 460                 paddr_next = (paddr & PAGE_MASK) + PAGE_SIZE;
 461                 if (paddr >= paddr_end) {
 462                         if (!after_bootmem &&
 463                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 464                                              E820_TYPE_RAM) &&
 465                             !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
 466                                              E820_TYPE_RESERVED_KERN))
 467                                 set_pte_init(pte, __pte(0), init);
 468                         continue;
 469                 }
 470
 471                 /*
 472                  * We will re-use the existing mapping.
 473                  * Xen for example has some special requirements, like mapping
 474                  * pagetable pages as RO. So assume someone who pre-setup
 475                  * these mappings are more intelligent.
 476                  */
 477                 if (!pte_none(*pte)) {
 478                         if (!after_bootmem)
 479                                 pages++;
 480                         continue;
 481                 }
 482
 483                 if (0)
 484                         pr_info("   pte=%p addr=%lx pte=%016lx\n", pte, paddr,
 485                                 pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 486                 pages++;
 487                 set_pte_init(pte, pfn_pte(paddr >> PAGE_SHIFT, prot), init);
 488                 paddr_last = (paddr & PAGE_MASK) + PAGE_SIZE;
 489         }
 490
 491         update_page_count(PG_LEVEL_4K, pages);
 492
 493         return paddr_last;
 494 }
 495
 496 /*
 497  * Create PMD level page table mapping for physical addresses. The virtual
 498  * and physical address have to be aligned at this level.
 499  * It returns the last physical address mapped.
 500  */
 501 static unsigned long __meminit
 502 phys_pmd_init(pmd_t *pmd_page, unsigned long paddr, unsigned long paddr_end,
 503               unsigned long page_size_mask, pgprot_t prot, bool init)
 504 {
 505         unsigned long pages = 0, paddr_next;
 506         unsigned long paddr_last = paddr_end;
 507
 508         int i = pmd_index(paddr);
 509
 510         for (; i < PTRS_PER_PMD; i++, paddr = paddr_next) {
 511                 pmd_t *pmd = pmd_page + pmd_index(paddr);
 512                 pte_t *pte;
 513                 pgprot_t new_prot = prot;
 514
 515                 paddr_next = (paddr & PMD_MASK) + PMD_SIZE;
 516                 if (paddr >= paddr_end) {
 517                         if (!after_bootmem &&
 518                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 519                                              E820_TYPE_RAM) &&
 520                             !e820__mapped_any(paddr & PMD_MASK, paddr_next,
 521                                              E820_TYPE_RESERVED_KERN))
 522                                 set_pmd_init(pmd, __pmd(0), init);
 523                         continue;
 524                 }
 525
 526                 if (!pmd_none(*pmd)) {
 527                         if (!pmd_large(*pmd)) {
 528                                 spin_lock(&init_mm.page_table_lock);
 529                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
 530                                 paddr_last = phys_pte_init(pte, paddr,
 531                                                            paddr_end, prot,
 532                                                            init);
 533                                 spin_unlock(&init_mm.page_table_lock);
 534                                 continue;
 535                         }
 536                         /*
 537                          * If we are ok with PG_LEVEL_2M mapping, then we will
 538                          * use the existing mapping,
 539                          *
 540                          * Otherwise, we will split the large page mapping but
 541                          * use the same existing protection bits except for
 542                          * large page, so that we don't violate Intel's TLB
 543                          * Application note (317080) which says, while changing
 544                          * the page sizes, new and old translations should
 545                          * not differ with respect to page frame and
 546                          * attributes.
 547                          */
 548                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 549                                 if (!after_bootmem)
 550                                         pages++;
 551                                 paddr_last = paddr_next;
 552                                 continue;
 553                         }
 554                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 555                 }
 556
 557                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 558                         pages++;
 559                         spin_lock(&init_mm.page_table_lock);
 560                         set_pte_init((pte_t *)pmd,
 561                                      pfn_pte((paddr & PMD_MASK) >> PAGE_SHIFT,
 562                                              __pgprot(pgprot_val(prot) | _PAGE_PSE)),
 563                                      init);
 564                         spin_unlock(&init_mm.page_table_lock);
 565                         paddr_last = paddr_next;
 566                         continue;
 567                 }
 568
 569                 pte = alloc_low_page();
 570                 paddr_last = phys_pte_init(pte, paddr, paddr_end, new_prot, init);
 571
 572                 spin_lock(&init_mm.page_table_lock);
 573                 pmd_populate_kernel_init(&init_mm, pmd, pte, init);
 574                 spin_unlock(&init_mm.page_table_lock);
 575         }
 576         update_page_count(PG_LEVEL_2M, pages);
 577         return paddr_last;
 578 }
 579
 580 /*
 581  * Create PUD level page table mapping for physical addresses. The virtual
 582  * and physical address do not have to be aligned at this level. KASLR can
 583  * randomize virtual addresses up to this level.
 584  * It returns the last physical address mapped.
 585  */
 586 static unsigned long __meminit
 587 phys_pud_init(pud_t *pud_page, unsigned long paddr, unsigned long paddr_end,
 588               unsigned long page_size_mask, bool init)
 589 {
 590         unsigned long pages = 0, paddr_next;
 591         unsigned long paddr_last = paddr_end;
 592         unsigned long vaddr = (unsigned long)__va(paddr);
 593         int i = pud_index(vaddr);
 594
 595         for (; i < PTRS_PER_PUD; i++, paddr = paddr_next) {
 596                 pud_t *pud;
 597                 pmd_t *pmd;
 598                 pgprot_t prot = PAGE_KERNEL;
 599
 600                 vaddr = (unsigned long)__va(paddr);
 601                 pud = pud_page + pud_index(vaddr);
 602                 paddr_next = (paddr & PUD_MASK) + PUD_SIZE;
 603
 604                 if (paddr >= paddr_end) {
 605                         if (!after_bootmem &&
 606                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 607                                              E820_TYPE_RAM) &&
 608                             !e820__mapped_any(paddr & PUD_MASK, paddr_next,
 609                                              E820_TYPE_RESERVED_KERN))
 610                                 set_pud_init(pud, __pud(0), init);
 611                         continue;
 612                 }
 613
 614                 if (!pud_none(*pud)) {
 615                         if (!pud_large(*pud)) {
 616                                 pmd = pmd_offset(pud, 0);
 617                                 paddr_last = phys_pmd_init(pmd, paddr,
 618                                                            paddr_end,
 619                                                            page_size_mask,
 620                                                            prot, init);
 621                                 continue;
 622                         }
 623                         /*
 624                          * If we are ok with PG_LEVEL_1G mapping, then we will
 625                          * use the existing mapping.
 626                          *
 627                          * Otherwise, we will split the gbpage mapping but use
 628                          * the same existing protection  bits except for large
 629                          * page, so that we don't violate Intel's TLB
 630                          * Application note (317080) which says, while changing
 631                          * the page sizes, new and old translations should
 632                          * not differ with respect to page frame and
 633                          * attributes.
 634                          */
 635                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 636                                 if (!after_bootmem)
 637                                         pages++;
 638                                 paddr_last = paddr_next;
 639                                 continue;
 640                         }
 641                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 642                 }
 643
 644                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 645                         pages++;
 646                         spin_lock(&init_mm.page_table_lock);
 647                         set_pte_init((pte_t *)pud,
 648                                      pfn_pte((paddr & PUD_MASK) >> PAGE_SHIFT,
 649                                              PAGE_KERNEL_LARGE),
 650                                      init);
 651                         spin_unlock(&init_mm.page_table_lock);
 652                         paddr_last = paddr_next;
 653                         continue;
 654                 }
 655
 656                 pmd = alloc_low_page();
 657                 paddr_last = phys_pmd_init(pmd, paddr, paddr_end,
 658                                            page_size_mask, prot, init);
 659
 660                 spin_lock(&init_mm.page_table_lock);
 661                 pud_populate_init(&init_mm, pud, pmd, init);
 662                 spin_unlock(&init_mm.page_table_lock);
 663         }
 664
 665         update_page_count(PG_LEVEL_1G, pages);
 666
 667         return paddr_last;
 668 }
 669
 670 static unsigned long __meminit
 671 phys_p4d_init(p4d_t *p4d_page, unsigned long paddr, unsigned long paddr_end,
 672               unsigned long page_size_mask, bool init)
 673 {
 674         unsigned long vaddr, vaddr_end, vaddr_next, paddr_next, paddr_last;
 675
 676         paddr_last = paddr_end;
 677         vaddr = (unsigned long)__va(paddr);
 678         vaddr_end = (unsigned long)__va(paddr_end);
 679
 680         if (!pgtable_l5_enabled())
 681                 return phys_pud_init((pud_t *) p4d_page, paddr, paddr_end,
 682                                      page_size_mask, init);
 683
 684         for (; vaddr < vaddr_end; vaddr = vaddr_next) {
 685                 p4d_t *p4d = p4d_page + p4d_index(vaddr);
 686                 pud_t *pud;
 687
 688                 vaddr_next = (vaddr & P4D_MASK) + P4D_SIZE;
 689                 paddr = __pa(vaddr);
 690
 691                 if (paddr >= paddr_end) {
 692                         paddr_next = __pa(vaddr_next);
 693                         if (!after_bootmem &&
 694                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 695                                              E820_TYPE_RAM) &&
 696                             !e820__mapped_any(paddr & P4D_MASK, paddr_next,
 697                                              E820_TYPE_RESERVED_KERN))
 698                                 set_p4d_init(p4d, __p4d(0), init);
 699                         continue;
 700                 }
 701
 702                 if (!p4d_none(*p4d)) {
 703                         pud = pud_offset(p4d, 0);
 704                         paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
 705                                         page_size_mask, init);
 706                         continue;
 707                 }
 708
 709                 pud = alloc_low_page();
 710                 paddr_last = phys_pud_init(pud, paddr, __pa(vaddr_end),
 711                                            page_size_mask, init);
 712
 713                 spin_lock(&init_mm.page_table_lock);
 714                 p4d_populate_init(&init_mm, p4d, pud, init);
 715                 spin_unlock(&init_mm.page_table_lock);
 716         }
 717
 718         return paddr_last;
 719 }
 720
 721 static unsigned long __meminit
 722 __kernel_physical_mapping_init(unsigned long paddr_start,
 723                                unsigned long paddr_end,
 724                                unsigned long page_size_mask,
 725                                bool init)
 726 {
 727         bool pgd_changed = false;
 728         unsigned long vaddr, vaddr_start, vaddr_end, vaddr_next, paddr_last;
 729
 730         paddr_last = paddr_end;
 731         vaddr = (unsigned long)__va(paddr_start);
 732         vaddr_end = (unsigned long)__va(paddr_end);
 733         vaddr_start = vaddr;
 734
 735         for (; vaddr < vaddr_end; vaddr = vaddr_next) {
 736                 pgd_t *pgd = pgd_offset_k(vaddr);
 737                 p4d_t *p4d;
 738
 739                 vaddr_next = (vaddr & PGDIR_MASK) + PGDIR_SIZE;
 740
 741                 if (pgd_val(*pgd)) {
 742                         p4d = (p4d_t *)pgd_page_vaddr(*pgd);
 743                         paddr_last = phys_p4d_init(p4d, __pa(vaddr),
 744                                                    __pa(vaddr_end),
 745                                                    page_size_mask,
 746                                                    init);
 747                         continue;
 748                 }
 749
 750                 p4d = alloc_low_page();
 751                 paddr_last = phys_p4d_init(p4d, __pa(vaddr), __pa(vaddr_end),
 752                                            page_size_mask, init);
 753
 754                 spin_lock(&init_mm.page_table_lock);
 755                 if (pgtable_l5_enabled())
 756                         pgd_populate_init(&init_mm, pgd, p4d, init);
 757                 else
 758                         p4d_populate_init(&init_mm, p4d_offset(pgd, vaddr),
 759                                           (pud_t *) p4d, init);
 760
 761                 spin_unlock(&init_mm.page_table_lock);
 762                 pgd_changed = true;
 763         }
 764
 765         if (pgd_changed)
 766                 sync_global_pgds(vaddr_start, vaddr_end - 1);
 767
 768         return paddr_last;
 769 }
 770
 771
 772 /*
 773  * Create page table mapping for the physical memory for specific physical
 774  * addresses. Note that it can only be used to populate non-present entries.
 775  * The virtual and physical addresses have to be aligned on PMD level
 776  * down. It returns the last physical address mapped.
 777  */
 778 unsigned long __meminit
 779 kernel_physical_mapping_init(unsigned long paddr_start,
 780                              unsigned long paddr_end,
 781                              unsigned long page_size_mask)
 782 {
 783         return __kernel_physical_mapping_init(paddr_start, paddr_end,
 784                                               page_size_mask, true);
 785 }
 786
 787 /*
 788  * This function is similar to kernel_physical_mapping_init() above with the
 789  * exception that it uses set_{pud,pmd}() instead of the set_{pud,pte}_safe()
 790  * when updating the mapping. The caller is responsible to flush the TLBs after
 791  * the function returns.
 792  */
 793 unsigned long __meminit
 794 kernel_physical_mapping_change(unsigned long paddr_start,
 795                                unsigned long paddr_end,
 796                                unsigned long page_size_mask)
 797 {
 798         return __kernel_physical_mapping_init(paddr_start, paddr_end,
 799                                               page_size_mask, false);
 800 }
 801
 802 #ifndef CONFIG_NUMA
 803 void __init initmem_init(void)
 804 {
 805         memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
 806 }
 807 #endif
 808
 809 void __init paging_init(void)
 810 {
 811         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 812         sparse_init();
 813
 814         /*
 815          * clear the default setting with node 0
 816          * note: don't use nodes_clear here, that is really clearing when
 817          *       numa support is not compiled in, and later node_set_state
 818          *       will not set it back.
 819          */
 820         node_clear_state(0, N_MEMORY);
 821         if (N_MEMORY != N_NORMAL_MEMORY)
 822                 node_clear_state(0, N_NORMAL_MEMORY);
 823
 824         zone_sizes_init();
 825 }
 826
 827 /*
 828  * Memory hotplug specific functions
 829  */
 830 #ifdef CONFIG_MEMORY_HOTPLUG
 831 /*
 832  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 833  * updating.
 834  */
 835 static void update_end_of_memory_vars(u64 start, u64 size)
 836 {
 837         unsigned long end_pfn = PFN_UP(start + size);
 838
 839         if (end_pfn > max_pfn) {
 840                 max_pfn = end_pfn;
 841                 max_low_pfn = end_pfn;
 842                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 843         }
 844 }
 845
 846 int add_pages(int nid, unsigned long start_pfn, unsigned long nr_pages,
 847                                 struct mhp_restrictions *restrictions)
 848 {
 849         int ret;
 850
 851         ret = __add_pages(nid, start_pfn, nr_pages, restrictions);
 852         WARN_ON_ONCE(ret);
 853
 854         /* update max_pfn, max_low_pfn and high_memory */
 855         update_end_of_memory_vars(start_pfn << PAGE_SHIFT,
 856                                   nr_pages << PAGE_SHIFT);
 857
 858         return ret;
 859 }
 860
 861 int arch_add_memory(int nid, u64 start, u64 size,
 862                         struct mhp_restrictions *restrictions)
 863 {
 864         unsigned long start_pfn = start >> PAGE_SHIFT;
 865         unsigned long nr_pages = size >> PAGE_SHIFT;
 866
 867         init_memory_mapping(start, start + size);
 868
 869         return add_pages(nid, start_pfn, nr_pages, restrictions);
 870 }
 871
 872 #define PAGE_INUSE 0xFD
 873
 874 static void __meminit free_pagetable(struct page *page, int order)
 875 {
 876         unsigned long magic;
 877         unsigned int nr_pages = 1 << order;
 878
 879         /* bootmem page has reserved flag */
 880         if (PageReserved(page)) {
 881                 __ClearPageReserved(page);
 882
 883                 magic = (unsigned long)page->freelist;
 884                 if (magic == SECTION_INFO || magic == MIX_SECTION_INFO) {
 885                         while (nr_pages--)
 886                                 put_page_bootmem(page++);
 887                 } else
 888                         while (nr_pages--)
 889                                 free_reserved_page(page++);
 890         } else
 891                 free_pages((unsigned long)page_address(page), order);
 892 }
 893
 894 static void __meminit free_hugepage_table(struct page *page,
 895                 struct vmem_altmap *altmap)
 896 {
 897         if (altmap)
 898                 vmem_altmap_free(altmap, PMD_SIZE / PAGE_SIZE);
 899         else
 900                 free_pagetable(page, get_order(PMD_SIZE));
 901 }
 902
 903 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
 904 {
 905         pte_t *pte;
 906         int i;
 907
 908         for (i = 0; i < PTRS_PER_PTE; i++) {
 909                 pte = pte_start + i;
 910                 if (!pte_none(*pte))
 911                         return;
 912         }
 913
 914         /* free a pte talbe */
 915         free_pagetable(pmd_page(*pmd), 0);
 916         spin_lock(&init_mm.page_table_lock);
 917         pmd_clear(pmd);
 918         spin_unlock(&init_mm.page_table_lock);
 919 }
 920
 921 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud)
 922 {
 923         pmd_t *pmd;
 924         int i;
 925
 926         for (i = 0; i < PTRS_PER_PMD; i++) {
 927                 pmd = pmd_start + i;
 928                 if (!pmd_none(*pmd))
 929                         return;
 930         }
 931
 932         /* free a pmd talbe */
 933         free_pagetable(pud_page(*pud), 0);
 934         spin_lock(&init_mm.page_table_lock);
 935         pud_clear(pud);
 936         spin_unlock(&init_mm.page_table_lock);
 937 }
 938
 939 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
 940 {
 941         pud_t *pud;
 942         int i;
 943
 944         for (i = 0; i < PTRS_PER_PUD; i++) {
 945                 pud = pud_start + i;
 946                 if (!pud_none(*pud))
 947                         return;
 948         }
 949
 950         /* free a pud talbe */
 951         free_pagetable(p4d_page(*p4d), 0);
 952         spin_lock(&init_mm.page_table_lock);
 953         p4d_clear(p4d);
 954         spin_unlock(&init_mm.page_table_lock);
 955 }
 956
 957 static void __meminit
 958 remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
 959                  bool direct)
 960 {
 961         unsigned long next, pages = 0;
 962         pte_t *pte;
 963         void *page_addr;
 964         phys_addr_t phys_addr;
 965
 966         pte = pte_start + pte_index(addr);
 967         for (; addr < end; addr = next, pte++) {
 968                 next = (addr + PAGE_SIZE) & PAGE_MASK;
 969                 if (next > end)
 970                         next = end;
 971
 972                 if (!pte_present(*pte))
 973                         continue;
 974
 975                 /*
 976                  * We mapped [0,1G) memory as identity mapping when
 977                  * initializing, in arch/x86/kernel/head_64.S. These
 978                  * pagetables cannot be removed.
 979                  */
 980                 phys_addr = pte_val(*pte) + (addr & PAGE_MASK);
 981                 if (phys_addr < (phys_addr_t)0x40000000)
 982                         return;
 983
 984                 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
 985                         /*
 986                          * Do not free direct mapping pages since they were
 987                          * freed when offlining, or simplely not in use.
 988                          */
 989                         if (!direct)
 990                                 free_pagetable(pte_page(*pte), 0);
 991
 992                         spin_lock(&init_mm.page_table_lock);
 993                         pte_clear(&init_mm, addr, pte);
 994                         spin_unlock(&init_mm.page_table_lock);
 995
 996                         /* For non-direct mapping, pages means nothing. */
 997                         pages++;
 998                 } else {
 999                         /*
1000                          * If we are here, we are freeing vmemmap pages since
1001                          * direct mapped memory ranges to be freed are aligned.
1002                          *
1003                          * If we are not removing the whole page, it means
1004                          * other page structs in this page are being used and
1005                          * we canot remove them. So fill the unused page_structs
1006                          * with 0xFD, and remove the page when it is wholly
1007                          * filled with 0xFD.
1008                          */
1009                         memset((void *)addr, PAGE_INUSE, next - addr);
1010
1011                         page_addr = page_address(pte_page(*pte));
1012                         if (!memchr_inv(page_addr, PAGE_INUSE, PAGE_SIZE)) {
1013                                 free_pagetable(pte_page(*pte), 0);
1014
1015                                 spin_lock(&init_mm.page_table_lock);
1016                                 pte_clear(&init_mm, addr, pte);
1017                                 spin_unlock(&init_mm.page_table_lock);
1018                         }
1019                 }
1020         }
1021
1022         /* Call free_pte_table() in remove_pmd_table(). */
1023         flush_tlb_all();
1024         if (direct)
1025                 update_page_count(PG_LEVEL_4K, -pages);
1026 }
1027
1028 static void __meminit
1029 remove_pmd_table(pmd_t *pmd_start, unsigned long addr, unsigned long end,
1030                  bool direct, struct vmem_altmap *altmap)
1031 {
1032         unsigned long next, pages = 0;
1033         pte_t *pte_base;
1034         pmd_t *pmd;
1035         void *page_addr;
1036
1037         pmd = pmd_start + pmd_index(addr);
1038         for (; addr < end; addr = next, pmd++) {
1039                 next = pmd_addr_end(addr, end);
1040
1041                 if (!pmd_present(*pmd))
1042                         continue;
1043
1044                 if (pmd_large(*pmd)) {
1045                         if (IS_ALIGNED(addr, PMD_SIZE) &&
1046                             IS_ALIGNED(next, PMD_SIZE)) {
1047                                 if (!direct)
1048                                         free_hugepage_table(pmd_page(*pmd),
1049                                                             altmap);
1050
1051                                 spin_lock(&init_mm.page_table_lock);
1052                                 pmd_clear(pmd);
1053                                 spin_unlock(&init_mm.page_table_lock);
1054                                 pages++;
1055                         } else {
1056                                 /* If here, we are freeing vmemmap pages. */
1057                                 memset((void *)addr, PAGE_INUSE, next - addr);
1058
1059                                 page_addr = page_address(pmd_page(*pmd));
1060                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1061                                                 PMD_SIZE)) {
1062                                         free_hugepage_table(pmd_page(*pmd),
1063                                                             altmap);
1064
1065                                         spin_lock(&init_mm.page_table_lock);
1066                                         pmd_clear(pmd);
1067                                         spin_unlock(&init_mm.page_table_lock);
1068                                 }
1069                         }
1070
1071                         continue;
1072                 }
1073
1074                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
1075                 remove_pte_table(pte_base, addr, next, direct);
1076                 free_pte_table(pte_base, pmd);
1077         }
1078
1079         /* Call free_pmd_table() in remove_pud_table(). */
1080         if (direct)
1081                 update_page_count(PG_LEVEL_2M, -pages);
1082 }
1083
1084 static void __meminit
1085 remove_pud_table(pud_t *pud_start, unsigned long addr, unsigned long end,
1086                  struct vmem_altmap *altmap, bool direct)
1087 {
1088         unsigned long next, pages = 0;
1089         pmd_t *pmd_base;
1090         pud_t *pud;
1091         void *page_addr;
1092
1093         pud = pud_start + pud_index(addr);
1094         for (; addr < end; addr = next, pud++) {
1095                 next = pud_addr_end(addr, end);
1096
1097                 if (!pud_present(*pud))
1098                         continue;
1099
1100                 if (pud_large(*pud)) {
1101                         if (IS_ALIGNED(addr, PUD_SIZE) &&
1102                             IS_ALIGNED(next, PUD_SIZE)) {
1103                                 if (!direct)
1104                                         free_pagetable(pud_page(*pud),
1105                                                        get_order(PUD_SIZE));
1106
1107                                 spin_lock(&init_mm.page_table_lock);
1108                                 pud_clear(pud);
1109                                 spin_unlock(&init_mm.page_table_lock);
1110                                 pages++;
1111                         } else {
1112                                 /* If here, we are freeing vmemmap pages. */
1113                                 memset((void *)addr, PAGE_INUSE, next - addr);
1114
1115                                 page_addr = page_address(pud_page(*pud));
1116                                 if (!memchr_inv(page_addr, PAGE_INUSE,
1117                                                 PUD_SIZE)) {
1118                                         free_pagetable(pud_page(*pud),
1119                                                        get_order(PUD_SIZE));
1120
1121                                         spin_lock(&init_mm.page_table_lock);
1122                                         pud_clear(pud);
1123                                         spin_unlock(&init_mm.page_table_lock);
1124                                 }
1125                         }
1126
1127                         continue;
1128                 }
1129
1130                 pmd_base = pmd_offset(pud, 0);
1131                 remove_pmd_table(pmd_base, addr, next, direct, altmap);
1132                 free_pmd_table(pmd_base, pud);
1133         }
1134
1135         if (direct)
1136                 update_page_count(PG_LEVEL_1G, -pages);
1137 }
1138
1139 static void __meminit
1140 remove_p4d_table(p4d_t *p4d_start, unsigned long addr, unsigned long end,
1141                  struct vmem_altmap *altmap, bool direct)
1142 {
1143         unsigned long next, pages = 0;
1144         pud_t *pud_base;
1145         p4d_t *p4d;
1146
1147         p4d = p4d_start + p4d_index(addr);
1148         for (; addr < end; addr = next, p4d++) {
1149                 next = p4d_addr_end(addr, end);
1150
1151                 if (!p4d_present(*p4d))
1152                         continue;
1153
1154                 BUILD_BUG_ON(p4d_large(*p4d));
1155
1156                 pud_base = pud_offset(p4d, 0);
1157                 remove_pud_table(pud_base, addr, next, altmap, direct);
1158                 /*
1159                  * For 4-level page tables we do not want to free PUDs, but in the
1160                  * 5-level case we should free them. This code will have to change
1161                  * to adapt for boot-time switching between 4 and 5 level page tables.
1162                  */
1163                 if (pgtable_l5_enabled())
1164                         free_pud_table(pud_base, p4d);
1165         }
1166
1167         if (direct)
1168                 update_page_count(PG_LEVEL_512G, -pages);
1169 }
1170
1171 /* start and end are both virtual address. */
1172 static void __meminit
1173 remove_pagetable(unsigned long start, unsigned long end, bool direct,
1174                 struct vmem_altmap *altmap)
1175 {
1176         unsigned long next;
1177         unsigned long addr;
1178         pgd_t *pgd;
1179         p4d_t *p4d;
1180
1181         for (addr = start; addr < end; addr = next) {
1182                 next = pgd_addr_end(addr, end);
1183
1184                 pgd = pgd_offset_k(addr);
1185                 if (!pgd_present(*pgd))
1186                         continue;
1187
1188                 p4d = p4d_offset(pgd, 0);
1189                 remove_p4d_table(p4d, addr, next, altmap, direct);
1190         }
1191
1192         flush_tlb_all();
1193 }
1194
1195 void __ref vmemmap_free(unsigned long start, unsigned long end,
1196                 struct vmem_altmap *altmap)
1197 {
1198         remove_pagetable(start, end, false, altmap);
1199 }
1200
1201 static void __meminit
1202 kernel_physical_mapping_remove(unsigned long start, unsigned long end)
1203 {
1204         start = (unsigned long)__va(start);
1205         end = (unsigned long)__va(end);
1206
1207         remove_pagetable(start, end, true, NULL);
1208 }
1209
1210 void __ref arch_remove_memory(int nid, u64 start, u64 size,
1211                               struct vmem_altmap *altmap)
1212 {
1213         unsigned long start_pfn = start >> PAGE_SHIFT;
1214         unsigned long nr_pages = size >> PAGE_SHIFT;
1215
1216         __remove_pages(start_pfn, nr_pages, altmap);
1217         kernel_physical_mapping_remove(start, start + size);
1218 }
1219 #endif /* CONFIG_MEMORY_HOTPLUG */
1220
1221 static struct kcore_list kcore_vsyscall;
1222
1223 static void __init register_page_bootmem_info(void)
1224 {
1225 #ifdef CONFIG_NUMA
1226         int i;
1227
1228         for_each_online_node(i)
1229                 register_page_bootmem_info_node(NODE_DATA(i));
1230 #endif
1231 }
1232
1233 void __init mem_init(void)
1234 {
1235         pci_iommu_alloc();
1236
1237         /* clear_bss() already clear the empty_zero_page */
1238
1239         /* this will put all memory onto the freelists */
1240         memblock_free_all();
1241         after_bootmem = 1;
1242         x86_init.hyper.init_after_bootmem();
1243
1244         /*
1245          * Must be done after boot memory is put on freelist, because here we
1246          * might set fields in deferred struct pages that have not yet been
1247          * initialized, and memblock_free_all() initializes all the reserved
1248          * deferred pages for us.
1249          */
1250         register_page_bootmem_info();
1251
1252         /* Register memory areas for /proc/kcore */
1253         if (get_gate_vma(&init_mm))
1254                 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_ADDR, PAGE_SIZE, KCORE_USER);
1255
1256         mem_init_print_info(NULL);
1257 }
1258
1259 int kernel_set_to_readonly;
1260
1261 void set_kernel_text_rw(void)
1262 {
1263         unsigned long start = PFN_ALIGN(_text);
1264         unsigned long end = PFN_ALIGN(__stop___ex_table);
1265
1266         if (!kernel_set_to_readonly)
1267                 return;
1268
1269         pr_debug("Set kernel text: %lx - %lx for read write\n",
1270                  start, end);
1271
1272         /*
1273          * Make the kernel identity mapping for text RW. Kernel text
1274          * mapping will always be RO. Refer to the comment in
1275          * static_protections() in pageattr.c
1276          */
1277         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
1278 }
1279
1280 void set_kernel_text_ro(void)
1281 {
1282         unsigned long start = PFN_ALIGN(_text);
1283         unsigned long end = PFN_ALIGN(__stop___ex_table);
1284
1285         if (!kernel_set_to_readonly)
1286                 return;
1287
1288         pr_debug("Set kernel text: %lx - %lx for read only\n",
1289                  start, end);
1290
1291         /*
1292          * Set the kernel identity mapping for text RO.
1293          */
1294         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1295 }
1296
1297 void mark_rodata_ro(void)
1298 {
1299         unsigned long start = PFN_ALIGN(_text);
1300         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
1301         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
1302         unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
1303         unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
1304         unsigned long all_end;
1305
1306         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
1307                (end - start) >> 10);
1308         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
1309
1310         kernel_set_to_readonly = 1;
1311
1312         /*
1313          * The rodata/data/bss/brk section (but not the kernel text!)
1314          * should also be not-executable.
1315          *
1316          * We align all_end to PMD_SIZE because the existing mapping
1317          * is a full PMD. If we would align _brk_end to PAGE_SIZE we
1318          * split the PMD and the reminder between _brk_end and the end
1319          * of the PMD will remain mapped executable.
1320          *
1321          * Any PMD which was setup after the one which covers _brk_end
1322          * has been zapped already via cleanup_highmem().
1323          */
1324         all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
1325         set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
1326
1327 #ifdef CONFIG_CPA_DEBUG
1328         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
1329         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
1330
1331         printk(KERN_INFO "Testing CPA: again\n");
1332         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
1333 #endif
1334
1335         free_kernel_image_pages((void *)text_end, (void *)rodata_start);
1336         free_kernel_image_pages((void *)rodata_end, (void *)_sdata);
1337
1338         debug_checkwx();
1339 }
1340
1341 int kern_addr_valid(unsigned long addr)
1342 {
1343         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1344         pgd_t *pgd;
1345         p4d_t *p4d;
1346         pud_t *pud;
1347         pmd_t *pmd;
1348         pte_t *pte;
1349
1350         if (above != 0 && above != -1UL)
1351                 return 0;
1352
1353         pgd = pgd_offset_k(addr);
1354         if (pgd_none(*pgd))
1355                 return 0;
1356
1357         p4d = p4d_offset(pgd, addr);
1358         if (!p4d_present(*p4d))
1359                 return 0;
1360
1361         pud = pud_offset(p4d, addr);
1362         if (!pud_present(*pud))
1363                 return 0;
1364
1365         if (pud_large(*pud))
1366                 return pfn_valid(pud_pfn(*pud));
1367
1368         pmd = pmd_offset(pud, addr);
1369         if (!pmd_present(*pmd))
1370                 return 0;
1371
1372         if (pmd_large(*pmd))
1373                 return pfn_valid(pmd_pfn(*pmd));
1374
1375         pte = pte_offset_kernel(pmd, addr);
1376         if (pte_none(*pte))
1377                 return 0;
1378
1379         return pfn_valid(pte_pfn(*pte));
1380 }
1381
1382 /*
1383  * Block size is the minimum amount of memory which can be hotplugged or
1384  * hotremoved. It must be power of two and must be equal or larger than
1385  * MIN_MEMORY_BLOCK_SIZE.
1386  */
1387 #define MAX_BLOCK_SIZE (2UL << 30)
1388
1389 /* Amount of ram needed to start using large blocks */
1390 #define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
1391
1392 /* Adjustable memory block size */
1393 static unsigned long set_memory_block_size;
1394 int __init set_memory_block_size_order(unsigned int order)
1395 {
1396         unsigned long size = 1UL << order;
1397
1398         if (size > MEM_SIZE_FOR_LARGE_BLOCK || size < MIN_MEMORY_BLOCK_SIZE)
1399                 return -EINVAL;
1400
1401         set_memory_block_size = size;
1402         return 0;
1403 }
1404
1405 static unsigned long probe_memory_block_size(void)
1406 {
1407         unsigned long boot_mem_end = max_pfn << PAGE_SHIFT;
1408         unsigned long bz;
1409
1410         /* If memory block size has been set, then use it */
1411         bz = set_memory_block_size;
1412         if (bz)
1413                 goto done;
1414
1415         /* Use regular block if RAM is smaller than MEM_SIZE_FOR_LARGE_BLOCK */
1416         if (boot_mem_end < MEM_SIZE_FOR_LARGE_BLOCK) {
1417                 bz = MIN_MEMORY_BLOCK_SIZE;
1418                 goto done;
1419         }
1420
1421         /* Find the largest allowed block size that aligns to memory end */
1422         for (bz = MAX_BLOCK_SIZE; bz > MIN_MEMORY_BLOCK_SIZE; bz >>= 1) {
1423                 if (IS_ALIGNED(boot_mem_end, bz))
1424                         break;
1425         }
1426 done:
1427         pr_info("x86/mm: Memory block size: %ldMB\n", bz >> 20);
1428
1429         return bz;
1430 }
1431
1432 static unsigned long memory_block_size_probed;
1433 unsigned long memory_block_size_bytes(void)
1434 {
1435         if (!memory_block_size_probed)
1436                 memory_block_size_probed = probe_memory_block_size();
1437
1438         return memory_block_size_probed;
1439 }
1440
1441 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1442 /*
1443  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1444  */
1445 static long __meminitdata addr_start, addr_end;
1446 static void __meminitdata *p_start, *p_end;
1447 static int __meminitdata node_start;
1448
1449 static int __meminit vmemmap_populate_hugepages(unsigned long start,
1450                 unsigned long end, int node, struct vmem_altmap *altmap)
1451 {
1452         unsigned long addr;
1453         unsigned long next;
1454         pgd_t *pgd;
1455         p4d_t *p4d;
1456         pud_t *pud;
1457         pmd_t *pmd;
1458
1459         for (addr = start; addr < end; addr = next) {
1460                 next = pmd_addr_end(addr, end);
1461
1462                 pgd = vmemmap_pgd_populate(addr, node);
1463                 if (!pgd)
1464                         return -ENOMEM;
1465
1466                 p4d = vmemmap_p4d_populate(pgd, addr, node);
1467                 if (!p4d)
1468                         return -ENOMEM;
1469
1470                 pud = vmemmap_pud_populate(p4d, addr, node);
1471                 if (!pud)
1472                         return -ENOMEM;
1473
1474                 pmd = pmd_offset(pud, addr);
1475                 if (pmd_none(*pmd)) {
1476                         void *p;
1477
1478                         if (altmap)
1479                                 p = altmap_alloc_block_buf(PMD_SIZE, altmap);
1480                         else
1481                                 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1482                         if (p) {
1483                                 pte_t entry;
1484
1485                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1486                                                 PAGE_KERNEL_LARGE);
1487                                 set_pmd(pmd, __pmd(pte_val(entry)));
1488
1489                                 /* check to see if we have contiguous blocks */
1490                                 if (p_end != p || node_start != node) {
1491                                         if (p_start)
1492                                                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1493                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1494                                         addr_start = addr;
1495                                         node_start = node;
1496                                         p_start = p;
1497                                 }
1498
1499                                 addr_end = addr + PMD_SIZE;
1500                                 p_end = p + PMD_SIZE;
1501                                 continue;
1502                         } else if (altmap)
1503                                 return -ENOMEM; /* no fallback */
1504                 } else if (pmd_large(*pmd)) {
1505                         vmemmap_verify((pte_t *)pmd, node, addr, next);
1506                         continue;
1507                 }
1508                 if (vmemmap_populate_basepages(addr, next, node))
1509                         return -ENOMEM;
1510         }
1511         return 0;
1512 }
1513
1514 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1515                 struct vmem_altmap *altmap)
1516 {
1517         int err;
1518
1519         if (end - start < PAGES_PER_SECTION * sizeof(struct page))
1520                 err = vmemmap_populate_basepages(start, end, node);
1521         else if (boot_cpu_has(X86_FEATURE_PSE))
1522                 err = vmemmap_populate_hugepages(start, end, node, altmap);
1523         else if (altmap) {
1524                 pr_err_once("%s: no cpu support for altmap allocations\n",
1525                                 __func__);
1526                 err = -ENOMEM;
1527         } else
1528                 err = vmemmap_populate_basepages(start, end, node);
1529         if (!err)
1530                 sync_global_pgds(start, end - 1);
1531         return err;
1532 }
1533
1534 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1535 void register_page_bootmem_memmap(unsigned long section_nr,
1536                                   struct page *start_page, unsigned long nr_pages)
1537 {
1538         unsigned long addr = (unsigned long)start_page;
1539         unsigned long end = (unsigned long)(start_page + nr_pages);
1540         unsigned long next;
1541         pgd_t *pgd;
1542         p4d_t *p4d;
1543         pud_t *pud;
1544         pmd_t *pmd;
1545         unsigned int nr_pmd_pages;
1546         struct page *page;
1547
1548         for (; addr < end; addr = next) {
1549                 pte_t *pte = NULL;
1550
1551                 pgd = pgd_offset_k(addr);
1552                 if (pgd_none(*pgd)) {
1553                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1554                         continue;
1555                 }
1556                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1557
1558                 p4d = p4d_offset(pgd, addr);
1559                 if (p4d_none(*p4d)) {
1560                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1561                         continue;
1562                 }
1563                 get_page_bootmem(section_nr, p4d_page(*p4d), MIX_SECTION_INFO);
1564
1565                 pud = pud_offset(p4d, addr);
1566                 if (pud_none(*pud)) {
1567                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1568                         continue;
1569                 }
1570                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1571
1572                 if (!boot_cpu_has(X86_FEATURE_PSE)) {
1573                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1574                         pmd = pmd_offset(pud, addr);
1575                         if (pmd_none(*pmd))
1576                                 continue;
1577                         get_page_bootmem(section_nr, pmd_page(*pmd),
1578                                          MIX_SECTION_INFO);
1579
1580                         pte = pte_offset_kernel(pmd, addr);
1581                         if (pte_none(*pte))
1582                                 continue;
1583                         get_page_bootmem(section_nr, pte_page(*pte),
1584                                          SECTION_INFO);
1585                 } else {
1586                         next = pmd_addr_end(addr, end);
1587
1588                         pmd = pmd_offset(pud, addr);
1589                         if (pmd_none(*pmd))
1590                                 continue;
1591
1592                         nr_pmd_pages = 1 << get_order(PMD_SIZE);
1593                         page = pmd_page(*pmd);
1594                         while (nr_pmd_pages--)
1595                                 get_page_bootmem(section_nr, page++,
1596                                                  SECTION_INFO);
1597                 }
1598         }
1599 }
1600 #endif
1601
1602 void __meminit vmemmap_populate_print_last(void)
1603 {
1604         if (p_start) {
1605                 pr_debug(" [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1606                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1607                 p_start = NULL;
1608                 p_end = NULL;
1609                 node_start = 0;
1610         }
1611 }
1612 #endif