arch/x86/mm/pageattr.c

   1 /*
   2  * Copyright 2002 Andi Kleen, SuSE Labs.
   3  * Thanks to Ben LaHaise for precious feedback.
   4  */
   5 #include <linux/highmem.h>
   6 #include <linux/bootmem.h>
   7 #include <linux/sched.h>
   8 #include <linux/mm.h>
   9 #include <linux/interrupt.h>
  10 #include <linux/seq_file.h>
  11 #include <linux/debugfs.h>
  12 #include <linux/pfn.h>
  13 #include <linux/percpu.h>
  14 #include <linux/gfp.h>
  15 #include <linux/pci.h>
  16 #include <linux/vmalloc.h>
  17
  18 #include <asm/e820.h>
  19 #include <asm/processor.h>
  20 #include <asm/tlbflush.h>
  21 #include <asm/sections.h>
  22 #include <asm/setup.h>
  23 #include <asm/uaccess.h>
  24 #include <asm/pgalloc.h>
  25 #include <asm/proto.h>
  26 #include <asm/pat.h>
  27
  28 /*
  29  * The current flushing context - we pass it instead of 5 arguments:
  30  */
  31 struct cpa_data {
  32         unsigned long   *vaddr;
  33         pgd_t           *pgd;
  34         pgprot_t        mask_set;
  35         pgprot_t        mask_clr;
  36         unsigned long   numpages;
  37         int             flags;
  38         unsigned long   pfn;
  39         unsigned        force_split : 1;
  40         int             curpage;
  41         struct page     **pages;
  42 };
  43
  44 /*
  45  * Serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity mappings)
  46  * using cpa_lock. So that we don't allow any other cpu, with stale large tlb
  47  * entries change the page attribute in parallel to some other cpu
  48  * splitting a large page entry along with changing the attribute.
  49  */
  50 static DEFINE_SPINLOCK(cpa_lock);
  51
  52 #define CPA_FLUSHTLB 1
  53 #define CPA_ARRAY 2
  54 #define CPA_PAGES_ARRAY 4
  55
  56 #ifdef CONFIG_PROC_FS
  57 static unsigned long direct_pages_count[PG_LEVEL_NUM];
  58
  59 void update_page_count(int level, unsigned long pages)
  60 {
  61         /* Protect against CPA */
  62         spin_lock(&pgd_lock);
  63         direct_pages_count[level] += pages;
  64         spin_unlock(&pgd_lock);
  65 }
  66
  67 static void split_page_count(int level)
  68 {
  69         if (direct_pages_count[level] == 0)
  70                 return;
  71
  72         direct_pages_count[level]--;
  73         direct_pages_count[level - 1] += PTRS_PER_PTE;
  74 }
  75
  76 void arch_report_meminfo(struct seq_file *m)
  77 {
  78         seq_printf(m, "DirectMap4k:    %8lu kB\n",
  79                         direct_pages_count[PG_LEVEL_4K] << 2);
  80 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
  81         seq_printf(m, "DirectMap2M:    %8lu kB\n",
  82                         direct_pages_count[PG_LEVEL_2M] << 11);
  83 #else
  84         seq_printf(m, "DirectMap4M:    %8lu kB\n",
  85                         direct_pages_count[PG_LEVEL_2M] << 12);
  86 #endif
  87         if (direct_gbpages)
  88                 seq_printf(m, "DirectMap1G:    %8lu kB\n",
  89                         direct_pages_count[PG_LEVEL_1G] << 20);
  90 }
  91 #else
  92 static inline void split_page_count(int level) { }
  93 #endif
  94
  95 #ifdef CONFIG_X86_64
  96
  97 static inline unsigned long highmap_start_pfn(void)
  98 {
  99         return __pa_symbol(_text) >> PAGE_SHIFT;
 100 }
 101
 102 static inline unsigned long highmap_end_pfn(void)
 103 {
 104         /* Do not reference physical address outside the kernel. */
 105         return __pa_symbol(roundup(_brk_end, PMD_SIZE) - 1) >> PAGE_SHIFT;
 106 }
 107
 108 #endif
 109
 110 static inline int
 111 within(unsigned long addr, unsigned long start, unsigned long end)
 112 {
 113         return addr >= start && addr < end;
 114 }
 115
 116 static inline int
 117 within_inclusive(unsigned long addr, unsigned long start, unsigned long end)
 118 {
 119         return addr >= start && addr <= end;
 120 }
 121
 122 /*
 123  * Flushing functions
 124  */
 125
 126 /**
 127  * clflush_cache_range - flush a cache range with clflush
 128  * @vaddr:      virtual start address
 129  * @size:       number of bytes to flush
 130  *
 131  * clflushopt is an unordered instruction which needs fencing with mfence or
 132  * sfence to avoid ordering issues.
 133  */
 134 void clflush_cache_range(void *vaddr, unsigned int size)
 135 {
 136         const unsigned long clflush_size = boot_cpu_data.x86_clflush_size;
 137         void *p = (void *)((unsigned long)vaddr & ~(clflush_size - 1));
 138         void *vend = vaddr + size;
 139
 140         if (p >= vend)
 141                 return;
 142
 143         mb();
 144
 145         for (; p < vend; p += clflush_size)
 146                 clflushopt(p);
 147
 148         mb();
 149 }
 150 EXPORT_SYMBOL_GPL(clflush_cache_range);
 151
 152 static void __cpa_flush_all(void *arg)
 153 {
 154         unsigned long cache = (unsigned long)arg;
 155
 156         /*
 157          * Flush all to work around Errata in early athlons regarding
 158          * large page flushing.
 159          */
 160         __flush_tlb_all();
 161
 162         if (cache && boot_cpu_data.x86 >= 4)
 163                 wbinvd();
 164 }
 165
 166 static void cpa_flush_all(unsigned long cache)
 167 {
 168         BUG_ON(irqs_disabled());
 169
 170         on_each_cpu(__cpa_flush_all, (void *) cache, 1);
 171 }
 172
 173 static void __cpa_flush_range(void *arg)
 174 {
 175         /*
 176          * We could optimize that further and do individual per page
 177          * tlb invalidates for a low number of pages. Caveat: we must
 178          * flush the high aliases on 64bit as well.
 179          */
 180         __flush_tlb_all();
 181 }
 182
 183 static void cpa_flush_range(unsigned long start, int numpages, int cache)
 184 {
 185         unsigned int i, level;
 186         unsigned long addr;
 187
 188         BUG_ON(irqs_disabled());
 189         WARN_ON(PAGE_ALIGN(start) != start);
 190
 191         on_each_cpu(__cpa_flush_range, NULL, 1);
 192
 193         if (!cache)
 194                 return;
 195
 196         /*
 197          * We only need to flush on one CPU,
 198          * clflush is a MESI-coherent instruction that
 199          * will cause all other CPUs to flush the same
 200          * cachelines:
 201          */
 202         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
 203                 pte_t *pte = lookup_address(addr, &level);
 204
 205                 /*
 206                  * Only flush present addresses:
 207                  */
 208                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 209                         clflush_cache_range((void *) addr, PAGE_SIZE);
 210         }
 211 }
 212
 213 static void cpa_flush_array(unsigned long *start, int numpages, int cache,
 214                             int in_flags, struct page **pages)
 215 {
 216         unsigned int i, level;
 217         unsigned long do_wbinvd = cache && numpages >= 1024; /* 4M threshold */
 218
 219         BUG_ON(irqs_disabled());
 220
 221         on_each_cpu(__cpa_flush_all, (void *) do_wbinvd, 1);
 222
 223         if (!cache || do_wbinvd)
 224                 return;
 225
 226         /*
 227          * We only need to flush on one CPU,
 228          * clflush is a MESI-coherent instruction that
 229          * will cause all other CPUs to flush the same
 230          * cachelines:
 231          */
 232         for (i = 0; i < numpages; i++) {
 233                 unsigned long addr;
 234                 pte_t *pte;
 235
 236                 if (in_flags & CPA_PAGES_ARRAY)
 237                         addr = (unsigned long)page_address(pages[i]);
 238                 else
 239                         addr = start[i];
 240
 241                 pte = lookup_address(addr, &level);
 242
 243                 /*
 244                  * Only flush present addresses:
 245                  */
 246                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 247                         clflush_cache_range((void *)addr, PAGE_SIZE);
 248         }
 249 }
 250
 251 /*
 252  * Certain areas of memory on x86 require very specific protection flags,
 253  * for example the BIOS area or kernel text. Callers don't always get this
 254  * right (again, ioremap() on BIOS memory is not uncommon) so this function
 255  * checks and fixes these known static required protection bits.
 256  */
 257 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
 258                                    unsigned long pfn)
 259 {
 260         pgprot_t forbidden = __pgprot(0);
 261
 262         /*
 263          * The BIOS area between 640k and 1Mb needs to be executable for
 264          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
 265          */
 266 #ifdef CONFIG_PCI_BIOS
 267         if (pcibios_enabled && within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
 268                 pgprot_val(forbidden) |= _PAGE_NX;
 269 #endif
 270
 271         /*
 272          * The kernel text needs to be executable for obvious reasons
 273          * Does not cover __inittext since that is gone later on. On
 274          * 64bit we do not enforce !NX on the low mapping
 275          */
 276         if (within(address, (unsigned long)_text, (unsigned long)_etext))
 277                 pgprot_val(forbidden) |= _PAGE_NX;
 278
 279         /*
 280          * The .rodata section needs to be read-only. Using the pfn
 281          * catches all aliases.
 282          */
 283         if (within(pfn, __pa_symbol(__start_rodata) >> PAGE_SHIFT,
 284                    __pa_symbol(__end_rodata) >> PAGE_SHIFT))
 285                 pgprot_val(forbidden) |= _PAGE_RW;
 286
 287 #if defined(CONFIG_X86_64)
 288         /*
 289          * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
 290          * kernel text mappings for the large page aligned text, rodata sections
 291          * will be always read-only. For the kernel identity mappings covering
 292          * the holes caused by this alignment can be anything that user asks.
 293          *
 294          * This will preserve the large page mappings for kernel text/data
 295          * at no extra cost.
 296          */
 297         if (kernel_set_to_readonly &&
 298             within(address, (unsigned long)_text,
 299                    (unsigned long)__end_rodata_hpage_align)) {
 300                 unsigned int level;
 301
 302                 /*
 303                  * Don't enforce the !RW mapping for the kernel text mapping,
 304                  * if the current mapping is already using small page mapping.
 305                  * No need to work hard to preserve large page mappings in this
 306                  * case.
 307                  *
 308                  * This also fixes the Linux Xen paravirt guest boot failure
 309                  * (because of unexpected read-only mappings for kernel identity
 310                  * mappings). In this paravirt guest case, the kernel text
 311                  * mapping and the kernel identity mapping share the same
 312                  * page-table pages. Thus we can't really use different
 313                  * protections for the kernel text and identity mappings. Also,
 314                  * these shared mappings are made of small page mappings.
 315                  * Thus this don't enforce !RW mapping for small page kernel
 316                  * text mapping logic will help Linux Xen parvirt guest boot
 317                  * as well.
 318                  */
 319                 if (lookup_address(address, &level) && (level != PG_LEVEL_4K))
 320                         pgprot_val(forbidden) |= _PAGE_RW;
 321         }
 322 #endif
 323
 324         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
 325
 326         return prot;
 327 }
 328
 329 /*
 330  * Lookup the page table entry for a virtual address in a specific pgd.
 331  * Return a pointer to the entry and the level of the mapping.
 332  */
 333 pte_t *lookup_address_in_pgd(pgd_t *pgd, unsigned long address,
 334                              unsigned int *level)
 335 {
 336         pud_t *pud;
 337         pmd_t *pmd;
 338
 339         *level = PG_LEVEL_NONE;
 340
 341         if (pgd_none(*pgd))
 342                 return NULL;
 343
 344         pud = pud_offset(pgd, address);
 345         if (pud_none(*pud))
 346                 return NULL;
 347
 348         *level = PG_LEVEL_1G;
 349         if (pud_large(*pud) || !pud_present(*pud))
 350                 return (pte_t *)pud;
 351
 352         pmd = pmd_offset(pud, address);
 353         if (pmd_none(*pmd))
 354                 return NULL;
 355
 356         *level = PG_LEVEL_2M;
 357         if (pmd_large(*pmd) || !pmd_present(*pmd))
 358                 return (pte_t *)pmd;
 359
 360         *level = PG_LEVEL_4K;
 361
 362         return pte_offset_kernel(pmd, address);
 363 }
 364
 365 /*
 366  * Lookup the page table entry for a virtual address. Return a pointer
 367  * to the entry and the level of the mapping.
 368  *
 369  * Note: We return pud and pmd either when the entry is marked large
 370  * or when the present bit is not set. Otherwise we would return a
 371  * pointer to a nonexisting mapping.
 372  */
 373 pte_t *lookup_address(unsigned long address, unsigned int *level)
 374 {
 375         return lookup_address_in_pgd(pgd_offset_k(address), address, level);
 376 }
 377 EXPORT_SYMBOL_GPL(lookup_address);
 378
 379 static pte_t *_lookup_address_cpa(struct cpa_data *cpa, unsigned long address,
 380                                   unsigned int *level)
 381 {
 382         if (cpa->pgd)
 383                 return lookup_address_in_pgd(cpa->pgd + pgd_index(address),
 384                                                address, level);
 385
 386         return lookup_address(address, level);
 387 }
 388
 389 /*
 390  * Lookup the PMD entry for a virtual address. Return a pointer to the entry
 391  * or NULL if not present.
 392  */
 393 pmd_t *lookup_pmd_address(unsigned long address)
 394 {
 395         pgd_t *pgd;
 396         pud_t *pud;
 397
 398         pgd = pgd_offset_k(address);
 399         if (pgd_none(*pgd))
 400                 return NULL;
 401
 402         pud = pud_offset(pgd, address);
 403         if (pud_none(*pud) || pud_large(*pud) || !pud_present(*pud))
 404                 return NULL;
 405
 406         return pmd_offset(pud, address);
 407 }
 408
 409 /*
 410  * This is necessary because __pa() does not work on some
 411  * kinds of memory, like vmalloc() or the alloc_remap()
 412  * areas on 32-bit NUMA systems.  The percpu areas can
 413  * end up in this kind of memory, for instance.
 414  *
 415  * This could be optimized, but it is only intended to be
 416  * used at inititalization time, and keeping it
 417  * unoptimized should increase the testing coverage for
 418  * the more obscure platforms.
 419  */
 420 phys_addr_t slow_virt_to_phys(void *__virt_addr)
 421 {
 422         unsigned long virt_addr = (unsigned long)__virt_addr;
 423         phys_addr_t phys_addr;
 424         unsigned long offset;
 425         enum pg_level level;
 426         pte_t *pte;
 427
 428         pte = lookup_address(virt_addr, &level);
 429         BUG_ON(!pte);
 430
 431         /*
 432          * pXX_pfn() returns unsigned long, which must be cast to phys_addr_t
 433          * before being left-shifted PAGE_SHIFT bits -- this trick is to
 434          * make 32-PAE kernel work correctly.
 435          */
 436         switch (level) {
 437         case PG_LEVEL_1G:
 438                 phys_addr = (phys_addr_t)pud_pfn(*(pud_t *)pte) << PAGE_SHIFT;
 439                 offset = virt_addr & ~PUD_PAGE_MASK;
 440                 break;
 441         case PG_LEVEL_2M:
 442                 phys_addr = (phys_addr_t)pmd_pfn(*(pmd_t *)pte) << PAGE_SHIFT;
 443                 offset = virt_addr & ~PMD_PAGE_MASK;
 444                 break;
 445         default:
 446                 phys_addr = (phys_addr_t)pte_pfn(*pte) << PAGE_SHIFT;
 447                 offset = virt_addr & ~PAGE_MASK;
 448         }
 449
 450         return (phys_addr_t)(phys_addr | offset);
 451 }
 452 EXPORT_SYMBOL_GPL(slow_virt_to_phys);
 453
 454 /*
 455  * Set the new pmd in all the pgds we know about:
 456  */
 457 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 458 {
 459         /* change init_mm */
 460         set_pte_atomic(kpte, pte);
 461 #ifdef CONFIG_X86_32
 462         if (!SHARED_KERNEL_PMD) {
 463                 struct page *page;
 464
 465                 list_for_each_entry(page, &pgd_list, lru) {
 466                         pgd_t *pgd;
 467                         pud_t *pud;
 468                         pmd_t *pmd;
 469
 470                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 471                         pud = pud_offset(pgd, address);
 472                         pmd = pmd_offset(pud, address);
 473                         set_pte_atomic((pte_t *)pmd, pte);
 474                 }
 475         }
 476 #endif
 477 }
 478
 479 static int
 480 try_preserve_large_page(pte_t *kpte, unsigned long address,
 481                         struct cpa_data *cpa)
 482 {
 483         unsigned long nextpage_addr, numpages, pmask, psize, addr, pfn, old_pfn;
 484         pte_t new_pte, old_pte, *tmp;
 485         pgprot_t old_prot, new_prot, req_prot;
 486         int i, do_split = 1;
 487         enum pg_level level;
 488
 489         if (cpa->force_split)
 490                 return 1;
 491
 492         spin_lock(&pgd_lock);
 493         /*
 494          * Check for races, another CPU might have split this page
 495          * up already:
 496          */
 497         tmp = _lookup_address_cpa(cpa, address, &level);
 498         if (tmp != kpte)
 499                 goto out_unlock;
 500
 501         switch (level) {
 502         case PG_LEVEL_2M:
 503                 old_prot = pmd_pgprot(*(pmd_t *)kpte);
 504                 old_pfn = pmd_pfn(*(pmd_t *)kpte);
 505                 break;
 506         case PG_LEVEL_1G:
 507                 old_prot = pud_pgprot(*(pud_t *)kpte);
 508                 old_pfn = pud_pfn(*(pud_t *)kpte);
 509                 break;
 510         default:
 511                 do_split = -EINVAL;
 512                 goto out_unlock;
 513         }
 514
 515         psize = page_level_size(level);
 516         pmask = page_level_mask(level);
 517
 518         /*
 519          * Calculate the number of pages, which fit into this large
 520          * page starting at address:
 521          */
 522         nextpage_addr = (address + psize) & pmask;
 523         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
 524         if (numpages < cpa->numpages)
 525                 cpa->numpages = numpages;
 526
 527         /*
 528          * We are safe now. Check whether the new pgprot is the same:
 529          * Convert protection attributes to 4k-format, as cpa->mask* are set
 530          * up accordingly.
 531          */
 532         old_pte = *kpte;
 533         req_prot = pgprot_large_2_4k(old_prot);
 534
 535         pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
 536         pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
 537
 538         /*
 539          * req_prot is in format of 4k pages. It must be converted to large
 540          * page format: the caching mode includes the PAT bit located at
 541          * different bit positions in the two formats.
 542          */
 543         req_prot = pgprot_4k_2_large(req_prot);
 544
 545         /*
 546          * Set the PSE and GLOBAL flags only if the PRESENT flag is
 547          * set otherwise pmd_present/pmd_huge will return true even on
 548          * a non present pmd. The canon_pgprot will clear _PAGE_GLOBAL
 549          * for the ancient hardware that doesn't support it.
 550          */
 551         if (pgprot_val(req_prot) & _PAGE_PRESENT)
 552                 pgprot_val(req_prot) |= _PAGE_PSE | _PAGE_GLOBAL;
 553         else
 554                 pgprot_val(req_prot) &= ~(_PAGE_PSE | _PAGE_GLOBAL);
 555
 556         req_prot = canon_pgprot(req_prot);
 557
 558         /*
 559          * old_pfn points to the large page base pfn. So we need
 560          * to add the offset of the virtual address:
 561          */
 562         pfn = old_pfn + ((address & (psize - 1)) >> PAGE_SHIFT);
 563         cpa->pfn = pfn;
 564
 565         new_prot = static_protections(req_prot, address, pfn);
 566
 567         /*
 568          * We need to check the full range, whether
 569          * static_protection() requires a different pgprot for one of
 570          * the pages in the range we try to preserve:
 571          */
 572         addr = address & pmask;
 573         pfn = old_pfn;
 574         for (i = 0; i < (psize >> PAGE_SHIFT); i++, addr += PAGE_SIZE, pfn++) {
 575                 pgprot_t chk_prot = static_protections(req_prot, addr, pfn);
 576
 577                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
 578                         goto out_unlock;
 579         }
 580
 581         /*
 582          * If there are no changes, return. maxpages has been updated
 583          * above:
 584          */
 585         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
 586                 do_split = 0;
 587                 goto out_unlock;
 588         }
 589
 590         /*
 591          * We need to change the attributes. Check, whether we can
 592          * change the large page in one go. We request a split, when
 593          * the address is not aligned and the number of pages is
 594          * smaller than the number of pages in the large page. Note
 595          * that we limited the number of possible pages already to
 596          * the number of pages in the large page.
 597          */
 598         if (address == (address & pmask) && cpa->numpages == (psize >> PAGE_SHIFT)) {
 599                 /*
 600                  * The address is aligned and the number of pages
 601                  * covers the full page.
 602                  */
 603                 new_pte = pfn_pte(old_pfn, new_prot);
 604                 __set_pmd_pte(kpte, address, new_pte);
 605                 cpa->flags |= CPA_FLUSHTLB;
 606                 do_split = 0;
 607         }
 608
 609 out_unlock:
 610         spin_unlock(&pgd_lock);
 611
 612         return do_split;
 613 }
 614
 615 static int
 616 __split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
 617                    struct page *base)
 618 {
 619         pte_t *pbase = (pte_t *)page_address(base);
 620         unsigned long ref_pfn, pfn, pfninc = 1;
 621         unsigned int i, level;
 622         pte_t *tmp;
 623         pgprot_t ref_prot;
 624
 625         spin_lock(&pgd_lock);
 626         /*
 627          * Check for races, another CPU might have split this page
 628          * up for us already:
 629          */
 630         tmp = _lookup_address_cpa(cpa, address, &level);
 631         if (tmp != kpte) {
 632                 spin_unlock(&pgd_lock);
 633                 return 1;
 634         }
 635
 636         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
 637
 638         switch (level) {
 639         case PG_LEVEL_2M:
 640                 ref_prot = pmd_pgprot(*(pmd_t *)kpte);
 641                 /* clear PSE and promote PAT bit to correct position */
 642                 ref_prot = pgprot_large_2_4k(ref_prot);
 643                 ref_pfn = pmd_pfn(*(pmd_t *)kpte);
 644                 break;
 645
 646         case PG_LEVEL_1G:
 647                 ref_prot = pud_pgprot(*(pud_t *)kpte);
 648                 ref_pfn = pud_pfn(*(pud_t *)kpte);
 649                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
 650
 651                 /*
 652                  * Clear the PSE flags if the PRESENT flag is not set
 653                  * otherwise pmd_present/pmd_huge will return true
 654                  * even on a non present pmd.
 655                  */
 656                 if (!(pgprot_val(ref_prot) & _PAGE_PRESENT))
 657                         pgprot_val(ref_prot) &= ~_PAGE_PSE;
 658                 break;
 659
 660         default:
 661                 spin_unlock(&pgd_lock);
 662                 return 1;
 663         }
 664
 665         /*
 666          * Set the GLOBAL flags only if the PRESENT flag is set
 667          * otherwise pmd/pte_present will return true even on a non
 668          * present pmd/pte. The canon_pgprot will clear _PAGE_GLOBAL
 669          * for the ancient hardware that doesn't support it.
 670          */
 671         if (pgprot_val(ref_prot) & _PAGE_PRESENT)
 672                 pgprot_val(ref_prot) |= _PAGE_GLOBAL;
 673         else
 674                 pgprot_val(ref_prot) &= ~_PAGE_GLOBAL;
 675
 676         /*
 677          * Get the target pfn from the original entry:
 678          */
 679         pfn = ref_pfn;
 680         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
 681                 set_pte(&pbase[i], pfn_pte(pfn, canon_pgprot(ref_prot)));
 682
 683         if (virt_addr_valid(address)) {
 684                 unsigned long pfn = PFN_DOWN(__pa(address));
 685
 686                 if (pfn_range_is_mapped(pfn, pfn + 1))
 687                         split_page_count(level);
 688         }
 689
 690         /*
 691          * Install the new, split up pagetable.
 692          *
 693          * We use the standard kernel pagetable protections for the new
 694          * pagetable protections, the actual ptes set above control the
 695          * primary protection behavior:
 696          */
 697         __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
 698
 699         /*
 700          * Intel Atom errata AAH41 workaround.
 701          *
 702          * The real fix should be in hw or in a microcode update, but
 703          * we also probabilistically try to reduce the window of having
 704          * a large TLB mixed with 4K TLBs while instruction fetches are
 705          * going on.
 706          */
 707         __flush_tlb_all();
 708         spin_unlock(&pgd_lock);
 709
 710         return 0;
 711 }
 712
 713 static int split_large_page(struct cpa_data *cpa, pte_t *kpte,
 714                             unsigned long address)
 715 {
 716         struct page *base;
 717
 718         if (!debug_pagealloc_enabled())
 719                 spin_unlock(&cpa_lock);
 720         base = alloc_pages(GFP_KERNEL | __GFP_NOTRACK, 0);
 721         if (!debug_pagealloc_enabled())
 722                 spin_lock(&cpa_lock);
 723         if (!base)
 724                 return -ENOMEM;
 725
 726         if (__split_large_page(cpa, kpte, address, base))
 727                 __free_page(base);
 728
 729         return 0;
 730 }
 731
 732 static bool try_to_free_pte_page(pte_t *pte)
 733 {
 734         int i;
 735
 736         for (i = 0; i < PTRS_PER_PTE; i++)
 737                 if (!pte_none(pte[i]))
 738                         return false;
 739
 740         free_page((unsigned long)pte);
 741         return true;
 742 }
 743
 744 static bool try_to_free_pmd_page(pmd_t *pmd)
 745 {
 746         int i;
 747
 748         for (i = 0; i < PTRS_PER_PMD; i++)
 749                 if (!pmd_none(pmd[i]))
 750                         return false;
 751
 752         free_page((unsigned long)pmd);
 753         return true;
 754 }
 755
 756 static bool unmap_pte_range(pmd_t *pmd, unsigned long start, unsigned long end)
 757 {
 758         pte_t *pte = pte_offset_kernel(pmd, start);
 759
 760         while (start < end) {
 761                 set_pte(pte, __pte(0));
 762
 763                 start += PAGE_SIZE;
 764                 pte++;
 765         }
 766
 767         if (try_to_free_pte_page((pte_t *)pmd_page_vaddr(*pmd))) {
 768                 pmd_clear(pmd);
 769                 return true;
 770         }
 771         return false;
 772 }
 773
 774 static void __unmap_pmd_range(pud_t *pud, pmd_t *pmd,
 775                               unsigned long start, unsigned long end)
 776 {
 777         if (unmap_pte_range(pmd, start, end))
 778                 if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
 779                         pud_clear(pud);
 780 }
 781
 782 static void unmap_pmd_range(pud_t *pud, unsigned long start, unsigned long end)
 783 {
 784         pmd_t *pmd = pmd_offset(pud, start);
 785
 786         /*
 787          * Not on a 2MB page boundary?
 788          */
 789         if (start & (PMD_SIZE - 1)) {
 790                 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
 791                 unsigned long pre_end = min_t(unsigned long, end, next_page);
 792
 793                 __unmap_pmd_range(pud, pmd, start, pre_end);
 794
 795                 start = pre_end;
 796                 pmd++;
 797         }
 798
 799         /*
 800          * Try to unmap in 2M chunks.
 801          */
 802         while (end - start >= PMD_SIZE) {
 803                 if (pmd_large(*pmd))
 804                         pmd_clear(pmd);
 805                 else
 806                         __unmap_pmd_range(pud, pmd, start, start + PMD_SIZE);
 807
 808                 start += PMD_SIZE;
 809                 pmd++;
 810         }
 811
 812         /*
 813          * 4K leftovers?
 814          */
 815         if (start < end)
 816                 return __unmap_pmd_range(pud, pmd, start, end);
 817
 818         /*
 819          * Try again to free the PMD page if haven't succeeded above.
 820          */
 821         if (!pud_none(*pud))
 822                 if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
 823                         pud_clear(pud);
 824 }
 825
 826 static void unmap_pud_range(pgd_t *pgd, unsigned long start, unsigned long end)
 827 {
 828         pud_t *pud = pud_offset(pgd, start);
 829
 830         /*
 831          * Not on a GB page boundary?
 832          */
 833         if (start & (PUD_SIZE - 1)) {
 834                 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
 835                 unsigned long pre_end   = min_t(unsigned long, end, next_page);
 836
 837                 unmap_pmd_range(pud, start, pre_end);
 838
 839                 start = pre_end;
 840                 pud++;
 841         }
 842
 843         /*
 844          * Try to unmap in 1G chunks?
 845          */
 846         while (end - start >= PUD_SIZE) {
 847
 848                 if (pud_large(*pud))
 849                         pud_clear(pud);
 850                 else
 851                         unmap_pmd_range(pud, start, start + PUD_SIZE);
 852
 853                 start += PUD_SIZE;
 854                 pud++;
 855         }
 856
 857         /*
 858          * 2M leftovers?
 859          */
 860         if (start < end)
 861                 unmap_pmd_range(pud, start, end);
 862
 863         /*
 864          * No need to try to free the PUD page because we'll free it in
 865          * populate_pgd's error path
 866          */
 867 }
 868
 869 static int alloc_pte_page(pmd_t *pmd)
 870 {
 871         pte_t *pte = (pte_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
 872         if (!pte)
 873                 return -1;
 874
 875         set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
 876         return 0;
 877 }
 878
 879 static int alloc_pmd_page(pud_t *pud)
 880 {
 881         pmd_t *pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
 882         if (!pmd)
 883                 return -1;
 884
 885         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
 886         return 0;
 887 }
 888
 889 static void populate_pte(struct cpa_data *cpa,
 890                          unsigned long start, unsigned long end,
 891                          unsigned num_pages, pmd_t *pmd, pgprot_t pgprot)
 892 {
 893         pte_t *pte;
 894
 895         pte = pte_offset_kernel(pmd, start);
 896
 897         /*
 898          * Set the GLOBAL flags only if the PRESENT flag is
 899          * set otherwise pte_present will return true even on
 900          * a non present pte. The canon_pgprot will clear
 901          * _PAGE_GLOBAL for the ancient hardware that doesn't
 902          * support it.
 903          */
 904         if (pgprot_val(pgprot) & _PAGE_PRESENT)
 905                 pgprot_val(pgprot) |= _PAGE_GLOBAL;
 906         else
 907                 pgprot_val(pgprot) &= ~_PAGE_GLOBAL;
 908
 909         pgprot = canon_pgprot(pgprot);
 910
 911         while (num_pages-- && start < end) {
 912                 set_pte(pte, pfn_pte(cpa->pfn, pgprot));
 913
 914                 start    += PAGE_SIZE;
 915                 cpa->pfn++;
 916                 pte++;
 917         }
 918 }
 919
 920 static int populate_pmd(struct cpa_data *cpa,
 921                         unsigned long start, unsigned long end,
 922                         unsigned num_pages, pud_t *pud, pgprot_t pgprot)
 923 {
 924         unsigned int cur_pages = 0;
 925         pmd_t *pmd;
 926         pgprot_t pmd_pgprot;
 927
 928         /*
 929          * Not on a 2M boundary?
 930          */
 931         if (start & (PMD_SIZE - 1)) {
 932                 unsigned long pre_end = start + (num_pages << PAGE_SHIFT);
 933                 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
 934
 935                 pre_end   = min_t(unsigned long, pre_end, next_page);
 936                 cur_pages = (pre_end - start) >> PAGE_SHIFT;
 937                 cur_pages = min_t(unsigned int, num_pages, cur_pages);
 938
 939                 /*
 940                  * Need a PTE page?
 941                  */
 942                 pmd = pmd_offset(pud, start);
 943                 if (pmd_none(*pmd))
 944                         if (alloc_pte_page(pmd))
 945                                 return -1;
 946
 947                 populate_pte(cpa, start, pre_end, cur_pages, pmd, pgprot);
 948
 949                 start = pre_end;
 950         }
 951
 952         /*
 953          * We mapped them all?
 954          */
 955         if (num_pages == cur_pages)
 956                 return cur_pages;
 957
 958         pmd_pgprot = pgprot_4k_2_large(pgprot);
 959
 960         while (end - start >= PMD_SIZE) {
 961
 962                 /*
 963                  * We cannot use a 1G page so allocate a PMD page if needed.
 964                  */
 965                 if (pud_none(*pud))
 966                         if (alloc_pmd_page(pud))
 967                                 return -1;
 968
 969                 pmd = pmd_offset(pud, start);
 970
 971                 set_pmd(pmd, __pmd(cpa->pfn << PAGE_SHIFT | _PAGE_PSE |
 972                                    massage_pgprot(pmd_pgprot)));
 973
 974                 start     += PMD_SIZE;
 975                 cpa->pfn  += PMD_SIZE >> PAGE_SHIFT;
 976                 cur_pages += PMD_SIZE >> PAGE_SHIFT;
 977         }
 978
 979         /*
 980          * Map trailing 4K pages.
 981          */
 982         if (start < end) {
 983                 pmd = pmd_offset(pud, start);
 984                 if (pmd_none(*pmd))
 985                         if (alloc_pte_page(pmd))
 986                                 return -1;
 987
 988                 populate_pte(cpa, start, end, num_pages - cur_pages,
 989                              pmd, pgprot);
 990         }
 991         return num_pages;
 992 }
 993
 994 static int populate_pud(struct cpa_data *cpa, unsigned long start, pgd_t *pgd,
 995                         pgprot_t pgprot)
 996 {
 997         pud_t *pud;
 998         unsigned long end;
 999         int cur_pages = 0;
1000         pgprot_t pud_pgprot;
1001
1002         end = start + (cpa->numpages << PAGE_SHIFT);
1003
1004         /*
1005          * Not on a Gb page boundary? => map everything up to it with
1006          * smaller pages.
1007          */
1008         if (start & (PUD_SIZE - 1)) {
1009                 unsigned long pre_end;
1010                 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1011
1012                 pre_end   = min_t(unsigned long, end, next_page);
1013                 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1014                 cur_pages = min_t(int, (int)cpa->numpages, cur_pages);
1015
1016                 pud = pud_offset(pgd, start);
1017
1018                 /*
1019                  * Need a PMD page?
1020                  */
1021                 if (pud_none(*pud))
1022                         if (alloc_pmd_page(pud))
1023                                 return -1;
1024
1025                 cur_pages = populate_pmd(cpa, start, pre_end, cur_pages,
1026                                          pud, pgprot);
1027                 if (cur_pages < 0)
1028                         return cur_pages;
1029
1030                 start = pre_end;
1031         }
1032
1033         /* We mapped them all? */
1034         if (cpa->numpages == cur_pages)
1035                 return cur_pages;
1036
1037         pud = pud_offset(pgd, start);
1038         pud_pgprot = pgprot_4k_2_large(pgprot);
1039
1040         /*
1041          * Map everything starting from the Gb boundary, possibly with 1G pages
1042          */
1043         while (boot_cpu_has(X86_FEATURE_GBPAGES) && end - start >= PUD_SIZE) {
1044                 set_pud(pud, __pud(cpa->pfn << PAGE_SHIFT | _PAGE_PSE |
1045                                    massage_pgprot(pud_pgprot)));
1046
1047                 start     += PUD_SIZE;
1048                 cpa->pfn  += PUD_SIZE >> PAGE_SHIFT;
1049                 cur_pages += PUD_SIZE >> PAGE_SHIFT;
1050                 pud++;
1051         }
1052
1053         /* Map trailing leftover */
1054         if (start < end) {
1055                 int tmp;
1056
1057                 pud = pud_offset(pgd, start);
1058                 if (pud_none(*pud))
1059                         if (alloc_pmd_page(pud))
1060                                 return -1;
1061
1062                 tmp = populate_pmd(cpa, start, end, cpa->numpages - cur_pages,
1063                                    pud, pgprot);
1064                 if (tmp < 0)
1065                         return cur_pages;
1066
1067                 cur_pages += tmp;
1068         }
1069         return cur_pages;
1070 }
1071
1072 /*
1073  * Restrictions for kernel page table do not necessarily apply when mapping in
1074  * an alternate PGD.
1075  */
1076 static int populate_pgd(struct cpa_data *cpa, unsigned long addr)
1077 {
1078         pgprot_t pgprot = __pgprot(_KERNPG_TABLE);
1079         pud_t *pud = NULL;      /* shut up gcc */
1080         pgd_t *pgd_entry;
1081         int ret;
1082
1083         pgd_entry = cpa->pgd + pgd_index(addr);
1084
1085         /*
1086          * Allocate a PUD page and hand it down for mapping.
1087          */
1088         if (pgd_none(*pgd_entry)) {
1089                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
1090                 if (!pud)
1091                         return -1;
1092
1093                 set_pgd(pgd_entry, __pgd(__pa(pud) | _KERNPG_TABLE));
1094         }
1095
1096         pgprot_val(pgprot) &= ~pgprot_val(cpa->mask_clr);
1097         pgprot_val(pgprot) |=  pgprot_val(cpa->mask_set);
1098
1099         ret = populate_pud(cpa, addr, pgd_entry, pgprot);
1100         if (ret < 0) {
1101                 /*
1102                  * Leave the PUD page in place in case some other CPU or thread
1103                  * already found it, but remove any useless entries we just
1104                  * added to it.
1105                  */
1106                 unmap_pud_range(pgd_entry, addr,
1107                                 addr + (cpa->numpages << PAGE_SHIFT));
1108                 return ret;
1109         }
1110
1111         cpa->numpages = ret;
1112         return 0;
1113 }
1114
1115 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
1116                                int primary)
1117 {
1118         if (cpa->pgd) {
1119                 /*
1120                  * Right now, we only execute this code path when mapping
1121                  * the EFI virtual memory map regions, no other users
1122                  * provide a ->pgd value. This may change in the future.
1123                  */
1124                 return populate_pgd(cpa, vaddr);
1125         }
1126
1127         /*
1128          * Ignore all non primary paths.
1129          */
1130         if (!primary) {
1131                 cpa->numpages = 1;
1132                 return 0;
1133         }
1134
1135         /*
1136          * Ignore the NULL PTE for kernel identity mapping, as it is expected
1137          * to have holes.
1138          * Also set numpages to '1' indicating that we processed cpa req for
1139          * one virtual address page and its pfn. TBD: numpages can be set based
1140          * on the initial value and the level returned by lookup_address().
1141          */
1142         if (within(vaddr, PAGE_OFFSET,
1143                    PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
1144                 cpa->numpages = 1;
1145                 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
1146                 return 0;
1147         } else {
1148                 WARN(1, KERN_WARNING "CPA: called for zero pte. "
1149                         "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
1150                         *cpa->vaddr);
1151
1152                 return -EFAULT;
1153         }
1154 }
1155
1156 static int __change_page_attr(struct cpa_data *cpa, int primary)
1157 {
1158         unsigned long address;
1159         int do_split, err;
1160         unsigned int level;
1161         pte_t *kpte, old_pte;
1162
1163         if (cpa->flags & CPA_PAGES_ARRAY) {
1164                 struct page *page = cpa->pages[cpa->curpage];
1165                 if (unlikely(PageHighMem(page)))
1166                         return 0;
1167                 address = (unsigned long)page_address(page);
1168         } else if (cpa->flags & CPA_ARRAY)
1169                 address = cpa->vaddr[cpa->curpage];
1170         else
1171                 address = *cpa->vaddr;
1172 repeat:
1173         kpte = _lookup_address_cpa(cpa, address, &level);
1174         if (!kpte)
1175                 return __cpa_process_fault(cpa, address, primary);
1176
1177         old_pte = *kpte;
1178         if (pte_none(old_pte))
1179                 return __cpa_process_fault(cpa, address, primary);
1180
1181         if (level == PG_LEVEL_4K) {
1182                 pte_t new_pte;
1183                 pgprot_t new_prot = pte_pgprot(old_pte);
1184                 unsigned long pfn = pte_pfn(old_pte);
1185
1186                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
1187                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
1188
1189                 new_prot = static_protections(new_prot, address, pfn);
1190
1191                 /*
1192                  * Set the GLOBAL flags only if the PRESENT flag is
1193                  * set otherwise pte_present will return true even on
1194                  * a non present pte. The canon_pgprot will clear
1195                  * _PAGE_GLOBAL for the ancient hardware that doesn't
1196                  * support it.
1197                  */
1198                 if (pgprot_val(new_prot) & _PAGE_PRESENT)
1199                         pgprot_val(new_prot) |= _PAGE_GLOBAL;
1200                 else
1201                         pgprot_val(new_prot) &= ~_PAGE_GLOBAL;
1202
1203                 /*
1204                  * We need to keep the pfn from the existing PTE,
1205                  * after all we're only going to change it's attributes
1206                  * not the memory it points to
1207                  */
1208                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
1209                 cpa->pfn = pfn;
1210                 /*
1211                  * Do we really change anything ?
1212                  */
1213                 if (pte_val(old_pte) != pte_val(new_pte)) {
1214                         set_pte_atomic(kpte, new_pte);
1215                         cpa->flags |= CPA_FLUSHTLB;
1216                 }
1217                 cpa->numpages = 1;
1218                 return 0;
1219         }
1220
1221         /*
1222          * Check, whether we can keep the large page intact
1223          * and just change the pte:
1224          */
1225         do_split = try_preserve_large_page(kpte, address, cpa);
1226         /*
1227          * When the range fits into the existing large page,
1228          * return. cp->numpages and cpa->tlbflush have been updated in
1229          * try_large_page:
1230          */
1231         if (do_split <= 0)
1232                 return do_split;
1233
1234         /*
1235          * We have to split the large page:
1236          */
1237         err = split_large_page(cpa, kpte, address);
1238         if (!err) {
1239                 /*
1240                  * Do a global flush tlb after splitting the large page
1241                  * and before we do the actual change page attribute in the PTE.
1242                  *
1243                  * With out this, we violate the TLB application note, that says
1244                  * "The TLBs may contain both ordinary and large-page
1245                  *  translations for a 4-KByte range of linear addresses. This
1246                  *  may occur if software modifies the paging structures so that
1247                  *  the page size used for the address range changes. If the two
1248                  *  translations differ with respect to page frame or attributes
1249                  *  (e.g., permissions), processor behavior is undefined and may
1250                  *  be implementation-specific."
1251                  *
1252                  * We do this global tlb flush inside the cpa_lock, so that we
1253                  * don't allow any other cpu, with stale tlb entries change the
1254                  * page attribute in parallel, that also falls into the
1255                  * just split large page entry.
1256                  */
1257                 flush_tlb_all();
1258                 goto repeat;
1259         }
1260
1261         return err;
1262 }
1263
1264 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
1265
1266 static int cpa_process_alias(struct cpa_data *cpa)
1267 {
1268         struct cpa_data alias_cpa;
1269         unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
1270         unsigned long vaddr;
1271         int ret;
1272
1273         if (!pfn_range_is_mapped(cpa->pfn, cpa->pfn + 1))
1274                 return 0;
1275
1276         /*
1277          * No need to redo, when the primary call touched the direct
1278          * mapping already:
1279          */
1280         if (cpa->flags & CPA_PAGES_ARRAY) {
1281                 struct page *page = cpa->pages[cpa->curpage];
1282                 if (unlikely(PageHighMem(page)))
1283                         return 0;
1284                 vaddr = (unsigned long)page_address(page);
1285         } else if (cpa->flags & CPA_ARRAY)
1286                 vaddr = cpa->vaddr[cpa->curpage];
1287         else
1288                 vaddr = *cpa->vaddr;
1289
1290         if (!(within(vaddr, PAGE_OFFSET,
1291                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
1292
1293                 alias_cpa = *cpa;
1294                 alias_cpa.vaddr = &laddr;
1295                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1296
1297                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
1298                 if (ret)
1299                         return ret;
1300         }
1301
1302 #ifdef CONFIG_X86_64
1303         /*
1304          * If the primary call didn't touch the high mapping already
1305          * and the physical address is inside the kernel map, we need
1306          * to touch the high mapped kernel as well:
1307          */
1308         if (!within(vaddr, (unsigned long)_text, _brk_end) &&
1309             within_inclusive(cpa->pfn, highmap_start_pfn(),
1310                              highmap_end_pfn())) {
1311                 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
1312                                                __START_KERNEL_map - phys_base;
1313                 alias_cpa = *cpa;
1314                 alias_cpa.vaddr = &temp_cpa_vaddr;
1315                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1316
1317                 /*
1318                  * The high mapping range is imprecise, so ignore the
1319                  * return value.
1320                  */
1321                 __change_page_attr_set_clr(&alias_cpa, 0);
1322         }
1323 #endif
1324
1325         return 0;
1326 }
1327
1328 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
1329 {
1330         int ret, numpages = cpa->numpages;
1331
1332         while (numpages) {
1333                 /*
1334                  * Store the remaining nr of pages for the large page
1335                  * preservation check.
1336                  */
1337                 cpa->numpages = numpages;
1338                 /* for array changes, we can't use large page */
1339                 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
1340                         cpa->numpages = 1;
1341
1342                 if (!debug_pagealloc_enabled())
1343                         spin_lock(&cpa_lock);
1344                 ret = __change_page_attr(cpa, checkalias);
1345                 if (!debug_pagealloc_enabled())
1346                         spin_unlock(&cpa_lock);
1347                 if (ret)
1348                         return ret;
1349
1350                 if (checkalias) {
1351                         ret = cpa_process_alias(cpa);
1352                         if (ret)
1353                                 return ret;
1354                 }
1355
1356                 /*
1357                  * Adjust the number of pages with the result of the
1358                  * CPA operation. Either a large page has been
1359                  * preserved or a single page update happened.
1360                  */
1361                 BUG_ON(cpa->numpages > numpages || !cpa->numpages);
1362                 numpages -= cpa->numpages;
1363                 if (cpa->flags & (CPA_PAGES_ARRAY | CPA_ARRAY))
1364                         cpa->curpage++;
1365                 else
1366                         *cpa->vaddr += cpa->numpages * PAGE_SIZE;
1367
1368         }
1369         return 0;
1370 }
1371
1372 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
1373                                     pgprot_t mask_set, pgprot_t mask_clr,
1374                                     int force_split, int in_flag,
1375                                     struct page **pages)
1376 {
1377         struct cpa_data cpa;
1378         int ret, cache, checkalias;
1379         unsigned long baddr = 0;
1380
1381         memset(&cpa, 0, sizeof(cpa));
1382
1383         /*
1384          * Check, if we are requested to change a not supported
1385          * feature:
1386          */
1387         mask_set = canon_pgprot(mask_set);
1388         mask_clr = canon_pgprot(mask_clr);
1389         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
1390                 return 0;
1391
1392         /* Ensure we are PAGE_SIZE aligned */
1393         if (in_flag & CPA_ARRAY) {
1394                 int i;
1395                 for (i = 0; i < numpages; i++) {
1396                         if (addr[i] & ~PAGE_MASK) {
1397                                 addr[i] &= PAGE_MASK;
1398                                 WARN_ON_ONCE(1);
1399                         }
1400                 }
1401         } else if (!(in_flag & CPA_PAGES_ARRAY)) {
1402                 /*
1403                  * in_flag of CPA_PAGES_ARRAY implies it is aligned.
1404                  * No need to cehck in that case
1405                  */
1406                 if (*addr & ~PAGE_MASK) {
1407                         *addr &= PAGE_MASK;
1408                         /*
1409                          * People should not be passing in unaligned addresses:
1410                          */
1411                         WARN_ON_ONCE(1);
1412                 }
1413                 /*
1414                  * Save address for cache flush. *addr is modified in the call
1415                  * to __change_page_attr_set_clr() below.
1416                  */
1417                 baddr = *addr;
1418         }
1419
1420         /* Must avoid aliasing mappings in the highmem code */
1421         kmap_flush_unused();
1422
1423         vm_unmap_aliases();
1424
1425         cpa.vaddr = addr;
1426         cpa.pages = pages;
1427         cpa.numpages = numpages;
1428         cpa.mask_set = mask_set;
1429         cpa.mask_clr = mask_clr;
1430         cpa.flags = 0;
1431         cpa.curpage = 0;
1432         cpa.force_split = force_split;
1433
1434         if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
1435                 cpa.flags |= in_flag;
1436
1437         /* No alias checking for _NX bit modifications */
1438         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
1439
1440         ret = __change_page_attr_set_clr(&cpa, checkalias);
1441
1442         /*
1443          * Check whether we really changed something:
1444          */
1445         if (!(cpa.flags & CPA_FLUSHTLB))
1446                 goto out;
1447
1448         /*
1449          * No need to flush, when we did not set any of the caching
1450          * attributes:
1451          */
1452         cache = !!pgprot2cachemode(mask_set);
1453
1454         /*
1455          * On success we use CLFLUSH, when the CPU supports it to
1456          * avoid the WBINVD. If the CPU does not support it and in the
1457          * error case we fall back to cpa_flush_all (which uses
1458          * WBINVD):
1459          */
1460         if (!ret && boot_cpu_has(X86_FEATURE_CLFLUSH)) {
1461                 if (cpa.flags & (CPA_PAGES_ARRAY | CPA_ARRAY)) {
1462                         cpa_flush_array(addr, numpages, cache,
1463                                         cpa.flags, pages);
1464                 } else
1465                         cpa_flush_range(baddr, numpages, cache);
1466         } else
1467                 cpa_flush_all(cache);
1468
1469 out:
1470         return ret;
1471 }
1472
1473 static inline int change_page_attr_set(unsigned long *addr, int numpages,
1474                                        pgprot_t mask, int array)
1475 {
1476         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
1477                 (array ? CPA_ARRAY : 0), NULL);
1478 }
1479
1480 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
1481                                          pgprot_t mask, int array)
1482 {
1483         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
1484                 (array ? CPA_ARRAY : 0), NULL);
1485 }
1486
1487 static inline int cpa_set_pages_array(struct page **pages, int numpages,
1488                                        pgprot_t mask)
1489 {
1490         return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
1491                 CPA_PAGES_ARRAY, pages);
1492 }
1493
1494 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1495                                          pgprot_t mask)
1496 {
1497         return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1498                 CPA_PAGES_ARRAY, pages);
1499 }
1500
1501 int _set_memory_uc(unsigned long addr, int numpages)
1502 {
1503         /*
1504          * for now UC MINUS. see comments in ioremap_nocache()
1505          * If you really need strong UC use ioremap_uc(), but note
1506          * that you cannot override IO areas with set_memory_*() as
1507          * these helpers cannot work with IO memory.
1508          */
1509         return change_page_attr_set(&addr, numpages,
1510                                     cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1511                                     0);
1512 }
1513
1514 int set_memory_uc(unsigned long addr, int numpages)
1515 {
1516         int ret;
1517
1518         /*
1519          * for now UC MINUS. see comments in ioremap_nocache()
1520          */
1521         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1522                               _PAGE_CACHE_MODE_UC_MINUS, NULL);
1523         if (ret)
1524                 goto out_err;
1525
1526         ret = _set_memory_uc(addr, numpages);
1527         if (ret)
1528                 goto out_free;
1529
1530         return 0;
1531
1532 out_free:
1533         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1534 out_err:
1535         return ret;
1536 }
1537 EXPORT_SYMBOL(set_memory_uc);
1538
1539 static int _set_memory_array(unsigned long *addr, int addrinarray,
1540                 enum page_cache_mode new_type)
1541 {
1542         enum page_cache_mode set_type;
1543         int i, j;
1544         int ret;
1545
1546         for (i = 0; i < addrinarray; i++) {
1547                 ret = reserve_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE,
1548                                         new_type, NULL);
1549                 if (ret)
1550                         goto out_free;
1551         }
1552
1553         /* If WC, set to UC- first and then WC */
1554         set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
1555                                 _PAGE_CACHE_MODE_UC_MINUS : new_type;
1556
1557         ret = change_page_attr_set(addr, addrinarray,
1558                                    cachemode2pgprot(set_type), 1);
1559
1560         if (!ret && new_type == _PAGE_CACHE_MODE_WC)
1561                 ret = change_page_attr_set_clr(addr, addrinarray,
1562                                                cachemode2pgprot(
1563                                                 _PAGE_CACHE_MODE_WC),
1564                                                __pgprot(_PAGE_CACHE_MASK),
1565                                                0, CPA_ARRAY, NULL);
1566         if (ret)
1567                 goto out_free;
1568
1569         return 0;
1570
1571 out_free:
1572         for (j = 0; j < i; j++)
1573                 free_memtype(__pa(addr[j]), __pa(addr[j]) + PAGE_SIZE);
1574
1575         return ret;
1576 }
1577
1578 int set_memory_array_uc(unsigned long *addr, int addrinarray)
1579 {
1580         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_UC_MINUS);
1581 }
1582 EXPORT_SYMBOL(set_memory_array_uc);
1583
1584 int set_memory_array_wc(unsigned long *addr, int addrinarray)
1585 {
1586         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_WC);
1587 }
1588 EXPORT_SYMBOL(set_memory_array_wc);
1589
1590 int set_memory_array_wt(unsigned long *addr, int addrinarray)
1591 {
1592         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_WT);
1593 }
1594 EXPORT_SYMBOL_GPL(set_memory_array_wt);
1595
1596 int _set_memory_wc(unsigned long addr, int numpages)
1597 {
1598         int ret;
1599         unsigned long addr_copy = addr;
1600
1601         ret = change_page_attr_set(&addr, numpages,
1602                                    cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1603                                    0);
1604         if (!ret) {
1605                 ret = change_page_attr_set_clr(&addr_copy, numpages,
1606                                                cachemode2pgprot(
1607                                                 _PAGE_CACHE_MODE_WC),
1608                                                __pgprot(_PAGE_CACHE_MASK),
1609                                                0, 0, NULL);
1610         }
1611         return ret;
1612 }
1613
1614 int set_memory_wc(unsigned long addr, int numpages)
1615 {
1616         int ret;
1617
1618         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1619                 _PAGE_CACHE_MODE_WC, NULL);
1620         if (ret)
1621                 return ret;
1622
1623         ret = _set_memory_wc(addr, numpages);
1624         if (ret)
1625                 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1626
1627         return ret;
1628 }
1629 EXPORT_SYMBOL(set_memory_wc);
1630
1631 int _set_memory_wt(unsigned long addr, int numpages)
1632 {
1633         return change_page_attr_set(&addr, numpages,
1634                                     cachemode2pgprot(_PAGE_CACHE_MODE_WT), 0);
1635 }
1636
1637 int set_memory_wt(unsigned long addr, int numpages)
1638 {
1639         int ret;
1640
1641         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1642                               _PAGE_CACHE_MODE_WT, NULL);
1643         if (ret)
1644                 return ret;
1645
1646         ret = _set_memory_wt(addr, numpages);
1647         if (ret)
1648                 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1649
1650         return ret;
1651 }
1652 EXPORT_SYMBOL_GPL(set_memory_wt);
1653
1654 int _set_memory_wb(unsigned long addr, int numpages)
1655 {
1656         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1657         return change_page_attr_clear(&addr, numpages,
1658                                       __pgprot(_PAGE_CACHE_MASK), 0);
1659 }
1660
1661 int set_memory_wb(unsigned long addr, int numpages)
1662 {
1663         int ret;
1664
1665         ret = _set_memory_wb(addr, numpages);
1666         if (ret)
1667                 return ret;
1668
1669         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1670         return 0;
1671 }
1672 EXPORT_SYMBOL(set_memory_wb);
1673
1674 int set_memory_array_wb(unsigned long *addr, int addrinarray)
1675 {
1676         int i;
1677         int ret;
1678
1679         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1680         ret = change_page_attr_clear(addr, addrinarray,
1681                                       __pgprot(_PAGE_CACHE_MASK), 1);
1682         if (ret)
1683                 return ret;
1684
1685         for (i = 0; i < addrinarray; i++)
1686                 free_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE);
1687
1688         return 0;
1689 }
1690 EXPORT_SYMBOL(set_memory_array_wb);
1691
1692 int set_memory_x(unsigned long addr, int numpages)
1693 {
1694         if (!(__supported_pte_mask & _PAGE_NX))
1695                 return 0;
1696
1697         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1698 }
1699 EXPORT_SYMBOL(set_memory_x);
1700
1701 int set_memory_nx(unsigned long addr, int numpages)
1702 {
1703         if (!(__supported_pte_mask & _PAGE_NX))
1704                 return 0;
1705
1706         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1707 }
1708 EXPORT_SYMBOL(set_memory_nx);
1709
1710 int set_memory_ro(unsigned long addr, int numpages)
1711 {
1712         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1713 }
1714
1715 int set_memory_rw(unsigned long addr, int numpages)
1716 {
1717         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1718 }
1719
1720 int set_memory_np(unsigned long addr, int numpages)
1721 {
1722         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1723 }
1724
1725 int set_memory_4k(unsigned long addr, int numpages)
1726 {
1727         return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1728                                         __pgprot(0), 1, 0, NULL);
1729 }
1730
1731 int set_pages_uc(struct page *page, int numpages)
1732 {
1733         unsigned long addr = (unsigned long)page_address(page);
1734
1735         return set_memory_uc(addr, numpages);
1736 }
1737 EXPORT_SYMBOL(set_pages_uc);
1738
1739 static int _set_pages_array(struct page **pages, int addrinarray,
1740                 enum page_cache_mode new_type)
1741 {
1742         unsigned long start;
1743         unsigned long end;
1744         enum page_cache_mode set_type;
1745         int i;
1746         int free_idx;
1747         int ret;
1748
1749         for (i = 0; i < addrinarray; i++) {
1750                 if (PageHighMem(pages[i]))
1751                         continue;
1752                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1753                 end = start + PAGE_SIZE;
1754                 if (reserve_memtype(start, end, new_type, NULL))
1755                         goto err_out;
1756         }
1757
1758         /* If WC, set to UC- first and then WC */
1759         set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
1760                                 _PAGE_CACHE_MODE_UC_MINUS : new_type;
1761
1762         ret = cpa_set_pages_array(pages, addrinarray,
1763                                   cachemode2pgprot(set_type));
1764         if (!ret && new_type == _PAGE_CACHE_MODE_WC)
1765                 ret = change_page_attr_set_clr(NULL, addrinarray,
1766                                                cachemode2pgprot(
1767                                                 _PAGE_CACHE_MODE_WC),
1768                                                __pgprot(_PAGE_CACHE_MASK),
1769                                                0, CPA_PAGES_ARRAY, pages);
1770         if (ret)
1771                 goto err_out;
1772         return 0; /* Success */
1773 err_out:
1774         free_idx = i;
1775         for (i = 0; i < free_idx; i++) {
1776                 if (PageHighMem(pages[i]))
1777                         continue;
1778                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1779                 end = start + PAGE_SIZE;
1780                 free_memtype(start, end);
1781         }
1782         return -EINVAL;
1783 }
1784
1785 int set_pages_array_uc(struct page **pages, int addrinarray)
1786 {
1787         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_UC_MINUS);
1788 }
1789 EXPORT_SYMBOL(set_pages_array_uc);
1790
1791 int set_pages_array_wc(struct page **pages, int addrinarray)
1792 {
1793         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_WC);
1794 }
1795 EXPORT_SYMBOL(set_pages_array_wc);
1796
1797 int set_pages_array_wt(struct page **pages, int addrinarray)
1798 {
1799         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_WT);
1800 }
1801 EXPORT_SYMBOL_GPL(set_pages_array_wt);
1802
1803 int set_pages_wb(struct page *page, int numpages)
1804 {
1805         unsigned long addr = (unsigned long)page_address(page);
1806
1807         return set_memory_wb(addr, numpages);
1808 }
1809 EXPORT_SYMBOL(set_pages_wb);
1810
1811 int set_pages_array_wb(struct page **pages, int addrinarray)
1812 {
1813         int retval;
1814         unsigned long start;
1815         unsigned long end;
1816         int i;
1817
1818         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1819         retval = cpa_clear_pages_array(pages, addrinarray,
1820                         __pgprot(_PAGE_CACHE_MASK));
1821         if (retval)
1822                 return retval;
1823
1824         for (i = 0; i < addrinarray; i++) {
1825                 if (PageHighMem(pages[i]))
1826                         continue;
1827                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1828                 end = start + PAGE_SIZE;
1829                 free_memtype(start, end);
1830         }
1831
1832         return 0;
1833 }
1834 EXPORT_SYMBOL(set_pages_array_wb);
1835
1836 int set_pages_x(struct page *page, int numpages)
1837 {
1838         unsigned long addr = (unsigned long)page_address(page);
1839
1840         return set_memory_x(addr, numpages);
1841 }
1842 EXPORT_SYMBOL(set_pages_x);
1843
1844 int set_pages_nx(struct page *page, int numpages)
1845 {
1846         unsigned long addr = (unsigned long)page_address(page);
1847
1848         return set_memory_nx(addr, numpages);
1849 }
1850 EXPORT_SYMBOL(set_pages_nx);
1851
1852 int set_pages_ro(struct page *page, int numpages)
1853 {
1854         unsigned long addr = (unsigned long)page_address(page);
1855
1856         return set_memory_ro(addr, numpages);
1857 }
1858
1859 int set_pages_rw(struct page *page, int numpages)
1860 {
1861         unsigned long addr = (unsigned long)page_address(page);
1862
1863         return set_memory_rw(addr, numpages);
1864 }
1865
1866 #ifdef CONFIG_DEBUG_PAGEALLOC
1867
1868 static int __set_pages_p(struct page *page, int numpages)
1869 {
1870         unsigned long tempaddr = (unsigned long) page_address(page);
1871         struct cpa_data cpa = { .vaddr = &tempaddr,
1872                                 .pgd = NULL,
1873                                 .numpages = numpages,
1874                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1875                                 .mask_clr = __pgprot(0),
1876                                 .flags = 0};
1877
1878         /*
1879          * No alias checking needed for setting present flag. otherwise,
1880          * we may need to break large pages for 64-bit kernel text
1881          * mappings (this adds to complexity if we want to do this from
1882          * atomic context especially). Let's keep it simple!
1883          */
1884         return __change_page_attr_set_clr(&cpa, 0);
1885 }
1886
1887 static int __set_pages_np(struct page *page, int numpages)
1888 {
1889         unsigned long tempaddr = (unsigned long) page_address(page);
1890         struct cpa_data cpa = { .vaddr = &tempaddr,
1891                                 .pgd = NULL,
1892                                 .numpages = numpages,
1893                                 .mask_set = __pgprot(0),
1894                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1895                                 .flags = 0};
1896
1897         /*
1898          * No alias checking needed for setting not present flag. otherwise,
1899          * we may need to break large pages for 64-bit kernel text
1900          * mappings (this adds to complexity if we want to do this from
1901          * atomic context especially). Let's keep it simple!
1902          */
1903         return __change_page_attr_set_clr(&cpa, 0);
1904 }
1905
1906 void __kernel_map_pages(struct page *page, int numpages, int enable)
1907 {
1908         if (PageHighMem(page))
1909                 return;
1910         if (!enable) {
1911                 debug_check_no_locks_freed(page_address(page),
1912                                            numpages * PAGE_SIZE);
1913         }
1914
1915         /*
1916          * The return value is ignored as the calls cannot fail.
1917          * Large pages for identity mappings are not used at boot time
1918          * and hence no memory allocations during large page split.
1919          */
1920         if (enable)
1921                 __set_pages_p(page, numpages);
1922         else
1923                 __set_pages_np(page, numpages);
1924
1925         /*
1926          * We should perform an IPI and flush all tlbs,
1927          * but that can deadlock->flush only current cpu:
1928          */
1929         __flush_tlb_all();
1930
1931         arch_flush_lazy_mmu_mode();
1932 }
1933
1934 #ifdef CONFIG_HIBERNATION
1935
1936 bool kernel_page_present(struct page *page)
1937 {
1938         unsigned int level;
1939         pte_t *pte;
1940
1941         if (PageHighMem(page))
1942                 return false;
1943
1944         pte = lookup_address((unsigned long)page_address(page), &level);
1945         return (pte_val(*pte) & _PAGE_PRESENT);
1946 }
1947
1948 #endif /* CONFIG_HIBERNATION */
1949
1950 #endif /* CONFIG_DEBUG_PAGEALLOC */
1951
1952 int kernel_map_pages_in_pgd(pgd_t *pgd, u64 pfn, unsigned long address,
1953                             unsigned numpages, unsigned long page_flags)
1954 {
1955         int retval = -EINVAL;
1956
1957         struct cpa_data cpa = {
1958                 .vaddr = &address,
1959                 .pfn = pfn,
1960                 .pgd = pgd,
1961                 .numpages = numpages,
1962                 .mask_set = __pgprot(0),
1963                 .mask_clr = __pgprot(0),
1964                 .flags = 0,
1965         };
1966
1967         if (!(__supported_pte_mask & _PAGE_NX))
1968                 goto out;
1969
1970         if (!(page_flags & _PAGE_NX))
1971                 cpa.mask_clr = __pgprot(_PAGE_NX);
1972
1973         if (!(page_flags & _PAGE_RW))
1974                 cpa.mask_clr = __pgprot(_PAGE_RW);
1975
1976         cpa.mask_set = __pgprot(_PAGE_PRESENT | page_flags);
1977
1978         retval = __change_page_attr_set_clr(&cpa, 0);
1979         __flush_tlb_all();
1980
1981 out:
1982         return retval;
1983 }
1984
1985 /*
1986  * The testcases use internal knowledge of the implementation that shouldn't
1987  * be exposed to the rest of the kernel. Include these directly here.
1988  */
1989 #ifdef CONFIG_CPA_DEBUG
1990 #include "pageattr-test.c"
1991 #endif