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