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 (virt_addr_valid(address)) {
 672                 unsigned long pfn = PFN_DOWN(__pa(address));
 673
 674                 if (pfn_range_is_mapped(pfn, pfn + 1))
 675                         split_page_count(level);
 676         }
 677
 678         /*
 679          * Install the new, split up pagetable.
 680          *
 681          * We use the standard kernel pagetable protections for the new
 682          * pagetable protections, the actual ptes set above control the
 683          * primary protection behavior:
 684          */
 685         __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
 686
 687         /*
 688          * Intel Atom errata AAH41 workaround.
 689          *
 690          * The real fix should be in hw or in a microcode update, but
 691          * we also probabilistically try to reduce the window of having
 692          * a large TLB mixed with 4K TLBs while instruction fetches are
 693          * going on.
 694          */
 695         __flush_tlb_all();
 696         spin_unlock(&pgd_lock);
 697
 698         return 0;
 699 }
 700
 701 static int split_large_page(struct cpa_data *cpa, pte_t *kpte,
 702                             unsigned long address)
 703 {
 704         struct page *base;
 705
 706         if (!debug_pagealloc)
 707                 spin_unlock(&cpa_lock);
 708         base = alloc_pages(GFP_KERNEL | __GFP_NOTRACK, 0);
 709         if (!debug_pagealloc)
 710                 spin_lock(&cpa_lock);
 711         if (!base)
 712                 return -ENOMEM;
 713
 714         if (__split_large_page(cpa, kpte, address, base))
 715                 __free_page(base);
 716
 717         return 0;
 718 }
 719
 720 static bool try_to_free_pte_page(pte_t *pte)
 721 {
 722         int i;
 723
 724         for (i = 0; i < PTRS_PER_PTE; i++)
 725                 if (!pte_none(pte[i]))
 726                         return false;
 727
 728         free_page((unsigned long)pte);
 729         return true;
 730 }
 731
 732 static bool try_to_free_pmd_page(pmd_t *pmd)
 733 {
 734         int i;
 735
 736         for (i = 0; i < PTRS_PER_PMD; i++)
 737                 if (!pmd_none(pmd[i]))
 738                         return false;
 739
 740         free_page((unsigned long)pmd);
 741         return true;
 742 }
 743
 744 static bool try_to_free_pud_page(pud_t *pud)
 745 {
 746         int i;
 747
 748         for (i = 0; i < PTRS_PER_PUD; i++)
 749                 if (!pud_none(pud[i]))
 750                         return false;
 751
 752         free_page((unsigned long)pud);
 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 void unmap_pgd_range(pgd_t *root, unsigned long addr, unsigned long end)
 870 {
 871         pgd_t *pgd_entry = root + pgd_index(addr);
 872
 873         unmap_pud_range(pgd_entry, addr, end);
 874
 875         if (try_to_free_pud_page((pud_t *)pgd_page_vaddr(*pgd_entry)))
 876                 pgd_clear(pgd_entry);
 877 }
 878
 879 static int alloc_pte_page(pmd_t *pmd)
 880 {
 881         pte_t *pte = (pte_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
 882         if (!pte)
 883                 return -1;
 884
 885         set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
 886         return 0;
 887 }
 888
 889 static int alloc_pmd_page(pud_t *pud)
 890 {
 891         pmd_t *pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
 892         if (!pmd)
 893                 return -1;
 894
 895         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
 896         return 0;
 897 }
 898
 899 static void populate_pte(struct cpa_data *cpa,
 900                          unsigned long start, unsigned long end,
 901                          unsigned num_pages, pmd_t *pmd, pgprot_t pgprot)
 902 {
 903         pte_t *pte;
 904
 905         pte = pte_offset_kernel(pmd, start);
 906
 907         while (num_pages-- && start < end) {
 908
 909                 /* deal with the NX bit */
 910                 if (!(pgprot_val(pgprot) & _PAGE_NX))
 911                         cpa->pfn &= ~_PAGE_NX;
 912
 913                 set_pte(pte, pfn_pte(cpa->pfn >> PAGE_SHIFT, pgprot));
 914
 915                 start    += PAGE_SIZE;
 916                 cpa->pfn += PAGE_SIZE;
 917                 pte++;
 918         }
 919 }
 920
 921 static int populate_pmd(struct cpa_data *cpa,
 922                         unsigned long start, unsigned long end,
 923                         unsigned num_pages, pud_t *pud, pgprot_t pgprot)
 924 {
 925         unsigned int cur_pages = 0;
 926         pmd_t *pmd;
 927         pgprot_t pmd_pgprot;
 928
 929         /*
 930          * Not on a 2M boundary?
 931          */
 932         if (start & (PMD_SIZE - 1)) {
 933                 unsigned long pre_end = start + (num_pages << PAGE_SHIFT);
 934                 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
 935
 936                 pre_end   = min_t(unsigned long, pre_end, next_page);
 937                 cur_pages = (pre_end - start) >> PAGE_SHIFT;
 938                 cur_pages = min_t(unsigned int, num_pages, cur_pages);
 939
 940                 /*
 941                  * Need a PTE page?
 942                  */
 943                 pmd = pmd_offset(pud, start);
 944                 if (pmd_none(*pmd))
 945                         if (alloc_pte_page(pmd))
 946                                 return -1;
 947
 948                 populate_pte(cpa, start, pre_end, cur_pages, pmd, pgprot);
 949
 950                 start = pre_end;
 951         }
 952
 953         /*
 954          * We mapped them all?
 955          */
 956         if (num_pages == cur_pages)
 957                 return cur_pages;
 958
 959         pmd_pgprot = pgprot_4k_2_large(pgprot);
 960
 961         while (end - start >= PMD_SIZE) {
 962
 963                 /*
 964                  * We cannot use a 1G page so allocate a PMD page if needed.
 965                  */
 966                 if (pud_none(*pud))
 967                         if (alloc_pmd_page(pud))
 968                                 return -1;
 969
 970                 pmd = pmd_offset(pud, start);
 971
 972                 set_pmd(pmd, __pmd(cpa->pfn | _PAGE_PSE |
 973                                    massage_pgprot(pmd_pgprot)));
 974
 975                 start     += PMD_SIZE;
 976                 cpa->pfn  += PMD_SIZE;
 977                 cur_pages += PMD_SIZE >> PAGE_SHIFT;
 978         }
 979
 980         /*
 981          * Map trailing 4K pages.
 982          */
 983         if (start < end) {
 984                 pmd = pmd_offset(pud, start);
 985                 if (pmd_none(*pmd))
 986                         if (alloc_pte_page(pmd))
 987                                 return -1;
 988
 989                 populate_pte(cpa, start, end, num_pages - cur_pages,
 990                              pmd, pgprot);
 991         }
 992         return num_pages;
 993 }
 994
 995 static int populate_pud(struct cpa_data *cpa, unsigned long start, pgd_t *pgd,
 996                         pgprot_t pgprot)
 997 {
 998         pud_t *pud;
 999         unsigned long end;
1000         int cur_pages = 0;
1001         pgprot_t pud_pgprot;
1002
1003         end = start + (cpa->numpages << PAGE_SHIFT);
1004
1005         /*
1006          * Not on a Gb page boundary? => map everything up to it with
1007          * smaller pages.
1008          */
1009         if (start & (PUD_SIZE - 1)) {
1010                 unsigned long pre_end;
1011                 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1012
1013                 pre_end   = min_t(unsigned long, end, next_page);
1014                 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1015                 cur_pages = min_t(int, (int)cpa->numpages, cur_pages);
1016
1017                 pud = pud_offset(pgd, start);
1018
1019                 /*
1020                  * Need a PMD page?
1021                  */
1022                 if (pud_none(*pud))
1023                         if (alloc_pmd_page(pud))
1024                                 return -1;
1025
1026                 cur_pages = populate_pmd(cpa, start, pre_end, cur_pages,
1027                                          pud, pgprot);
1028                 if (cur_pages < 0)
1029                         return cur_pages;
1030
1031                 start = pre_end;
1032         }
1033
1034         /* We mapped them all? */
1035         if (cpa->numpages == cur_pages)
1036                 return cur_pages;
1037
1038         pud = pud_offset(pgd, start);
1039         pud_pgprot = pgprot_4k_2_large(pgprot);
1040
1041         /*
1042          * Map everything starting from the Gb boundary, possibly with 1G pages
1043          */
1044         while (end - start >= PUD_SIZE) {
1045                 set_pud(pud, __pud(cpa->pfn | _PAGE_PSE |
1046                                    massage_pgprot(pud_pgprot)));
1047
1048                 start     += PUD_SIZE;
1049                 cpa->pfn  += PUD_SIZE;
1050                 cur_pages += PUD_SIZE >> PAGE_SHIFT;
1051                 pud++;
1052         }
1053
1054         /* Map trailing leftover */
1055         if (start < end) {
1056                 int tmp;
1057
1058                 pud = pud_offset(pgd, start);
1059                 if (pud_none(*pud))
1060                         if (alloc_pmd_page(pud))
1061                                 return -1;
1062
1063                 tmp = populate_pmd(cpa, start, end, cpa->numpages - cur_pages,
1064                                    pud, pgprot);
1065                 if (tmp < 0)
1066                         return cur_pages;
1067
1068                 cur_pages += tmp;
1069         }
1070         return cur_pages;
1071 }
1072
1073 /*
1074  * Restrictions for kernel page table do not necessarily apply when mapping in
1075  * an alternate PGD.
1076  */
1077 static int populate_pgd(struct cpa_data *cpa, unsigned long addr)
1078 {
1079         pgprot_t pgprot = __pgprot(_KERNPG_TABLE);
1080         pud_t *pud = NULL;      /* shut up gcc */
1081         pgd_t *pgd_entry;
1082         int ret;
1083
1084         pgd_entry = cpa->pgd + pgd_index(addr);
1085
1086         /*
1087          * Allocate a PUD page and hand it down for mapping.
1088          */
1089         if (pgd_none(*pgd_entry)) {
1090                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
1091                 if (!pud)
1092                         return -1;
1093
1094                 set_pgd(pgd_entry, __pgd(__pa(pud) | _KERNPG_TABLE));
1095         }
1096
1097         pgprot_val(pgprot) &= ~pgprot_val(cpa->mask_clr);
1098         pgprot_val(pgprot) |=  pgprot_val(cpa->mask_set);
1099
1100         ret = populate_pud(cpa, addr, pgd_entry, pgprot);
1101         if (ret < 0) {
1102                 unmap_pgd_range(cpa->pgd, addr,
1103                                 addr + (cpa->numpages << PAGE_SHIFT));
1104                 return ret;
1105         }
1106
1107         cpa->numpages = ret;
1108         return 0;
1109 }
1110
1111 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
1112                                int primary)
1113 {
1114         if (cpa->pgd)
1115                 return populate_pgd(cpa, vaddr);
1116
1117         /*
1118          * Ignore all non primary paths.
1119          */
1120         if (!primary)
1121                 return 0;
1122
1123         /*
1124          * Ignore the NULL PTE for kernel identity mapping, as it is expected
1125          * to have holes.
1126          * Also set numpages to '1' indicating that we processed cpa req for
1127          * one virtual address page and its pfn. TBD: numpages can be set based
1128          * on the initial value and the level returned by lookup_address().
1129          */
1130         if (within(vaddr, PAGE_OFFSET,
1131                    PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
1132                 cpa->numpages = 1;
1133                 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
1134                 return 0;
1135         } else {
1136                 WARN(1, KERN_WARNING "CPA: called for zero pte. "
1137                         "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
1138                         *cpa->vaddr);
1139
1140                 return -EFAULT;
1141         }
1142 }
1143
1144 static int __change_page_attr(struct cpa_data *cpa, int primary)
1145 {
1146         unsigned long address;
1147         int do_split, err;
1148         unsigned int level;
1149         pte_t *kpte, old_pte;
1150
1151         if (cpa->flags & CPA_PAGES_ARRAY) {
1152                 struct page *page = cpa->pages[cpa->curpage];
1153                 if (unlikely(PageHighMem(page)))
1154                         return 0;
1155                 address = (unsigned long)page_address(page);
1156         } else if (cpa->flags & CPA_ARRAY)
1157                 address = cpa->vaddr[cpa->curpage];
1158         else
1159                 address = *cpa->vaddr;
1160 repeat:
1161         kpte = _lookup_address_cpa(cpa, address, &level);
1162         if (!kpte)
1163                 return __cpa_process_fault(cpa, address, primary);
1164
1165         old_pte = *kpte;
1166         if (!pte_val(old_pte))
1167                 return __cpa_process_fault(cpa, address, primary);
1168
1169         if (level == PG_LEVEL_4K) {
1170                 pte_t new_pte;
1171                 pgprot_t new_prot = pte_pgprot(old_pte);
1172                 unsigned long pfn = pte_pfn(old_pte);
1173
1174                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
1175                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
1176
1177                 new_prot = static_protections(new_prot, address, pfn);
1178
1179                 /*
1180                  * Set the GLOBAL flags only if the PRESENT flag is
1181                  * set otherwise pte_present will return true even on
1182                  * a non present pte. The canon_pgprot will clear
1183                  * _PAGE_GLOBAL for the ancient hardware that doesn't
1184                  * support it.
1185                  */
1186                 if (pgprot_val(new_prot) & _PAGE_PRESENT)
1187                         pgprot_val(new_prot) |= _PAGE_GLOBAL;
1188                 else
1189                         pgprot_val(new_prot) &= ~_PAGE_GLOBAL;
1190
1191                 /*
1192                  * We need to keep the pfn from the existing PTE,
1193                  * after all we're only going to change it's attributes
1194                  * not the memory it points to
1195                  */
1196                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
1197                 cpa->pfn = pfn;
1198                 /*
1199                  * Do we really change anything ?
1200                  */
1201                 if (pte_val(old_pte) != pte_val(new_pte)) {
1202                         set_pte_atomic(kpte, new_pte);
1203                         cpa->flags |= CPA_FLUSHTLB;
1204                 }
1205                 cpa->numpages = 1;
1206                 return 0;
1207         }
1208
1209         /*
1210          * Check, whether we can keep the large page intact
1211          * and just change the pte:
1212          */
1213         do_split = try_preserve_large_page(kpte, address, cpa);
1214         /*
1215          * When the range fits into the existing large page,
1216          * return. cp->numpages and cpa->tlbflush have been updated in
1217          * try_large_page:
1218          */
1219         if (do_split <= 0)
1220                 return do_split;
1221
1222         /*
1223          * We have to split the large page:
1224          */
1225         err = split_large_page(cpa, kpte, address);
1226         if (!err) {
1227                 /*
1228                  * Do a global flush tlb after splitting the large page
1229                  * and before we do the actual change page attribute in the PTE.
1230                  *
1231                  * With out this, we violate the TLB application note, that says
1232                  * "The TLBs may contain both ordinary and large-page
1233                  *  translations for a 4-KByte range of linear addresses. This
1234                  *  may occur if software modifies the paging structures so that
1235                  *  the page size used for the address range changes. If the two
1236                  *  translations differ with respect to page frame or attributes
1237                  *  (e.g., permissions), processor behavior is undefined and may
1238                  *  be implementation-specific."
1239                  *
1240                  * We do this global tlb flush inside the cpa_lock, so that we
1241                  * don't allow any other cpu, with stale tlb entries change the
1242                  * page attribute in parallel, that also falls into the
1243                  * just split large page entry.
1244                  */
1245                 flush_tlb_all();
1246                 goto repeat;
1247         }
1248
1249         return err;
1250 }
1251
1252 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
1253
1254 static int cpa_process_alias(struct cpa_data *cpa)
1255 {
1256         struct cpa_data alias_cpa;
1257         unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
1258         unsigned long vaddr;
1259         int ret;
1260
1261         if (!pfn_range_is_mapped(cpa->pfn, cpa->pfn + 1))
1262                 return 0;
1263
1264         /*
1265          * No need to redo, when the primary call touched the direct
1266          * mapping already:
1267          */
1268         if (cpa->flags & CPA_PAGES_ARRAY) {
1269                 struct page *page = cpa->pages[cpa->curpage];
1270                 if (unlikely(PageHighMem(page)))
1271                         return 0;
1272                 vaddr = (unsigned long)page_address(page);
1273         } else if (cpa->flags & CPA_ARRAY)
1274                 vaddr = cpa->vaddr[cpa->curpage];
1275         else
1276                 vaddr = *cpa->vaddr;
1277
1278         if (!(within(vaddr, PAGE_OFFSET,
1279                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
1280
1281                 alias_cpa = *cpa;
1282                 alias_cpa.vaddr = &laddr;
1283                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1284
1285                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
1286                 if (ret)
1287                         return ret;
1288         }
1289
1290 #ifdef CONFIG_X86_64
1291         /*
1292          * If the primary call didn't touch the high mapping already
1293          * and the physical address is inside the kernel map, we need
1294          * to touch the high mapped kernel as well:
1295          */
1296         if (!within(vaddr, (unsigned long)_text, _brk_end) &&
1297             within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn())) {
1298                 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
1299                                                __START_KERNEL_map - phys_base;
1300                 alias_cpa = *cpa;
1301                 alias_cpa.vaddr = &temp_cpa_vaddr;
1302                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1303
1304                 /*
1305                  * The high mapping range is imprecise, so ignore the
1306                  * return value.
1307                  */
1308                 __change_page_attr_set_clr(&alias_cpa, 0);
1309         }
1310 #endif
1311
1312         return 0;
1313 }
1314
1315 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
1316 {
1317         int ret, numpages = cpa->numpages;
1318
1319         while (numpages) {
1320                 /*
1321                  * Store the remaining nr of pages for the large page
1322                  * preservation check.
1323                  */
1324                 cpa->numpages = numpages;
1325                 /* for array changes, we can't use large page */
1326                 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
1327                         cpa->numpages = 1;
1328
1329                 if (!debug_pagealloc)
1330                         spin_lock(&cpa_lock);
1331                 ret = __change_page_attr(cpa, checkalias);
1332                 if (!debug_pagealloc)
1333                         spin_unlock(&cpa_lock);
1334                 if (ret)
1335                         return ret;
1336
1337                 if (checkalias) {
1338                         ret = cpa_process_alias(cpa);
1339                         if (ret)
1340                                 return ret;
1341                 }
1342
1343                 /*
1344                  * Adjust the number of pages with the result of the
1345                  * CPA operation. Either a large page has been
1346                  * preserved or a single page update happened.
1347                  */
1348                 BUG_ON(cpa->numpages > numpages);
1349                 numpages -= cpa->numpages;
1350                 if (cpa->flags & (CPA_PAGES_ARRAY | CPA_ARRAY))
1351                         cpa->curpage++;
1352                 else
1353                         *cpa->vaddr += cpa->numpages * PAGE_SIZE;
1354
1355         }
1356         return 0;
1357 }
1358
1359 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
1360                                     pgprot_t mask_set, pgprot_t mask_clr,
1361                                     int force_split, int in_flag,
1362                                     struct page **pages)
1363 {
1364         struct cpa_data cpa;
1365         int ret, cache, checkalias;
1366         unsigned long baddr = 0;
1367
1368         memset(&cpa, 0, sizeof(cpa));
1369
1370         /*
1371          * Check, if we are requested to change a not supported
1372          * feature:
1373          */
1374         mask_set = canon_pgprot(mask_set);
1375         mask_clr = canon_pgprot(mask_clr);
1376         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
1377                 return 0;
1378
1379         /* Ensure we are PAGE_SIZE aligned */
1380         if (in_flag & CPA_ARRAY) {
1381                 int i;
1382                 for (i = 0; i < numpages; i++) {
1383                         if (addr[i] & ~PAGE_MASK) {
1384                                 addr[i] &= PAGE_MASK;
1385                                 WARN_ON_ONCE(1);
1386                         }
1387                 }
1388         } else if (!(in_flag & CPA_PAGES_ARRAY)) {
1389                 /*
1390                  * in_flag of CPA_PAGES_ARRAY implies it is aligned.
1391                  * No need to cehck in that case
1392                  */
1393                 if (*addr & ~PAGE_MASK) {
1394                         *addr &= PAGE_MASK;
1395                         /*
1396                          * People should not be passing in unaligned addresses:
1397                          */
1398                         WARN_ON_ONCE(1);
1399                 }
1400                 /*
1401                  * Save address for cache flush. *addr is modified in the call
1402                  * to __change_page_attr_set_clr() below.
1403                  */
1404                 baddr = *addr;
1405         }
1406
1407         /* Must avoid aliasing mappings in the highmem code */
1408         kmap_flush_unused();
1409
1410         vm_unmap_aliases();
1411
1412         cpa.vaddr = addr;
1413         cpa.pages = pages;
1414         cpa.numpages = numpages;
1415         cpa.mask_set = mask_set;
1416         cpa.mask_clr = mask_clr;
1417         cpa.flags = 0;
1418         cpa.curpage = 0;
1419         cpa.force_split = force_split;
1420
1421         if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
1422                 cpa.flags |= in_flag;
1423
1424         /* No alias checking for _NX bit modifications */
1425         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
1426
1427         ret = __change_page_attr_set_clr(&cpa, checkalias);
1428
1429         /*
1430          * Check whether we really changed something:
1431          */
1432         if (!(cpa.flags & CPA_FLUSHTLB))
1433                 goto out;
1434
1435         /*
1436          * No need to flush, when we did not set any of the caching
1437          * attributes:
1438          */
1439         cache = !!pgprot2cachemode(mask_set);
1440
1441         /*
1442          * On success we use CLFLUSH, when the CPU supports it to
1443          * avoid the WBINVD. If the CPU does not support it and in the
1444          * error case we fall back to cpa_flush_all (which uses
1445          * WBINVD):
1446          */
1447         if (!ret && cpu_has_clflush) {
1448                 if (cpa.flags & (CPA_PAGES_ARRAY | CPA_ARRAY)) {
1449                         cpa_flush_array(addr, numpages, cache,
1450                                         cpa.flags, pages);
1451                 } else
1452                         cpa_flush_range(baddr, numpages, cache);
1453         } else
1454                 cpa_flush_all(cache);
1455
1456 out:
1457         return ret;
1458 }
1459
1460 static inline int change_page_attr_set(unsigned long *addr, int numpages,
1461                                        pgprot_t mask, int array)
1462 {
1463         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
1464                 (array ? CPA_ARRAY : 0), NULL);
1465 }
1466
1467 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
1468                                          pgprot_t mask, int array)
1469 {
1470         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
1471                 (array ? CPA_ARRAY : 0), NULL);
1472 }
1473
1474 static inline int cpa_set_pages_array(struct page **pages, int numpages,
1475                                        pgprot_t mask)
1476 {
1477         return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
1478                 CPA_PAGES_ARRAY, pages);
1479 }
1480
1481 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1482                                          pgprot_t mask)
1483 {
1484         return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1485                 CPA_PAGES_ARRAY, pages);
1486 }
1487
1488 int _set_memory_uc(unsigned long addr, int numpages)
1489 {
1490         /*
1491          * for now UC MINUS. see comments in ioremap_nocache()
1492          * If you really need strong UC use ioremap_uc(), but note
1493          * that you cannot override IO areas with set_memory_*() as
1494          * these helpers cannot work with IO memory.
1495          */
1496         return change_page_attr_set(&addr, numpages,
1497                                     cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1498                                     0);
1499 }
1500
1501 int set_memory_uc(unsigned long addr, int numpages)
1502 {
1503         int ret;
1504
1505         /*
1506          * for now UC MINUS. see comments in ioremap_nocache()
1507          */
1508         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1509                               _PAGE_CACHE_MODE_UC_MINUS, NULL);
1510         if (ret)
1511                 goto out_err;
1512
1513         ret = _set_memory_uc(addr, numpages);
1514         if (ret)
1515                 goto out_free;
1516
1517         return 0;
1518
1519 out_free:
1520         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1521 out_err:
1522         return ret;
1523 }
1524 EXPORT_SYMBOL(set_memory_uc);
1525
1526 static int _set_memory_array(unsigned long *addr, int addrinarray,
1527                 enum page_cache_mode new_type)
1528 {
1529         enum page_cache_mode set_type;
1530         int i, j;
1531         int ret;
1532
1533         for (i = 0; i < addrinarray; i++) {
1534                 ret = reserve_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE,
1535                                         new_type, NULL);
1536                 if (ret)
1537                         goto out_free;
1538         }
1539
1540         /* If WC, set to UC- first and then WC */
1541         set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
1542                                 _PAGE_CACHE_MODE_UC_MINUS : new_type;
1543
1544         ret = change_page_attr_set(addr, addrinarray,
1545                                    cachemode2pgprot(set_type), 1);
1546
1547         if (!ret && new_type == _PAGE_CACHE_MODE_WC)
1548                 ret = change_page_attr_set_clr(addr, addrinarray,
1549                                                cachemode2pgprot(
1550                                                 _PAGE_CACHE_MODE_WC),
1551                                                __pgprot(_PAGE_CACHE_MASK),
1552                                                0, CPA_ARRAY, NULL);
1553         if (ret)
1554                 goto out_free;
1555
1556         return 0;
1557
1558 out_free:
1559         for (j = 0; j < i; j++)
1560                 free_memtype(__pa(addr[j]), __pa(addr[j]) + PAGE_SIZE);
1561
1562         return ret;
1563 }
1564
1565 int set_memory_array_uc(unsigned long *addr, int addrinarray)
1566 {
1567         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_UC_MINUS);
1568 }
1569 EXPORT_SYMBOL(set_memory_array_uc);
1570
1571 int set_memory_array_wc(unsigned long *addr, int addrinarray)
1572 {
1573         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_WC);
1574 }
1575 EXPORT_SYMBOL(set_memory_array_wc);
1576
1577 int set_memory_array_wt(unsigned long *addr, int addrinarray)
1578 {
1579         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_WT);
1580 }
1581 EXPORT_SYMBOL_GPL(set_memory_array_wt);
1582
1583 int _set_memory_wc(unsigned long addr, int numpages)
1584 {
1585         int ret;
1586         unsigned long addr_copy = addr;
1587
1588         ret = change_page_attr_set(&addr, numpages,
1589                                    cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1590                                    0);
1591         if (!ret) {
1592                 ret = change_page_attr_set_clr(&addr_copy, numpages,
1593                                                cachemode2pgprot(
1594                                                 _PAGE_CACHE_MODE_WC),
1595                                                __pgprot(_PAGE_CACHE_MASK),
1596                                                0, 0, NULL);
1597         }
1598         return ret;
1599 }
1600
1601 int set_memory_wc(unsigned long addr, int numpages)
1602 {
1603         int ret;
1604
1605         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1606                 _PAGE_CACHE_MODE_WC, NULL);
1607         if (ret)
1608                 return ret;
1609
1610         ret = _set_memory_wc(addr, numpages);
1611         if (ret)
1612                 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1613
1614         return ret;
1615 }
1616 EXPORT_SYMBOL(set_memory_wc);
1617
1618 int _set_memory_wt(unsigned long addr, int numpages)
1619 {
1620         return change_page_attr_set(&addr, numpages,
1621                                     cachemode2pgprot(_PAGE_CACHE_MODE_WT), 0);
1622 }
1623
1624 int set_memory_wt(unsigned long addr, int numpages)
1625 {
1626         int ret;
1627
1628         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1629                               _PAGE_CACHE_MODE_WT, NULL);
1630         if (ret)
1631                 return ret;
1632
1633         ret = _set_memory_wt(addr, numpages);
1634         if (ret)
1635                 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1636
1637         return ret;
1638 }
1639 EXPORT_SYMBOL_GPL(set_memory_wt);
1640
1641 int _set_memory_wb(unsigned long addr, int numpages)
1642 {
1643         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1644         return change_page_attr_clear(&addr, numpages,
1645                                       __pgprot(_PAGE_CACHE_MASK), 0);
1646 }
1647
1648 int set_memory_wb(unsigned long addr, int numpages)
1649 {
1650         int ret;
1651
1652         ret = _set_memory_wb(addr, numpages);
1653         if (ret)
1654                 return ret;
1655
1656         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1657         return 0;
1658 }
1659 EXPORT_SYMBOL(set_memory_wb);
1660
1661 int set_memory_array_wb(unsigned long *addr, int addrinarray)
1662 {
1663         int i;
1664         int ret;
1665
1666         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1667         ret = change_page_attr_clear(addr, addrinarray,
1668                                       __pgprot(_PAGE_CACHE_MASK), 1);
1669         if (ret)
1670                 return ret;
1671
1672         for (i = 0; i < addrinarray; i++)
1673                 free_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE);
1674
1675         return 0;
1676 }
1677 EXPORT_SYMBOL(set_memory_array_wb);
1678
1679 int set_memory_x(unsigned long addr, int numpages)
1680 {
1681         if (!(__supported_pte_mask & _PAGE_NX))
1682                 return 0;
1683
1684         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1685 }
1686 EXPORT_SYMBOL(set_memory_x);
1687
1688 int set_memory_nx(unsigned long addr, int numpages)
1689 {
1690         if (!(__supported_pte_mask & _PAGE_NX))
1691                 return 0;
1692
1693         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1694 }
1695 EXPORT_SYMBOL(set_memory_nx);
1696
1697 int set_memory_ro(unsigned long addr, int numpages)
1698 {
1699         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1700 }
1701
1702 int set_memory_rw(unsigned long addr, int numpages)
1703 {
1704         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1705 }
1706
1707 int set_memory_np(unsigned long addr, int numpages)
1708 {
1709         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1710 }
1711
1712 int set_memory_4k(unsigned long addr, int numpages)
1713 {
1714         return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1715                                         __pgprot(0), 1, 0, NULL);
1716 }
1717
1718 int set_pages_uc(struct page *page, int numpages)
1719 {
1720         unsigned long addr = (unsigned long)page_address(page);
1721
1722         return set_memory_uc(addr, numpages);
1723 }
1724 EXPORT_SYMBOL(set_pages_uc);
1725
1726 static int _set_pages_array(struct page **pages, int addrinarray,
1727                 enum page_cache_mode new_type)
1728 {
1729         unsigned long start;
1730         unsigned long end;
1731         enum page_cache_mode set_type;
1732         int i;
1733         int free_idx;
1734         int ret;
1735
1736         for (i = 0; i < addrinarray; i++) {
1737                 if (PageHighMem(pages[i]))
1738                         continue;
1739                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1740                 end = start + PAGE_SIZE;
1741                 if (reserve_memtype(start, end, new_type, NULL))
1742                         goto err_out;
1743         }
1744
1745         /* If WC, set to UC- first and then WC */
1746         set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
1747                                 _PAGE_CACHE_MODE_UC_MINUS : new_type;
1748
1749         ret = cpa_set_pages_array(pages, addrinarray,
1750                                   cachemode2pgprot(set_type));
1751         if (!ret && new_type == _PAGE_CACHE_MODE_WC)
1752                 ret = change_page_attr_set_clr(NULL, addrinarray,
1753                                                cachemode2pgprot(
1754                                                 _PAGE_CACHE_MODE_WC),
1755                                                __pgprot(_PAGE_CACHE_MASK),
1756                                                0, CPA_PAGES_ARRAY, pages);
1757         if (ret)
1758                 goto err_out;
1759         return 0; /* Success */
1760 err_out:
1761         free_idx = i;
1762         for (i = 0; i < free_idx; i++) {
1763                 if (PageHighMem(pages[i]))
1764                         continue;
1765                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1766                 end = start + PAGE_SIZE;
1767                 free_memtype(start, end);
1768         }
1769         return -EINVAL;
1770 }
1771
1772 int set_pages_array_uc(struct page **pages, int addrinarray)
1773 {
1774         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_UC_MINUS);
1775 }
1776 EXPORT_SYMBOL(set_pages_array_uc);
1777
1778 int set_pages_array_wc(struct page **pages, int addrinarray)
1779 {
1780         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_WC);
1781 }
1782 EXPORT_SYMBOL(set_pages_array_wc);
1783
1784 int set_pages_array_wt(struct page **pages, int addrinarray)
1785 {
1786         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_WT);
1787 }
1788 EXPORT_SYMBOL_GPL(set_pages_array_wt);
1789
1790 int set_pages_wb(struct page *page, int numpages)
1791 {
1792         unsigned long addr = (unsigned long)page_address(page);
1793
1794         return set_memory_wb(addr, numpages);
1795 }
1796 EXPORT_SYMBOL(set_pages_wb);
1797
1798 int set_pages_array_wb(struct page **pages, int addrinarray)
1799 {
1800         int retval;
1801         unsigned long start;
1802         unsigned long end;
1803         int i;
1804
1805         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1806         retval = cpa_clear_pages_array(pages, addrinarray,
1807                         __pgprot(_PAGE_CACHE_MASK));
1808         if (retval)
1809                 return retval;
1810
1811         for (i = 0; i < addrinarray; i++) {
1812                 if (PageHighMem(pages[i]))
1813                         continue;
1814                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1815                 end = start + PAGE_SIZE;
1816                 free_memtype(start, end);
1817         }
1818
1819         return 0;
1820 }
1821 EXPORT_SYMBOL(set_pages_array_wb);
1822
1823 int set_pages_x(struct page *page, int numpages)
1824 {
1825         unsigned long addr = (unsigned long)page_address(page);
1826
1827         return set_memory_x(addr, numpages);
1828 }
1829 EXPORT_SYMBOL(set_pages_x);
1830
1831 int set_pages_nx(struct page *page, int numpages)
1832 {
1833         unsigned long addr = (unsigned long)page_address(page);
1834
1835         return set_memory_nx(addr, numpages);
1836 }
1837 EXPORT_SYMBOL(set_pages_nx);
1838
1839 int set_pages_ro(struct page *page, int numpages)
1840 {
1841         unsigned long addr = (unsigned long)page_address(page);
1842
1843         return set_memory_ro(addr, numpages);
1844 }
1845
1846 int set_pages_rw(struct page *page, int numpages)
1847 {
1848         unsigned long addr = (unsigned long)page_address(page);
1849
1850         return set_memory_rw(addr, numpages);
1851 }
1852
1853 #ifdef CONFIG_DEBUG_PAGEALLOC
1854
1855 static int __set_pages_p(struct page *page, int numpages)
1856 {
1857         unsigned long tempaddr = (unsigned long) page_address(page);
1858         struct cpa_data cpa = { .vaddr = &tempaddr,
1859                                 .pgd = NULL,
1860                                 .numpages = numpages,
1861                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1862                                 .mask_clr = __pgprot(0),
1863                                 .flags = 0};
1864
1865         /*
1866          * No alias checking needed for setting present flag. otherwise,
1867          * we may need to break large pages for 64-bit kernel text
1868          * mappings (this adds to complexity if we want to do this from
1869          * atomic context especially). Let's keep it simple!
1870          */
1871         return __change_page_attr_set_clr(&cpa, 0);
1872 }
1873
1874 static int __set_pages_np(struct page *page, int numpages)
1875 {
1876         unsigned long tempaddr = (unsigned long) page_address(page);
1877         struct cpa_data cpa = { .vaddr = &tempaddr,
1878                                 .pgd = NULL,
1879                                 .numpages = numpages,
1880                                 .mask_set = __pgprot(0),
1881                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1882                                 .flags = 0};
1883
1884         /*
1885          * No alias checking needed for setting not present flag. otherwise,
1886          * we may need to break large pages for 64-bit kernel text
1887          * mappings (this adds to complexity if we want to do this from
1888          * atomic context especially). Let's keep it simple!
1889          */
1890         return __change_page_attr_set_clr(&cpa, 0);
1891 }
1892
1893 void __kernel_map_pages(struct page *page, int numpages, int enable)
1894 {
1895         if (PageHighMem(page))
1896                 return;
1897         if (!enable) {
1898                 debug_check_no_locks_freed(page_address(page),
1899                                            numpages * PAGE_SIZE);
1900         }
1901
1902         /*
1903          * The return value is ignored as the calls cannot fail.
1904          * Large pages for identity mappings are not used at boot time
1905          * and hence no memory allocations during large page split.
1906          */
1907         if (enable)
1908                 __set_pages_p(page, numpages);
1909         else
1910                 __set_pages_np(page, numpages);
1911
1912         /*
1913          * We should perform an IPI and flush all tlbs,
1914          * but that can deadlock->flush only current cpu:
1915          */
1916         __flush_tlb_all();
1917
1918         arch_flush_lazy_mmu_mode();
1919 }
1920
1921 #ifdef CONFIG_HIBERNATION
1922
1923 bool kernel_page_present(struct page *page)
1924 {
1925         unsigned int level;
1926         pte_t *pte;
1927
1928         if (PageHighMem(page))
1929                 return false;
1930
1931         pte = lookup_address((unsigned long)page_address(page), &level);
1932         return (pte_val(*pte) & _PAGE_PRESENT);
1933 }
1934
1935 #endif /* CONFIG_HIBERNATION */
1936
1937 #endif /* CONFIG_DEBUG_PAGEALLOC */
1938
1939 int kernel_map_pages_in_pgd(pgd_t *pgd, u64 pfn, unsigned long address,
1940                             unsigned numpages, unsigned long page_flags)
1941 {
1942         int retval = -EINVAL;
1943
1944         struct cpa_data cpa = {
1945                 .vaddr = &address,
1946                 .pfn = pfn,
1947                 .pgd = pgd,
1948                 .numpages = numpages,
1949                 .mask_set = __pgprot(0),
1950                 .mask_clr = __pgprot(0),
1951                 .flags = 0,
1952         };
1953
1954         if (!(__supported_pte_mask & _PAGE_NX))
1955                 goto out;
1956
1957         if (!(page_flags & _PAGE_NX))
1958                 cpa.mask_clr = __pgprot(_PAGE_NX);
1959
1960         cpa.mask_set = __pgprot(_PAGE_PRESENT | page_flags);
1961
1962         retval = __change_page_attr_set_clr(&cpa, 0);
1963         __flush_tlb_all();
1964
1965 out:
1966         return retval;
1967 }
1968
1969 void kernel_unmap_pages_in_pgd(pgd_t *root, unsigned long address,
1970                                unsigned numpages)
1971 {
1972         unmap_pgd_range(root, address, address + (numpages << PAGE_SHIFT));
1973 }
1974
1975 /*
1976  * The testcases use internal knowledge of the implementation that shouldn't
1977  * be exposed to the rest of the kernel. Include these directly here.
1978  */
1979 #ifdef CONFIG_CPA_DEBUG
1980 #include "pageattr-test.c"
1981 #endif