arch/x86/mm/pageattr.c

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