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