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