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