arch/x86/mm/pageattr.c

   1 /*
   2  * Copyright 2002 Andi Kleen, SuSE Labs.
   3  * Thanks to Ben LaHaise for precious feedback.
   4  */
   5 #include <linux/highmem.h>
   6 #include <linux/bootmem.h>
   7 #include <linux/sched.h>
   8 #include <linux/mm.h>
   9 #include <linux/interrupt.h>
  10 #include <linux/seq_file.h>
  11 #include <linux/debugfs.h>
  12 #include <linux/pfn.h>
  13 #include <linux/percpu.h>
  14 #include <linux/gfp.h>
  15 #include <linux/pci.h>
  16 #include <linux/vmalloc.h>
  17
  18 #include <asm/e820.h>
  19 #include <asm/processor.h>
  20 #include <asm/tlbflush.h>
  21 #include <asm/sections.h>
  22 #include <asm/setup.h>
  23 #include <asm/uaccess.h>
  24 #include <asm/pgalloc.h>
  25 #include <asm/proto.h>
  26 #include <asm/pat.h>
  27
  28 /*
  29  * The current flushing context - we pass it instead of 5 arguments:
  30  */
  31 struct cpa_data {
  32         unsigned long   *vaddr;
  33         pgd_t           *pgd;
  34         pgprot_t        mask_set;
  35         pgprot_t        mask_clr;
  36         unsigned long   numpages;
  37         int             flags;
  38         unsigned long   pfn;
  39         unsigned        force_split : 1;
  40         int             curpage;
  41         struct page     **pages;
  42 };
  43
  44 /*
  45  * Serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity mappings)
  46  * using cpa_lock. So that we don't allow any other cpu, with stale large tlb
  47  * entries change the page attribute in parallel to some other cpu
  48  * splitting a large page entry along with changing the attribute.
  49  */
  50 static DEFINE_SPINLOCK(cpa_lock);
  51
  52 #define CPA_FLUSHTLB 1
  53 #define CPA_ARRAY 2
  54 #define CPA_PAGES_ARRAY 4
  55 #define CPA_FREE_PAGETABLES 8
  56
  57 #ifdef CONFIG_PROC_FS
  58 static unsigned long direct_pages_count[PG_LEVEL_NUM];
  59
  60 void update_page_count(int level, unsigned long pages)
  61 {
  62         /* Protect against CPA */
  63         spin_lock(&pgd_lock);
  64         direct_pages_count[level] += pages;
  65         spin_unlock(&pgd_lock);
  66 }
  67
  68 static void split_page_count(int level)
  69 {
  70         direct_pages_count[level]--;
  71         direct_pages_count[level - 1] += PTRS_PER_PTE;
  72 }
  73
  74 void arch_report_meminfo(struct seq_file *m)
  75 {
  76         seq_printf(m, "DirectMap4k:    %8lu kB\n",
  77                         direct_pages_count[PG_LEVEL_4K] << 2);
  78 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
  79         seq_printf(m, "DirectMap2M:    %8lu kB\n",
  80                         direct_pages_count[PG_LEVEL_2M] << 11);
  81 #else
  82         seq_printf(m, "DirectMap4M:    %8lu kB\n",
  83                         direct_pages_count[PG_LEVEL_2M] << 12);
  84 #endif
  85         if (direct_gbpages)
  86                 seq_printf(m, "DirectMap1G:    %8lu kB\n",
  87                         direct_pages_count[PG_LEVEL_1G] << 20);
  88 }
  89 #else
  90 static inline void split_page_count(int level) { }
  91 #endif
  92
  93 #ifdef CONFIG_X86_64
  94
  95 static inline unsigned long highmap_start_pfn(void)
  96 {
  97         return __pa_symbol(_text) >> PAGE_SHIFT;
  98 }
  99
 100 static inline unsigned long highmap_end_pfn(void)
 101 {
 102         return __pa_symbol(roundup(_brk_end, PMD_SIZE)) >> PAGE_SHIFT;
 103 }
 104
 105 #endif
 106
 107 #ifdef CONFIG_DEBUG_PAGEALLOC
 108 # define debug_pagealloc 1
 109 #else
 110 # define debug_pagealloc 0
 111 #endif
 112
 113 static inline int
 114 within(unsigned long addr, unsigned long start, unsigned long end)
 115 {
 116         return addr >= start && addr < end;
 117 }
 118
 119 /*
 120  * Flushing functions
 121  */
 122
 123 /**
 124  * clflush_cache_range - flush a cache range with clflush
 125  * @vaddr:      virtual start address
 126  * @size:       number of bytes to flush
 127  *
 128  * clflushopt is an unordered instruction which needs fencing with mfence or
 129  * sfence to avoid ordering issues.
 130  */
 131 void clflush_cache_range(void *vaddr, unsigned int size)
 132 {
 133         unsigned long clflush_mask = boot_cpu_data.x86_clflush_size - 1;
 134         void *vend = vaddr + size;
 135         void *p;
 136
 137         mb();
 138
 139         for (p = (void *)((unsigned long)vaddr & ~clflush_mask);
 140              p < vend; p += boot_cpu_data.x86_clflush_size)
 141                 clflushopt(p);
 142
 143         mb();
 144 }
 145 EXPORT_SYMBOL_GPL(clflush_cache_range);
 146
 147 static void __cpa_flush_all(void *arg)
 148 {
 149         unsigned long cache = (unsigned long)arg;
 150
 151         /*
 152          * Flush all to work around Errata in early athlons regarding
 153          * large page flushing.
 154          */
 155         __flush_tlb_all();
 156
 157         if (cache && boot_cpu_data.x86 >= 4)
 158                 wbinvd();
 159 }
 160
 161 static void cpa_flush_all(unsigned long cache)
 162 {
 163         BUG_ON(irqs_disabled());
 164
 165         on_each_cpu(__cpa_flush_all, (void *) cache, 1);
 166 }
 167
 168 static void __cpa_flush_range(void *arg)
 169 {
 170         /*
 171          * We could optimize that further and do individual per page
 172          * tlb invalidates for a low number of pages. Caveat: we must
 173          * flush the high aliases on 64bit as well.
 174          */
 175         __flush_tlb_all();
 176 }
 177
 178 static void cpa_flush_range(unsigned long start, int numpages, int cache)
 179 {
 180         unsigned int i, level;
 181         unsigned long addr;
 182
 183         BUG_ON(irqs_disabled());
 184         WARN_ON(PAGE_ALIGN(start) != start);
 185
 186         on_each_cpu(__cpa_flush_range, NULL, 1);
 187
 188         if (!cache)
 189                 return;
 190
 191         /*
 192          * We only need to flush on one CPU,
 193          * clflush is a MESI-coherent instruction that
 194          * will cause all other CPUs to flush the same
 195          * cachelines:
 196          */
 197         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
 198                 pte_t *pte = lookup_address(addr, &level);
 199
 200                 /*
 201                  * Only flush present addresses:
 202                  */
 203                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 204                         clflush_cache_range((void *) addr, PAGE_SIZE);
 205         }
 206 }
 207
 208 static void cpa_flush_array(unsigned long *start, int numpages, int cache,
 209                             int in_flags, struct page **pages)
 210 {
 211         unsigned int i, level;
 212         unsigned long do_wbinvd = cache && numpages >= 1024; /* 4M threshold */
 213
 214         BUG_ON(irqs_disabled());
 215
 216         on_each_cpu(__cpa_flush_all, (void *) do_wbinvd, 1);
 217
 218         if (!cache || do_wbinvd)
 219                 return;
 220
 221         /*
 222          * We only need to flush on one CPU,
 223          * clflush is a MESI-coherent instruction that
 224          * will cause all other CPUs to flush the same
 225          * cachelines:
 226          */
 227         for (i = 0; i < numpages; i++) {
 228                 unsigned long addr;
 229                 pte_t *pte;
 230
 231                 if (in_flags & CPA_PAGES_ARRAY)
 232                         addr = (unsigned long)page_address(pages[i]);
 233                 else
 234                         addr = start[i];
 235
 236                 pte = lookup_address(addr, &level);
 237
 238                 /*
 239                  * Only flush present addresses:
 240                  */
 241                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 242                         clflush_cache_range((void *)addr, PAGE_SIZE);
 243         }
 244 }
 245
 246 /*
 247  * Certain areas of memory on x86 require very specific protection flags,
 248  * for example the BIOS area or kernel text. Callers don't always get this
 249  * right (again, ioremap() on BIOS memory is not uncommon) so this function
 250  * checks and fixes these known static required protection bits.
 251  */
 252 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
 253                                    unsigned long pfn)
 254 {
 255         pgprot_t forbidden = __pgprot(0);
 256
 257         /*
 258          * The BIOS area between 640k and 1Mb needs to be executable for
 259          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
 260          */
 261 #ifdef CONFIG_PCI_BIOS
 262         if (pcibios_enabled && within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
 263                 pgprot_val(forbidden) |= _PAGE_NX;
 264 #endif
 265
 266         /*
 267          * The kernel text needs to be executable for obvious reasons
 268          * Does not cover __inittext since that is gone later on. On
 269          * 64bit we do not enforce !NX on the low mapping
 270          */
 271         if (within(address, (unsigned long)_text, (unsigned long)_etext))
 272                 pgprot_val(forbidden) |= _PAGE_NX;
 273
 274         /*
 275          * The .rodata section needs to be read-only. Using the pfn
 276          * catches all aliases.
 277          */
 278         if (within(pfn, __pa_symbol(__start_rodata) >> PAGE_SHIFT,
 279                    __pa_symbol(__end_rodata) >> PAGE_SHIFT))
 280                 pgprot_val(forbidden) |= _PAGE_RW;
 281
 282 #if defined(CONFIG_X86_64) && defined(CONFIG_DEBUG_RODATA)
 283         /*
 284          * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
 285          * kernel text mappings for the large page aligned text, rodata sections
 286          * will be always read-only. For the kernel identity mappings covering
 287          * the holes caused by this alignment can be anything that user asks.
 288          *
 289          * This will preserve the large page mappings for kernel text/data
 290          * at no extra cost.
 291          */
 292         if (kernel_set_to_readonly &&
 293             within(address, (unsigned long)_text,
 294                    (unsigned long)__end_rodata_hpage_align)) {
 295                 unsigned int level;
 296
 297                 /*
 298                  * Don't enforce the !RW mapping for the kernel text mapping,
 299                  * if the current mapping is already using small page mapping.
 300                  * No need to work hard to preserve large page mappings in this
 301                  * case.
 302                  *
 303                  * This also fixes the Linux Xen paravirt guest boot failure
 304                  * (because of unexpected read-only mappings for kernel identity
 305                  * mappings). In this paravirt guest case, the kernel text
 306                  * mapping and the kernel identity mapping share the same
 307                  * page-table pages. Thus we can't really use different
 308                  * protections for the kernel text and identity mappings. Also,
 309                  * these shared mappings are made of small page mappings.
 310                  * Thus this don't enforce !RW mapping for small page kernel
 311                  * text mapping logic will help Linux Xen parvirt guest boot
 312                  * as well.
 313                  */
 314                 if (lookup_address(address, &level) && (level != PG_LEVEL_4K))
 315                         pgprot_val(forbidden) |= _PAGE_RW;
 316         }
 317 #endif
 318
 319         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
 320
 321         return prot;
 322 }
 323
 324 /*
 325  * Lookup the page table entry for a virtual address in a specific pgd.
 326  * Return a pointer to the entry and the level of the mapping.
 327  */
 328 pte_t *lookup_address_in_pgd(pgd_t *pgd, unsigned long address,
 329                              unsigned int *level)
 330 {
 331         pud_t *pud;
 332         pmd_t *pmd;
 333
 334         *level = PG_LEVEL_NONE;
 335
 336         if (pgd_none(*pgd))
 337                 return NULL;
 338
 339         pud = pud_offset(pgd, address);
 340         if (pud_none(*pud))
 341                 return NULL;
 342
 343         *level = PG_LEVEL_1G;
 344         if (pud_large(*pud) || !pud_present(*pud))
 345                 return (pte_t *)pud;
 346
 347         pmd = pmd_offset(pud, address);
 348         if (pmd_none(*pmd))
 349                 return NULL;
 350
 351         *level = PG_LEVEL_2M;
 352         if (pmd_large(*pmd) || !pmd_present(*pmd))
 353                 return (pte_t *)pmd;
 354
 355         *level = PG_LEVEL_4K;
 356
 357         return pte_offset_kernel(pmd, address);
 358 }
 359
 360 /*
 361  * Lookup the page table entry for a virtual address. Return a pointer
 362  * to the entry and the level of the mapping.
 363  *
 364  * Note: We return pud and pmd either when the entry is marked large
 365  * or when the present bit is not set. Otherwise we would return a
 366  * pointer to a nonexisting mapping.
 367  */
 368 pte_t *lookup_address(unsigned long address, unsigned int *level)
 369 {
 370         return lookup_address_in_pgd(pgd_offset_k(address), address, level);
 371 }
 372 EXPORT_SYMBOL_GPL(lookup_address);
 373
 374 static pte_t *_lookup_address_cpa(struct cpa_data *cpa, unsigned long address,
 375                                   unsigned int *level)
 376 {
 377         if (cpa->pgd)
 378                 return lookup_address_in_pgd(cpa->pgd + pgd_index(address),
 379                                                address, level);
 380
 381         return lookup_address(address, level);
 382 }
 383
 384 /*
 385  * Lookup the PMD entry for a virtual address. Return a pointer to the entry
 386  * or NULL if not present.
 387  */
 388 pmd_t *lookup_pmd_address(unsigned long address)
 389 {
 390         pgd_t *pgd;
 391         pud_t *pud;
 392
 393         pgd = pgd_offset_k(address);
 394         if (pgd_none(*pgd))
 395                 return NULL;
 396
 397         pud = pud_offset(pgd, address);
 398         if (pud_none(*pud) || pud_large(*pud) || !pud_present(*pud))
 399                 return NULL;
 400
 401         return pmd_offset(pud, address);
 402 }
 403
 404 /*
 405  * This is necessary because __pa() does not work on some
 406  * kinds of memory, like vmalloc() or the alloc_remap()
 407  * areas on 32-bit NUMA systems.  The percpu areas can
 408  * end up in this kind of memory, for instance.
 409  *
 410  * This could be optimized, but it is only intended to be
 411  * used at inititalization time, and keeping it
 412  * unoptimized should increase the testing coverage for
 413  * the more obscure platforms.
 414  */
 415 phys_addr_t slow_virt_to_phys(void *__virt_addr)
 416 {
 417         unsigned long virt_addr = (unsigned long)__virt_addr;
 418         phys_addr_t phys_addr;
 419         unsigned long offset;
 420         enum pg_level level;
 421         pte_t *pte;
 422
 423         pte = lookup_address(virt_addr, &level);
 424         BUG_ON(!pte);
 425
 426         /*
 427          * pXX_pfn() returns unsigned long, which must be cast to phys_addr_t
 428          * before being left-shifted PAGE_SHIFT bits -- this trick is to
 429          * make 32-PAE kernel work correctly.
 430          */
 431         switch (level) {
 432         case PG_LEVEL_1G:
 433                 phys_addr = (phys_addr_t)pud_pfn(*(pud_t *)pte) << PAGE_SHIFT;
 434                 offset = virt_addr & ~PUD_PAGE_MASK;
 435                 break;
 436         case PG_LEVEL_2M:
 437                 phys_addr = (phys_addr_t)pmd_pfn(*(pmd_t *)pte) << PAGE_SHIFT;
 438                 offset = virt_addr & ~PMD_PAGE_MASK;
 439                 break;
 440         default:
 441                 phys_addr = (phys_addr_t)pte_pfn(*pte) << PAGE_SHIFT;
 442                 offset = virt_addr & ~PAGE_MASK;
 443         }
 444
 445         return (phys_addr_t)(phys_addr | offset);
 446 }
 447 EXPORT_SYMBOL_GPL(slow_virt_to_phys);
 448
 449 /*
 450  * Set the new pmd in all the pgds we know about:
 451  */
 452 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 453 {
 454         /* change init_mm */
 455         set_pte_atomic(kpte, pte);
 456 #ifdef CONFIG_X86_32
 457         if (!SHARED_KERNEL_PMD) {
 458                 struct page *page;
 459
 460                 list_for_each_entry(page, &pgd_list, lru) {
 461                         pgd_t *pgd;
 462                         pud_t *pud;
 463                         pmd_t *pmd;
 464
 465                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 466                         pud = pud_offset(pgd, address);
 467                         pmd = pmd_offset(pud, address);
 468                         set_pte_atomic((pte_t *)pmd, pte);
 469                 }
 470         }
 471 #endif
 472 }
 473
 474 static int
 475 try_preserve_large_page(pte_t *kpte, unsigned long address,
 476                         struct cpa_data *cpa)
 477 {
 478         unsigned long nextpage_addr, numpages, pmask, psize, addr, pfn, old_pfn;
 479         pte_t new_pte, old_pte, *tmp;
 480         pgprot_t old_prot, new_prot, req_prot;
 481         int i, do_split = 1;
 482         enum pg_level level;
 483
 484         if (cpa->force_split)
 485                 return 1;
 486
 487         spin_lock(&pgd_lock);
 488         /*
 489          * Check for races, another CPU might have split this page
 490          * up already:
 491          */
 492         tmp = _lookup_address_cpa(cpa, address, &level);
 493         if (tmp != kpte)
 494                 goto out_unlock;
 495
 496         switch (level) {
 497         case PG_LEVEL_2M:
 498                 old_prot = pmd_pgprot(*(pmd_t *)kpte);
 499                 old_pfn = pmd_pfn(*(pmd_t *)kpte);
 500                 break;
 501         case PG_LEVEL_1G:
 502                 old_prot = pud_pgprot(*(pud_t *)kpte);
 503                 old_pfn = pud_pfn(*(pud_t *)kpte);
 504                 break;
 505         default:
 506                 do_split = -EINVAL;
 507                 goto out_unlock;
 508         }
 509
 510         psize = page_level_size(level);
 511         pmask = page_level_mask(level);
 512
 513         /*
 514          * Calculate the number of pages, which fit into this large
 515          * page starting at address:
 516          */
 517         nextpage_addr = (address + psize) & pmask;
 518         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
 519         if (numpages < cpa->numpages)
 520                 cpa->numpages = numpages;
 521
 522         /*
 523          * We are safe now. Check whether the new pgprot is the same:
 524          * Convert protection attributes to 4k-format, as cpa->mask* are set
 525          * up accordingly.
 526          */
 527         old_pte = *kpte;
 528         req_prot = pgprot_large_2_4k(old_prot);
 529
 530         pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
 531         pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
 532
 533         /*
 534          * req_prot is in format of 4k pages. It must be converted to large
 535          * page format: the caching mode includes the PAT bit located at
 536          * different bit positions in the two formats.
 537          */
 538         req_prot = pgprot_4k_2_large(req_prot);
 539
 540         /*
 541          * Set the PSE and GLOBAL flags only if the PRESENT flag is
 542          * set otherwise pmd_present/pmd_huge will return true even on
 543          * a non present pmd. The canon_pgprot will clear _PAGE_GLOBAL
 544          * for the ancient hardware that doesn't support it.
 545          */
 546         if (pgprot_val(req_prot) & _PAGE_PRESENT)
 547                 pgprot_val(req_prot) |= _PAGE_PSE | _PAGE_GLOBAL;
 548         else
 549                 pgprot_val(req_prot) &= ~(_PAGE_PSE | _PAGE_GLOBAL);
 550
 551         req_prot = canon_pgprot(req_prot);
 552
 553         /*
 554          * old_pfn points to the large page base pfn. So we need
 555          * to add the offset of the virtual address:
 556          */
 557         pfn = old_pfn + ((address & (psize - 1)) >> PAGE_SHIFT);
 558         cpa->pfn = pfn;
 559
 560         new_prot = static_protections(req_prot, address, pfn);
 561
 562         /*
 563          * We need to check the full range, whether
 564          * static_protection() requires a different pgprot for one of
 565          * the pages in the range we try to preserve:
 566          */
 567         addr = address & pmask;
 568         pfn = old_pfn;
 569         for (i = 0; i < (psize >> PAGE_SHIFT); i++, addr += PAGE_SIZE, pfn++) {
 570                 pgprot_t chk_prot = static_protections(req_prot, addr, pfn);
 571
 572                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
 573                         goto out_unlock;
 574         }
 575
 576         /*
 577          * If there are no changes, return. maxpages has been updated
 578          * above:
 579          */
 580         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
 581                 do_split = 0;
 582                 goto out_unlock;
 583         }
 584
 585         /*
 586          * We need to change the attributes. Check, whether we can
 587          * change the large page in one go. We request a split, when
 588          * the address is not aligned and the number of pages is
 589          * smaller than the number of pages in the large page. Note
 590          * that we limited the number of possible pages already to
 591          * the number of pages in the large page.
 592          */
 593         if (address == (address & pmask) && cpa->numpages == (psize >> PAGE_SHIFT)) {
 594                 /*
 595                  * The address is aligned and the number of pages
 596                  * covers the full page.
 597                  */
 598                 new_pte = pfn_pte(old_pfn, new_prot);
 599                 __set_pmd_pte(kpte, address, new_pte);
 600                 cpa->flags |= CPA_FLUSHTLB;
 601                 do_split = 0;
 602         }
 603
 604 out_unlock:
 605         spin_unlock(&pgd_lock);
 606
 607         return do_split;
 608 }
 609
 610 static int
 611 __split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
 612                    struct page *base)
 613 {
 614         pte_t *pbase = (pte_t *)page_address(base);
 615         unsigned long ref_pfn, pfn, pfninc = 1;
 616         unsigned int i, level;
 617         pte_t *tmp;
 618         pgprot_t ref_prot;
 619
 620         spin_lock(&pgd_lock);
 621         /*
 622          * Check for races, another CPU might have split this page
 623          * up for us already:
 624          */
 625         tmp = _lookup_address_cpa(cpa, address, &level);
 626         if (tmp != kpte) {
 627                 spin_unlock(&pgd_lock);
 628                 return 1;
 629         }
 630
 631         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
 632
 633         switch (level) {
 634         case PG_LEVEL_2M:
 635                 ref_prot = pmd_pgprot(*(pmd_t *)kpte);
 636                 /* clear PSE and promote PAT bit to correct position */
 637                 ref_prot = pgprot_large_2_4k(ref_prot);
 638                 ref_pfn = pmd_pfn(*(pmd_t *)kpte);
 639                 break;
 640
 641         case PG_LEVEL_1G:
 642                 ref_prot = pud_pgprot(*(pud_t *)kpte);
 643                 ref_pfn = pud_pfn(*(pud_t *)kpte);
 644                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
 645
 646                 /*
 647                  * Clear the PSE flags if the PRESENT flag is not set
 648                  * otherwise pmd_present/pmd_huge will return true
 649                  * even on a non present pmd.
 650                  */
 651                 if (!(pgprot_val(ref_prot) & _PAGE_PRESENT))
 652                         pgprot_val(ref_prot) &= ~_PAGE_PSE;
 653                 break;
 654
 655         default:
 656                 spin_unlock(&pgd_lock);
 657                 return 1;
 658         }
 659
 660         /*
 661          * Set the GLOBAL flags only if the PRESENT flag is set
 662          * otherwise pmd/pte_present will return true even on a non
 663          * present pmd/pte. The canon_pgprot will clear _PAGE_GLOBAL
 664          * for the ancient hardware that doesn't support it.
 665          */
 666         if (pgprot_val(ref_prot) & _PAGE_PRESENT)
 667                 pgprot_val(ref_prot) |= _PAGE_GLOBAL;
 668         else
 669                 pgprot_val(ref_prot) &= ~_PAGE_GLOBAL;
 670
 671         /*
 672          * Get the target pfn from the original entry:
 673          */
 674         pfn = ref_pfn;
 675         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
 676                 set_pte(&pbase[i], pfn_pte(pfn, canon_pgprot(ref_prot)));
 677
 678         if (virt_addr_valid(address)) {
 679                 unsigned long pfn = PFN_DOWN(__pa(address));
 680
 681                 if (pfn_range_is_mapped(pfn, pfn + 1))
 682                         split_page_count(level);
 683         }
 684
 685         /*
 686          * Install the new, split up pagetable.
 687          *
 688          * We use the standard kernel pagetable protections for the new
 689          * pagetable protections, the actual ptes set above control the
 690          * primary protection behavior:
 691          */
 692         __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
 693
 694         /*
 695          * Intel Atom errata AAH41 workaround.
 696          *
 697          * The real fix should be in hw or in a microcode update, but
 698          * we also probabilistically try to reduce the window of having
 699          * a large TLB mixed with 4K TLBs while instruction fetches are
 700          * going on.
 701          */
 702         __flush_tlb_all();
 703         spin_unlock(&pgd_lock);
 704
 705         return 0;
 706 }
 707
 708 static int split_large_page(struct cpa_data *cpa, pte_t *kpte,
 709                             unsigned long address)
 710 {
 711         struct page *base;
 712
 713         if (!debug_pagealloc)
 714                 spin_unlock(&cpa_lock);
 715         base = alloc_pages(GFP_KERNEL | __GFP_NOTRACK, 0);
 716         if (!debug_pagealloc)
 717                 spin_lock(&cpa_lock);
 718         if (!base)
 719                 return -ENOMEM;
 720
 721         if (__split_large_page(cpa, kpte, address, base))
 722                 __free_page(base);
 723
 724         return 0;
 725 }
 726
 727 static bool try_to_free_pte_page(struct cpa_data *cpa, pte_t *pte)
 728 {
 729         int i;
 730
 731         if (!(cpa->flags & CPA_FREE_PAGETABLES))
 732                 return false;
 733
 734         for (i = 0; i < PTRS_PER_PTE; i++)
 735                 if (!pte_none(pte[i]))
 736                         return false;
 737
 738         free_page((unsigned long)pte);
 739         return true;
 740 }
 741
 742 static bool try_to_free_pmd_page(struct cpa_data *cpa, pmd_t *pmd)
 743 {
 744         int i;
 745
 746         if (!(cpa->flags & CPA_FREE_PAGETABLES))
 747                 return false;
 748
 749         for (i = 0; i < PTRS_PER_PMD; i++)
 750                 if (!pmd_none(pmd[i]))
 751                         return false;
 752
 753         free_page((unsigned long)pmd);
 754         return true;
 755 }
 756
 757 static bool try_to_free_pud_page(pud_t *pud)
 758 {
 759         int i;
 760
 761         for (i = 0; i < PTRS_PER_PUD; i++)
 762                 if (!pud_none(pud[i]))
 763                         return false;
 764
 765         free_page((unsigned long)pud);
 766         return true;
 767 }
 768
 769 static bool unmap_pte_range(struct cpa_data *cpa, pmd_t *pmd,
 770                             unsigned long start,
 771                             unsigned long end)
 772 {
 773         pte_t *pte = pte_offset_kernel(pmd, start);
 774
 775         while (start < end) {
 776                 set_pte(pte, __pte(0));
 777
 778                 start += PAGE_SIZE;
 779                 pte++;
 780         }
 781
 782         if (try_to_free_pte_page(cpa, (pte_t *)pmd_page_vaddr(*pmd))) {
 783                 pmd_clear(pmd);
 784                 return true;
 785         }
 786         return false;
 787 }
 788
 789 static void __unmap_pmd_range(struct cpa_data *cpa, pud_t *pud, pmd_t *pmd,
 790                               unsigned long start, unsigned long end)
 791 {
 792         if (unmap_pte_range(cpa, pmd, start, end))
 793                 if (try_to_free_pmd_page(cpa, (pmd_t *)pud_page_vaddr(*pud)))
 794                         pud_clear(pud);
 795 }
 796
 797 static void unmap_pmd_range(struct cpa_data *cpa, pud_t *pud,
 798                             unsigned long start, unsigned long end)
 799 {
 800         pmd_t *pmd = pmd_offset(pud, start);
 801
 802         /*
 803          * Not on a 2MB page boundary?
 804          */
 805         if (start & (PMD_SIZE - 1)) {
 806                 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
 807                 unsigned long pre_end = min_t(unsigned long, end, next_page);
 808
 809                 __unmap_pmd_range(cpa, pud, pmd, start, pre_end);
 810
 811                 start = pre_end;
 812                 pmd++;
 813         }
 814
 815         /*
 816          * Try to unmap in 2M chunks.
 817          */
 818         while (end - start >= PMD_SIZE) {
 819                 if (pmd_large(*pmd))
 820                         pmd_clear(pmd);
 821                 else
 822                         __unmap_pmd_range(cpa, pud, pmd,
 823                                           start, start + PMD_SIZE);
 824
 825                 start += PMD_SIZE;
 826                 pmd++;
 827         }
 828
 829         /*
 830          * 4K leftovers?
 831          */
 832         if (start < end)
 833                 return __unmap_pmd_range(cpa, pud, pmd, start, end);
 834
 835         /*
 836          * Try again to free the PMD page if haven't succeeded above.
 837          */
 838         if (!pud_none(*pud))
 839                 if (try_to_free_pmd_page(cpa, (pmd_t *)pud_page_vaddr(*pud)))
 840                         pud_clear(pud);
 841 }
 842
 843 static void __unmap_pud_range(struct cpa_data *cpa, pgd_t *pgd,
 844                               unsigned long start,
 845                               unsigned long end)
 846 {
 847         pud_t *pud = pud_offset(pgd, start);
 848
 849         /*
 850          * Not on a GB page boundary?
 851          */
 852         if (start & (PUD_SIZE - 1)) {
 853                 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
 854                 unsigned long pre_end   = min_t(unsigned long, end, next_page);
 855
 856                 unmap_pmd_range(cpa, pud, start, pre_end);
 857
 858                 start = pre_end;
 859                 pud++;
 860         }
 861
 862         /*
 863          * Try to unmap in 1G chunks?
 864          */
 865         while (end - start >= PUD_SIZE) {
 866
 867                 if (pud_large(*pud))
 868                         pud_clear(pud);
 869                 else
 870                         unmap_pmd_range(cpa, pud, start, start + PUD_SIZE);
 871
 872                 start += PUD_SIZE;
 873                 pud++;
 874         }
 875
 876         /*
 877          * 2M leftovers?
 878          */
 879         if (start < end)
 880                 unmap_pmd_range(cpa, pud, start, end);
 881
 882         /*
 883          * No need to try to free the PUD page because we'll free it in
 884          * populate_pgd's error path
 885          */
 886 }
 887
 888 static void unmap_pud_range(pgd_t *pgd, unsigned long start, unsigned long end)
 889 {
 890         struct cpa_data cpa = {
 891                 .flags = CPA_FREE_PAGETABLES,
 892         };
 893
 894         __unmap_pud_range(&cpa, pgd, start, end);
 895 }
 896
 897 void unmap_pud_range_nofree(pgd_t *pgd, unsigned long start, unsigned long end)
 898 {
 899         struct cpa_data cpa = {
 900                 .flags = 0,
 901         };
 902
 903         __unmap_pud_range(&cpa, pgd, start, end);
 904 }
 905
 906 static void unmap_pgd_range(pgd_t *root, unsigned long addr, unsigned long end)
 907 {
 908         pgd_t *pgd_entry = root + pgd_index(addr);
 909
 910         unmap_pud_range(pgd_entry, addr, end);
 911
 912         if (try_to_free_pud_page((pud_t *)pgd_page_vaddr(*pgd_entry)))
 913                 pgd_clear(pgd_entry);
 914 }
 915
 916 static int alloc_pte_page(pmd_t *pmd)
 917 {
 918         pte_t *pte = (pte_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
 919         if (!pte)
 920                 return -1;
 921
 922         set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
 923         return 0;
 924 }
 925
 926 static int alloc_pmd_page(pud_t *pud)
 927 {
 928         pmd_t *pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
 929         if (!pmd)
 930                 return -1;
 931
 932         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
 933         return 0;
 934 }
 935
 936 static void populate_pte(struct cpa_data *cpa,
 937                          unsigned long start, unsigned long end,
 938                          unsigned num_pages, pmd_t *pmd, pgprot_t pgprot)
 939 {
 940         pte_t *pte;
 941
 942         pte = pte_offset_kernel(pmd, start);
 943
 944         while (num_pages-- && start < end) {
 945
 946                 /* deal with the NX bit */
 947                 if (!(pgprot_val(pgprot) & _PAGE_NX))
 948                         cpa->pfn &= ~_PAGE_NX;
 949
 950                 set_pte(pte, pfn_pte(cpa->pfn >> PAGE_SHIFT, pgprot));
 951
 952                 start    += PAGE_SIZE;
 953                 cpa->pfn += PAGE_SIZE;
 954                 pte++;
 955         }
 956 }
 957
 958 static int populate_pmd(struct cpa_data *cpa,
 959                         unsigned long start, unsigned long end,
 960                         unsigned num_pages, pud_t *pud, pgprot_t pgprot)
 961 {
 962         unsigned int cur_pages = 0;
 963         pmd_t *pmd;
 964         pgprot_t pmd_pgprot;
 965
 966         /*
 967          * Not on a 2M boundary?
 968          */
 969         if (start & (PMD_SIZE - 1)) {
 970                 unsigned long pre_end = start + (num_pages << PAGE_SHIFT);
 971                 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
 972
 973                 pre_end   = min_t(unsigned long, pre_end, next_page);
 974                 cur_pages = (pre_end - start) >> PAGE_SHIFT;
 975                 cur_pages = min_t(unsigned int, num_pages, cur_pages);
 976
 977                 /*
 978                  * Need a PTE page?
 979                  */
 980                 pmd = pmd_offset(pud, start);
 981                 if (pmd_none(*pmd))
 982                         if (alloc_pte_page(pmd))
 983                                 return -1;
 984
 985                 populate_pte(cpa, start, pre_end, cur_pages, pmd, pgprot);
 986
 987                 start = pre_end;
 988         }
 989
 990         /*
 991          * We mapped them all?
 992          */
 993         if (num_pages == cur_pages)
 994                 return cur_pages;
 995
 996         pmd_pgprot = pgprot_4k_2_large(pgprot);
 997
 998         while (end - start >= PMD_SIZE) {
 999
1000                 /*
1001                  * We cannot use a 1G page so allocate a PMD page if needed.
1002                  */
1003                 if (pud_none(*pud))
1004                         if (alloc_pmd_page(pud))
1005                                 return -1;
1006
1007                 pmd = pmd_offset(pud, start);
1008
1009                 set_pmd(pmd, __pmd(cpa->pfn | _PAGE_PSE |
1010                                    massage_pgprot(pmd_pgprot)));
1011
1012                 start     += PMD_SIZE;
1013                 cpa->pfn  += PMD_SIZE;
1014                 cur_pages += PMD_SIZE >> PAGE_SHIFT;
1015         }
1016
1017         /*
1018          * Map trailing 4K pages.
1019          */
1020         if (start < end) {
1021                 pmd = pmd_offset(pud, start);
1022                 if (pmd_none(*pmd))
1023                         if (alloc_pte_page(pmd))
1024                                 return -1;
1025
1026                 populate_pte(cpa, start, end, num_pages - cur_pages,
1027                              pmd, pgprot);
1028         }
1029         return num_pages;
1030 }
1031
1032 static int populate_pud(struct cpa_data *cpa, unsigned long start, pgd_t *pgd,
1033                         pgprot_t pgprot)
1034 {
1035         pud_t *pud;
1036         unsigned long end;
1037         int cur_pages = 0;
1038         pgprot_t pud_pgprot;
1039
1040         end = start + (cpa->numpages << PAGE_SHIFT);
1041
1042         /*
1043          * Not on a Gb page boundary? => map everything up to it with
1044          * smaller pages.
1045          */
1046         if (start & (PUD_SIZE - 1)) {
1047                 unsigned long pre_end;
1048                 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1049
1050                 pre_end   = min_t(unsigned long, end, next_page);
1051                 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1052                 cur_pages = min_t(int, (int)cpa->numpages, cur_pages);
1053
1054                 pud = pud_offset(pgd, start);
1055
1056                 /*
1057                  * Need a PMD page?
1058                  */
1059                 if (pud_none(*pud))
1060                         if (alloc_pmd_page(pud))
1061                                 return -1;
1062
1063                 cur_pages = populate_pmd(cpa, start, pre_end, cur_pages,
1064                                          pud, pgprot);
1065                 if (cur_pages < 0)
1066                         return cur_pages;
1067
1068                 start = pre_end;
1069         }
1070
1071         /* We mapped them all? */
1072         if (cpa->numpages == cur_pages)
1073                 return cur_pages;
1074
1075         pud = pud_offset(pgd, start);
1076         pud_pgprot = pgprot_4k_2_large(pgprot);
1077
1078         /*
1079          * Map everything starting from the Gb boundary, possibly with 1G pages
1080          */
1081         while (end - start >= PUD_SIZE) {
1082                 set_pud(pud, __pud(cpa->pfn | _PAGE_PSE |
1083                                    massage_pgprot(pud_pgprot)));
1084
1085                 start     += PUD_SIZE;
1086                 cpa->pfn  += PUD_SIZE;
1087                 cur_pages += PUD_SIZE >> PAGE_SHIFT;
1088                 pud++;
1089         }
1090
1091         /* Map trailing leftover */
1092         if (start < end) {
1093                 int tmp;
1094
1095                 pud = pud_offset(pgd, start);
1096                 if (pud_none(*pud))
1097                         if (alloc_pmd_page(pud))
1098                                 return -1;
1099
1100                 tmp = populate_pmd(cpa, start, end, cpa->numpages - cur_pages,
1101                                    pud, pgprot);
1102                 if (tmp < 0)
1103                         return cur_pages;
1104
1105                 cur_pages += tmp;
1106         }
1107         return cur_pages;
1108 }
1109
1110 /*
1111  * Restrictions for kernel page table do not necessarily apply when mapping in
1112  * an alternate PGD.
1113  */
1114 static int populate_pgd(struct cpa_data *cpa, unsigned long addr)
1115 {
1116         pgprot_t pgprot = __pgprot(_KERNPG_TABLE);
1117         pud_t *pud = NULL;      /* shut up gcc */
1118         pgd_t *pgd_entry;
1119         int ret;
1120
1121         pgd_entry = cpa->pgd + pgd_index(addr);
1122
1123         /*
1124          * Allocate a PUD page and hand it down for mapping.
1125          */
1126         if (pgd_none(*pgd_entry)) {
1127                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
1128                 if (!pud)
1129                         return -1;
1130
1131                 set_pgd(pgd_entry, __pgd(__pa(pud) | _KERNPG_TABLE));
1132         }
1133
1134         pgprot_val(pgprot) &= ~pgprot_val(cpa->mask_clr);
1135         pgprot_val(pgprot) |=  pgprot_val(cpa->mask_set);
1136
1137         ret = populate_pud(cpa, addr, pgd_entry, pgprot);
1138         if (ret < 0) {
1139                 unmap_pgd_range(cpa->pgd, addr,
1140                                 addr + (cpa->numpages << PAGE_SHIFT));
1141                 return ret;
1142         }
1143
1144         cpa->numpages = ret;
1145         return 0;
1146 }
1147
1148 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
1149                                int primary)
1150 {
1151         if (cpa->pgd)
1152                 return populate_pgd(cpa, vaddr);
1153
1154         /*
1155          * Ignore all non primary paths.
1156          */
1157         if (!primary)
1158                 return 0;
1159
1160         /*
1161          * Ignore the NULL PTE for kernel identity mapping, as it is expected
1162          * to have holes.
1163          * Also set numpages to '1' indicating that we processed cpa req for
1164          * one virtual address page and its pfn. TBD: numpages can be set based
1165          * on the initial value and the level returned by lookup_address().
1166          */
1167         if (within(vaddr, PAGE_OFFSET,
1168                    PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
1169                 cpa->numpages = 1;
1170                 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
1171                 return 0;
1172         } else {
1173                 WARN(1, KERN_WARNING "CPA: called for zero pte. "
1174                         "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
1175                         *cpa->vaddr);
1176
1177                 return -EFAULT;
1178         }
1179 }
1180
1181 static int __change_page_attr(struct cpa_data *cpa, int primary)
1182 {
1183         unsigned long address;
1184         int do_split, err;
1185         unsigned int level;
1186         pte_t *kpte, old_pte;
1187
1188         if (cpa->flags & CPA_PAGES_ARRAY) {
1189                 struct page *page = cpa->pages[cpa->curpage];
1190                 if (unlikely(PageHighMem(page)))
1191                         return 0;
1192                 address = (unsigned long)page_address(page);
1193         } else if (cpa->flags & CPA_ARRAY)
1194                 address = cpa->vaddr[cpa->curpage];
1195         else
1196                 address = *cpa->vaddr;
1197 repeat:
1198         kpte = _lookup_address_cpa(cpa, address, &level);
1199         if (!kpte)
1200                 return __cpa_process_fault(cpa, address, primary);
1201
1202         old_pte = *kpte;
1203         if (!pte_val(old_pte))
1204                 return __cpa_process_fault(cpa, address, primary);
1205
1206         if (level == PG_LEVEL_4K) {
1207                 pte_t new_pte;
1208                 pgprot_t new_prot = pte_pgprot(old_pte);
1209                 unsigned long pfn = pte_pfn(old_pte);
1210
1211                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
1212                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
1213
1214                 new_prot = static_protections(new_prot, address, pfn);
1215
1216                 /*
1217                  * Set the GLOBAL flags only if the PRESENT flag is
1218                  * set otherwise pte_present will return true even on
1219                  * a non present pte. The canon_pgprot will clear
1220                  * _PAGE_GLOBAL for the ancient hardware that doesn't
1221                  * support it.
1222                  */
1223                 if (pgprot_val(new_prot) & _PAGE_PRESENT)
1224                         pgprot_val(new_prot) |= _PAGE_GLOBAL;
1225                 else
1226                         pgprot_val(new_prot) &= ~_PAGE_GLOBAL;
1227
1228                 /*
1229                  * We need to keep the pfn from the existing PTE,
1230                  * after all we're only going to change it's attributes
1231                  * not the memory it points to
1232                  */
1233                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
1234                 cpa->pfn = pfn;
1235                 /*
1236                  * Do we really change anything ?
1237                  */
1238                 if (pte_val(old_pte) != pte_val(new_pte)) {
1239                         set_pte_atomic(kpte, new_pte);
1240                         cpa->flags |= CPA_FLUSHTLB;
1241                 }
1242                 cpa->numpages = 1;
1243                 return 0;
1244         }
1245
1246         /*
1247          * Check, whether we can keep the large page intact
1248          * and just change the pte:
1249          */
1250         do_split = try_preserve_large_page(kpte, address, cpa);
1251         /*
1252          * When the range fits into the existing large page,
1253          * return. cp->numpages and cpa->tlbflush have been updated in
1254          * try_large_page:
1255          */
1256         if (do_split <= 0)
1257                 return do_split;
1258
1259         /*
1260          * We have to split the large page:
1261          */
1262         err = split_large_page(cpa, kpte, address);
1263         if (!err) {
1264                 /*
1265                  * Do a global flush tlb after splitting the large page
1266                  * and before we do the actual change page attribute in the PTE.
1267                  *
1268                  * With out this, we violate the TLB application note, that says
1269                  * "The TLBs may contain both ordinary and large-page
1270                  *  translations for a 4-KByte range of linear addresses. This
1271                  *  may occur if software modifies the paging structures so that
1272                  *  the page size used for the address range changes. If the two
1273                  *  translations differ with respect to page frame or attributes
1274                  *  (e.g., permissions), processor behavior is undefined and may
1275                  *  be implementation-specific."
1276                  *
1277                  * We do this global tlb flush inside the cpa_lock, so that we
1278                  * don't allow any other cpu, with stale tlb entries change the
1279                  * page attribute in parallel, that also falls into the
1280                  * just split large page entry.
1281                  */
1282                 flush_tlb_all();
1283                 goto repeat;
1284         }
1285
1286         return err;
1287 }
1288
1289 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
1290
1291 static int cpa_process_alias(struct cpa_data *cpa)
1292 {
1293         struct cpa_data alias_cpa;
1294         unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
1295         unsigned long vaddr;
1296         int ret;
1297
1298         if (!pfn_range_is_mapped(cpa->pfn, cpa->pfn + 1))
1299                 return 0;
1300
1301         /*
1302          * No need to redo, when the primary call touched the direct
1303          * mapping already:
1304          */
1305         if (cpa->flags & CPA_PAGES_ARRAY) {
1306                 struct page *page = cpa->pages[cpa->curpage];
1307                 if (unlikely(PageHighMem(page)))
1308                         return 0;
1309                 vaddr = (unsigned long)page_address(page);
1310         } else if (cpa->flags & CPA_ARRAY)
1311                 vaddr = cpa->vaddr[cpa->curpage];
1312         else
1313                 vaddr = *cpa->vaddr;
1314
1315         if (!(within(vaddr, PAGE_OFFSET,
1316                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
1317
1318                 alias_cpa = *cpa;
1319                 alias_cpa.vaddr = &laddr;
1320                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1321
1322                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
1323                 if (ret)
1324                         return ret;
1325         }
1326
1327 #ifdef CONFIG_X86_64
1328         /*
1329          * If the primary call didn't touch the high mapping already
1330          * and the physical address is inside the kernel map, we need
1331          * to touch the high mapped kernel as well:
1332          */
1333         if (!within(vaddr, (unsigned long)_text, _brk_end) &&
1334             within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn())) {
1335                 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
1336                                                __START_KERNEL_map - phys_base;
1337                 alias_cpa = *cpa;
1338                 alias_cpa.vaddr = &temp_cpa_vaddr;
1339                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1340
1341                 /*
1342                  * The high mapping range is imprecise, so ignore the
1343                  * return value.
1344                  */
1345                 __change_page_attr_set_clr(&alias_cpa, 0);
1346         }
1347 #endif
1348
1349         return 0;
1350 }
1351
1352 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
1353 {
1354         int ret, numpages = cpa->numpages;
1355
1356         while (numpages) {
1357                 /*
1358                  * Store the remaining nr of pages for the large page
1359                  * preservation check.
1360                  */
1361                 cpa->numpages = numpages;
1362                 /* for array changes, we can't use large page */
1363                 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
1364                         cpa->numpages = 1;
1365
1366                 if (!debug_pagealloc)
1367                         spin_lock(&cpa_lock);
1368                 ret = __change_page_attr(cpa, checkalias);
1369                 if (!debug_pagealloc)
1370                         spin_unlock(&cpa_lock);
1371                 if (ret)
1372                         return ret;
1373
1374                 if (checkalias) {
1375                         ret = cpa_process_alias(cpa);
1376                         if (ret)
1377                                 return ret;
1378                 }
1379
1380                 /*
1381                  * Adjust the number of pages with the result of the
1382                  * CPA operation. Either a large page has been
1383                  * preserved or a single page update happened.
1384                  */
1385                 BUG_ON(cpa->numpages > numpages || !cpa->numpages);
1386                 numpages -= cpa->numpages;
1387                 if (cpa->flags & (CPA_PAGES_ARRAY | CPA_ARRAY))
1388                         cpa->curpage++;
1389                 else
1390                         *cpa->vaddr += cpa->numpages * PAGE_SIZE;
1391
1392         }
1393         return 0;
1394 }
1395
1396 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
1397                                     pgprot_t mask_set, pgprot_t mask_clr,
1398                                     int force_split, int in_flag,
1399                                     struct page **pages)
1400 {
1401         struct cpa_data cpa;
1402         int ret, cache, checkalias;
1403         unsigned long baddr = 0;
1404
1405         memset(&cpa, 0, sizeof(cpa));
1406
1407         /*
1408          * Check, if we are requested to change a not supported
1409          * feature:
1410          */
1411         mask_set = canon_pgprot(mask_set);
1412         mask_clr = canon_pgprot(mask_clr);
1413         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
1414                 return 0;
1415
1416         /* Ensure we are PAGE_SIZE aligned */
1417         if (in_flag & CPA_ARRAY) {
1418                 int i;
1419                 for (i = 0; i < numpages; i++) {
1420                         if (addr[i] & ~PAGE_MASK) {
1421                                 addr[i] &= PAGE_MASK;
1422                                 WARN_ON_ONCE(1);
1423                         }
1424                 }
1425         } else if (!(in_flag & CPA_PAGES_ARRAY)) {
1426                 /*
1427                  * in_flag of CPA_PAGES_ARRAY implies it is aligned.
1428                  * No need to cehck in that case
1429                  */
1430                 if (*addr & ~PAGE_MASK) {
1431                         *addr &= PAGE_MASK;
1432                         /*
1433                          * People should not be passing in unaligned addresses:
1434                          */
1435                         WARN_ON_ONCE(1);
1436                 }
1437                 /*
1438                  * Save address for cache flush. *addr is modified in the call
1439                  * to __change_page_attr_set_clr() below.
1440                  */
1441                 baddr = *addr;
1442         }
1443
1444         /* Must avoid aliasing mappings in the highmem code */
1445         kmap_flush_unused();
1446
1447         vm_unmap_aliases();
1448
1449         cpa.vaddr = addr;
1450         cpa.pages = pages;
1451         cpa.numpages = numpages;
1452         cpa.mask_set = mask_set;
1453         cpa.mask_clr = mask_clr;
1454         cpa.flags = 0;
1455         cpa.curpage = 0;
1456         cpa.force_split = force_split;
1457
1458         if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
1459                 cpa.flags |= in_flag;
1460
1461         /* No alias checking for _NX bit modifications */
1462         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
1463
1464         ret = __change_page_attr_set_clr(&cpa, checkalias);
1465
1466         /*
1467          * Check whether we really changed something:
1468          */
1469         if (!(cpa.flags & CPA_FLUSHTLB))
1470                 goto out;
1471
1472         /*
1473          * No need to flush, when we did not set any of the caching
1474          * attributes:
1475          */
1476         cache = !!pgprot2cachemode(mask_set);
1477
1478         /*
1479          * On success we use CLFLUSH, when the CPU supports it to
1480          * avoid the WBINVD. If the CPU does not support it and in the
1481          * error case we fall back to cpa_flush_all (which uses
1482          * WBINVD):
1483          */
1484         if (!ret && cpu_has_clflush) {
1485                 if (cpa.flags & (CPA_PAGES_ARRAY | CPA_ARRAY)) {
1486                         cpa_flush_array(addr, numpages, cache,
1487                                         cpa.flags, pages);
1488                 } else
1489                         cpa_flush_range(baddr, numpages, cache);
1490         } else
1491                 cpa_flush_all(cache);
1492
1493 out:
1494         return ret;
1495 }
1496
1497 static inline int change_page_attr_set(unsigned long *addr, int numpages,
1498                                        pgprot_t mask, int array)
1499 {
1500         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
1501                 (array ? CPA_ARRAY : 0), NULL);
1502 }
1503
1504 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
1505                                          pgprot_t mask, int array)
1506 {
1507         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
1508                 (array ? CPA_ARRAY : 0), NULL);
1509 }
1510
1511 static inline int cpa_set_pages_array(struct page **pages, int numpages,
1512                                        pgprot_t mask)
1513 {
1514         return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
1515                 CPA_PAGES_ARRAY, pages);
1516 }
1517
1518 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1519                                          pgprot_t mask)
1520 {
1521         return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1522                 CPA_PAGES_ARRAY, pages);
1523 }
1524
1525 int _set_memory_uc(unsigned long addr, int numpages)
1526 {
1527         /*
1528          * for now UC MINUS. see comments in ioremap_nocache()
1529          * If you really need strong UC use ioremap_uc(), but note
1530          * that you cannot override IO areas with set_memory_*() as
1531          * these helpers cannot work with IO memory.
1532          */
1533         return change_page_attr_set(&addr, numpages,
1534                                     cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1535                                     0);
1536 }
1537
1538 int set_memory_uc(unsigned long addr, int numpages)
1539 {
1540         int ret;
1541
1542         /*
1543          * for now UC MINUS. see comments in ioremap_nocache()
1544          */
1545         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1546                               _PAGE_CACHE_MODE_UC_MINUS, NULL);
1547         if (ret)
1548                 goto out_err;
1549
1550         ret = _set_memory_uc(addr, numpages);
1551         if (ret)
1552                 goto out_free;
1553
1554         return 0;
1555
1556 out_free:
1557         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1558 out_err:
1559         return ret;
1560 }
1561 EXPORT_SYMBOL(set_memory_uc);
1562
1563 static int _set_memory_array(unsigned long *addr, int addrinarray,
1564                 enum page_cache_mode new_type)
1565 {
1566         enum page_cache_mode set_type;
1567         int i, j;
1568         int ret;
1569
1570         for (i = 0; i < addrinarray; i++) {
1571                 ret = reserve_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE,
1572                                         new_type, NULL);
1573                 if (ret)
1574                         goto out_free;
1575         }
1576
1577         /* If WC, set to UC- first and then WC */
1578         set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
1579                                 _PAGE_CACHE_MODE_UC_MINUS : new_type;
1580
1581         ret = change_page_attr_set(addr, addrinarray,
1582                                    cachemode2pgprot(set_type), 1);
1583
1584         if (!ret && new_type == _PAGE_CACHE_MODE_WC)
1585                 ret = change_page_attr_set_clr(addr, addrinarray,
1586                                                cachemode2pgprot(
1587                                                 _PAGE_CACHE_MODE_WC),
1588                                                __pgprot(_PAGE_CACHE_MASK),
1589                                                0, CPA_ARRAY, NULL);
1590         if (ret)
1591                 goto out_free;
1592
1593         return 0;
1594
1595 out_free:
1596         for (j = 0; j < i; j++)
1597                 free_memtype(__pa(addr[j]), __pa(addr[j]) + PAGE_SIZE);
1598
1599         return ret;
1600 }
1601
1602 int set_memory_array_uc(unsigned long *addr, int addrinarray)
1603 {
1604         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_UC_MINUS);
1605 }
1606 EXPORT_SYMBOL(set_memory_array_uc);
1607
1608 int set_memory_array_wc(unsigned long *addr, int addrinarray)
1609 {
1610         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_WC);
1611 }
1612 EXPORT_SYMBOL(set_memory_array_wc);
1613
1614 int set_memory_array_wt(unsigned long *addr, int addrinarray)
1615 {
1616         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_WT);
1617 }
1618 EXPORT_SYMBOL_GPL(set_memory_array_wt);
1619
1620 int _set_memory_wc(unsigned long addr, int numpages)
1621 {
1622         int ret;
1623         unsigned long addr_copy = addr;
1624
1625         ret = change_page_attr_set(&addr, numpages,
1626                                    cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1627                                    0);
1628         if (!ret) {
1629                 ret = change_page_attr_set_clr(&addr_copy, numpages,
1630                                                cachemode2pgprot(
1631                                                 _PAGE_CACHE_MODE_WC),
1632                                                __pgprot(_PAGE_CACHE_MASK),
1633                                                0, 0, NULL);
1634         }
1635         return ret;
1636 }
1637
1638 int set_memory_wc(unsigned long addr, int numpages)
1639 {
1640         int ret;
1641
1642         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1643                 _PAGE_CACHE_MODE_WC, NULL);
1644         if (ret)
1645                 return ret;
1646
1647         ret = _set_memory_wc(addr, numpages);
1648         if (ret)
1649                 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1650
1651         return ret;
1652 }
1653 EXPORT_SYMBOL(set_memory_wc);
1654
1655 int _set_memory_wt(unsigned long addr, int numpages)
1656 {
1657         return change_page_attr_set(&addr, numpages,
1658                                     cachemode2pgprot(_PAGE_CACHE_MODE_WT), 0);
1659 }
1660
1661 int set_memory_wt(unsigned long addr, int numpages)
1662 {
1663         int ret;
1664
1665         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1666                               _PAGE_CACHE_MODE_WT, NULL);
1667         if (ret)
1668                 return ret;
1669
1670         ret = _set_memory_wt(addr, numpages);
1671         if (ret)
1672                 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1673
1674         return ret;
1675 }
1676 EXPORT_SYMBOL_GPL(set_memory_wt);
1677
1678 int _set_memory_wb(unsigned long addr, int numpages)
1679 {
1680         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1681         return change_page_attr_clear(&addr, numpages,
1682                                       __pgprot(_PAGE_CACHE_MASK), 0);
1683 }
1684
1685 int set_memory_wb(unsigned long addr, int numpages)
1686 {
1687         int ret;
1688
1689         ret = _set_memory_wb(addr, numpages);
1690         if (ret)
1691                 return ret;
1692
1693         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1694         return 0;
1695 }
1696 EXPORT_SYMBOL(set_memory_wb);
1697
1698 int set_memory_array_wb(unsigned long *addr, int addrinarray)
1699 {
1700         int i;
1701         int ret;
1702
1703         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1704         ret = change_page_attr_clear(addr, addrinarray,
1705                                       __pgprot(_PAGE_CACHE_MASK), 1);
1706         if (ret)
1707                 return ret;
1708
1709         for (i = 0; i < addrinarray; i++)
1710                 free_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE);
1711
1712         return 0;
1713 }
1714 EXPORT_SYMBOL(set_memory_array_wb);
1715
1716 int set_memory_x(unsigned long addr, int numpages)
1717 {
1718         if (!(__supported_pte_mask & _PAGE_NX))
1719                 return 0;
1720
1721         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1722 }
1723 EXPORT_SYMBOL(set_memory_x);
1724
1725 int set_memory_nx(unsigned long addr, int numpages)
1726 {
1727         if (!(__supported_pte_mask & _PAGE_NX))
1728                 return 0;
1729
1730         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1731 }
1732 EXPORT_SYMBOL(set_memory_nx);
1733
1734 int set_memory_ro(unsigned long addr, int numpages)
1735 {
1736         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1737 }
1738
1739 int set_memory_rw(unsigned long addr, int numpages)
1740 {
1741         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1742 }
1743
1744 int set_memory_np(unsigned long addr, int numpages)
1745 {
1746         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1747 }
1748
1749 int set_memory_4k(unsigned long addr, int numpages)
1750 {
1751         return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1752                                         __pgprot(0), 1, 0, NULL);
1753 }
1754
1755 int set_pages_uc(struct page *page, int numpages)
1756 {
1757         unsigned long addr = (unsigned long)page_address(page);
1758
1759         return set_memory_uc(addr, numpages);
1760 }
1761 EXPORT_SYMBOL(set_pages_uc);
1762
1763 static int _set_pages_array(struct page **pages, int addrinarray,
1764                 enum page_cache_mode new_type)
1765 {
1766         unsigned long start;
1767         unsigned long end;
1768         enum page_cache_mode set_type;
1769         int i;
1770         int free_idx;
1771         int ret;
1772
1773         for (i = 0; i < addrinarray; i++) {
1774                 if (PageHighMem(pages[i]))
1775                         continue;
1776                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1777                 end = start + PAGE_SIZE;
1778                 if (reserve_memtype(start, end, new_type, NULL))
1779                         goto err_out;
1780         }
1781
1782         /* If WC, set to UC- first and then WC */
1783         set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
1784                                 _PAGE_CACHE_MODE_UC_MINUS : new_type;
1785
1786         ret = cpa_set_pages_array(pages, addrinarray,
1787                                   cachemode2pgprot(set_type));
1788         if (!ret && new_type == _PAGE_CACHE_MODE_WC)
1789                 ret = change_page_attr_set_clr(NULL, addrinarray,
1790                                                cachemode2pgprot(
1791                                                 _PAGE_CACHE_MODE_WC),
1792                                                __pgprot(_PAGE_CACHE_MASK),
1793                                                0, CPA_PAGES_ARRAY, pages);
1794         if (ret)
1795                 goto err_out;
1796         return 0; /* Success */
1797 err_out:
1798         free_idx = i;
1799         for (i = 0; i < free_idx; i++) {
1800                 if (PageHighMem(pages[i]))
1801                         continue;
1802                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1803                 end = start + PAGE_SIZE;
1804                 free_memtype(start, end);
1805         }
1806         return -EINVAL;
1807 }
1808
1809 int set_pages_array_uc(struct page **pages, int addrinarray)
1810 {
1811         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_UC_MINUS);
1812 }
1813 EXPORT_SYMBOL(set_pages_array_uc);
1814
1815 int set_pages_array_wc(struct page **pages, int addrinarray)
1816 {
1817         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_WC);
1818 }
1819 EXPORT_SYMBOL(set_pages_array_wc);
1820
1821 int set_pages_array_wt(struct page **pages, int addrinarray)
1822 {
1823         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_WT);
1824 }
1825 EXPORT_SYMBOL_GPL(set_pages_array_wt);
1826
1827 int set_pages_wb(struct page *page, int numpages)
1828 {
1829         unsigned long addr = (unsigned long)page_address(page);
1830
1831         return set_memory_wb(addr, numpages);
1832 }
1833 EXPORT_SYMBOL(set_pages_wb);
1834
1835 int set_pages_array_wb(struct page **pages, int addrinarray)
1836 {
1837         int retval;
1838         unsigned long start;
1839         unsigned long end;
1840         int i;
1841
1842         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1843         retval = cpa_clear_pages_array(pages, addrinarray,
1844                         __pgprot(_PAGE_CACHE_MASK));
1845         if (retval)
1846                 return retval;
1847
1848         for (i = 0; i < addrinarray; i++) {
1849                 if (PageHighMem(pages[i]))
1850                         continue;
1851                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1852                 end = start + PAGE_SIZE;
1853                 free_memtype(start, end);
1854         }
1855
1856         return 0;
1857 }
1858 EXPORT_SYMBOL(set_pages_array_wb);
1859
1860 int set_pages_x(struct page *page, int numpages)
1861 {
1862         unsigned long addr = (unsigned long)page_address(page);
1863
1864         return set_memory_x(addr, numpages);
1865 }
1866 EXPORT_SYMBOL(set_pages_x);
1867
1868 int set_pages_nx(struct page *page, int numpages)
1869 {
1870         unsigned long addr = (unsigned long)page_address(page);
1871
1872         return set_memory_nx(addr, numpages);
1873 }
1874 EXPORT_SYMBOL(set_pages_nx);
1875
1876 int set_pages_ro(struct page *page, int numpages)
1877 {
1878         unsigned long addr = (unsigned long)page_address(page);
1879
1880         return set_memory_ro(addr, numpages);
1881 }
1882
1883 int set_pages_rw(struct page *page, int numpages)
1884 {
1885         unsigned long addr = (unsigned long)page_address(page);
1886
1887         return set_memory_rw(addr, numpages);
1888 }
1889
1890 #ifdef CONFIG_DEBUG_PAGEALLOC
1891
1892 static int __set_pages_p(struct page *page, int numpages)
1893 {
1894         unsigned long tempaddr = (unsigned long) page_address(page);
1895         struct cpa_data cpa = { .vaddr = &tempaddr,
1896                                 .pgd = NULL,
1897                                 .numpages = numpages,
1898                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1899                                 .mask_clr = __pgprot(0),
1900                                 .flags = 0};
1901
1902         /*
1903          * No alias checking needed for setting present flag. otherwise,
1904          * we may need to break large pages for 64-bit kernel text
1905          * mappings (this adds to complexity if we want to do this from
1906          * atomic context especially). Let's keep it simple!
1907          */
1908         return __change_page_attr_set_clr(&cpa, 0);
1909 }
1910
1911 static int __set_pages_np(struct page *page, int numpages)
1912 {
1913         unsigned long tempaddr = (unsigned long) page_address(page);
1914         struct cpa_data cpa = { .vaddr = &tempaddr,
1915                                 .pgd = NULL,
1916                                 .numpages = numpages,
1917                                 .mask_set = __pgprot(0),
1918                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1919                                 .flags = 0};
1920
1921         /*
1922          * No alias checking needed for setting not present flag. otherwise,
1923          * we may need to break large pages for 64-bit kernel text
1924          * mappings (this adds to complexity if we want to do this from
1925          * atomic context especially). Let's keep it simple!
1926          */
1927         return __change_page_attr_set_clr(&cpa, 0);
1928 }
1929
1930 void __kernel_map_pages(struct page *page, int numpages, int enable)
1931 {
1932         if (PageHighMem(page))
1933                 return;
1934         if (!enable) {
1935                 debug_check_no_locks_freed(page_address(page),
1936                                            numpages * PAGE_SIZE);
1937         }
1938
1939         /*
1940          * The return value is ignored as the calls cannot fail.
1941          * Large pages for identity mappings are not used at boot time
1942          * and hence no memory allocations during large page split.
1943          */
1944         if (enable)
1945                 __set_pages_p(page, numpages);
1946         else
1947                 __set_pages_np(page, numpages);
1948
1949         /*
1950          * We should perform an IPI and flush all tlbs,
1951          * but that can deadlock->flush only current cpu:
1952          */
1953         __flush_tlb_all();
1954
1955         arch_flush_lazy_mmu_mode();
1956 }
1957
1958 #ifdef CONFIG_HIBERNATION
1959
1960 bool kernel_page_present(struct page *page)
1961 {
1962         unsigned int level;
1963         pte_t *pte;
1964
1965         if (PageHighMem(page))
1966                 return false;
1967
1968         pte = lookup_address((unsigned long)page_address(page), &level);
1969         return (pte_val(*pte) & _PAGE_PRESENT);
1970 }
1971
1972 #endif /* CONFIG_HIBERNATION */
1973
1974 #endif /* CONFIG_DEBUG_PAGEALLOC */
1975
1976 int kernel_map_pages_in_pgd(pgd_t *pgd, u64 pfn, unsigned long address,
1977                             unsigned numpages, unsigned long page_flags)
1978 {
1979         int retval = -EINVAL;
1980
1981         struct cpa_data cpa = {
1982                 .vaddr = &address,
1983                 .pfn = pfn,
1984                 .pgd = pgd,
1985                 .numpages = numpages,
1986                 .mask_set = __pgprot(0),
1987                 .mask_clr = __pgprot(0),
1988                 .flags = 0,
1989         };
1990
1991         if (!(__supported_pte_mask & _PAGE_NX))
1992                 goto out;
1993
1994         if (!(page_flags & _PAGE_NX))
1995                 cpa.mask_clr = __pgprot(_PAGE_NX);
1996
1997         cpa.mask_set = __pgprot(_PAGE_PRESENT | page_flags);
1998
1999         retval = __change_page_attr_set_clr(&cpa, 0);
2000         __flush_tlb_all();
2001
2002 out:
2003         return retval;
2004 }
2005
2006 void kernel_unmap_pages_in_pgd(pgd_t *root, unsigned long address,
2007                                unsigned numpages)
2008 {
2009         unmap_pgd_range(root, address, address + (numpages << PAGE_SHIFT));
2010 }
2011
2012 /*
2013  * The testcases use internal knowledge of the implementation that shouldn't
2014  * be exposed to the rest of the kernel. Include these directly here.
2015  */
2016 #ifdef CONFIG_CPA_DEBUG
2017 #include "pageattr-test.c"
2018 #endif