kernel/arch/x86/mm/pageattr.c

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