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         if (direct_pages_count[level] == 0)
  71                 return;
  72
  73         direct_pages_count[level]--;
  74         direct_pages_count[level - 1] += PTRS_PER_PTE;
  75 }
  76
  77 void arch_report_meminfo(struct seq_file *m)
  78 {
  79         seq_printf(m, "DirectMap4k:    %8lu kB\n",
  80                         direct_pages_count[PG_LEVEL_4K] << 2);
  81 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
  82         seq_printf(m, "DirectMap2M:    %8lu kB\n",
  83                         direct_pages_count[PG_LEVEL_2M] << 11);
  84 #else
  85         seq_printf(m, "DirectMap4M:    %8lu kB\n",
  86                         direct_pages_count[PG_LEVEL_2M] << 12);
  87 #endif
  88         if (direct_gbpages)
  89                 seq_printf(m, "DirectMap1G:    %8lu kB\n",
  90                         direct_pages_count[PG_LEVEL_1G] << 20);
  91 }
  92 #else
  93 static inline void split_page_count(int level) { }
  94 #endif
  95
  96 #ifdef CONFIG_X86_64
  97
  98 static inline unsigned long highmap_start_pfn(void)
  99 {
 100         return __pa_symbol(_text) >> PAGE_SHIFT;
 101 }
 102
 103 static inline unsigned long highmap_end_pfn(void)
 104 {
 105         /* Do not reference physical address outside the kernel. */
 106         return __pa_symbol(roundup(_brk_end, PMD_SIZE) - 1) >> PAGE_SHIFT;
 107 }
 108
 109 #endif
 110
 111 static inline int
 112 within(unsigned long addr, unsigned long start, unsigned long end)
 113 {
 114         return addr >= start && addr < end;
 115 }
 116
 117 static inline int
 118 within_inclusive(unsigned long addr, unsigned long start, unsigned long end)
 119 {
 120         return addr >= start && addr <= end;
 121 }
 122
 123 /*
 124  * Flushing functions
 125  */
 126
 127 /**
 128  * clflush_cache_range - flush a cache range with clflush
 129  * @vaddr:      virtual start address
 130  * @size:       number of bytes to flush
 131  *
 132  * clflushopt is an unordered instruction which needs fencing with mfence or
 133  * sfence to avoid ordering issues.
 134  */
 135 void clflush_cache_range(void *vaddr, unsigned int size)
 136 {
 137         const unsigned long clflush_size = boot_cpu_data.x86_clflush_size;
 138         void *p = (void *)((unsigned long)vaddr & ~(clflush_size - 1));
 139         void *vend = vaddr + size;
 140
 141         if (p >= vend)
 142                 return;
 143
 144         mb();
 145
 146         for (; p < vend; p += clflush_size)
 147                 clflushopt(p);
 148
 149         mb();
 150 }
 151 EXPORT_SYMBOL_GPL(clflush_cache_range);
 152
 153 static void __cpa_flush_all(void *arg)
 154 {
 155         unsigned long cache = (unsigned long)arg;
 156
 157         /*
 158          * Flush all to work around Errata in early athlons regarding
 159          * large page flushing.
 160          */
 161         __flush_tlb_all();
 162
 163         if (cache && boot_cpu_data.x86 >= 4)
 164                 wbinvd();
 165 }
 166
 167 static void cpa_flush_all(unsigned long cache)
 168 {
 169         BUG_ON(irqs_disabled());
 170
 171         on_each_cpu(__cpa_flush_all, (void *) cache, 1);
 172 }
 173
 174 static void __cpa_flush_range(void *arg)
 175 {
 176         /*
 177          * We could optimize that further and do individual per page
 178          * tlb invalidates for a low number of pages. Caveat: we must
 179          * flush the high aliases on 64bit as well.
 180          */
 181         __flush_tlb_all();
 182 }
 183
 184 static void cpa_flush_range(unsigned long start, int numpages, int cache)
 185 {
 186         unsigned int i, level;
 187         unsigned long addr;
 188
 189         BUG_ON(irqs_disabled());
 190         WARN_ON(PAGE_ALIGN(start) != start);
 191
 192         on_each_cpu(__cpa_flush_range, NULL, 1);
 193
 194         if (!cache)
 195                 return;
 196
 197         /*
 198          * We only need to flush on one CPU,
 199          * clflush is a MESI-coherent instruction that
 200          * will cause all other CPUs to flush the same
 201          * cachelines:
 202          */
 203         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
 204                 pte_t *pte = lookup_address(addr, &level);
 205
 206                 /*
 207                  * Only flush present addresses:
 208                  */
 209                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 210                         clflush_cache_range((void *) addr, PAGE_SIZE);
 211         }
 212 }
 213
 214 static void cpa_flush_array(unsigned long *start, int numpages, int cache,
 215                             int in_flags, struct page **pages)
 216 {
 217         unsigned int i, level;
 218         unsigned long do_wbinvd = cache && numpages >= 1024; /* 4M threshold */
 219
 220         BUG_ON(irqs_disabled());
 221
 222         on_each_cpu(__cpa_flush_all, (void *) do_wbinvd, 1);
 223
 224         if (!cache || do_wbinvd)
 225                 return;
 226
 227         /*
 228          * We only need to flush on one CPU,
 229          * clflush is a MESI-coherent instruction that
 230          * will cause all other CPUs to flush the same
 231          * cachelines:
 232          */
 233         for (i = 0; i < numpages; i++) {
 234                 unsigned long addr;
 235                 pte_t *pte;
 236
 237                 if (in_flags & CPA_PAGES_ARRAY)
 238                         addr = (unsigned long)page_address(pages[i]);
 239                 else
 240                         addr = start[i];
 241
 242                 pte = lookup_address(addr, &level);
 243
 244                 /*
 245                  * Only flush present addresses:
 246                  */
 247                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 248                         clflush_cache_range((void *)addr, PAGE_SIZE);
 249         }
 250 }
 251
 252 /*
 253  * Certain areas of memory on x86 require very specific protection flags,
 254  * for example the BIOS area or kernel text. Callers don't always get this
 255  * right (again, ioremap() on BIOS memory is not uncommon) so this function
 256  * checks and fixes these known static required protection bits.
 257  */
 258 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
 259                                    unsigned long pfn)
 260 {
 261         pgprot_t forbidden = __pgprot(0);
 262
 263         /*
 264          * The BIOS area between 640k and 1Mb needs to be executable for
 265          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
 266          */
 267 #ifdef CONFIG_PCI_BIOS
 268         if (pcibios_enabled && within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
 269                 pgprot_val(forbidden) |= _PAGE_NX;
 270 #endif
 271
 272         /*
 273          * The kernel text needs to be executable for obvious reasons
 274          * Does not cover __inittext since that is gone later on. On
 275          * 64bit we do not enforce !NX on the low mapping
 276          */
 277         if (within(address, (unsigned long)_text, (unsigned long)_etext))
 278                 pgprot_val(forbidden) |= _PAGE_NX;
 279
 280         /*
 281          * The .rodata section needs to be read-only. Using the pfn
 282          * catches all aliases.
 283          */
 284         if (within(pfn, __pa_symbol(__start_rodata) >> PAGE_SHIFT,
 285                    __pa_symbol(__end_rodata) >> PAGE_SHIFT))
 286                 pgprot_val(forbidden) |= _PAGE_RW;
 287
 288 #if defined(CONFIG_X86_64)
 289         /*
 290          * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
 291          * kernel text mappings for the large page aligned text, rodata sections
 292          * will be always read-only. For the kernel identity mappings covering
 293          * the holes caused by this alignment can be anything that user asks.
 294          *
 295          * This will preserve the large page mappings for kernel text/data
 296          * at no extra cost.
 297          */
 298         if (kernel_set_to_readonly &&
 299             within(address, (unsigned long)_text,
 300                    (unsigned long)__end_rodata_hpage_align)) {
 301                 unsigned int level;
 302
 303                 /*
 304                  * Don't enforce the !RW mapping for the kernel text mapping,
 305                  * if the current mapping is already using small page mapping.
 306                  * No need to work hard to preserve large page mappings in this
 307                  * case.
 308                  *
 309                  * This also fixes the Linux Xen paravirt guest boot failure
 310                  * (because of unexpected read-only mappings for kernel identity
 311                  * mappings). In this paravirt guest case, the kernel text
 312                  * mapping and the kernel identity mapping share the same
 313                  * page-table pages. Thus we can't really use different
 314                  * protections for the kernel text and identity mappings. Also,
 315                  * these shared mappings are made of small page mappings.
 316                  * Thus this don't enforce !RW mapping for small page kernel
 317                  * text mapping logic will help Linux Xen parvirt guest boot
 318                  * as well.
 319                  */
 320                 if (lookup_address(address, &level) && (level != PG_LEVEL_4K))
 321                         pgprot_val(forbidden) |= _PAGE_RW;
 322         }
 323 #endif
 324
 325         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
 326
 327         return prot;
 328 }
 329
 330 /*
 331  * Lookup the page table entry for a virtual address in a specific pgd.
 332  * Return a pointer to the entry and the level of the mapping.
 333  */
 334 pte_t *lookup_address_in_pgd(pgd_t *pgd, unsigned long address,
 335                              unsigned int *level)
 336 {
 337         pud_t *pud;
 338         pmd_t *pmd;
 339
 340         *level = PG_LEVEL_NONE;
 341
 342         if (pgd_none(*pgd))
 343                 return NULL;
 344
 345         pud = pud_offset(pgd, address);
 346         if (pud_none(*pud))
 347                 return NULL;
 348
 349         *level = PG_LEVEL_1G;
 350         if (pud_large(*pud) || !pud_present(*pud))
 351                 return (pte_t *)pud;
 352
 353         pmd = pmd_offset(pud, address);
 354         if (pmd_none(*pmd))
 355                 return NULL;
 356
 357         *level = PG_LEVEL_2M;
 358         if (pmd_large(*pmd) || !pmd_present(*pmd))
 359                 return (pte_t *)pmd;
 360
 361         *level = PG_LEVEL_4K;
 362
 363         return pte_offset_kernel(pmd, address);
 364 }
 365
 366 /*
 367  * Lookup the page table entry for a virtual address. Return a pointer
 368  * to the entry and the level of the mapping.
 369  *
 370  * Note: We return pud and pmd either when the entry is marked large
 371  * or when the present bit is not set. Otherwise we would return a
 372  * pointer to a nonexisting mapping.
 373  */
 374 pte_t *lookup_address(unsigned long address, unsigned int *level)
 375 {
 376         return lookup_address_in_pgd(pgd_offset_k(address), address, level);
 377 }
 378 EXPORT_SYMBOL_GPL(lookup_address);
 379
 380 static pte_t *_lookup_address_cpa(struct cpa_data *cpa, unsigned long address,
 381                                   unsigned int *level)
 382 {
 383         if (cpa->pgd)
 384                 return lookup_address_in_pgd(cpa->pgd + pgd_index(address),
 385                                                address, level);
 386
 387         return lookup_address(address, level);
 388 }
 389
 390 /*
 391  * Lookup the PMD entry for a virtual address. Return a pointer to the entry
 392  * or NULL if not present.
 393  */
 394 pmd_t *lookup_pmd_address(unsigned long address)
 395 {
 396         pgd_t *pgd;
 397         pud_t *pud;
 398
 399         pgd = pgd_offset_k(address);
 400         if (pgd_none(*pgd))
 401                 return NULL;
 402
 403         pud = pud_offset(pgd, address);
 404         if (pud_none(*pud) || pud_large(*pud) || !pud_present(*pud))
 405                 return NULL;
 406
 407         return pmd_offset(pud, address);
 408 }
 409
 410 /*
 411  * This is necessary because __pa() does not work on some
 412  * kinds of memory, like vmalloc() or the alloc_remap()
 413  * areas on 32-bit NUMA systems.  The percpu areas can
 414  * end up in this kind of memory, for instance.
 415  *
 416  * This could be optimized, but it is only intended to be
 417  * used at inititalization time, and keeping it
 418  * unoptimized should increase the testing coverage for
 419  * the more obscure platforms.
 420  */
 421 phys_addr_t slow_virt_to_phys(void *__virt_addr)
 422 {
 423         unsigned long virt_addr = (unsigned long)__virt_addr;
 424         phys_addr_t phys_addr;
 425         unsigned long offset;
 426         enum pg_level level;
 427         pte_t *pte;
 428
 429         pte = lookup_address(virt_addr, &level);
 430         BUG_ON(!pte);
 431
 432         /*
 433          * pXX_pfn() returns unsigned long, which must be cast to phys_addr_t
 434          * before being left-shifted PAGE_SHIFT bits -- this trick is to
 435          * make 32-PAE kernel work correctly.
 436          */
 437         switch (level) {
 438         case PG_LEVEL_1G:
 439                 phys_addr = (phys_addr_t)pud_pfn(*(pud_t *)pte) << PAGE_SHIFT;
 440                 offset = virt_addr & ~PUD_PAGE_MASK;
 441                 break;
 442         case PG_LEVEL_2M:
 443                 phys_addr = (phys_addr_t)pmd_pfn(*(pmd_t *)pte) << PAGE_SHIFT;
 444                 offset = virt_addr & ~PMD_PAGE_MASK;
 445                 break;
 446         default:
 447                 phys_addr = (phys_addr_t)pte_pfn(*pte) << PAGE_SHIFT;
 448                 offset = virt_addr & ~PAGE_MASK;
 449         }
 450
 451         return (phys_addr_t)(phys_addr | offset);
 452 }
 453 EXPORT_SYMBOL_GPL(slow_virt_to_phys);
 454
 455 /*
 456  * Set the new pmd in all the pgds we know about:
 457  */
 458 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 459 {
 460         /* change init_mm */
 461         set_pte_atomic(kpte, pte);
 462 #ifdef CONFIG_X86_32
 463         if (!SHARED_KERNEL_PMD) {
 464                 struct page *page;
 465
 466                 list_for_each_entry(page, &pgd_list, lru) {
 467                         pgd_t *pgd;
 468                         pud_t *pud;
 469                         pmd_t *pmd;
 470
 471                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 472                         pud = pud_offset(pgd, address);
 473                         pmd = pmd_offset(pud, address);
 474                         set_pte_atomic((pte_t *)pmd, pte);
 475                 }
 476         }
 477 #endif
 478 }
 479
 480 static int
 481 try_preserve_large_page(pte_t *kpte, unsigned long address,
 482                         struct cpa_data *cpa)
 483 {
 484         unsigned long nextpage_addr, numpages, pmask, psize, addr, pfn, old_pfn;
 485         pte_t new_pte, old_pte, *tmp;
 486         pgprot_t old_prot, new_prot, req_prot;
 487         int i, do_split = 1;
 488         enum pg_level level;
 489
 490         if (cpa->force_split)
 491                 return 1;
 492
 493         spin_lock(&pgd_lock);
 494         /*
 495          * Check for races, another CPU might have split this page
 496          * up already:
 497          */
 498         tmp = _lookup_address_cpa(cpa, address, &level);
 499         if (tmp != kpte)
 500                 goto out_unlock;
 501
 502         switch (level) {
 503         case PG_LEVEL_2M:
 504                 old_prot = pmd_pgprot(*(pmd_t *)kpte);
 505                 old_pfn = pmd_pfn(*(pmd_t *)kpte);
 506                 break;
 507         case PG_LEVEL_1G:
 508                 old_prot = pud_pgprot(*(pud_t *)kpte);
 509                 old_pfn = pud_pfn(*(pud_t *)kpte);
 510                 break;
 511         default:
 512                 do_split = -EINVAL;
 513                 goto out_unlock;
 514         }
 515
 516         psize = page_level_size(level);
 517         pmask = page_level_mask(level);
 518
 519         /*
 520          * Calculate the number of pages, which fit into this large
 521          * page starting at address:
 522          */
 523         nextpage_addr = (address + psize) & pmask;
 524         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
 525         if (numpages < cpa->numpages)
 526                 cpa->numpages = numpages;
 527
 528         /*
 529          * We are safe now. Check whether the new pgprot is the same:
 530          * Convert protection attributes to 4k-format, as cpa->mask* are set
 531          * up accordingly.
 532          */
 533         old_pte = *kpte;
 534         req_prot = pgprot_large_2_4k(old_prot);
 535
 536         pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
 537         pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
 538
 539         /*
 540          * req_prot is in format of 4k pages. It must be converted to large
 541          * page format: the caching mode includes the PAT bit located at
 542          * different bit positions in the two formats.
 543          */
 544         req_prot = pgprot_4k_2_large(req_prot);
 545
 546         /*
 547          * Set the PSE and GLOBAL flags only if the PRESENT flag is
 548          * set otherwise pmd_present/pmd_huge will return true even on
 549          * a non present pmd. The canon_pgprot will clear _PAGE_GLOBAL
 550          * for the ancient hardware that doesn't support it.
 551          */
 552         if (pgprot_val(req_prot) & _PAGE_PRESENT)
 553                 pgprot_val(req_prot) |= _PAGE_PSE | _PAGE_GLOBAL;
 554         else
 555                 pgprot_val(req_prot) &= ~(_PAGE_PSE | _PAGE_GLOBAL);
 556
 557         req_prot = canon_pgprot(req_prot);
 558
 559         /*
 560          * old_pfn points to the large page base pfn. So we need
 561          * to add the offset of the virtual address:
 562          */
 563         pfn = old_pfn + ((address & (psize - 1)) >> PAGE_SHIFT);
 564         cpa->pfn = pfn;
 565
 566         new_prot = static_protections(req_prot, address, pfn);
 567
 568         /*
 569          * We need to check the full range, whether
 570          * static_protection() requires a different pgprot for one of
 571          * the pages in the range we try to preserve:
 572          */
 573         addr = address & pmask;
 574         pfn = old_pfn;
 575         for (i = 0; i < (psize >> PAGE_SHIFT); i++, addr += PAGE_SIZE, pfn++) {
 576                 pgprot_t chk_prot = static_protections(req_prot, addr, pfn);
 577
 578                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
 579                         goto out_unlock;
 580         }
 581
 582         /*
 583          * If there are no changes, return. maxpages has been updated
 584          * above:
 585          */
 586         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
 587                 do_split = 0;
 588                 goto out_unlock;
 589         }
 590
 591         /*
 592          * We need to change the attributes. Check, whether we can
 593          * change the large page in one go. We request a split, when
 594          * the address is not aligned and the number of pages is
 595          * smaller than the number of pages in the large page. Note
 596          * that we limited the number of possible pages already to
 597          * the number of pages in the large page.
 598          */
 599         if (address == (address & pmask) && cpa->numpages == (psize >> PAGE_SHIFT)) {
 600                 /*
 601                  * The address is aligned and the number of pages
 602                  * covers the full page.
 603                  */
 604                 new_pte = pfn_pte(old_pfn, new_prot);
 605                 __set_pmd_pte(kpte, address, new_pte);
 606                 cpa->flags |= CPA_FLUSHTLB;
 607                 do_split = 0;
 608         }
 609
 610 out_unlock:
 611         spin_unlock(&pgd_lock);
 612
 613         return do_split;
 614 }
 615
 616 static int
 617 __split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
 618                    struct page *base)
 619 {
 620         pte_t *pbase = (pte_t *)page_address(base);
 621         unsigned long ref_pfn, pfn, pfninc = 1;
 622         unsigned int i, level;
 623         pte_t *tmp;
 624         pgprot_t ref_prot;
 625
 626         spin_lock(&pgd_lock);
 627         /*
 628          * Check for races, another CPU might have split this page
 629          * up for us already:
 630          */
 631         tmp = _lookup_address_cpa(cpa, address, &level);
 632         if (tmp != kpte) {
 633                 spin_unlock(&pgd_lock);
 634                 return 1;
 635         }
 636
 637         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
 638
 639         switch (level) {
 640         case PG_LEVEL_2M:
 641                 ref_prot = pmd_pgprot(*(pmd_t *)kpte);
 642                 /* clear PSE and promote PAT bit to correct position */
 643                 ref_prot = pgprot_large_2_4k(ref_prot);
 644                 ref_pfn = pmd_pfn(*(pmd_t *)kpte);
 645                 break;
 646
 647         case PG_LEVEL_1G:
 648                 ref_prot = pud_pgprot(*(pud_t *)kpte);
 649                 ref_pfn = pud_pfn(*(pud_t *)kpte);
 650                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
 651
 652                 /*
 653                  * Clear the PSE flags if the PRESENT flag is not set
 654                  * otherwise pmd_present/pmd_huge will return true
 655                  * even on a non present pmd.
 656                  */
 657                 if (!(pgprot_val(ref_prot) & _PAGE_PRESENT))
 658                         pgprot_val(ref_prot) &= ~_PAGE_PSE;
 659                 break;
 660
 661         default:
 662                 spin_unlock(&pgd_lock);
 663                 return 1;
 664         }
 665
 666         /*
 667          * Set the GLOBAL flags only if the PRESENT flag is set
 668          * otherwise pmd/pte_present will return true even on a non
 669          * present pmd/pte. The canon_pgprot will clear _PAGE_GLOBAL
 670          * for the ancient hardware that doesn't support it.
 671          */
 672         if (pgprot_val(ref_prot) & _PAGE_PRESENT)
 673                 pgprot_val(ref_prot) |= _PAGE_GLOBAL;
 674         else
 675                 pgprot_val(ref_prot) &= ~_PAGE_GLOBAL;
 676
 677         /*
 678          * Get the target pfn from the original entry:
 679          */
 680         pfn = ref_pfn;
 681         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
 682                 set_pte(&pbase[i], pfn_pte(pfn, canon_pgprot(ref_prot)));
 683
 684         if (virt_addr_valid(address)) {
 685                 unsigned long pfn = PFN_DOWN(__pa(address));
 686
 687                 if (pfn_range_is_mapped(pfn, pfn + 1))
 688                         split_page_count(level);
 689         }
 690
 691         /*
 692          * Install the new, split up pagetable.
 693          *
 694          * We use the standard kernel pagetable protections for the new
 695          * pagetable protections, the actual ptes set above control the
 696          * primary protection behavior:
 697          */
 698         __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
 699
 700         /*
 701          * Intel Atom errata AAH41 workaround.
 702          *
 703          * The real fix should be in hw or in a microcode update, but
 704          * we also probabilistically try to reduce the window of having
 705          * a large TLB mixed with 4K TLBs while instruction fetches are
 706          * going on.
 707          */
 708         __flush_tlb_all();
 709         spin_unlock(&pgd_lock);
 710
 711         return 0;
 712 }
 713
 714 static int split_large_page(struct cpa_data *cpa, pte_t *kpte,
 715                             unsigned long address)
 716 {
 717         struct page *base;
 718
 719         if (!debug_pagealloc_enabled())
 720                 spin_unlock(&cpa_lock);
 721         base = alloc_pages(GFP_KERNEL | __GFP_NOTRACK, 0);
 722         if (!debug_pagealloc_enabled())
 723                 spin_lock(&cpa_lock);
 724         if (!base)
 725                 return -ENOMEM;
 726
 727         if (__split_large_page(cpa, kpte, address, base))
 728                 __free_page(base);
 729
 730         return 0;
 731 }
 732
 733 static bool try_to_free_pte_page(struct cpa_data *cpa, pte_t *pte)
 734 {
 735         int i;
 736
 737         if (!(cpa->flags & CPA_FREE_PAGETABLES))
 738                 return false;
 739
 740         for (i = 0; i < PTRS_PER_PTE; i++)
 741                 if (!pte_none(pte[i]))
 742                         return false;
 743
 744         free_page((unsigned long)pte);
 745         return true;
 746 }
 747
 748 static bool try_to_free_pmd_page(struct cpa_data *cpa, pmd_t *pmd)
 749 {
 750         int i;
 751
 752         if (!(cpa->flags & CPA_FREE_PAGETABLES))
 753                 return false;
 754
 755         for (i = 0; i < PTRS_PER_PMD; i++)
 756                 if (!pmd_none(pmd[i]))
 757                         return false;
 758
 759         free_page((unsigned long)pmd);
 760         return true;
 761 }
 762
 763 static bool unmap_pte_range(struct cpa_data *cpa, pmd_t *pmd,
 764                             unsigned long start,
 765                             unsigned long end)
 766 {
 767         pte_t *pte = pte_offset_kernel(pmd, start);
 768
 769         while (start < end) {
 770                 set_pte(pte, __pte(0));
 771
 772                 start += PAGE_SIZE;
 773                 pte++;
 774         }
 775
 776         if (try_to_free_pte_page(cpa, (pte_t *)pmd_page_vaddr(*pmd))) {
 777                 pmd_clear(pmd);
 778                 return true;
 779         }
 780         return false;
 781 }
 782
 783 static void __unmap_pmd_range(struct cpa_data *cpa, pud_t *pud, pmd_t *pmd,
 784                               unsigned long start, unsigned long end)
 785 {
 786         if (unmap_pte_range(cpa, pmd, start, end))
 787                 if (try_to_free_pmd_page(cpa, (pmd_t *)pud_page_vaddr(*pud)))
 788                         pud_clear(pud);
 789 }
 790
 791 static void unmap_pmd_range(struct cpa_data *cpa, pud_t *pud,
 792                             unsigned long start, unsigned long end)
 793 {
 794         pmd_t *pmd = pmd_offset(pud, start);
 795
 796         /*
 797          * Not on a 2MB page boundary?
 798          */
 799         if (start & (PMD_SIZE - 1)) {
 800                 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
 801                 unsigned long pre_end = min_t(unsigned long, end, next_page);
 802
 803                 __unmap_pmd_range(cpa, pud, pmd, start, pre_end);
 804
 805                 start = pre_end;
 806                 pmd++;
 807         }
 808
 809         /*
 810          * Try to unmap in 2M chunks.
 811          */
 812         while (end - start >= PMD_SIZE) {
 813                 if (pmd_large(*pmd))
 814                         pmd_clear(pmd);
 815                 else
 816                         __unmap_pmd_range(cpa, pud, pmd,
 817                                           start, start + PMD_SIZE);
 818
 819                 start += PMD_SIZE;
 820                 pmd++;
 821         }
 822
 823         /*
 824          * 4K leftovers?
 825          */
 826         if (start < end)
 827                 return __unmap_pmd_range(cpa, pud, pmd, start, end);
 828
 829         /*
 830          * Try again to free the PMD page if haven't succeeded above.
 831          */
 832         if (!pud_none(*pud))
 833                 if (try_to_free_pmd_page(cpa, (pmd_t *)pud_page_vaddr(*pud)))
 834                         pud_clear(pud);
 835 }
 836
 837 static void __unmap_pud_range(struct cpa_data *cpa, pgd_t *pgd,
 838                               unsigned long start,
 839                               unsigned long end)
 840 {
 841         pud_t *pud = pud_offset(pgd, start);
 842
 843         /*
 844          * Not on a GB page boundary?
 845          */
 846         if (start & (PUD_SIZE - 1)) {
 847                 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
 848                 unsigned long pre_end   = min_t(unsigned long, end, next_page);
 849
 850                 unmap_pmd_range(cpa, pud, start, pre_end);
 851
 852                 start = pre_end;
 853                 pud++;
 854         }
 855
 856         /*
 857          * Try to unmap in 1G chunks?
 858          */
 859         while (end - start >= PUD_SIZE) {
 860
 861                 if (pud_large(*pud))
 862                         pud_clear(pud);
 863                 else
 864                         unmap_pmd_range(cpa, pud, start, start + PUD_SIZE);
 865
 866                 start += PUD_SIZE;
 867                 pud++;
 868         }
 869
 870         /*
 871          * 2M leftovers?
 872          */
 873         if (start < end)
 874                 unmap_pmd_range(cpa, pud, start, end);
 875
 876         /*
 877          * No need to try to free the PUD page because we'll free it in
 878          * populate_pgd's error path
 879          */
 880 }
 881
 882 static void unmap_pud_range(pgd_t *pgd, unsigned long start, unsigned long end)
 883 {
 884         struct cpa_data cpa = {
 885                 .flags = CPA_FREE_PAGETABLES,
 886         };
 887
 888         __unmap_pud_range(&cpa, pgd, start, end);
 889 }
 890
 891 void unmap_pud_range_nofree(pgd_t *pgd, unsigned long start, unsigned long end)
 892 {
 893         struct cpa_data cpa = {
 894                 .flags = 0,
 895         };
 896
 897         __unmap_pud_range(&cpa, pgd, start, end);
 898 }
 899
 900 static int alloc_pte_page(pmd_t *pmd)
 901 {
 902         pte_t *pte = (pte_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
 903         if (!pte)
 904                 return -1;
 905
 906         set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
 907         return 0;
 908 }
 909
 910 static int alloc_pmd_page(pud_t *pud)
 911 {
 912         pmd_t *pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
 913         if (!pmd)
 914                 return -1;
 915
 916         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
 917         return 0;
 918 }
 919
 920 static void populate_pte(struct cpa_data *cpa,
 921                          unsigned long start, unsigned long end,
 922                          unsigned num_pages, pmd_t *pmd, pgprot_t pgprot)
 923 {
 924         pte_t *pte;
 925
 926         pte = pte_offset_kernel(pmd, start);
 927
 928         /*
 929          * Set the GLOBAL flags only if the PRESENT flag is
 930          * set otherwise pte_present will return true even on
 931          * a non present pte. The canon_pgprot will clear
 932          * _PAGE_GLOBAL for the ancient hardware that doesn't
 933          * support it.
 934          */
 935         if (pgprot_val(pgprot) & _PAGE_PRESENT)
 936                 pgprot_val(pgprot) |= _PAGE_GLOBAL;
 937         else
 938                 pgprot_val(pgprot) &= ~_PAGE_GLOBAL;
 939
 940         pgprot = canon_pgprot(pgprot);
 941
 942         while (num_pages-- && start < end) {
 943                 set_pte(pte, pfn_pte(cpa->pfn, pgprot));
 944
 945                 start    += PAGE_SIZE;
 946                 cpa->pfn++;
 947                 pte++;
 948         }
 949 }
 950
 951 static long populate_pmd(struct cpa_data *cpa,
 952                          unsigned long start, unsigned long end,
 953                          unsigned num_pages, pud_t *pud, pgprot_t pgprot)
 954 {
 955         long cur_pages = 0;
 956         pmd_t *pmd;
 957         pgprot_t pmd_pgprot;
 958
 959         /*
 960          * Not on a 2M boundary?
 961          */
 962         if (start & (PMD_SIZE - 1)) {
 963                 unsigned long pre_end = start + (num_pages << PAGE_SHIFT);
 964                 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
 965
 966                 pre_end   = min_t(unsigned long, pre_end, next_page);
 967                 cur_pages = (pre_end - start) >> PAGE_SHIFT;
 968                 cur_pages = min_t(unsigned int, num_pages, cur_pages);
 969
 970                 /*
 971                  * Need a PTE page?
 972                  */
 973                 pmd = pmd_offset(pud, start);
 974                 if (pmd_none(*pmd))
 975                         if (alloc_pte_page(pmd))
 976                                 return -1;
 977
 978                 populate_pte(cpa, start, pre_end, cur_pages, pmd, pgprot);
 979
 980                 start = pre_end;
 981         }
 982
 983         /*
 984          * We mapped them all?
 985          */
 986         if (num_pages == cur_pages)
 987                 return cur_pages;
 988
 989         pmd_pgprot = pgprot_4k_2_large(pgprot);
 990
 991         while (end - start >= PMD_SIZE) {
 992
 993                 /*
 994                  * We cannot use a 1G page so allocate a PMD page if needed.
 995                  */
 996                 if (pud_none(*pud))
 997                         if (alloc_pmd_page(pud))
 998                                 return -1;
 999
1000                 pmd = pmd_offset(pud, start);
1001
1002                 set_pmd(pmd, __pmd(cpa->pfn << PAGE_SHIFT | _PAGE_PSE |
1003                                    massage_pgprot(pmd_pgprot)));
1004
1005                 start     += PMD_SIZE;
1006                 cpa->pfn  += PMD_SIZE >> PAGE_SHIFT;
1007                 cur_pages += PMD_SIZE >> PAGE_SHIFT;
1008         }
1009
1010         /*
1011          * Map trailing 4K pages.
1012          */
1013         if (start < end) {
1014                 pmd = pmd_offset(pud, start);
1015                 if (pmd_none(*pmd))
1016                         if (alloc_pte_page(pmd))
1017                                 return -1;
1018
1019                 populate_pte(cpa, start, end, num_pages - cur_pages,
1020                              pmd, pgprot);
1021         }
1022         return num_pages;
1023 }
1024
1025 static long populate_pud(struct cpa_data *cpa, unsigned long start, pgd_t *pgd,
1026                          pgprot_t pgprot)
1027 {
1028         pud_t *pud;
1029         unsigned long end;
1030         long cur_pages = 0;
1031         pgprot_t pud_pgprot;
1032
1033         end = start + (cpa->numpages << PAGE_SHIFT);
1034
1035         /*
1036          * Not on a Gb page boundary? => map everything up to it with
1037          * smaller pages.
1038          */
1039         if (start & (PUD_SIZE - 1)) {
1040                 unsigned long pre_end;
1041                 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1042
1043                 pre_end   = min_t(unsigned long, end, next_page);
1044                 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1045                 cur_pages = min_t(int, (int)cpa->numpages, cur_pages);
1046
1047                 pud = pud_offset(pgd, start);
1048
1049                 /*
1050                  * Need a PMD page?
1051                  */
1052                 if (pud_none(*pud))
1053                         if (alloc_pmd_page(pud))
1054                                 return -1;
1055
1056                 cur_pages = populate_pmd(cpa, start, pre_end, cur_pages,
1057                                          pud, pgprot);
1058                 if (cur_pages < 0)
1059                         return cur_pages;
1060
1061                 start = pre_end;
1062         }
1063
1064         /* We mapped them all? */
1065         if (cpa->numpages == cur_pages)
1066                 return cur_pages;
1067
1068         pud = pud_offset(pgd, start);
1069         pud_pgprot = pgprot_4k_2_large(pgprot);
1070
1071         /*
1072          * Map everything starting from the Gb boundary, possibly with 1G pages
1073          */
1074         while (boot_cpu_has(X86_FEATURE_GBPAGES) && end - start >= PUD_SIZE) {
1075                 set_pud(pud, __pud(cpa->pfn << PAGE_SHIFT | _PAGE_PSE |
1076                                    massage_pgprot(pud_pgprot)));
1077
1078                 start     += PUD_SIZE;
1079                 cpa->pfn  += PUD_SIZE >> PAGE_SHIFT;
1080                 cur_pages += PUD_SIZE >> PAGE_SHIFT;
1081                 pud++;
1082         }
1083
1084         /* Map trailing leftover */
1085         if (start < end) {
1086                 long tmp;
1087
1088                 pud = pud_offset(pgd, start);
1089                 if (pud_none(*pud))
1090                         if (alloc_pmd_page(pud))
1091                                 return -1;
1092
1093                 tmp = populate_pmd(cpa, start, end, cpa->numpages - cur_pages,
1094                                    pud, pgprot);
1095                 if (tmp < 0)
1096                         return cur_pages;
1097
1098                 cur_pages += tmp;
1099         }
1100         return cur_pages;
1101 }
1102
1103 /*
1104  * Restrictions for kernel page table do not necessarily apply when mapping in
1105  * an alternate PGD.
1106  */
1107 static int populate_pgd(struct cpa_data *cpa, unsigned long addr)
1108 {
1109         pgprot_t pgprot = __pgprot(_KERNPG_TABLE);
1110         pud_t *pud = NULL;      /* shut up gcc */
1111         pgd_t *pgd_entry;
1112         long ret;
1113
1114         pgd_entry = cpa->pgd + pgd_index(addr);
1115
1116         /*
1117          * Allocate a PUD page and hand it down for mapping.
1118          */
1119         if (pgd_none(*pgd_entry)) {
1120                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
1121                 if (!pud)
1122                         return -1;
1123
1124                 set_pgd(pgd_entry, __pgd(__pa(pud) | _KERNPG_TABLE));
1125         }
1126
1127         pgprot_val(pgprot) &= ~pgprot_val(cpa->mask_clr);
1128         pgprot_val(pgprot) |=  pgprot_val(cpa->mask_set);
1129
1130         ret = populate_pud(cpa, addr, pgd_entry, pgprot);
1131         if (ret < 0) {
1132                 /*
1133                  * Leave the PUD page in place in case some other CPU or thread
1134                  * already found it, but remove any useless entries we just
1135                  * added to it.
1136                  */
1137                 unmap_pud_range(pgd_entry, addr,
1138                                 addr + (cpa->numpages << PAGE_SHIFT));
1139                 return ret;
1140         }
1141
1142         cpa->numpages = ret;
1143         return 0;
1144 }
1145
1146 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
1147                                int primary)
1148 {
1149         if (cpa->pgd) {
1150                 /*
1151                  * Right now, we only execute this code path when mapping
1152                  * the EFI virtual memory map regions, no other users
1153                  * provide a ->pgd value. This may change in the future.
1154                  */
1155                 return populate_pgd(cpa, vaddr);
1156         }
1157
1158         /*
1159          * Ignore all non primary paths.
1160          */
1161         if (!primary) {
1162                 cpa->numpages = 1;
1163                 return 0;
1164         }
1165
1166         /*
1167          * Ignore the NULL PTE for kernel identity mapping, as it is expected
1168          * to have holes.
1169          * Also set numpages to '1' indicating that we processed cpa req for
1170          * one virtual address page and its pfn. TBD: numpages can be set based
1171          * on the initial value and the level returned by lookup_address().
1172          */
1173         if (within(vaddr, PAGE_OFFSET,
1174                    PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
1175                 cpa->numpages = 1;
1176                 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
1177                 return 0;
1178         } else {
1179                 WARN(1, KERN_WARNING "CPA: called for zero pte. "
1180                         "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
1181                         *cpa->vaddr);
1182
1183                 return -EFAULT;
1184         }
1185 }
1186
1187 static int __change_page_attr(struct cpa_data *cpa, int primary)
1188 {
1189         unsigned long address;
1190         int do_split, err;
1191         unsigned int level;
1192         pte_t *kpte, old_pte;
1193
1194         if (cpa->flags & CPA_PAGES_ARRAY) {
1195                 struct page *page = cpa->pages[cpa->curpage];
1196                 if (unlikely(PageHighMem(page)))
1197                         return 0;
1198                 address = (unsigned long)page_address(page);
1199         } else if (cpa->flags & CPA_ARRAY)
1200                 address = cpa->vaddr[cpa->curpage];
1201         else
1202                 address = *cpa->vaddr;
1203 repeat:
1204         kpte = _lookup_address_cpa(cpa, address, &level);
1205         if (!kpte)
1206                 return __cpa_process_fault(cpa, address, primary);
1207
1208         old_pte = *kpte;
1209         if (pte_none(old_pte))
1210                 return __cpa_process_fault(cpa, address, primary);
1211
1212         if (level == PG_LEVEL_4K) {
1213                 pte_t new_pte;
1214                 pgprot_t new_prot = pte_pgprot(old_pte);
1215                 unsigned long pfn = pte_pfn(old_pte);
1216
1217                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
1218                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
1219
1220                 new_prot = static_protections(new_prot, address, pfn);
1221
1222                 /*
1223                  * Set the GLOBAL flags only if the PRESENT flag is
1224                  * set otherwise pte_present will return true even on
1225                  * a non present pte. The canon_pgprot will clear
1226                  * _PAGE_GLOBAL for the ancient hardware that doesn't
1227                  * support it.
1228                  */
1229                 if (pgprot_val(new_prot) & _PAGE_PRESENT)
1230                         pgprot_val(new_prot) |= _PAGE_GLOBAL;
1231                 else
1232                         pgprot_val(new_prot) &= ~_PAGE_GLOBAL;
1233
1234                 /*
1235                  * We need to keep the pfn from the existing PTE,
1236                  * after all we're only going to change it's attributes
1237                  * not the memory it points to
1238                  */
1239                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
1240                 cpa->pfn = pfn;
1241                 /*
1242                  * Do we really change anything ?
1243                  */
1244                 if (pte_val(old_pte) != pte_val(new_pte)) {
1245                         set_pte_atomic(kpte, new_pte);
1246                         cpa->flags |= CPA_FLUSHTLB;
1247                 }
1248                 cpa->numpages = 1;
1249                 return 0;
1250         }
1251
1252         /*
1253          * Check, whether we can keep the large page intact
1254          * and just change the pte:
1255          */
1256         do_split = try_preserve_large_page(kpte, address, cpa);
1257         /*
1258          * When the range fits into the existing large page,
1259          * return. cp->numpages and cpa->tlbflush have been updated in
1260          * try_large_page:
1261          */
1262         if (do_split <= 0)
1263                 return do_split;
1264
1265         /*
1266          * We have to split the large page:
1267          */
1268         err = split_large_page(cpa, kpte, address);
1269         if (!err) {
1270                 /*
1271                  * Do a global flush tlb after splitting the large page
1272                  * and before we do the actual change page attribute in the PTE.
1273                  *
1274                  * With out this, we violate the TLB application note, that says
1275                  * "The TLBs may contain both ordinary and large-page
1276                  *  translations for a 4-KByte range of linear addresses. This
1277                  *  may occur if software modifies the paging structures so that
1278                  *  the page size used for the address range changes. If the two
1279                  *  translations differ with respect to page frame or attributes
1280                  *  (e.g., permissions), processor behavior is undefined and may
1281                  *  be implementation-specific."
1282                  *
1283                  * We do this global tlb flush inside the cpa_lock, so that we
1284                  * don't allow any other cpu, with stale tlb entries change the
1285                  * page attribute in parallel, that also falls into the
1286                  * just split large page entry.
1287                  */
1288                 flush_tlb_all();
1289                 goto repeat;
1290         }
1291
1292         return err;
1293 }
1294
1295 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
1296
1297 static int cpa_process_alias(struct cpa_data *cpa)
1298 {
1299         struct cpa_data alias_cpa;
1300         unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
1301         unsigned long vaddr;
1302         int ret;
1303
1304         if (!pfn_range_is_mapped(cpa->pfn, cpa->pfn + 1))
1305                 return 0;
1306
1307         /*
1308          * No need to redo, when the primary call touched the direct
1309          * mapping already:
1310          */
1311         if (cpa->flags & CPA_PAGES_ARRAY) {
1312                 struct page *page = cpa->pages[cpa->curpage];
1313                 if (unlikely(PageHighMem(page)))
1314                         return 0;
1315                 vaddr = (unsigned long)page_address(page);
1316         } else if (cpa->flags & CPA_ARRAY)
1317                 vaddr = cpa->vaddr[cpa->curpage];
1318         else
1319                 vaddr = *cpa->vaddr;
1320
1321         if (!(within(vaddr, PAGE_OFFSET,
1322                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
1323
1324                 alias_cpa = *cpa;
1325                 alias_cpa.vaddr = &laddr;
1326                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1327
1328                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
1329                 if (ret)
1330                         return ret;
1331         }
1332
1333 #ifdef CONFIG_X86_64
1334         /*
1335          * If the primary call didn't touch the high mapping already
1336          * and the physical address is inside the kernel map, we need
1337          * to touch the high mapped kernel as well:
1338          */
1339         if (!within(vaddr, (unsigned long)_text, _brk_end) &&
1340             within_inclusive(cpa->pfn, highmap_start_pfn(),
1341                              highmap_end_pfn())) {
1342                 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
1343                                                __START_KERNEL_map - phys_base;
1344                 alias_cpa = *cpa;
1345                 alias_cpa.vaddr = &temp_cpa_vaddr;
1346                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1347
1348                 /*
1349                  * The high mapping range is imprecise, so ignore the
1350                  * return value.
1351                  */
1352                 __change_page_attr_set_clr(&alias_cpa, 0);
1353         }
1354 #endif
1355
1356         return 0;
1357 }
1358
1359 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
1360 {
1361         unsigned long numpages = cpa->numpages;
1362         int ret;
1363
1364         while (numpages) {
1365                 /*
1366                  * Store the remaining nr of pages for the large page
1367                  * preservation check.
1368                  */
1369                 cpa->numpages = numpages;
1370                 /* for array changes, we can't use large page */
1371                 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
1372                         cpa->numpages = 1;
1373
1374                 if (!debug_pagealloc_enabled())
1375                         spin_lock(&cpa_lock);
1376                 ret = __change_page_attr(cpa, checkalias);
1377                 if (!debug_pagealloc_enabled())
1378                         spin_unlock(&cpa_lock);
1379                 if (ret)
1380                         return ret;
1381
1382                 if (checkalias) {
1383                         ret = cpa_process_alias(cpa);
1384                         if (ret)
1385                                 return ret;
1386                 }
1387
1388                 /*
1389                  * Adjust the number of pages with the result of the
1390                  * CPA operation. Either a large page has been
1391                  * preserved or a single page update happened.
1392                  */
1393                 BUG_ON(cpa->numpages > numpages || !cpa->numpages);
1394                 numpages -= cpa->numpages;
1395                 if (cpa->flags & (CPA_PAGES_ARRAY | CPA_ARRAY))
1396                         cpa->curpage++;
1397                 else
1398                         *cpa->vaddr += cpa->numpages * PAGE_SIZE;
1399
1400         }
1401         return 0;
1402 }
1403
1404 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
1405                                     pgprot_t mask_set, pgprot_t mask_clr,
1406                                     int force_split, int in_flag,
1407                                     struct page **pages)
1408 {
1409         struct cpa_data cpa;
1410         int ret, cache, checkalias;
1411         unsigned long baddr = 0;
1412
1413         memset(&cpa, 0, sizeof(cpa));
1414
1415         /*
1416          * Check, if we are requested to change a not supported
1417          * feature:
1418          */
1419         mask_set = canon_pgprot(mask_set);
1420         mask_clr = canon_pgprot(mask_clr);
1421         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
1422                 return 0;
1423
1424         /* Ensure we are PAGE_SIZE aligned */
1425         if (in_flag & CPA_ARRAY) {
1426                 int i;
1427                 for (i = 0; i < numpages; i++) {
1428                         if (addr[i] & ~PAGE_MASK) {
1429                                 addr[i] &= PAGE_MASK;
1430                                 WARN_ON_ONCE(1);
1431                         }
1432                 }
1433         } else if (!(in_flag & CPA_PAGES_ARRAY)) {
1434                 /*
1435                  * in_flag of CPA_PAGES_ARRAY implies it is aligned.
1436                  * No need to cehck in that case
1437                  */
1438                 if (*addr & ~PAGE_MASK) {
1439                         *addr &= PAGE_MASK;
1440                         /*
1441                          * People should not be passing in unaligned addresses:
1442                          */
1443                         WARN_ON_ONCE(1);
1444                 }
1445                 /*
1446                  * Save address for cache flush. *addr is modified in the call
1447                  * to __change_page_attr_set_clr() below.
1448                  */
1449                 baddr = *addr;
1450         }
1451
1452         /* Must avoid aliasing mappings in the highmem code */
1453         kmap_flush_unused();
1454
1455         vm_unmap_aliases();
1456
1457         cpa.vaddr = addr;
1458         cpa.pages = pages;
1459         cpa.numpages = numpages;
1460         cpa.mask_set = mask_set;
1461         cpa.mask_clr = mask_clr;
1462         cpa.flags = 0;
1463         cpa.curpage = 0;
1464         cpa.force_split = force_split;
1465
1466         if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
1467                 cpa.flags |= in_flag;
1468
1469         /* No alias checking for _NX bit modifications */
1470         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
1471
1472         ret = __change_page_attr_set_clr(&cpa, checkalias);
1473
1474         /*
1475          * Check whether we really changed something:
1476          */
1477         if (!(cpa.flags & CPA_FLUSHTLB))
1478                 goto out;
1479
1480         /*
1481          * No need to flush, when we did not set any of the caching
1482          * attributes:
1483          */
1484         cache = !!pgprot2cachemode(mask_set);
1485
1486         /*
1487          * On success we use CLFLUSH, when the CPU supports it to
1488          * avoid the WBINVD. If the CPU does not support it and in the
1489          * error case we fall back to cpa_flush_all (which uses
1490          * WBINVD):
1491          */
1492         if (!ret && boot_cpu_has(X86_FEATURE_CLFLUSH)) {
1493                 if (cpa.flags & (CPA_PAGES_ARRAY | CPA_ARRAY)) {
1494                         cpa_flush_array(addr, numpages, cache,
1495                                         cpa.flags, pages);
1496                 } else
1497                         cpa_flush_range(baddr, numpages, cache);
1498         } else
1499                 cpa_flush_all(cache);
1500
1501 out:
1502         return ret;
1503 }
1504
1505 static inline int change_page_attr_set(unsigned long *addr, int numpages,
1506                                        pgprot_t mask, int array)
1507 {
1508         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
1509                 (array ? CPA_ARRAY : 0), NULL);
1510 }
1511
1512 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
1513                                          pgprot_t mask, int array)
1514 {
1515         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
1516                 (array ? CPA_ARRAY : 0), NULL);
1517 }
1518
1519 static inline int cpa_set_pages_array(struct page **pages, int numpages,
1520                                        pgprot_t mask)
1521 {
1522         return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
1523                 CPA_PAGES_ARRAY, pages);
1524 }
1525
1526 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1527                                          pgprot_t mask)
1528 {
1529         return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1530                 CPA_PAGES_ARRAY, pages);
1531 }
1532
1533 int _set_memory_uc(unsigned long addr, int numpages)
1534 {
1535         /*
1536          * for now UC MINUS. see comments in ioremap_nocache()
1537          * If you really need strong UC use ioremap_uc(), but note
1538          * that you cannot override IO areas with set_memory_*() as
1539          * these helpers cannot work with IO memory.
1540          */
1541         return change_page_attr_set(&addr, numpages,
1542                                     cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1543                                     0);
1544 }
1545
1546 int set_memory_uc(unsigned long addr, int numpages)
1547 {
1548         int ret;
1549
1550         /*
1551          * for now UC MINUS. see comments in ioremap_nocache()
1552          */
1553         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1554                               _PAGE_CACHE_MODE_UC_MINUS, NULL);
1555         if (ret)
1556                 goto out_err;
1557
1558         ret = _set_memory_uc(addr, numpages);
1559         if (ret)
1560                 goto out_free;
1561
1562         return 0;
1563
1564 out_free:
1565         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1566 out_err:
1567         return ret;
1568 }
1569 EXPORT_SYMBOL(set_memory_uc);
1570
1571 static int _set_memory_array(unsigned long *addr, int addrinarray,
1572                 enum page_cache_mode new_type)
1573 {
1574         enum page_cache_mode set_type;
1575         int i, j;
1576         int ret;
1577
1578         for (i = 0; i < addrinarray; i++) {
1579                 ret = reserve_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE,
1580                                         new_type, NULL);
1581                 if (ret)
1582                         goto out_free;
1583         }
1584
1585         /* If WC, set to UC- first and then WC */
1586         set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
1587                                 _PAGE_CACHE_MODE_UC_MINUS : new_type;
1588
1589         ret = change_page_attr_set(addr, addrinarray,
1590                                    cachemode2pgprot(set_type), 1);
1591
1592         if (!ret && new_type == _PAGE_CACHE_MODE_WC)
1593                 ret = change_page_attr_set_clr(addr, addrinarray,
1594                                                cachemode2pgprot(
1595                                                 _PAGE_CACHE_MODE_WC),
1596                                                __pgprot(_PAGE_CACHE_MASK),
1597                                                0, CPA_ARRAY, NULL);
1598         if (ret)
1599                 goto out_free;
1600
1601         return 0;
1602
1603 out_free:
1604         for (j = 0; j < i; j++)
1605                 free_memtype(__pa(addr[j]), __pa(addr[j]) + PAGE_SIZE);
1606
1607         return ret;
1608 }
1609
1610 int set_memory_array_uc(unsigned long *addr, int addrinarray)
1611 {
1612         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_UC_MINUS);
1613 }
1614 EXPORT_SYMBOL(set_memory_array_uc);
1615
1616 int set_memory_array_wc(unsigned long *addr, int addrinarray)
1617 {
1618         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_WC);
1619 }
1620 EXPORT_SYMBOL(set_memory_array_wc);
1621
1622 int set_memory_array_wt(unsigned long *addr, int addrinarray)
1623 {
1624         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_MODE_WT);
1625 }
1626 EXPORT_SYMBOL_GPL(set_memory_array_wt);
1627
1628 int _set_memory_wc(unsigned long addr, int numpages)
1629 {
1630         int ret;
1631         unsigned long addr_copy = addr;
1632
1633         ret = change_page_attr_set(&addr, numpages,
1634                                    cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1635                                    0);
1636         if (!ret) {
1637                 ret = change_page_attr_set_clr(&addr_copy, numpages,
1638                                                cachemode2pgprot(
1639                                                 _PAGE_CACHE_MODE_WC),
1640                                                __pgprot(_PAGE_CACHE_MASK),
1641                                                0, 0, NULL);
1642         }
1643         return ret;
1644 }
1645
1646 int set_memory_wc(unsigned long addr, int numpages)
1647 {
1648         int ret;
1649
1650         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1651                 _PAGE_CACHE_MODE_WC, NULL);
1652         if (ret)
1653                 return ret;
1654
1655         ret = _set_memory_wc(addr, numpages);
1656         if (ret)
1657                 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1658
1659         return ret;
1660 }
1661 EXPORT_SYMBOL(set_memory_wc);
1662
1663 int _set_memory_wt(unsigned long addr, int numpages)
1664 {
1665         return change_page_attr_set(&addr, numpages,
1666                                     cachemode2pgprot(_PAGE_CACHE_MODE_WT), 0);
1667 }
1668
1669 int set_memory_wt(unsigned long addr, int numpages)
1670 {
1671         int ret;
1672
1673         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1674                               _PAGE_CACHE_MODE_WT, NULL);
1675         if (ret)
1676                 return ret;
1677
1678         ret = _set_memory_wt(addr, numpages);
1679         if (ret)
1680                 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1681
1682         return ret;
1683 }
1684 EXPORT_SYMBOL_GPL(set_memory_wt);
1685
1686 int _set_memory_wb(unsigned long addr, int numpages)
1687 {
1688         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1689         return change_page_attr_clear(&addr, numpages,
1690                                       __pgprot(_PAGE_CACHE_MASK), 0);
1691 }
1692
1693 int set_memory_wb(unsigned long addr, int numpages)
1694 {
1695         int ret;
1696
1697         ret = _set_memory_wb(addr, numpages);
1698         if (ret)
1699                 return ret;
1700
1701         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1702         return 0;
1703 }
1704 EXPORT_SYMBOL(set_memory_wb);
1705
1706 int set_memory_array_wb(unsigned long *addr, int addrinarray)
1707 {
1708         int i;
1709         int ret;
1710
1711         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1712         ret = change_page_attr_clear(addr, addrinarray,
1713                                       __pgprot(_PAGE_CACHE_MASK), 1);
1714         if (ret)
1715                 return ret;
1716
1717         for (i = 0; i < addrinarray; i++)
1718                 free_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE);
1719
1720         return 0;
1721 }
1722 EXPORT_SYMBOL(set_memory_array_wb);
1723
1724 int set_memory_x(unsigned long addr, int numpages)
1725 {
1726         if (!(__supported_pte_mask & _PAGE_NX))
1727                 return 0;
1728
1729         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1730 }
1731 EXPORT_SYMBOL(set_memory_x);
1732
1733 int set_memory_nx(unsigned long addr, int numpages)
1734 {
1735         if (!(__supported_pte_mask & _PAGE_NX))
1736                 return 0;
1737
1738         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1739 }
1740 EXPORT_SYMBOL(set_memory_nx);
1741
1742 int set_memory_ro(unsigned long addr, int numpages)
1743 {
1744         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1745 }
1746
1747 int set_memory_rw(unsigned long addr, int numpages)
1748 {
1749         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1750 }
1751
1752 int set_memory_np(unsigned long addr, int numpages)
1753 {
1754         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1755 }
1756
1757 int set_memory_4k(unsigned long addr, int numpages)
1758 {
1759         return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1760                                         __pgprot(0), 1, 0, NULL);
1761 }
1762
1763 int set_pages_uc(struct page *page, int numpages)
1764 {
1765         unsigned long addr = (unsigned long)page_address(page);
1766
1767         return set_memory_uc(addr, numpages);
1768 }
1769 EXPORT_SYMBOL(set_pages_uc);
1770
1771 static int _set_pages_array(struct page **pages, int addrinarray,
1772                 enum page_cache_mode new_type)
1773 {
1774         unsigned long start;
1775         unsigned long end;
1776         enum page_cache_mode set_type;
1777         int i;
1778         int free_idx;
1779         int ret;
1780
1781         for (i = 0; i < addrinarray; i++) {
1782                 if (PageHighMem(pages[i]))
1783                         continue;
1784                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1785                 end = start + PAGE_SIZE;
1786                 if (reserve_memtype(start, end, new_type, NULL))
1787                         goto err_out;
1788         }
1789
1790         /* If WC, set to UC- first and then WC */
1791         set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
1792                                 _PAGE_CACHE_MODE_UC_MINUS : new_type;
1793
1794         ret = cpa_set_pages_array(pages, addrinarray,
1795                                   cachemode2pgprot(set_type));
1796         if (!ret && new_type == _PAGE_CACHE_MODE_WC)
1797                 ret = change_page_attr_set_clr(NULL, addrinarray,
1798                                                cachemode2pgprot(
1799                                                 _PAGE_CACHE_MODE_WC),
1800                                                __pgprot(_PAGE_CACHE_MASK),
1801                                                0, CPA_PAGES_ARRAY, pages);
1802         if (ret)
1803                 goto err_out;
1804         return 0; /* Success */
1805 err_out:
1806         free_idx = i;
1807         for (i = 0; i < free_idx; i++) {
1808                 if (PageHighMem(pages[i]))
1809                         continue;
1810                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1811                 end = start + PAGE_SIZE;
1812                 free_memtype(start, end);
1813         }
1814         return -EINVAL;
1815 }
1816
1817 int set_pages_array_uc(struct page **pages, int addrinarray)
1818 {
1819         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_UC_MINUS);
1820 }
1821 EXPORT_SYMBOL(set_pages_array_uc);
1822
1823 int set_pages_array_wc(struct page **pages, int addrinarray)
1824 {
1825         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_WC);
1826 }
1827 EXPORT_SYMBOL(set_pages_array_wc);
1828
1829 int set_pages_array_wt(struct page **pages, int addrinarray)
1830 {
1831         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_MODE_WT);
1832 }
1833 EXPORT_SYMBOL_GPL(set_pages_array_wt);
1834
1835 int set_pages_wb(struct page *page, int numpages)
1836 {
1837         unsigned long addr = (unsigned long)page_address(page);
1838
1839         return set_memory_wb(addr, numpages);
1840 }
1841 EXPORT_SYMBOL(set_pages_wb);
1842
1843 int set_pages_array_wb(struct page **pages, int addrinarray)
1844 {
1845         int retval;
1846         unsigned long start;
1847         unsigned long end;
1848         int i;
1849
1850         /* WB cache mode is hard wired to all cache attribute bits being 0 */
1851         retval = cpa_clear_pages_array(pages, addrinarray,
1852                         __pgprot(_PAGE_CACHE_MASK));
1853         if (retval)
1854                 return retval;
1855
1856         for (i = 0; i < addrinarray; i++) {
1857                 if (PageHighMem(pages[i]))
1858                         continue;
1859                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1860                 end = start + PAGE_SIZE;
1861                 free_memtype(start, end);
1862         }
1863
1864         return 0;
1865 }
1866 EXPORT_SYMBOL(set_pages_array_wb);
1867
1868 int set_pages_x(struct page *page, int numpages)
1869 {
1870         unsigned long addr = (unsigned long)page_address(page);
1871
1872         return set_memory_x(addr, numpages);
1873 }
1874 EXPORT_SYMBOL(set_pages_x);
1875
1876 int set_pages_nx(struct page *page, int numpages)
1877 {
1878         unsigned long addr = (unsigned long)page_address(page);
1879
1880         return set_memory_nx(addr, numpages);
1881 }
1882 EXPORT_SYMBOL(set_pages_nx);
1883
1884 int set_pages_ro(struct page *page, int numpages)
1885 {
1886         unsigned long addr = (unsigned long)page_address(page);
1887
1888         return set_memory_ro(addr, numpages);
1889 }
1890
1891 int set_pages_rw(struct page *page, int numpages)
1892 {
1893         unsigned long addr = (unsigned long)page_address(page);
1894
1895         return set_memory_rw(addr, numpages);
1896 }
1897
1898 #ifdef CONFIG_DEBUG_PAGEALLOC
1899
1900 static int __set_pages_p(struct page *page, int numpages)
1901 {
1902         unsigned long tempaddr = (unsigned long) page_address(page);
1903         struct cpa_data cpa = { .vaddr = &tempaddr,
1904                                 .pgd = NULL,
1905                                 .numpages = numpages,
1906                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1907                                 .mask_clr = __pgprot(0),
1908                                 .flags = 0};
1909
1910         /*
1911          * No alias checking needed for setting present flag. otherwise,
1912          * we may need to break large pages for 64-bit kernel text
1913          * mappings (this adds to complexity if we want to do this from
1914          * atomic context especially). Let's keep it simple!
1915          */
1916         return __change_page_attr_set_clr(&cpa, 0);
1917 }
1918
1919 static int __set_pages_np(struct page *page, int numpages)
1920 {
1921         unsigned long tempaddr = (unsigned long) page_address(page);
1922         struct cpa_data cpa = { .vaddr = &tempaddr,
1923                                 .pgd = NULL,
1924                                 .numpages = numpages,
1925                                 .mask_set = __pgprot(0),
1926                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1927                                 .flags = 0};
1928
1929         /*
1930          * No alias checking needed for setting not present flag. otherwise,
1931          * we may need to break large pages for 64-bit kernel text
1932          * mappings (this adds to complexity if we want to do this from
1933          * atomic context especially). Let's keep it simple!
1934          */
1935         return __change_page_attr_set_clr(&cpa, 0);
1936 }
1937
1938 void __kernel_map_pages(struct page *page, int numpages, int enable)
1939 {
1940         if (PageHighMem(page))
1941                 return;
1942         if (!enable) {
1943                 debug_check_no_locks_freed(page_address(page),
1944                                            numpages * PAGE_SIZE);
1945         }
1946
1947         /*
1948          * The return value is ignored as the calls cannot fail.
1949          * Large pages for identity mappings are not used at boot time
1950          * and hence no memory allocations during large page split.
1951          */
1952         if (enable)
1953                 __set_pages_p(page, numpages);
1954         else
1955                 __set_pages_np(page, numpages);
1956
1957         /*
1958          * We should perform an IPI and flush all tlbs,
1959          * but that can deadlock->flush only current cpu:
1960          */
1961         __flush_tlb_all();
1962
1963         arch_flush_lazy_mmu_mode();
1964 }
1965
1966 #ifdef CONFIG_HIBERNATION
1967
1968 bool kernel_page_present(struct page *page)
1969 {
1970         unsigned int level;
1971         pte_t *pte;
1972
1973         if (PageHighMem(page))
1974                 return false;
1975
1976         pte = lookup_address((unsigned long)page_address(page), &level);
1977         return (pte_val(*pte) & _PAGE_PRESENT);
1978 }
1979
1980 #endif /* CONFIG_HIBERNATION */
1981
1982 #endif /* CONFIG_DEBUG_PAGEALLOC */
1983
1984 int kernel_map_pages_in_pgd(pgd_t *pgd, u64 pfn, unsigned long address,
1985                             unsigned numpages, unsigned long page_flags)
1986 {
1987         int retval = -EINVAL;
1988
1989         struct cpa_data cpa = {
1990                 .vaddr = &address,
1991                 .pfn = pfn,
1992                 .pgd = pgd,
1993                 .numpages = numpages,
1994                 .mask_set = __pgprot(0),
1995                 .mask_clr = __pgprot(0),
1996                 .flags = 0,
1997         };
1998
1999         if (!(__supported_pte_mask & _PAGE_NX))
2000                 goto out;
2001
2002         if (!(page_flags & _PAGE_NX))
2003                 cpa.mask_clr = __pgprot(_PAGE_NX);
2004
2005         if (!(page_flags & _PAGE_RW))
2006                 cpa.mask_clr = __pgprot(_PAGE_RW);
2007
2008         cpa.mask_set = __pgprot(_PAGE_PRESENT | page_flags);
2009
2010         retval = __change_page_attr_set_clr(&cpa, 0);
2011         __flush_tlb_all();
2012
2013 out:
2014         return retval;
2015 }
2016
2017 /*
2018  * The testcases use internal knowledge of the implementation that shouldn't
2019  * be exposed to the rest of the kernel. Include these directly here.
2020  */
2021 #ifdef CONFIG_CPA_DEBUG
2022 #include "pageattr-test.c"
2023 #endif