arch/x86/mm/pageattr.c

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