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