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/slab.h>
  10 #include <linux/mm.h>
  11 #include <linux/interrupt.h>
  12 #include <linux/seq_file.h>
  13 #include <linux/debugfs.h>
  14
  15 #include <asm/e820.h>
  16 #include <asm/processor.h>
  17 #include <asm/tlbflush.h>
  18 #include <asm/sections.h>
  19 #include <asm/uaccess.h>
  20 #include <asm/pgalloc.h>
  21 #include <asm/proto.h>
  22 #include <asm/pat.h>
  23
  24 /*
  25  * The current flushing context - we pass it instead of 5 arguments:
  26  */
  27 struct cpa_data {
  28         unsigned long   vaddr;
  29         pgprot_t        mask_set;
  30         pgprot_t        mask_clr;
  31         int             numpages;
  32         int             flushtlb;
  33         unsigned long   pfn;
  34         unsigned        force_split : 1;
  35 };
  36
  37 #ifdef CONFIG_PROC_FS
  38 static unsigned long direct_pages_count[PG_LEVEL_NUM];
  39
  40 void update_page_count(int level, unsigned long pages)
  41 {
  42         unsigned long flags;
  43
  44         /* Protect against CPA */
  45         spin_lock_irqsave(&pgd_lock, flags);
  46         direct_pages_count[level] += pages;
  47         spin_unlock_irqrestore(&pgd_lock, flags);
  48 }
  49
  50 static void split_page_count(int level)
  51 {
  52         direct_pages_count[level]--;
  53         direct_pages_count[level - 1] += PTRS_PER_PTE;
  54 }
  55
  56 int arch_report_meminfo(char *page)
  57 {
  58         int n = sprintf(page, "DirectMap4k:  %8lu kB\n",
  59                         direct_pages_count[PG_LEVEL_4K] << 2);
  60 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
  61         n += sprintf(page + n, "DirectMap2M:  %8lu kB\n",
  62                         direct_pages_count[PG_LEVEL_2M] << 11);
  63 #else
  64         n += sprintf(page + n, "DirectMap4M:  %8lu kB\n",
  65                         direct_pages_count[PG_LEVEL_2M] << 12);
  66 #endif
  67 #ifdef CONFIG_X86_64
  68         if (direct_gbpages)
  69                 n += sprintf(page + n, "DirectMap1G:  %8lu kB\n",
  70                         direct_pages_count[PG_LEVEL_1G] << 20);
  71 #endif
  72         return n;
  73 }
  74 #else
  75 static inline void split_page_count(int level) { }
  76 #endif
  77
  78 #ifdef CONFIG_X86_64
  79
  80 static inline unsigned long highmap_start_pfn(void)
  81 {
  82         return __pa(_text) >> PAGE_SHIFT;
  83 }
  84
  85 static inline unsigned long highmap_end_pfn(void)
  86 {
  87         return __pa(roundup((unsigned long)_end, PMD_SIZE)) >> PAGE_SHIFT;
  88 }
  89
  90 #endif
  91
  92 #ifdef CONFIG_DEBUG_PAGEALLOC
  93 # define debug_pagealloc 1
  94 #else
  95 # define debug_pagealloc 0
  96 #endif
  97
  98 static inline int
  99 within(unsigned long addr, unsigned long start, unsigned long end)
 100 {
 101         return addr >= start && addr < end;
 102 }
 103
 104 /*
 105  * Flushing functions
 106  */
 107
 108 /**
 109  * clflush_cache_range - flush a cache range with clflush
 110  * @addr:       virtual start address
 111  * @size:       number of bytes to flush
 112  *
 113  * clflush is an unordered instruction which needs fencing with mfence
 114  * to avoid ordering issues.
 115  */
 116 void clflush_cache_range(void *vaddr, unsigned int size)
 117 {
 118         void *vend = vaddr + size - 1;
 119
 120         mb();
 121
 122         for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
 123                 clflush(vaddr);
 124         /*
 125          * Flush any possible final partial cacheline:
 126          */
 127         clflush(vend);
 128
 129         mb();
 130 }
 131
 132 static void __cpa_flush_all(void *arg)
 133 {
 134         unsigned long cache = (unsigned long)arg;
 135
 136         /*
 137          * Flush all to work around Errata in early athlons regarding
 138          * large page flushing.
 139          */
 140         __flush_tlb_all();
 141
 142         if (cache && boot_cpu_data.x86_model >= 4)
 143                 wbinvd();
 144 }
 145
 146 static void cpa_flush_all(unsigned long cache)
 147 {
 148         BUG_ON(irqs_disabled());
 149
 150         on_each_cpu(__cpa_flush_all, (void *) cache, 1);
 151 }
 152
 153 static void __cpa_flush_range(void *arg)
 154 {
 155         /*
 156          * We could optimize that further and do individual per page
 157          * tlb invalidates for a low number of pages. Caveat: we must
 158          * flush the high aliases on 64bit as well.
 159          */
 160         __flush_tlb_all();
 161 }
 162
 163 static void cpa_flush_range(unsigned long start, int numpages, int cache)
 164 {
 165         unsigned int i, level;
 166         unsigned long addr;
 167
 168         BUG_ON(irqs_disabled());
 169         WARN_ON(PAGE_ALIGN(start) != start);
 170
 171         on_each_cpu(__cpa_flush_range, NULL, 1);
 172
 173         if (!cache)
 174                 return;
 175
 176         /*
 177          * We only need to flush on one CPU,
 178          * clflush is a MESI-coherent instruction that
 179          * will cause all other CPUs to flush the same
 180          * cachelines:
 181          */
 182         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
 183                 pte_t *pte = lookup_address(addr, &level);
 184
 185                 /*
 186                  * Only flush present addresses:
 187                  */
 188                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 189                         clflush_cache_range((void *) addr, PAGE_SIZE);
 190         }
 191 }
 192
 193 /*
 194  * Certain areas of memory on x86 require very specific protection flags,
 195  * for example the BIOS area or kernel text. Callers don't always get this
 196  * right (again, ioremap() on BIOS memory is not uncommon) so this function
 197  * checks and fixes these known static required protection bits.
 198  */
 199 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
 200                                    unsigned long pfn)
 201 {
 202         pgprot_t forbidden = __pgprot(0);
 203
 204         /*
 205          * The BIOS area between 640k and 1Mb needs to be executable for
 206          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
 207          */
 208         if (within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
 209                 pgprot_val(forbidden) |= _PAGE_NX;
 210
 211         /*
 212          * The kernel text needs to be executable for obvious reasons
 213          * Does not cover __inittext since that is gone later on. On
 214          * 64bit we do not enforce !NX on the low mapping
 215          */
 216         if (within(address, (unsigned long)_text, (unsigned long)_etext))
 217                 pgprot_val(forbidden) |= _PAGE_NX;
 218
 219         /*
 220          * The .rodata section needs to be read-only. Using the pfn
 221          * catches all aliases.
 222          */
 223         if (within(pfn, __pa((unsigned long)__start_rodata) >> PAGE_SHIFT,
 224                    __pa((unsigned long)__end_rodata) >> PAGE_SHIFT))
 225                 pgprot_val(forbidden) |= _PAGE_RW;
 226
 227         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
 228
 229         return prot;
 230 }
 231
 232 /*
 233  * Lookup the page table entry for a virtual address. Return a pointer
 234  * to the entry and the level of the mapping.
 235  *
 236  * Note: We return pud and pmd either when the entry is marked large
 237  * or when the present bit is not set. Otherwise we would return a
 238  * pointer to a nonexisting mapping.
 239  */
 240 pte_t *lookup_address(unsigned long address, unsigned int *level)
 241 {
 242         pgd_t *pgd = pgd_offset_k(address);
 243         pud_t *pud;
 244         pmd_t *pmd;
 245
 246         *level = PG_LEVEL_NONE;
 247
 248         if (pgd_none(*pgd))
 249                 return NULL;
 250
 251         pud = pud_offset(pgd, address);
 252         if (pud_none(*pud))
 253                 return NULL;
 254
 255         *level = PG_LEVEL_1G;
 256         if (pud_large(*pud) || !pud_present(*pud))
 257                 return (pte_t *)pud;
 258
 259         pmd = pmd_offset(pud, address);
 260         if (pmd_none(*pmd))
 261                 return NULL;
 262
 263         *level = PG_LEVEL_2M;
 264         if (pmd_large(*pmd) || !pmd_present(*pmd))
 265                 return (pte_t *)pmd;
 266
 267         *level = PG_LEVEL_4K;
 268
 269         return pte_offset_kernel(pmd, address);
 270 }
 271 EXPORT_SYMBOL_GPL(lookup_address);
 272
 273 /*
 274  * Set the new pmd in all the pgds we know about:
 275  */
 276 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 277 {
 278         /* change init_mm */
 279         set_pte_atomic(kpte, pte);
 280 #ifdef CONFIG_X86_32
 281         if (!SHARED_KERNEL_PMD) {
 282                 struct page *page;
 283
 284                 list_for_each_entry(page, &pgd_list, lru) {
 285                         pgd_t *pgd;
 286                         pud_t *pud;
 287                         pmd_t *pmd;
 288
 289                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 290                         pud = pud_offset(pgd, address);
 291                         pmd = pmd_offset(pud, address);
 292                         set_pte_atomic((pte_t *)pmd, pte);
 293                 }
 294         }
 295 #endif
 296 }
 297
 298 static int
 299 try_preserve_large_page(pte_t *kpte, unsigned long address,
 300                         struct cpa_data *cpa)
 301 {
 302         unsigned long nextpage_addr, numpages, pmask, psize, flags, addr, pfn;
 303         pte_t new_pte, old_pte, *tmp;
 304         pgprot_t old_prot, new_prot;
 305         int i, do_split = 1;
 306         unsigned int level;
 307
 308         if (cpa->force_split)
 309                 return 1;
 310
 311         spin_lock_irqsave(&pgd_lock, flags);
 312         /*
 313          * Check for races, another CPU might have split this page
 314          * up already:
 315          */
 316         tmp = lookup_address(address, &level);
 317         if (tmp != kpte)
 318                 goto out_unlock;
 319
 320         switch (level) {
 321         case PG_LEVEL_2M:
 322                 psize = PMD_PAGE_SIZE;
 323                 pmask = PMD_PAGE_MASK;
 324                 break;
 325 #ifdef CONFIG_X86_64
 326         case PG_LEVEL_1G:
 327                 psize = PUD_PAGE_SIZE;
 328                 pmask = PUD_PAGE_MASK;
 329                 break;
 330 #endif
 331         default:
 332                 do_split = -EINVAL;
 333                 goto out_unlock;
 334         }
 335
 336         /*
 337          * Calculate the number of pages, which fit into this large
 338          * page starting at address:
 339          */
 340         nextpage_addr = (address + psize) & pmask;
 341         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
 342         if (numpages < cpa->numpages)
 343                 cpa->numpages = numpages;
 344
 345         /*
 346          * We are safe now. Check whether the new pgprot is the same:
 347          */
 348         old_pte = *kpte;
 349         old_prot = new_prot = pte_pgprot(old_pte);
 350
 351         pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
 352         pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
 353
 354         /*
 355          * old_pte points to the large page base address. So we need
 356          * to add the offset of the virtual address:
 357          */
 358         pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
 359         cpa->pfn = pfn;
 360
 361         new_prot = static_protections(new_prot, address, pfn);
 362
 363         /*
 364          * We need to check the full range, whether
 365          * static_protection() requires a different pgprot for one of
 366          * the pages in the range we try to preserve:
 367          */
 368         addr = address + PAGE_SIZE;
 369         pfn++;
 370         for (i = 1; i < cpa->numpages; i++, addr += PAGE_SIZE, pfn++) {
 371                 pgprot_t chk_prot = static_protections(new_prot, addr, pfn);
 372
 373                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
 374                         goto out_unlock;
 375         }
 376
 377         /*
 378          * If there are no changes, return. maxpages has been updated
 379          * above:
 380          */
 381         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
 382                 do_split = 0;
 383                 goto out_unlock;
 384         }
 385
 386         /*
 387          * We need to change the attributes. Check, whether we can
 388          * change the large page in one go. We request a split, when
 389          * the address is not aligned and the number of pages is
 390          * smaller than the number of pages in the large page. Note
 391          * that we limited the number of possible pages already to
 392          * the number of pages in the large page.
 393          */
 394         if (address == (nextpage_addr - psize) && cpa->numpages == numpages) {
 395                 /*
 396                  * The address is aligned and the number of pages
 397                  * covers the full page.
 398                  */
 399                 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
 400                 __set_pmd_pte(kpte, address, new_pte);
 401                 cpa->flushtlb = 1;
 402                 do_split = 0;
 403         }
 404
 405 out_unlock:
 406         spin_unlock_irqrestore(&pgd_lock, flags);
 407
 408         return do_split;
 409 }
 410
 411 static LIST_HEAD(page_pool);
 412 static unsigned long pool_size, pool_pages, pool_low;
 413 static unsigned long pool_used, pool_failed;
 414
 415 static void cpa_fill_pool(struct page **ret)
 416 {
 417         gfp_t gfp = GFP_KERNEL;
 418         unsigned long flags;
 419         struct page *p;
 420
 421         /*
 422          * Avoid recursion (on debug-pagealloc) and also signal
 423          * our priority to get to these pagetables:
 424          */
 425         if (current->flags & PF_MEMALLOC)
 426                 return;
 427         current->flags |= PF_MEMALLOC;
 428
 429         /*
 430          * Allocate atomically from atomic contexts:
 431          */
 432         if (in_atomic() || irqs_disabled() || debug_pagealloc)
 433                 gfp =  GFP_ATOMIC | __GFP_NORETRY | __GFP_NOWARN;
 434
 435         while (pool_pages < pool_size || (ret && !*ret)) {
 436                 p = alloc_pages(gfp, 0);
 437                 if (!p) {
 438                         pool_failed++;
 439                         break;
 440                 }
 441                 /*
 442                  * If the call site needs a page right now, provide it:
 443                  */
 444                 if (ret && !*ret) {
 445                         *ret = p;
 446                         continue;
 447                 }
 448                 spin_lock_irqsave(&pgd_lock, flags);
 449                 list_add(&p->lru, &page_pool);
 450                 pool_pages++;
 451                 spin_unlock_irqrestore(&pgd_lock, flags);
 452         }
 453
 454         current->flags &= ~PF_MEMALLOC;
 455 }
 456
 457 #define SHIFT_MB                (20 - PAGE_SHIFT)
 458 #define ROUND_MB_GB             ((1 << 10) - 1)
 459 #define SHIFT_MB_GB             10
 460 #define POOL_PAGES_PER_GB       16
 461
 462 void __init cpa_init(void)
 463 {
 464         struct sysinfo si;
 465         unsigned long gb;
 466
 467         si_meminfo(&si);
 468         /*
 469          * Calculate the number of pool pages:
 470          *
 471          * Convert totalram (nr of pages) to MiB and round to the next
 472          * GiB. Shift MiB to Gib and multiply the result by
 473          * POOL_PAGES_PER_GB:
 474          */
 475         if (debug_pagealloc) {
 476                 gb = ((si.totalram >> SHIFT_MB) + ROUND_MB_GB) >> SHIFT_MB_GB;
 477                 pool_size = POOL_PAGES_PER_GB * gb;
 478         } else {
 479                 pool_size = 1;
 480         }
 481         pool_low = pool_size;
 482
 483         cpa_fill_pool(NULL);
 484         printk(KERN_DEBUG
 485                "CPA: page pool initialized %lu of %lu pages preallocated\n",
 486                pool_pages, pool_size);
 487 }
 488
 489 static int split_large_page(pte_t *kpte, unsigned long address)
 490 {
 491         unsigned long flags, pfn, pfninc = 1;
 492         unsigned int i, level;
 493         pte_t *pbase, *tmp;
 494         pgprot_t ref_prot;
 495         struct page *base;
 496
 497         /*
 498          * Get a page from the pool. The pool list is protected by the
 499          * pgd_lock, which we have to take anyway for the split
 500          * operation:
 501          */
 502         spin_lock_irqsave(&pgd_lock, flags);
 503         if (list_empty(&page_pool)) {
 504                 spin_unlock_irqrestore(&pgd_lock, flags);
 505                 base = NULL;
 506                 cpa_fill_pool(&base);
 507                 if (!base)
 508                         return -ENOMEM;
 509                 spin_lock_irqsave(&pgd_lock, flags);
 510         } else {
 511                 base = list_first_entry(&page_pool, struct page, lru);
 512                 list_del(&base->lru);
 513                 pool_pages--;
 514
 515                 if (pool_pages < pool_low)
 516                         pool_low = pool_pages;
 517         }
 518
 519         /*
 520          * Check for races, another CPU might have split this page
 521          * up for us already:
 522          */
 523         tmp = lookup_address(address, &level);
 524         if (tmp != kpte)
 525                 goto out_unlock;
 526
 527         pbase = (pte_t *)page_address(base);
 528         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
 529         ref_prot = pte_pgprot(pte_clrhuge(*kpte));
 530
 531 #ifdef CONFIG_X86_64
 532         if (level == PG_LEVEL_1G) {
 533                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
 534                 pgprot_val(ref_prot) |= _PAGE_PSE;
 535         }
 536 #endif
 537
 538         /*
 539          * Get the target pfn from the original entry:
 540          */
 541         pfn = pte_pfn(*kpte);
 542         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
 543                 set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
 544
 545         if (address >= (unsigned long)__va(0) &&
 546                 address < (unsigned long)__va(max_low_pfn_mapped << PAGE_SHIFT))
 547                 split_page_count(level);
 548
 549 #ifdef CONFIG_X86_64
 550         if (address >= (unsigned long)__va(1UL<<32) &&
 551                 address < (unsigned long)__va(max_pfn_mapped << PAGE_SHIFT))
 552                 split_page_count(level);
 553 #endif
 554
 555         /*
 556          * Install the new, split up pagetable. Important details here:
 557          *
 558          * On Intel the NX bit of all levels must be cleared to make a
 559          * page executable. See section 4.13.2 of Intel 64 and IA-32
 560          * Architectures Software Developer's Manual).
 561          *
 562          * Mark the entry present. The current mapping might be
 563          * set to not present, which we preserved above.
 564          */
 565         ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
 566         pgprot_val(ref_prot) |= _PAGE_PRESENT;
 567         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
 568         base = NULL;
 569
 570 out_unlock:
 571         /*
 572          * If we dropped out via the lookup_address check under
 573          * pgd_lock then stick the page back into the pool:
 574          */
 575         if (base) {
 576                 list_add(&base->lru, &page_pool);
 577                 pool_pages++;
 578         } else
 579                 pool_used++;
 580         spin_unlock_irqrestore(&pgd_lock, flags);
 581
 582         return 0;
 583 }
 584
 585 static int __change_page_attr(struct cpa_data *cpa, int primary)
 586 {
 587         unsigned long address = cpa->vaddr;
 588         int do_split, err;
 589         unsigned int level;
 590         pte_t *kpte, old_pte;
 591
 592 repeat:
 593         kpte = lookup_address(address, &level);
 594         if (!kpte)
 595                 return 0;
 596
 597         old_pte = *kpte;
 598         if (!pte_val(old_pte)) {
 599                 if (!primary)
 600                         return 0;
 601                 WARN(1, KERN_WARNING "CPA: called for zero pte. "
 602                        "vaddr = %lx cpa->vaddr = %lx\n", address,
 603                        cpa->vaddr);
 604                 return -EINVAL;
 605         }
 606
 607         if (level == PG_LEVEL_4K) {
 608                 pte_t new_pte;
 609                 pgprot_t new_prot = pte_pgprot(old_pte);
 610                 unsigned long pfn = pte_pfn(old_pte);
 611
 612                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
 613                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
 614
 615                 new_prot = static_protections(new_prot, address, pfn);
 616
 617                 /*
 618                  * We need to keep the pfn from the existing PTE,
 619                  * after all we're only going to change it's attributes
 620                  * not the memory it points to
 621                  */
 622                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
 623                 cpa->pfn = pfn;
 624                 /*
 625                  * Do we really change anything ?
 626                  */
 627                 if (pte_val(old_pte) != pte_val(new_pte)) {
 628                         set_pte_atomic(kpte, new_pte);
 629                         cpa->flushtlb = 1;
 630                 }
 631                 cpa->numpages = 1;
 632                 return 0;
 633         }
 634
 635         /*
 636          * Check, whether we can keep the large page intact
 637          * and just change the pte:
 638          */
 639         do_split = try_preserve_large_page(kpte, address, cpa);
 640         /*
 641          * When the range fits into the existing large page,
 642          * return. cp->numpages and cpa->tlbflush have been updated in
 643          * try_large_page:
 644          */
 645         if (do_split <= 0)
 646                 return do_split;
 647
 648         /*
 649          * We have to split the large page:
 650          */
 651         err = split_large_page(kpte, address);
 652         if (!err) {
 653                 cpa->flushtlb = 1;
 654                 goto repeat;
 655         }
 656
 657         return err;
 658 }
 659
 660 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
 661
 662 static int cpa_process_alias(struct cpa_data *cpa)
 663 {
 664         struct cpa_data alias_cpa;
 665         int ret = 0;
 666
 667         if (cpa->pfn >= max_pfn_mapped)
 668                 return 0;
 669
 670 #ifdef CONFIG_X86_64
 671         if (cpa->pfn >= max_low_pfn_mapped && cpa->pfn < (1UL<<(32-PAGE_SHIFT)))
 672                 return 0;
 673 #endif
 674         /*
 675          * No need to redo, when the primary call touched the direct
 676          * mapping already:
 677          */
 678         if (!(within(cpa->vaddr, PAGE_OFFSET,
 679                     PAGE_OFFSET + (max_low_pfn_mapped << PAGE_SHIFT))
 680 #ifdef CONFIG_X86_64
 681                 || within(cpa->vaddr, PAGE_OFFSET + (1UL<<32),
 682                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))
 683 #endif
 684         )) {
 685
 686                 alias_cpa = *cpa;
 687                 alias_cpa.vaddr = (unsigned long) __va(cpa->pfn << PAGE_SHIFT);
 688
 689                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
 690         }
 691
 692 #ifdef CONFIG_X86_64
 693         if (ret)
 694                 return ret;
 695         /*
 696          * No need to redo, when the primary call touched the high
 697          * mapping already:
 698          */
 699         if (within(cpa->vaddr, (unsigned long) _text, (unsigned long) _end))
 700                 return 0;
 701
 702         /*
 703          * If the physical address is inside the kernel map, we need
 704          * to touch the high mapped kernel as well:
 705          */
 706         if (!within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn()))
 707                 return 0;
 708
 709         alias_cpa = *cpa;
 710         alias_cpa.vaddr =
 711                 (cpa->pfn << PAGE_SHIFT) + __START_KERNEL_map - phys_base;
 712
 713         /*
 714          * The high mapping range is imprecise, so ignore the return value.
 715          */
 716         __change_page_attr_set_clr(&alias_cpa, 0);
 717 #endif
 718         return ret;
 719 }
 720
 721 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
 722 {
 723         int ret, numpages = cpa->numpages;
 724
 725         while (numpages) {
 726                 /*
 727                  * Store the remaining nr of pages for the large page
 728                  * preservation check.
 729                  */
 730                 cpa->numpages = numpages;
 731
 732                 ret = __change_page_attr(cpa, checkalias);
 733                 if (ret)
 734                         return ret;
 735
 736                 if (checkalias) {
 737                         ret = cpa_process_alias(cpa);
 738                         if (ret)
 739                                 return ret;
 740                 }
 741
 742                 /*
 743                  * Adjust the number of pages with the result of the
 744                  * CPA operation. Either a large page has been
 745                  * preserved or a single page update happened.
 746                  */
 747                 BUG_ON(cpa->numpages > numpages);
 748                 numpages -= cpa->numpages;
 749                 cpa->vaddr += cpa->numpages * PAGE_SIZE;
 750         }
 751         return 0;
 752 }
 753
 754 static inline int cache_attr(pgprot_t attr)
 755 {
 756         return pgprot_val(attr) &
 757                 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
 758 }
 759
 760 static int change_page_attr_set_clr(unsigned long addr, int numpages,
 761                                     pgprot_t mask_set, pgprot_t mask_clr,
 762                                     int force_split)
 763 {
 764         struct cpa_data cpa;
 765         int ret, cache, checkalias;
 766
 767         /*
 768          * Check, if we are requested to change a not supported
 769          * feature:
 770          */
 771         mask_set = canon_pgprot(mask_set);
 772         mask_clr = canon_pgprot(mask_clr);
 773         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
 774                 return 0;
 775
 776         /* Ensure we are PAGE_SIZE aligned */
 777         if (addr & ~PAGE_MASK) {
 778                 addr &= PAGE_MASK;
 779                 /*
 780                  * People should not be passing in unaligned addresses:
 781                  */
 782                 WARN_ON_ONCE(1);
 783         }
 784
 785         cpa.vaddr = addr;
 786         cpa.numpages = numpages;
 787         cpa.mask_set = mask_set;
 788         cpa.mask_clr = mask_clr;
 789         cpa.flushtlb = 0;
 790         cpa.force_split = force_split;
 791
 792         /* No alias checking for _NX bit modifications */
 793         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
 794
 795         ret = __change_page_attr_set_clr(&cpa, checkalias);
 796
 797         /*
 798          * Check whether we really changed something:
 799          */
 800         if (!cpa.flushtlb)
 801                 goto out;
 802
 803         /*
 804          * No need to flush, when we did not set any of the caching
 805          * attributes:
 806          */
 807         cache = cache_attr(mask_set);
 808
 809         /*
 810          * On success we use clflush, when the CPU supports it to
 811          * avoid the wbindv. If the CPU does not support it and in the
 812          * error case we fall back to cpa_flush_all (which uses
 813          * wbindv):
 814          */
 815         if (!ret && cpu_has_clflush)
 816                 cpa_flush_range(addr, numpages, cache);
 817         else
 818                 cpa_flush_all(cache);
 819
 820 out:
 821         cpa_fill_pool(NULL);
 822
 823         return ret;
 824 }
 825
 826 static inline int change_page_attr_set(unsigned long addr, int numpages,
 827                                        pgprot_t mask)
 828 {
 829         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0);
 830 }
 831
 832 static inline int change_page_attr_clear(unsigned long addr, int numpages,
 833                                          pgprot_t mask)
 834 {
 835         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0);
 836 }
 837
 838 int _set_memory_uc(unsigned long addr, int numpages)
 839 {
 840         /*
 841          * for now UC MINUS. see comments in ioremap_nocache()
 842          */
 843         return change_page_attr_set(addr, numpages,
 844                                     __pgprot(_PAGE_CACHE_UC_MINUS));
 845 }
 846
 847 int set_memory_uc(unsigned long addr, int numpages)
 848 {
 849         /*
 850          * for now UC MINUS. see comments in ioremap_nocache()
 851          */
 852         if (reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
 853                             _PAGE_CACHE_UC_MINUS, NULL))
 854                 return -EINVAL;
 855
 856         return _set_memory_uc(addr, numpages);
 857 }
 858 EXPORT_SYMBOL(set_memory_uc);
 859
 860 int _set_memory_wc(unsigned long addr, int numpages)
 861 {
 862         return change_page_attr_set(addr, numpages,
 863                                     __pgprot(_PAGE_CACHE_WC));
 864 }
 865
 866 int set_memory_wc(unsigned long addr, int numpages)
 867 {
 868         if (!pat_enabled)
 869                 return set_memory_uc(addr, numpages);
 870
 871         if (reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
 872                 _PAGE_CACHE_WC, NULL))
 873                 return -EINVAL;
 874
 875         return _set_memory_wc(addr, numpages);
 876 }
 877 EXPORT_SYMBOL(set_memory_wc);
 878
 879 int _set_memory_wb(unsigned long addr, int numpages)
 880 {
 881         return change_page_attr_clear(addr, numpages,
 882                                       __pgprot(_PAGE_CACHE_MASK));
 883 }
 884
 885 int set_memory_wb(unsigned long addr, int numpages)
 886 {
 887         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
 888
 889         return _set_memory_wb(addr, numpages);
 890 }
 891 EXPORT_SYMBOL(set_memory_wb);
 892
 893 int set_memory_x(unsigned long addr, int numpages)
 894 {
 895         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
 896 }
 897 EXPORT_SYMBOL(set_memory_x);
 898
 899 int set_memory_nx(unsigned long addr, int numpages)
 900 {
 901         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
 902 }
 903 EXPORT_SYMBOL(set_memory_nx);
 904
 905 int set_memory_ro(unsigned long addr, int numpages)
 906 {
 907         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
 908 }
 909 EXPORT_SYMBOL_GPL(set_memory_ro);
 910
 911 int set_memory_rw(unsigned long addr, int numpages)
 912 {
 913         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
 914 }
 915 EXPORT_SYMBOL_GPL(set_memory_rw);
 916
 917 int set_memory_np(unsigned long addr, int numpages)
 918 {
 919         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
 920 }
 921
 922 int set_memory_4k(unsigned long addr, int numpages)
 923 {
 924         return change_page_attr_set_clr(addr, numpages, __pgprot(0),
 925                                         __pgprot(0), 1);
 926 }
 927
 928 int set_pages_uc(struct page *page, int numpages)
 929 {
 930         unsigned long addr = (unsigned long)page_address(page);
 931
 932         return set_memory_uc(addr, numpages);
 933 }
 934 EXPORT_SYMBOL(set_pages_uc);
 935
 936 int set_pages_wb(struct page *page, int numpages)
 937 {
 938         unsigned long addr = (unsigned long)page_address(page);
 939
 940         return set_memory_wb(addr, numpages);
 941 }
 942 EXPORT_SYMBOL(set_pages_wb);
 943
 944 int set_pages_x(struct page *page, int numpages)
 945 {
 946         unsigned long addr = (unsigned long)page_address(page);
 947
 948         return set_memory_x(addr, numpages);
 949 }
 950 EXPORT_SYMBOL(set_pages_x);
 951
 952 int set_pages_nx(struct page *page, int numpages)
 953 {
 954         unsigned long addr = (unsigned long)page_address(page);
 955
 956         return set_memory_nx(addr, numpages);
 957 }
 958 EXPORT_SYMBOL(set_pages_nx);
 959
 960 int set_pages_ro(struct page *page, int numpages)
 961 {
 962         unsigned long addr = (unsigned long)page_address(page);
 963
 964         return set_memory_ro(addr, numpages);
 965 }
 966
 967 int set_pages_rw(struct page *page, int numpages)
 968 {
 969         unsigned long addr = (unsigned long)page_address(page);
 970
 971         return set_memory_rw(addr, numpages);
 972 }
 973
 974 #ifdef CONFIG_DEBUG_PAGEALLOC
 975
 976 static int __set_pages_p(struct page *page, int numpages)
 977 {
 978         struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
 979                                 .numpages = numpages,
 980                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
 981                                 .mask_clr = __pgprot(0)};
 982
 983         return __change_page_attr_set_clr(&cpa, 1);
 984 }
 985
 986 static int __set_pages_np(struct page *page, int numpages)
 987 {
 988         struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
 989                                 .numpages = numpages,
 990                                 .mask_set = __pgprot(0),
 991                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW)};
 992
 993         return __change_page_attr_set_clr(&cpa, 1);
 994 }
 995
 996 void kernel_map_pages(struct page *page, int numpages, int enable)
 997 {
 998         if (PageHighMem(page))
 999                 return;
1000         if (!enable) {
1001                 debug_check_no_locks_freed(page_address(page),
1002                                            numpages * PAGE_SIZE);
1003         }
1004
1005         /*
1006          * If page allocator is not up yet then do not call c_p_a():
1007          */
1008         if (!debug_pagealloc_enabled)
1009                 return;
1010
1011         /*
1012          * The return value is ignored as the calls cannot fail.
1013          * Large pages are kept enabled at boot time, and are
1014          * split up quickly with DEBUG_PAGEALLOC. If a splitup
1015          * fails here (due to temporary memory shortage) no damage
1016          * is done because we just keep the largepage intact up
1017          * to the next attempt when it will likely be split up:
1018          */
1019         if (enable)
1020                 __set_pages_p(page, numpages);
1021         else
1022                 __set_pages_np(page, numpages);
1023
1024         /*
1025          * We should perform an IPI and flush all tlbs,
1026          * but that can deadlock->flush only current cpu:
1027          */
1028         __flush_tlb_all();
1029
1030         /*
1031          * Try to refill the page pool here. We can do this only after
1032          * the tlb flush.
1033          */
1034         cpa_fill_pool(NULL);
1035 }
1036
1037 #ifdef CONFIG_DEBUG_FS
1038 static int dpa_show(struct seq_file *m, void *v)
1039 {
1040         seq_puts(m, "DEBUG_PAGEALLOC\n");
1041         seq_printf(m, "pool_size     : %lu\n", pool_size);
1042         seq_printf(m, "pool_pages    : %lu\n", pool_pages);
1043         seq_printf(m, "pool_low      : %lu\n", pool_low);
1044         seq_printf(m, "pool_used     : %lu\n", pool_used);
1045         seq_printf(m, "pool_failed   : %lu\n", pool_failed);
1046
1047         return 0;
1048 }
1049
1050 static int dpa_open(struct inode *inode, struct file *filp)
1051 {
1052         return single_open(filp, dpa_show, NULL);
1053 }
1054
1055 static const struct file_operations dpa_fops = {
1056         .open           = dpa_open,
1057         .read           = seq_read,
1058         .llseek         = seq_lseek,
1059         .release        = single_release,
1060 };
1061
1062 static int __init debug_pagealloc_proc_init(void)
1063 {
1064         struct dentry *de;
1065
1066         de = debugfs_create_file("debug_pagealloc", 0600, NULL, NULL,
1067                                  &dpa_fops);
1068         if (!de)
1069                 return -ENOMEM;
1070
1071         return 0;
1072 }
1073 __initcall(debug_pagealloc_proc_init);
1074 #endif
1075
1076 #ifdef CONFIG_HIBERNATION
1077
1078 bool kernel_page_present(struct page *page)
1079 {
1080         unsigned int level;
1081         pte_t *pte;
1082
1083         if (PageHighMem(page))
1084                 return false;
1085
1086         pte = lookup_address((unsigned long)page_address(page), &level);
1087         return (pte_val(*pte) & _PAGE_PRESENT);
1088 }
1089
1090 #endif /* CONFIG_HIBERNATION */
1091
1092 #endif /* CONFIG_DEBUG_PAGEALLOC */
1093
1094 /*
1095  * The testcases use internal knowledge of the implementation that shouldn't
1096  * be exposed to the rest of the kernel. Include these directly here.
1097  */
1098 #ifdef CONFIG_CPA_DEBUG
1099 #include "pageattr-test.c"
1100 #endif