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             flags;
  33         unsigned long   pfn;
  34         unsigned        force_split : 1;
  35         int             curpage;
  36 };
  37
  38 #define CPA_FLUSHTLB 1
  39 #define CPA_ARRAY 2
  40
  41 #ifdef CONFIG_PROC_FS
  42 static unsigned long direct_pages_count[PG_LEVEL_NUM];
  43
  44 void update_page_count(int level, unsigned long pages)
  45 {
  46         unsigned long flags;
  47
  48         /* Protect against CPA */
  49         spin_lock_irqsave(&pgd_lock, flags);
  50         direct_pages_count[level] += pages;
  51         spin_unlock_irqrestore(&pgd_lock, flags);
  52 }
  53
  54 static void split_page_count(int level)
  55 {
  56         direct_pages_count[level]--;
  57         direct_pages_count[level - 1] += PTRS_PER_PTE;
  58 }
  59
  60 int arch_report_meminfo(char *page)
  61 {
  62         int n = sprintf(page, "DirectMap4k:  %8lu\n"
  63                         "DirectMap2M:  %8lu\n",
  64                         direct_pages_count[PG_LEVEL_4K],
  65                         direct_pages_count[PG_LEVEL_2M]);
  66 #ifdef CONFIG_X86_64
  67         n += sprintf(page + n, "DirectMap1G:  %8lu\n",
  68                      direct_pages_count[PG_LEVEL_1G]);
  69 #endif
  70         return n;
  71 }
  72 #else
  73 static inline void split_page_count(int level) { }
  74 #endif
  75
  76 #ifdef CONFIG_X86_64
  77
  78 static inline unsigned long highmap_start_pfn(void)
  79 {
  80         return __pa(_text) >> PAGE_SHIFT;
  81 }
  82
  83 static inline unsigned long highmap_end_pfn(void)
  84 {
  85         return __pa(round_up((unsigned long)_end, PMD_SIZE)) >> PAGE_SHIFT;
  86 }
  87
  88 #endif
  89
  90 #ifdef CONFIG_DEBUG_PAGEALLOC
  91 # define debug_pagealloc 1
  92 #else
  93 # define debug_pagealloc 0
  94 #endif
  95
  96 static inline int
  97 within(unsigned long addr, unsigned long start, unsigned long end)
  98 {
  99         return addr >= start && addr < end;
 100 }
 101
 102 /*
 103  * Flushing functions
 104  */
 105
 106 /**
 107  * clflush_cache_range - flush a cache range with clflush
 108  * @addr:       virtual start address
 109  * @size:       number of bytes to flush
 110  *
 111  * clflush is an unordered instruction which needs fencing with mfence
 112  * to avoid ordering issues.
 113  */
 114 void clflush_cache_range(void *vaddr, unsigned int size)
 115 {
 116         void *vend = vaddr + size - 1;
 117
 118         mb();
 119
 120         for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
 121                 clflush(vaddr);
 122         /*
 123          * Flush any possible final partial cacheline:
 124          */
 125         clflush(vend);
 126
 127         mb();
 128 }
 129
 130 static void __cpa_flush_all(void *arg)
 131 {
 132         unsigned long cache = (unsigned long)arg;
 133
 134         /*
 135          * Flush all to work around Errata in early athlons regarding
 136          * large page flushing.
 137          */
 138         __flush_tlb_all();
 139
 140         if (cache && boot_cpu_data.x86_model >= 4)
 141                 wbinvd();
 142 }
 143
 144 static void cpa_flush_all(unsigned long cache)
 145 {
 146         BUG_ON(irqs_disabled());
 147
 148         on_each_cpu(__cpa_flush_all, (void *) cache, 1);
 149 }
 150
 151 static void __cpa_flush_range(void *arg)
 152 {
 153         /*
 154          * We could optimize that further and do individual per page
 155          * tlb invalidates for a low number of pages. Caveat: we must
 156          * flush the high aliases on 64bit as well.
 157          */
 158         __flush_tlb_all();
 159 }
 160
 161 static void cpa_flush_range(unsigned long start, int numpages, int cache)
 162 {
 163         unsigned int i, level;
 164         unsigned long addr;
 165
 166         BUG_ON(irqs_disabled());
 167         WARN_ON(PAGE_ALIGN(start) != start);
 168
 169         on_each_cpu(__cpa_flush_range, NULL, 1);
 170
 171         if (!cache)
 172                 return;
 173
 174         /*
 175          * We only need to flush on one CPU,
 176          * clflush is a MESI-coherent instruction that
 177          * will cause all other CPUs to flush the same
 178          * cachelines:
 179          */
 180         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
 181                 pte_t *pte = lookup_address(addr, &level);
 182
 183                 /*
 184                  * Only flush present addresses:
 185                  */
 186                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 187                         clflush_cache_range((void *) addr, PAGE_SIZE);
 188         }
 189 }
 190
 191 static void cpa_flush_array(unsigned long *start, int numpages, int cache)
 192 {
 193         unsigned int i, level;
 194         unsigned long *addr;
 195
 196         BUG_ON(irqs_disabled());
 197
 198         on_each_cpu(__cpa_flush_range, NULL, 1);
 199
 200         if (!cache)
 201                 return;
 202
 203         /* 4M threshold */
 204         if (numpages >= 1024) {
 205                 if (boot_cpu_data.x86_model >= 4)
 206                         wbinvd();
 207                 return;
 208         }
 209         /*
 210          * We only need to flush on one CPU,
 211          * clflush is a MESI-coherent instruction that
 212          * will cause all other CPUs to flush the same
 213          * cachelines:
 214          */
 215         for (i = 0, addr = start; i < numpages; i++, addr++) {
 216                 pte_t *pte = lookup_address(*addr, &level);
 217
 218                 /*
 219                  * Only flush present addresses:
 220                  */
 221                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 222                         clflush_cache_range((void *) *addr, PAGE_SIZE);
 223         }
 224 }
 225
 226 /*
 227  * Certain areas of memory on x86 require very specific protection flags,
 228  * for example the BIOS area or kernel text. Callers don't always get this
 229  * right (again, ioremap() on BIOS memory is not uncommon) so this function
 230  * checks and fixes these known static required protection bits.
 231  */
 232 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
 233                                    unsigned long pfn)
 234 {
 235         pgprot_t forbidden = __pgprot(0);
 236
 237         /*
 238          * The BIOS area between 640k and 1Mb needs to be executable for
 239          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
 240          */
 241         if (within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
 242                 pgprot_val(forbidden) |= _PAGE_NX;
 243
 244         /*
 245          * The kernel text needs to be executable for obvious reasons
 246          * Does not cover __inittext since that is gone later on. On
 247          * 64bit we do not enforce !NX on the low mapping
 248          */
 249         if (within(address, (unsigned long)_text, (unsigned long)_etext))
 250                 pgprot_val(forbidden) |= _PAGE_NX;
 251
 252         /*
 253          * The .rodata section needs to be read-only. Using the pfn
 254          * catches all aliases.
 255          */
 256         if (within(pfn, __pa((unsigned long)__start_rodata) >> PAGE_SHIFT,
 257                    __pa((unsigned long)__end_rodata) >> PAGE_SHIFT))
 258                 pgprot_val(forbidden) |= _PAGE_RW;
 259
 260         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
 261
 262         return prot;
 263 }
 264
 265 /*
 266  * Lookup the page table entry for a virtual address. Return a pointer
 267  * to the entry and the level of the mapping.
 268  *
 269  * Note: We return pud and pmd either when the entry is marked large
 270  * or when the present bit is not set. Otherwise we would return a
 271  * pointer to a nonexisting mapping.
 272  */
 273 pte_t *lookup_address(unsigned long address, unsigned int *level)
 274 {
 275         pgd_t *pgd = pgd_offset_k(address);
 276         pud_t *pud;
 277         pmd_t *pmd;
 278
 279         *level = PG_LEVEL_NONE;
 280
 281         if (pgd_none(*pgd))
 282                 return NULL;
 283
 284         pud = pud_offset(pgd, address);
 285         if (pud_none(*pud))
 286                 return NULL;
 287
 288         *level = PG_LEVEL_1G;
 289         if (pud_large(*pud) || !pud_present(*pud))
 290                 return (pte_t *)pud;
 291
 292         pmd = pmd_offset(pud, address);
 293         if (pmd_none(*pmd))
 294                 return NULL;
 295
 296         *level = PG_LEVEL_2M;
 297         if (pmd_large(*pmd) || !pmd_present(*pmd))
 298                 return (pte_t *)pmd;
 299
 300         *level = PG_LEVEL_4K;
 301
 302         return pte_offset_kernel(pmd, address);
 303 }
 304 EXPORT_SYMBOL_GPL(lookup_address);
 305
 306 /*
 307  * Set the new pmd in all the pgds we know about:
 308  */
 309 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 310 {
 311         /* change init_mm */
 312         set_pte_atomic(kpte, pte);
 313 #ifdef CONFIG_X86_32
 314         if (!SHARED_KERNEL_PMD) {
 315                 struct page *page;
 316
 317                 list_for_each_entry(page, &pgd_list, lru) {
 318                         pgd_t *pgd;
 319                         pud_t *pud;
 320                         pmd_t *pmd;
 321
 322                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 323                         pud = pud_offset(pgd, address);
 324                         pmd = pmd_offset(pud, address);
 325                         set_pte_atomic((pte_t *)pmd, pte);
 326                 }
 327         }
 328 #endif
 329 }
 330
 331 static int
 332 try_preserve_large_page(pte_t *kpte, unsigned long address,
 333                         struct cpa_data *cpa)
 334 {
 335         unsigned long nextpage_addr, numpages, pmask, psize, flags, addr, pfn;
 336         pte_t new_pte, old_pte, *tmp;
 337         pgprot_t old_prot, new_prot;
 338         int i, do_split = 1;
 339         unsigned int level;
 340
 341         if (cpa->force_split)
 342                 return 1;
 343
 344         spin_lock_irqsave(&pgd_lock, flags);
 345         /*
 346          * Check for races, another CPU might have split this page
 347          * up already:
 348          */
 349         tmp = lookup_address(address, &level);
 350         if (tmp != kpte)
 351                 goto out_unlock;
 352
 353         switch (level) {
 354         case PG_LEVEL_2M:
 355                 psize = PMD_PAGE_SIZE;
 356                 pmask = PMD_PAGE_MASK;
 357                 break;
 358 #ifdef CONFIG_X86_64
 359         case PG_LEVEL_1G:
 360                 psize = PUD_PAGE_SIZE;
 361                 pmask = PUD_PAGE_MASK;
 362                 break;
 363 #endif
 364         default:
 365                 do_split = -EINVAL;
 366                 goto out_unlock;
 367         }
 368
 369         /*
 370          * Calculate the number of pages, which fit into this large
 371          * page starting at address:
 372          */
 373         nextpage_addr = (address + psize) & pmask;
 374         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
 375         if (numpages < cpa->numpages)
 376                 cpa->numpages = numpages;
 377
 378         /*
 379          * We are safe now. Check whether the new pgprot is the same:
 380          */
 381         old_pte = *kpte;
 382         old_prot = new_prot = pte_pgprot(old_pte);
 383
 384         pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
 385         pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
 386
 387         /*
 388          * old_pte points to the large page base address. So we need
 389          * to add the offset of the virtual address:
 390          */
 391         pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
 392         cpa->pfn = pfn;
 393
 394         new_prot = static_protections(new_prot, address, pfn);
 395
 396         /*
 397          * We need to check the full range, whether
 398          * static_protection() requires a different pgprot for one of
 399          * the pages in the range we try to preserve:
 400          */
 401         addr = address + PAGE_SIZE;
 402         pfn++;
 403         for (i = 1; i < cpa->numpages; i++, addr += PAGE_SIZE, pfn++) {
 404                 pgprot_t chk_prot = static_protections(new_prot, addr, pfn);
 405
 406                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
 407                         goto out_unlock;
 408         }
 409
 410         /*
 411          * If there are no changes, return. maxpages has been updated
 412          * above:
 413          */
 414         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
 415                 do_split = 0;
 416                 goto out_unlock;
 417         }
 418
 419         /*
 420          * We need to change the attributes. Check, whether we can
 421          * change the large page in one go. We request a split, when
 422          * the address is not aligned and the number of pages is
 423          * smaller than the number of pages in the large page. Note
 424          * that we limited the number of possible pages already to
 425          * the number of pages in the large page.
 426          */
 427         if (address == (nextpage_addr - psize) && cpa->numpages == numpages) {
 428                 /*
 429                  * The address is aligned and the number of pages
 430                  * covers the full page.
 431                  */
 432                 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
 433                 __set_pmd_pte(kpte, address, new_pte);
 434                 cpa->flags |= CPA_FLUSHTLB;
 435                 do_split = 0;
 436         }
 437
 438 out_unlock:
 439         spin_unlock_irqrestore(&pgd_lock, flags);
 440
 441         return do_split;
 442 }
 443
 444 static LIST_HEAD(page_pool);
 445 static unsigned long pool_size, pool_pages, pool_low;
 446 static unsigned long pool_used, pool_failed;
 447
 448 static void cpa_fill_pool(struct page **ret)
 449 {
 450         gfp_t gfp = GFP_KERNEL;
 451         unsigned long flags;
 452         struct page *p;
 453
 454         /*
 455          * Avoid recursion (on debug-pagealloc) and also signal
 456          * our priority to get to these pagetables:
 457          */
 458         if (current->flags & PF_MEMALLOC)
 459                 return;
 460         current->flags |= PF_MEMALLOC;
 461
 462         /*
 463          * Allocate atomically from atomic contexts:
 464          */
 465         if (in_atomic() || irqs_disabled() || debug_pagealloc)
 466                 gfp =  GFP_ATOMIC | __GFP_NORETRY | __GFP_NOWARN;
 467
 468         while (pool_pages < pool_size || (ret && !*ret)) {
 469                 p = alloc_pages(gfp, 0);
 470                 if (!p) {
 471                         pool_failed++;
 472                         break;
 473                 }
 474                 /*
 475                  * If the call site needs a page right now, provide it:
 476                  */
 477                 if (ret && !*ret) {
 478                         *ret = p;
 479                         continue;
 480                 }
 481                 spin_lock_irqsave(&pgd_lock, flags);
 482                 list_add(&p->lru, &page_pool);
 483                 pool_pages++;
 484                 spin_unlock_irqrestore(&pgd_lock, flags);
 485         }
 486
 487         current->flags &= ~PF_MEMALLOC;
 488 }
 489
 490 #define SHIFT_MB                (20 - PAGE_SHIFT)
 491 #define ROUND_MB_GB             ((1 << 10) - 1)
 492 #define SHIFT_MB_GB             10
 493 #define POOL_PAGES_PER_GB       16
 494
 495 void __init cpa_init(void)
 496 {
 497         struct sysinfo si;
 498         unsigned long gb;
 499
 500         si_meminfo(&si);
 501         /*
 502          * Calculate the number of pool pages:
 503          *
 504          * Convert totalram (nr of pages) to MiB and round to the next
 505          * GiB. Shift MiB to Gib and multiply the result by
 506          * POOL_PAGES_PER_GB:
 507          */
 508         if (debug_pagealloc) {
 509                 gb = ((si.totalram >> SHIFT_MB) + ROUND_MB_GB) >> SHIFT_MB_GB;
 510                 pool_size = POOL_PAGES_PER_GB * gb;
 511         } else {
 512                 pool_size = 1;
 513         }
 514         pool_low = pool_size;
 515
 516         cpa_fill_pool(NULL);
 517         printk(KERN_DEBUG
 518                "CPA: page pool initialized %lu of %lu pages preallocated\n",
 519                pool_pages, pool_size);
 520 }
 521
 522 static int split_large_page(pte_t *kpte, unsigned long address)
 523 {
 524         unsigned long flags, pfn, pfninc = 1;
 525         unsigned int i, level;
 526         pte_t *pbase, *tmp;
 527         pgprot_t ref_prot;
 528         struct page *base;
 529
 530         /*
 531          * Get a page from the pool. The pool list is protected by the
 532          * pgd_lock, which we have to take anyway for the split
 533          * operation:
 534          */
 535         spin_lock_irqsave(&pgd_lock, flags);
 536         if (list_empty(&page_pool)) {
 537                 spin_unlock_irqrestore(&pgd_lock, flags);
 538                 base = NULL;
 539                 cpa_fill_pool(&base);
 540                 if (!base)
 541                         return -ENOMEM;
 542                 spin_lock_irqsave(&pgd_lock, flags);
 543         } else {
 544                 base = list_first_entry(&page_pool, struct page, lru);
 545                 list_del(&base->lru);
 546                 pool_pages--;
 547
 548                 if (pool_pages < pool_low)
 549                         pool_low = pool_pages;
 550         }
 551
 552         /*
 553          * Check for races, another CPU might have split this page
 554          * up for us already:
 555          */
 556         tmp = lookup_address(address, &level);
 557         if (tmp != kpte)
 558                 goto out_unlock;
 559
 560         pbase = (pte_t *)page_address(base);
 561         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
 562         ref_prot = pte_pgprot(pte_clrhuge(*kpte));
 563
 564 #ifdef CONFIG_X86_64
 565         if (level == PG_LEVEL_1G) {
 566                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
 567                 pgprot_val(ref_prot) |= _PAGE_PSE;
 568         }
 569 #endif
 570
 571         /*
 572          * Get the target pfn from the original entry:
 573          */
 574         pfn = pte_pfn(*kpte);
 575         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
 576                 set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
 577
 578         if (address >= (unsigned long)__va(0) &&
 579                 address < (unsigned long)__va(max_low_pfn_mapped << PAGE_SHIFT))
 580                 split_page_count(level);
 581
 582 #ifdef CONFIG_X86_64
 583         if (address >= (unsigned long)__va(1UL<<32) &&
 584                 address < (unsigned long)__va(max_pfn_mapped << PAGE_SHIFT))
 585                 split_page_count(level);
 586 #endif
 587
 588         /*
 589          * Install the new, split up pagetable. Important details here:
 590          *
 591          * On Intel the NX bit of all levels must be cleared to make a
 592          * page executable. See section 4.13.2 of Intel 64 and IA-32
 593          * Architectures Software Developer's Manual).
 594          *
 595          * Mark the entry present. The current mapping might be
 596          * set to not present, which we preserved above.
 597          */
 598         ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
 599         pgprot_val(ref_prot) |= _PAGE_PRESENT;
 600         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
 601         base = NULL;
 602
 603 out_unlock:
 604         /*
 605          * If we dropped out via the lookup_address check under
 606          * pgd_lock then stick the page back into the pool:
 607          */
 608         if (base) {
 609                 list_add(&base->lru, &page_pool);
 610                 pool_pages++;
 611         } else
 612                 pool_used++;
 613         spin_unlock_irqrestore(&pgd_lock, flags);
 614
 615         return 0;
 616 }
 617
 618 static int __change_page_attr(struct cpa_data *cpa, int primary)
 619 {
 620         unsigned long address;
 621         int do_split, err;
 622         unsigned int level;
 623         pte_t *kpte, old_pte;
 624
 625         if (cpa->flags & CPA_ARRAY)
 626                 address = cpa->vaddr[cpa->curpage];
 627         else
 628                 address = *cpa->vaddr;
 629
 630 repeat:
 631         kpte = lookup_address(address, &level);
 632         if (!kpte)
 633                 return 0;
 634
 635         old_pte = *kpte;
 636         if (!pte_val(old_pte)) {
 637                 if (!primary)
 638                         return 0;
 639                 printk(KERN_WARNING "CPA: called for zero pte. "
 640                        "vaddr = %lx cpa->vaddr = %lx\n", address,
 641                 WARN_ON(1);
 642                        *cpa->vaddr);
 643                 return -EINVAL;
 644         }
 645
 646         if (level == PG_LEVEL_4K) {
 647                 pte_t new_pte;
 648                 pgprot_t new_prot = pte_pgprot(old_pte);
 649                 unsigned long pfn = pte_pfn(old_pte);
 650
 651                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
 652                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
 653
 654                 new_prot = static_protections(new_prot, address, pfn);
 655
 656                 /*
 657                  * We need to keep the pfn from the existing PTE,
 658                  * after all we're only going to change it's attributes
 659                  * not the memory it points to
 660                  */
 661                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
 662                 cpa->pfn = pfn;
 663                 /*
 664                  * Do we really change anything ?
 665                  */
 666                 if (pte_val(old_pte) != pte_val(new_pte)) {
 667                         set_pte_atomic(kpte, new_pte);
 668                         cpa->flags |= CPA_FLUSHTLB;
 669                 }
 670                 cpa->numpages = 1;
 671                 return 0;
 672         }
 673
 674         /*
 675          * Check, whether we can keep the large page intact
 676          * and just change the pte:
 677          */
 678         do_split = try_preserve_large_page(kpte, address, cpa);
 679         /*
 680          * When the range fits into the existing large page,
 681          * return. cp->numpages and cpa->tlbflush have been updated in
 682          * try_large_page:
 683          */
 684         if (do_split <= 0)
 685                 return do_split;
 686
 687         /*
 688          * We have to split the large page:
 689          */
 690         err = split_large_page(kpte, address);
 691         if (!err) {
 692                 cpa->flags |= CPA_FLUSHTLB;
 693                 goto repeat;
 694         }
 695
 696         return err;
 697 }
 698
 699 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
 700
 701 static int cpa_process_alias(struct cpa_data *cpa)
 702 {
 703         struct cpa_data alias_cpa;
 704         int ret = 0;
 705         unsigned long temp_cpa_vaddr, vaddr;
 706
 707         if (cpa->pfn >= max_pfn_mapped)
 708                 return 0;
 709
 710 #ifdef CONFIG_X86_64
 711         if (cpa->pfn >= max_low_pfn_mapped && cpa->pfn < (1UL<<(32-PAGE_SHIFT)))
 712                 return 0;
 713 #endif
 714         /*
 715          * No need to redo, when the primary call touched the direct
 716          * mapping already:
 717          */
 718         if (cpa->flags & CPA_ARRAY)
 719                 vaddr = cpa->vaddr[cpa->curpage];
 720         else
 721                 vaddr = *cpa->vaddr;
 722
 723         if (!(within(vaddr, PAGE_OFFSET,
 724                     PAGE_OFFSET + (max_low_pfn_mapped << PAGE_SHIFT))
 725 #ifdef CONFIG_X86_64
 726                 || within(vaddr, PAGE_OFFSET + (1UL<<32),
 727                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))
 728 #endif
 729         )) {
 730
 731                 alias_cpa = *cpa;
 732                 temp_cpa_vaddr = (unsigned long) __va(cpa->pfn << PAGE_SHIFT);
 733                 alias_cpa.vaddr = &temp_cpa_vaddr;
 734                 alias_cpa.flags &= ~CPA_ARRAY;
 735
 736
 737                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
 738         }
 739
 740 #ifdef CONFIG_X86_64
 741         if (ret)
 742                 return ret;
 743         /*
 744          * No need to redo, when the primary call touched the high
 745          * mapping already:
 746          */
 747         if (within(vaddr, (unsigned long) _text, (unsigned long) _end))
 748                 return 0;
 749
 750         /*
 751          * If the physical address is inside the kernel map, we need
 752          * to touch the high mapped kernel as well:
 753          */
 754         if (!within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn()))
 755                 return 0;
 756
 757         alias_cpa = *cpa;
 758         temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) + __START_KERNEL_map - phys_base;
 759         alias_cpa.vaddr = &temp_cpa_vaddr;
 760         alias_cpa.flags &= ~CPA_ARRAY;
 761
 762         /*
 763          * The high mapping range is imprecise, so ignore the return value.
 764          */
 765         __change_page_attr_set_clr(&alias_cpa, 0);
 766 #endif
 767         return ret;
 768 }
 769
 770 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
 771 {
 772         int ret, numpages = cpa->numpages;
 773
 774         while (numpages) {
 775                 /*
 776                  * Store the remaining nr of pages for the large page
 777                  * preservation check.
 778                  */
 779                 cpa->numpages = numpages;
 780                 /* for array changes, we can't use large page */
 781                 if (cpa->flags & CPA_ARRAY)
 782                         cpa->numpages = 1;
 783
 784                 ret = __change_page_attr(cpa, checkalias);
 785                 if (ret)
 786                         return ret;
 787
 788                 if (checkalias) {
 789                         ret = cpa_process_alias(cpa);
 790                         if (ret)
 791                                 return ret;
 792                 }
 793
 794                 /*
 795                  * Adjust the number of pages with the result of the
 796                  * CPA operation. Either a large page has been
 797                  * preserved or a single page update happened.
 798                  */
 799                 BUG_ON(cpa->numpages > numpages);
 800                 numpages -= cpa->numpages;
 801                 if (cpa->flags & CPA_ARRAY)
 802                         cpa->curpage++;
 803                 else
 804                         *cpa->vaddr += cpa->numpages * PAGE_SIZE;
 805
 806         }
 807         return 0;
 808 }
 809
 810 static inline int cache_attr(pgprot_t attr)
 811 {
 812         return pgprot_val(attr) &
 813                 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
 814 }
 815
 816 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
 817                                     pgprot_t mask_set, pgprot_t mask_clr,
 818                                     int force_split, int array)
 819 {
 820         struct cpa_data cpa;
 821         int ret, cache, checkalias;
 822
 823         /*
 824          * Check, if we are requested to change a not supported
 825          * feature:
 826          */
 827         mask_set = canon_pgprot(mask_set);
 828         mask_clr = canon_pgprot(mask_clr);
 829         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
 830                 return 0;
 831
 832         /* Ensure we are PAGE_SIZE aligned */
 833         if (!array) {
 834                 if (*addr & ~PAGE_MASK) {
 835                         *addr &= PAGE_MASK;
 836                         /*
 837                          * People should not be passing in unaligned addresses:
 838                          */
 839                         WARN_ON_ONCE(1);
 840                 }
 841         } else {
 842                 int i;
 843                 for (i = 0; i < numpages; i++) {
 844                         if (addr[i] & ~PAGE_MASK) {
 845                                 addr[i] &= PAGE_MASK;
 846                                 WARN_ON_ONCE(1);
 847                         }
 848                 }
 849         }
 850
 851         /* Must avoid aliasing mappings in the highmem code */
 852         kmap_flush_unused();
 853
 854         cpa.vaddr = addr;
 855         cpa.numpages = numpages;
 856         cpa.mask_set = mask_set;
 857         cpa.mask_clr = mask_clr;
 858         cpa.flags = 0;
 859         cpa.curpage = 0;
 860         cpa.force_split = force_split;
 861
 862         if (array)
 863                 cpa.flags |= CPA_ARRAY;
 864
 865         /* No alias checking for _NX bit modifications */
 866         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
 867
 868         ret = __change_page_attr_set_clr(&cpa, checkalias);
 869
 870         /*
 871          * Check whether we really changed something:
 872          */
 873         if (!(cpa.flags & CPA_FLUSHTLB))
 874                 goto out;
 875
 876         /*
 877          * No need to flush, when we did not set any of the caching
 878          * attributes:
 879          */
 880         cache = cache_attr(mask_set);
 881
 882         /*
 883          * On success we use clflush, when the CPU supports it to
 884          * avoid the wbindv. If the CPU does not support it and in the
 885          * error case we fall back to cpa_flush_all (which uses
 886          * wbindv):
 887          */
 888         if (!ret && cpu_has_clflush) {
 889                 if (cpa.flags & CPA_ARRAY)
 890                         cpa_flush_array(addr, numpages, cache);
 891                 else
 892                         cpa_flush_range(*addr, numpages, cache);
 893         } else
 894                 cpa_flush_all(cache);
 895
 896 out:
 897         cpa_fill_pool(NULL);
 898
 899         return ret;
 900 }
 901
 902 static inline int change_page_attr_set(unsigned long *addr, int numpages,
 903                                        pgprot_t mask, int array)
 904 {
 905         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
 906                 array);
 907 }
 908
 909 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
 910                                          pgprot_t mask, int array)
 911 {
 912         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
 913                 array);
 914 }
 915
 916 int _set_memory_uc(unsigned long addr, int numpages)
 917 {
 918         /*
 919          * for now UC MINUS. see comments in ioremap_nocache()
 920          */
 921         return change_page_attr_set(&addr, numpages,
 922                                     __pgprot(_PAGE_CACHE_UC_MINUS), 0);
 923 }
 924
 925 int set_memory_uc(unsigned long addr, int numpages)
 926 {
 927         /*
 928          * for now UC MINUS. see comments in ioremap_nocache()
 929          */
 930         if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
 931                             _PAGE_CACHE_UC_MINUS, NULL))
 932                 return -EINVAL;
 933
 934         return _set_memory_uc(addr, numpages);
 935 }
 936 EXPORT_SYMBOL(set_memory_uc);
 937
 938 int set_memory_array_uc(unsigned long *addr, int addrinarray)
 939 {
 940         int i;
 941         /*
 942          * for now UC MINUS. see comments in ioremap_nocache()
 943          */
 944         for (i = 0; i < addrinarray; i++) {
 945                 if (reserve_memtype(addr[i], addr[i] + PAGE_SIZE,
 946                             _PAGE_CACHE_UC_MINUS, NULL))
 947                         goto out;
 948         }
 949
 950         return change_page_attr_set(addr, addrinarray,
 951                                     __pgprot(_PAGE_CACHE_UC_MINUS), 1);
 952 out:
 953         while (--i >= 0)
 954                 free_memtype(addr[i], addr[i] + PAGE_SIZE);
 955         return -EINVAL;
 956 }
 957 EXPORT_SYMBOL(set_memory_array_uc);
 958
 959 int _set_memory_wc(unsigned long addr, int numpages)
 960 {
 961         return change_page_attr_set(&addr, numpages,
 962                                     __pgprot(_PAGE_CACHE_WC), 0);
 963 }
 964
 965 int set_memory_wc(unsigned long addr, int numpages)
 966 {
 967         if (!pat_enabled)
 968                 return set_memory_uc(addr, numpages);
 969
 970         if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
 971                 _PAGE_CACHE_WC, NULL))
 972                 return -EINVAL;
 973
 974         return _set_memory_wc(addr, numpages);
 975 }
 976 EXPORT_SYMBOL(set_memory_wc);
 977
 978 int _set_memory_wb(unsigned long addr, int numpages)
 979 {
 980         return change_page_attr_clear(&addr, numpages,
 981                                       __pgprot(_PAGE_CACHE_MASK), 0);
 982 }
 983
 984 int set_memory_wb(unsigned long addr, int numpages)
 985 {
 986         free_memtype(addr, addr + numpages * PAGE_SIZE);
 987
 988         return _set_memory_wb(addr, numpages);
 989 }
 990 EXPORT_SYMBOL(set_memory_wb);
 991
 992 int set_memory_array_wb(unsigned long *addr, int addrinarray)
 993 {
 994         int i;
 995         for (i = 0; i < addrinarray; i++)
 996                 free_memtype(addr[i], addr[i] + PAGE_SIZE);
 997
 998         return change_page_attr_clear(addr, addrinarray,
 999                                       __pgprot(_PAGE_CACHE_MASK), 1);
1000 }
1001 EXPORT_SYMBOL(set_memory_array_wb);
1002
1003 int set_memory_x(unsigned long addr, int numpages)
1004 {
1005         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1006 }
1007 EXPORT_SYMBOL(set_memory_x);
1008
1009 int set_memory_nx(unsigned long addr, int numpages)
1010 {
1011         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1012 }
1013 EXPORT_SYMBOL(set_memory_nx);
1014
1015 int set_memory_ro(unsigned long addr, int numpages)
1016 {
1017         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1018 }
1019
1020 int set_memory_rw(unsigned long addr, int numpages)
1021 {
1022         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1023 }
1024
1025 int set_memory_np(unsigned long addr, int numpages)
1026 {
1027         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1028 }
1029
1030 int set_memory_4k(unsigned long addr, int numpages)
1031 {
1032         return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1033                                         __pgprot(0), 1, 0);
1034 }
1035
1036 int set_pages_uc(struct page *page, int numpages)
1037 {
1038         unsigned long addr = (unsigned long)page_address(page);
1039
1040         return set_memory_uc(addr, numpages);
1041 }
1042 EXPORT_SYMBOL(set_pages_uc);
1043
1044 int set_pages_wb(struct page *page, int numpages)
1045 {
1046         unsigned long addr = (unsigned long)page_address(page);
1047
1048         return set_memory_wb(addr, numpages);
1049 }
1050 EXPORT_SYMBOL(set_pages_wb);
1051
1052 int set_pages_x(struct page *page, int numpages)
1053 {
1054         unsigned long addr = (unsigned long)page_address(page);
1055
1056         return set_memory_x(addr, numpages);
1057 }
1058 EXPORT_SYMBOL(set_pages_x);
1059
1060 int set_pages_nx(struct page *page, int numpages)
1061 {
1062         unsigned long addr = (unsigned long)page_address(page);
1063
1064         return set_memory_nx(addr, numpages);
1065 }
1066 EXPORT_SYMBOL(set_pages_nx);
1067
1068 int set_pages_ro(struct page *page, int numpages)
1069 {
1070         unsigned long addr = (unsigned long)page_address(page);
1071
1072         return set_memory_ro(addr, numpages);
1073 }
1074
1075 int set_pages_rw(struct page *page, int numpages)
1076 {
1077         unsigned long addr = (unsigned long)page_address(page);
1078
1079         return set_memory_rw(addr, numpages);
1080 }
1081
1082 #ifdef CONFIG_DEBUG_PAGEALLOC
1083
1084 static int __set_pages_p(struct page *page, int numpages)
1085 {
1086         unsigned long tempaddr = (unsigned long) page_address(page);
1087         struct cpa_data cpa = { .vaddr = &tempaddr,
1088                                 .numpages = numpages,
1089                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1090                                 .mask_clr = __pgprot(0),
1091                                 .flags = 0};
1092
1093         return __change_page_attr_set_clr(&cpa, 1);
1094 }
1095
1096 static int __set_pages_np(struct page *page, int numpages)
1097 {
1098         unsigned long tempaddr = (unsigned long) page_address(page);
1099         struct cpa_data cpa = { .vaddr = &tempaddr,
1100                                 .numpages = numpages,
1101                                 .mask_set = __pgprot(0),
1102                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1103                                 .flags = 0};
1104
1105         return __change_page_attr_set_clr(&cpa, 1);
1106 }
1107
1108 void kernel_map_pages(struct page *page, int numpages, int enable)
1109 {
1110         if (PageHighMem(page))
1111                 return;
1112         if (!enable) {
1113                 debug_check_no_locks_freed(page_address(page),
1114                                            numpages * PAGE_SIZE);
1115         }
1116
1117         /*
1118          * If page allocator is not up yet then do not call c_p_a():
1119          */
1120         if (!debug_pagealloc_enabled)
1121                 return;
1122
1123         /*
1124          * The return value is ignored as the calls cannot fail.
1125          * Large pages are kept enabled at boot time, and are
1126          * split up quickly with DEBUG_PAGEALLOC. If a splitup
1127          * fails here (due to temporary memory shortage) no damage
1128          * is done because we just keep the largepage intact up
1129          * to the next attempt when it will likely be split up:
1130          */
1131         if (enable)
1132                 __set_pages_p(page, numpages);
1133         else
1134                 __set_pages_np(page, numpages);
1135
1136         /*
1137          * We should perform an IPI and flush all tlbs,
1138          * but that can deadlock->flush only current cpu:
1139          */
1140         __flush_tlb_all();
1141
1142         /*
1143          * Try to refill the page pool here. We can do this only after
1144          * the tlb flush.
1145          */
1146         cpa_fill_pool(NULL);
1147 }
1148
1149 #ifdef CONFIG_DEBUG_FS
1150 static int dpa_show(struct seq_file *m, void *v)
1151 {
1152         seq_puts(m, "DEBUG_PAGEALLOC\n");
1153         seq_printf(m, "pool_size     : %lu\n", pool_size);
1154         seq_printf(m, "pool_pages    : %lu\n", pool_pages);
1155         seq_printf(m, "pool_low      : %lu\n", pool_low);
1156         seq_printf(m, "pool_used     : %lu\n", pool_used);
1157         seq_printf(m, "pool_failed   : %lu\n", pool_failed);
1158
1159         return 0;
1160 }
1161
1162 static int dpa_open(struct inode *inode, struct file *filp)
1163 {
1164         return single_open(filp, dpa_show, NULL);
1165 }
1166
1167 static const struct file_operations dpa_fops = {
1168         .open           = dpa_open,
1169         .read           = seq_read,
1170         .llseek         = seq_lseek,
1171         .release        = single_release,
1172 };
1173
1174 static int __init debug_pagealloc_proc_init(void)
1175 {
1176         struct dentry *de;
1177
1178         de = debugfs_create_file("debug_pagealloc", 0600, NULL, NULL,
1179                                  &dpa_fops);
1180         if (!de)
1181                 return -ENOMEM;
1182
1183         return 0;
1184 }
1185 __initcall(debug_pagealloc_proc_init);
1186 #endif
1187
1188 #ifdef CONFIG_HIBERNATION
1189
1190 bool kernel_page_present(struct page *page)
1191 {
1192         unsigned int level;
1193         pte_t *pte;
1194
1195         if (PageHighMem(page))
1196                 return false;
1197
1198         pte = lookup_address((unsigned long)page_address(page), &level);
1199         return (pte_val(*pte) & _PAGE_PRESENT);
1200 }
1201
1202 #endif /* CONFIG_HIBERNATION */
1203
1204 #endif /* CONFIG_DEBUG_PAGEALLOC */
1205
1206 /*
1207  * The testcases use internal knowledge of the implementation that shouldn't
1208  * be exposed to the rest of the kernel. Include these directly here.
1209  */
1210 #ifdef CONFIG_CPA_DEBUG
1211 #include "pageattr-test.c"
1212 #endif