SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 kB\n",
  63                         direct_pages_count[PG_LEVEL_4K] << 2);
  64 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
  65         n += sprintf(page + n, "DirectMap2M:  %8lu kB\n",
  66                         direct_pages_count[PG_LEVEL_2M] << 11);
  67 #else
  68         n += sprintf(page + n, "DirectMap4M:  %8lu kB\n",
  69                         direct_pages_count[PG_LEVEL_2M] << 12);
  70 #endif
  71 #ifdef CONFIG_X86_64
  72         if (direct_gbpages)
  73                 n += sprintf(page + n, "DirectMap1G:  %8lu kB\n",
  74                         direct_pages_count[PG_LEVEL_1G] << 20);
  75 #endif
  76         return n;
  77 }
  78 #else
  79 static inline void split_page_count(int level) { }
  80 #endif
  81
  82 #ifdef CONFIG_X86_64
  83
  84 static inline unsigned long highmap_start_pfn(void)
  85 {
  86         return __pa(_text) >> PAGE_SHIFT;
  87 }
  88
  89 static inline unsigned long highmap_end_pfn(void)
  90 {
  91         return __pa(round_up((unsigned long)_end, PMD_SIZE)) >> PAGE_SHIFT;
  92 }
  93
  94 #endif
  95
  96 #ifdef CONFIG_DEBUG_PAGEALLOC
  97 # define debug_pagealloc 1
  98 #else
  99 # define debug_pagealloc 0
 100 #endif
 101
 102 static inline int
 103 within(unsigned long addr, unsigned long start, unsigned long end)
 104 {
 105         return addr >= start && addr < end;
 106 }
 107
 108 /*
 109  * Flushing functions
 110  */
 111
 112 /**
 113  * clflush_cache_range - flush a cache range with clflush
 114  * @addr:       virtual start address
 115  * @size:       number of bytes to flush
 116  *
 117  * clflush is an unordered instruction which needs fencing with mfence
 118  * to avoid ordering issues.
 119  */
 120 void clflush_cache_range(void *vaddr, unsigned int size)
 121 {
 122         void *vend = vaddr + size - 1;
 123
 124         mb();
 125
 126         for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
 127                 clflush(vaddr);
 128         /*
 129          * Flush any possible final partial cacheline:
 130          */
 131         clflush(vend);
 132
 133         mb();
 134 }
 135
 136 static void __cpa_flush_all(void *arg)
 137 {
 138         unsigned long cache = (unsigned long)arg;
 139
 140         /*
 141          * Flush all to work around Errata in early athlons regarding
 142          * large page flushing.
 143          */
 144         __flush_tlb_all();
 145
 146         if (cache && boot_cpu_data.x86_model >= 4)
 147                 wbinvd();
 148 }
 149
 150 static void cpa_flush_all(unsigned long cache)
 151 {
 152         BUG_ON(irqs_disabled());
 153
 154         on_each_cpu(__cpa_flush_all, (void *) cache, 1);
 155 }
 156
 157 static void __cpa_flush_range(void *arg)
 158 {
 159         /*
 160          * We could optimize that further and do individual per page
 161          * tlb invalidates for a low number of pages. Caveat: we must
 162          * flush the high aliases on 64bit as well.
 163          */
 164         __flush_tlb_all();
 165 }
 166
 167 static void cpa_flush_range(unsigned long start, int numpages, int cache)
 168 {
 169         unsigned int i, level;
 170         unsigned long addr;
 171
 172         BUG_ON(irqs_disabled());
 173         WARN_ON(PAGE_ALIGN(start) != start);
 174
 175         on_each_cpu(__cpa_flush_range, NULL, 1);
 176
 177         if (!cache)
 178                 return;
 179
 180         /*
 181          * We only need to flush on one CPU,
 182          * clflush is a MESI-coherent instruction that
 183          * will cause all other CPUs to flush the same
 184          * cachelines:
 185          */
 186         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
 187                 pte_t *pte = lookup_address(addr, &level);
 188
 189                 /*
 190                  * Only flush present addresses:
 191                  */
 192                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 193                         clflush_cache_range((void *) addr, PAGE_SIZE);
 194         }
 195 }
 196
 197 static void cpa_flush_array(unsigned long *start, int numpages, int cache)
 198 {
 199         unsigned int i, level;
 200         unsigned long *addr;
 201
 202         BUG_ON(irqs_disabled());
 203
 204         on_each_cpu(__cpa_flush_range, NULL, 1);
 205
 206         if (!cache)
 207                 return;
 208
 209         /* 4M threshold */
 210         if (numpages >= 1024) {
 211                 if (boot_cpu_data.x86_model >= 4)
 212                         wbinvd();
 213                 return;
 214         }
 215         /*
 216          * We only need to flush on one CPU,
 217          * clflush is a MESI-coherent instruction that
 218          * will cause all other CPUs to flush the same
 219          * cachelines:
 220          */
 221         for (i = 0, addr = start; i < numpages; i++, addr++) {
 222                 pte_t *pte = lookup_address(*addr, &level);
 223
 224                 /*
 225                  * Only flush present addresses:
 226                  */
 227                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 228                         clflush_cache_range((void *) *addr, PAGE_SIZE);
 229         }
 230 }
 231
 232 /*
 233  * Certain areas of memory on x86 require very specific protection flags,
 234  * for example the BIOS area or kernel text. Callers don't always get this
 235  * right (again, ioremap() on BIOS memory is not uncommon) so this function
 236  * checks and fixes these known static required protection bits.
 237  */
 238 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
 239                                    unsigned long pfn)
 240 {
 241         pgprot_t forbidden = __pgprot(0);
 242
 243         /*
 244          * The BIOS area between 640k and 1Mb needs to be executable for
 245          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
 246          */
 247         if (within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
 248                 pgprot_val(forbidden) |= _PAGE_NX;
 249
 250         /*
 251          * The kernel text needs to be executable for obvious reasons
 252          * Does not cover __inittext since that is gone later on. On
 253          * 64bit we do not enforce !NX on the low mapping
 254          */
 255         if (within(address, (unsigned long)_text, (unsigned long)_etext))
 256                 pgprot_val(forbidden) |= _PAGE_NX;
 257
 258         /*
 259          * The .rodata section needs to be read-only. Using the pfn
 260          * catches all aliases.
 261          */
 262         if (within(pfn, __pa((unsigned long)__start_rodata) >> PAGE_SHIFT,
 263                    __pa((unsigned long)__end_rodata) >> PAGE_SHIFT))
 264                 pgprot_val(forbidden) |= _PAGE_RW;
 265
 266         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
 267
 268         return prot;
 269 }
 270
 271 /*
 272  * Lookup the page table entry for a virtual address. Return a pointer
 273  * to the entry and the level of the mapping.
 274  *
 275  * Note: We return pud and pmd either when the entry is marked large
 276  * or when the present bit is not set. Otherwise we would return a
 277  * pointer to a nonexisting mapping.
 278  */
 279 pte_t *lookup_address(unsigned long address, unsigned int *level)
 280 {
 281         pgd_t *pgd = pgd_offset_k(address);
 282         pud_t *pud;
 283         pmd_t *pmd;
 284
 285         *level = PG_LEVEL_NONE;
 286
 287         if (pgd_none(*pgd))
 288                 return NULL;
 289
 290         pud = pud_offset(pgd, address);
 291         if (pud_none(*pud))
 292                 return NULL;
 293
 294         *level = PG_LEVEL_1G;
 295         if (pud_large(*pud) || !pud_present(*pud))
 296                 return (pte_t *)pud;
 297
 298         pmd = pmd_offset(pud, address);
 299         if (pmd_none(*pmd))
 300                 return NULL;
 301
 302         *level = PG_LEVEL_2M;
 303         if (pmd_large(*pmd) || !pmd_present(*pmd))
 304                 return (pte_t *)pmd;
 305
 306         *level = PG_LEVEL_4K;
 307
 308         return pte_offset_kernel(pmd, address);
 309 }
 310 EXPORT_SYMBOL_GPL(lookup_address);
 311
 312 /*
 313  * Set the new pmd in all the pgds we know about:
 314  */
 315 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 316 {
 317         /* change init_mm */
 318         set_pte_atomic(kpte, pte);
 319 #ifdef CONFIG_X86_32
 320         if (!SHARED_KERNEL_PMD) {
 321                 struct page *page;
 322
 323                 list_for_each_entry(page, &pgd_list, lru) {
 324                         pgd_t *pgd;
 325                         pud_t *pud;
 326                         pmd_t *pmd;
 327
 328                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 329                         pud = pud_offset(pgd, address);
 330                         pmd = pmd_offset(pud, address);
 331                         set_pte_atomic((pte_t *)pmd, pte);
 332                 }
 333         }
 334 #endif
 335 }
 336
 337 static int
 338 try_preserve_large_page(pte_t *kpte, unsigned long address,
 339                         struct cpa_data *cpa)
 340 {
 341         unsigned long nextpage_addr, numpages, pmask, psize, flags, addr, pfn;
 342         pte_t new_pte, old_pte, *tmp;
 343         pgprot_t old_prot, new_prot;
 344         int i, do_split = 1;
 345         unsigned int level;
 346
 347         if (cpa->force_split)
 348                 return 1;
 349
 350         spin_lock_irqsave(&pgd_lock, flags);
 351         /*
 352          * Check for races, another CPU might have split this page
 353          * up already:
 354          */
 355         tmp = lookup_address(address, &level);
 356         if (tmp != kpte)
 357                 goto out_unlock;
 358
 359         switch (level) {
 360         case PG_LEVEL_2M:
 361                 psize = PMD_PAGE_SIZE;
 362                 pmask = PMD_PAGE_MASK;
 363                 break;
 364 #ifdef CONFIG_X86_64
 365         case PG_LEVEL_1G:
 366                 psize = PUD_PAGE_SIZE;
 367                 pmask = PUD_PAGE_MASK;
 368                 break;
 369 #endif
 370         default:
 371                 do_split = -EINVAL;
 372                 goto out_unlock;
 373         }
 374
 375         /*
 376          * Calculate the number of pages, which fit into this large
 377          * page starting at address:
 378          */
 379         nextpage_addr = (address + psize) & pmask;
 380         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
 381         if (numpages < cpa->numpages)
 382                 cpa->numpages = numpages;
 383
 384         /*
 385          * We are safe now. Check whether the new pgprot is the same:
 386          */
 387         old_pte = *kpte;
 388         old_prot = new_prot = pte_pgprot(old_pte);
 389
 390         pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
 391         pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
 392
 393         /*
 394          * old_pte points to the large page base address. So we need
 395          * to add the offset of the virtual address:
 396          */
 397         pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
 398         cpa->pfn = pfn;
 399
 400         new_prot = static_protections(new_prot, address, pfn);
 401
 402         /*
 403          * We need to check the full range, whether
 404          * static_protection() requires a different pgprot for one of
 405          * the pages in the range we try to preserve:
 406          */
 407         addr = address + PAGE_SIZE;
 408         pfn++;
 409         for (i = 1; i < cpa->numpages; i++, addr += PAGE_SIZE, pfn++) {
 410                 pgprot_t chk_prot = static_protections(new_prot, addr, pfn);
 411
 412                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
 413                         goto out_unlock;
 414         }
 415
 416         /*
 417          * If there are no changes, return. maxpages has been updated
 418          * above:
 419          */
 420         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
 421                 do_split = 0;
 422                 goto out_unlock;
 423         }
 424
 425         /*
 426          * We need to change the attributes. Check, whether we can
 427          * change the large page in one go. We request a split, when
 428          * the address is not aligned and the number of pages is
 429          * smaller than the number of pages in the large page. Note
 430          * that we limited the number of possible pages already to
 431          * the number of pages in the large page.
 432          */
 433         if (address == (nextpage_addr - psize) && cpa->numpages == numpages) {
 434                 /*
 435                  * The address is aligned and the number of pages
 436                  * covers the full page.
 437                  */
 438                 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
 439                 __set_pmd_pte(kpte, address, new_pte);
 440                 cpa->flags |= CPA_FLUSHTLB;
 441                 do_split = 0;
 442         }
 443
 444 out_unlock:
 445         spin_unlock_irqrestore(&pgd_lock, flags);
 446
 447         return do_split;
 448 }
 449
 450 static LIST_HEAD(page_pool);
 451 static unsigned long pool_size, pool_pages, pool_low;
 452 static unsigned long pool_used, pool_failed;
 453
 454 static void cpa_fill_pool(struct page **ret)
 455 {
 456         gfp_t gfp = GFP_KERNEL;
 457         unsigned long flags;
 458         struct page *p;
 459
 460         /*
 461          * Avoid recursion (on debug-pagealloc) and also signal
 462          * our priority to get to these pagetables:
 463          */
 464         if (current->flags & PF_MEMALLOC)
 465                 return;
 466         current->flags |= PF_MEMALLOC;
 467
 468         /*
 469          * Allocate atomically from atomic contexts:
 470          */
 471         if (in_atomic() || irqs_disabled() || debug_pagealloc)
 472                 gfp =  GFP_ATOMIC | __GFP_NORETRY | __GFP_NOWARN;
 473
 474         while (pool_pages < pool_size || (ret && !*ret)) {
 475                 p = alloc_pages(gfp, 0);
 476                 if (!p) {
 477                         pool_failed++;
 478                         break;
 479                 }
 480                 /*
 481                  * If the call site needs a page right now, provide it:
 482                  */
 483                 if (ret && !*ret) {
 484                         *ret = p;
 485                         continue;
 486                 }
 487                 spin_lock_irqsave(&pgd_lock, flags);
 488                 list_add(&p->lru, &page_pool);
 489                 pool_pages++;
 490                 spin_unlock_irqrestore(&pgd_lock, flags);
 491         }
 492
 493         current->flags &= ~PF_MEMALLOC;
 494 }
 495
 496 #define SHIFT_MB                (20 - PAGE_SHIFT)
 497 #define ROUND_MB_GB             ((1 << 10) - 1)
 498 #define SHIFT_MB_GB             10
 499 #define POOL_PAGES_PER_GB       16
 500
 501 void __init cpa_init(void)
 502 {
 503         struct sysinfo si;
 504         unsigned long gb;
 505
 506         si_meminfo(&si);
 507         /*
 508          * Calculate the number of pool pages:
 509          *
 510          * Convert totalram (nr of pages) to MiB and round to the next
 511          * GiB. Shift MiB to Gib and multiply the result by
 512          * POOL_PAGES_PER_GB:
 513          */
 514         if (debug_pagealloc) {
 515                 gb = ((si.totalram >> SHIFT_MB) + ROUND_MB_GB) >> SHIFT_MB_GB;
 516                 pool_size = POOL_PAGES_PER_GB * gb;
 517         } else {
 518                 pool_size = 1;
 519         }
 520         pool_low = pool_size;
 521
 522         cpa_fill_pool(NULL);
 523         printk(KERN_DEBUG
 524                "CPA: page pool initialized %lu of %lu pages preallocated\n",
 525                pool_pages, pool_size);
 526 }
 527
 528 static int split_large_page(pte_t *kpte, unsigned long address)
 529 {
 530         unsigned long flags, pfn, pfninc = 1;
 531         unsigned int i, level;
 532         pte_t *pbase, *tmp;
 533         pgprot_t ref_prot;
 534         struct page *base;
 535
 536         /*
 537          * Get a page from the pool. The pool list is protected by the
 538          * pgd_lock, which we have to take anyway for the split
 539          * operation:
 540          */
 541         spin_lock_irqsave(&pgd_lock, flags);
 542         if (list_empty(&page_pool)) {
 543                 spin_unlock_irqrestore(&pgd_lock, flags);
 544                 base = NULL;
 545                 cpa_fill_pool(&base);
 546                 if (!base)
 547                         return -ENOMEM;
 548                 spin_lock_irqsave(&pgd_lock, flags);
 549         } else {
 550                 base = list_first_entry(&page_pool, struct page, lru);
 551                 list_del(&base->lru);
 552                 pool_pages--;
 553
 554                 if (pool_pages < pool_low)
 555                         pool_low = pool_pages;
 556         }
 557
 558         /*
 559          * Check for races, another CPU might have split this page
 560          * up for us already:
 561          */
 562         tmp = lookup_address(address, &level);
 563         if (tmp != kpte)
 564                 goto out_unlock;
 565
 566         pbase = (pte_t *)page_address(base);
 567         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
 568         ref_prot = pte_pgprot(pte_clrhuge(*kpte));
 569
 570 #ifdef CONFIG_X86_64
 571         if (level == PG_LEVEL_1G) {
 572                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
 573                 pgprot_val(ref_prot) |= _PAGE_PSE;
 574         }
 575 #endif
 576
 577         /*
 578          * Get the target pfn from the original entry:
 579          */
 580         pfn = pte_pfn(*kpte);
 581         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
 582                 set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
 583
 584         if (address >= (unsigned long)__va(0) &&
 585                 address < (unsigned long)__va(max_low_pfn_mapped << PAGE_SHIFT))
 586                 split_page_count(level);
 587
 588 #ifdef CONFIG_X86_64
 589         if (address >= (unsigned long)__va(1UL<<32) &&
 590                 address < (unsigned long)__va(max_pfn_mapped << PAGE_SHIFT))
 591                 split_page_count(level);
 592 #endif
 593
 594         /*
 595          * Install the new, split up pagetable. Important details here:
 596          *
 597          * On Intel the NX bit of all levels must be cleared to make a
 598          * page executable. See section 4.13.2 of Intel 64 and IA-32
 599          * Architectures Software Developer's Manual).
 600          *
 601          * Mark the entry present. The current mapping might be
 602          * set to not present, which we preserved above.
 603          */
 604         ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
 605         pgprot_val(ref_prot) |= _PAGE_PRESENT;
 606         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
 607         base = NULL;
 608
 609 out_unlock:
 610         /*
 611          * If we dropped out via the lookup_address check under
 612          * pgd_lock then stick the page back into the pool:
 613          */
 614         if (base) {
 615                 list_add(&base->lru, &page_pool);
 616                 pool_pages++;
 617         } else
 618                 pool_used++;
 619         spin_unlock_irqrestore(&pgd_lock, flags);
 620
 621         return 0;
 622 }
 623
 624 static int __change_page_attr(struct cpa_data *cpa, int primary)
 625 {
 626         unsigned long address;
 627         int do_split, err;
 628         unsigned int level;
 629         pte_t *kpte, old_pte;
 630
 631         if (cpa->flags & CPA_ARRAY)
 632                 address = cpa->vaddr[cpa->curpage];
 633         else
 634                 address = *cpa->vaddr;
 635
 636 repeat:
 637         kpte = lookup_address(address, &level);
 638         if (!kpte)
 639                 return 0;
 640
 641         old_pte = *kpte;
 642         if (!pte_val(old_pte)) {
 643                 if (!primary)
 644                         return 0;
 645                 WARN(1, KERN_WARNING "CPA: called for zero pte. "
 646                        "vaddr = %lx cpa->vaddr = %lx\n", address,
 647                        *cpa->vaddr);
 648                 return -EINVAL;
 649         }
 650
 651         if (level == PG_LEVEL_4K) {
 652                 pte_t new_pte;
 653                 pgprot_t new_prot = pte_pgprot(old_pte);
 654                 unsigned long pfn = pte_pfn(old_pte);
 655
 656                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
 657                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
 658
 659                 new_prot = static_protections(new_prot, address, pfn);
 660
 661                 /*
 662                  * We need to keep the pfn from the existing PTE,
 663                  * after all we're only going to change it's attributes
 664                  * not the memory it points to
 665                  */
 666                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
 667                 cpa->pfn = pfn;
 668                 /*
 669                  * Do we really change anything ?
 670                  */
 671                 if (pte_val(old_pte) != pte_val(new_pte)) {
 672                         set_pte_atomic(kpte, new_pte);
 673                         cpa->flags |= CPA_FLUSHTLB;
 674                 }
 675                 cpa->numpages = 1;
 676                 return 0;
 677         }
 678
 679         /*
 680          * Check, whether we can keep the large page intact
 681          * and just change the pte:
 682          */
 683         do_split = try_preserve_large_page(kpte, address, cpa);
 684         /*
 685          * When the range fits into the existing large page,
 686          * return. cp->numpages and cpa->tlbflush have been updated in
 687          * try_large_page:
 688          */
 689         if (do_split <= 0)
 690                 return do_split;
 691
 692         /*
 693          * We have to split the large page:
 694          */
 695         err = split_large_page(kpte, address);
 696         if (!err) {
 697                 cpa->flags |= CPA_FLUSHTLB;
 698                 goto repeat;
 699         }
 700
 701         return err;
 702 }
 703
 704 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
 705
 706 static int cpa_process_alias(struct cpa_data *cpa)
 707 {
 708         struct cpa_data alias_cpa;
 709         int ret = 0;
 710         unsigned long temp_cpa_vaddr, vaddr;
 711
 712         if (cpa->pfn >= max_pfn_mapped)
 713                 return 0;
 714
 715 #ifdef CONFIG_X86_64
 716         if (cpa->pfn >= max_low_pfn_mapped && cpa->pfn < (1UL<<(32-PAGE_SHIFT)))
 717                 return 0;
 718 #endif
 719         /*
 720          * No need to redo, when the primary call touched the direct
 721          * mapping already:
 722          */
 723         if (cpa->flags & CPA_ARRAY)
 724                 vaddr = cpa->vaddr[cpa->curpage];
 725         else
 726                 vaddr = *cpa->vaddr;
 727
 728         if (!(within(vaddr, PAGE_OFFSET,
 729                     PAGE_OFFSET + (max_low_pfn_mapped << PAGE_SHIFT))
 730 #ifdef CONFIG_X86_64
 731                 || within(vaddr, PAGE_OFFSET + (1UL<<32),
 732                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))
 733 #endif
 734         )) {
 735
 736                 alias_cpa = *cpa;
 737                 temp_cpa_vaddr = (unsigned long) __va(cpa->pfn << PAGE_SHIFT);
 738                 alias_cpa.vaddr = &temp_cpa_vaddr;
 739                 alias_cpa.flags &= ~CPA_ARRAY;
 740
 741
 742                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
 743         }
 744
 745 #ifdef CONFIG_X86_64
 746         if (ret)
 747                 return ret;
 748         /*
 749          * No need to redo, when the primary call touched the high
 750          * mapping already:
 751          */
 752         if (within(vaddr, (unsigned long) _text, (unsigned long) _end))
 753                 return 0;
 754
 755         /*
 756          * If the physical address is inside the kernel map, we need
 757          * to touch the high mapped kernel as well:
 758          */
 759         if (!within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn()))
 760                 return 0;
 761
 762         alias_cpa = *cpa;
 763         temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) + __START_KERNEL_map - phys_base;
 764         alias_cpa.vaddr = &temp_cpa_vaddr;
 765         alias_cpa.flags &= ~CPA_ARRAY;
 766
 767         /*
 768          * The high mapping range is imprecise, so ignore the return value.
 769          */
 770         __change_page_attr_set_clr(&alias_cpa, 0);
 771 #endif
 772         return ret;
 773 }
 774
 775 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
 776 {
 777         int ret, numpages = cpa->numpages;
 778
 779         while (numpages) {
 780                 /*
 781                  * Store the remaining nr of pages for the large page
 782                  * preservation check.
 783                  */
 784                 cpa->numpages = numpages;
 785                 /* for array changes, we can't use large page */
 786                 if (cpa->flags & CPA_ARRAY)
 787                         cpa->numpages = 1;
 788
 789                 ret = __change_page_attr(cpa, checkalias);
 790                 if (ret)
 791                         return ret;
 792
 793                 if (checkalias) {
 794                         ret = cpa_process_alias(cpa);
 795                         if (ret)
 796                                 return ret;
 797                 }
 798
 799                 /*
 800                  * Adjust the number of pages with the result of the
 801                  * CPA operation. Either a large page has been
 802                  * preserved or a single page update happened.
 803                  */
 804                 BUG_ON(cpa->numpages > numpages);
 805                 numpages -= cpa->numpages;
 806                 if (cpa->flags & CPA_ARRAY)
 807                         cpa->curpage++;
 808                 else
 809                         *cpa->vaddr += cpa->numpages * PAGE_SIZE;
 810
 811         }
 812         return 0;
 813 }
 814
 815 static inline int cache_attr(pgprot_t attr)
 816 {
 817         return pgprot_val(attr) &
 818                 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
 819 }
 820
 821 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
 822                                     pgprot_t mask_set, pgprot_t mask_clr,
 823                                     int force_split, int array)
 824 {
 825         struct cpa_data cpa;
 826         int ret, cache, checkalias;
 827
 828         /*
 829          * Check, if we are requested to change a not supported
 830          * feature:
 831          */
 832         mask_set = canon_pgprot(mask_set);
 833         mask_clr = canon_pgprot(mask_clr);
 834         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
 835                 return 0;
 836
 837         /* Ensure we are PAGE_SIZE aligned */
 838         if (!array) {
 839                 if (*addr & ~PAGE_MASK) {
 840                         *addr &= PAGE_MASK;
 841                         /*
 842                          * People should not be passing in unaligned addresses:
 843                          */
 844                         WARN_ON_ONCE(1);
 845                 }
 846         } else {
 847                 int i;
 848                 for (i = 0; i < numpages; i++) {
 849                         if (addr[i] & ~PAGE_MASK) {
 850                                 addr[i] &= PAGE_MASK;
 851                                 WARN_ON_ONCE(1);
 852                         }
 853                 }
 854         }
 855
 856         /* Must avoid aliasing mappings in the highmem code */
 857         kmap_flush_unused();
 858
 859         cpa.vaddr = addr;
 860         cpa.numpages = numpages;
 861         cpa.mask_set = mask_set;
 862         cpa.mask_clr = mask_clr;
 863         cpa.flags = 0;
 864         cpa.curpage = 0;
 865         cpa.force_split = force_split;
 866
 867         if (array)
 868                 cpa.flags |= CPA_ARRAY;
 869
 870         /* No alias checking for _NX bit modifications */
 871         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
 872
 873         ret = __change_page_attr_set_clr(&cpa, checkalias);
 874
 875         /*
 876          * Check whether we really changed something:
 877          */
 878         if (!(cpa.flags & CPA_FLUSHTLB))
 879                 goto out;
 880
 881         /*
 882          * No need to flush, when we did not set any of the caching
 883          * attributes:
 884          */
 885         cache = cache_attr(mask_set);
 886
 887         /*
 888          * On success we use clflush, when the CPU supports it to
 889          * avoid the wbindv. If the CPU does not support it and in the
 890          * error case we fall back to cpa_flush_all (which uses
 891          * wbindv):
 892          */
 893         if (!ret && cpu_has_clflush) {
 894                 if (cpa.flags & CPA_ARRAY)
 895                         cpa_flush_array(addr, numpages, cache);
 896                 else
 897                         cpa_flush_range(*addr, numpages, cache);
 898         } else
 899                 cpa_flush_all(cache);
 900
 901 out:
 902         cpa_fill_pool(NULL);
 903
 904         return ret;
 905 }
 906
 907 static inline int change_page_attr_set(unsigned long *addr, int numpages,
 908                                        pgprot_t mask, int array)
 909 {
 910         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
 911                 array);
 912 }
 913
 914 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
 915                                          pgprot_t mask, int array)
 916 {
 917         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
 918                 array);
 919 }
 920
 921 int _set_memory_uc(unsigned long addr, int numpages)
 922 {
 923         /*
 924          * for now UC MINUS. see comments in ioremap_nocache()
 925          */
 926         return change_page_attr_set(&addr, numpages,
 927                                     __pgprot(_PAGE_CACHE_UC_MINUS), 0);
 928 }
 929
 930 int set_memory_uc(unsigned long addr, int numpages)
 931 {
 932         /*
 933          * for now UC MINUS. see comments in ioremap_nocache()
 934          */
 935         if (reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
 936                             _PAGE_CACHE_UC_MINUS, NULL))
 937                 return -EINVAL;
 938
 939         return _set_memory_uc(addr, numpages);
 940 }
 941 EXPORT_SYMBOL(set_memory_uc);
 942
 943 int set_memory_array_uc(unsigned long *addr, int addrinarray)
 944 {
 945         unsigned long start;
 946         unsigned long end;
 947         int i;
 948         /*
 949          * for now UC MINUS. see comments in ioremap_nocache()
 950          */
 951         for (i = 0; i < addrinarray; i++) {
 952                 start = __pa(addr[i]);
 953                 for (end = start + PAGE_SIZE; i < addrinarray - 1; end += PAGE_SIZE) {
 954                         if (end != __pa(addr[i + 1]))
 955                                 break;
 956                         i++;
 957                 }
 958                 if (reserve_memtype(start, end, _PAGE_CACHE_UC_MINUS, NULL))
 959                         goto out;
 960         }
 961
 962         return change_page_attr_set(addr, addrinarray,
 963                                     __pgprot(_PAGE_CACHE_UC_MINUS), 1);
 964 out:
 965         for (i = 0; i < addrinarray; i++) {
 966                 unsigned long tmp = __pa(addr[i]);
 967
 968                 if (tmp == start)
 969                         break;
 970                 for (end = tmp + PAGE_SIZE; i < addrinarray - 1; end += PAGE_SIZE) {
 971                         if (end != __pa(addr[i + 1]))
 972                                 break;
 973                         i++;
 974                 }
 975                 free_memtype(tmp, end);
 976         }
 977         return -EINVAL;
 978 }
 979 EXPORT_SYMBOL(set_memory_array_uc);
 980
 981 int _set_memory_wc(unsigned long addr, int numpages)
 982 {
 983         return change_page_attr_set(&addr, numpages,
 984                                     __pgprot(_PAGE_CACHE_WC), 0);
 985 }
 986
 987 int set_memory_wc(unsigned long addr, int numpages)
 988 {
 989         if (!pat_enabled)
 990                 return set_memory_uc(addr, numpages);
 991
 992         if (reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
 993                 _PAGE_CACHE_WC, NULL))
 994                 return -EINVAL;
 995
 996         return _set_memory_wc(addr, numpages);
 997 }
 998 EXPORT_SYMBOL(set_memory_wc);
 999
1000 int _set_memory_wb(unsigned long addr, int numpages)
1001 {
1002         return change_page_attr_clear(&addr, numpages,
1003                                       __pgprot(_PAGE_CACHE_MASK), 0);
1004 }
1005
1006 int set_memory_wb(unsigned long addr, int numpages)
1007 {
1008         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1009
1010         return _set_memory_wb(addr, numpages);
1011 }
1012 EXPORT_SYMBOL(set_memory_wb);
1013
1014 int set_memory_array_wb(unsigned long *addr, int addrinarray)
1015 {
1016         int i;
1017
1018         for (i = 0; i < addrinarray; i++) {
1019                 unsigned long start = __pa(addr[i]);
1020                 unsigned long end;
1021
1022                 for (end = start + PAGE_SIZE; i < addrinarray - 1; end += PAGE_SIZE) {
1023                         if (end != __pa(addr[i + 1]))
1024                                 break;
1025                         i++;
1026                 }
1027                 free_memtype(start, end);
1028         }
1029         return change_page_attr_clear(addr, addrinarray,
1030                                       __pgprot(_PAGE_CACHE_MASK), 1);
1031 }
1032 EXPORT_SYMBOL(set_memory_array_wb);
1033
1034 int set_memory_x(unsigned long addr, int numpages)
1035 {
1036         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1037 }
1038 EXPORT_SYMBOL(set_memory_x);
1039
1040 int set_memory_nx(unsigned long addr, int numpages)
1041 {
1042         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1043 }
1044 EXPORT_SYMBOL(set_memory_nx);
1045
1046 int set_memory_ro(unsigned long addr, int numpages)
1047 {
1048         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1049 }
1050
1051 int set_memory_rw(unsigned long addr, int numpages)
1052 {
1053         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1054 }
1055
1056 int set_memory_np(unsigned long addr, int numpages)
1057 {
1058         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1059 }
1060
1061 int set_memory_4k(unsigned long addr, int numpages)
1062 {
1063         return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1064                                         __pgprot(0), 1, 0);
1065 }
1066
1067 int set_pages_uc(struct page *page, int numpages)
1068 {
1069         unsigned long addr = (unsigned long)page_address(page);
1070
1071         return set_memory_uc(addr, numpages);
1072 }
1073 EXPORT_SYMBOL(set_pages_uc);
1074
1075 int set_pages_wb(struct page *page, int numpages)
1076 {
1077         unsigned long addr = (unsigned long)page_address(page);
1078
1079         return set_memory_wb(addr, numpages);
1080 }
1081 EXPORT_SYMBOL(set_pages_wb);
1082
1083 int set_pages_x(struct page *page, int numpages)
1084 {
1085         unsigned long addr = (unsigned long)page_address(page);
1086
1087         return set_memory_x(addr, numpages);
1088 }
1089 EXPORT_SYMBOL(set_pages_x);
1090
1091 int set_pages_nx(struct page *page, int numpages)
1092 {
1093         unsigned long addr = (unsigned long)page_address(page);
1094
1095         return set_memory_nx(addr, numpages);
1096 }
1097 EXPORT_SYMBOL(set_pages_nx);
1098
1099 int set_pages_ro(struct page *page, int numpages)
1100 {
1101         unsigned long addr = (unsigned long)page_address(page);
1102
1103         return set_memory_ro(addr, numpages);
1104 }
1105
1106 int set_pages_rw(struct page *page, int numpages)
1107 {
1108         unsigned long addr = (unsigned long)page_address(page);
1109
1110         return set_memory_rw(addr, numpages);
1111 }
1112
1113 #ifdef CONFIG_DEBUG_PAGEALLOC
1114
1115 static int __set_pages_p(struct page *page, int numpages)
1116 {
1117         unsigned long tempaddr = (unsigned long) page_address(page);
1118         struct cpa_data cpa = { .vaddr = &tempaddr,
1119                                 .numpages = numpages,
1120                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1121                                 .mask_clr = __pgprot(0),
1122                                 .flags = 0};
1123
1124         return __change_page_attr_set_clr(&cpa, 1);
1125 }
1126
1127 static int __set_pages_np(struct page *page, int numpages)
1128 {
1129         unsigned long tempaddr = (unsigned long) page_address(page);
1130         struct cpa_data cpa = { .vaddr = &tempaddr,
1131                                 .numpages = numpages,
1132                                 .mask_set = __pgprot(0),
1133                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1134                                 .flags = 0};
1135
1136         return __change_page_attr_set_clr(&cpa, 1);
1137 }
1138
1139 void kernel_map_pages(struct page *page, int numpages, int enable)
1140 {
1141         if (PageHighMem(page))
1142                 return;
1143         if (!enable) {
1144                 debug_check_no_locks_freed(page_address(page),
1145                                            numpages * PAGE_SIZE);
1146         }
1147
1148         /*
1149          * If page allocator is not up yet then do not call c_p_a():
1150          */
1151         if (!debug_pagealloc_enabled)
1152                 return;
1153
1154         /*
1155          * The return value is ignored as the calls cannot fail.
1156          * Large pages are kept enabled at boot time, and are
1157          * split up quickly with DEBUG_PAGEALLOC. If a splitup
1158          * fails here (due to temporary memory shortage) no damage
1159          * is done because we just keep the largepage intact up
1160          * to the next attempt when it will likely be split up:
1161          */
1162         if (enable)
1163                 __set_pages_p(page, numpages);
1164         else
1165                 __set_pages_np(page, numpages);
1166
1167         /*
1168          * We should perform an IPI and flush all tlbs,
1169          * but that can deadlock->flush only current cpu:
1170          */
1171         __flush_tlb_all();
1172
1173         /*
1174          * Try to refill the page pool here. We can do this only after
1175          * the tlb flush.
1176          */
1177         cpa_fill_pool(NULL);
1178 }
1179
1180 #ifdef CONFIG_DEBUG_FS
1181 static int dpa_show(struct seq_file *m, void *v)
1182 {
1183         seq_puts(m, "DEBUG_PAGEALLOC\n");
1184         seq_printf(m, "pool_size     : %lu\n", pool_size);
1185         seq_printf(m, "pool_pages    : %lu\n", pool_pages);
1186         seq_printf(m, "pool_low      : %lu\n", pool_low);
1187         seq_printf(m, "pool_used     : %lu\n", pool_used);
1188         seq_printf(m, "pool_failed   : %lu\n", pool_failed);
1189
1190         return 0;
1191 }
1192
1193 static int dpa_open(struct inode *inode, struct file *filp)
1194 {
1195         return single_open(filp, dpa_show, NULL);
1196 }
1197
1198 static const struct file_operations dpa_fops = {
1199         .open           = dpa_open,
1200         .read           = seq_read,
1201         .llseek         = seq_lseek,
1202         .release        = single_release,
1203 };
1204
1205 static int __init debug_pagealloc_proc_init(void)
1206 {
1207         struct dentry *de;
1208
1209         de = debugfs_create_file("debug_pagealloc", 0600, NULL, NULL,
1210                                  &dpa_fops);
1211         if (!de)
1212                 return -ENOMEM;
1213
1214         return 0;
1215 }
1216 __initcall(debug_pagealloc_proc_init);
1217 #endif
1218
1219 #ifdef CONFIG_HIBERNATION
1220
1221 bool kernel_page_present(struct page *page)
1222 {
1223         unsigned int level;
1224         pte_t *pte;
1225
1226         if (PageHighMem(page))
1227                 return false;
1228
1229         pte = lookup_address((unsigned long)page_address(page), &level);
1230         return (pte_val(*pte) & _PAGE_PRESENT);
1231 }
1232
1233 #endif /* CONFIG_HIBERNATION */
1234
1235 #endif /* CONFIG_DEBUG_PAGEALLOC */
1236
1237 /*
1238  * The testcases use internal knowledge of the implementation that shouldn't
1239  * be exposed to the rest of the kernel. Include these directly here.
1240  */
1241 #ifdef CONFIG_CPA_DEBUG
1242 #include "pageattr-test.c"
1243 #endif