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
  12 #include <asm/e820.h>
  13 #include <asm/processor.h>
  14 #include <asm/tlbflush.h>
  15 #include <asm/sections.h>
  16 #include <asm/uaccess.h>
  17 #include <asm/pgalloc.h>
  18
  19 static inline int
  20 within(unsigned long addr, unsigned long start, unsigned long end)
  21 {
  22         return addr >= start && addr < end;
  23 }
  24
  25 /*
  26  * Flushing functions
  27  */
  28 void clflush_cache_range(void *addr, int size)
  29 {
  30         int i;
  31
  32         for (i = 0; i < size; i += boot_cpu_data.x86_clflush_size)
  33                 clflush(addr+i);
  34 }
  35
  36 static void flush_kernel_map(void *arg)
  37 {
  38         /*
  39          * Flush all to work around Errata in early athlons regarding
  40          * large page flushing.
  41          */
  42         __flush_tlb_all();
  43
  44         if (boot_cpu_data.x86_model >= 4)
  45                 wbinvd();
  46 }
  47
  48 static void global_flush_tlb(void)
  49 {
  50         BUG_ON(irqs_disabled());
  51
  52         on_each_cpu(flush_kernel_map, NULL, 1, 1);
  53 }
  54
  55 /*
  56  * Certain areas of memory on x86 require very specific protection flags,
  57  * for example the BIOS area or kernel text. Callers don't always get this
  58  * right (again, ioremap() on BIOS memory is not uncommon) so this function
  59  * checks and fixes these known static required protection bits.
  60  */
  61 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address)
  62 {
  63         pgprot_t forbidden = __pgprot(0);
  64
  65         /*
  66          * The BIOS area between 640k and 1Mb needs to be executable for
  67          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
  68          */
  69         if (within(__pa(address), BIOS_BEGIN, BIOS_END))
  70                 pgprot_val(forbidden) |= _PAGE_NX;
  71
  72         /*
  73          * The kernel text needs to be executable for obvious reasons
  74          * Does not cover __inittext since that is gone later on
  75          */
  76         if (within(address, (unsigned long)_text, (unsigned long)_etext))
  77                 pgprot_val(forbidden) |= _PAGE_NX;
  78
  79 #ifdef CONFIG_DEBUG_RODATA
  80         /* The .rodata section needs to be read-only */
  81         if (within(address, (unsigned long)__start_rodata,
  82                                 (unsigned long)__end_rodata))
  83                 pgprot_val(forbidden) |= _PAGE_RW;
  84 #endif
  85
  86         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
  87
  88         return prot;
  89 }
  90
  91 pte_t *lookup_address(unsigned long address, int *level)
  92 {
  93         pgd_t *pgd = pgd_offset_k(address);
  94         pud_t *pud;
  95         pmd_t *pmd;
  96
  97         *level = PG_LEVEL_NONE;
  98
  99         if (pgd_none(*pgd))
 100                 return NULL;
 101         pud = pud_offset(pgd, address);
 102         if (pud_none(*pud))
 103                 return NULL;
 104         pmd = pmd_offset(pud, address);
 105         if (pmd_none(*pmd))
 106                 return NULL;
 107
 108         *level = PG_LEVEL_2M;
 109         if (pmd_large(*pmd))
 110                 return (pte_t *)pmd;
 111
 112         *level = PG_LEVEL_4K;
 113         return pte_offset_kernel(pmd, address);
 114 }
 115
 116 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 117 {
 118         /* change init_mm */
 119         set_pte_atomic(kpte, pte);
 120 #ifdef CONFIG_X86_32
 121         if (!SHARED_KERNEL_PMD) {
 122                 struct page *page;
 123
 124                 for (page = pgd_list; page; page = (struct page *)page->index) {
 125                         pgd_t *pgd;
 126                         pud_t *pud;
 127                         pmd_t *pmd;
 128
 129                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 130                         pud = pud_offset(pgd, address);
 131                         pmd = pmd_offset(pud, address);
 132                         set_pte_atomic((pte_t *)pmd, pte);
 133                 }
 134         }
 135 #endif
 136 }
 137
 138 static int split_large_page(pte_t *kpte, unsigned long address)
 139 {
 140         pgprot_t ref_prot = pte_pgprot(pte_clrhuge(*kpte));
 141         gfp_t gfp_flags = GFP_KERNEL;
 142         unsigned long flags;
 143         unsigned long addr;
 144         pte_t *pbase, *tmp;
 145         struct page *base;
 146         int i, level;
 147
 148 #ifdef CONFIG_DEBUG_PAGEALLOC
 149         gfp_flags = GFP_ATOMIC;
 150 #endif
 151         base = alloc_pages(gfp_flags, 0);
 152         if (!base)
 153                 return -ENOMEM;
 154
 155         spin_lock_irqsave(&pgd_lock, flags);
 156         /*
 157          * Check for races, another CPU might have split this page
 158          * up for us already:
 159          */
 160         tmp = lookup_address(address, &level);
 161         if (tmp != kpte) {
 162                 WARN_ON_ONCE(1);
 163                 goto out_unlock;
 164         }
 165
 166         address = __pa(address);
 167         addr = address & LARGE_PAGE_MASK;
 168         pbase = (pte_t *)page_address(base);
 169 #ifdef CONFIG_X86_32
 170         paravirt_alloc_pt(&init_mm, page_to_pfn(base));
 171 #endif
 172
 173         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
 174                 set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT, ref_prot));
 175
 176         /*
 177          * Install the new, split up pagetable. Important detail here:
 178          *
 179          * On Intel the NX bit of all levels must be cleared to make a
 180          * page executable. See section 4.13.2 of Intel 64 and IA-32
 181          * Architectures Software Developer's Manual).
 182          */
 183         ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
 184         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
 185         base = NULL;
 186
 187 out_unlock:
 188         spin_unlock_irqrestore(&pgd_lock, flags);
 189
 190         if (base)
 191                 __free_pages(base, 0);
 192
 193         return 0;
 194 }
 195
 196 static int
 197 __change_page_attr(unsigned long address, unsigned long pfn, pgprot_t prot)
 198 {
 199         struct page *kpte_page;
 200         int level, err = 0;
 201         pte_t *kpte;
 202
 203 #ifdef CONFIG_X86_32
 204         BUG_ON(pfn > max_low_pfn);
 205 #endif
 206
 207 repeat:
 208         kpte = lookup_address(address, &level);
 209         if (!kpte)
 210                 return -EINVAL;
 211
 212         kpte_page = virt_to_page(kpte);
 213         BUG_ON(PageLRU(kpte_page));
 214         BUG_ON(PageCompound(kpte_page));
 215
 216         prot = static_protections(prot, address);
 217
 218         if (level == PG_LEVEL_4K) {
 219                 WARN_ON_ONCE(pgprot_val(prot) & _PAGE_PSE);
 220                 set_pte_atomic(kpte, pfn_pte(pfn, canon_pgprot(prot)));
 221         } else {
 222                 /* Clear the PSE bit for the 4k level pages ! */
 223                 pgprot_val(prot) = pgprot_val(prot) & ~_PAGE_PSE;
 224
 225                 err = split_large_page(kpte, address);
 226                 if (!err)
 227                         goto repeat;
 228         }
 229         return err;
 230 }
 231
 232 /**
 233  * change_page_attr_addr - Change page table attributes in linear mapping
 234  * @address: Virtual address in linear mapping.
 235  * @prot:    New page table attribute (PAGE_*)
 236  *
 237  * Change page attributes of a page in the direct mapping. This is a variant
 238  * of change_page_attr() that also works on memory holes that do not have
 239  * mem_map entry (pfn_valid() is false).
 240  *
 241  * See change_page_attr() documentation for more details.
 242  *
 243  * Modules and drivers should use the set_memory_* APIs instead.
 244  */
 245
 246 static int change_page_attr_addr(unsigned long address, pgprot_t prot)
 247 {
 248         int err = 0, kernel_map = 0;
 249         unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 250
 251 #ifdef CONFIG_X86_64
 252         if (address >= __START_KERNEL_map &&
 253                         address < __START_KERNEL_map + KERNEL_TEXT_SIZE) {
 254
 255                 address = (unsigned long)__va(__pa(address));
 256                 kernel_map = 1;
 257         }
 258 #endif
 259
 260         if (!kernel_map || pte_present(pfn_pte(0, prot))) {
 261                 err = __change_page_attr(address, pfn, prot);
 262                 if (err)
 263                         return err;
 264         }
 265
 266 #ifdef CONFIG_X86_64
 267         /*
 268          * Handle kernel mapping too which aliases part of
 269          * lowmem:
 270          */
 271         if (__pa(address) < KERNEL_TEXT_SIZE) {
 272                 unsigned long addr2;
 273                 pgprot_t prot2;
 274
 275                 addr2 = __START_KERNEL_map + __pa(address);
 276                 /* Make sure the kernel mappings stay executable */
 277                 prot2 = pte_pgprot(pte_mkexec(pfn_pte(0, prot)));
 278                 err = __change_page_attr(addr2, pfn, prot2);
 279         }
 280 #endif
 281
 282         return err;
 283 }
 284
 285 /**
 286  * change_page_attr_set - Change page table attributes in the linear mapping.
 287  * @addr: Virtual address in linear mapping.
 288  * @numpages: Number of pages to change
 289  * @prot: Protection/caching type bits to set (PAGE_*)
 290  *
 291  * Returns 0 on success, otherwise a negated errno.
 292  *
 293  * This should be used when a page is mapped with a different caching policy
 294  * than write-back somewhere - some CPUs do not like it when mappings with
 295  * different caching policies exist. This changes the page attributes of the
 296  * in kernel linear mapping too.
 297  *
 298  * The caller needs to ensure that there are no conflicting mappings elsewhere
 299  * (e.g. in user space) * This function only deals with the kernel linear map.
 300  *
 301  * This function is different from change_page_attr() in that only selected bits
 302  * are impacted, all other bits remain as is.
 303  */
 304 static int change_page_attr_set(unsigned long addr, int numpages,
 305                                                                 pgprot_t prot)
 306 {
 307         pgprot_t current_prot, new_prot;
 308         int level;
 309         pte_t *pte;
 310         int i, ret;
 311
 312         for (i = 0; i < numpages ; i++) {
 313
 314                 pte = lookup_address(addr, &level);
 315                 if (!pte)
 316                         return -EINVAL;
 317
 318                 current_prot = pte_pgprot(*pte);
 319
 320                 pgprot_val(new_prot) =
 321                         pgprot_val(current_prot) | pgprot_val(prot);
 322
 323                 ret = change_page_attr_addr(addr, new_prot);
 324                 if (ret)
 325                         return ret;
 326                 addr += PAGE_SIZE;
 327         }
 328         return 0;
 329 }
 330
 331 /**
 332  * change_page_attr_clear - Change page table attributes in the linear mapping.
 333  * @addr: Virtual address in linear mapping.
 334  * @numpages: Number of pages to change
 335  * @prot: Protection/caching type bits to clear (PAGE_*)
 336  *
 337  * Returns 0 on success, otherwise a negated errno.
 338  *
 339  * This should be used when a page is mapped with a different caching policy
 340  * than write-back somewhere - some CPUs do not like it when mappings with
 341  * different caching policies exist. This changes the page attributes of the
 342  * in kernel linear mapping too.
 343  *
 344  * The caller needs to ensure that there are no conflicting mappings elsewhere
 345  * (e.g. in user space) * This function only deals with the kernel linear map.
 346  *
 347  * This function is different from change_page_attr() in that only selected bits
 348  * are impacted, all other bits remain as is.
 349  */
 350 static int change_page_attr_clear(unsigned long addr, int numpages,
 351                                                                 pgprot_t prot)
 352 {
 353         pgprot_t current_prot, new_prot;
 354         int level;
 355         pte_t *pte;
 356         int i, ret;
 357
 358         for (i = 0; i < numpages; i++) {
 359
 360                 pte = lookup_address(addr, &level);
 361                 if (!pte)
 362                         return -EINVAL;
 363
 364                 current_prot = pte_pgprot(*pte);
 365
 366                 pgprot_val(new_prot) =
 367                                 pgprot_val(current_prot) & ~pgprot_val(prot);
 368
 369                 ret = change_page_attr_addr(addr, new_prot);
 370                 if (ret)
 371                         return ret;
 372                 addr += PAGE_SIZE;
 373         }
 374         return 0;
 375 }
 376
 377 int set_memory_uc(unsigned long addr, int numpages)
 378 {
 379         int err;
 380
 381         err = change_page_attr_set(addr, numpages,
 382                                 __pgprot(_PAGE_PCD | _PAGE_PWT));
 383         global_flush_tlb();
 384         return err;
 385 }
 386 EXPORT_SYMBOL(set_memory_uc);
 387
 388 int set_memory_wb(unsigned long addr, int numpages)
 389 {
 390         int err;
 391
 392         err = change_page_attr_clear(addr, numpages,
 393                                 __pgprot(_PAGE_PCD | _PAGE_PWT));
 394         global_flush_tlb();
 395         return err;
 396 }
 397 EXPORT_SYMBOL(set_memory_wb);
 398
 399 int set_memory_x(unsigned long addr, int numpages)
 400 {
 401         int err;
 402
 403         err = change_page_attr_clear(addr, numpages,
 404                                 __pgprot(_PAGE_NX));
 405         global_flush_tlb();
 406         return err;
 407 }
 408 EXPORT_SYMBOL(set_memory_x);
 409
 410 int set_memory_nx(unsigned long addr, int numpages)
 411 {
 412         int err;
 413
 414         err = change_page_attr_set(addr, numpages,
 415                                 __pgprot(_PAGE_NX));
 416         global_flush_tlb();
 417         return err;
 418 }
 419 EXPORT_SYMBOL(set_memory_nx);
 420
 421 int set_memory_ro(unsigned long addr, int numpages)
 422 {
 423         int err;
 424
 425         err = change_page_attr_clear(addr, numpages,
 426                                 __pgprot(_PAGE_RW));
 427         global_flush_tlb();
 428         return err;
 429 }
 430
 431 int set_memory_rw(unsigned long addr, int numpages)
 432 {
 433         int err;
 434
 435         err = change_page_attr_set(addr, numpages,
 436                                 __pgprot(_PAGE_RW));
 437         global_flush_tlb();
 438         return err;
 439 }
 440
 441 int set_memory_np(unsigned long addr, int numpages)
 442 {
 443         int err;
 444
 445         err = change_page_attr_clear(addr, numpages,
 446                                 __pgprot(_PAGE_PRESENT));
 447         global_flush_tlb();
 448         return err;
 449 }
 450
 451 int set_pages_uc(struct page *page, int numpages)
 452 {
 453         unsigned long addr = (unsigned long)page_address(page);
 454
 455         return set_memory_uc(addr, numpages);
 456 }
 457 EXPORT_SYMBOL(set_pages_uc);
 458
 459 int set_pages_wb(struct page *page, int numpages)
 460 {
 461         unsigned long addr = (unsigned long)page_address(page);
 462
 463         return set_memory_wb(addr, numpages);
 464 }
 465 EXPORT_SYMBOL(set_pages_wb);
 466
 467 int set_pages_x(struct page *page, int numpages)
 468 {
 469         unsigned long addr = (unsigned long)page_address(page);
 470
 471         return set_memory_x(addr, numpages);
 472 }
 473 EXPORT_SYMBOL(set_pages_x);
 474
 475 int set_pages_nx(struct page *page, int numpages)
 476 {
 477         unsigned long addr = (unsigned long)page_address(page);
 478
 479         return set_memory_nx(addr, numpages);
 480 }
 481 EXPORT_SYMBOL(set_pages_nx);
 482
 483 int set_pages_ro(struct page *page, int numpages)
 484 {
 485         unsigned long addr = (unsigned long)page_address(page);
 486
 487         return set_memory_ro(addr, numpages);
 488 }
 489
 490 int set_pages_rw(struct page *page, int numpages)
 491 {
 492         unsigned long addr = (unsigned long)page_address(page);
 493
 494         return set_memory_rw(addr, numpages);
 495 }
 496
 497
 498 #ifdef CONFIG_DEBUG_PAGEALLOC
 499
 500 static int __set_pages_p(struct page *page, int numpages)
 501 {
 502         unsigned long addr = (unsigned long)page_address(page);
 503         return change_page_attr_set(addr, numpages,
 504                                 __pgprot(_PAGE_PRESENT | _PAGE_RW));
 505 }
 506
 507 static int __set_pages_np(struct page *page, int numpages)
 508 {
 509         unsigned long addr = (unsigned long)page_address(page);
 510         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
 511 }
 512
 513 void kernel_map_pages(struct page *page, int numpages, int enable)
 514 {
 515         if (PageHighMem(page))
 516                 return;
 517         if (!enable) {
 518                 debug_check_no_locks_freed(page_address(page),
 519                                            numpages * PAGE_SIZE);
 520         }
 521
 522         /*
 523          * If page allocator is not up yet then do not call c_p_a():
 524          */
 525         if (!debug_pagealloc_enabled)
 526                 return;
 527
 528         /*
 529          * The return value is ignored - the calls cannot fail,
 530          * large pages are disabled at boot time:
 531          */
 532         if (enable)
 533                 __set_pages_p(page, numpages);
 534         else
 535                 __set_pages_np(page, numpages);
 536
 537         /*
 538          * We should perform an IPI and flush all tlbs,
 539          * but that can deadlock->flush only current cpu:
 540          */
 541         __flush_tlb_all();
 542 }
 543 #endif
 544
 545 /*
 546  * The testcases use internal knowledge of the implementation that shouldn't
 547  * be exposed to the rest of the kernel. Include these directly here.
 548  */
 549 #ifdef CONFIG_CPA_DEBUG
 550 #include "pageattr-test.c"
 551 #endif