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                 set_pte_atomic(kpte, pfn_pte(pfn, canon_pgprot(prot)));
 220         } else {
 221                 err = split_large_page(kpte, address);
 222                 if (!err)
 223                         goto repeat;
 224         }
 225         return err;
 226 }
 227
 228 /**
 229  * change_page_attr_addr - Change page table attributes in linear mapping
 230  * @address: Virtual address in linear mapping.
 231  * @prot:    New page table attribute (PAGE_*)
 232  *
 233  * Change page attributes of a page in the direct mapping. This is a variant
 234  * of change_page_attr() that also works on memory holes that do not have
 235  * mem_map entry (pfn_valid() is false).
 236  *
 237  * See change_page_attr() documentation for more details.
 238  *
 239  * Modules and drivers should use the set_memory_* APIs instead.
 240  */
 241
 242 static int change_page_attr_addr(unsigned long address, pgprot_t prot)
 243 {
 244         int err = 0, kernel_map = 0;
 245         unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 246
 247 #ifdef CONFIG_X86_64
 248         if (address >= __START_KERNEL_map &&
 249                         address < __START_KERNEL_map + KERNEL_TEXT_SIZE) {
 250
 251                 address = (unsigned long)__va(__pa(address));
 252                 kernel_map = 1;
 253         }
 254 #endif
 255
 256         if (!kernel_map || pte_present(pfn_pte(0, prot))) {
 257                 err = __change_page_attr(address, pfn, prot);
 258                 if (err)
 259                         return err;
 260         }
 261
 262 #ifdef CONFIG_X86_64
 263         /*
 264          * Handle kernel mapping too which aliases part of
 265          * lowmem:
 266          */
 267         if (__pa(address) < KERNEL_TEXT_SIZE) {
 268                 unsigned long addr2;
 269                 pgprot_t prot2;
 270
 271                 addr2 = __START_KERNEL_map + __pa(address);
 272                 /* Make sure the kernel mappings stay executable */
 273                 prot2 = pte_pgprot(pte_mkexec(pfn_pte(0, prot)));
 274                 err = __change_page_attr(addr2, pfn, prot2);
 275         }
 276 #endif
 277
 278         return err;
 279 }
 280
 281 /**
 282  * change_page_attr_set - Change page table attributes in the linear mapping.
 283  * @addr: Virtual address in linear mapping.
 284  * @numpages: Number of pages to change
 285  * @prot: Protection/caching type bits to set (PAGE_*)
 286  *
 287  * Returns 0 on success, otherwise a negated errno.
 288  *
 289  * This should be used when a page is mapped with a different caching policy
 290  * than write-back somewhere - some CPUs do not like it when mappings with
 291  * different caching policies exist. This changes the page attributes of the
 292  * in kernel linear mapping too.
 293  *
 294  * The caller needs to ensure that there are no conflicting mappings elsewhere
 295  * (e.g. in user space) * This function only deals with the kernel linear map.
 296  *
 297  * This function is different from change_page_attr() in that only selected bits
 298  * are impacted, all other bits remain as is.
 299  */
 300 static int change_page_attr_set(unsigned long addr, int numpages,
 301                                                                 pgprot_t prot)
 302 {
 303         pgprot_t current_prot;
 304         int level;
 305         pte_t *pte;
 306         int i, ret;
 307
 308         for (i = 0; i < numpages ; i++) {
 309
 310                 pte = lookup_address(addr, &level);
 311                 if (pte)
 312                         current_prot = pte_pgprot(*pte);
 313                 else
 314                         pgprot_val(current_prot) = 0;
 315
 316                 pgprot_val(prot) = pgprot_val(current_prot) | pgprot_val(prot);
 317
 318                 ret = change_page_attr_addr(addr, prot);
 319                 if (ret)
 320                         return ret;
 321                 addr += PAGE_SIZE;
 322         }
 323         return 0;
 324 }
 325
 326 /**
 327  * change_page_attr_clear - Change page table attributes in the linear mapping.
 328  * @addr: Virtual address in linear mapping.
 329  * @numpages: Number of pages to change
 330  * @prot: Protection/caching type bits to clear (PAGE_*)
 331  *
 332  * Returns 0 on success, otherwise a negated errno.
 333  *
 334  * This should be used when a page is mapped with a different caching policy
 335  * than write-back somewhere - some CPUs do not like it when mappings with
 336  * different caching policies exist. This changes the page attributes of the
 337  * in kernel linear mapping too.
 338  *
 339  * The caller needs to ensure that there are no conflicting mappings elsewhere
 340  * (e.g. in user space) * This function only deals with the kernel linear map.
 341  *
 342  * This function is different from change_page_attr() in that only selected bits
 343  * are impacted, all other bits remain as is.
 344  */
 345 static int change_page_attr_clear(unsigned long addr, int numpages,
 346                                                                 pgprot_t prot)
 347 {
 348         pgprot_t current_prot;
 349         int level;
 350         pte_t *pte;
 351         int i, ret;
 352
 353         for (i = 0; i < numpages; i++) {
 354                 pte = lookup_address(addr, &level);
 355                 if (pte)
 356                         current_prot = pte_pgprot(*pte);
 357                 else
 358                         pgprot_val(current_prot) = 0;
 359
 360                 pgprot_val(prot) =
 361                                 pgprot_val(current_prot) & ~pgprot_val(prot);
 362
 363                 ret = change_page_attr_addr(addr, prot);
 364                 if (ret)
 365                         return ret;
 366                 addr += PAGE_SIZE;
 367         }
 368         return 0;
 369 }
 370
 371 int set_memory_uc(unsigned long addr, int numpages)
 372 {
 373         int err;
 374
 375         err = change_page_attr_set(addr, numpages,
 376                                 __pgprot(_PAGE_PCD | _PAGE_PWT));
 377         global_flush_tlb();
 378         return err;
 379 }
 380 EXPORT_SYMBOL(set_memory_uc);
 381
 382 int set_memory_wb(unsigned long addr, int numpages)
 383 {
 384         int err;
 385
 386         err = change_page_attr_clear(addr, numpages,
 387                                 __pgprot(_PAGE_PCD | _PAGE_PWT));
 388         global_flush_tlb();
 389         return err;
 390 }
 391 EXPORT_SYMBOL(set_memory_wb);
 392
 393 int set_memory_x(unsigned long addr, int numpages)
 394 {
 395         int err;
 396
 397         err = change_page_attr_clear(addr, numpages,
 398                                 __pgprot(_PAGE_NX));
 399         global_flush_tlb();
 400         return err;
 401 }
 402 EXPORT_SYMBOL(set_memory_x);
 403
 404 int set_memory_nx(unsigned long addr, int numpages)
 405 {
 406         int err;
 407
 408         err = change_page_attr_set(addr, numpages,
 409                                 __pgprot(_PAGE_NX));
 410         global_flush_tlb();
 411         return err;
 412 }
 413 EXPORT_SYMBOL(set_memory_nx);
 414
 415 int set_memory_ro(unsigned long addr, int numpages)
 416 {
 417         int err;
 418
 419         err = change_page_attr_clear(addr, numpages,
 420                                 __pgprot(_PAGE_RW));
 421         global_flush_tlb();
 422         return err;
 423 }
 424
 425 int set_memory_rw(unsigned long addr, int numpages)
 426 {
 427         int err;
 428
 429         err = change_page_attr_set(addr, numpages,
 430                                 __pgprot(_PAGE_RW));
 431         global_flush_tlb();
 432         return err;
 433 }
 434
 435 int set_memory_np(unsigned long addr, int numpages)
 436 {
 437         int err;
 438
 439         err = change_page_attr_clear(addr, numpages,
 440                                 __pgprot(_PAGE_PRESENT));
 441         global_flush_tlb();
 442         return err;
 443 }
 444
 445 int set_pages_uc(struct page *page, int numpages)
 446 {
 447         unsigned long addr = (unsigned long)page_address(page);
 448
 449         return set_memory_uc(addr, numpages);
 450 }
 451 EXPORT_SYMBOL(set_pages_uc);
 452
 453 int set_pages_wb(struct page *page, int numpages)
 454 {
 455         unsigned long addr = (unsigned long)page_address(page);
 456
 457         return set_memory_wb(addr, numpages);
 458 }
 459 EXPORT_SYMBOL(set_pages_wb);
 460
 461 int set_pages_x(struct page *page, int numpages)
 462 {
 463         unsigned long addr = (unsigned long)page_address(page);
 464
 465         return set_memory_x(addr, numpages);
 466 }
 467 EXPORT_SYMBOL(set_pages_x);
 468
 469 int set_pages_nx(struct page *page, int numpages)
 470 {
 471         unsigned long addr = (unsigned long)page_address(page);
 472
 473         return set_memory_nx(addr, numpages);
 474 }
 475 EXPORT_SYMBOL(set_pages_nx);
 476
 477 int set_pages_ro(struct page *page, int numpages)
 478 {
 479         unsigned long addr = (unsigned long)page_address(page);
 480
 481         return set_memory_ro(addr, numpages);
 482 }
 483
 484 int set_pages_rw(struct page *page, int numpages)
 485 {
 486         unsigned long addr = (unsigned long)page_address(page);
 487
 488         return set_memory_rw(addr, numpages);
 489 }
 490
 491
 492 #ifdef CONFIG_DEBUG_PAGEALLOC
 493
 494 static int __set_pages_p(struct page *page, int numpages)
 495 {
 496         unsigned long addr = (unsigned long)page_address(page);
 497         return change_page_attr_set(addr, numpages,
 498                                 __pgprot(_PAGE_PRESENT | _PAGE_RW));
 499 }
 500
 501 static int __set_pages_np(struct page *page, int numpages)
 502 {
 503         unsigned long addr = (unsigned long)page_address(page);
 504         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
 505 }
 506
 507 void kernel_map_pages(struct page *page, int numpages, int enable)
 508 {
 509         if (PageHighMem(page))
 510                 return;
 511         if (!enable) {
 512                 debug_check_no_locks_freed(page_address(page),
 513                                            numpages * PAGE_SIZE);
 514         }
 515
 516         /*
 517          * If page allocator is not up yet then do not call c_p_a():
 518          */
 519         if (!debug_pagealloc_enabled)
 520                 return;
 521
 522         /*
 523          * The return value is ignored - the calls cannot fail,
 524          * large pages are disabled at boot time:
 525          */
 526         if (enable)
 527                 __set_pages_p(page, numpages);
 528         else
 529                 __set_pages_np(page, numpages);
 530
 531         /*
 532          * We should perform an IPI and flush all tlbs,
 533          * but that can deadlock->flush only current cpu:
 534          */
 535         __flush_tlb_all();
 536 }
 537 #endif
 538
 539 /*
 540  * The testcases use internal knowledge of the implementation that shouldn't
 541  * be exposed to the rest of the kernel. Include these directly here.
 542  */
 543 #ifdef CONFIG_CPA_DEBUG
 544 #include "pageattr-test.c"
 545 #endif