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