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
  29 /**
  30  * clflush_cache_range - flush a cache range with clflush
  31  * @addr:       virtual start address
  32  * @size:       number of bytes to flush
  33  *
  34  * clflush is an unordered instruction which needs fencing with mfence
  35  * to avoid ordering issues.
  36  */
  37 void clflush_cache_range(void *vaddr, unsigned int size)
  38 {
  39         void *vend = vaddr + size - 1;
  40
  41         mb();
  42
  43         for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
  44                 clflush(vaddr);
  45         /*
  46          * Flush any possible final partial cacheline:
  47          */
  48         clflush(vend);
  49
  50         mb();
  51 }
  52
  53 static void __cpa_flush_all(void *arg)
  54 {
  55         /*
  56          * Flush all to work around Errata in early athlons regarding
  57          * large page flushing.
  58          */
  59         __flush_tlb_all();
  60
  61         if (boot_cpu_data.x86_model >= 4)
  62                 wbinvd();
  63 }
  64
  65 static void cpa_flush_all(void)
  66 {
  67         BUG_ON(irqs_disabled());
  68
  69         on_each_cpu(__cpa_flush_all, NULL, 1, 1);
  70 }
  71
  72 static void __cpa_flush_range(void *arg)
  73 {
  74         /*
  75          * We could optimize that further and do individual per page
  76          * tlb invalidates for a low number of pages. Caveat: we must
  77          * flush the high aliases on 64bit as well.
  78          */
  79         __flush_tlb_all();
  80 }
  81
  82 static void cpa_flush_range(unsigned long start, int numpages)
  83 {
  84         unsigned int i, level;
  85         unsigned long addr;
  86
  87         BUG_ON(irqs_disabled());
  88         WARN_ON(PAGE_ALIGN(start) != start);
  89
  90         on_each_cpu(__cpa_flush_range, NULL, 1, 1);
  91
  92         /*
  93          * We only need to flush on one CPU,
  94          * clflush is a MESI-coherent instruction that
  95          * will cause all other CPUs to flush the same
  96          * cachelines:
  97          */
  98         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
  99                 pte_t *pte = lookup_address(addr, &level);
 100
 101                 /*
 102                  * Only flush present addresses:
 103                  */
 104                 if (pte && pte_present(*pte))
 105                         clflush_cache_range((void *) addr, PAGE_SIZE);
 106         }
 107 }
 108
 109 #define HIGH_MAP_START  __START_KERNEL_map
 110 #define HIGH_MAP_END    (__START_KERNEL_map + KERNEL_TEXT_SIZE)
 111
 112
 113 /*
 114  * Converts a virtual address to a X86-64 highmap address
 115  */
 116 static unsigned long virt_to_highmap(void *address)
 117 {
 118 #ifdef CONFIG_X86_64
 119         return __pa((unsigned long)address) + HIGH_MAP_START - phys_base;
 120 #else
 121         return (unsigned long)address;
 122 #endif
 123 }
 124
 125 /*
 126  * Certain areas of memory on x86 require very specific protection flags,
 127  * for example the BIOS area or kernel text. Callers don't always get this
 128  * right (again, ioremap() on BIOS memory is not uncommon) so this function
 129  * checks and fixes these known static required protection bits.
 130  */
 131 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address)
 132 {
 133         pgprot_t forbidden = __pgprot(0);
 134
 135         /*
 136          * The BIOS area between 640k and 1Mb needs to be executable for
 137          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
 138          */
 139         if (within(__pa(address), BIOS_BEGIN, BIOS_END))
 140                 pgprot_val(forbidden) |= _PAGE_NX;
 141
 142         /*
 143          * The kernel text needs to be executable for obvious reasons
 144          * Does not cover __inittext since that is gone later on
 145          */
 146         if (within(address, (unsigned long)_text, (unsigned long)_etext))
 147                 pgprot_val(forbidden) |= _PAGE_NX;
 148         /*
 149          * Do the same for the x86-64 high kernel mapping
 150          */
 151         if (within(address, virt_to_highmap(_text), virt_to_highmap(_etext)))
 152                 pgprot_val(forbidden) |= _PAGE_NX;
 153
 154
 155 #ifdef CONFIG_DEBUG_RODATA
 156         /* The .rodata section needs to be read-only */
 157         if (within(address, (unsigned long)__start_rodata,
 158                                 (unsigned long)__end_rodata))
 159                 pgprot_val(forbidden) |= _PAGE_RW;
 160         /*
 161          * Do the same for the x86-64 high kernel mapping
 162          */
 163         if (within(address, virt_to_highmap(__start_rodata),
 164                                 virt_to_highmap(__end_rodata)))
 165                 pgprot_val(forbidden) |= _PAGE_RW;
 166 #endif
 167
 168         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
 169
 170         return prot;
 171 }
 172
 173 pte_t *lookup_address(unsigned long address, int *level)
 174 {
 175         pgd_t *pgd = pgd_offset_k(address);
 176         pud_t *pud;
 177         pmd_t *pmd;
 178
 179         *level = PG_LEVEL_NONE;
 180
 181         if (pgd_none(*pgd))
 182                 return NULL;
 183         pud = pud_offset(pgd, address);
 184         if (pud_none(*pud))
 185                 return NULL;
 186         pmd = pmd_offset(pud, address);
 187         if (pmd_none(*pmd))
 188                 return NULL;
 189
 190         *level = PG_LEVEL_2M;
 191         if (pmd_large(*pmd))
 192                 return (pte_t *)pmd;
 193
 194         *level = PG_LEVEL_4K;
 195         return pte_offset_kernel(pmd, address);
 196 }
 197
 198 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 199 {
 200         /* change init_mm */
 201         set_pte_atomic(kpte, pte);
 202 #ifdef CONFIG_X86_32
 203         if (!SHARED_KERNEL_PMD) {
 204                 struct page *page;
 205
 206                 list_for_each_entry(page, &pgd_list, lru) {
 207                         pgd_t *pgd;
 208                         pud_t *pud;
 209                         pmd_t *pmd;
 210
 211                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 212                         pud = pud_offset(pgd, address);
 213                         pmd = pmd_offset(pud, address);
 214                         set_pte_atomic((pte_t *)pmd, pte);
 215                 }
 216         }
 217 #endif
 218 }
 219
 220 static int split_large_page(pte_t *kpte, unsigned long address)
 221 {
 222         pgprot_t ref_prot = pte_pgprot(pte_clrhuge(*kpte));
 223         gfp_t gfp_flags = GFP_KERNEL;
 224         unsigned long flags;
 225         unsigned long addr;
 226         pte_t *pbase, *tmp;
 227         struct page *base;
 228         unsigned int i, level;
 229
 230 #ifdef CONFIG_DEBUG_PAGEALLOC
 231         gfp_flags = __GFP_HIGH | __GFP_NOFAIL | __GFP_NOWARN;
 232         gfp_flags = GFP_ATOMIC | __GFP_NOWARN;
 233 #endif
 234         base = alloc_pages(gfp_flags, 0);
 235         if (!base)
 236                 return -ENOMEM;
 237
 238         spin_lock_irqsave(&pgd_lock, flags);
 239         /*
 240          * Check for races, another CPU might have split this page
 241          * up for us already:
 242          */
 243         tmp = lookup_address(address, &level);
 244         if (tmp != kpte) {
 245                 WARN_ON_ONCE(1);
 246                 goto out_unlock;
 247         }
 248
 249         address = __pa(address);
 250         addr = address & LARGE_PAGE_MASK;
 251         pbase = (pte_t *)page_address(base);
 252 #ifdef CONFIG_X86_32
 253         paravirt_alloc_pt(&init_mm, page_to_pfn(base));
 254 #endif
 255
 256         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
 257                 set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT, ref_prot));
 258
 259         /*
 260          * Install the new, split up pagetable. Important detail here:
 261          *
 262          * On Intel the NX bit of all levels must be cleared to make a
 263          * page executable. See section 4.13.2 of Intel 64 and IA-32
 264          * Architectures Software Developer's Manual).
 265          */
 266         ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
 267         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
 268         base = NULL;
 269
 270 out_unlock:
 271         spin_unlock_irqrestore(&pgd_lock, flags);
 272
 273         if (base)
 274                 __free_pages(base, 0);
 275
 276         return 0;
 277 }
 278
 279 static int
 280 __change_page_attr(unsigned long address, pgprot_t mask_set, pgprot_t mask_clr)
 281 {
 282         struct page *kpte_page;
 283         int level, err = 0;
 284         pte_t *kpte;
 285
 286 repeat:
 287         kpte = lookup_address(address, &level);
 288         if (!kpte)
 289                 return -EINVAL;
 290
 291         kpte_page = virt_to_page(kpte);
 292         BUG_ON(PageLRU(kpte_page));
 293         BUG_ON(PageCompound(kpte_page));
 294
 295         if (level == PG_LEVEL_4K) {
 296                 pte_t new_pte, old_pte = *kpte;
 297                 pgprot_t new_prot = pte_pgprot(old_pte);
 298
 299                 if(!pte_val(old_pte)) {
 300                         WARN_ON_ONCE(1);
 301                         return -EINVAL;
 302                 }
 303
 304                 pgprot_val(new_prot) &= ~pgprot_val(mask_clr);
 305                 pgprot_val(new_prot) |= pgprot_val(mask_set);
 306
 307                 new_prot = static_protections(new_prot, address);
 308
 309                 /*
 310                  * We need to keep the pfn from the existing PTE,
 311                  * after all we're only going to change it's attributes
 312                  * not the memory it points to
 313                  */
 314                 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
 315                 set_pte_atomic(kpte, new_pte);
 316         } else {
 317                 err = split_large_page(kpte, address);
 318                 if (!err)
 319                         goto repeat;
 320         }
 321         return err;
 322 }
 323
 324 /**
 325  * change_page_attr_addr - Change page table attributes in linear mapping
 326  * @address: Virtual address in linear mapping.
 327  * @prot:    New page table attribute (PAGE_*)
 328  *
 329  * Change page attributes of a page in the direct mapping. This is a variant
 330  * of change_page_attr() that also works on memory holes that do not have
 331  * mem_map entry (pfn_valid() is false).
 332  *
 333  * See change_page_attr() documentation for more details.
 334  *
 335  * Modules and drivers should use the set_memory_* APIs instead.
 336  */
 337
 338
 339 static int
 340 change_page_attr_addr(unsigned long address, pgprot_t mask_set,
 341                       pgprot_t mask_clr)
 342 {
 343         int err;
 344
 345 #ifdef CONFIG_X86_64
 346         unsigned long phys_addr = __pa(address);
 347
 348         /*
 349          * If we are inside the high mapped kernel range, then we
 350          * fixup the low mapping first. __va() returns the virtual
 351          * address in the linear mapping:
 352          */
 353         if (within(address, HIGH_MAP_START, HIGH_MAP_END))
 354                 address = (unsigned long) __va(phys_addr);
 355 #endif
 356
 357         err = __change_page_attr(address, mask_set, mask_clr);
 358         if (err)
 359                 return err;
 360
 361 #ifdef CONFIG_X86_64
 362         /*
 363          * If the physical address is inside the kernel map, we need
 364          * to touch the high mapped kernel as well:
 365          */
 366         if (within(phys_addr, 0, KERNEL_TEXT_SIZE)) {
 367                 /*
 368                  * Calc the high mapping address. See __phys_addr()
 369                  * for the non obvious details.
 370                  *
 371                  * Note that NX and other required permissions are
 372                  * checked in static_protections().
 373                  */
 374                 address = phys_addr + HIGH_MAP_START - phys_base;
 375
 376                 /*
 377                  * Our high aliases are imprecise, because we check
 378                  * everything between 0 and KERNEL_TEXT_SIZE, so do
 379                  * not propagate lookup failures back to users:
 380                  */
 381                 __change_page_attr(address, mask_set, mask_clr);
 382         }
 383 #endif
 384         return err;
 385 }
 386
 387 static int __change_page_attr_set_clr(unsigned long addr, int numpages,
 388                                       pgprot_t mask_set, pgprot_t mask_clr)
 389 {
 390         unsigned int i;
 391         int ret;
 392
 393         for (i = 0; i < numpages ; i++, addr += PAGE_SIZE) {
 394                 ret = change_page_attr_addr(addr, mask_set, mask_clr);
 395                 if (ret)
 396                         return ret;
 397         }
 398
 399         return 0;
 400 }
 401
 402 static int change_page_attr_set_clr(unsigned long addr, int numpages,
 403                                     pgprot_t mask_set, pgprot_t mask_clr)
 404 {
 405         int ret = __change_page_attr_set_clr(addr, numpages, mask_set,
 406                                              mask_clr);
 407
 408         /*
 409          * On success we use clflush, when the CPU supports it to
 410          * avoid the wbindv. If the CPU does not support it and in the
 411          * error case we fall back to cpa_flush_all (which uses
 412          * wbindv):
 413          */
 414         if (!ret && cpu_has_clflush)
 415                 cpa_flush_range(addr, numpages);
 416         else
 417                 cpa_flush_all();
 418
 419         return ret;
 420 }
 421
 422 static inline int change_page_attr_set(unsigned long addr, int numpages,
 423                                        pgprot_t mask)
 424 {
 425         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0));
 426 }
 427
 428 static inline int change_page_attr_clear(unsigned long addr, int numpages,
 429                                          pgprot_t mask)
 430 {
 431         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask);
 432 }
 433
 434 int set_memory_uc(unsigned long addr, int numpages)
 435 {
 436         return change_page_attr_set(addr, numpages,
 437                                     __pgprot(_PAGE_PCD | _PAGE_PWT));
 438 }
 439 EXPORT_SYMBOL(set_memory_uc);
 440
 441 int set_memory_wb(unsigned long addr, int numpages)
 442 {
 443         return change_page_attr_clear(addr, numpages,
 444                                       __pgprot(_PAGE_PCD | _PAGE_PWT));
 445 }
 446 EXPORT_SYMBOL(set_memory_wb);
 447
 448 int set_memory_x(unsigned long addr, int numpages)
 449 {
 450         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
 451 }
 452 EXPORT_SYMBOL(set_memory_x);
 453
 454 int set_memory_nx(unsigned long addr, int numpages)
 455 {
 456         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
 457 }
 458 EXPORT_SYMBOL(set_memory_nx);
 459
 460 int set_memory_ro(unsigned long addr, int numpages)
 461 {
 462         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
 463 }
 464
 465 int set_memory_rw(unsigned long addr, int numpages)
 466 {
 467         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
 468 }
 469
 470 int set_memory_np(unsigned long addr, int numpages)
 471 {
 472         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
 473 }
 474
 475 int set_pages_uc(struct page *page, int numpages)
 476 {
 477         unsigned long addr = (unsigned long)page_address(page);
 478
 479         return set_memory_uc(addr, numpages);
 480 }
 481 EXPORT_SYMBOL(set_pages_uc);
 482
 483 int set_pages_wb(struct page *page, int numpages)
 484 {
 485         unsigned long addr = (unsigned long)page_address(page);
 486
 487         return set_memory_wb(addr, numpages);
 488 }
 489 EXPORT_SYMBOL(set_pages_wb);
 490
 491 int set_pages_x(struct page *page, int numpages)
 492 {
 493         unsigned long addr = (unsigned long)page_address(page);
 494
 495         return set_memory_x(addr, numpages);
 496 }
 497 EXPORT_SYMBOL(set_pages_x);
 498
 499 int set_pages_nx(struct page *page, int numpages)
 500 {
 501         unsigned long addr = (unsigned long)page_address(page);
 502
 503         return set_memory_nx(addr, numpages);
 504 }
 505 EXPORT_SYMBOL(set_pages_nx);
 506
 507 int set_pages_ro(struct page *page, int numpages)
 508 {
 509         unsigned long addr = (unsigned long)page_address(page);
 510
 511         return set_memory_ro(addr, numpages);
 512 }
 513
 514 int set_pages_rw(struct page *page, int numpages)
 515 {
 516         unsigned long addr = (unsigned long)page_address(page);
 517
 518         return set_memory_rw(addr, numpages);
 519 }
 520
 521
 522 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_CPA_DEBUG)
 523 static inline int __change_page_attr_set(unsigned long addr, int numpages,
 524                                          pgprot_t mask)
 525 {
 526         return __change_page_attr_set_clr(addr, numpages, mask, __pgprot(0));
 527 }
 528
 529 static inline int __change_page_attr_clear(unsigned long addr, int numpages,
 530                                            pgprot_t mask)
 531 {
 532         return __change_page_attr_set_clr(addr, numpages, __pgprot(0), mask);
 533 }
 534 #endif
 535
 536 #ifdef CONFIG_DEBUG_PAGEALLOC
 537
 538 static int __set_pages_p(struct page *page, int numpages)
 539 {
 540         unsigned long addr = (unsigned long)page_address(page);
 541
 542         return __change_page_attr_set(addr, numpages,
 543                                       __pgprot(_PAGE_PRESENT | _PAGE_RW));
 544 }
 545
 546 static int __set_pages_np(struct page *page, int numpages)
 547 {
 548         unsigned long addr = (unsigned long)page_address(page);
 549
 550         return __change_page_attr_clear(addr, numpages,
 551                                         __pgprot(_PAGE_PRESENT));
 552 }
 553
 554 void kernel_map_pages(struct page *page, int numpages, int enable)
 555 {
 556         if (PageHighMem(page))
 557                 return;
 558         if (!enable) {
 559                 debug_check_no_locks_freed(page_address(page),
 560                                            numpages * PAGE_SIZE);
 561         }
 562
 563         /*
 564          * If page allocator is not up yet then do not call c_p_a():
 565          */
 566         if (!debug_pagealloc_enabled)
 567                 return;
 568
 569         /*
 570          * The return value is ignored - the calls cannot fail,
 571          * large pages are disabled at boot time:
 572          */
 573         if (enable)
 574                 __set_pages_p(page, numpages);
 575         else
 576                 __set_pages_np(page, numpages);
 577
 578         /*
 579          * We should perform an IPI and flush all tlbs,
 580          * but that can deadlock->flush only current cpu:
 581          */
 582         __flush_tlb_all();
 583 }
 584 #endif
 585
 586 /*
 587  * The testcases use internal knowledge of the implementation that shouldn't
 588  * be exposed to the rest of the kernel. Include these directly here.
 589  */
 590 #ifdef CONFIG_CPA_DEBUG
 591 #include "pageattr-test.c"
 592 #endif