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 /*
 110  * Certain areas of memory on x86 require very specific protection flags,
 111  * for example the BIOS area or kernel text. Callers don't always get this
 112  * right (again, ioremap() on BIOS memory is not uncommon) so this function
 113  * checks and fixes these known static required protection bits.
 114  */
 115 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address)
 116 {
 117         pgprot_t forbidden = __pgprot(0);
 118
 119         /*
 120          * The BIOS area between 640k and 1Mb needs to be executable for
 121          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
 122          */
 123         if (within(__pa(address), BIOS_BEGIN, BIOS_END))
 124                 pgprot_val(forbidden) |= _PAGE_NX;
 125
 126         /*
 127          * The kernel text needs to be executable for obvious reasons
 128          * Does not cover __inittext since that is gone later on
 129          */
 130         if (within(address, (unsigned long)_text, (unsigned long)_etext))
 131                 pgprot_val(forbidden) |= _PAGE_NX;
 132
 133 #ifdef CONFIG_DEBUG_RODATA
 134         /* The .rodata section needs to be read-only */
 135         if (within(address, (unsigned long)__start_rodata,
 136                                 (unsigned long)__end_rodata))
 137                 pgprot_val(forbidden) |= _PAGE_RW;
 138 #endif
 139
 140         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
 141
 142         return prot;
 143 }
 144
 145 pte_t *lookup_address(unsigned long address, int *level)
 146 {
 147         pgd_t *pgd = pgd_offset_k(address);
 148         pud_t *pud;
 149         pmd_t *pmd;
 150
 151         *level = PG_LEVEL_NONE;
 152
 153         if (pgd_none(*pgd))
 154                 return NULL;
 155         pud = pud_offset(pgd, address);
 156         if (pud_none(*pud))
 157                 return NULL;
 158         pmd = pmd_offset(pud, address);
 159         if (pmd_none(*pmd))
 160                 return NULL;
 161
 162         *level = PG_LEVEL_2M;
 163         if (pmd_large(*pmd))
 164                 return (pte_t *)pmd;
 165
 166         *level = PG_LEVEL_4K;
 167         return pte_offset_kernel(pmd, address);
 168 }
 169
 170 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 171 {
 172         /* change init_mm */
 173         set_pte_atomic(kpte, pte);
 174 #ifdef CONFIG_X86_32
 175         if (!SHARED_KERNEL_PMD) {
 176                 struct page *page;
 177
 178                 for (page = pgd_list; page; page = (struct page *)page->index) {
 179                         pgd_t *pgd;
 180                         pud_t *pud;
 181                         pmd_t *pmd;
 182
 183                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 184                         pud = pud_offset(pgd, address);
 185                         pmd = pmd_offset(pud, address);
 186                         set_pte_atomic((pte_t *)pmd, pte);
 187                 }
 188         }
 189 #endif
 190 }
 191
 192 static int split_large_page(pte_t *kpte, unsigned long address)
 193 {
 194         pgprot_t ref_prot = pte_pgprot(pte_clrhuge(*kpte));
 195         gfp_t gfp_flags = GFP_KERNEL;
 196         unsigned long flags;
 197         unsigned long addr;
 198         pte_t *pbase, *tmp;
 199         struct page *base;
 200         unsigned int i, level;
 201
 202 #ifdef CONFIG_DEBUG_PAGEALLOC
 203         gfp_flags = __GFP_HIGH | __GFP_NOFAIL | __GFP_NOWARN;
 204         gfp_flags = GFP_ATOMIC | __GFP_NOWARN;
 205 #endif
 206         base = alloc_pages(gfp_flags, 0);
 207         if (!base)
 208                 return -ENOMEM;
 209
 210         spin_lock_irqsave(&pgd_lock, flags);
 211         /*
 212          * Check for races, another CPU might have split this page
 213          * up for us already:
 214          */
 215         tmp = lookup_address(address, &level);
 216         if (tmp != kpte) {
 217                 WARN_ON_ONCE(1);
 218                 goto out_unlock;
 219         }
 220
 221         address = __pa(address);
 222         addr = address & LARGE_PAGE_MASK;
 223         pbase = (pte_t *)page_address(base);
 224 #ifdef CONFIG_X86_32
 225         paravirt_alloc_pt(&init_mm, page_to_pfn(base));
 226 #endif
 227
 228         pgprot_val(ref_prot) &= ~_PAGE_NX;
 229         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
 230                 set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT, ref_prot));
 231
 232         /*
 233          * Install the new, split up pagetable. Important detail here:
 234          *
 235          * On Intel the NX bit of all levels must be cleared to make a
 236          * page executable. See section 4.13.2 of Intel 64 and IA-32
 237          * Architectures Software Developer's Manual).
 238          */
 239         ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
 240         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
 241         base = NULL;
 242
 243 out_unlock:
 244         spin_unlock_irqrestore(&pgd_lock, flags);
 245
 246         if (base)
 247                 __free_pages(base, 0);
 248
 249         return 0;
 250 }
 251
 252 static int
 253 __change_page_attr(unsigned long address, unsigned long pfn,
 254                    pgprot_t mask_set, pgprot_t mask_clr)
 255 {
 256         struct page *kpte_page;
 257         int level, err = 0;
 258         pte_t *kpte;
 259
 260 #ifdef CONFIG_X86_32
 261         BUG_ON(pfn > max_low_pfn);
 262 #endif
 263
 264 repeat:
 265         kpte = lookup_address(address, &level);
 266         if (!kpte)
 267                 return -EINVAL;
 268
 269         kpte_page = virt_to_page(kpte);
 270         BUG_ON(PageLRU(kpte_page));
 271         BUG_ON(PageCompound(kpte_page));
 272
 273         if (level == PG_LEVEL_4K) {
 274                 pgprot_t new_prot = pte_pgprot(*kpte);
 275                 pte_t new_pte, old_pte = *kpte;
 276
 277                 pgprot_val(new_prot) &= ~pgprot_val(mask_clr);
 278                 pgprot_val(new_prot) |= pgprot_val(mask_set);
 279
 280                 new_prot = static_protections(new_prot, address);
 281
 282                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
 283                 BUG_ON(pte_pfn(new_pte) != pte_pfn(old_pte));
 284
 285                 set_pte_atomic(kpte, new_pte);
 286         } else {
 287                 err = split_large_page(kpte, address);
 288                 if (!err)
 289                         goto repeat;
 290         }
 291         return err;
 292 }
 293
 294 /**
 295  * change_page_attr_addr - Change page table attributes in linear mapping
 296  * @address: Virtual address in linear mapping.
 297  * @prot:    New page table attribute (PAGE_*)
 298  *
 299  * Change page attributes of a page in the direct mapping. This is a variant
 300  * of change_page_attr() that also works on memory holes that do not have
 301  * mem_map entry (pfn_valid() is false).
 302  *
 303  * See change_page_attr() documentation for more details.
 304  *
 305  * Modules and drivers should use the set_memory_* APIs instead.
 306  */
 307
 308 static int
 309 change_page_attr_addr(unsigned long address, pgprot_t mask_set,
 310                                                         pgprot_t mask_clr)
 311 {
 312         int err = 0, kernel_map = 0;
 313         unsigned long pfn;
 314
 315 #ifdef CONFIG_X86_64
 316         if (address >= __START_KERNEL_map &&
 317                         address < __START_KERNEL_map + KERNEL_TEXT_SIZE) {
 318
 319                 address = (unsigned long)__va(__pa((void *)address));
 320                 kernel_map = 1;
 321         }
 322 #endif
 323
 324         pfn = __pa(address) >> PAGE_SHIFT;
 325
 326         if (!kernel_map || 1) {
 327                 err = __change_page_attr(address, pfn, mask_set, mask_clr);
 328                 if (err)
 329                         return err;
 330         }
 331
 332 #ifdef CONFIG_X86_64
 333         /*
 334          * Handle kernel mapping too which aliases part of
 335          * lowmem:
 336          */
 337         if (__pa(address) < KERNEL_TEXT_SIZE) {
 338                 unsigned long addr2;
 339
 340                 addr2 = __pa(address) + __START_KERNEL_map - phys_base;
 341                 /* Make sure the kernel mappings stay executable */
 342                 pgprot_val(mask_clr) |= _PAGE_NX;
 343                 /*
 344                  * Our high aliases are imprecise, so do not propagate
 345                  * failures back to users:
 346                  */
 347                 __change_page_attr(addr2, pfn, mask_set, mask_clr);
 348         }
 349 #endif
 350
 351         return err;
 352 }
 353
 354 static int __change_page_attr_set_clr(unsigned long addr, int numpages,
 355                                       pgprot_t mask_set, pgprot_t mask_clr)
 356 {
 357         unsigned int i;
 358         int ret;
 359
 360         for (i = 0; i < numpages ; i++, addr += PAGE_SIZE) {
 361                 ret = change_page_attr_addr(addr, mask_set, mask_clr);
 362                 if (ret)
 363                         return ret;
 364         }
 365
 366         return 0;
 367 }
 368
 369 static int change_page_attr_set_clr(unsigned long addr, int numpages,
 370                                     pgprot_t mask_set, pgprot_t mask_clr)
 371 {
 372         int ret = __change_page_attr_set_clr(addr, numpages, mask_set,
 373                                              mask_clr);
 374
 375         /*
 376          * On success we use clflush, when the CPU supports it to
 377          * avoid the wbindv. If the CPU does not support it and in the
 378          * error case we fall back to cpa_flush_all (which uses
 379          * wbindv):
 380          */
 381         if (!ret && cpu_has_clflush)
 382                 cpa_flush_range(addr, numpages);
 383         else
 384                 cpa_flush_all();
 385
 386         return ret;
 387 }
 388
 389 static inline int change_page_attr_set(unsigned long addr, int numpages,
 390                                        pgprot_t mask)
 391 {
 392         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0));
 393 }
 394
 395 static inline int change_page_attr_clear(unsigned long addr, int numpages,
 396                                          pgprot_t mask)
 397 {
 398         return __change_page_attr_set_clr(addr, numpages, __pgprot(0), mask);
 399
 400 }
 401
 402 int set_memory_uc(unsigned long addr, int numpages)
 403 {
 404         return change_page_attr_set(addr, numpages,
 405                                     __pgprot(_PAGE_PCD | _PAGE_PWT));
 406 }
 407 EXPORT_SYMBOL(set_memory_uc);
 408
 409 int set_memory_wb(unsigned long addr, int numpages)
 410 {
 411         return change_page_attr_clear(addr, numpages,
 412                                       __pgprot(_PAGE_PCD | _PAGE_PWT));
 413 }
 414 EXPORT_SYMBOL(set_memory_wb);
 415
 416 int set_memory_x(unsigned long addr, int numpages)
 417 {
 418         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
 419 }
 420 EXPORT_SYMBOL(set_memory_x);
 421
 422 int set_memory_nx(unsigned long addr, int numpages)
 423 {
 424         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
 425 }
 426 EXPORT_SYMBOL(set_memory_nx);
 427
 428 int set_memory_ro(unsigned long addr, int numpages)
 429 {
 430         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
 431 }
 432
 433 int set_memory_rw(unsigned long addr, int numpages)
 434 {
 435         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
 436 }
 437
 438 int set_memory_np(unsigned long addr, int numpages)
 439 {
 440         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
 441 }
 442
 443 int set_pages_uc(struct page *page, int numpages)
 444 {
 445         unsigned long addr = (unsigned long)page_address(page);
 446
 447         return set_memory_uc(addr, numpages);
 448 }
 449 EXPORT_SYMBOL(set_pages_uc);
 450
 451 int set_pages_wb(struct page *page, int numpages)
 452 {
 453         unsigned long addr = (unsigned long)page_address(page);
 454
 455         return set_memory_wb(addr, numpages);
 456 }
 457 EXPORT_SYMBOL(set_pages_wb);
 458
 459 int set_pages_x(struct page *page, int numpages)
 460 {
 461         unsigned long addr = (unsigned long)page_address(page);
 462
 463         return set_memory_x(addr, numpages);
 464 }
 465 EXPORT_SYMBOL(set_pages_x);
 466
 467 int set_pages_nx(struct page *page, int numpages)
 468 {
 469         unsigned long addr = (unsigned long)page_address(page);
 470
 471         return set_memory_nx(addr, numpages);
 472 }
 473 EXPORT_SYMBOL(set_pages_nx);
 474
 475 int set_pages_ro(struct page *page, int numpages)
 476 {
 477         unsigned long addr = (unsigned long)page_address(page);
 478
 479         return set_memory_ro(addr, numpages);
 480 }
 481
 482 int set_pages_rw(struct page *page, int numpages)
 483 {
 484         unsigned long addr = (unsigned long)page_address(page);
 485
 486         return set_memory_rw(addr, numpages);
 487 }
 488
 489
 490 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_CPA_DEBUG)
 491 static inline int __change_page_attr_set(unsigned long addr, int numpages,
 492                                          pgprot_t mask)
 493 {
 494         return __change_page_attr_set_clr(addr, numpages, mask, __pgprot(0));
 495 }
 496
 497 static inline int __change_page_attr_clear(unsigned long addr, int numpages,
 498                                            pgprot_t mask)
 499 {
 500         return __change_page_attr_set_clr(addr, numpages, __pgprot(0), mask);
 501 }
 502 #endif
 503
 504 #ifdef CONFIG_DEBUG_PAGEALLOC
 505
 506 static int __set_pages_p(struct page *page, int numpages)
 507 {
 508         unsigned long addr = (unsigned long)page_address(page);
 509
 510         return __change_page_attr_set(addr, numpages,
 511                                       __pgprot(_PAGE_PRESENT | _PAGE_RW));
 512 }
 513
 514 static int __set_pages_np(struct page *page, int numpages)
 515 {
 516         unsigned long addr = (unsigned long)page_address(page);
 517
 518         return __change_page_attr_clear(addr, numpages,
 519                                         __pgprot(_PAGE_PRESENT));
 520 }
 521
 522 void kernel_map_pages(struct page *page, int numpages, int enable)
 523 {
 524         if (PageHighMem(page))
 525                 return;
 526         if (!enable) {
 527                 debug_check_no_locks_freed(page_address(page),
 528                                            numpages * PAGE_SIZE);
 529         }
 530
 531         /*
 532          * If page allocator is not up yet then do not call c_p_a():
 533          */
 534         if (!debug_pagealloc_enabled)
 535                 return;
 536
 537         /*
 538          * The return value is ignored - the calls cannot fail,
 539          * large pages are disabled at boot time:
 540          */
 541         if (enable)
 542                 __set_pages_p(page, numpages);
 543         else
 544                 __set_pages_np(page, numpages);
 545
 546         /*
 547          * We should perform an IPI and flush all tlbs,
 548          * but that can deadlock->flush only current cpu:
 549          */
 550         __flush_tlb_all();
 551 }
 552 #endif
 553
 554 /*
 555  * The testcases use internal knowledge of the implementation that shouldn't
 556  * be exposed to the rest of the kernel. Include these directly here.
 557  */
 558 #ifdef CONFIG_CPA_DEBUG
 559 #include "pageattr-test.c"
 560 #endif