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                 list_for_each_entry(page, &pgd_list, lru) {
 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         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
 229                 set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT, ref_prot));
 230
 231         /*
 232          * Install the new, split up pagetable. Important detail here:
 233          *
 234          * On Intel the NX bit of all levels must be cleared to make a
 235          * page executable. See section 4.13.2 of Intel 64 and IA-32
 236          * Architectures Software Developer's Manual).
 237          */
 238         ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
 239         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
 240         base = NULL;
 241
 242 out_unlock:
 243         spin_unlock_irqrestore(&pgd_lock, flags);
 244
 245         if (base)
 246                 __free_pages(base, 0);
 247
 248         return 0;
 249 }
 250
 251 static int
 252 __change_page_attr(unsigned long address, unsigned long pfn,
 253                    pgprot_t mask_set, pgprot_t mask_clr)
 254 {
 255         struct page *kpte_page;
 256         int level, err = 0;
 257         pte_t *kpte;
 258
 259 #ifdef CONFIG_X86_32
 260         BUG_ON(pfn > max_low_pfn);
 261 #endif
 262
 263 repeat:
 264         kpte = lookup_address(address, &level);
 265         if (!kpte)
 266                 return -EINVAL;
 267
 268         kpte_page = virt_to_page(kpte);
 269         BUG_ON(PageLRU(kpte_page));
 270         BUG_ON(PageCompound(kpte_page));
 271
 272         if (level == PG_LEVEL_4K) {
 273                 pgprot_t new_prot = pte_pgprot(*kpte);
 274                 pte_t new_pte, old_pte = *kpte;
 275
 276                 pgprot_val(new_prot) &= ~pgprot_val(mask_clr);
 277                 pgprot_val(new_prot) |= pgprot_val(mask_set);
 278
 279                 new_prot = static_protections(new_prot, address);
 280
 281                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
 282                 BUG_ON(pte_pfn(new_pte) != pte_pfn(old_pte));
 283
 284                 set_pte_atomic(kpte, new_pte);
 285         } else {
 286                 err = split_large_page(kpte, address);
 287                 if (!err)
 288                         goto repeat;
 289         }
 290         return err;
 291 }
 292
 293 /**
 294  * change_page_attr_addr - Change page table attributes in linear mapping
 295  * @address: Virtual address in linear mapping.
 296  * @prot:    New page table attribute (PAGE_*)
 297  *
 298  * Change page attributes of a page in the direct mapping. This is a variant
 299  * of change_page_attr() that also works on memory holes that do not have
 300  * mem_map entry (pfn_valid() is false).
 301  *
 302  * See change_page_attr() documentation for more details.
 303  *
 304  * Modules and drivers should use the set_memory_* APIs instead.
 305  */
 306
 307 #define HIGH_MAP_START  __START_KERNEL_map
 308 #define HIGH_MAP_END    (__START_KERNEL_map + KERNEL_TEXT_SIZE)
 309
 310 static int
 311 change_page_attr_addr(unsigned long address, pgprot_t mask_set,
 312                       pgprot_t mask_clr)
 313 {
 314         unsigned long phys_addr = __pa(address);
 315         unsigned long pfn = phys_addr >> PAGE_SHIFT;
 316         int err;
 317
 318 #ifdef CONFIG_X86_64
 319         /*
 320          * If we are inside the high mapped kernel range, then we
 321          * fixup the low mapping first. __va() returns the virtual
 322          * address in the linear mapping:
 323          */
 324         if (within(address, HIGH_MAP_START, HIGH_MAP_END))
 325                 address = (unsigned long) __va(phys_addr);
 326 #endif
 327
 328         err = __change_page_attr(address, pfn, mask_set, mask_clr);
 329         if (err)
 330                 return err;
 331
 332 #ifdef CONFIG_X86_64
 333         /*
 334          * If the physical address is inside the kernel map, we need
 335          * to touch the high mapped kernel as well:
 336          */
 337         if (within(phys_addr, 0, KERNEL_TEXT_SIZE)) {
 338                 /*
 339                  * Calc the high mapping address. See __phys_addr()
 340                  * for the non obvious details.
 341                  */
 342                 address = phys_addr + HIGH_MAP_START - phys_base;
 343                 /* Make sure the kernel mappings stay executable */
 344                 pgprot_val(mask_clr) |= _PAGE_NX;
 345
 346                 /*
 347                  * Our high aliases are imprecise, because we check
 348                  * everything between 0 and KERNEL_TEXT_SIZE, so do
 349                  * not propagate lookup failures back to users:
 350                  */
 351                 __change_page_attr(address, pfn, mask_set, mask_clr);
 352         }
 353 #endif
 354         return err;
 355 }
 356
 357 static int __change_page_attr_set_clr(unsigned long addr, int numpages,
 358                                       pgprot_t mask_set, pgprot_t mask_clr)
 359 {
 360         unsigned int i;
 361         int ret;
 362
 363         for (i = 0; i < numpages ; i++, addr += PAGE_SIZE) {
 364                 ret = change_page_attr_addr(addr, mask_set, mask_clr);
 365                 if (ret)
 366                         return ret;
 367         }
 368
 369         return 0;
 370 }
 371
 372 static int change_page_attr_set_clr(unsigned long addr, int numpages,
 373                                     pgprot_t mask_set, pgprot_t mask_clr)
 374 {
 375         int ret = __change_page_attr_set_clr(addr, numpages, mask_set,
 376                                              mask_clr);
 377
 378         /*
 379          * On success we use clflush, when the CPU supports it to
 380          * avoid the wbindv. If the CPU does not support it and in the
 381          * error case we fall back to cpa_flush_all (which uses
 382          * wbindv):
 383          */
 384         if (!ret && cpu_has_clflush)
 385                 cpa_flush_range(addr, numpages);
 386         else
 387                 cpa_flush_all();
 388
 389         return ret;
 390 }
 391
 392 static inline int change_page_attr_set(unsigned long addr, int numpages,
 393                                        pgprot_t mask)
 394 {
 395         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0));
 396 }
 397
 398 static inline int change_page_attr_clear(unsigned long addr, int numpages,
 399                                          pgprot_t mask)
 400 {
 401         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask);
 402 }
 403
 404 int set_memory_uc(unsigned long addr, int numpages)
 405 {
 406         return change_page_attr_set(addr, numpages,
 407                                     __pgprot(_PAGE_PCD | _PAGE_PWT));
 408 }
 409 EXPORT_SYMBOL(set_memory_uc);
 410
 411 int set_memory_wb(unsigned long addr, int numpages)
 412 {
 413         return change_page_attr_clear(addr, numpages,
 414                                       __pgprot(_PAGE_PCD | _PAGE_PWT));
 415 }
 416 EXPORT_SYMBOL(set_memory_wb);
 417
 418 int set_memory_x(unsigned long addr, int numpages)
 419 {
 420         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
 421 }
 422 EXPORT_SYMBOL(set_memory_x);
 423
 424 int set_memory_nx(unsigned long addr, int numpages)
 425 {
 426         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
 427 }
 428 EXPORT_SYMBOL(set_memory_nx);
 429
 430 int set_memory_ro(unsigned long addr, int numpages)
 431 {
 432         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
 433 }
 434
 435 int set_memory_rw(unsigned long addr, int numpages)
 436 {
 437         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
 438 }
 439
 440 int set_memory_np(unsigned long addr, int numpages)
 441 {
 442         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
 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 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_CPA_DEBUG)
 493 static inline int __change_page_attr_set(unsigned long addr, int numpages,
 494                                          pgprot_t mask)
 495 {
 496         return __change_page_attr_set_clr(addr, numpages, mask, __pgprot(0));
 497 }
 498
 499 static inline int __change_page_attr_clear(unsigned long addr, int numpages,
 500                                            pgprot_t mask)
 501 {
 502         return __change_page_attr_set_clr(addr, numpages, __pgprot(0), mask);
 503 }
 504 #endif
 505
 506 #ifdef CONFIG_DEBUG_PAGEALLOC
 507
 508 static int __set_pages_p(struct page *page, int numpages)
 509 {
 510         unsigned long addr = (unsigned long)page_address(page);
 511
 512         return __change_page_attr_set(addr, numpages,
 513                                       __pgprot(_PAGE_PRESENT | _PAGE_RW));
 514 }
 515
 516 static int __set_pages_np(struct page *page, int numpages)
 517 {
 518         unsigned long addr = (unsigned long)page_address(page);
 519
 520         return __change_page_attr_clear(addr, numpages,
 521                                         __pgprot(_PAGE_PRESENT));
 522 }
 523
 524 void kernel_map_pages(struct page *page, int numpages, int enable)
 525 {
 526         if (PageHighMem(page))
 527                 return;
 528         if (!enable) {
 529                 debug_check_no_locks_freed(page_address(page),
 530                                            numpages * PAGE_SIZE);
 531         }
 532
 533         /*
 534          * If page allocator is not up yet then do not call c_p_a():
 535          */
 536         if (!debug_pagealloc_enabled)
 537                 return;
 538
 539         /*
 540          * The return value is ignored - the calls cannot fail,
 541          * large pages are disabled at boot time:
 542          */
 543         if (enable)
 544                 __set_pages_p(page, numpages);
 545         else
 546                 __set_pages_np(page, numpages);
 547
 548         /*
 549          * We should perform an IPI and flush all tlbs,
 550          * but that can deadlock->flush only current cpu:
 551          */
 552         __flush_tlb_all();
 553 }
 554 #endif
 555
 556 /*
 557  * The testcases use internal knowledge of the implementation that shouldn't
 558  * be exposed to the rest of the kernel. Include these directly here.
 559  */
 560 #ifdef CONFIG_CPA_DEBUG
 561 #include "pageattr-test.c"
 562 #endif