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 /**
  31  * clflush_cache_range - flush a cache range with clflush
  32  * @addr:       virtual start address
  33  * @size:       number of bytes to flush
  34  *
  35  * clflush is an unordered instruction which needs fencing with mfence
  36  * to avoid ordering issues.
  37  */
  38 void clflush_cache_range(void *addr, int size)
  39 {
  40         int i;
  41
  42         mb();
  43         for (i = 0; i < size; i += boot_cpu_data.x86_clflush_size)
  44                 clflush(addr+i);
  45         mb();
  46 }
  47
  48 static void __cpa_flush_all(void *arg)
  49 {
  50         /*
  51          * Flush all to work around Errata in early athlons regarding
  52          * large page flushing.
  53          */
  54         __flush_tlb_all();
  55
  56         if (boot_cpu_data.x86_model >= 4)
  57                 wbinvd();
  58 }
  59
  60 static void cpa_flush_all(void)
  61 {
  62         BUG_ON(irqs_disabled());
  63
  64         on_each_cpu(__cpa_flush_all, NULL, 1, 1);
  65 }
  66
  67 struct clflush_data {
  68         unsigned long addr;
  69         int numpages;
  70 };
  71
  72 static void __cpa_flush_range(void *arg)
  73 {
  74         struct clflush_data *cld = arg;
  75
  76         /*
  77          * We could optimize that further and do individual per page
  78          * tlb invalidates for a low number of pages. Caveat: we must
  79          * flush the high aliases on 64bit as well.
  80          */
  81         __flush_tlb_all();
  82
  83         clflush_cache_range((void *) cld->addr, cld->numpages * PAGE_SIZE);
  84 }
  85
  86 static void cpa_flush_range(unsigned long addr, int numpages)
  87 {
  88         struct clflush_data cld;
  89
  90         BUG_ON(irqs_disabled());
  91
  92         cld.addr = addr;
  93         cld.numpages = numpages;
  94
  95         on_each_cpu(__cpa_flush_range, &cld, 1, 1);
  96 }
  97
  98 /*
  99  * Certain areas of memory on x86 require very specific protection flags,
 100  * for example the BIOS area or kernel text. Callers don't always get this
 101  * right (again, ioremap() on BIOS memory is not uncommon) so this function
 102  * checks and fixes these known static required protection bits.
 103  */
 104 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address)
 105 {
 106         pgprot_t forbidden = __pgprot(0);
 107
 108         /*
 109          * The BIOS area between 640k and 1Mb needs to be executable for
 110          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
 111          */
 112         if (within(__pa(address), BIOS_BEGIN, BIOS_END))
 113                 pgprot_val(forbidden) |= _PAGE_NX;
 114
 115         /*
 116          * The kernel text needs to be executable for obvious reasons
 117          * Does not cover __inittext since that is gone later on
 118          */
 119         if (within(address, (unsigned long)_text, (unsigned long)_etext))
 120                 pgprot_val(forbidden) |= _PAGE_NX;
 121
 122 #ifdef CONFIG_DEBUG_RODATA
 123         /* The .rodata section needs to be read-only */
 124         if (within(address, (unsigned long)__start_rodata,
 125                                 (unsigned long)__end_rodata))
 126                 pgprot_val(forbidden) |= _PAGE_RW;
 127 #endif
 128
 129         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
 130
 131         return prot;
 132 }
 133
 134 pte_t *lookup_address(unsigned long address, int *level)
 135 {
 136         pgd_t *pgd = pgd_offset_k(address);
 137         pud_t *pud;
 138         pmd_t *pmd;
 139
 140         *level = PG_LEVEL_NONE;
 141
 142         if (pgd_none(*pgd))
 143                 return NULL;
 144         pud = pud_offset(pgd, address);
 145         if (pud_none(*pud))
 146                 return NULL;
 147         pmd = pmd_offset(pud, address);
 148         if (pmd_none(*pmd))
 149                 return NULL;
 150
 151         *level = PG_LEVEL_2M;
 152         if (pmd_large(*pmd))
 153                 return (pte_t *)pmd;
 154
 155         *level = PG_LEVEL_4K;
 156         return pte_offset_kernel(pmd, address);
 157 }
 158
 159 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 160 {
 161         /* change init_mm */
 162         set_pte_atomic(kpte, pte);
 163 #ifdef CONFIG_X86_32
 164         if (!SHARED_KERNEL_PMD) {
 165                 struct page *page;
 166
 167                 for (page = pgd_list; page; page = (struct page *)page->index) {
 168                         pgd_t *pgd;
 169                         pud_t *pud;
 170                         pmd_t *pmd;
 171
 172                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 173                         pud = pud_offset(pgd, address);
 174                         pmd = pmd_offset(pud, address);
 175                         set_pte_atomic((pte_t *)pmd, pte);
 176                 }
 177         }
 178 #endif
 179 }
 180
 181 static int split_large_page(pte_t *kpte, unsigned long address)
 182 {
 183         pgprot_t ref_prot = pte_pgprot(pte_clrhuge(*kpte));
 184         gfp_t gfp_flags = GFP_KERNEL;
 185         unsigned long flags;
 186         unsigned long addr;
 187         pte_t *pbase, *tmp;
 188         struct page *base;
 189         int i, level;
 190
 191 #ifdef CONFIG_DEBUG_PAGEALLOC
 192         gfp_flags = GFP_ATOMIC;
 193 #endif
 194         base = alloc_pages(gfp_flags, 0);
 195         if (!base)
 196                 return -ENOMEM;
 197
 198         spin_lock_irqsave(&pgd_lock, flags);
 199         /*
 200          * Check for races, another CPU might have split this page
 201          * up for us already:
 202          */
 203         tmp = lookup_address(address, &level);
 204         if (tmp != kpte) {
 205                 WARN_ON_ONCE(1);
 206                 goto out_unlock;
 207         }
 208
 209         address = __pa(address);
 210         addr = address & LARGE_PAGE_MASK;
 211         pbase = (pte_t *)page_address(base);
 212 #ifdef CONFIG_X86_32
 213         paravirt_alloc_pt(&init_mm, page_to_pfn(base));
 214 #endif
 215
 216         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
 217                 set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT, ref_prot));
 218
 219         /*
 220          * Install the new, split up pagetable. Important detail here:
 221          *
 222          * On Intel the NX bit of all levels must be cleared to make a
 223          * page executable. See section 4.13.2 of Intel 64 and IA-32
 224          * Architectures Software Developer's Manual).
 225          */
 226         ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
 227         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
 228         base = NULL;
 229
 230 out_unlock:
 231         spin_unlock_irqrestore(&pgd_lock, flags);
 232
 233         if (base)
 234                 __free_pages(base, 0);
 235
 236         return 0;
 237 }
 238
 239 static int
 240 __change_page_attr(unsigned long address, unsigned long pfn, pgprot_t prot)
 241 {
 242         struct page *kpte_page;
 243         int level, err = 0;
 244         pte_t *kpte;
 245
 246 #ifdef CONFIG_X86_32
 247         BUG_ON(pfn > max_low_pfn);
 248 #endif
 249
 250 repeat:
 251         kpte = lookup_address(address, &level);
 252         if (!kpte)
 253                 return -EINVAL;
 254
 255         kpte_page = virt_to_page(kpte);
 256         BUG_ON(PageLRU(kpte_page));
 257         BUG_ON(PageCompound(kpte_page));
 258
 259         prot = static_protections(prot, address);
 260
 261         if (level == PG_LEVEL_4K) {
 262                 WARN_ON_ONCE(pgprot_val(prot) & _PAGE_PSE);
 263                 set_pte_atomic(kpte, pfn_pte(pfn, canon_pgprot(prot)));
 264         } else {
 265                 /* Clear the PSE bit for the 4k level pages ! */
 266                 pgprot_val(prot) = pgprot_val(prot) & ~_PAGE_PSE;
 267
 268                 err = split_large_page(kpte, address);
 269                 if (!err)
 270                         goto repeat;
 271         }
 272         return err;
 273 }
 274
 275 /**
 276  * change_page_attr_addr - Change page table attributes in linear mapping
 277  * @address: Virtual address in linear mapping.
 278  * @prot:    New page table attribute (PAGE_*)
 279  *
 280  * Change page attributes of a page in the direct mapping. This is a variant
 281  * of change_page_attr() that also works on memory holes that do not have
 282  * mem_map entry (pfn_valid() is false).
 283  *
 284  * See change_page_attr() documentation for more details.
 285  *
 286  * Modules and drivers should use the set_memory_* APIs instead.
 287  */
 288
 289 static int change_page_attr_addr(unsigned long address, pgprot_t prot)
 290 {
 291         int err = 0, kernel_map = 0;
 292         unsigned long pfn = __pa(address) >> PAGE_SHIFT;
 293
 294 #ifdef CONFIG_X86_64
 295         if (address >= __START_KERNEL_map &&
 296                         address < __START_KERNEL_map + KERNEL_TEXT_SIZE) {
 297
 298                 address = (unsigned long)__va(__pa(address));
 299                 kernel_map = 1;
 300         }
 301 #endif
 302
 303         if (!kernel_map || pte_present(pfn_pte(0, prot))) {
 304                 err = __change_page_attr(address, pfn, prot);
 305                 if (err)
 306                         return err;
 307         }
 308
 309 #ifdef CONFIG_X86_64
 310         /*
 311          * Handle kernel mapping too which aliases part of
 312          * lowmem:
 313          */
 314         if (__pa(address) < KERNEL_TEXT_SIZE) {
 315                 unsigned long addr2;
 316                 pgprot_t prot2;
 317
 318                 addr2 = __START_KERNEL_map + __pa(address);
 319                 /* Make sure the kernel mappings stay executable */
 320                 prot2 = pte_pgprot(pte_mkexec(pfn_pte(0, prot)));
 321                 err = __change_page_attr(addr2, pfn, prot2);
 322         }
 323 #endif
 324
 325         return err;
 326 }
 327
 328 static int __change_page_attr_set_clr(unsigned long addr, int numpages,
 329                                       pgprot_t mask_set, pgprot_t mask_clr)
 330 {
 331         pgprot_t new_prot;
 332         int level;
 333         pte_t *pte;
 334         int i, ret;
 335
 336         for (i = 0; i < numpages ; i++) {
 337
 338                 pte = lookup_address(addr, &level);
 339                 if (!pte)
 340                         return -EINVAL;
 341
 342                 new_prot = pte_pgprot(*pte);
 343
 344                 pgprot_val(new_prot) &= ~pgprot_val(mask_clr);
 345                 pgprot_val(new_prot) |= pgprot_val(mask_set);
 346
 347                 ret = change_page_attr_addr(addr, new_prot);
 348                 if (ret)
 349                         return ret;
 350                 addr += PAGE_SIZE;
 351         }
 352
 353         return 0;
 354 }
 355
 356 static int change_page_attr_set_clr(unsigned long addr, int numpages,
 357                                     pgprot_t mask_set, pgprot_t mask_clr)
 358 {
 359         int ret = __change_page_attr_set_clr(addr, numpages, mask_set,
 360                                              mask_clr);
 361
 362         /*
 363          * On success we use clflush, when the CPU supports it to
 364          * avoid the wbindv. If the CPU does not support it and in the
 365          * error case we fall back to cpa_flush_all (which uses
 366          * wbindv):
 367          */
 368         if (!ret && cpu_has_clflush)
 369                 cpa_flush_range(addr, numpages);
 370         else
 371                 cpa_flush_all();
 372
 373         return ret;
 374 }
 375
 376 static inline int change_page_attr_set(unsigned long addr, int numpages,
 377                                        pgprot_t mask)
 378 {
 379         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0));
 380 }
 381
 382 static inline int change_page_attr_clear(unsigned long addr, int numpages,
 383                                          pgprot_t mask)
 384 {
 385         return __change_page_attr_set_clr(addr, numpages, __pgprot(0), mask);
 386
 387 }
 388
 389 int set_memory_uc(unsigned long addr, int numpages)
 390 {
 391         return change_page_attr_set(addr, numpages,
 392                                     __pgprot(_PAGE_PCD | _PAGE_PWT));
 393 }
 394 EXPORT_SYMBOL(set_memory_uc);
 395
 396 int set_memory_wb(unsigned long addr, int numpages)
 397 {
 398         return change_page_attr_clear(addr, numpages,
 399                                       __pgprot(_PAGE_PCD | _PAGE_PWT));
 400 }
 401 EXPORT_SYMBOL(set_memory_wb);
 402
 403 int set_memory_x(unsigned long addr, int numpages)
 404 {
 405         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
 406 }
 407 EXPORT_SYMBOL(set_memory_x);
 408
 409 int set_memory_nx(unsigned long addr, int numpages)
 410 {
 411         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
 412 }
 413 EXPORT_SYMBOL(set_memory_nx);
 414
 415 int set_memory_ro(unsigned long addr, int numpages)
 416 {
 417         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
 418 }
 419
 420 int set_memory_rw(unsigned long addr, int numpages)
 421 {
 422         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
 423 }
 424
 425 int set_memory_np(unsigned long addr, int numpages)
 426 {
 427         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
 428 }
 429
 430 int set_pages_uc(struct page *page, int numpages)
 431 {
 432         unsigned long addr = (unsigned long)page_address(page);
 433
 434         return set_memory_uc(addr, numpages);
 435 }
 436 EXPORT_SYMBOL(set_pages_uc);
 437
 438 int set_pages_wb(struct page *page, int numpages)
 439 {
 440         unsigned long addr = (unsigned long)page_address(page);
 441
 442         return set_memory_wb(addr, numpages);
 443 }
 444 EXPORT_SYMBOL(set_pages_wb);
 445
 446 int set_pages_x(struct page *page, int numpages)
 447 {
 448         unsigned long addr = (unsigned long)page_address(page);
 449
 450         return set_memory_x(addr, numpages);
 451 }
 452 EXPORT_SYMBOL(set_pages_x);
 453
 454 int set_pages_nx(struct page *page, int numpages)
 455 {
 456         unsigned long addr = (unsigned long)page_address(page);
 457
 458         return set_memory_nx(addr, numpages);
 459 }
 460 EXPORT_SYMBOL(set_pages_nx);
 461
 462 int set_pages_ro(struct page *page, int numpages)
 463 {
 464         unsigned long addr = (unsigned long)page_address(page);
 465
 466         return set_memory_ro(addr, numpages);
 467 }
 468
 469 int set_pages_rw(struct page *page, int numpages)
 470 {
 471         unsigned long addr = (unsigned long)page_address(page);
 472
 473         return set_memory_rw(addr, numpages);
 474 }
 475
 476
 477 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_CPA_DEBUG)
 478 static inline int __change_page_attr_set(unsigned long addr, int numpages,
 479                                          pgprot_t mask)
 480 {
 481         return __change_page_attr_set_clr(addr, numpages, mask, __pgprot(0));
 482 }
 483
 484 static inline int __change_page_attr_clear(unsigned long addr, int numpages,
 485                                            pgprot_t mask)
 486 {
 487         return __change_page_attr_set_clr(addr, numpages, __pgprot(0), mask);
 488 }
 489 #endif
 490
 491 #ifdef CONFIG_DEBUG_PAGEALLOC
 492
 493 static int __set_pages_p(struct page *page, int numpages)
 494 {
 495         unsigned long addr = (unsigned long)page_address(page);
 496
 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
 505         return __change_page_attr_clear(addr, numpages,
 506                                         __pgprot(_PAGE_PRESENT));
 507 }
 508
 509 void kernel_map_pages(struct page *page, int numpages, int enable)
 510 {
 511         if (PageHighMem(page))
 512                 return;
 513         if (!enable) {
 514                 debug_check_no_locks_freed(page_address(page),
 515                                            numpages * PAGE_SIZE);
 516         }
 517
 518         /*
 519          * If page allocator is not up yet then do not call c_p_a():
 520          */
 521         if (!debug_pagealloc_enabled)
 522                 return;
 523
 524         /*
 525          * The return value is ignored - the calls cannot fail,
 526          * large pages are disabled at boot time:
 527          */
 528         if (enable)
 529                 __set_pages_p(page, numpages);
 530         else
 531                 __set_pages_np(page, numpages);
 532
 533         /*
 534          * We should perform an IPI and flush all tlbs,
 535          * but that can deadlock->flush only current cpu:
 536          */
 537         __flush_tlb_all();
 538 }
 539 #endif
 540
 541 /*
 542  * The testcases use internal knowledge of the implementation that shouldn't
 543  * be exposed to the rest of the kernel. Include these directly here.
 544  */
 545 #ifdef CONFIG_CPA_DEBUG
 546 #include "pageattr-test.c"
 547 #endif