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 struct cpa_data {
  20         unsigned long   vaddr;
  21         int             numpages;
  22         pgprot_t        mask_set;
  23         pgprot_t        mask_clr;
  24         int             flushtlb;
  25 };
  26
  27 static inline int
  28 within(unsigned long addr, unsigned long start, unsigned long end)
  29 {
  30         return addr >= start && addr < end;
  31 }
  32
  33 /*
  34  * Flushing functions
  35  */
  36
  37 /**
  38  * clflush_cache_range - flush a cache range with clflush
  39  * @addr:       virtual start address
  40  * @size:       number of bytes to flush
  41  *
  42  * clflush is an unordered instruction which needs fencing with mfence
  43  * to avoid ordering issues.
  44  */
  45 void clflush_cache_range(void *vaddr, unsigned int size)
  46 {
  47         void *vend = vaddr + size - 1;
  48
  49         mb();
  50
  51         for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
  52                 clflush(vaddr);
  53         /*
  54          * Flush any possible final partial cacheline:
  55          */
  56         clflush(vend);
  57
  58         mb();
  59 }
  60
  61 static void __cpa_flush_all(void *arg)
  62 {
  63         unsigned long cache = (unsigned long)arg;
  64
  65         /*
  66          * Flush all to work around Errata in early athlons regarding
  67          * large page flushing.
  68          */
  69         __flush_tlb_all();
  70
  71         if (cache && boot_cpu_data.x86_model >= 4)
  72                 wbinvd();
  73 }
  74
  75 static void cpa_flush_all(unsigned long cache)
  76 {
  77         BUG_ON(irqs_disabled());
  78
  79         on_each_cpu(__cpa_flush_all, (void *) cache, 1, 1);
  80 }
  81
  82 static void __cpa_flush_range(void *arg)
  83 {
  84         /*
  85          * We could optimize that further and do individual per page
  86          * tlb invalidates for a low number of pages. Caveat: we must
  87          * flush the high aliases on 64bit as well.
  88          */
  89         __flush_tlb_all();
  90 }
  91
  92 static void cpa_flush_range(unsigned long start, int numpages, int cache)
  93 {
  94         unsigned int i, level;
  95         unsigned long addr;
  96
  97         BUG_ON(irqs_disabled());
  98         WARN_ON(PAGE_ALIGN(start) != start);
  99
 100         on_each_cpu(__cpa_flush_range, NULL, 1, 1);
 101
 102         if (!cache)
 103                 return;
 104
 105         /*
 106          * We only need to flush on one CPU,
 107          * clflush is a MESI-coherent instruction that
 108          * will cause all other CPUs to flush the same
 109          * cachelines:
 110          */
 111         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
 112                 pte_t *pte = lookup_address(addr, &level);
 113
 114                 /*
 115                  * Only flush present addresses:
 116                  */
 117                 if (pte && pte_present(*pte))
 118                         clflush_cache_range((void *) addr, PAGE_SIZE);
 119         }
 120 }
 121
 122 #define HIGH_MAP_START  __START_KERNEL_map
 123 #define HIGH_MAP_END    (__START_KERNEL_map + KERNEL_TEXT_SIZE)
 124
 125
 126 /*
 127  * Converts a virtual address to a X86-64 highmap address
 128  */
 129 static unsigned long virt_to_highmap(void *address)
 130 {
 131 #ifdef CONFIG_X86_64
 132         return __pa((unsigned long)address) + HIGH_MAP_START - phys_base;
 133 #else
 134         return (unsigned long)address;
 135 #endif
 136 }
 137
 138 /*
 139  * Certain areas of memory on x86 require very specific protection flags,
 140  * for example the BIOS area or kernel text. Callers don't always get this
 141  * right (again, ioremap() on BIOS memory is not uncommon) so this function
 142  * checks and fixes these known static required protection bits.
 143  */
 144 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address)
 145 {
 146         pgprot_t forbidden = __pgprot(0);
 147
 148         /*
 149          * The BIOS area between 640k and 1Mb needs to be executable for
 150          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
 151          */
 152         if (within(__pa(address), BIOS_BEGIN, BIOS_END))
 153                 pgprot_val(forbidden) |= _PAGE_NX;
 154
 155         /*
 156          * The kernel text needs to be executable for obvious reasons
 157          * Does not cover __inittext since that is gone later on
 158          */
 159         if (within(address, (unsigned long)_text, (unsigned long)_etext))
 160                 pgprot_val(forbidden) |= _PAGE_NX;
 161         /*
 162          * Do the same for the x86-64 high kernel mapping
 163          */
 164         if (within(address, virt_to_highmap(_text), virt_to_highmap(_etext)))
 165                 pgprot_val(forbidden) |= _PAGE_NX;
 166
 167
 168 #ifdef CONFIG_DEBUG_RODATA
 169         /* The .rodata section needs to be read-only */
 170         if (within(address, (unsigned long)__start_rodata,
 171                                 (unsigned long)__end_rodata))
 172                 pgprot_val(forbidden) |= _PAGE_RW;
 173         /*
 174          * Do the same for the x86-64 high kernel mapping
 175          */
 176         if (within(address, virt_to_highmap(__start_rodata),
 177                                 virt_to_highmap(__end_rodata)))
 178                 pgprot_val(forbidden) |= _PAGE_RW;
 179 #endif
 180
 181         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
 182
 183         return prot;
 184 }
 185
 186 pte_t *lookup_address(unsigned long address, int *level)
 187 {
 188         pgd_t *pgd = pgd_offset_k(address);
 189         pud_t *pud;
 190         pmd_t *pmd;
 191
 192         *level = PG_LEVEL_NONE;
 193
 194         if (pgd_none(*pgd))
 195                 return NULL;
 196         pud = pud_offset(pgd, address);
 197         if (pud_none(*pud))
 198                 return NULL;
 199         pmd = pmd_offset(pud, address);
 200         if (pmd_none(*pmd))
 201                 return NULL;
 202
 203         *level = PG_LEVEL_2M;
 204         if (pmd_large(*pmd))
 205                 return (pte_t *)pmd;
 206
 207         *level = PG_LEVEL_4K;
 208         return pte_offset_kernel(pmd, address);
 209 }
 210
 211 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 212 {
 213         /* change init_mm */
 214         set_pte_atomic(kpte, pte);
 215 #ifdef CONFIG_X86_32
 216         if (!SHARED_KERNEL_PMD) {
 217                 struct page *page;
 218
 219                 list_for_each_entry(page, &pgd_list, lru) {
 220                         pgd_t *pgd;
 221                         pud_t *pud;
 222                         pmd_t *pmd;
 223
 224                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 225                         pud = pud_offset(pgd, address);
 226                         pmd = pmd_offset(pud, address);
 227                         set_pte_atomic((pte_t *)pmd, pte);
 228                 }
 229         }
 230 #endif
 231 }
 232
 233 static int split_large_page(pte_t *kpte, unsigned long address)
 234 {
 235         pgprot_t ref_prot = pte_pgprot(pte_clrhuge(*kpte));
 236         gfp_t gfp_flags = GFP_KERNEL;
 237         unsigned long flags, addr, pfn;
 238         pte_t *pbase, *tmp;
 239         struct page *base;
 240         unsigned int i, level;
 241
 242 #ifdef CONFIG_DEBUG_PAGEALLOC
 243         gfp_flags = __GFP_HIGH | __GFP_NOFAIL | __GFP_NOWARN;
 244         gfp_flags = GFP_ATOMIC | __GFP_NOWARN;
 245 #endif
 246         base = alloc_pages(gfp_flags, 0);
 247         if (!base)
 248                 return -ENOMEM;
 249
 250         spin_lock_irqsave(&pgd_lock, flags);
 251         /*
 252          * Check for races, another CPU might have split this page
 253          * up for us already:
 254          */
 255         tmp = lookup_address(address, &level);
 256         if (tmp != kpte) {
 257                 WARN_ON_ONCE(1);
 258                 goto out_unlock;
 259         }
 260
 261         address = __pa(address);
 262         addr = address & LARGE_PAGE_MASK;
 263         pbase = (pte_t *)page_address(base);
 264 #ifdef CONFIG_X86_32
 265         paravirt_alloc_pt(&init_mm, page_to_pfn(base));
 266 #endif
 267
 268         /*
 269          * Get the target pfn from the original entry:
 270          */
 271         pfn = pte_pfn(*kpte);
 272         for (i = 0; i < PTRS_PER_PTE; i++, pfn++)
 273                 set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
 274
 275         /*
 276          * Install the new, split up pagetable. Important detail here:
 277          *
 278          * On Intel the NX bit of all levels must be cleared to make a
 279          * page executable. See section 4.13.2 of Intel 64 and IA-32
 280          * Architectures Software Developer's Manual).
 281          */
 282         ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
 283         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
 284         base = NULL;
 285
 286 out_unlock:
 287         spin_unlock_irqrestore(&pgd_lock, flags);
 288
 289         if (base)
 290                 __free_pages(base, 0);
 291
 292         return 0;
 293 }
 294
 295 static int __change_page_attr(unsigned long address, struct cpa_data *cpa)
 296 {
 297         struct page *kpte_page;
 298         int level, err = 0;
 299         pte_t *kpte;
 300
 301 repeat:
 302         kpte = lookup_address(address, &level);
 303         if (!kpte)
 304                 return -EINVAL;
 305
 306         kpte_page = virt_to_page(kpte);
 307         BUG_ON(PageLRU(kpte_page));
 308         BUG_ON(PageCompound(kpte_page));
 309
 310         if (level == PG_LEVEL_4K) {
 311                 pte_t new_pte, old_pte = *kpte;
 312                 pgprot_t new_prot = pte_pgprot(old_pte);
 313
 314                 if(!pte_val(old_pte)) {
 315                         printk(KERN_WARNING "CPA: called for zero pte. "
 316                                "vaddr = %lx cpa->vaddr = %lx\n", address,
 317                                 cpa->vaddr);
 318                         WARN_ON(1);
 319                         return -EINVAL;
 320                 }
 321
 322                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
 323                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
 324
 325                 new_prot = static_protections(new_prot, address);
 326
 327                 /*
 328                  * We need to keep the pfn from the existing PTE,
 329                  * after all we're only going to change it's attributes
 330                  * not the memory it points to
 331                  */
 332                 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
 333
 334                 /*
 335                  * Do we really change anything ?
 336                  */
 337                 if (pte_val(old_pte) != pte_val(new_pte)) {
 338                         set_pte_atomic(kpte, new_pte);
 339                         cpa->flushtlb = 1;
 340                 }
 341         } else {
 342                 err = split_large_page(kpte, address);
 343                 if (!err)
 344                         goto repeat;
 345                 cpa->flushtlb = 1;
 346         }
 347         return err;
 348 }
 349
 350 /**
 351  * change_page_attr_addr - Change page table attributes in linear mapping
 352  * @address: Virtual address in linear mapping.
 353  * @prot:    New page table attribute (PAGE_*)
 354  *
 355  * Change page attributes of a page in the direct mapping. This is a variant
 356  * of change_page_attr() that also works on memory holes that do not have
 357  * mem_map entry (pfn_valid() is false).
 358  *
 359  * See change_page_attr() documentation for more details.
 360  *
 361  * Modules and drivers should use the set_memory_* APIs instead.
 362  */
 363
 364 static int change_page_attr_addr(struct cpa_data *cpa)
 365 {
 366         int err;
 367         unsigned long address = cpa->vaddr;
 368
 369 #ifdef CONFIG_X86_64
 370         unsigned long phys_addr = __pa(address);
 371
 372         /*
 373          * If we are inside the high mapped kernel range, then we
 374          * fixup the low mapping first. __va() returns the virtual
 375          * address in the linear mapping:
 376          */
 377         if (within(address, HIGH_MAP_START, HIGH_MAP_END))
 378                 address = (unsigned long) __va(phys_addr);
 379 #endif
 380
 381         err = __change_page_attr(address, cpa);
 382         if (err)
 383                 return err;
 384
 385 #ifdef CONFIG_X86_64
 386         /*
 387          * If the physical address is inside the kernel map, we need
 388          * to touch the high mapped kernel as well:
 389          */
 390         if (within(phys_addr, 0, KERNEL_TEXT_SIZE)) {
 391                 /*
 392                  * Calc the high mapping address. See __phys_addr()
 393                  * for the non obvious details.
 394                  *
 395                  * Note that NX and other required permissions are
 396                  * checked in static_protections().
 397                  */
 398                 address = phys_addr + HIGH_MAP_START - phys_base;
 399
 400                 /*
 401                  * Our high aliases are imprecise, because we check
 402                  * everything between 0 and KERNEL_TEXT_SIZE, so do
 403                  * not propagate lookup failures back to users:
 404                  */
 405                 __change_page_attr(address, cpa);
 406         }
 407 #endif
 408         return err;
 409 }
 410
 411 static int __change_page_attr_set_clr(struct cpa_data *cpa)
 412 {
 413         unsigned int i;
 414         int ret;
 415
 416         for (i = 0; i < cpa->numpages ; i++, cpa->vaddr += PAGE_SIZE) {
 417                 ret = change_page_attr_addr(cpa);
 418                 if (ret)
 419                         return ret;
 420         }
 421
 422         return 0;
 423 }
 424
 425 static inline int cache_attr(pgprot_t attr)
 426 {
 427         return pgprot_val(attr) &
 428                 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
 429 }
 430
 431 static int change_page_attr_set_clr(unsigned long addr, int numpages,
 432                                     pgprot_t mask_set, pgprot_t mask_clr)
 433 {
 434         struct cpa_data cpa;
 435         int ret, cache;
 436
 437         /*
 438          * Check, if we are requested to change a not supported
 439          * feature:
 440          */
 441         mask_set = canon_pgprot(mask_set);
 442         mask_clr = canon_pgprot(mask_clr);
 443         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr))
 444                 return 0;
 445
 446         cpa.vaddr = addr;
 447         cpa.numpages = numpages;
 448         cpa.mask_set = mask_set;
 449         cpa.mask_clr = mask_clr;
 450         cpa.flushtlb = 0;
 451
 452         ret = __change_page_attr_set_clr(&cpa);
 453
 454         /*
 455          * Check whether we really changed something:
 456          */
 457         if (!cpa.flushtlb)
 458                 return ret;
 459
 460         /*
 461          * No need to flush, when we did not set any of the caching
 462          * attributes:
 463          */
 464         cache = cache_attr(mask_set);
 465
 466         /*
 467          * On success we use clflush, when the CPU supports it to
 468          * avoid the wbindv. If the CPU does not support it and in the
 469          * error case we fall back to cpa_flush_all (which uses
 470          * wbindv):
 471          */
 472         if (!ret && cpu_has_clflush)
 473                 cpa_flush_range(addr, numpages, cache);
 474         else
 475                 cpa_flush_all(cache);
 476
 477         return ret;
 478 }
 479
 480 static inline int change_page_attr_set(unsigned long addr, int numpages,
 481                                        pgprot_t mask)
 482 {
 483         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0));
 484 }
 485
 486 static inline int change_page_attr_clear(unsigned long addr, int numpages,
 487                                          pgprot_t mask)
 488 {
 489         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask);
 490 }
 491
 492 int set_memory_uc(unsigned long addr, int numpages)
 493 {
 494         return change_page_attr_set(addr, numpages,
 495                                     __pgprot(_PAGE_PCD | _PAGE_PWT));
 496 }
 497 EXPORT_SYMBOL(set_memory_uc);
 498
 499 int set_memory_wb(unsigned long addr, int numpages)
 500 {
 501         return change_page_attr_clear(addr, numpages,
 502                                       __pgprot(_PAGE_PCD | _PAGE_PWT));
 503 }
 504 EXPORT_SYMBOL(set_memory_wb);
 505
 506 int set_memory_x(unsigned long addr, int numpages)
 507 {
 508         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
 509 }
 510 EXPORT_SYMBOL(set_memory_x);
 511
 512 int set_memory_nx(unsigned long addr, int numpages)
 513 {
 514         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
 515 }
 516 EXPORT_SYMBOL(set_memory_nx);
 517
 518 int set_memory_ro(unsigned long addr, int numpages)
 519 {
 520         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
 521 }
 522
 523 int set_memory_rw(unsigned long addr, int numpages)
 524 {
 525         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
 526 }
 527
 528 int set_memory_np(unsigned long addr, int numpages)
 529 {
 530         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
 531 }
 532
 533 int set_pages_uc(struct page *page, int numpages)
 534 {
 535         unsigned long addr = (unsigned long)page_address(page);
 536
 537         return set_memory_uc(addr, numpages);
 538 }
 539 EXPORT_SYMBOL(set_pages_uc);
 540
 541 int set_pages_wb(struct page *page, int numpages)
 542 {
 543         unsigned long addr = (unsigned long)page_address(page);
 544
 545         return set_memory_wb(addr, numpages);
 546 }
 547 EXPORT_SYMBOL(set_pages_wb);
 548
 549 int set_pages_x(struct page *page, int numpages)
 550 {
 551         unsigned long addr = (unsigned long)page_address(page);
 552
 553         return set_memory_x(addr, numpages);
 554 }
 555 EXPORT_SYMBOL(set_pages_x);
 556
 557 int set_pages_nx(struct page *page, int numpages)
 558 {
 559         unsigned long addr = (unsigned long)page_address(page);
 560
 561         return set_memory_nx(addr, numpages);
 562 }
 563 EXPORT_SYMBOL(set_pages_nx);
 564
 565 int set_pages_ro(struct page *page, int numpages)
 566 {
 567         unsigned long addr = (unsigned long)page_address(page);
 568
 569         return set_memory_ro(addr, numpages);
 570 }
 571
 572 int set_pages_rw(struct page *page, int numpages)
 573 {
 574         unsigned long addr = (unsigned long)page_address(page);
 575
 576         return set_memory_rw(addr, numpages);
 577 }
 578
 579 #ifdef CONFIG_DEBUG_PAGEALLOC
 580
 581 static int __set_pages_p(struct page *page, int numpages)
 582 {
 583         struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
 584                                 .numpages = numpages,
 585                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
 586                                 .mask_clr = __pgprot(0)};
 587
 588         return __change_page_attr_set_clr(&cpa);
 589 }
 590
 591 static int __set_pages_np(struct page *page, int numpages)
 592 {
 593         struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
 594                                 .numpages = numpages,
 595                                 .mask_set = __pgprot(0),
 596                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW)};
 597
 598         return __change_page_attr_set_clr(&cpa);
 599 }
 600
 601 void kernel_map_pages(struct page *page, int numpages, int enable)
 602 {
 603         if (PageHighMem(page))
 604                 return;
 605         if (!enable) {
 606                 debug_check_no_locks_freed(page_address(page),
 607                                            numpages * PAGE_SIZE);
 608         }
 609
 610         /*
 611          * If page allocator is not up yet then do not call c_p_a():
 612          */
 613         if (!debug_pagealloc_enabled)
 614                 return;
 615
 616         /*
 617          * The return value is ignored - the calls cannot fail,
 618          * large pages are disabled at boot time:
 619          */
 620         if (enable)
 621                 __set_pages_p(page, numpages);
 622         else
 623                 __set_pages_np(page, numpages);
 624
 625         /*
 626          * We should perform an IPI and flush all tlbs,
 627          * but that can deadlock->flush only current cpu:
 628          */
 629         __flush_tlb_all();
 630 }
 631 #endif
 632
 633 /*
 634  * The testcases use internal knowledge of the implementation that shouldn't
 635  * be exposed to the rest of the kernel. Include these directly here.
 636  */
 637 #ifdef CONFIG_CPA_DEBUG
 638 #include "pageattr-test.c"
 639 #endif