Pileus Git - ~andy/linux/blob - 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/mm.h>
  10 #include <linux/interrupt.h>
  11 #include <linux/seq_file.h>
  12 #include <linux/debugfs.h>
  13 #include <linux/pfn.h>
  14 #include <linux/percpu.h>
  15 #include <linux/gfp.h>
  16 #include <linux/pci.h>
  17
  18 #include <asm/e820.h>
  19 #include <asm/processor.h>
  20 #include <asm/tlbflush.h>
  21 #include <asm/sections.h>
  22 #include <asm/setup.h>
  23 #include <asm/uaccess.h>
  24 #include <asm/pgalloc.h>
  25 #include <asm/proto.h>
  26 #include <asm/pat.h>
  27
  28 /*
  29  * The current flushing context - we pass it instead of 5 arguments:
  30  */
  31 struct cpa_data {
  32         unsigned long   *vaddr;
  33         pgd_t           *pgd;
  34         pgprot_t        mask_set;
  35         pgprot_t        mask_clr;
  36         int             numpages;
  37         int             flags;
  38         unsigned long   pfn;
  39         unsigned        force_split : 1;
  40         int             curpage;
  41         struct page     **pages;
  42 };
  43
  44 /*
  45  * Serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity mappings)
  46  * using cpa_lock. So that we don't allow any other cpu, with stale large tlb
  47  * entries change the page attribute in parallel to some other cpu
  48  * splitting a large page entry along with changing the attribute.
  49  */
  50 static DEFINE_SPINLOCK(cpa_lock);
  51
  52 #define CPA_FLUSHTLB 1
  53 #define CPA_ARRAY 2
  54 #define CPA_PAGES_ARRAY 4
  55
  56 #ifdef CONFIG_PROC_FS
  57 static unsigned long direct_pages_count[PG_LEVEL_NUM];
  58
  59 void update_page_count(int level, unsigned long pages)
  60 {
  61         /* Protect against CPA */
  62         spin_lock(&pgd_lock);
  63         direct_pages_count[level] += pages;
  64         spin_unlock(&pgd_lock);
  65 }
  66
  67 static void split_page_count(int level)
  68 {
  69         direct_pages_count[level]--;
  70         direct_pages_count[level - 1] += PTRS_PER_PTE;
  71 }
  72
  73 void arch_report_meminfo(struct seq_file *m)
  74 {
  75         seq_printf(m, "DirectMap4k:    %8lu kB\n",
  76                         direct_pages_count[PG_LEVEL_4K] << 2);
  77 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
  78         seq_printf(m, "DirectMap2M:    %8lu kB\n",
  79                         direct_pages_count[PG_LEVEL_2M] << 11);
  80 #else
  81         seq_printf(m, "DirectMap4M:    %8lu kB\n",
  82                         direct_pages_count[PG_LEVEL_2M] << 12);
  83 #endif
  84 #ifdef CONFIG_X86_64
  85         if (direct_gbpages)
  86                 seq_printf(m, "DirectMap1G:    %8lu kB\n",
  87                         direct_pages_count[PG_LEVEL_1G] << 20);
  88 #endif
  89 }
  90 #else
  91 static inline void split_page_count(int level) { }
  92 #endif
  93
  94 #ifdef CONFIG_X86_64
  95
  96 static inline unsigned long highmap_start_pfn(void)
  97 {
  98         return __pa_symbol(_text) >> PAGE_SHIFT;
  99 }
 100
 101 static inline unsigned long highmap_end_pfn(void)
 102 {
 103         return __pa_symbol(roundup(_brk_end, PMD_SIZE)) >> PAGE_SHIFT;
 104 }
 105
 106 #endif
 107
 108 #ifdef CONFIG_DEBUG_PAGEALLOC
 109 # define debug_pagealloc 1
 110 #else
 111 # define debug_pagealloc 0
 112 #endif
 113
 114 static inline int
 115 within(unsigned long addr, unsigned long start, unsigned long end)
 116 {
 117         return addr >= start && addr < end;
 118 }
 119
 120 /*
 121  * Flushing functions
 122  */
 123
 124 /**
 125  * clflush_cache_range - flush a cache range with clflush
 126  * @vaddr:      virtual start address
 127  * @size:       number of bytes to flush
 128  *
 129  * clflush is an unordered instruction which needs fencing with mfence
 130  * to avoid ordering issues.
 131  */
 132 void clflush_cache_range(void *vaddr, unsigned int size)
 133 {
 134         void *vend = vaddr + size - 1;
 135
 136         mb();
 137
 138         for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
 139                 clflush(vaddr);
 140         /*
 141          * Flush any possible final partial cacheline:
 142          */
 143         clflush(vend);
 144
 145         mb();
 146 }
 147 EXPORT_SYMBOL_GPL(clflush_cache_range);
 148
 149 static void __cpa_flush_all(void *arg)
 150 {
 151         unsigned long cache = (unsigned long)arg;
 152
 153         /*
 154          * Flush all to work around Errata in early athlons regarding
 155          * large page flushing.
 156          */
 157         __flush_tlb_all();
 158
 159         if (cache && boot_cpu_data.x86 >= 4)
 160                 wbinvd();
 161 }
 162
 163 static void cpa_flush_all(unsigned long cache)
 164 {
 165         BUG_ON(irqs_disabled());
 166
 167         on_each_cpu(__cpa_flush_all, (void *) cache, 1);
 168 }
 169
 170 static void __cpa_flush_range(void *arg)
 171 {
 172         /*
 173          * We could optimize that further and do individual per page
 174          * tlb invalidates for a low number of pages. Caveat: we must
 175          * flush the high aliases on 64bit as well.
 176          */
 177         __flush_tlb_all();
 178 }
 179
 180 static void cpa_flush_range(unsigned long start, int numpages, int cache)
 181 {
 182         unsigned int i, level;
 183         unsigned long addr;
 184
 185         BUG_ON(irqs_disabled());
 186         WARN_ON(PAGE_ALIGN(start) != start);
 187
 188         on_each_cpu(__cpa_flush_range, NULL, 1);
 189
 190         if (!cache)
 191                 return;
 192
 193         /*
 194          * We only need to flush on one CPU,
 195          * clflush is a MESI-coherent instruction that
 196          * will cause all other CPUs to flush the same
 197          * cachelines:
 198          */
 199         for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
 200                 pte_t *pte = lookup_address(addr, &level);
 201
 202                 /*
 203                  * Only flush present addresses:
 204                  */
 205                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 206                         clflush_cache_range((void *) addr, PAGE_SIZE);
 207         }
 208 }
 209
 210 static void cpa_flush_array(unsigned long *start, int numpages, int cache,
 211                             int in_flags, struct page **pages)
 212 {
 213         unsigned int i, level;
 214         unsigned long do_wbinvd = cache && numpages >= 1024; /* 4M threshold */
 215
 216         BUG_ON(irqs_disabled());
 217
 218         on_each_cpu(__cpa_flush_all, (void *) do_wbinvd, 1);
 219
 220         if (!cache || do_wbinvd)
 221                 return;
 222
 223         /*
 224          * We only need to flush on one CPU,
 225          * clflush is a MESI-coherent instruction that
 226          * will cause all other CPUs to flush the same
 227          * cachelines:
 228          */
 229         for (i = 0; i < numpages; i++) {
 230                 unsigned long addr;
 231                 pte_t *pte;
 232
 233                 if (in_flags & CPA_PAGES_ARRAY)
 234                         addr = (unsigned long)page_address(pages[i]);
 235                 else
 236                         addr = start[i];
 237
 238                 pte = lookup_address(addr, &level);
 239
 240                 /*
 241                  * Only flush present addresses:
 242                  */
 243                 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 244                         clflush_cache_range((void *)addr, PAGE_SIZE);
 245         }
 246 }
 247
 248 /*
 249  * Certain areas of memory on x86 require very specific protection flags,
 250  * for example the BIOS area or kernel text. Callers don't always get this
 251  * right (again, ioremap() on BIOS memory is not uncommon) so this function
 252  * checks and fixes these known static required protection bits.
 253  */
 254 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
 255                                    unsigned long pfn)
 256 {
 257         pgprot_t forbidden = __pgprot(0);
 258
 259         /*
 260          * The BIOS area between 640k and 1Mb needs to be executable for
 261          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
 262          */
 263 #ifdef CONFIG_PCI_BIOS
 264         if (pcibios_enabled && within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
 265                 pgprot_val(forbidden) |= _PAGE_NX;
 266 #endif
 267
 268         /*
 269          * The kernel text needs to be executable for obvious reasons
 270          * Does not cover __inittext since that is gone later on. On
 271          * 64bit we do not enforce !NX on the low mapping
 272          */
 273         if (within(address, (unsigned long)_text, (unsigned long)_etext))
 274                 pgprot_val(forbidden) |= _PAGE_NX;
 275
 276         /*
 277          * The .rodata section needs to be read-only. Using the pfn
 278          * catches all aliases.
 279          */
 280         if (within(pfn, __pa_symbol(__start_rodata) >> PAGE_SHIFT,
 281                    __pa_symbol(__end_rodata) >> PAGE_SHIFT))
 282                 pgprot_val(forbidden) |= _PAGE_RW;
 283
 284 #if defined(CONFIG_X86_64) && defined(CONFIG_DEBUG_RODATA)
 285         /*
 286          * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
 287          * kernel text mappings for the large page aligned text, rodata sections
 288          * will be always read-only. For the kernel identity mappings covering
 289          * the holes caused by this alignment can be anything that user asks.
 290          *
 291          * This will preserve the large page mappings for kernel text/data
 292          * at no extra cost.
 293          */
 294         if (kernel_set_to_readonly &&
 295             within(address, (unsigned long)_text,
 296                    (unsigned long)__end_rodata_hpage_align)) {
 297                 unsigned int level;
 298
 299                 /*
 300                  * Don't enforce the !RW mapping for the kernel text mapping,
 301                  * if the current mapping is already using small page mapping.
 302                  * No need to work hard to preserve large page mappings in this
 303                  * case.
 304                  *
 305                  * This also fixes the Linux Xen paravirt guest boot failure
 306                  * (because of unexpected read-only mappings for kernel identity
 307                  * mappings). In this paravirt guest case, the kernel text
 308                  * mapping and the kernel identity mapping share the same
 309                  * page-table pages. Thus we can't really use different
 310                  * protections for the kernel text and identity mappings. Also,
 311                  * these shared mappings are made of small page mappings.
 312                  * Thus this don't enforce !RW mapping for small page kernel
 313                  * text mapping logic will help Linux Xen parvirt guest boot
 314                  * as well.
 315                  */
 316                 if (lookup_address(address, &level) && (level != PG_LEVEL_4K))
 317                         pgprot_val(forbidden) |= _PAGE_RW;
 318         }
 319 #endif
 320
 321         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
 322
 323         return prot;
 324 }
 325
 326 static pte_t *__lookup_address_in_pgd(pgd_t *pgd, unsigned long address,
 327                                       unsigned int *level)
 328 {
 329         pud_t *pud;
 330         pmd_t *pmd;
 331
 332         *level = PG_LEVEL_NONE;
 333
 334         if (pgd_none(*pgd))
 335                 return NULL;
 336
 337         pud = pud_offset(pgd, address);
 338         if (pud_none(*pud))
 339                 return NULL;
 340
 341         *level = PG_LEVEL_1G;
 342         if (pud_large(*pud) || !pud_present(*pud))
 343                 return (pte_t *)pud;
 344
 345         pmd = pmd_offset(pud, address);
 346         if (pmd_none(*pmd))
 347                 return NULL;
 348
 349         *level = PG_LEVEL_2M;
 350         if (pmd_large(*pmd) || !pmd_present(*pmd))
 351                 return (pte_t *)pmd;
 352
 353         *level = PG_LEVEL_4K;
 354
 355         return pte_offset_kernel(pmd, address);
 356 }
 357
 358 /*
 359  * Lookup the page table entry for a virtual address. Return a pointer
 360  * to the entry and the level of the mapping.
 361  *
 362  * Note: We return pud and pmd either when the entry is marked large
 363  * or when the present bit is not set. Otherwise we would return a
 364  * pointer to a nonexisting mapping.
 365  */
 366 pte_t *lookup_address(unsigned long address, unsigned int *level)
 367 {
 368         return __lookup_address_in_pgd(pgd_offset_k(address), address, level);
 369 }
 370 EXPORT_SYMBOL_GPL(lookup_address);
 371
 372 static pte_t *_lookup_address_cpa(struct cpa_data *cpa, unsigned long address,
 373                                   unsigned int *level)
 374 {
 375         if (cpa->pgd)
 376                 return __lookup_address_in_pgd(cpa->pgd + pgd_index(address),
 377                                                address, level);
 378
 379         return lookup_address(address, level);
 380 }
 381
 382 /*
 383  * This is necessary because __pa() does not work on some
 384  * kinds of memory, like vmalloc() or the alloc_remap()
 385  * areas on 32-bit NUMA systems.  The percpu areas can
 386  * end up in this kind of memory, for instance.
 387  *
 388  * This could be optimized, but it is only intended to be
 389  * used at inititalization time, and keeping it
 390  * unoptimized should increase the testing coverage for
 391  * the more obscure platforms.
 392  */
 393 phys_addr_t slow_virt_to_phys(void *__virt_addr)
 394 {
 395         unsigned long virt_addr = (unsigned long)__virt_addr;
 396         phys_addr_t phys_addr;
 397         unsigned long offset;
 398         enum pg_level level;
 399         unsigned long psize;
 400         unsigned long pmask;
 401         pte_t *pte;
 402
 403         pte = lookup_address(virt_addr, &level);
 404         BUG_ON(!pte);
 405         psize = page_level_size(level);
 406         pmask = page_level_mask(level);
 407         offset = virt_addr & ~pmask;
 408         phys_addr = pte_pfn(*pte) << PAGE_SHIFT;
 409         return (phys_addr | offset);
 410 }
 411 EXPORT_SYMBOL_GPL(slow_virt_to_phys);
 412
 413 /*
 414  * Set the new pmd in all the pgds we know about:
 415  */
 416 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 417 {
 418         /* change init_mm */
 419         set_pte_atomic(kpte, pte);
 420 #ifdef CONFIG_X86_32
 421         if (!SHARED_KERNEL_PMD) {
 422                 struct page *page;
 423
 424                 list_for_each_entry(page, &pgd_list, lru) {
 425                         pgd_t *pgd;
 426                         pud_t *pud;
 427                         pmd_t *pmd;
 428
 429                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 430                         pud = pud_offset(pgd, address);
 431                         pmd = pmd_offset(pud, address);
 432                         set_pte_atomic((pte_t *)pmd, pte);
 433                 }
 434         }
 435 #endif
 436 }
 437
 438 static int
 439 try_preserve_large_page(pte_t *kpte, unsigned long address,
 440                         struct cpa_data *cpa)
 441 {
 442         unsigned long nextpage_addr, numpages, pmask, psize, addr, pfn;
 443         pte_t new_pte, old_pte, *tmp;
 444         pgprot_t old_prot, new_prot, req_prot;
 445         int i, do_split = 1;
 446         enum pg_level level;
 447
 448         if (cpa->force_split)
 449                 return 1;
 450
 451         spin_lock(&pgd_lock);
 452         /*
 453          * Check for races, another CPU might have split this page
 454          * up already:
 455          */
 456         tmp = _lookup_address_cpa(cpa, address, &level);
 457         if (tmp != kpte)
 458                 goto out_unlock;
 459
 460         switch (level) {
 461         case PG_LEVEL_2M:
 462 #ifdef CONFIG_X86_64
 463         case PG_LEVEL_1G:
 464 #endif
 465                 psize = page_level_size(level);
 466                 pmask = page_level_mask(level);
 467                 break;
 468         default:
 469                 do_split = -EINVAL;
 470                 goto out_unlock;
 471         }
 472
 473         /*
 474          * Calculate the number of pages, which fit into this large
 475          * page starting at address:
 476          */
 477         nextpage_addr = (address + psize) & pmask;
 478         numpages = (nextpage_addr - address) >> PAGE_SHIFT;
 479         if (numpages < cpa->numpages)
 480                 cpa->numpages = numpages;
 481
 482         /*
 483          * We are safe now. Check whether the new pgprot is the same:
 484          */
 485         old_pte = *kpte;
 486         old_prot = req_prot = pte_pgprot(old_pte);
 487
 488         pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
 489         pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
 490
 491         /*
 492          * Set the PSE and GLOBAL flags only if the PRESENT flag is
 493          * set otherwise pmd_present/pmd_huge will return true even on
 494          * a non present pmd. The canon_pgprot will clear _PAGE_GLOBAL
 495          * for the ancient hardware that doesn't support it.
 496          */
 497         if (pgprot_val(req_prot) & _PAGE_PRESENT)
 498                 pgprot_val(req_prot) |= _PAGE_PSE | _PAGE_GLOBAL;
 499         else
 500                 pgprot_val(req_prot) &= ~(_PAGE_PSE | _PAGE_GLOBAL);
 501
 502         req_prot = canon_pgprot(req_prot);
 503
 504         /*
 505          * old_pte points to the large page base address. So we need
 506          * to add the offset of the virtual address:
 507          */
 508         pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
 509         cpa->pfn = pfn;
 510
 511         new_prot = static_protections(req_prot, address, pfn);
 512
 513         /*
 514          * We need to check the full range, whether
 515          * static_protection() requires a different pgprot for one of
 516          * the pages in the range we try to preserve:
 517          */
 518         addr = address & pmask;
 519         pfn = pte_pfn(old_pte);
 520         for (i = 0; i < (psize >> PAGE_SHIFT); i++, addr += PAGE_SIZE, pfn++) {
 521                 pgprot_t chk_prot = static_protections(req_prot, addr, pfn);
 522
 523                 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
 524                         goto out_unlock;
 525         }
 526
 527         /*
 528          * If there are no changes, return. maxpages has been updated
 529          * above:
 530          */
 531         if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
 532                 do_split = 0;
 533                 goto out_unlock;
 534         }
 535
 536         /*
 537          * We need to change the attributes. Check, whether we can
 538          * change the large page in one go. We request a split, when
 539          * the address is not aligned and the number of pages is
 540          * smaller than the number of pages in the large page. Note
 541          * that we limited the number of possible pages already to
 542          * the number of pages in the large page.
 543          */
 544         if (address == (address & pmask) && cpa->numpages == (psize >> PAGE_SHIFT)) {
 545                 /*
 546                  * The address is aligned and the number of pages
 547                  * covers the full page.
 548                  */
 549                 new_pte = pfn_pte(pte_pfn(old_pte), new_prot);
 550                 __set_pmd_pte(kpte, address, new_pte);
 551                 cpa->flags |= CPA_FLUSHTLB;
 552                 do_split = 0;
 553         }
 554
 555 out_unlock:
 556         spin_unlock(&pgd_lock);
 557
 558         return do_split;
 559 }
 560
 561 static int
 562 __split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
 563                    struct page *base)
 564 {
 565         pte_t *pbase = (pte_t *)page_address(base);
 566         unsigned long pfn, pfninc = 1;
 567         unsigned int i, level;
 568         pte_t *tmp;
 569         pgprot_t ref_prot;
 570
 571         spin_lock(&pgd_lock);
 572         /*
 573          * Check for races, another CPU might have split this page
 574          * up for us already:
 575          */
 576         tmp = _lookup_address_cpa(cpa, address, &level);
 577         if (tmp != kpte) {
 578                 spin_unlock(&pgd_lock);
 579                 return 1;
 580         }
 581
 582         paravirt_alloc_pte(&init_mm, page_to_pfn(base));
 583         ref_prot = pte_pgprot(pte_clrhuge(*kpte));
 584         /*
 585          * If we ever want to utilize the PAT bit, we need to
 586          * update this function to make sure it's converted from
 587          * bit 12 to bit 7 when we cross from the 2MB level to
 588          * the 4K level:
 589          */
 590         WARN_ON_ONCE(pgprot_val(ref_prot) & _PAGE_PAT_LARGE);
 591
 592 #ifdef CONFIG_X86_64
 593         if (level == PG_LEVEL_1G) {
 594                 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
 595                 /*
 596                  * Set the PSE flags only if the PRESENT flag is set
 597                  * otherwise pmd_present/pmd_huge will return true
 598                  * even on a non present pmd.
 599                  */
 600                 if (pgprot_val(ref_prot) & _PAGE_PRESENT)
 601                         pgprot_val(ref_prot) |= _PAGE_PSE;
 602                 else
 603                         pgprot_val(ref_prot) &= ~_PAGE_PSE;
 604         }
 605 #endif
 606
 607         /*
 608          * Set the GLOBAL flags only if the PRESENT flag is set
 609          * otherwise pmd/pte_present will return true even on a non
 610          * present pmd/pte. The canon_pgprot will clear _PAGE_GLOBAL
 611          * for the ancient hardware that doesn't support it.
 612          */
 613         if (pgprot_val(ref_prot) & _PAGE_PRESENT)
 614                 pgprot_val(ref_prot) |= _PAGE_GLOBAL;
 615         else
 616                 pgprot_val(ref_prot) &= ~_PAGE_GLOBAL;
 617
 618         /*
 619          * Get the target pfn from the original entry:
 620          */
 621         pfn = pte_pfn(*kpte);
 622         for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
 623                 set_pte(&pbase[i], pfn_pte(pfn, canon_pgprot(ref_prot)));
 624
 625         if (pfn_range_is_mapped(PFN_DOWN(__pa(address)),
 626                                 PFN_DOWN(__pa(address)) + 1))
 627                 split_page_count(level);
 628
 629         /*
 630          * Install the new, split up pagetable.
 631          *
 632          * We use the standard kernel pagetable protections for the new
 633          * pagetable protections, the actual ptes set above control the
 634          * primary protection behavior:
 635          */
 636         __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
 637
 638         /*
 639          * Intel Atom errata AAH41 workaround.
 640          *
 641          * The real fix should be in hw or in a microcode update, but
 642          * we also probabilistically try to reduce the window of having
 643          * a large TLB mixed with 4K TLBs while instruction fetches are
 644          * going on.
 645          */
 646         __flush_tlb_all();
 647         spin_unlock(&pgd_lock);
 648
 649         return 0;
 650 }
 651
 652 static int split_large_page(struct cpa_data *cpa, pte_t *kpte,
 653                             unsigned long address)
 654 {
 655         struct page *base;
 656
 657         if (!debug_pagealloc)
 658                 spin_unlock(&cpa_lock);
 659         base = alloc_pages(GFP_KERNEL | __GFP_NOTRACK, 0);
 660         if (!debug_pagealloc)
 661                 spin_lock(&cpa_lock);
 662         if (!base)
 663                 return -ENOMEM;
 664
 665         if (__split_large_page(cpa, kpte, address, base))
 666                 __free_page(base);
 667
 668         return 0;
 669 }
 670
 671 static bool try_to_free_pte_page(pte_t *pte)
 672 {
 673         int i;
 674
 675         for (i = 0; i < PTRS_PER_PTE; i++)
 676                 if (!pte_none(pte[i]))
 677                         return false;
 678
 679         free_page((unsigned long)pte);
 680         return true;
 681 }
 682
 683 static bool try_to_free_pmd_page(pmd_t *pmd)
 684 {
 685         int i;
 686
 687         for (i = 0; i < PTRS_PER_PMD; i++)
 688                 if (!pmd_none(pmd[i]))
 689                         return false;
 690
 691         free_page((unsigned long)pmd);
 692         return true;
 693 }
 694
 695 static bool unmap_pte_range(pmd_t *pmd, unsigned long start, unsigned long end)
 696 {
 697         pte_t *pte = pte_offset_kernel(pmd, start);
 698
 699         while (start < end) {
 700                 set_pte(pte, __pte(0));
 701
 702                 start += PAGE_SIZE;
 703                 pte++;
 704         }
 705
 706         if (try_to_free_pte_page((pte_t *)pmd_page_vaddr(*pmd))) {
 707                 pmd_clear(pmd);
 708                 return true;
 709         }
 710         return false;
 711 }
 712
 713 static void __unmap_pmd_range(pud_t *pud, pmd_t *pmd,
 714                               unsigned long start, unsigned long end)
 715 {
 716         if (unmap_pte_range(pmd, start, end))
 717                 if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
 718                         pud_clear(pud);
 719 }
 720
 721 static void unmap_pmd_range(pud_t *pud, unsigned long start, unsigned long end)
 722 {
 723         pmd_t *pmd = pmd_offset(pud, start);
 724
 725         /*
 726          * Not on a 2MB page boundary?
 727          */
 728         if (start & (PMD_SIZE - 1)) {
 729                 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
 730                 unsigned long pre_end = min_t(unsigned long, end, next_page);
 731
 732                 __unmap_pmd_range(pud, pmd, start, pre_end);
 733
 734                 start = pre_end;
 735                 pmd++;
 736         }
 737
 738         /*
 739          * Try to unmap in 2M chunks.
 740          */
 741         while (end - start >= PMD_SIZE) {
 742                 if (pmd_large(*pmd))
 743                         pmd_clear(pmd);
 744                 else
 745                         __unmap_pmd_range(pud, pmd, start, start + PMD_SIZE);
 746
 747                 start += PMD_SIZE;
 748                 pmd++;
 749         }
 750
 751         /*
 752          * 4K leftovers?
 753          */
 754         if (start < end)
 755                 return __unmap_pmd_range(pud, pmd, start, end);
 756
 757         /*
 758          * Try again to free the PMD page if haven't succeeded above.
 759          */
 760         if (!pud_none(*pud))
 761                 if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
 762                         pud_clear(pud);
 763 }
 764
 765 static void unmap_pud_range(pgd_t *pgd, unsigned long start, unsigned long end)
 766 {
 767         pud_t *pud = pud_offset(pgd, start);
 768
 769         /*
 770          * Not on a GB page boundary?
 771          */
 772         if (start & (PUD_SIZE - 1)) {
 773                 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
 774                 unsigned long pre_end   = min_t(unsigned long, end, next_page);
 775
 776                 unmap_pmd_range(pud, start, pre_end);
 777
 778                 start = pre_end;
 779                 pud++;
 780         }
 781
 782         /*
 783          * Try to unmap in 1G chunks?
 784          */
 785         while (end - start >= PUD_SIZE) {
 786
 787                 if (pud_large(*pud))
 788                         pud_clear(pud);
 789                 else
 790                         unmap_pmd_range(pud, start, start + PUD_SIZE);
 791
 792                 start += PUD_SIZE;
 793                 pud++;
 794         }
 795
 796         /*
 797          * 2M leftovers?
 798          */
 799         if (start < end)
 800                 unmap_pmd_range(pud, start, end);
 801
 802         /*
 803          * No need to try to free the PUD page because we'll free it in
 804          * populate_pgd's error path
 805          */
 806 }
 807
 808 static int alloc_pte_page(pmd_t *pmd)
 809 {
 810         pte_t *pte = (pte_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
 811         if (!pte)
 812                 return -1;
 813
 814         set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
 815         return 0;
 816 }
 817
 818 static int alloc_pmd_page(pud_t *pud)
 819 {
 820         pmd_t *pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
 821         if (!pmd)
 822                 return -1;
 823
 824         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
 825         return 0;
 826 }
 827
 828 static void populate_pte(struct cpa_data *cpa,
 829                          unsigned long start, unsigned long end,
 830                          unsigned num_pages, pmd_t *pmd, pgprot_t pgprot)
 831 {
 832         pte_t *pte;
 833
 834         pte = pte_offset_kernel(pmd, start);
 835
 836         while (num_pages-- && start < end) {
 837
 838                 /* deal with the NX bit */
 839                 if (!(pgprot_val(pgprot) & _PAGE_NX))
 840                         cpa->pfn &= ~_PAGE_NX;
 841
 842                 set_pte(pte, pfn_pte(cpa->pfn >> PAGE_SHIFT, pgprot));
 843
 844                 start    += PAGE_SIZE;
 845                 cpa->pfn += PAGE_SIZE;
 846                 pte++;
 847         }
 848 }
 849
 850 static int populate_pmd(struct cpa_data *cpa,
 851                         unsigned long start, unsigned long end,
 852                         unsigned num_pages, pud_t *pud, pgprot_t pgprot)
 853 {
 854         unsigned int cur_pages = 0;
 855         pmd_t *pmd;
 856
 857         /*
 858          * Not on a 2M boundary?
 859          */
 860         if (start & (PMD_SIZE - 1)) {
 861                 unsigned long pre_end = start + (num_pages << PAGE_SHIFT);
 862                 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
 863
 864                 pre_end   = min_t(unsigned long, pre_end, next_page);
 865                 cur_pages = (pre_end - start) >> PAGE_SHIFT;
 866                 cur_pages = min_t(unsigned int, num_pages, cur_pages);
 867
 868                 /*
 869                  * Need a PTE page?
 870                  */
 871                 pmd = pmd_offset(pud, start);
 872                 if (pmd_none(*pmd))
 873                         if (alloc_pte_page(pmd))
 874                                 return -1;
 875
 876                 populate_pte(cpa, start, pre_end, cur_pages, pmd, pgprot);
 877
 878                 start = pre_end;
 879         }
 880
 881         /*
 882          * We mapped them all?
 883          */
 884         if (num_pages == cur_pages)
 885                 return cur_pages;
 886
 887         while (end - start >= PMD_SIZE) {
 888
 889                 /*
 890                  * We cannot use a 1G page so allocate a PMD page if needed.
 891                  */
 892                 if (pud_none(*pud))
 893                         if (alloc_pmd_page(pud))
 894                                 return -1;
 895
 896                 pmd = pmd_offset(pud, start);
 897
 898                 set_pmd(pmd, __pmd(cpa->pfn | _PAGE_PSE | massage_pgprot(pgprot)));
 899
 900                 start     += PMD_SIZE;
 901                 cpa->pfn  += PMD_SIZE;
 902                 cur_pages += PMD_SIZE >> PAGE_SHIFT;
 903         }
 904
 905         /*
 906          * Map trailing 4K pages.
 907          */
 908         if (start < end) {
 909                 pmd = pmd_offset(pud, start);
 910                 if (pmd_none(*pmd))
 911                         if (alloc_pte_page(pmd))
 912                                 return -1;
 913
 914                 populate_pte(cpa, start, end, num_pages - cur_pages,
 915                              pmd, pgprot);
 916         }
 917         return num_pages;
 918 }
 919
 920 static int populate_pud(struct cpa_data *cpa, unsigned long start, pgd_t *pgd,
 921                         pgprot_t pgprot)
 922 {
 923         pud_t *pud;
 924         unsigned long end;
 925         int cur_pages = 0;
 926
 927         end = start + (cpa->numpages << PAGE_SHIFT);
 928
 929         /*
 930          * Not on a Gb page boundary? => map everything up to it with
 931          * smaller pages.
 932          */
 933         if (start & (PUD_SIZE - 1)) {
 934                 unsigned long pre_end;
 935                 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
 936
 937                 pre_end   = min_t(unsigned long, end, next_page);
 938                 cur_pages = (pre_end - start) >> PAGE_SHIFT;
 939                 cur_pages = min_t(int, (int)cpa->numpages, cur_pages);
 940
 941                 pud = pud_offset(pgd, start);
 942
 943                 /*
 944                  * Need a PMD page?
 945                  */
 946                 if (pud_none(*pud))
 947                         if (alloc_pmd_page(pud))
 948                                 return -1;
 949
 950                 cur_pages = populate_pmd(cpa, start, pre_end, cur_pages,
 951                                          pud, pgprot);
 952                 if (cur_pages < 0)
 953                         return cur_pages;
 954
 955                 start = pre_end;
 956         }
 957
 958         /* We mapped them all? */
 959         if (cpa->numpages == cur_pages)
 960                 return cur_pages;
 961
 962         pud = pud_offset(pgd, start);
 963
 964         /*
 965          * Map everything starting from the Gb boundary, possibly with 1G pages
 966          */
 967         while (end - start >= PUD_SIZE) {
 968                 set_pud(pud, __pud(cpa->pfn | _PAGE_PSE | massage_pgprot(pgprot)));
 969
 970                 start     += PUD_SIZE;
 971                 cpa->pfn  += PUD_SIZE;
 972                 cur_pages += PUD_SIZE >> PAGE_SHIFT;
 973                 pud++;
 974         }
 975
 976         /* Map trailing leftover */
 977         if (start < end) {
 978                 int tmp;
 979
 980                 pud = pud_offset(pgd, start);
 981                 if (pud_none(*pud))
 982                         if (alloc_pmd_page(pud))
 983                                 return -1;
 984
 985                 tmp = populate_pmd(cpa, start, end, cpa->numpages - cur_pages,
 986                                    pud, pgprot);
 987                 if (tmp < 0)
 988                         return cur_pages;
 989
 990                 cur_pages += tmp;
 991         }
 992         return cur_pages;
 993 }
 994
 995 /*
 996  * Restrictions for kernel page table do not necessarily apply when mapping in
 997  * an alternate PGD.
 998  */
 999 static int populate_pgd(struct cpa_data *cpa, unsigned long addr)
1000 {
1001         pgprot_t pgprot = __pgprot(_KERNPG_TABLE);
1002         bool allocd_pgd = false;
1003         pgd_t *pgd_entry;
1004         pud_t *pud = NULL;      /* shut up gcc */
1005         int ret;
1006
1007         pgd_entry = cpa->pgd + pgd_index(addr);
1008
1009         /*
1010          * Allocate a PUD page and hand it down for mapping.
1011          */
1012         if (pgd_none(*pgd_entry)) {
1013                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
1014                 if (!pud)
1015                         return -1;
1016
1017                 set_pgd(pgd_entry, __pgd(__pa(pud) | _KERNPG_TABLE));
1018                 allocd_pgd = true;
1019         }
1020
1021         pgprot_val(pgprot) &= ~pgprot_val(cpa->mask_clr);
1022         pgprot_val(pgprot) |=  pgprot_val(cpa->mask_set);
1023
1024         ret = populate_pud(cpa, addr, pgd_entry, pgprot);
1025         if (ret < 0) {
1026                 unmap_pud_range(pgd_entry, addr,
1027                                 addr + (cpa->numpages << PAGE_SHIFT));
1028
1029                 if (allocd_pgd) {
1030                         /*
1031                          * If I allocated this PUD page, I can just as well
1032                          * free it in this error path.
1033                          */
1034                         pgd_clear(pgd_entry);
1035                         free_page((unsigned long)pud);
1036                 }
1037                 return ret;
1038         }
1039         cpa->numpages = ret;
1040         return 0;
1041 }
1042
1043 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
1044                                int primary)
1045 {
1046         if (cpa->pgd)
1047                 return populate_pgd(cpa, vaddr);
1048
1049         /*
1050          * Ignore all non primary paths.
1051          */
1052         if (!primary)
1053                 return 0;
1054
1055         /*
1056          * Ignore the NULL PTE for kernel identity mapping, as it is expected
1057          * to have holes.
1058          * Also set numpages to '1' indicating that we processed cpa req for
1059          * one virtual address page and its pfn. TBD: numpages can be set based
1060          * on the initial value and the level returned by lookup_address().
1061          */
1062         if (within(vaddr, PAGE_OFFSET,
1063                    PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
1064                 cpa->numpages = 1;
1065                 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
1066                 return 0;
1067         } else {
1068                 WARN(1, KERN_WARNING "CPA: called for zero pte. "
1069                         "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
1070                         *cpa->vaddr);
1071
1072                 return -EFAULT;
1073         }
1074 }
1075
1076 static int __change_page_attr(struct cpa_data *cpa, int primary)
1077 {
1078         unsigned long address;
1079         int do_split, err;
1080         unsigned int level;
1081         pte_t *kpte, old_pte;
1082
1083         if (cpa->flags & CPA_PAGES_ARRAY) {
1084                 struct page *page = cpa->pages[cpa->curpage];
1085                 if (unlikely(PageHighMem(page)))
1086                         return 0;
1087                 address = (unsigned long)page_address(page);
1088         } else if (cpa->flags & CPA_ARRAY)
1089                 address = cpa->vaddr[cpa->curpage];
1090         else
1091                 address = *cpa->vaddr;
1092 repeat:
1093         kpte = _lookup_address_cpa(cpa, address, &level);
1094         if (!kpte)
1095                 return __cpa_process_fault(cpa, address, primary);
1096
1097         old_pte = *kpte;
1098         if (!pte_val(old_pte))
1099                 return __cpa_process_fault(cpa, address, primary);
1100
1101         if (level == PG_LEVEL_4K) {
1102                 pte_t new_pte;
1103                 pgprot_t new_prot = pte_pgprot(old_pte);
1104                 unsigned long pfn = pte_pfn(old_pte);
1105
1106                 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
1107                 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
1108
1109                 new_prot = static_protections(new_prot, address, pfn);
1110
1111                 /*
1112                  * Set the GLOBAL flags only if the PRESENT flag is
1113                  * set otherwise pte_present will return true even on
1114                  * a non present pte. The canon_pgprot will clear
1115                  * _PAGE_GLOBAL for the ancient hardware that doesn't
1116                  * support it.
1117                  */
1118                 if (pgprot_val(new_prot) & _PAGE_PRESENT)
1119                         pgprot_val(new_prot) |= _PAGE_GLOBAL;
1120                 else
1121                         pgprot_val(new_prot) &= ~_PAGE_GLOBAL;
1122
1123                 /*
1124                  * We need to keep the pfn from the existing PTE,
1125                  * after all we're only going to change it's attributes
1126                  * not the memory it points to
1127                  */
1128                 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
1129                 cpa->pfn = pfn;
1130                 /*
1131                  * Do we really change anything ?
1132                  */
1133                 if (pte_val(old_pte) != pte_val(new_pte)) {
1134                         set_pte_atomic(kpte, new_pte);
1135                         cpa->flags |= CPA_FLUSHTLB;
1136                 }
1137                 cpa->numpages = 1;
1138                 return 0;
1139         }
1140
1141         /*
1142          * Check, whether we can keep the large page intact
1143          * and just change the pte:
1144          */
1145         do_split = try_preserve_large_page(kpte, address, cpa);
1146         /*
1147          * When the range fits into the existing large page,
1148          * return. cp->numpages and cpa->tlbflush have been updated in
1149          * try_large_page:
1150          */
1151         if (do_split <= 0)
1152                 return do_split;
1153
1154         /*
1155          * We have to split the large page:
1156          */
1157         err = split_large_page(cpa, kpte, address);
1158         if (!err) {
1159                 /*
1160                  * Do a global flush tlb after splitting the large page
1161                  * and before we do the actual change page attribute in the PTE.
1162                  *
1163                  * With out this, we violate the TLB application note, that says
1164                  * "The TLBs may contain both ordinary and large-page
1165                  *  translations for a 4-KByte range of linear addresses. This
1166                  *  may occur if software modifies the paging structures so that
1167                  *  the page size used for the address range changes. If the two
1168                  *  translations differ with respect to page frame or attributes
1169                  *  (e.g., permissions), processor behavior is undefined and may
1170                  *  be implementation-specific."
1171                  *
1172                  * We do this global tlb flush inside the cpa_lock, so that we
1173                  * don't allow any other cpu, with stale tlb entries change the
1174                  * page attribute in parallel, that also falls into the
1175                  * just split large page entry.
1176                  */
1177                 flush_tlb_all();
1178                 goto repeat;
1179         }
1180
1181         return err;
1182 }
1183
1184 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
1185
1186 static int cpa_process_alias(struct cpa_data *cpa)
1187 {
1188         struct cpa_data alias_cpa;
1189         unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
1190         unsigned long vaddr;
1191         int ret;
1192
1193         if (!pfn_range_is_mapped(cpa->pfn, cpa->pfn + 1))
1194                 return 0;
1195
1196         /*
1197          * No need to redo, when the primary call touched the direct
1198          * mapping already:
1199          */
1200         if (cpa->flags & CPA_PAGES_ARRAY) {
1201                 struct page *page = cpa->pages[cpa->curpage];
1202                 if (unlikely(PageHighMem(page)))
1203                         return 0;
1204                 vaddr = (unsigned long)page_address(page);
1205         } else if (cpa->flags & CPA_ARRAY)
1206                 vaddr = cpa->vaddr[cpa->curpage];
1207         else
1208                 vaddr = *cpa->vaddr;
1209
1210         if (!(within(vaddr, PAGE_OFFSET,
1211                     PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
1212
1213                 alias_cpa = *cpa;
1214                 alias_cpa.vaddr = &laddr;
1215                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1216
1217                 ret = __change_page_attr_set_clr(&alias_cpa, 0);
1218                 if (ret)
1219                         return ret;
1220         }
1221
1222 #ifdef CONFIG_X86_64
1223         /*
1224          * If the primary call didn't touch the high mapping already
1225          * and the physical address is inside the kernel map, we need
1226          * to touch the high mapped kernel as well:
1227          */
1228         if (!within(vaddr, (unsigned long)_text, _brk_end) &&
1229             within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn())) {
1230                 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
1231                                                __START_KERNEL_map - phys_base;
1232                 alias_cpa = *cpa;
1233                 alias_cpa.vaddr = &temp_cpa_vaddr;
1234                 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1235
1236                 /*
1237                  * The high mapping range is imprecise, so ignore the
1238                  * return value.
1239                  */
1240                 __change_page_attr_set_clr(&alias_cpa, 0);
1241         }
1242 #endif
1243
1244         return 0;
1245 }
1246
1247 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
1248 {
1249         int ret, numpages = cpa->numpages;
1250
1251         while (numpages) {
1252                 /*
1253                  * Store the remaining nr of pages for the large page
1254                  * preservation check.
1255                  */
1256                 cpa->numpages = numpages;
1257                 /* for array changes, we can't use large page */
1258                 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
1259                         cpa->numpages = 1;
1260
1261                 if (!debug_pagealloc)
1262                         spin_lock(&cpa_lock);
1263                 ret = __change_page_attr(cpa, checkalias);
1264                 if (!debug_pagealloc)
1265                         spin_unlock(&cpa_lock);
1266                 if (ret)
1267                         return ret;
1268
1269                 if (checkalias) {
1270                         ret = cpa_process_alias(cpa);
1271                         if (ret)
1272                                 return ret;
1273                 }
1274
1275                 /*
1276                  * Adjust the number of pages with the result of the
1277                  * CPA operation. Either a large page has been
1278                  * preserved or a single page update happened.
1279                  */
1280                 BUG_ON(cpa->numpages > numpages);
1281                 numpages -= cpa->numpages;
1282                 if (cpa->flags & (CPA_PAGES_ARRAY | CPA_ARRAY))
1283                         cpa->curpage++;
1284                 else
1285                         *cpa->vaddr += cpa->numpages * PAGE_SIZE;
1286
1287         }
1288         return 0;
1289 }
1290
1291 static inline int cache_attr(pgprot_t attr)
1292 {
1293         return pgprot_val(attr) &
1294                 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
1295 }
1296
1297 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
1298                                     pgprot_t mask_set, pgprot_t mask_clr,
1299                                     int force_split, int in_flag,
1300                                     struct page **pages)
1301 {
1302         struct cpa_data cpa;
1303         int ret, cache, checkalias;
1304         unsigned long baddr = 0;
1305
1306         memset(&cpa, 0, sizeof(cpa));
1307
1308         /*
1309          * Check, if we are requested to change a not supported
1310          * feature:
1311          */
1312         mask_set = canon_pgprot(mask_set);
1313         mask_clr = canon_pgprot(mask_clr);
1314         if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
1315                 return 0;
1316
1317         /* Ensure we are PAGE_SIZE aligned */
1318         if (in_flag & CPA_ARRAY) {
1319                 int i;
1320                 for (i = 0; i < numpages; i++) {
1321                         if (addr[i] & ~PAGE_MASK) {
1322                                 addr[i] &= PAGE_MASK;
1323                                 WARN_ON_ONCE(1);
1324                         }
1325                 }
1326         } else if (!(in_flag & CPA_PAGES_ARRAY)) {
1327                 /*
1328                  * in_flag of CPA_PAGES_ARRAY implies it is aligned.
1329                  * No need to cehck in that case
1330                  */
1331                 if (*addr & ~PAGE_MASK) {
1332                         *addr &= PAGE_MASK;
1333                         /*
1334                          * People should not be passing in unaligned addresses:
1335                          */
1336                         WARN_ON_ONCE(1);
1337                 }
1338                 /*
1339                  * Save address for cache flush. *addr is modified in the call
1340                  * to __change_page_attr_set_clr() below.
1341                  */
1342                 baddr = *addr;
1343         }
1344
1345         /* Must avoid aliasing mappings in the highmem code */
1346         kmap_flush_unused();
1347
1348         vm_unmap_aliases();
1349
1350         cpa.vaddr = addr;
1351         cpa.pages = pages;
1352         cpa.numpages = numpages;
1353         cpa.mask_set = mask_set;
1354         cpa.mask_clr = mask_clr;
1355         cpa.flags = 0;
1356         cpa.curpage = 0;
1357         cpa.force_split = force_split;
1358
1359         if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
1360                 cpa.flags |= in_flag;
1361
1362         /* No alias checking for _NX bit modifications */
1363         checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
1364
1365         ret = __change_page_attr_set_clr(&cpa, checkalias);
1366
1367         /*
1368          * Check whether we really changed something:
1369          */
1370         if (!(cpa.flags & CPA_FLUSHTLB))
1371                 goto out;
1372
1373         /*
1374          * No need to flush, when we did not set any of the caching
1375          * attributes:
1376          */
1377         cache = cache_attr(mask_set);
1378
1379         /*
1380          * On success we use clflush, when the CPU supports it to
1381          * avoid the wbindv. If the CPU does not support it and in the
1382          * error case we fall back to cpa_flush_all (which uses
1383          * wbindv):
1384          */
1385         if (!ret && cpu_has_clflush) {
1386                 if (cpa.flags & (CPA_PAGES_ARRAY | CPA_ARRAY)) {
1387                         cpa_flush_array(addr, numpages, cache,
1388                                         cpa.flags, pages);
1389                 } else
1390                         cpa_flush_range(baddr, numpages, cache);
1391         } else
1392                 cpa_flush_all(cache);
1393
1394 out:
1395         return ret;
1396 }
1397
1398 static inline int change_page_attr_set(unsigned long *addr, int numpages,
1399                                        pgprot_t mask, int array)
1400 {
1401         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
1402                 (array ? CPA_ARRAY : 0), NULL);
1403 }
1404
1405 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
1406                                          pgprot_t mask, int array)
1407 {
1408         return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
1409                 (array ? CPA_ARRAY : 0), NULL);
1410 }
1411
1412 static inline int cpa_set_pages_array(struct page **pages, int numpages,
1413                                        pgprot_t mask)
1414 {
1415         return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
1416                 CPA_PAGES_ARRAY, pages);
1417 }
1418
1419 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1420                                          pgprot_t mask)
1421 {
1422         return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1423                 CPA_PAGES_ARRAY, pages);
1424 }
1425
1426 int _set_memory_uc(unsigned long addr, int numpages)
1427 {
1428         /*
1429          * for now UC MINUS. see comments in ioremap_nocache()
1430          */
1431         return change_page_attr_set(&addr, numpages,
1432                                     __pgprot(_PAGE_CACHE_UC_MINUS), 0);
1433 }
1434
1435 int set_memory_uc(unsigned long addr, int numpages)
1436 {
1437         int ret;
1438
1439         /*
1440          * for now UC MINUS. see comments in ioremap_nocache()
1441          */
1442         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1443                             _PAGE_CACHE_UC_MINUS, NULL);
1444         if (ret)
1445                 goto out_err;
1446
1447         ret = _set_memory_uc(addr, numpages);
1448         if (ret)
1449                 goto out_free;
1450
1451         return 0;
1452
1453 out_free:
1454         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1455 out_err:
1456         return ret;
1457 }
1458 EXPORT_SYMBOL(set_memory_uc);
1459
1460 static int _set_memory_array(unsigned long *addr, int addrinarray,
1461                 unsigned long new_type)
1462 {
1463         int i, j;
1464         int ret;
1465
1466         /*
1467          * for now UC MINUS. see comments in ioremap_nocache()
1468          */
1469         for (i = 0; i < addrinarray; i++) {
1470                 ret = reserve_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE,
1471                                         new_type, NULL);
1472                 if (ret)
1473                         goto out_free;
1474         }
1475
1476         ret = change_page_attr_set(addr, addrinarray,
1477                                     __pgprot(_PAGE_CACHE_UC_MINUS), 1);
1478
1479         if (!ret && new_type == _PAGE_CACHE_WC)
1480                 ret = change_page_attr_set_clr(addr, addrinarray,
1481                                                __pgprot(_PAGE_CACHE_WC),
1482                                                __pgprot(_PAGE_CACHE_MASK),
1483                                                0, CPA_ARRAY, NULL);
1484         if (ret)
1485                 goto out_free;
1486
1487         return 0;
1488
1489 out_free:
1490         for (j = 0; j < i; j++)
1491                 free_memtype(__pa(addr[j]), __pa(addr[j]) + PAGE_SIZE);
1492
1493         return ret;
1494 }
1495
1496 int set_memory_array_uc(unsigned long *addr, int addrinarray)
1497 {
1498         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_UC_MINUS);
1499 }
1500 EXPORT_SYMBOL(set_memory_array_uc);
1501
1502 int set_memory_array_wc(unsigned long *addr, int addrinarray)
1503 {
1504         return _set_memory_array(addr, addrinarray, _PAGE_CACHE_WC);
1505 }
1506 EXPORT_SYMBOL(set_memory_array_wc);
1507
1508 int _set_memory_wc(unsigned long addr, int numpages)
1509 {
1510         int ret;
1511         unsigned long addr_copy = addr;
1512
1513         ret = change_page_attr_set(&addr, numpages,
1514                                     __pgprot(_PAGE_CACHE_UC_MINUS), 0);
1515         if (!ret) {
1516                 ret = change_page_attr_set_clr(&addr_copy, numpages,
1517                                                __pgprot(_PAGE_CACHE_WC),
1518                                                __pgprot(_PAGE_CACHE_MASK),
1519                                                0, 0, NULL);
1520         }
1521         return ret;
1522 }
1523
1524 int set_memory_wc(unsigned long addr, int numpages)
1525 {
1526         int ret;
1527
1528         if (!pat_enabled)
1529                 return set_memory_uc(addr, numpages);
1530
1531         ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1532                 _PAGE_CACHE_WC, NULL);
1533         if (ret)
1534                 goto out_err;
1535
1536         ret = _set_memory_wc(addr, numpages);
1537         if (ret)
1538                 goto out_free;
1539
1540         return 0;
1541
1542 out_free:
1543         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1544 out_err:
1545         return ret;
1546 }
1547 EXPORT_SYMBOL(set_memory_wc);
1548
1549 int _set_memory_wb(unsigned long addr, int numpages)
1550 {
1551         return change_page_attr_clear(&addr, numpages,
1552                                       __pgprot(_PAGE_CACHE_MASK), 0);
1553 }
1554
1555 int set_memory_wb(unsigned long addr, int numpages)
1556 {
1557         int ret;
1558
1559         ret = _set_memory_wb(addr, numpages);
1560         if (ret)
1561                 return ret;
1562
1563         free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1564         return 0;
1565 }
1566 EXPORT_SYMBOL(set_memory_wb);
1567
1568 int set_memory_array_wb(unsigned long *addr, int addrinarray)
1569 {
1570         int i;
1571         int ret;
1572
1573         ret = change_page_attr_clear(addr, addrinarray,
1574                                       __pgprot(_PAGE_CACHE_MASK), 1);
1575         if (ret)
1576                 return ret;
1577
1578         for (i = 0; i < addrinarray; i++)
1579                 free_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE);
1580
1581         return 0;
1582 }
1583 EXPORT_SYMBOL(set_memory_array_wb);
1584
1585 int set_memory_x(unsigned long addr, int numpages)
1586 {
1587         if (!(__supported_pte_mask & _PAGE_NX))
1588                 return 0;
1589
1590         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1591 }
1592 EXPORT_SYMBOL(set_memory_x);
1593
1594 int set_memory_nx(unsigned long addr, int numpages)
1595 {
1596         if (!(__supported_pte_mask & _PAGE_NX))
1597                 return 0;
1598
1599         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1600 }
1601 EXPORT_SYMBOL(set_memory_nx);
1602
1603 int set_memory_ro(unsigned long addr, int numpages)
1604 {
1605         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1606 }
1607 EXPORT_SYMBOL_GPL(set_memory_ro);
1608
1609 int set_memory_rw(unsigned long addr, int numpages)
1610 {
1611         return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1612 }
1613 EXPORT_SYMBOL_GPL(set_memory_rw);
1614
1615 int set_memory_np(unsigned long addr, int numpages)
1616 {
1617         return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1618 }
1619
1620 int set_memory_4k(unsigned long addr, int numpages)
1621 {
1622         return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1623                                         __pgprot(0), 1, 0, NULL);
1624 }
1625
1626 int set_pages_uc(struct page *page, int numpages)
1627 {
1628         unsigned long addr = (unsigned long)page_address(page);
1629
1630         return set_memory_uc(addr, numpages);
1631 }
1632 EXPORT_SYMBOL(set_pages_uc);
1633
1634 static int _set_pages_array(struct page **pages, int addrinarray,
1635                 unsigned long new_type)
1636 {
1637         unsigned long start;
1638         unsigned long end;
1639         int i;
1640         int free_idx;
1641         int ret;
1642
1643         for (i = 0; i < addrinarray; i++) {
1644                 if (PageHighMem(pages[i]))
1645                         continue;
1646                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1647                 end = start + PAGE_SIZE;
1648                 if (reserve_memtype(start, end, new_type, NULL))
1649                         goto err_out;
1650         }
1651
1652         ret = cpa_set_pages_array(pages, addrinarray,
1653                         __pgprot(_PAGE_CACHE_UC_MINUS));
1654         if (!ret && new_type == _PAGE_CACHE_WC)
1655                 ret = change_page_attr_set_clr(NULL, addrinarray,
1656                                                __pgprot(_PAGE_CACHE_WC),
1657                                                __pgprot(_PAGE_CACHE_MASK),
1658                                                0, CPA_PAGES_ARRAY, pages);
1659         if (ret)
1660                 goto err_out;
1661         return 0; /* Success */
1662 err_out:
1663         free_idx = i;
1664         for (i = 0; i < free_idx; i++) {
1665                 if (PageHighMem(pages[i]))
1666                         continue;
1667                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1668                 end = start + PAGE_SIZE;
1669                 free_memtype(start, end);
1670         }
1671         return -EINVAL;
1672 }
1673
1674 int set_pages_array_uc(struct page **pages, int addrinarray)
1675 {
1676         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_UC_MINUS);
1677 }
1678 EXPORT_SYMBOL(set_pages_array_uc);
1679
1680 int set_pages_array_wc(struct page **pages, int addrinarray)
1681 {
1682         return _set_pages_array(pages, addrinarray, _PAGE_CACHE_WC);
1683 }
1684 EXPORT_SYMBOL(set_pages_array_wc);
1685
1686 int set_pages_wb(struct page *page, int numpages)
1687 {
1688         unsigned long addr = (unsigned long)page_address(page);
1689
1690         return set_memory_wb(addr, numpages);
1691 }
1692 EXPORT_SYMBOL(set_pages_wb);
1693
1694 int set_pages_array_wb(struct page **pages, int addrinarray)
1695 {
1696         int retval;
1697         unsigned long start;
1698         unsigned long end;
1699         int i;
1700
1701         retval = cpa_clear_pages_array(pages, addrinarray,
1702                         __pgprot(_PAGE_CACHE_MASK));
1703         if (retval)
1704                 return retval;
1705
1706         for (i = 0; i < addrinarray; i++) {
1707                 if (PageHighMem(pages[i]))
1708                         continue;
1709                 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1710                 end = start + PAGE_SIZE;
1711                 free_memtype(start, end);
1712         }
1713
1714         return 0;
1715 }
1716 EXPORT_SYMBOL(set_pages_array_wb);
1717
1718 int set_pages_x(struct page *page, int numpages)
1719 {
1720         unsigned long addr = (unsigned long)page_address(page);
1721
1722         return set_memory_x(addr, numpages);
1723 }
1724 EXPORT_SYMBOL(set_pages_x);
1725
1726 int set_pages_nx(struct page *page, int numpages)
1727 {
1728         unsigned long addr = (unsigned long)page_address(page);
1729
1730         return set_memory_nx(addr, numpages);
1731 }
1732 EXPORT_SYMBOL(set_pages_nx);
1733
1734 int set_pages_ro(struct page *page, int numpages)
1735 {
1736         unsigned long addr = (unsigned long)page_address(page);
1737
1738         return set_memory_ro(addr, numpages);
1739 }
1740
1741 int set_pages_rw(struct page *page, int numpages)
1742 {
1743         unsigned long addr = (unsigned long)page_address(page);
1744
1745         return set_memory_rw(addr, numpages);
1746 }
1747
1748 #ifdef CONFIG_DEBUG_PAGEALLOC
1749
1750 static int __set_pages_p(struct page *page, int numpages)
1751 {
1752         unsigned long tempaddr = (unsigned long) page_address(page);
1753         struct cpa_data cpa = { .vaddr = &tempaddr,
1754                                 .pgd = NULL,
1755                                 .numpages = numpages,
1756                                 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1757                                 .mask_clr = __pgprot(0),
1758                                 .flags = 0};
1759
1760         /*
1761          * No alias checking needed for setting present flag. otherwise,
1762          * we may need to break large pages for 64-bit kernel text
1763          * mappings (this adds to complexity if we want to do this from
1764          * atomic context especially). Let's keep it simple!
1765          */
1766         return __change_page_attr_set_clr(&cpa, 0);
1767 }
1768
1769 static int __set_pages_np(struct page *page, int numpages)
1770 {
1771         unsigned long tempaddr = (unsigned long) page_address(page);
1772         struct cpa_data cpa = { .vaddr = &tempaddr,
1773                                 .pgd = NULL,
1774                                 .numpages = numpages,
1775                                 .mask_set = __pgprot(0),
1776                                 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1777                                 .flags = 0};
1778
1779         /*
1780          * No alias checking needed for setting not present flag. otherwise,
1781          * we may need to break large pages for 64-bit kernel text
1782          * mappings (this adds to complexity if we want to do this from
1783          * atomic context especially). Let's keep it simple!
1784          */
1785         return __change_page_attr_set_clr(&cpa, 0);
1786 }
1787
1788 void kernel_map_pages(struct page *page, int numpages, int enable)
1789 {
1790         if (PageHighMem(page))
1791                 return;
1792         if (!enable) {
1793                 debug_check_no_locks_freed(page_address(page),
1794                                            numpages * PAGE_SIZE);
1795         }
1796
1797         /*
1798          * The return value is ignored as the calls cannot fail.
1799          * Large pages for identity mappings are not used at boot time
1800          * and hence no memory allocations during large page split.
1801          */
1802         if (enable)
1803                 __set_pages_p(page, numpages);
1804         else
1805                 __set_pages_np(page, numpages);
1806
1807         /*
1808          * We should perform an IPI and flush all tlbs,
1809          * but that can deadlock->flush only current cpu:
1810          */
1811         __flush_tlb_all();
1812
1813         arch_flush_lazy_mmu_mode();
1814 }
1815
1816 #ifdef CONFIG_HIBERNATION
1817
1818 bool kernel_page_present(struct page *page)
1819 {
1820         unsigned int level;
1821         pte_t *pte;
1822
1823         if (PageHighMem(page))
1824                 return false;
1825
1826         pte = lookup_address((unsigned long)page_address(page), &level);
1827         return (pte_val(*pte) & _PAGE_PRESENT);
1828 }
1829
1830 #endif /* CONFIG_HIBERNATION */
1831
1832 #endif /* CONFIG_DEBUG_PAGEALLOC */
1833
1834 int kernel_map_pages_in_pgd(pgd_t *pgd, u64 pfn, unsigned long address,
1835                             unsigned numpages, unsigned long page_flags)
1836 {
1837         int retval = -EINVAL;
1838
1839         struct cpa_data cpa = {
1840                 .vaddr = &address,
1841                 .pfn = pfn,
1842                 .pgd = pgd,
1843                 .numpages = numpages,
1844                 .mask_set = __pgprot(0),
1845                 .mask_clr = __pgprot(0),
1846                 .flags = 0,
1847         };
1848
1849         if (!(__supported_pte_mask & _PAGE_NX))
1850                 goto out;
1851
1852         if (!(page_flags & _PAGE_NX))
1853                 cpa.mask_clr = __pgprot(_PAGE_NX);
1854
1855         cpa.mask_set = __pgprot(_PAGE_PRESENT | page_flags);
1856
1857         retval = __change_page_attr_set_clr(&cpa, 0);
1858         __flush_tlb_all();
1859
1860 out:
1861         return retval;
1862 }
1863
1864 /*
1865  * The testcases use internal knowledge of the implementation that shouldn't
1866  * be exposed to the rest of the kernel. Include these directly here.
1867  */
1868 #ifdef CONFIG_CPA_DEBUG
1869 #include "pageattr-test.c"
1870 #endif