linux/arch/x86/mm/pageattr.c
<<
>>
Prefs
   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 */
  31struct 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 */
  50static 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
  57static unsigned long direct_pages_count[PG_LEVEL_NUM];
  58
  59void 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
  67static void split_page_count(int level)
  68{
  69        direct_pages_count[level]--;
  70        direct_pages_count[level - 1] += PTRS_PER_PTE;
  71}
  72
  73void 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
  91static inline void split_page_count(int level) { }
  92#endif
  93
  94#ifdef CONFIG_X86_64
  95
  96static inline unsigned long highmap_start_pfn(void)
  97{
  98        return __pa_symbol(_text) >> PAGE_SHIFT;
  99}
 100
 101static 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
 114static inline int
 115within(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 */
 132void 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}
 147EXPORT_SYMBOL_GPL(clflush_cache_range);
 148
 149static 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
 163static 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
 170static 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
 180static 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
 210static 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 */
 254static 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
 326static 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 */
 366pte_t *lookup_address(unsigned long address, unsigned int *level)
 367{
 368        return __lookup_address_in_pgd(pgd_offset_k(address), address, level);
 369}
 370EXPORT_SYMBOL_GPL(lookup_address);
 371
 372static 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 */
 393phys_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}
 411EXPORT_SYMBOL_GPL(slow_virt_to_phys);
 412
 413/*
 414 * Set the new pmd in all the pgds we know about:
 415 */
 416static 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
 438static int
 439try_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
 555out_unlock:
 556        spin_unlock(&pgd_lock);
 557
 558        return do_split;
 559}
 560
 561static 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
 652static 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
 671static 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
 683static 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
 695static bool try_to_free_pud_page(pud_t *pud)
 696{
 697        int i;
 698
 699        for (i = 0; i < PTRS_PER_PUD; i++)
 700                if (!pud_none(pud[i]))
 701                        return false;
 702
 703        free_page((unsigned long)pud);
 704        return true;
 705}
 706
 707static bool unmap_pte_range(pmd_t *pmd, unsigned long start, unsigned long end)
 708{
 709        pte_t *pte = pte_offset_kernel(pmd, start);
 710
 711        while (start < end) {
 712                set_pte(pte, __pte(0));
 713
 714                start += PAGE_SIZE;
 715                pte++;
 716        }
 717
 718        if (try_to_free_pte_page((pte_t *)pmd_page_vaddr(*pmd))) {
 719                pmd_clear(pmd);
 720                return true;
 721        }
 722        return false;
 723}
 724
 725static void __unmap_pmd_range(pud_t *pud, pmd_t *pmd,
 726                              unsigned long start, unsigned long end)
 727{
 728        if (unmap_pte_range(pmd, start, end))
 729                if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
 730                        pud_clear(pud);
 731}
 732
 733static void unmap_pmd_range(pud_t *pud, unsigned long start, unsigned long end)
 734{
 735        pmd_t *pmd = pmd_offset(pud, start);
 736
 737        /*
 738         * Not on a 2MB page boundary?
 739         */
 740        if (start & (PMD_SIZE - 1)) {
 741                unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
 742                unsigned long pre_end = min_t(unsigned long, end, next_page);
 743
 744                __unmap_pmd_range(pud, pmd, start, pre_end);
 745
 746                start = pre_end;
 747                pmd++;
 748        }
 749
 750        /*
 751         * Try to unmap in 2M chunks.
 752         */
 753        while (end - start >= PMD_SIZE) {
 754                if (pmd_large(*pmd))
 755                        pmd_clear(pmd);
 756                else
 757                        __unmap_pmd_range(pud, pmd, start, start + PMD_SIZE);
 758
 759                start += PMD_SIZE;
 760                pmd++;
 761        }
 762
 763        /*
 764         * 4K leftovers?
 765         */
 766        if (start < end)
 767                return __unmap_pmd_range(pud, pmd, start, end);
 768
 769        /*
 770         * Try again to free the PMD page if haven't succeeded above.
 771         */
 772        if (!pud_none(*pud))
 773                if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
 774                        pud_clear(pud);
 775}
 776
 777static void unmap_pud_range(pgd_t *pgd, unsigned long start, unsigned long end)
 778{
 779        pud_t *pud = pud_offset(pgd, start);
 780
 781        /*
 782         * Not on a GB page boundary?
 783         */
 784        if (start & (PUD_SIZE - 1)) {
 785                unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
 786                unsigned long pre_end   = min_t(unsigned long, end, next_page);
 787
 788                unmap_pmd_range(pud, start, pre_end);
 789
 790                start = pre_end;
 791                pud++;
 792        }
 793
 794        /*
 795         * Try to unmap in 1G chunks?
 796         */
 797        while (end - start >= PUD_SIZE) {
 798
 799                if (pud_large(*pud))
 800                        pud_clear(pud);
 801                else
 802                        unmap_pmd_range(pud, start, start + PUD_SIZE);
 803
 804                start += PUD_SIZE;
 805                pud++;
 806        }
 807
 808        /*
 809         * 2M leftovers?
 810         */
 811        if (start < end)
 812                unmap_pmd_range(pud, start, end);
 813
 814        /*
 815         * No need to try to free the PUD page because we'll free it in
 816         * populate_pgd's error path
 817         */
 818}
 819
 820static void unmap_pgd_range(pgd_t *root, unsigned long addr, unsigned long end)
 821{
 822        pgd_t *pgd_entry = root + pgd_index(addr);
 823
 824        unmap_pud_range(pgd_entry, addr, end);
 825
 826        if (try_to_free_pud_page((pud_t *)pgd_page_vaddr(*pgd_entry)))
 827                pgd_clear(pgd_entry);
 828}
 829
 830static int alloc_pte_page(pmd_t *pmd)
 831{
 832        pte_t *pte = (pte_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
 833        if (!pte)
 834                return -1;
 835
 836        set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
 837        return 0;
 838}
 839
 840static int alloc_pmd_page(pud_t *pud)
 841{
 842        pmd_t *pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
 843        if (!pmd)
 844                return -1;
 845
 846        set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
 847        return 0;
 848}
 849
 850static void populate_pte(struct cpa_data *cpa,
 851                         unsigned long start, unsigned long end,
 852                         unsigned num_pages, pmd_t *pmd, pgprot_t pgprot)
 853{
 854        pte_t *pte;
 855
 856        pte = pte_offset_kernel(pmd, start);
 857
 858        while (num_pages-- && start < end) {
 859
 860                /* deal with the NX bit */
 861                if (!(pgprot_val(pgprot) & _PAGE_NX))
 862                        cpa->pfn &= ~_PAGE_NX;
 863
 864                set_pte(pte, pfn_pte(cpa->pfn >> PAGE_SHIFT, pgprot));
 865
 866                start    += PAGE_SIZE;
 867                cpa->pfn += PAGE_SIZE;
 868                pte++;
 869        }
 870}
 871
 872static int populate_pmd(struct cpa_data *cpa,
 873                        unsigned long start, unsigned long end,
 874                        unsigned num_pages, pud_t *pud, pgprot_t pgprot)
 875{
 876        unsigned int cur_pages = 0;
 877        pmd_t *pmd;
 878
 879        /*
 880         * Not on a 2M boundary?
 881         */
 882        if (start & (PMD_SIZE - 1)) {
 883                unsigned long pre_end = start + (num_pages << PAGE_SHIFT);
 884                unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
 885
 886                pre_end   = min_t(unsigned long, pre_end, next_page);
 887                cur_pages = (pre_end - start) >> PAGE_SHIFT;
 888                cur_pages = min_t(unsigned int, num_pages, cur_pages);
 889
 890                /*
 891                 * Need a PTE page?
 892                 */
 893                pmd = pmd_offset(pud, start);
 894                if (pmd_none(*pmd))
 895                        if (alloc_pte_page(pmd))
 896                                return -1;
 897
 898                populate_pte(cpa, start, pre_end, cur_pages, pmd, pgprot);
 899
 900                start = pre_end;
 901        }
 902
 903        /*
 904         * We mapped them all?
 905         */
 906        if (num_pages == cur_pages)
 907                return cur_pages;
 908
 909        while (end - start >= PMD_SIZE) {
 910
 911                /*
 912                 * We cannot use a 1G page so allocate a PMD page if needed.
 913                 */
 914                if (pud_none(*pud))
 915                        if (alloc_pmd_page(pud))
 916                                return -1;
 917
 918                pmd = pmd_offset(pud, start);
 919
 920                set_pmd(pmd, __pmd(cpa->pfn | _PAGE_PSE | massage_pgprot(pgprot)));
 921
 922                start     += PMD_SIZE;
 923                cpa->pfn  += PMD_SIZE;
 924                cur_pages += PMD_SIZE >> PAGE_SHIFT;
 925        }
 926
 927        /*
 928         * Map trailing 4K pages.
 929         */
 930        if (start < end) {
 931                pmd = pmd_offset(pud, start);
 932                if (pmd_none(*pmd))
 933                        if (alloc_pte_page(pmd))
 934                                return -1;
 935
 936                populate_pte(cpa, start, end, num_pages - cur_pages,
 937                             pmd, pgprot);
 938        }
 939        return num_pages;
 940}
 941
 942static int populate_pud(struct cpa_data *cpa, unsigned long start, pgd_t *pgd,
 943                        pgprot_t pgprot)
 944{
 945        pud_t *pud;
 946        unsigned long end;
 947        int cur_pages = 0;
 948
 949        end = start + (cpa->numpages << PAGE_SHIFT);
 950
 951        /*
 952         * Not on a Gb page boundary? => map everything up to it with
 953         * smaller pages.
 954         */
 955        if (start & (PUD_SIZE - 1)) {
 956                unsigned long pre_end;
 957                unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
 958
 959                pre_end   = min_t(unsigned long, end, next_page);
 960                cur_pages = (pre_end - start) >> PAGE_SHIFT;
 961                cur_pages = min_t(int, (int)cpa->numpages, cur_pages);
 962
 963                pud = pud_offset(pgd, start);
 964
 965                /*
 966                 * Need a PMD page?
 967                 */
 968                if (pud_none(*pud))
 969                        if (alloc_pmd_page(pud))
 970                                return -1;
 971
 972                cur_pages = populate_pmd(cpa, start, pre_end, cur_pages,
 973                                         pud, pgprot);
 974                if (cur_pages < 0)
 975                        return cur_pages;
 976
 977                start = pre_end;
 978        }
 979
 980        /* We mapped them all? */
 981        if (cpa->numpages == cur_pages)
 982                return cur_pages;
 983
 984        pud = pud_offset(pgd, start);
 985
 986        /*
 987         * Map everything starting from the Gb boundary, possibly with 1G pages
 988         */
 989        while (end - start >= PUD_SIZE) {
 990                set_pud(pud, __pud(cpa->pfn | _PAGE_PSE | massage_pgprot(pgprot)));
 991
 992                start     += PUD_SIZE;
 993                cpa->pfn  += PUD_SIZE;
 994                cur_pages += PUD_SIZE >> PAGE_SHIFT;
 995                pud++;
 996        }
 997
 998        /* Map trailing leftover */
 999        if (start < end) {
1000                int tmp;
1001
1002                pud = pud_offset(pgd, start);
1003                if (pud_none(*pud))
1004                        if (alloc_pmd_page(pud))
1005                                return -1;
1006
1007                tmp = populate_pmd(cpa, start, end, cpa->numpages - cur_pages,
1008                                   pud, pgprot);
1009                if (tmp < 0)
1010                        return cur_pages;
1011
1012                cur_pages += tmp;
1013        }
1014        return cur_pages;
1015}
1016
1017/*
1018 * Restrictions for kernel page table do not necessarily apply when mapping in
1019 * an alternate PGD.
1020 */
1021static int populate_pgd(struct cpa_data *cpa, unsigned long addr)
1022{
1023        pgprot_t pgprot = __pgprot(_KERNPG_TABLE);
1024        pud_t *pud = NULL;      /* shut up gcc */
1025        pgd_t *pgd_entry;
1026        int ret;
1027
1028        pgd_entry = cpa->pgd + pgd_index(addr);
1029
1030        /*
1031         * Allocate a PUD page and hand it down for mapping.
1032         */
1033        if (pgd_none(*pgd_entry)) {
1034                pud = (pud_t *)get_zeroed_page(GFP_KERNEL | __GFP_NOTRACK);
1035                if (!pud)
1036                        return -1;
1037
1038                set_pgd(pgd_entry, __pgd(__pa(pud) | _KERNPG_TABLE));
1039        }
1040
1041        pgprot_val(pgprot) &= ~pgprot_val(cpa->mask_clr);
1042        pgprot_val(pgprot) |=  pgprot_val(cpa->mask_set);
1043
1044        ret = populate_pud(cpa, addr, pgd_entry, pgprot);
1045        if (ret < 0) {
1046                unmap_pgd_range(cpa->pgd, addr,
1047                                addr + (cpa->numpages << PAGE_SHIFT));
1048                return ret;
1049        }
1050
1051        cpa->numpages = ret;
1052        return 0;
1053}
1054
1055static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
1056                               int primary)
1057{
1058        if (cpa->pgd)
1059                return populate_pgd(cpa, vaddr);
1060
1061        /*
1062         * Ignore all non primary paths.
1063         */
1064        if (!primary)
1065                return 0;
1066
1067        /*
1068         * Ignore the NULL PTE for kernel identity mapping, as it is expected
1069         * to have holes.
1070         * Also set numpages to '1' indicating that we processed cpa req for
1071         * one virtual address page and its pfn. TBD: numpages can be set based
1072         * on the initial value and the level returned by lookup_address().
1073         */
1074        if (within(vaddr, PAGE_OFFSET,
1075                   PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
1076                cpa->numpages = 1;
1077                cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
1078                return 0;
1079        } else {
1080                WARN(1, KERN_WARNING "CPA: called for zero pte. "
1081                        "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
1082                        *cpa->vaddr);
1083
1084                return -EFAULT;
1085        }
1086}
1087
1088static int __change_page_attr(struct cpa_data *cpa, int primary)
1089{
1090        unsigned long address;
1091        int do_split, err;
1092        unsigned int level;
1093        pte_t *kpte, old_pte;
1094
1095        if (cpa->flags & CPA_PAGES_ARRAY) {
1096                struct page *page = cpa->pages[cpa->curpage];
1097                if (unlikely(PageHighMem(page)))
1098                        return 0;
1099                address = (unsigned long)page_address(page);
1100        } else if (cpa->flags & CPA_ARRAY)
1101                address = cpa->vaddr[cpa->curpage];
1102        else
1103                address = *cpa->vaddr;
1104repeat:
1105        kpte = _lookup_address_cpa(cpa, address, &level);
1106        if (!kpte)
1107                return __cpa_process_fault(cpa, address, primary);
1108
1109        old_pte = *kpte;
1110        if (!pte_val(old_pte))
1111                return __cpa_process_fault(cpa, address, primary);
1112
1113        if (level == PG_LEVEL_4K) {
1114                pte_t new_pte;
1115                pgprot_t new_prot = pte_pgprot(old_pte);
1116                unsigned long pfn = pte_pfn(old_pte);
1117
1118                pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
1119                pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
1120
1121                new_prot = static_protections(new_prot, address, pfn);
1122
1123                /*
1124                 * Set the GLOBAL flags only if the PRESENT flag is
1125                 * set otherwise pte_present will return true even on
1126                 * a non present pte. The canon_pgprot will clear
1127                 * _PAGE_GLOBAL for the ancient hardware that doesn't
1128                 * support it.
1129                 */
1130                if (pgprot_val(new_prot) & _PAGE_PRESENT)
1131                        pgprot_val(new_prot) |= _PAGE_GLOBAL;
1132                else
1133                        pgprot_val(new_prot) &= ~_PAGE_GLOBAL;
1134
1135                /*
1136                 * We need to keep the pfn from the existing PTE,
1137                 * after all we're only going to change it's attributes
1138                 * not the memory it points to
1139                 */
1140                new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
1141                cpa->pfn = pfn;
1142                /*
1143                 * Do we really change anything ?
1144                 */
1145                if (pte_val(old_pte) != pte_val(new_pte)) {
1146                        set_pte_atomic(kpte, new_pte);
1147                        cpa->flags |= CPA_FLUSHTLB;
1148                }
1149                cpa->numpages = 1;
1150                return 0;
1151        }
1152
1153        /*
1154         * Check, whether we can keep the large page intact
1155         * and just change the pte:
1156         */
1157        do_split = try_preserve_large_page(kpte, address, cpa);
1158        /*
1159         * When the range fits into the existing large page,
1160         * return. cp->numpages and cpa->tlbflush have been updated in
1161         * try_large_page:
1162         */
1163        if (do_split <= 0)
1164                return do_split;
1165
1166        /*
1167         * We have to split the large page:
1168         */
1169        err = split_large_page(cpa, kpte, address);
1170        if (!err) {
1171                /*
1172                 * Do a global flush tlb after splitting the large page
1173                 * and before we do the actual change page attribute in the PTE.
1174                 *
1175                 * With out this, we violate the TLB application note, that says
1176                 * "The TLBs may contain both ordinary and large-page
1177                 *  translations for a 4-KByte range of linear addresses. This
1178                 *  may occur if software modifies the paging structures so that
1179                 *  the page size used for the address range changes. If the two
1180                 *  translations differ with respect to page frame or attributes
1181                 *  (e.g., permissions), processor behavior is undefined and may
1182                 *  be implementation-specific."
1183                 *
1184                 * We do this global tlb flush inside the cpa_lock, so that we
1185                 * don't allow any other cpu, with stale tlb entries change the
1186                 * page attribute in parallel, that also falls into the
1187                 * just split large page entry.
1188                 */
1189                flush_tlb_all();
1190                goto repeat;
1191        }
1192
1193        return err;
1194}
1195
1196static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
1197
1198static int cpa_process_alias(struct cpa_data *cpa)
1199{
1200        struct cpa_data alias_cpa;
1201        unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
1202        unsigned long vaddr;
1203        int ret;
1204
1205        if (!pfn_range_is_mapped(cpa->pfn, cpa->pfn + 1))
1206                return 0;
1207
1208        /*
1209         * No need to redo, when the primary call touched the direct
1210         * mapping already:
1211         */
1212        if (cpa->flags & CPA_PAGES_ARRAY) {
1213                struct page *page = cpa->pages[cpa->curpage];
1214                if (unlikely(PageHighMem(page)))
1215                        return 0;
1216                vaddr = (unsigned long)page_address(page);
1217        } else if (cpa->flags & CPA_ARRAY)
1218                vaddr = cpa->vaddr[cpa->curpage];
1219        else
1220                vaddr = *cpa->vaddr;
1221
1222        if (!(within(vaddr, PAGE_OFFSET,
1223                    PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
1224
1225                alias_cpa = *cpa;
1226                alias_cpa.vaddr = &laddr;
1227                alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1228
1229                ret = __change_page_attr_set_clr(&alias_cpa, 0);
1230                if (ret)
1231                        return ret;
1232        }
1233
1234#ifdef CONFIG_X86_64
1235        /*
1236         * If the primary call didn't touch the high mapping already
1237         * and the physical address is inside the kernel map, we need
1238         * to touch the high mapped kernel as well:
1239         */
1240        if (!within(vaddr, (unsigned long)_text, _brk_end) &&
1241            within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn())) {
1242                unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
1243                                               __START_KERNEL_map - phys_base;
1244                alias_cpa = *cpa;
1245                alias_cpa.vaddr = &temp_cpa_vaddr;
1246                alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1247
1248                /*
1249                 * The high mapping range is imprecise, so ignore the
1250                 * return value.
1251                 */
1252                __change_page_attr_set_clr(&alias_cpa, 0);
1253        }
1254#endif
1255
1256        return 0;
1257}
1258
1259static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
1260{
1261        int ret, numpages = cpa->numpages;
1262
1263        while (numpages) {
1264                /*
1265                 * Store the remaining nr of pages for the large page
1266                 * preservation check.
1267                 */
1268                cpa->numpages = numpages;
1269                /* for array changes, we can't use large page */
1270                if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
1271                        cpa->numpages = 1;
1272
1273                if (!debug_pagealloc)
1274                        spin_lock(&cpa_lock);
1275                ret = __change_page_attr(cpa, checkalias);
1276                if (!debug_pagealloc)
1277                        spin_unlock(&cpa_lock);
1278                if (ret)
1279                        return ret;
1280
1281                if (checkalias) {
1282                        ret = cpa_process_alias(cpa);
1283                        if (ret)
1284                                return ret;
1285                }
1286
1287                /*
1288                 * Adjust the number of pages with the result of the
1289                 * CPA operation. Either a large page has been
1290                 * preserved or a single page update happened.
1291                 */
1292                BUG_ON(cpa->numpages > numpages);
1293                numpages -= cpa->numpages;
1294                if (cpa->flags & (CPA_PAGES_ARRAY | CPA_ARRAY))
1295                        cpa->curpage++;
1296                else
1297                        *cpa->vaddr += cpa->numpages * PAGE_SIZE;
1298
1299        }
1300        return 0;
1301}
1302
1303static inline int cache_attr(pgprot_t attr)
1304{
1305        return pgprot_val(attr) &
1306                (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
1307}
1308
1309static int change_page_attr_set_clr(unsigned long *addr, int numpages,
1310                                    pgprot_t mask_set, pgprot_t mask_clr,
1311                                    int force_split, int in_flag,
1312                                    struct page **pages)
1313{
1314        struct cpa_data cpa;
1315        int ret, cache, checkalias;
1316        unsigned long baddr = 0;
1317
1318        memset(&cpa, 0, sizeof(cpa));
1319
1320        /*
1321         * Check, if we are requested to change a not supported
1322         * feature:
1323         */
1324        mask_set = canon_pgprot(mask_set);
1325        mask_clr = canon_pgprot(mask_clr);
1326        if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
1327                return 0;
1328
1329        /* Ensure we are PAGE_SIZE aligned */
1330        if (in_flag & CPA_ARRAY) {
1331                int i;
1332                for (i = 0; i < numpages; i++) {
1333                        if (addr[i] & ~PAGE_MASK) {
1334                                addr[i] &= PAGE_MASK;
1335                                WARN_ON_ONCE(1);
1336                        }
1337                }
1338        } else if (!(in_flag & CPA_PAGES_ARRAY)) {
1339                /*
1340                 * in_flag of CPA_PAGES_ARRAY implies it is aligned.
1341                 * No need to cehck in that case
1342                 */
1343                if (*addr & ~PAGE_MASK) {
1344                        *addr &= PAGE_MASK;
1345                        /*
1346                         * People should not be passing in unaligned addresses:
1347                         */
1348                        WARN_ON_ONCE(1);
1349                }
1350                /*
1351                 * Save address for cache flush. *addr is modified in the call
1352                 * to __change_page_attr_set_clr() below.
1353                 */
1354                baddr = *addr;
1355        }
1356
1357        /* Must avoid aliasing mappings in the highmem code */
1358        kmap_flush_unused();
1359
1360        vm_unmap_aliases();
1361
1362        cpa.vaddr = addr;
1363        cpa.pages = pages;
1364        cpa.numpages = numpages;
1365        cpa.mask_set = mask_set;
1366        cpa.mask_clr = mask_clr;
1367        cpa.flags = 0;
1368        cpa.curpage = 0;
1369        cpa.force_split = force_split;
1370
1371        if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
1372                cpa.flags |= in_flag;
1373
1374        /* No alias checking for _NX bit modifications */
1375        checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
1376
1377        ret = __change_page_attr_set_clr(&cpa, checkalias);
1378
1379        /*
1380         * Check whether we really changed something:
1381         */
1382        if (!(cpa.flags & CPA_FLUSHTLB))
1383                goto out;
1384
1385        /*
1386         * No need to flush, when we did not set any of the caching
1387         * attributes:
1388         */
1389        cache = cache_attr(mask_set);
1390
1391        /*
1392         * On success we use clflush, when the CPU supports it to
1393         * avoid the wbindv. If the CPU does not support it and in the
1394         * error case we fall back to cpa_flush_all (which uses
1395         * wbindv):
1396         */
1397        if (!ret && cpu_has_clflush) {
1398                if (cpa.flags & (CPA_PAGES_ARRAY | CPA_ARRAY)) {
1399                        cpa_flush_array(addr, numpages, cache,
1400                                        cpa.flags, pages);
1401                } else
1402                        cpa_flush_range(baddr, numpages, cache);
1403        } else
1404                cpa_flush_all(cache);
1405
1406out:
1407        return ret;
1408}
1409
1410static inline int change_page_attr_set(unsigned long *addr, int numpages,
1411                                       pgprot_t mask, int array)
1412{
1413        return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
1414                (array ? CPA_ARRAY : 0), NULL);
1415}
1416
1417static inline int change_page_attr_clear(unsigned long *addr, int numpages,
1418                                         pgprot_t mask, int array)
1419{
1420        return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
1421                (array ? CPA_ARRAY : 0), NULL);
1422}
1423
1424static inline int cpa_set_pages_array(struct page **pages, int numpages,
1425                                       pgprot_t mask)
1426{
1427        return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
1428                CPA_PAGES_ARRAY, pages);
1429}
1430
1431static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1432                                         pgprot_t mask)
1433{
1434        return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1435                CPA_PAGES_ARRAY, pages);
1436}
1437
1438int _set_memory_uc(unsigned long addr, int numpages)
1439{
1440        /*
1441         * for now UC MINUS. see comments in ioremap_nocache()
1442         */
1443        return change_page_attr_set(&addr, numpages,
1444                                    __pgprot(_PAGE_CACHE_UC_MINUS), 0);
1445}
1446
1447int set_memory_uc(unsigned long addr, int numpages)
1448{
1449        int ret;
1450
1451        /*
1452         * for now UC MINUS. see comments in ioremap_nocache()
1453         */
1454        ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1455                            _PAGE_CACHE_UC_MINUS, NULL);
1456        if (ret)
1457                goto out_err;
1458
1459        ret = _set_memory_uc(addr, numpages);
1460        if (ret)
1461                goto out_free;
1462
1463        return 0;
1464
1465out_free:
1466        free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1467out_err:
1468        return ret;
1469}
1470EXPORT_SYMBOL(set_memory_uc);
1471
1472static int _set_memory_array(unsigned long *addr, int addrinarray,
1473                unsigned long new_type)
1474{
1475        int i, j;
1476        int ret;
1477
1478        /*
1479         * for now UC MINUS. see comments in ioremap_nocache()
1480         */
1481        for (i = 0; i < addrinarray; i++) {
1482                ret = reserve_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE,
1483                                        new_type, NULL);
1484                if (ret)
1485                        goto out_free;
1486        }
1487
1488        ret = change_page_attr_set(addr, addrinarray,
1489                                    __pgprot(_PAGE_CACHE_UC_MINUS), 1);
1490
1491        if (!ret && new_type == _PAGE_CACHE_WC)
1492                ret = change_page_attr_set_clr(addr, addrinarray,
1493                                               __pgprot(_PAGE_CACHE_WC),
1494                                               __pgprot(_PAGE_CACHE_MASK),
1495                                               0, CPA_ARRAY, NULL);
1496        if (ret)
1497                goto out_free;
1498
1499        return 0;
1500
1501out_free:
1502        for (j = 0; j < i; j++)
1503                free_memtype(__pa(addr[j]), __pa(addr[j]) + PAGE_SIZE);
1504
1505        return ret;
1506}
1507
1508int set_memory_array_uc(unsigned long *addr, int addrinarray)
1509{
1510        return _set_memory_array(addr, addrinarray, _PAGE_CACHE_UC_MINUS);
1511}
1512EXPORT_SYMBOL(set_memory_array_uc);
1513
1514int set_memory_array_wc(unsigned long *addr, int addrinarray)
1515{
1516        return _set_memory_array(addr, addrinarray, _PAGE_CACHE_WC);
1517}
1518EXPORT_SYMBOL(set_memory_array_wc);
1519
1520int _set_memory_wc(unsigned long addr, int numpages)
1521{
1522        int ret;
1523        unsigned long addr_copy = addr;
1524
1525        ret = change_page_attr_set(&addr, numpages,
1526                                    __pgprot(_PAGE_CACHE_UC_MINUS), 0);
1527        if (!ret) {
1528                ret = change_page_attr_set_clr(&addr_copy, numpages,
1529                                               __pgprot(_PAGE_CACHE_WC),
1530                                               __pgprot(_PAGE_CACHE_MASK),
1531                                               0, 0, NULL);
1532        }
1533        return ret;
1534}
1535
1536int set_memory_wc(unsigned long addr, int numpages)
1537{
1538        int ret;
1539
1540        if (!pat_enabled)
1541                return set_memory_uc(addr, numpages);
1542
1543        ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1544                _PAGE_CACHE_WC, NULL);
1545        if (ret)
1546                goto out_err;
1547
1548        ret = _set_memory_wc(addr, numpages);
1549        if (ret)
1550                goto out_free;
1551
1552        return 0;
1553
1554out_free:
1555        free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1556out_err:
1557        return ret;
1558}
1559EXPORT_SYMBOL(set_memory_wc);
1560
1561int _set_memory_wb(unsigned long addr, int numpages)
1562{
1563        return change_page_attr_clear(&addr, numpages,
1564                                      __pgprot(_PAGE_CACHE_MASK), 0);
1565}
1566
1567int set_memory_wb(unsigned long addr, int numpages)
1568{
1569        int ret;
1570
1571        ret = _set_memory_wb(addr, numpages);
1572        if (ret)
1573                return ret;
1574
1575        free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1576        return 0;
1577}
1578EXPORT_SYMBOL(set_memory_wb);
1579
1580int set_memory_array_wb(unsigned long *addr, int addrinarray)
1581{
1582        int i;
1583        int ret;
1584
1585        ret = change_page_attr_clear(addr, addrinarray,
1586                                      __pgprot(_PAGE_CACHE_MASK), 1);
1587        if (ret)
1588                return ret;
1589
1590        for (i = 0; i < addrinarray; i++)
1591                free_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE);
1592
1593        return 0;
1594}
1595EXPORT_SYMBOL(set_memory_array_wb);
1596
1597int set_memory_x(unsigned long addr, int numpages)
1598{
1599        if (!(__supported_pte_mask & _PAGE_NX))
1600                return 0;
1601
1602        return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1603}
1604EXPORT_SYMBOL(set_memory_x);
1605
1606int set_memory_nx(unsigned long addr, int numpages)
1607{
1608        if (!(__supported_pte_mask & _PAGE_NX))
1609                return 0;
1610
1611        return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1612}
1613EXPORT_SYMBOL(set_memory_nx);
1614
1615int set_memory_ro(unsigned long addr, int numpages)
1616{
1617        return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1618}
1619EXPORT_SYMBOL_GPL(set_memory_ro);
1620
1621int set_memory_rw(unsigned long addr, int numpages)
1622{
1623        return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1624}
1625EXPORT_SYMBOL_GPL(set_memory_rw);
1626
1627int set_memory_np(unsigned long addr, int numpages)
1628{
1629        return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1630}
1631
1632int set_memory_4k(unsigned long addr, int numpages)
1633{
1634        return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1635                                        __pgprot(0), 1, 0, NULL);
1636}
1637
1638int set_pages_uc(struct page *page, int numpages)
1639{
1640        unsigned long addr = (unsigned long)page_address(page);
1641
1642        return set_memory_uc(addr, numpages);
1643}
1644EXPORT_SYMBOL(set_pages_uc);
1645
1646static int _set_pages_array(struct page **pages, int addrinarray,
1647                unsigned long new_type)
1648{
1649        unsigned long start;
1650        unsigned long end;
1651        int i;
1652        int free_idx;
1653        int ret;
1654
1655        for (i = 0; i < addrinarray; i++) {
1656                if (PageHighMem(pages[i]))
1657                        continue;
1658                start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1659                end = start + PAGE_SIZE;
1660                if (reserve_memtype(start, end, new_type, NULL))
1661                        goto err_out;
1662        }
1663
1664        ret = cpa_set_pages_array(pages, addrinarray,
1665                        __pgprot(_PAGE_CACHE_UC_MINUS));
1666        if (!ret && new_type == _PAGE_CACHE_WC)
1667                ret = change_page_attr_set_clr(NULL, addrinarray,
1668                                               __pgprot(_PAGE_CACHE_WC),
1669                                               __pgprot(_PAGE_CACHE_MASK),
1670                                               0, CPA_PAGES_ARRAY, pages);
1671        if (ret)
1672                goto err_out;
1673        return 0; /* Success */
1674err_out:
1675        free_idx = i;
1676        for (i = 0; i < free_idx; i++) {
1677                if (PageHighMem(pages[i]))
1678                        continue;
1679                start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1680                end = start + PAGE_SIZE;
1681                free_memtype(start, end);
1682        }
1683        return -EINVAL;
1684}
1685
1686int set_pages_array_uc(struct page **pages, int addrinarray)
1687{
1688        return _set_pages_array(pages, addrinarray, _PAGE_CACHE_UC_MINUS);
1689}
1690EXPORT_SYMBOL(set_pages_array_uc);
1691
1692int set_pages_array_wc(struct page **pages, int addrinarray)
1693{
1694        return _set_pages_array(pages, addrinarray, _PAGE_CACHE_WC);
1695}
1696EXPORT_SYMBOL(set_pages_array_wc);
1697
1698int set_pages_wb(struct page *page, int numpages)
1699{
1700        unsigned long addr = (unsigned long)page_address(page);
1701
1702        return set_memory_wb(addr, numpages);
1703}
1704EXPORT_SYMBOL(set_pages_wb);
1705
1706int set_pages_array_wb(struct page **pages, int addrinarray)
1707{
1708        int retval;
1709        unsigned long start;
1710        unsigned long end;
1711        int i;
1712
1713        retval = cpa_clear_pages_array(pages, addrinarray,
1714                        __pgprot(_PAGE_CACHE_MASK));
1715        if (retval)
1716                return retval;
1717
1718        for (i = 0; i < addrinarray; i++) {
1719                if (PageHighMem(pages[i]))
1720                        continue;
1721                start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1722                end = start + PAGE_SIZE;
1723                free_memtype(start, end);
1724        }
1725
1726        return 0;
1727}
1728EXPORT_SYMBOL(set_pages_array_wb);
1729
1730int set_pages_x(struct page *page, int numpages)
1731{
1732        unsigned long addr = (unsigned long)page_address(page);
1733
1734        return set_memory_x(addr, numpages);
1735}
1736EXPORT_SYMBOL(set_pages_x);
1737
1738int set_pages_nx(struct page *page, int numpages)
1739{
1740        unsigned long addr = (unsigned long)page_address(page);
1741
1742        return set_memory_nx(addr, numpages);
1743}
1744EXPORT_SYMBOL(set_pages_nx);
1745
1746int set_pages_ro(struct page *page, int numpages)
1747{
1748        unsigned long addr = (unsigned long)page_address(page);
1749
1750        return set_memory_ro(addr, numpages);
1751}
1752
1753int set_pages_rw(struct page *page, int numpages)
1754{
1755        unsigned long addr = (unsigned long)page_address(page);
1756
1757        return set_memory_rw(addr, numpages);
1758}
1759
1760#ifdef CONFIG_DEBUG_PAGEALLOC
1761
1762static int __set_pages_p(struct page *page, int numpages)
1763{
1764        unsigned long tempaddr = (unsigned long) page_address(page);
1765        struct cpa_data cpa = { .vaddr = &tempaddr,
1766                                .pgd = NULL,
1767                                .numpages = numpages,
1768                                .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1769                                .mask_clr = __pgprot(0),
1770                                .flags = 0};
1771
1772        /*
1773         * No alias checking needed for setting present flag. otherwise,
1774         * we may need to break large pages for 64-bit kernel text
1775         * mappings (this adds to complexity if we want to do this from
1776         * atomic context especially). Let's keep it simple!
1777         */
1778        return __change_page_attr_set_clr(&cpa, 0);
1779}
1780
1781static int __set_pages_np(struct page *page, int numpages)
1782{
1783        unsigned long tempaddr = (unsigned long) page_address(page);
1784        struct cpa_data cpa = { .vaddr = &tempaddr,
1785                                .pgd = NULL,
1786                                .numpages = numpages,
1787                                .mask_set = __pgprot(0),
1788                                .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1789                                .flags = 0};
1790
1791        /*
1792         * No alias checking needed for setting not present flag. otherwise,
1793         * we may need to break large pages for 64-bit kernel text
1794         * mappings (this adds to complexity if we want to do this from
1795         * atomic context especially). Let's keep it simple!
1796         */
1797        return __change_page_attr_set_clr(&cpa, 0);
1798}
1799
1800void kernel_map_pages(struct page *page, int numpages, int enable)
1801{
1802        if (PageHighMem(page))
1803                return;
1804        if (!enable) {
1805                debug_check_no_locks_freed(page_address(page),
1806                                           numpages * PAGE_SIZE);
1807        }
1808
1809        /*
1810         * The return value is ignored as the calls cannot fail.
1811         * Large pages for identity mappings are not used at boot time
1812         * and hence no memory allocations during large page split.
1813         */
1814        if (enable)
1815                __set_pages_p(page, numpages);
1816        else
1817                __set_pages_np(page, numpages);
1818
1819        /*
1820         * We should perform an IPI and flush all tlbs,
1821         * but that can deadlock->flush only current cpu:
1822         */
1823        __flush_tlb_all();
1824
1825        arch_flush_lazy_mmu_mode();
1826}
1827
1828#ifdef CONFIG_HIBERNATION
1829
1830bool kernel_page_present(struct page *page)
1831{
1832        unsigned int level;
1833        pte_t *pte;
1834
1835        if (PageHighMem(page))
1836                return false;
1837
1838        pte = lookup_address((unsigned long)page_address(page), &level);
1839        return (pte_val(*pte) & _PAGE_PRESENT);
1840}
1841
1842#endif /* CONFIG_HIBERNATION */
1843
1844#endif /* CONFIG_DEBUG_PAGEALLOC */
1845
1846int kernel_map_pages_in_pgd(pgd_t *pgd, u64 pfn, unsigned long address,
1847                            unsigned numpages, unsigned long page_flags)
1848{
1849        int retval = -EINVAL;
1850
1851        struct cpa_data cpa = {
1852                .vaddr = &address,
1853                .pfn = pfn,
1854                .pgd = pgd,
1855                .numpages = numpages,
1856                .mask_set = __pgprot(0),
1857                .mask_clr = __pgprot(0),
1858                .flags = 0,
1859        };
1860
1861        if (!(__supported_pte_mask & _PAGE_NX))
1862                goto out;
1863
1864        if (!(page_flags & _PAGE_NX))
1865                cpa.mask_clr = __pgprot(_PAGE_NX);
1866
1867        cpa.mask_set = __pgprot(_PAGE_PRESENT | page_flags);
1868
1869        retval = __change_page_attr_set_clr(&cpa, 0);
1870        __flush_tlb_all();
1871
1872out:
1873        return retval;
1874}
1875
1876void kernel_unmap_pages_in_pgd(pgd_t *root, unsigned long address,
1877                               unsigned numpages)
1878{
1879        unmap_pgd_range(root, address, address + (numpages << PAGE_SHIFT));
1880}
1881
1882/*
1883 * The testcases use internal knowledge of the implementation that shouldn't
1884 * be exposed to the rest of the kernel. Include these directly here.
1885 */
1886#ifdef CONFIG_CPA_DEBUG
1887#include "pageattr-test.c"
1888#endif
1889
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.