mm/fremap.c

   1 /*
   2  *   linux/mm/fremap.c
   3  *
   4  * Explicit pagetable population and nonlinear (random) mappings support.
   5  *
   6  * started by Ingo Molnar, Copyright (C) 2002, 2003
   7  */
   8
   9 #include <linux/mm.h>
  10 #include <linux/swap.h>
  11 #include <linux/file.h>
  12 #include <linux/mman.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/swapops.h>
  15 #include <linux/rmap-locking.h>
  16 #include <linux/module.h>
  17
  18 #include <asm/mmu_context.h>
  19 #include <asm/cacheflush.h>
  20 #include <asm/tlbflush.h>
  21
  22 static inline int zap_pte(struct mm_struct *mm, struct vm_area_struct *vma,
  23                         unsigned long addr, pte_t *ptep)
  24 {
  25         pte_t pte = *ptep;
  26
  27         if (pte_none(pte))
  28                 return 0;
  29         if (pte_present(pte)) {
  30                 unsigned long pfn = pte_pfn(pte);
  31
  32                 flush_cache_page(vma, addr);
  33                 pte = ptep_get_and_clear(ptep);
  34                 if (pfn_valid(pfn)) {
  35                         struct page *page = pfn_to_page(pfn);
  36                         if (!PageReserved(page)) {
  37                                 if (pte_dirty(pte))
  38                                         set_page_dirty(page);
  39                                 page_remove_rmap(page, ptep);
  40                                 page_cache_release(page);
  41                                 mm->rss--;
  42                         }
  43                 }
  44                 return 1;
  45         } else {
  46                 if (!pte_file(pte))
  47                         free_swap_and_cache(pte_to_swp_entry(pte));
  48                 pte_clear(ptep);
  49                 return 0;
  50         }
  51 }
  52
  53 /*
  54  * Install a page to a given virtual memory address, release any
  55  * previously existing mapping.
  56  */
  57 int install_page(struct mm_struct *mm, struct vm_area_struct *vma,
  58                 unsigned long addr, struct page *page, pgprot_t prot)
  59 {
  60         int err = -ENOMEM, flush;
  61         pte_t *pte;
  62         pgd_t *pgd;
  63         pmd_t *pmd;
  64         pte_t pte_val;
  65         struct pte_chain *pte_chain;
  66
  67         pte_chain = pte_chain_alloc(GFP_KERNEL);
  68         if (!pte_chain)
  69                 goto err;
  70         pgd = pgd_offset(mm, addr);
  71         spin_lock(&mm->page_table_lock);
  72
  73         pmd = pmd_alloc(mm, pgd, addr);
  74         if (!pmd)
  75                 goto err_unlock;
  76
  77         pte = pte_alloc_map(mm, pmd, addr);
  78         if (!pte)
  79                 goto err_unlock;
  80
  81         flush = zap_pte(mm, vma, addr, pte);
  82
  83         mm->rss++;
  84         flush_icache_page(vma, page);
  85         set_pte(pte, mk_pte(page, prot));
  86         pte_chain = page_add_rmap(page, pte, pte_chain);
  87         pte_val = *pte;
  88         pte_unmap(pte);
  89         if (flush)
  90                 flush_tlb_page(vma, addr);
  91         update_mmu_cache(vma, addr, pte_val);
  92         spin_unlock(&mm->page_table_lock);
  93         pte_chain_free(pte_chain);
  94         return 0;
  95
  96 err_unlock:
  97         spin_unlock(&mm->page_table_lock);
  98         pte_chain_free(pte_chain);
  99 err:
 100         return err;
 101 }
 102 EXPORT_SYMBOL(install_page);
 103
 104
 105 /*
 106  * Install a file pte to a given virtual memory address, release any
 107  * previously existing mapping.
 108  */
 109 int install_file_pte(struct mm_struct *mm, struct vm_area_struct *vma,
 110                 unsigned long addr, unsigned long pgoff, pgprot_t prot)
 111 {
 112         int err = -ENOMEM, flush;
 113         pte_t *pte;
 114         pgd_t *pgd;
 115         pmd_t *pmd;
 116         pte_t pte_val;
 117
 118         pgd = pgd_offset(mm, addr);
 119         spin_lock(&mm->page_table_lock);
 120
 121         pmd = pmd_alloc(mm, pgd, addr);
 122         if (!pmd)
 123                 goto err_unlock;
 124
 125         pte = pte_alloc_map(mm, pmd, addr);
 126         if (!pte)
 127                 goto err_unlock;
 128
 129         flush = zap_pte(mm, vma, addr, pte);
 130
 131         set_pte(pte, pgoff_to_pte(pgoff));
 132         pte_val = *pte;
 133         pte_unmap(pte);
 134         if (flush)
 135                 flush_tlb_page(vma, addr);
 136         update_mmu_cache(vma, addr, pte_val);
 137         spin_unlock(&mm->page_table_lock);
 138         return 0;
 139
 140 err_unlock:
 141         spin_unlock(&mm->page_table_lock);
 142         return err;
 143 }
 144
 145
 146 /***
 147  * sys_remap_file_pages - remap arbitrary pages of a shared backing store
 148  *                        file within an existing vma.
 149  * @start: start of the remapped virtual memory range
 150  * @size: size of the remapped virtual memory range
 151  * @prot: new protection bits of the range
 152  * @pgoff: to be mapped page of the backing store file
 153  * @flags: 0 or MAP_NONBLOCKED - the later will cause no IO.
 154  *
 155  * this syscall works purely via pagetables, so it's the most efficient
 156  * way to map the same (large) file into a given virtual window. Unlike
 157  * mmap()/mremap() it does not create any new vmas. The new mappings are
 158  * also safe across swapout.
 159  *
 160  * NOTE: the 'prot' parameter right now is ignored, and the vma's default
 161  * protection is used. Arbitrary protections might be implemented in the
 162  * future.
 163  */
 164 long sys_remap_file_pages(unsigned long start, unsigned long size,
 165         unsigned long __prot, unsigned long pgoff, unsigned long flags)
 166 {
 167         struct mm_struct *mm = current->mm;
 168         unsigned long end = start + size;
 169         struct vm_area_struct *vma;
 170         int err = -EINVAL;
 171
 172         if (__prot)
 173                 return err;
 174         /*
 175          * Sanitize the syscall parameters:
 176          */
 177         start = start & PAGE_MASK;
 178         size = size & PAGE_MASK;
 179
 180         /* Does the address range wrap, or is the span zero-sized? */
 181         if (start + size <= start)
 182                 return err;
 183
 184         /* Can we represent this offset inside this architecture's pte's? */
 185 #if PTE_FILE_MAX_BITS < BITS_PER_LONG
 186         if (pgoff + (size >> PAGE_SHIFT) >= (1UL << PTE_FILE_MAX_BITS))
 187                 return err;
 188 #endif
 189
 190         /* We need down_write() to change vma->vm_flags. */
 191         down_write(&mm->mmap_sem);
 192         vma = find_vma(mm, start);
 193
 194         /*
 195          * Make sure the vma is shared, that it supports prefaulting,
 196          * and that the remapped range is valid and fully within
 197          * the single existing vma:
 198          */
 199         if (vma && (vma->vm_flags & VM_SHARED) &&
 200                 vma->vm_ops && vma->vm_ops->populate &&
 201                         end > start && start >= vma->vm_start &&
 202                                 end <= vma->vm_end) {
 203
 204                 /* Must set VM_NONLINEAR before any pages are populated. */
 205                 if (pgoff != ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff)
 206                         vma->vm_flags |= VM_NONLINEAR;
 207
 208                 /* ->populate can take a long time, so downgrade the lock. */
 209                 downgrade_write(&mm->mmap_sem);
 210                 err = vma->vm_ops->populate(vma, start, size,
 211                                             vma->vm_page_prot,
 212                                             pgoff, flags & MAP_NONBLOCK);
 213
 214                 /*
 215                  * We can't clear VM_NONLINEAR because we'd have to do
 216                  * it after ->populate completes, and that would prevent
 217                  * downgrading the lock.  (Locks can't be upgraded).
 218                  */
 219                 up_read(&mm->mmap_sem);
 220         } else {
 221                 up_write(&mm->mmap_sem);
 222         }
 223
 224         return err;
 225 }
 226