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 #include <linux/backing-dev.h>
   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.h>
  16 #include <linux/module.h>
  17 #include <linux/syscalls.h>
  18 #include <linux/mmu_notifier.h>
  19
  20 #include <asm/mmu_context.h>
  21 #include <asm/cacheflush.h>
  22 #include <asm/tlbflush.h>
  23
  24 #include "internal.h"
  25
  26 static void zap_pte(struct mm_struct *mm, struct vm_area_struct *vma,
  27                         unsigned long addr, pte_t *ptep)
  28 {
  29         pte_t pte = *ptep;
  30
  31         if (pte_present(pte)) {
  32                 struct page *page;
  33
  34                 flush_cache_page(vma, addr, pte_pfn(pte));
  35                 pte = ptep_clear_flush(vma, addr, ptep);
  36                 page = vm_normal_page(vma, addr, pte);
  37                 if (page) {
  38                         if (pte_dirty(pte))
  39                                 set_page_dirty(page);
  40                         page_remove_rmap(page);
  41                         page_cache_release(page);
  42                         update_hiwater_rss(mm);
  43                         dec_mm_counter(mm, file_rss);
  44                 }
  45         } else {
  46                 if (!pte_file(pte))
  47                         free_swap_and_cache(pte_to_swp_entry(pte));
  48                 pte_clear_not_present_full(mm, addr, ptep, 0);
  49         }
  50 }
  51
  52 /*
  53  * Install a file pte to a given virtual memory address, release any
  54  * previously existing mapping.
  55  */
  56 static int install_file_pte(struct mm_struct *mm, struct vm_area_struct *vma,
  57                 unsigned long addr, unsigned long pgoff, pgprot_t prot)
  58 {
  59         int err = -ENOMEM;
  60         pte_t *pte;
  61         spinlock_t *ptl;
  62
  63         pte = get_locked_pte(mm, addr, &ptl);
  64         if (!pte)
  65                 goto out;
  66
  67         if (!pte_none(*pte))
  68                 zap_pte(mm, vma, addr, pte);
  69
  70         set_pte_at(mm, addr, pte, pgoff_to_pte(pgoff));
  71         /*
  72          * We don't need to run update_mmu_cache() here because the "file pte"
  73          * being installed by install_file_pte() is not a real pte - it's a
  74          * non-present entry (like a swap entry), noting what file offset should
  75          * be mapped there when there's a fault (in a non-linear vma where
  76          * that's not obvious).
  77          */
  78         pte_unmap_unlock(pte, ptl);
  79         err = 0;
  80 out:
  81         return err;
  82 }
  83
  84 static int populate_range(struct mm_struct *mm, struct vm_area_struct *vma,
  85                         unsigned long addr, unsigned long size, pgoff_t pgoff)
  86 {
  87         int err;
  88
  89         do {
  90                 err = install_file_pte(mm, vma, addr, pgoff, vma->vm_page_prot);
  91                 if (err)
  92                         return err;
  93
  94                 size -= PAGE_SIZE;
  95                 addr += PAGE_SIZE;
  96                 pgoff++;
  97         } while (size);
  98
  99         return 0;
 100
 101 }
 102
 103 /**
 104  * sys_remap_file_pages - remap arbitrary pages of an existing VM_SHARED vma
 105  * @start: start of the remapped virtual memory range
 106  * @size: size of the remapped virtual memory range
 107  * @prot: new protection bits of the range (see NOTE)
 108  * @pgoff: to-be-mapped page of the backing store file
 109  * @flags: 0 or MAP_NONBLOCKED - the later will cause no IO.
 110  *
 111  * sys_remap_file_pages remaps arbitrary pages of an existing VM_SHARED vma
 112  * (shared backing store file).
 113  *
 114  * This syscall works purely via pagetables, so it's the most efficient
 115  * way to map the same (large) file into a given virtual window. Unlike
 116  * mmap()/mremap() it does not create any new vmas. The new mappings are
 117  * also safe across swapout.
 118  *
 119  * NOTE: the @prot parameter right now is ignored (but must be zero),
 120  * and the vma's default protection is used. Arbitrary protections
 121  * might be implemented in the future.
 122  */
 123 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
 124                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
 125 {
 126         struct mm_struct *mm = current->mm;
 127         struct address_space *mapping;
 128         unsigned long end = start + size;
 129         struct vm_area_struct *vma;
 130         int err = -EINVAL;
 131         int has_write_lock = 0;
 132
 133         if (prot)
 134                 return err;
 135         /*
 136          * Sanitize the syscall parameters:
 137          */
 138         start = start & PAGE_MASK;
 139         size = size & PAGE_MASK;
 140
 141         /* Does the address range wrap, or is the span zero-sized? */
 142         if (start + size <= start)
 143                 return err;
 144
 145         /* Can we represent this offset inside this architecture's pte's? */
 146 #if PTE_FILE_MAX_BITS < BITS_PER_LONG
 147         if (pgoff + (size >> PAGE_SHIFT) >= (1UL << PTE_FILE_MAX_BITS))
 148                 return err;
 149 #endif
 150
 151         /* We need down_write() to change vma->vm_flags. */
 152         down_read(&mm->mmap_sem);
 153  retry:
 154         vma = find_vma(mm, start);
 155
 156         /*
 157          * Make sure the vma is shared, that it supports prefaulting,
 158          * and that the remapped range is valid and fully within
 159          * the single existing vma.  vm_private_data is used as a
 160          * swapout cursor in a VM_NONLINEAR vma.
 161          */
 162         if (!vma || !(vma->vm_flags & VM_SHARED))
 163                 goto out;
 164
 165         if (vma->vm_private_data && !(vma->vm_flags & VM_NONLINEAR))
 166                 goto out;
 167
 168         if (!(vma->vm_flags & VM_CAN_NONLINEAR))
 169                 goto out;
 170
 171         if (end <= start || start < vma->vm_start || end > vma->vm_end)
 172                 goto out;
 173
 174         /* Must set VM_NONLINEAR before any pages are populated. */
 175         if (!(vma->vm_flags & VM_NONLINEAR)) {
 176                 /* Don't need a nonlinear mapping, exit success */
 177                 if (pgoff == linear_page_index(vma, start)) {
 178                         err = 0;
 179                         goto out;
 180                 }
 181
 182                 if (!has_write_lock) {
 183                         up_read(&mm->mmap_sem);
 184                         down_write(&mm->mmap_sem);
 185                         has_write_lock = 1;
 186                         goto retry;
 187                 }
 188                 mapping = vma->vm_file->f_mapping;
 189                 /*
 190                  * page_mkclean doesn't work on nonlinear vmas, so if
 191                  * dirty pages need to be accounted, emulate with linear
 192                  * vmas.
 193                  */
 194                 if (mapping_cap_account_dirty(mapping)) {
 195                         unsigned long addr;
 196                         struct file *file = vma->vm_file;
 197
 198                         flags &= MAP_NONBLOCK;
 199                         get_file(file);
 200                         addr = mmap_region(file, start, size,
 201                                         flags, vma->vm_flags, pgoff, 1);
 202                         fput(file);
 203                         if (IS_ERR_VALUE(addr)) {
 204                                 err = addr;
 205                         } else {
 206                                 BUG_ON(addr != start);
 207                                 err = 0;
 208                         }
 209                         goto out;
 210                 }
 211                 spin_lock(&mapping->i_mmap_lock);
 212                 flush_dcache_mmap_lock(mapping);
 213                 vma->vm_flags |= VM_NONLINEAR;
 214                 vma_prio_tree_remove(vma, &mapping->i_mmap);
 215                 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
 216                 flush_dcache_mmap_unlock(mapping);
 217                 spin_unlock(&mapping->i_mmap_lock);
 218         }
 219
 220         if (vma->vm_flags & VM_LOCKED) {
 221                 /*
 222                  * drop PG_Mlocked flag for over-mapped range
 223                  */
 224                 unsigned int saved_flags = vma->vm_flags;
 225                 munlock_vma_pages_range(vma, start, start + size);
 226                 vma->vm_flags = saved_flags;
 227         }
 228
 229         mmu_notifier_invalidate_range_start(mm, start, start + size);
 230         err = populate_range(mm, vma, start, size, pgoff);
 231         mmu_notifier_invalidate_range_end(mm, start, start + size);
 232         if (!err && !(flags & MAP_NONBLOCK)) {
 233                 if (vma->vm_flags & VM_LOCKED) {
 234                         /*
 235                          * might be mapping previously unmapped range of file
 236                          */
 237                         mlock_vma_pages_range(vma, start, start + size);
 238                 } else {
 239                         if (unlikely(has_write_lock)) {
 240                                 downgrade_write(&mm->mmap_sem);
 241                                 has_write_lock = 0;
 242                         }
 243                         make_pages_present(start, start+size);
 244                 }
 245         }
 246
 247         /*
 248          * We can't clear VM_NONLINEAR because we'd have to do
 249          * it after ->populate completes, and that would prevent
 250          * downgrading the lock.  (Locks can't be upgraded).
 251          */
 252
 253 out:
 254         if (likely(!has_write_lock))
 255                 up_read(&mm->mmap_sem);
 256         else
 257                 up_write(&mm->mmap_sem);
 258
 259         return err;
 260 }