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/export.h>
   9 #include <linux/backing-dev.h>
  10 #include <linux/mm.h>
  11 #include <linux/swap.h>
  12 #include <linux/file.h>
  13 #include <linux/mman.h>
  14 #include <linux/pagemap.h>
  15 #include <linux/swapops.h>
  16 #include <linux/rmap.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, MM_FILEPAGES);
  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 int generic_file_remap_pages(struct vm_area_struct *vma, unsigned long addr,
  85                              unsigned long size, pgoff_t pgoff)
  86 {
  87         struct mm_struct *mm = vma->vm_mm;
  88         int err;
  89
  90         do {
  91                 err = install_file_pte(mm, vma, addr, pgoff, vma->vm_page_prot);
  92                 if (err)
  93                         return err;
  94
  95                 size -= PAGE_SIZE;
  96                 addr += PAGE_SIZE;
  97                 pgoff++;
  98         } while (size);
  99
 100         return 0;
 101 }
 102 EXPORT_SYMBOL(generic_file_remap_pages);
 103
 104 /**
 105  * sys_remap_file_pages - remap arbitrary pages of an existing VM_SHARED vma
 106  * @start: start of the remapped virtual memory range
 107  * @size: size of the remapped virtual memory range
 108  * @prot: new protection bits of the range (see NOTE)
 109  * @pgoff: to-be-mapped page of the backing store file
 110  * @flags: 0 or MAP_NONBLOCKED - the later will cause no IO.
 111  *
 112  * sys_remap_file_pages remaps arbitrary pages of an existing VM_SHARED vma
 113  * (shared backing store file).
 114  *
 115  * This syscall works purely via pagetables, so it's the most efficient
 116  * way to map the same (large) file into a given virtual window. Unlike
 117  * mmap()/mremap() it does not create any new vmas. The new mappings are
 118  * also safe across swapout.
 119  *
 120  * NOTE: the @prot parameter right now is ignored (but must be zero),
 121  * and the vma's default protection is used. Arbitrary protections
 122  * might be implemented in the future.
 123  */
 124 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
 125                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
 126 {
 127         struct mm_struct *mm = current->mm;
 128         struct address_space *mapping;
 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         /* Does pgoff wrap? */
 146         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
 147                 return err;
 148
 149         /* Can we represent this offset inside this architecture's pte's? */
 150 #if PTE_FILE_MAX_BITS < BITS_PER_LONG
 151         if (pgoff + (size >> PAGE_SHIFT) >= (1UL << PTE_FILE_MAX_BITS))
 152                 return err;
 153 #endif
 154
 155         /* We need down_write() to change vma->vm_flags. */
 156         down_read(&mm->mmap_sem);
 157  retry:
 158         vma = find_vma(mm, start);
 159
 160         /*
 161          * Make sure the vma is shared, that it supports prefaulting,
 162          * and that the remapped range is valid and fully within
 163          * the single existing vma.
 164          */
 165         if (!vma || !(vma->vm_flags & VM_SHARED))
 166                 goto out;
 167
 168         if (!vma->vm_ops || !vma->vm_ops->remap_pages)
 169                 goto out;
 170
 171         if (start < vma->vm_start || start + size > 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                 /*
 177                  * vm_private_data is used as a swapout cursor
 178                  * in a VM_NONLINEAR vma.
 179                  */
 180                 if (vma->vm_private_data)
 181                         goto out;
 182
 183                 /* Don't need a nonlinear mapping, exit success */
 184                 if (pgoff == linear_page_index(vma, start)) {
 185                         err = 0;
 186                         goto out;
 187                 }
 188
 189                 if (!has_write_lock) {
 190 get_write_lock:
 191                         up_read(&mm->mmap_sem);
 192                         down_write(&mm->mmap_sem);
 193                         has_write_lock = 1;
 194                         goto retry;
 195                 }
 196                 mapping = vma->vm_file->f_mapping;
 197                 /*
 198                  * page_mkclean doesn't work on nonlinear vmas, so if
 199                  * dirty pages need to be accounted, emulate with linear
 200                  * vmas.
 201                  */
 202                 if (mapping_cap_account_dirty(mapping)) {
 203                         unsigned long addr;
 204                         struct file *file = get_file(vma->vm_file);
 205
 206                         flags = (flags & MAP_NONBLOCK) | MAP_POPULATE;
 207                         addr = mmap_region(file, start, size,
 208                                         flags, vma->vm_flags, pgoff);
 209                         fput(file);
 210                         if (IS_ERR_VALUE(addr)) {
 211                                 err = addr;
 212                         } else {
 213                                 BUG_ON(addr != start);
 214                                 err = 0;
 215                         }
 216                         goto out;
 217                 }
 218                 mutex_lock(&mapping->i_mmap_mutex);
 219                 flush_dcache_mmap_lock(mapping);
 220                 vma->vm_flags |= VM_NONLINEAR;
 221                 vma_interval_tree_remove(vma, &mapping->i_mmap);
 222                 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
 223                 flush_dcache_mmap_unlock(mapping);
 224                 mutex_unlock(&mapping->i_mmap_mutex);
 225         }
 226
 227         if (vma->vm_flags & VM_LOCKED) {
 228                 /*
 229                  * drop PG_Mlocked flag for over-mapped range
 230                  */
 231                 vm_flags_t saved_flags = vma->vm_flags;
 232                 if (!has_write_lock)
 233                         goto get_write_lock;
 234                 munlock_vma_pages_range(vma, start, start + size);
 235                 vma->vm_flags = saved_flags;
 236         }
 237
 238         mmu_notifier_invalidate_range_start(mm, start, start + size);
 239         err = vma->vm_ops->remap_pages(vma, start, size, pgoff);
 240         mmu_notifier_invalidate_range_end(mm, start, start + size);
 241         if (!err) {
 242                 if (vma->vm_flags & VM_LOCKED) {
 243                         /*
 244                          * might be mapping previously unmapped range of file
 245                          */
 246                         mlock_vma_pages_range(vma, start, start + size);
 247                 } else if (!(flags & MAP_NONBLOCK)) {
 248                         if (unlikely(has_write_lock)) {
 249                                 downgrade_write(&mm->mmap_sem);
 250                                 has_write_lock = 0;
 251                         }
 252                         make_pages_present(start, start+size);
 253                 }
 254         }
 255
 256         /*
 257          * We can't clear VM_NONLINEAR because we'd have to do
 258          * it after ->populate completes, and that would prevent
 259          * downgrading the lock.  (Locks can't be upgraded).
 260          */
 261
 262 out:
 263         if (likely(!has_write_lock))
 264                 up_read(&mm->mmap_sem);
 265         else
 266                 up_write(&mm->mmap_sem);
 267
 268         return err;
 269 }