mm/mprotect.c

   1 /*
   2  *  mm/mprotect.c
   3  *
   4  *  (C) Copyright 1994 Linus Torvalds
   5  *  (C) Copyright 2002 Christoph Hellwig
   6  *
   7  *  Address space accounting code       <alan@redhat.com>
   8  *  (C) Copyright 2002 Red Hat Inc, All Rights Reserved
   9  */
  10
  11 #include <linux/mm.h>
  12 #include <linux/hugetlb.h>
  13 #include <linux/slab.h>
  14 #include <linux/shm.h>
  15 #include <linux/mman.h>
  16 #include <linux/fs.h>
  17 #include <linux/highmem.h>
  18 #include <linux/security.h>
  19 #include <linux/mempolicy.h>
  20 #include <linux/personality.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/swap.h>
  23 #include <linux/swapops.h>
  24 #include <asm/uaccess.h>
  25 #include <asm/pgtable.h>
  26 #include <asm/pgalloc.h>
  27 #include <asm/cacheflush.h>
  28 #include <asm/tlbflush.h>
  29
  30 static void change_pte_range(struct mm_struct *mm, pmd_t *pmd,
  31                 unsigned long addr, unsigned long end, pgprot_t newprot,
  32                 int dirty_accountable)
  33 {
  34         pte_t *pte, oldpte;
  35         spinlock_t *ptl;
  36
  37         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
  38         arch_enter_lazy_mmu_mode();
  39         do {
  40                 oldpte = *pte;
  41                 if (pte_present(oldpte)) {
  42                         pte_t ptent;
  43
  44                         /* Avoid an SMP race with hardware updated dirty/clean
  45                          * bits by wiping the pte and then setting the new pte
  46                          * into place.
  47                          */
  48                         ptent = ptep_get_and_clear(mm, addr, pte);
  49                         ptent = pte_modify(ptent, newprot);
  50                         /*
  51                          * Avoid taking write faults for pages we know to be
  52                          * dirty.
  53                          */
  54                         if (dirty_accountable && pte_dirty(ptent))
  55                                 ptent = pte_mkwrite(ptent);
  56                         set_pte_at(mm, addr, pte, ptent);
  57                         lazy_mmu_prot_update(ptent);
  58 #ifdef CONFIG_MIGRATION
  59                 } else if (!pte_file(oldpte)) {
  60                         swp_entry_t entry = pte_to_swp_entry(oldpte);
  61
  62                         if (is_write_migration_entry(entry)) {
  63                                 /*
  64                                  * A protection check is difficult so
  65                                  * just be safe and disable write
  66                                  */
  67                                 make_migration_entry_read(&entry);
  68                                 set_pte_at(mm, addr, pte,
  69                                         swp_entry_to_pte(entry));
  70                         }
  71 #endif
  72                 }
  73
  74         } while (pte++, addr += PAGE_SIZE, addr != end);
  75         arch_leave_lazy_mmu_mode();
  76         pte_unmap_unlock(pte - 1, ptl);
  77 }
  78
  79 static inline void change_pmd_range(struct mm_struct *mm, pud_t *pud,
  80                 unsigned long addr, unsigned long end, pgprot_t newprot,
  81                 int dirty_accountable)
  82 {
  83         pmd_t *pmd;
  84         unsigned long next;
  85
  86         pmd = pmd_offset(pud, addr);
  87         do {
  88                 next = pmd_addr_end(addr, end);
  89                 if (pmd_none_or_clear_bad(pmd))
  90                         continue;
  91                 change_pte_range(mm, pmd, addr, next, newprot, dirty_accountable);
  92         } while (pmd++, addr = next, addr != end);
  93 }
  94
  95 static inline void change_pud_range(struct mm_struct *mm, pgd_t *pgd,
  96                 unsigned long addr, unsigned long end, pgprot_t newprot,
  97                 int dirty_accountable)
  98 {
  99         pud_t *pud;
 100         unsigned long next;
 101
 102         pud = pud_offset(pgd, addr);
 103         do {
 104                 next = pud_addr_end(addr, end);
 105                 if (pud_none_or_clear_bad(pud))
 106                         continue;
 107                 change_pmd_range(mm, pud, addr, next, newprot, dirty_accountable);
 108         } while (pud++, addr = next, addr != end);
 109 }
 110
 111 static void change_protection(struct vm_area_struct *vma,
 112                 unsigned long addr, unsigned long end, pgprot_t newprot,
 113                 int dirty_accountable)
 114 {
 115         struct mm_struct *mm = vma->vm_mm;
 116         pgd_t *pgd;
 117         unsigned long next;
 118         unsigned long start = addr;
 119
 120         BUG_ON(addr >= end);
 121         pgd = pgd_offset(mm, addr);
 122         flush_cache_range(vma, addr, end);
 123         do {
 124                 next = pgd_addr_end(addr, end);
 125                 if (pgd_none_or_clear_bad(pgd))
 126                         continue;
 127                 change_pud_range(mm, pgd, addr, next, newprot, dirty_accountable);
 128         } while (pgd++, addr = next, addr != end);
 129         flush_tlb_range(vma, start, end);
 130 }
 131
 132 static int
 133 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 134         unsigned long start, unsigned long end, unsigned long newflags)
 135 {
 136         struct mm_struct *mm = vma->vm_mm;
 137         unsigned long oldflags = vma->vm_flags;
 138         long nrpages = (end - start) >> PAGE_SHIFT;
 139         unsigned long charged = 0, old_end = vma->vm_end;
 140         pgoff_t pgoff;
 141         int error;
 142         int dirty_accountable = 0;
 143
 144         if (newflags == oldflags) {
 145                 *pprev = vma;
 146                 return 0;
 147         }
 148
 149         /*
 150          * If we make a private mapping writable we increase our commit;
 151          * but (without finer accounting) cannot reduce our commit if we
 152          * make it unwritable again.
 153          *
 154          * FIXME? We haven't defined a VM_NORESERVE flag, so mprotecting
 155          * a MAP_NORESERVE private mapping to writable will now reserve.
 156          */
 157         if (newflags & VM_WRITE) {
 158                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_SHARED))) {
 159                         charged = nrpages;
 160                         if (security_vm_enough_memory(charged))
 161                                 return -ENOMEM;
 162                         newflags |= VM_ACCOUNT;
 163                 }
 164         }
 165
 166         /*
 167          * First try to merge with previous and/or next vma.
 168          */
 169         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 170         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 171                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
 172         if (*pprev) {
 173                 vma = *pprev;
 174                 goto success;
 175         }
 176
 177         *pprev = vma;
 178
 179         if (start != vma->vm_start) {
 180                 error = split_vma(mm, vma, start, 1);
 181                 if (error)
 182                         goto fail;
 183         }
 184
 185         if (end != vma->vm_end) {
 186                 error = split_vma(mm, vma, end, 0);
 187                 if (error)
 188                         goto fail;
 189         }
 190
 191 success:
 192         /*
 193          * vm_flags and vm_page_prot are protected by the mmap_sem
 194          * held in write mode.
 195          */
 196         vma->vm_flags = newflags;
 197         vma->vm_page_prot = protection_map[newflags &
 198                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
 199         if (vma_wants_writenotify(vma)) {
 200                 vma->vm_page_prot = protection_map[newflags &
 201                         (VM_READ|VM_WRITE|VM_EXEC)];
 202                 dirty_accountable = 1;
 203         }
 204
 205         if (oldflags & VM_EXEC)
 206                 arch_remove_exec_range(current->mm, old_end);
 207
 208         if (is_vm_hugetlb_page(vma))
 209                 hugetlb_change_protection(vma, start, end, vma->vm_page_prot);
 210         else
 211                 change_protection(vma, start, end, vma->vm_page_prot, dirty_accountable);
 212         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
 213         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
 214         return 0;
 215
 216 fail:
 217         vm_unacct_memory(charged);
 218         return error;
 219 }
 220
 221 asmlinkage long
 222 sys_mprotect(unsigned long start, size_t len, unsigned long prot)
 223 {
 224         unsigned long vm_flags, nstart, end, tmp, reqprot;
 225         struct vm_area_struct *vma, *prev;
 226         int error = -EINVAL;
 227         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 228         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 229         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 230                 return -EINVAL;
 231
 232         if (start & ~PAGE_MASK)
 233                 return -EINVAL;
 234         if (!len)
 235                 return 0;
 236         len = PAGE_ALIGN(len);
 237         end = start + len;
 238         if (end <= start)
 239                 return -ENOMEM;
 240         if (prot & ~(PROT_READ | PROT_WRITE | PROT_EXEC | PROT_SEM))
 241                 return -EINVAL;
 242
 243         reqprot = prot;
 244         /*
 245          * Does the application expect PROT_READ to imply PROT_EXEC:
 246          */
 247         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 248                 prot |= PROT_EXEC;
 249
 250         vm_flags = calc_vm_prot_bits(prot);
 251
 252         down_write(&current->mm->mmap_sem);
 253
 254         vma = find_vma_prev(current->mm, start, &prev);
 255         error = -ENOMEM;
 256         if (!vma)
 257                 goto out;
 258         if (unlikely(grows & PROT_GROWSDOWN)) {
 259                 if (vma->vm_start >= end)
 260                         goto out;
 261                 start = vma->vm_start;
 262                 error = -EINVAL;
 263                 if (!(vma->vm_flags & VM_GROWSDOWN))
 264                         goto out;
 265         }
 266         else {
 267                 if (vma->vm_start > start)
 268                         goto out;
 269                 if (unlikely(grows & PROT_GROWSUP)) {
 270                         end = vma->vm_end;
 271                         error = -EINVAL;
 272                         if (!(vma->vm_flags & VM_GROWSUP))
 273                                 goto out;
 274                 }
 275         }
 276         if (start > vma->vm_start)
 277                 prev = vma;
 278
 279         for (nstart = start ; ; ) {
 280                 unsigned long newflags;
 281
 282                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
 283
 284                 newflags = vm_flags | (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 285
 286                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 287                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 288                         error = -EACCES;
 289                         goto out;
 290                 }
 291
 292                 error = security_file_mprotect(vma, reqprot, prot);
 293                 if (error)
 294                         goto out;
 295
 296                 tmp = vma->vm_end;
 297                 if (tmp > end)
 298                         tmp = end;
 299                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 300                 if (error)
 301                         goto out;
 302                 nstart = tmp;
 303
 304                 if (nstart < prev->vm_end)
 305                         nstart = prev->vm_end;
 306                 if (nstart >= end)
 307                         goto out;
 308
 309                 vma = prev->vm_next;
 310                 if (!vma || vma->vm_start != nstart) {
 311                         error = -ENOMEM;
 312                         goto out;
 313                 }
 314         }
 315 out:
 316         up_write(&current->mm->mmap_sem);
 317         return error;
 318 }