Pileus Git - ~andy/linux/blob - 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@lxorguk.ukuu.org.uk>
   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/shm.h>
  14 #include <linux/mman.h>
  15 #include <linux/fs.h>
  16 #include <linux/highmem.h>
  17 #include <linux/security.h>
  18 #include <linux/mempolicy.h>
  19 #include <linux/personality.h>
  20 #include <linux/syscalls.h>
  21 #include <linux/swap.h>
  22 #include <linux/swapops.h>
  23 #include <linux/mmu_notifier.h>
  24 #include <linux/migrate.h>
  25 #include <linux/perf_event.h>
  26 #include <asm/uaccess.h>
  27 #include <asm/pgtable.h>
  28 #include <asm/cacheflush.h>
  29 #include <asm/tlbflush.h>
  30
  31 #ifndef pgprot_modify
  32 static inline pgprot_t pgprot_modify(pgprot_t oldprot, pgprot_t newprot)
  33 {
  34         return newprot;
  35 }
  36 #endif
  37
  38 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  39                 unsigned long addr, unsigned long end, pgprot_t newprot,
  40                 int dirty_accountable, int prot_numa)
  41 {
  42         struct mm_struct *mm = vma->vm_mm;
  43         pte_t *pte, oldpte;
  44         spinlock_t *ptl;
  45         unsigned long pages = 0;
  46
  47         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
  48         arch_enter_lazy_mmu_mode();
  49         do {
  50                 oldpte = *pte;
  51                 if (pte_present(oldpte)) {
  52                         pte_t ptent;
  53                         bool updated = false;
  54
  55                         if (!prot_numa) {
  56                                 ptent = ptep_modify_prot_start(mm, addr, pte);
  57                                 if (pte_numa(ptent))
  58                                         ptent = pte_mknonnuma(ptent);
  59                                 ptent = pte_modify(ptent, newprot);
  60                                 updated = true;
  61                         } else {
  62                                 struct page *page;
  63
  64                                 ptent = *pte;
  65                                 page = vm_normal_page(vma, addr, oldpte);
  66                                 if (page) {
  67                                         if (!pte_numa(oldpte)) {
  68                                                 ptent = pte_mknuma(ptent);
  69                                                 set_pte_at(mm, addr, pte, ptent);
  70                                                 updated = true;
  71                                         }
  72                                 }
  73                         }
  74
  75                         /*
  76                          * Avoid taking write faults for pages we know to be
  77                          * dirty.
  78                          */
  79                         if (dirty_accountable && pte_dirty(ptent)) {
  80                                 ptent = pte_mkwrite(ptent);
  81                                 updated = true;
  82                         }
  83
  84                         if (updated)
  85                                 pages++;
  86
  87                         /* Only !prot_numa always clears the pte */
  88                         if (!prot_numa)
  89                                 ptep_modify_prot_commit(mm, addr, pte, ptent);
  90                 } else if (IS_ENABLED(CONFIG_MIGRATION) && !pte_file(oldpte)) {
  91                         swp_entry_t entry = pte_to_swp_entry(oldpte);
  92
  93                         if (is_write_migration_entry(entry)) {
  94                                 pte_t newpte;
  95                                 /*
  96                                  * A protection check is difficult so
  97                                  * just be safe and disable write
  98                                  */
  99                                 make_migration_entry_read(&entry);
 100                                 newpte = swp_entry_to_pte(entry);
 101                                 if (pte_swp_soft_dirty(oldpte))
 102                                         newpte = pte_swp_mksoft_dirty(newpte);
 103                                 set_pte_at(mm, addr, pte, newpte);
 104
 105                                 pages++;
 106                         }
 107                 }
 108         } while (pte++, addr += PAGE_SIZE, addr != end);
 109         arch_leave_lazy_mmu_mode();
 110         pte_unmap_unlock(pte - 1, ptl);
 111
 112         return pages;
 113 }
 114
 115 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 116                 pud_t *pud, unsigned long addr, unsigned long end,
 117                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 118 {
 119         pmd_t *pmd;
 120         unsigned long next;
 121         unsigned long pages = 0;
 122         unsigned long nr_huge_updates = 0;
 123
 124         pmd = pmd_offset(pud, addr);
 125         do {
 126                 unsigned long this_pages;
 127
 128                 next = pmd_addr_end(addr, end);
 129                 if (pmd_trans_huge(*pmd)) {
 130                         if (next - addr != HPAGE_PMD_SIZE)
 131                                 split_huge_page_pmd(vma, addr, pmd);
 132                         else {
 133                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 134                                                 newprot, prot_numa);
 135
 136                                 if (nr_ptes) {
 137                                         if (nr_ptes == HPAGE_PMD_NR) {
 138                                                 pages += HPAGE_PMD_NR;
 139                                                 nr_huge_updates++;
 140                                         }
 141                                         continue;
 142                                 }
 143                         }
 144                         /* fall through */
 145                 }
 146                 if (pmd_none_or_clear_bad(pmd))
 147                         continue;
 148                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 149                                  dirty_accountable, prot_numa);
 150                 pages += this_pages;
 151         } while (pmd++, addr = next, addr != end);
 152
 153         if (nr_huge_updates)
 154                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 155         return pages;
 156 }
 157
 158 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 159                 pgd_t *pgd, unsigned long addr, unsigned long end,
 160                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 161 {
 162         pud_t *pud;
 163         unsigned long next;
 164         unsigned long pages = 0;
 165
 166         pud = pud_offset(pgd, addr);
 167         do {
 168                 next = pud_addr_end(addr, end);
 169                 if (pud_none_or_clear_bad(pud))
 170                         continue;
 171                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 172                                  dirty_accountable, prot_numa);
 173         } while (pud++, addr = next, addr != end);
 174
 175         return pages;
 176 }
 177
 178 static unsigned long change_protection_range(struct vm_area_struct *vma,
 179                 unsigned long addr, unsigned long end, pgprot_t newprot,
 180                 int dirty_accountable, int prot_numa)
 181 {
 182         struct mm_struct *mm = vma->vm_mm;
 183         pgd_t *pgd;
 184         unsigned long next;
 185         unsigned long start = addr;
 186         unsigned long pages = 0;
 187
 188         BUG_ON(addr >= end);
 189         pgd = pgd_offset(mm, addr);
 190         flush_cache_range(vma, addr, end);
 191         set_tlb_flush_pending(mm);
 192         do {
 193                 next = pgd_addr_end(addr, end);
 194                 if (pgd_none_or_clear_bad(pgd))
 195                         continue;
 196                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 197                                  dirty_accountable, prot_numa);
 198         } while (pgd++, addr = next, addr != end);
 199
 200         /* Only flush the TLB if we actually modified any entries: */
 201         if (pages)
 202                 flush_tlb_range(vma, start, end);
 203         clear_tlb_flush_pending(mm);
 204
 205         return pages;
 206 }
 207
 208 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 209                        unsigned long end, pgprot_t newprot,
 210                        int dirty_accountable, int prot_numa)
 211 {
 212         struct mm_struct *mm = vma->vm_mm;
 213         unsigned long pages;
 214
 215         mmu_notifier_invalidate_range_start(mm, start, end);
 216         if (is_vm_hugetlb_page(vma))
 217                 pages = hugetlb_change_protection(vma, start, end, newprot);
 218         else
 219                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 220         mmu_notifier_invalidate_range_end(mm, start, end);
 221
 222         return pages;
 223 }
 224
 225 int
 226 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 227         unsigned long start, unsigned long end, unsigned long newflags)
 228 {
 229         struct mm_struct *mm = vma->vm_mm;
 230         unsigned long oldflags = vma->vm_flags;
 231         long nrpages = (end - start) >> PAGE_SHIFT;
 232         unsigned long charged = 0;
 233         pgoff_t pgoff;
 234         int error;
 235         int dirty_accountable = 0;
 236
 237         if (newflags == oldflags) {
 238                 *pprev = vma;
 239                 return 0;
 240         }
 241
 242         /*
 243          * If we make a private mapping writable we increase our commit;
 244          * but (without finer accounting) cannot reduce our commit if we
 245          * make it unwritable again. hugetlb mapping were accounted for
 246          * even if read-only so there is no need to account for them here
 247          */
 248         if (newflags & VM_WRITE) {
 249                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 250                                                 VM_SHARED|VM_NORESERVE))) {
 251                         charged = nrpages;
 252                         if (security_vm_enough_memory_mm(mm, charged))
 253                                 return -ENOMEM;
 254                         newflags |= VM_ACCOUNT;
 255                 }
 256         }
 257
 258         /*
 259          * First try to merge with previous and/or next vma.
 260          */
 261         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 262         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 263                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
 264         if (*pprev) {
 265                 vma = *pprev;
 266                 goto success;
 267         }
 268
 269         *pprev = vma;
 270
 271         if (start != vma->vm_start) {
 272                 error = split_vma(mm, vma, start, 1);
 273                 if (error)
 274                         goto fail;
 275         }
 276
 277         if (end != vma->vm_end) {
 278                 error = split_vma(mm, vma, end, 0);
 279                 if (error)
 280                         goto fail;
 281         }
 282
 283 success:
 284         /*
 285          * vm_flags and vm_page_prot are protected by the mmap_sem
 286          * held in write mode.
 287          */
 288         vma->vm_flags = newflags;
 289         vma->vm_page_prot = pgprot_modify(vma->vm_page_prot,
 290                                           vm_get_page_prot(newflags));
 291
 292         if (vma_wants_writenotify(vma)) {
 293                 vma->vm_page_prot = vm_get_page_prot(newflags & ~VM_SHARED);
 294                 dirty_accountable = 1;
 295         }
 296
 297         change_protection(vma, start, end, vma->vm_page_prot,
 298                           dirty_accountable, 0);
 299
 300         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
 301         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
 302         perf_event_mmap(vma);
 303         return 0;
 304
 305 fail:
 306         vm_unacct_memory(charged);
 307         return error;
 308 }
 309
 310 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 311                 unsigned long, prot)
 312 {
 313         unsigned long vm_flags, nstart, end, tmp, reqprot;
 314         struct vm_area_struct *vma, *prev;
 315         int error = -EINVAL;
 316         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 317         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 318         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 319                 return -EINVAL;
 320
 321         if (start & ~PAGE_MASK)
 322                 return -EINVAL;
 323         if (!len)
 324                 return 0;
 325         len = PAGE_ALIGN(len);
 326         end = start + len;
 327         if (end <= start)
 328                 return -ENOMEM;
 329         if (!arch_validate_prot(prot))
 330                 return -EINVAL;
 331
 332         reqprot = prot;
 333         /*
 334          * Does the application expect PROT_READ to imply PROT_EXEC:
 335          */
 336         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 337                 prot |= PROT_EXEC;
 338
 339         vm_flags = calc_vm_prot_bits(prot);
 340
 341         down_write(&current->mm->mmap_sem);
 342
 343         vma = find_vma(current->mm, start);
 344         error = -ENOMEM;
 345         if (!vma)
 346                 goto out;
 347         prev = vma->vm_prev;
 348         if (unlikely(grows & PROT_GROWSDOWN)) {
 349                 if (vma->vm_start >= end)
 350                         goto out;
 351                 start = vma->vm_start;
 352                 error = -EINVAL;
 353                 if (!(vma->vm_flags & VM_GROWSDOWN))
 354                         goto out;
 355         } else {
 356                 if (vma->vm_start > start)
 357                         goto out;
 358                 if (unlikely(grows & PROT_GROWSUP)) {
 359                         end = vma->vm_end;
 360                         error = -EINVAL;
 361                         if (!(vma->vm_flags & VM_GROWSUP))
 362                                 goto out;
 363                 }
 364         }
 365         if (start > vma->vm_start)
 366                 prev = vma;
 367
 368         for (nstart = start ; ; ) {
 369                 unsigned long newflags;
 370
 371                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 372
 373                 newflags = vm_flags;
 374                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 375
 376                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 377                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 378                         error = -EACCES;
 379                         goto out;
 380                 }
 381
 382                 error = security_file_mprotect(vma, reqprot, prot);
 383                 if (error)
 384                         goto out;
 385
 386                 tmp = vma->vm_end;
 387                 if (tmp > end)
 388                         tmp = end;
 389                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 390                 if (error)
 391                         goto out;
 392                 nstart = tmp;
 393
 394                 if (nstart < prev->vm_end)
 395                         nstart = prev->vm_end;
 396                 if (nstart >= end)
 397                         goto out;
 398
 399                 vma = prev->vm_next;
 400                 if (!vma || vma->vm_start != nstart) {
 401                         error = -ENOMEM;
 402                         goto out;
 403                 }
 404         }
 405 out:
 406         up_write(&current->mm->mmap_sem);
 407         return error;
 408 }