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 <linux/ksm.h>
  27 #include <asm/uaccess.h>
  28 #include <asm/pgtable.h>
  29 #include <asm/cacheflush.h>
  30 #include <asm/tlbflush.h>
  31
  32 /*
  33  * For a prot_numa update we only hold mmap_sem for read so there is a
  34  * potential race with faulting where a pmd was temporarily none. This
  35  * function checks for a transhuge pmd under the appropriate lock. It
  36  * returns a pte if it was successfully locked or NULL if it raced with
  37  * a transhuge insertion.
  38  */
  39 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
  40                         unsigned long addr, int prot_numa, spinlock_t **ptl)
  41 {
  42         pte_t *pte;
  43         spinlock_t *pmdl;
  44
  45         /* !prot_numa is protected by mmap_sem held for write */
  46         if (!prot_numa)
  47                 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  48
  49         pmdl = pmd_lock(vma->vm_mm, pmd);
  50         if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
  51                 spin_unlock(pmdl);
  52                 return NULL;
  53         }
  54
  55         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  56         spin_unlock(pmdl);
  57         return pte;
  58 }
  59
  60 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  61                 unsigned long addr, unsigned long end, pgprot_t newprot,
  62                 int dirty_accountable, int prot_numa)
  63 {
  64         struct mm_struct *mm = vma->vm_mm;
  65         pte_t *pte, oldpte;
  66         spinlock_t *ptl;
  67         unsigned long pages = 0;
  68
  69         pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
  70         if (!pte)
  71                 return 0;
  72
  73         arch_enter_lazy_mmu_mode();
  74         do {
  75                 oldpte = *pte;
  76                 if (pte_present(oldpte)) {
  77                         pte_t ptent;
  78
  79                         /*
  80                          * Avoid trapping faults against the zero or KSM
  81                          * pages. See similar comment in change_huge_pmd.
  82                          */
  83                         if (prot_numa) {
  84                                 struct page *page;
  85
  86                                 page = vm_normal_page(vma, addr, oldpte);
  87                                 if (!page || PageKsm(page))
  88                                         continue;
  89                         }
  90
  91                         ptent = ptep_modify_prot_start(mm, addr, pte);
  92                         ptent = pte_modify(ptent, newprot);
  93
  94                         /* Avoid taking write faults for known dirty pages */
  95                         if (dirty_accountable && pte_dirty(ptent) &&
  96                                         (pte_soft_dirty(ptent) ||
  97                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
  98                                 ptent = pte_mkwrite(ptent);
  99                         }
 100                         ptep_modify_prot_commit(mm, addr, pte, ptent);
 101                         pages++;
 102                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 103                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 104
 105                         if (is_write_migration_entry(entry)) {
 106                                 pte_t newpte;
 107                                 /*
 108                                  * A protection check is difficult so
 109                                  * just be safe and disable write
 110                                  */
 111                                 make_migration_entry_read(&entry);
 112                                 newpte = swp_entry_to_pte(entry);
 113                                 if (pte_swp_soft_dirty(oldpte))
 114                                         newpte = pte_swp_mksoft_dirty(newpte);
 115                                 set_pte_at(mm, addr, pte, newpte);
 116
 117                                 pages++;
 118                         }
 119                 }
 120         } while (pte++, addr += PAGE_SIZE, addr != end);
 121         arch_leave_lazy_mmu_mode();
 122         pte_unmap_unlock(pte - 1, ptl);
 123
 124         return pages;
 125 }
 126
 127 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 128                 pud_t *pud, unsigned long addr, unsigned long end,
 129                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 130 {
 131         pmd_t *pmd;
 132         struct mm_struct *mm = vma->vm_mm;
 133         unsigned long next;
 134         unsigned long pages = 0;
 135         unsigned long nr_huge_updates = 0;
 136         unsigned long mni_start = 0;
 137
 138         pmd = pmd_offset(pud, addr);
 139         do {
 140                 unsigned long this_pages;
 141
 142                 next = pmd_addr_end(addr, end);
 143                 if (!pmd_trans_huge(*pmd) && pmd_none_or_clear_bad(pmd))
 144                         continue;
 145
 146                 /* invoke the mmu notifier if the pmd is populated */
 147                 if (!mni_start) {
 148                         mni_start = addr;
 149                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 150                 }
 151
 152                 if (pmd_trans_huge(*pmd)) {
 153                         if (next - addr != HPAGE_PMD_SIZE)
 154                                 split_huge_page_pmd(vma, addr, pmd);
 155                         else {
 156                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 157                                                 newprot, prot_numa);
 158
 159                                 if (nr_ptes) {
 160                                         if (nr_ptes == HPAGE_PMD_NR) {
 161                                                 pages += HPAGE_PMD_NR;
 162                                                 nr_huge_updates++;
 163                                         }
 164
 165                                         /* huge pmd was handled */
 166                                         continue;
 167                                 }
 168                         }
 169                         /* fall through, the trans huge pmd just split */
 170                 }
 171                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 172                                  dirty_accountable, prot_numa);
 173                 pages += this_pages;
 174         } while (pmd++, addr = next, addr != end);
 175
 176         if (mni_start)
 177                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 178
 179         if (nr_huge_updates)
 180                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 181         return pages;
 182 }
 183
 184 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 185                 pgd_t *pgd, unsigned long addr, unsigned long end,
 186                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 187 {
 188         pud_t *pud;
 189         unsigned long next;
 190         unsigned long pages = 0;
 191
 192         pud = pud_offset(pgd, addr);
 193         do {
 194                 next = pud_addr_end(addr, end);
 195                 if (pud_none_or_clear_bad(pud))
 196                         continue;
 197                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 198                                  dirty_accountable, prot_numa);
 199         } while (pud++, addr = next, addr != end);
 200
 201         return pages;
 202 }
 203
 204 static unsigned long change_protection_range(struct vm_area_struct *vma,
 205                 unsigned long addr, unsigned long end, pgprot_t newprot,
 206                 int dirty_accountable, int prot_numa)
 207 {
 208         struct mm_struct *mm = vma->vm_mm;
 209         pgd_t *pgd;
 210         unsigned long next;
 211         unsigned long start = addr;
 212         unsigned long pages = 0;
 213
 214         BUG_ON(addr >= end);
 215         pgd = pgd_offset(mm, addr);
 216         flush_cache_range(vma, addr, end);
 217         set_tlb_flush_pending(mm);
 218         do {
 219                 next = pgd_addr_end(addr, end);
 220                 if (pgd_none_or_clear_bad(pgd))
 221                         continue;
 222                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 223                                  dirty_accountable, prot_numa);
 224         } while (pgd++, addr = next, addr != end);
 225
 226         /* Only flush the TLB if we actually modified any entries: */
 227         if (pages)
 228                 flush_tlb_range(vma, start, end);
 229         clear_tlb_flush_pending(mm);
 230
 231         return pages;
 232 }
 233
 234 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 235                        unsigned long end, pgprot_t newprot,
 236                        int dirty_accountable, int prot_numa)
 237 {
 238         unsigned long pages;
 239
 240         if (is_vm_hugetlb_page(vma))
 241                 pages = hugetlb_change_protection(vma, start, end, newprot);
 242         else
 243                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 244
 245         return pages;
 246 }
 247
 248 int
 249 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 250         unsigned long start, unsigned long end, unsigned long newflags)
 251 {
 252         struct mm_struct *mm = vma->vm_mm;
 253         unsigned long oldflags = vma->vm_flags;
 254         long nrpages = (end - start) >> PAGE_SHIFT;
 255         unsigned long charged = 0;
 256         pgoff_t pgoff;
 257         int error;
 258         int dirty_accountable = 0;
 259
 260         if (newflags == oldflags) {
 261                 *pprev = vma;
 262                 return 0;
 263         }
 264
 265         /*
 266          * If we make a private mapping writable we increase our commit;
 267          * but (without finer accounting) cannot reduce our commit if we
 268          * make it unwritable again. hugetlb mapping were accounted for
 269          * even if read-only so there is no need to account for them here
 270          */
 271         if (newflags & VM_WRITE) {
 272                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 273                                                 VM_SHARED|VM_NORESERVE))) {
 274                         charged = nrpages;
 275                         if (security_vm_enough_memory_mm(mm, charged))
 276                                 return -ENOMEM;
 277                         newflags |= VM_ACCOUNT;
 278                 }
 279         }
 280
 281         /*
 282          * First try to merge with previous and/or next vma.
 283          */
 284         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 285         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 286                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
 287         if (*pprev) {
 288                 vma = *pprev;
 289                 goto success;
 290         }
 291
 292         *pprev = vma;
 293
 294         if (start != vma->vm_start) {
 295                 error = split_vma(mm, vma, start, 1);
 296                 if (error)
 297                         goto fail;
 298         }
 299
 300         if (end != vma->vm_end) {
 301                 error = split_vma(mm, vma, end, 0);
 302                 if (error)
 303                         goto fail;
 304         }
 305
 306 success:
 307         /*
 308          * vm_flags and vm_page_prot are protected by the mmap_sem
 309          * held in write mode.
 310          */
 311         vma->vm_flags = newflags;
 312         dirty_accountable = vma_wants_writenotify(vma);
 313         vma_set_page_prot(vma);
 314
 315         change_protection(vma, start, end, vma->vm_page_prot,
 316                           dirty_accountable, 0);
 317
 318         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
 319         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
 320         perf_event_mmap(vma);
 321         return 0;
 322
 323 fail:
 324         vm_unacct_memory(charged);
 325         return error;
 326 }
 327
 328 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 329                 unsigned long, prot)
 330 {
 331         unsigned long vm_flags, nstart, end, tmp, reqprot;
 332         struct vm_area_struct *vma, *prev;
 333         int error = -EINVAL;
 334         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 335         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 336         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 337                 return -EINVAL;
 338
 339         if (start & ~PAGE_MASK)
 340                 return -EINVAL;
 341         if (!len)
 342                 return 0;
 343         len = PAGE_ALIGN(len);
 344         end = start + len;
 345         if (end <= start)
 346                 return -ENOMEM;
 347         if (!arch_validate_prot(prot))
 348                 return -EINVAL;
 349
 350         reqprot = prot;
 351         /*
 352          * Does the application expect PROT_READ to imply PROT_EXEC:
 353          */
 354         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 355                 prot |= PROT_EXEC;
 356
 357         vm_flags = calc_vm_prot_bits(prot);
 358
 359         down_write(&current->mm->mmap_sem);
 360
 361         vma = find_vma(current->mm, start);
 362         error = -ENOMEM;
 363         if (!vma)
 364                 goto out;
 365         prev = vma->vm_prev;
 366         if (unlikely(grows & PROT_GROWSDOWN)) {
 367                 if (vma->vm_start >= end)
 368                         goto out;
 369                 start = vma->vm_start;
 370                 error = -EINVAL;
 371                 if (!(vma->vm_flags & VM_GROWSDOWN))
 372                         goto out;
 373         } else {
 374                 if (vma->vm_start > start)
 375                         goto out;
 376                 if (unlikely(grows & PROT_GROWSUP)) {
 377                         end = vma->vm_end;
 378                         error = -EINVAL;
 379                         if (!(vma->vm_flags & VM_GROWSUP))
 380                                 goto out;
 381                 }
 382         }
 383         if (start > vma->vm_start)
 384                 prev = vma;
 385
 386         for (nstart = start ; ; ) {
 387                 unsigned long newflags;
 388
 389                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 390
 391                 newflags = vm_flags;
 392                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 393
 394                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 395                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 396                         error = -EACCES;
 397                         goto out;
 398                 }
 399
 400                 error = security_file_mprotect(vma, reqprot, prot);
 401                 if (error)
 402                         goto out;
 403
 404                 tmp = vma->vm_end;
 405                 if (tmp > end)
 406                         tmp = end;
 407                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 408                 if (error)
 409                         goto out;
 410                 nstart = tmp;
 411
 412                 if (nstart < prev->vm_end)
 413                         nstart = prev->vm_end;
 414                 if (nstart >= end)
 415                         goto out;
 416
 417                 vma = prev->vm_next;
 418                 if (!vma || vma->vm_start != nstart) {
 419                         error = -ENOMEM;
 420                         goto out;
 421                 }
 422         }
 423 out:
 424         up_write(&current->mm->mmap_sem);
 425         return error;
 426 }