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                                 /* Avoid TLB flush if possible */
  91                                 if (pte_protnone(oldpte))
  92                                         continue;
  93                         }
  94
  95                         ptent = ptep_modify_prot_start(mm, addr, pte);
  96                         ptent = pte_modify(ptent, newprot);
  97
  98                         /* Avoid taking write faults for known dirty pages */
  99                         if (dirty_accountable && pte_dirty(ptent) &&
 100                                         (pte_soft_dirty(ptent) ||
 101                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 102                                 ptent = pte_mkwrite(ptent);
 103                         }
 104                         ptep_modify_prot_commit(mm, addr, pte, ptent);
 105                         pages++;
 106                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 107                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 108
 109                         if (is_write_migration_entry(entry)) {
 110                                 pte_t newpte;
 111                                 /*
 112                                  * A protection check is difficult so
 113                                  * just be safe and disable write
 114                                  */
 115                                 make_migration_entry_read(&entry);
 116                                 newpte = swp_entry_to_pte(entry);
 117                                 if (pte_swp_soft_dirty(oldpte))
 118                                         newpte = pte_swp_mksoft_dirty(newpte);
 119                                 set_pte_at(mm, addr, pte, newpte);
 120
 121                                 pages++;
 122                         }
 123                 }
 124         } while (pte++, addr += PAGE_SIZE, addr != end);
 125         arch_leave_lazy_mmu_mode();
 126         pte_unmap_unlock(pte - 1, ptl);
 127
 128         return pages;
 129 }
 130
 131 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 132                 pud_t *pud, unsigned long addr, unsigned long end,
 133                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 134 {
 135         pmd_t *pmd;
 136         struct mm_struct *mm = vma->vm_mm;
 137         unsigned long next;
 138         unsigned long pages = 0;
 139         unsigned long nr_huge_updates = 0;
 140         unsigned long mni_start = 0;
 141
 142         pmd = pmd_offset(pud, addr);
 143         do {
 144                 unsigned long this_pages;
 145
 146                 next = pmd_addr_end(addr, end);
 147                 if (!pmd_trans_huge(*pmd) && pmd_none_or_clear_bad(pmd))
 148                         continue;
 149
 150                 /* invoke the mmu notifier if the pmd is populated */
 151                 if (!mni_start) {
 152                         mni_start = addr;
 153                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 154                 }
 155
 156                 if (pmd_trans_huge(*pmd)) {
 157                         if (next - addr != HPAGE_PMD_SIZE)
 158                                 split_huge_page_pmd(vma, addr, pmd);
 159                         else {
 160                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 161                                                 newprot, prot_numa);
 162
 163                                 if (nr_ptes) {
 164                                         if (nr_ptes == HPAGE_PMD_NR) {
 165                                                 pages += HPAGE_PMD_NR;
 166                                                 nr_huge_updates++;
 167                                         }
 168
 169                                         /* huge pmd was handled */
 170                                         continue;
 171                                 }
 172                         }
 173                         /* fall through, the trans huge pmd just split */
 174                 }
 175                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 176                                  dirty_accountable, prot_numa);
 177                 pages += this_pages;
 178         } while (pmd++, addr = next, addr != end);
 179
 180         if (mni_start)
 181                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 182
 183         if (nr_huge_updates)
 184                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 185         return pages;
 186 }
 187
 188 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 189                 pgd_t *pgd, unsigned long addr, unsigned long end,
 190                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 191 {
 192         pud_t *pud;
 193         unsigned long next;
 194         unsigned long pages = 0;
 195
 196         pud = pud_offset(pgd, addr);
 197         do {
 198                 next = pud_addr_end(addr, end);
 199                 if (pud_none_or_clear_bad(pud))
 200                         continue;
 201                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 202                                  dirty_accountable, prot_numa);
 203         } while (pud++, addr = next, addr != end);
 204
 205         return pages;
 206 }
 207
 208 static unsigned long change_protection_range(struct vm_area_struct *vma,
 209                 unsigned long addr, unsigned long end, pgprot_t newprot,
 210                 int dirty_accountable, int prot_numa)
 211 {
 212         struct mm_struct *mm = vma->vm_mm;
 213         pgd_t *pgd;
 214         unsigned long next;
 215         unsigned long start = addr;
 216         unsigned long pages = 0;
 217
 218         BUG_ON(addr >= end);
 219         pgd = pgd_offset(mm, addr);
 220         flush_cache_range(vma, addr, end);
 221         set_tlb_flush_pending(mm);
 222         do {
 223                 next = pgd_addr_end(addr, end);
 224                 if (pgd_none_or_clear_bad(pgd))
 225                         continue;
 226                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 227                                  dirty_accountable, prot_numa);
 228         } while (pgd++, addr = next, addr != end);
 229
 230         /* Only flush the TLB if we actually modified any entries: */
 231         if (pages)
 232                 flush_tlb_range(vma, start, end);
 233         clear_tlb_flush_pending(mm);
 234
 235         return pages;
 236 }
 237
 238 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 239                        unsigned long end, pgprot_t newprot,
 240                        int dirty_accountable, int prot_numa)
 241 {
 242         unsigned long pages;
 243
 244         if (is_vm_hugetlb_page(vma))
 245                 pages = hugetlb_change_protection(vma, start, end, newprot);
 246         else
 247                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 248
 249         return pages;
 250 }
 251
 252 int
 253 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 254         unsigned long start, unsigned long end, unsigned long newflags)
 255 {
 256         struct mm_struct *mm = vma->vm_mm;
 257         unsigned long oldflags = vma->vm_flags;
 258         long nrpages = (end - start) >> PAGE_SHIFT;
 259         unsigned long charged = 0;
 260         pgoff_t pgoff;
 261         int error;
 262         int dirty_accountable = 0;
 263
 264         if (newflags == oldflags) {
 265                 *pprev = vma;
 266                 return 0;
 267         }
 268
 269         /*
 270          * If we make a private mapping writable we increase our commit;
 271          * but (without finer accounting) cannot reduce our commit if we
 272          * make it unwritable again. hugetlb mapping were accounted for
 273          * even if read-only so there is no need to account for them here
 274          */
 275         if (newflags & VM_WRITE) {
 276                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 277                                                 VM_SHARED|VM_NORESERVE))) {
 278                         charged = nrpages;
 279                         if (security_vm_enough_memory_mm(mm, charged))
 280                                 return -ENOMEM;
 281                         newflags |= VM_ACCOUNT;
 282                 }
 283         }
 284
 285         /*
 286          * First try to merge with previous and/or next vma.
 287          */
 288         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 289         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 290                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
 291         if (*pprev) {
 292                 vma = *pprev;
 293                 goto success;
 294         }
 295
 296         *pprev = vma;
 297
 298         if (start != vma->vm_start) {
 299                 error = split_vma(mm, vma, start, 1);
 300                 if (error)
 301                         goto fail;
 302         }
 303
 304         if (end != vma->vm_end) {
 305                 error = split_vma(mm, vma, end, 0);
 306                 if (error)
 307                         goto fail;
 308         }
 309
 310 success:
 311         /*
 312          * vm_flags and vm_page_prot are protected by the mmap_sem
 313          * held in write mode.
 314          */
 315         vma->vm_flags = newflags;
 316         dirty_accountable = vma_wants_writenotify(vma);
 317         vma_set_page_prot(vma);
 318
 319         change_protection(vma, start, end, vma->vm_page_prot,
 320                           dirty_accountable, 0);
 321
 322         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
 323         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
 324         perf_event_mmap(vma);
 325         return 0;
 326
 327 fail:
 328         vm_unacct_memory(charged);
 329         return error;
 330 }
 331
 332 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 333                 unsigned long, prot)
 334 {
 335         unsigned long vm_flags, nstart, end, tmp, reqprot;
 336         struct vm_area_struct *vma, *prev;
 337         int error = -EINVAL;
 338         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 339         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 340         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 341                 return -EINVAL;
 342
 343         if (start & ~PAGE_MASK)
 344                 return -EINVAL;
 345         if (!len)
 346                 return 0;
 347         len = PAGE_ALIGN(len);
 348         end = start + len;
 349         if (end <= start)
 350                 return -ENOMEM;
 351         if (!arch_validate_prot(prot))
 352                 return -EINVAL;
 353
 354         reqprot = prot;
 355         /*
 356          * Does the application expect PROT_READ to imply PROT_EXEC:
 357          */
 358         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 359                 prot |= PROT_EXEC;
 360
 361         vm_flags = calc_vm_prot_bits(prot);
 362
 363         down_write(&current->mm->mmap_sem);
 364
 365         vma = find_vma(current->mm, start);
 366         error = -ENOMEM;
 367         if (!vma)
 368                 goto out;
 369         prev = vma->vm_prev;
 370         if (unlikely(grows & PROT_GROWSDOWN)) {
 371                 if (vma->vm_start >= end)
 372                         goto out;
 373                 start = vma->vm_start;
 374                 error = -EINVAL;
 375                 if (!(vma->vm_flags & VM_GROWSDOWN))
 376                         goto out;
 377         } else {
 378                 if (vma->vm_start > start)
 379                         goto out;
 380                 if (unlikely(grows & PROT_GROWSUP)) {
 381                         end = vma->vm_end;
 382                         error = -EINVAL;
 383                         if (!(vma->vm_flags & VM_GROWSUP))
 384                                 goto out;
 385                 }
 386         }
 387         if (start > vma->vm_start)
 388                 prev = vma;
 389
 390         for (nstart = start ; ; ) {
 391                 unsigned long newflags;
 392
 393                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 394
 395                 newflags = vm_flags;
 396                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 397
 398                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 399                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 400                         error = -EACCES;
 401                         goto out;
 402                 }
 403
 404                 error = security_file_mprotect(vma, reqprot, prot);
 405                 if (error)
 406                         goto out;
 407
 408                 tmp = vma->vm_end;
 409                 if (tmp > end)
 410                         tmp = end;
 411                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 412                 if (error)
 413                         goto out;
 414                 nstart = tmp;
 415
 416                 if (nstart < prev->vm_end)
 417                         nstart = prev->vm_end;
 418                 if (nstart >= end)
 419                         goto out;
 420
 421                 vma = prev->vm_next;
 422                 if (!vma || vma->vm_start != nstart) {
 423                         error = -ENOMEM;
 424                         goto out;
 425                 }
 426         }
 427 out:
 428         up_write(&current->mm->mmap_sem);
 429         return error;
 430 }