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 #include "internal.h"
  33
  34 /*
  35  * For a prot_numa update we only hold mmap_sem for read so there is a
  36  * potential race with faulting where a pmd was temporarily none. This
  37  * function checks for a transhuge pmd under the appropriate lock. It
  38  * returns a pte if it was successfully locked or NULL if it raced with
  39  * a transhuge insertion.
  40  */
  41 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
  42                         unsigned long addr, int prot_numa, spinlock_t **ptl)
  43 {
  44         pte_t *pte;
  45         spinlock_t *pmdl;
  46
  47         /* !prot_numa is protected by mmap_sem held for write */
  48         if (!prot_numa)
  49                 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  50
  51         pmdl = pmd_lock(vma->vm_mm, pmd);
  52         if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
  53                 spin_unlock(pmdl);
  54                 return NULL;
  55         }
  56
  57         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  58         spin_unlock(pmdl);
  59         return pte;
  60 }
  61
  62 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  63                 unsigned long addr, unsigned long end, pgprot_t newprot,
  64                 int dirty_accountable, int prot_numa)
  65 {
  66         struct mm_struct *mm = vma->vm_mm;
  67         pte_t *pte, oldpte;
  68         spinlock_t *ptl;
  69         unsigned long pages = 0;
  70
  71         pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
  72         if (!pte)
  73                 return 0;
  74
  75         arch_enter_lazy_mmu_mode();
  76         do {
  77                 oldpte = *pte;
  78                 if (pte_present(oldpte)) {
  79                         pte_t ptent;
  80                         bool preserve_write = prot_numa && pte_write(oldpte);
  81
  82                         /*
  83                          * Avoid trapping faults against the zero or KSM
  84                          * pages. See similar comment in change_huge_pmd.
  85                          */
  86                         if (prot_numa) {
  87                                 struct page *page;
  88
  89                                 page = vm_normal_page(vma, addr, oldpte);
  90                                 if (!page || PageKsm(page))
  91                                         continue;
  92
  93                                 /* Avoid TLB flush if possible */
  94                                 if (pte_protnone(oldpte))
  95                                         continue;
  96                         }
  97
  98                         ptent = ptep_modify_prot_start(mm, addr, pte);
  99                         ptent = pte_modify(ptent, newprot);
 100                         if (preserve_write)
 101                                 ptent = pte_mkwrite(ptent);
 102
 103                         /* Avoid taking write faults for known dirty pages */
 104                         if (dirty_accountable && pte_dirty(ptent) &&
 105                                         (pte_soft_dirty(ptent) ||
 106                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 107                                 ptent = pte_mkwrite(ptent);
 108                         }
 109                         ptep_modify_prot_commit(mm, addr, pte, ptent);
 110                         pages++;
 111                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 112                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 113
 114                         if (is_write_migration_entry(entry)) {
 115                                 pte_t newpte;
 116                                 /*
 117                                  * A protection check is difficult so
 118                                  * just be safe and disable write
 119                                  */
 120                                 make_migration_entry_read(&entry);
 121                                 newpte = swp_entry_to_pte(entry);
 122                                 if (pte_swp_soft_dirty(oldpte))
 123                                         newpte = pte_swp_mksoft_dirty(newpte);
 124                                 set_pte_at(mm, addr, pte, newpte);
 125
 126                                 pages++;
 127                         }
 128                 }
 129         } while (pte++, addr += PAGE_SIZE, addr != end);
 130         arch_leave_lazy_mmu_mode();
 131         pte_unmap_unlock(pte - 1, ptl);
 132
 133         return pages;
 134 }
 135
 136 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 137                 pud_t *pud, unsigned long addr, unsigned long end,
 138                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 139 {
 140         pmd_t *pmd;
 141         struct mm_struct *mm = vma->vm_mm;
 142         unsigned long next;
 143         unsigned long pages = 0;
 144         unsigned long nr_huge_updates = 0;
 145         unsigned long mni_start = 0;
 146
 147         pmd = pmd_offset(pud, addr);
 148         do {
 149                 unsigned long this_pages;
 150
 151                 next = pmd_addr_end(addr, end);
 152                 if (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
 153                                 && pmd_none_or_clear_bad(pmd))
 154                         continue;
 155
 156                 /* invoke the mmu notifier if the pmd is populated */
 157                 if (!mni_start) {
 158                         mni_start = addr;
 159                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 160                 }
 161
 162                 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 163                         if (next - addr != HPAGE_PMD_SIZE) {
 164                                 split_huge_pmd(vma, pmd, addr);
 165                                 if (pmd_none(*pmd))
 166                                         continue;
 167                         } else {
 168                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 169                                                 newprot, prot_numa);
 170
 171                                 if (nr_ptes) {
 172                                         if (nr_ptes == HPAGE_PMD_NR) {
 173                                                 pages += HPAGE_PMD_NR;
 174                                                 nr_huge_updates++;
 175                                         }
 176
 177                                         /* huge pmd was handled */
 178                                         continue;
 179                                 }
 180                         }
 181                         /* fall through, the trans huge pmd just split */
 182                 }
 183                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 184                                  dirty_accountable, prot_numa);
 185                 pages += this_pages;
 186         } while (pmd++, addr = next, addr != end);
 187
 188         if (mni_start)
 189                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 190
 191         if (nr_huge_updates)
 192                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 193         return pages;
 194 }
 195
 196 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 197                 pgd_t *pgd, unsigned long addr, unsigned long end,
 198                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 199 {
 200         pud_t *pud;
 201         unsigned long next;
 202         unsigned long pages = 0;
 203
 204         pud = pud_offset(pgd, addr);
 205         do {
 206                 next = pud_addr_end(addr, end);
 207                 if (pud_none_or_clear_bad(pud))
 208                         continue;
 209                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 210                                  dirty_accountable, prot_numa);
 211         } while (pud++, addr = next, addr != end);
 212
 213         return pages;
 214 }
 215
 216 static unsigned long change_protection_range(struct vm_area_struct *vma,
 217                 unsigned long addr, unsigned long end, pgprot_t newprot,
 218                 int dirty_accountable, int prot_numa)
 219 {
 220         struct mm_struct *mm = vma->vm_mm;
 221         pgd_t *pgd;
 222         unsigned long next;
 223         unsigned long start = addr;
 224         unsigned long pages = 0;
 225
 226         BUG_ON(addr >= end);
 227         pgd = pgd_offset(mm, addr);
 228         flush_cache_range(vma, addr, end);
 229         set_tlb_flush_pending(mm);
 230         do {
 231                 next = pgd_addr_end(addr, end);
 232                 if (pgd_none_or_clear_bad(pgd))
 233                         continue;
 234                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 235                                  dirty_accountable, prot_numa);
 236         } while (pgd++, addr = next, addr != end);
 237
 238         /* Only flush the TLB if we actually modified any entries: */
 239         if (pages)
 240                 flush_tlb_range(vma, start, end);
 241         clear_tlb_flush_pending(mm);
 242
 243         return pages;
 244 }
 245
 246 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 247                        unsigned long end, pgprot_t newprot,
 248                        int dirty_accountable, int prot_numa)
 249 {
 250         unsigned long pages;
 251
 252         if (is_vm_hugetlb_page(vma))
 253                 pages = hugetlb_change_protection(vma, start, end, newprot);
 254         else
 255                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 256
 257         return pages;
 258 }
 259
 260 int
 261 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 262         unsigned long start, unsigned long end, unsigned long newflags)
 263 {
 264         struct mm_struct *mm = vma->vm_mm;
 265         unsigned long oldflags = vma->vm_flags;
 266         long nrpages = (end - start) >> PAGE_SHIFT;
 267         unsigned long charged = 0;
 268         pgoff_t pgoff;
 269         int error;
 270         int dirty_accountable = 0;
 271
 272         if (newflags == oldflags) {
 273                 *pprev = vma;
 274                 return 0;
 275         }
 276
 277         /*
 278          * If we make a private mapping writable we increase our commit;
 279          * but (without finer accounting) cannot reduce our commit if we
 280          * make it unwritable again. hugetlb mapping were accounted for
 281          * even if read-only so there is no need to account for them here
 282          */
 283         if (newflags & VM_WRITE) {
 284                 /* Check space limits when area turns into data. */
 285                 if (!may_expand_vm(mm, newflags, nrpages) &&
 286                                 may_expand_vm(mm, oldflags, nrpages))
 287                         return -ENOMEM;
 288                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 289                                                 VM_SHARED|VM_NORESERVE))) {
 290                         charged = nrpages;
 291                         if (security_vm_enough_memory_mm(mm, charged))
 292                                 return -ENOMEM;
 293                         newflags |= VM_ACCOUNT;
 294                 }
 295         }
 296
 297         /*
 298          * First try to merge with previous and/or next vma.
 299          */
 300         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 301         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 302                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 303                            vma->vm_userfaultfd_ctx);
 304         if (*pprev) {
 305                 vma = *pprev;
 306                 goto success;
 307         }
 308
 309         *pprev = vma;
 310
 311         if (start != vma->vm_start) {
 312                 error = split_vma(mm, vma, start, 1);
 313                 if (error)
 314                         goto fail;
 315         }
 316
 317         if (end != vma->vm_end) {
 318                 error = split_vma(mm, vma, end, 0);
 319                 if (error)
 320                         goto fail;
 321         }
 322
 323 success:
 324         /*
 325          * vm_flags and vm_page_prot are protected by the mmap_sem
 326          * held in write mode.
 327          */
 328         vma->vm_flags = newflags;
 329         dirty_accountable = vma_wants_writenotify(vma);
 330         vma_set_page_prot(vma);
 331
 332         change_protection(vma, start, end, vma->vm_page_prot,
 333                           dirty_accountable, 0);
 334
 335         /*
 336          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 337          * fault on access.
 338          */
 339         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 340                         (newflags & VM_WRITE)) {
 341                 populate_vma_page_range(vma, start, end, NULL);
 342         }
 343
 344         vm_stat_account(mm, oldflags, -nrpages);
 345         vm_stat_account(mm, newflags, nrpages);
 346         perf_event_mmap(vma);
 347         return 0;
 348
 349 fail:
 350         vm_unacct_memory(charged);
 351         return error;
 352 }
 353
 354 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 355                 unsigned long, prot)
 356 {
 357         unsigned long nstart, end, tmp, reqprot;
 358         struct vm_area_struct *vma, *prev;
 359         int error = -EINVAL;
 360         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 361         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 362                                 (prot & PROT_READ);
 363         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 364         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 365                 return -EINVAL;
 366
 367         if (start & ~PAGE_MASK)
 368                 return -EINVAL;
 369         if (!len)
 370                 return 0;
 371         len = PAGE_ALIGN(len);
 372         end = start + len;
 373         if (end <= start)
 374                 return -ENOMEM;
 375         if (!arch_validate_prot(prot))
 376                 return -EINVAL;
 377
 378         reqprot = prot;
 379
 380         down_write(&current->mm->mmap_sem);
 381
 382         vma = find_vma(current->mm, start);
 383         error = -ENOMEM;
 384         if (!vma)
 385                 goto out;
 386         prev = vma->vm_prev;
 387         if (unlikely(grows & PROT_GROWSDOWN)) {
 388                 if (vma->vm_start >= end)
 389                         goto out;
 390                 start = vma->vm_start;
 391                 error = -EINVAL;
 392                 if (!(vma->vm_flags & VM_GROWSDOWN))
 393                         goto out;
 394         } else {
 395                 if (vma->vm_start > start)
 396                         goto out;
 397                 if (unlikely(grows & PROT_GROWSUP)) {
 398                         end = vma->vm_end;
 399                         error = -EINVAL;
 400                         if (!(vma->vm_flags & VM_GROWSUP))
 401                                 goto out;
 402                 }
 403         }
 404         if (start > vma->vm_start)
 405                 prev = vma;
 406
 407         for (nstart = start ; ; ) {
 408                 unsigned long newflags;
 409
 410                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 411
 412                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 413                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 414                         prot |= PROT_EXEC;
 415
 416                 newflags = calc_vm_prot_bits(prot);
 417                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 418
 419                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 420                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 421                         error = -EACCES;
 422                         goto out;
 423                 }
 424
 425                 error = security_file_mprotect(vma, reqprot, prot);
 426                 if (error)
 427                         goto out;
 428
 429                 tmp = vma->vm_end;
 430                 if (tmp > end)
 431                         tmp = end;
 432                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 433                 if (error)
 434                         goto out;
 435                 nstart = tmp;
 436
 437                 if (nstart < prev->vm_end)
 438                         nstart = prev->vm_end;
 439                 if (nstart >= end)
 440                         goto out;
 441
 442                 vma = prev->vm_next;
 443                 if (!vma || vma->vm_start != nstart) {
 444                         error = -ENOMEM;
 445                         goto out;
 446                 }
 447                 prot = reqprot;
 448         }
 449 out:
 450         up_write(&current->mm->mmap_sem);
 451         return error;
 452 }