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/pkeys.h>
  27 #include <linux/ksm.h>
  28 #include <linux/uaccess.h>
  29 #include <asm/pgtable.h>
  30 #include <asm/cacheflush.h>
  31 #include <asm/mmu_context.h>
  32 #include <asm/tlbflush.h>
  33
  34 #include "internal.h"
  35
  36 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  37                 unsigned long addr, unsigned long end, pgprot_t newprot,
  38                 int dirty_accountable, int prot_numa)
  39 {
  40         struct mm_struct *mm = vma->vm_mm;
  41         pte_t *pte, oldpte;
  42         spinlock_t *ptl;
  43         unsigned long pages = 0;
  44         int target_node = NUMA_NO_NODE;
  45
  46         /*
  47          * Can be called with only the mmap_sem for reading by
  48          * prot_numa so we must check the pmd isn't constantly
  49          * changing from under us from pmd_none to pmd_trans_huge
  50          * and/or the other way around.
  51          */
  52         if (pmd_trans_unstable(pmd))
  53                 return 0;
  54
  55         /*
  56          * The pmd points to a regular pte so the pmd can't change
  57          * from under us even if the mmap_sem is only hold for
  58          * reading.
  59          */
  60         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
  61         if (!pte)
  62                 return 0;
  63
  64         /* Get target node for single threaded private VMAs */
  65         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
  66             atomic_read(&vma->vm_mm->mm_users) == 1)
  67                 target_node = numa_node_id();
  68
  69         arch_enter_lazy_mmu_mode();
  70         do {
  71                 oldpte = *pte;
  72                 if (pte_present(oldpte)) {
  73                         pte_t ptent;
  74                         bool preserve_write = prot_numa && pte_write(oldpte);
  75
  76                         /*
  77                          * Avoid trapping faults against the zero or KSM
  78                          * pages. See similar comment in change_huge_pmd.
  79                          */
  80                         if (prot_numa) {
  81                                 struct page *page;
  82
  83                                 page = vm_normal_page(vma, addr, oldpte);
  84                                 if (!page || PageKsm(page))
  85                                         continue;
  86
  87                                 /* Avoid TLB flush if possible */
  88                                 if (pte_protnone(oldpte))
  89                                         continue;
  90
  91                                 /*
  92                                  * Don't mess with PTEs if page is already on the node
  93                                  * a single-threaded process is running on.
  94                                  */
  95                                 if (target_node == page_to_nid(page))
  96                                         continue;
  97                         }
  98
  99                         ptent = ptep_modify_prot_start(mm, addr, pte);
 100                         ptent = pte_modify(ptent, newprot);
 101                         if (preserve_write)
 102                                 ptent = pte_mk_savedwrite(ptent);
 103
 104                         /* Avoid taking write faults for known dirty pages */
 105                         if (dirty_accountable && pte_dirty(ptent) &&
 106                                         (pte_soft_dirty(ptent) ||
 107                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 108                                 ptent = pte_mkwrite(ptent);
 109                         }
 110                         ptep_modify_prot_commit(mm, addr, pte, ptent);
 111                         pages++;
 112                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 113                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 114
 115                         if (is_write_migration_entry(entry)) {
 116                                 pte_t newpte;
 117                                 /*
 118                                  * A protection check is difficult so
 119                                  * just be safe and disable write
 120                                  */
 121                                 make_migration_entry_read(&entry);
 122                                 newpte = swp_entry_to_pte(entry);
 123                                 if (pte_swp_soft_dirty(oldpte))
 124                                         newpte = pte_swp_mksoft_dirty(newpte);
 125                                 set_pte_at(mm, addr, pte, newpte);
 126
 127                                 pages++;
 128                         }
 129                 }
 130         } while (pte++, addr += PAGE_SIZE, addr != end);
 131         arch_leave_lazy_mmu_mode();
 132         pte_unmap_unlock(pte - 1, ptl);
 133
 134         return pages;
 135 }
 136
 137 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 138                 pud_t *pud, unsigned long addr, unsigned long end,
 139                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 140 {
 141         pmd_t *pmd;
 142         struct mm_struct *mm = vma->vm_mm;
 143         unsigned long next;
 144         unsigned long pages = 0;
 145         unsigned long nr_huge_updates = 0;
 146         unsigned long mni_start = 0;
 147
 148         pmd = pmd_offset(pud, addr);
 149         do {
 150                 unsigned long this_pages;
 151
 152                 next = pmd_addr_end(addr, end);
 153                 if (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
 154                                 && pmd_none_or_clear_bad(pmd))
 155                         continue;
 156
 157                 /* invoke the mmu notifier if the pmd is populated */
 158                 if (!mni_start) {
 159                         mni_start = addr;
 160                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 161                 }
 162
 163                 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 164                         if (next - addr != HPAGE_PMD_SIZE) {
 165                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 166                         } else {
 167                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 168                                                 newprot, prot_numa);
 169
 170                                 if (nr_ptes) {
 171                                         if (nr_ptes == HPAGE_PMD_NR) {
 172                                                 pages += HPAGE_PMD_NR;
 173                                                 nr_huge_updates++;
 174                                         }
 175
 176                                         /* huge pmd was handled */
 177                                         continue;
 178                                 }
 179                         }
 180                         /* fall through, the trans huge pmd just split */
 181                 }
 182                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 183                                  dirty_accountable, prot_numa);
 184                 pages += this_pages;
 185         } while (pmd++, addr = next, addr != end);
 186
 187         if (mni_start)
 188                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 189
 190         if (nr_huge_updates)
 191                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 192         return pages;
 193 }
 194
 195 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 196                 pgd_t *pgd, unsigned long addr, unsigned long end,
 197                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 198 {
 199         pud_t *pud;
 200         unsigned long next;
 201         unsigned long pages = 0;
 202
 203         pud = pud_offset(pgd, addr);
 204         do {
 205                 next = pud_addr_end(addr, end);
 206                 if (pud_none_or_clear_bad(pud))
 207                         continue;
 208                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 209                                  dirty_accountable, prot_numa);
 210         } while (pud++, addr = next, addr != end);
 211
 212         return pages;
 213 }
 214
 215 static unsigned long change_protection_range(struct vm_area_struct *vma,
 216                 unsigned long addr, unsigned long end, pgprot_t newprot,
 217                 int dirty_accountable, int prot_numa)
 218 {
 219         struct mm_struct *mm = vma->vm_mm;
 220         pgd_t *pgd;
 221         unsigned long next;
 222         unsigned long start = addr;
 223         unsigned long pages = 0;
 224
 225         BUG_ON(addr >= end);
 226         pgd = pgd_offset(mm, addr);
 227         flush_cache_range(vma, addr, end);
 228         set_tlb_flush_pending(mm);
 229         do {
 230                 next = pgd_addr_end(addr, end);
 231                 if (pgd_none_or_clear_bad(pgd))
 232                         continue;
 233                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 234                                  dirty_accountable, prot_numa);
 235         } while (pgd++, addr = next, addr != end);
 236
 237         /* Only flush the TLB if we actually modified any entries: */
 238         if (pages)
 239                 flush_tlb_range(vma, start, end);
 240         clear_tlb_flush_pending(mm);
 241
 242         return pages;
 243 }
 244
 245 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 246                        unsigned long end, pgprot_t newprot,
 247                        int dirty_accountable, int prot_numa)
 248 {
 249         unsigned long pages;
 250
 251         if (is_vm_hugetlb_page(vma))
 252                 pages = hugetlb_change_protection(vma, start, end, newprot);
 253         else
 254                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 255
 256         return pages;
 257 }
 258
 259 int
 260 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 261         unsigned long start, unsigned long end, unsigned long newflags)
 262 {
 263         struct mm_struct *mm = vma->vm_mm;
 264         unsigned long oldflags = vma->vm_flags;
 265         long nrpages = (end - start) >> PAGE_SHIFT;
 266         unsigned long charged = 0;
 267         pgoff_t pgoff;
 268         int error;
 269         int dirty_accountable = 0;
 270
 271         if (newflags == oldflags) {
 272                 *pprev = vma;
 273                 return 0;
 274         }
 275
 276         /*
 277          * If we make a private mapping writable we increase our commit;
 278          * but (without finer accounting) cannot reduce our commit if we
 279          * make it unwritable again. hugetlb mapping were accounted for
 280          * even if read-only so there is no need to account for them here
 281          */
 282         if (newflags & VM_WRITE) {
 283                 /* Check space limits when area turns into data. */
 284                 if (!may_expand_vm(mm, newflags, nrpages) &&
 285                                 may_expand_vm(mm, oldflags, nrpages))
 286                         return -ENOMEM;
 287                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 288                                                 VM_SHARED|VM_NORESERVE))) {
 289                         charged = nrpages;
 290                         if (security_vm_enough_memory_mm(mm, charged))
 291                                 return -ENOMEM;
 292                         newflags |= VM_ACCOUNT;
 293                 }
 294         }
 295
 296         /*
 297          * First try to merge with previous and/or next vma.
 298          */
 299         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 300         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 301                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 302                            vma->vm_userfaultfd_ctx);
 303         if (*pprev) {
 304                 vma = *pprev;
 305                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 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, vma->vm_page_prot);
 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 /*
 355  * pkey==-1 when doing a legacy mprotect()
 356  */
 357 static int do_mprotect_pkey(unsigned long start, size_t len,
 358                 unsigned long prot, int pkey)
 359 {
 360         unsigned long nstart, end, tmp, reqprot;
 361         struct vm_area_struct *vma, *prev;
 362         int error = -EINVAL;
 363         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 364         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 365                                 (prot & PROT_READ);
 366
 367         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 368         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 369                 return -EINVAL;
 370
 371         if (start & ~PAGE_MASK)
 372                 return -EINVAL;
 373         if (!len)
 374                 return 0;
 375         len = PAGE_ALIGN(len);
 376         end = start + len;
 377         if (end <= start)
 378                 return -ENOMEM;
 379         if (!arch_validate_prot(prot))
 380                 return -EINVAL;
 381
 382         reqprot = prot;
 383
 384         if (down_write_killable(&current->mm->mmap_sem))
 385                 return -EINTR;
 386
 387         /*
 388          * If userspace did not allocate the pkey, do not let
 389          * them use it here.
 390          */
 391         error = -EINVAL;
 392         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 393                 goto out;
 394
 395         vma = find_vma(current->mm, start);
 396         error = -ENOMEM;
 397         if (!vma)
 398                 goto out;
 399         prev = vma->vm_prev;
 400         if (unlikely(grows & PROT_GROWSDOWN)) {
 401                 if (vma->vm_start >= end)
 402                         goto out;
 403                 start = vma->vm_start;
 404                 error = -EINVAL;
 405                 if (!(vma->vm_flags & VM_GROWSDOWN))
 406                         goto out;
 407         } else {
 408                 if (vma->vm_start > start)
 409                         goto out;
 410                 if (unlikely(grows & PROT_GROWSUP)) {
 411                         end = vma->vm_end;
 412                         error = -EINVAL;
 413                         if (!(vma->vm_flags & VM_GROWSUP))
 414                                 goto out;
 415                 }
 416         }
 417         if (start > vma->vm_start)
 418                 prev = vma;
 419
 420         for (nstart = start ; ; ) {
 421                 unsigned long mask_off_old_flags;
 422                 unsigned long newflags;
 423                 int new_vma_pkey;
 424
 425                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 426
 427                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 428                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 429                         prot |= PROT_EXEC;
 430
 431                 /*
 432                  * Each mprotect() call explicitly passes r/w/x permissions.
 433                  * If a permission is not passed to mprotect(), it must be
 434                  * cleared from the VMA.
 435                  */
 436                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 437                                         ARCH_VM_PKEY_FLAGS;
 438
 439                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 440                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 441                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 442
 443                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 444                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 445                         error = -EACCES;
 446                         goto out;
 447                 }
 448
 449                 error = security_file_mprotect(vma, reqprot, prot);
 450                 if (error)
 451                         goto out;
 452
 453                 tmp = vma->vm_end;
 454                 if (tmp > end)
 455                         tmp = end;
 456                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 457                 if (error)
 458                         goto out;
 459                 nstart = tmp;
 460
 461                 if (nstart < prev->vm_end)
 462                         nstart = prev->vm_end;
 463                 if (nstart >= end)
 464                         goto out;
 465
 466                 vma = prev->vm_next;
 467                 if (!vma || vma->vm_start != nstart) {
 468                         error = -ENOMEM;
 469                         goto out;
 470                 }
 471                 prot = reqprot;
 472         }
 473 out:
 474         up_write(&current->mm->mmap_sem);
 475         return error;
 476 }
 477
 478 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 479                 unsigned long, prot)
 480 {
 481         return do_mprotect_pkey(start, len, prot, -1);
 482 }
 483
 484 #ifdef CONFIG_ARCH_HAS_PKEYS
 485
 486 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 487                 unsigned long, prot, int, pkey)
 488 {
 489         return do_mprotect_pkey(start, len, prot, pkey);
 490 }
 491
 492 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 493 {
 494         int pkey;
 495         int ret;
 496
 497         /* No flags supported yet. */
 498         if (flags)
 499                 return -EINVAL;
 500         /* check for unsupported init values */
 501         if (init_val & ~PKEY_ACCESS_MASK)
 502                 return -EINVAL;
 503
 504         down_write(&current->mm->mmap_sem);
 505         pkey = mm_pkey_alloc(current->mm);
 506
 507         ret = -ENOSPC;
 508         if (pkey == -1)
 509                 goto out;
 510
 511         ret = arch_set_user_pkey_access(current, pkey, init_val);
 512         if (ret) {
 513                 mm_pkey_free(current->mm, pkey);
 514                 goto out;
 515         }
 516         ret = pkey;
 517 out:
 518         up_write(&current->mm->mmap_sem);
 519         return ret;
 520 }
 521
 522 SYSCALL_DEFINE1(pkey_free, int, pkey)
 523 {
 524         int ret;
 525
 526         down_write(&current->mm->mmap_sem);
 527         ret = mm_pkey_free(current->mm, pkey);
 528         up_write(&current->mm->mmap_sem);
 529
 530         /*
 531          * We could provie warnings or errors if any VMA still
 532          * has the pkey set here.
 533          */
 534         return ret;
 535 }
 536
 537 #endif /* CONFIG_ARCH_HAS_PKEYS */