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                 p4d_t *p4d, 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(p4d, 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 inline unsigned long change_p4d_range(struct vm_area_struct *vma,
 216                 pgd_t *pgd, unsigned long addr, unsigned long end,
 217                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 218 {
 219         p4d_t *p4d;
 220         unsigned long next;
 221         unsigned long pages = 0;
 222
 223         p4d = p4d_offset(pgd, addr);
 224         do {
 225                 next = p4d_addr_end(addr, end);
 226                 if (p4d_none_or_clear_bad(p4d))
 227                         continue;
 228                 pages += change_pud_range(vma, p4d, addr, next, newprot,
 229                                  dirty_accountable, prot_numa);
 230         } while (p4d++, addr = next, addr != end);
 231
 232         return pages;
 233 }
 234
 235 static unsigned long change_protection_range(struct vm_area_struct *vma,
 236                 unsigned long addr, unsigned long end, pgprot_t newprot,
 237                 int dirty_accountable, int prot_numa)
 238 {
 239         struct mm_struct *mm = vma->vm_mm;
 240         pgd_t *pgd;
 241         unsigned long next;
 242         unsigned long start = addr;
 243         unsigned long pages = 0;
 244
 245         BUG_ON(addr >= end);
 246         pgd = pgd_offset(mm, addr);
 247         flush_cache_range(vma, addr, end);
 248         set_tlb_flush_pending(mm);
 249         do {
 250                 next = pgd_addr_end(addr, end);
 251                 if (pgd_none_or_clear_bad(pgd))
 252                         continue;
 253                 pages += change_p4d_range(vma, pgd, addr, next, newprot,
 254                                  dirty_accountable, prot_numa);
 255         } while (pgd++, addr = next, addr != end);
 256
 257         /* Only flush the TLB if we actually modified any entries: */
 258         if (pages)
 259                 flush_tlb_range(vma, start, end);
 260         clear_tlb_flush_pending(mm);
 261
 262         return pages;
 263 }
 264
 265 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 266                        unsigned long end, pgprot_t newprot,
 267                        int dirty_accountable, int prot_numa)
 268 {
 269         unsigned long pages;
 270
 271         if (is_vm_hugetlb_page(vma))
 272                 pages = hugetlb_change_protection(vma, start, end, newprot);
 273         else
 274                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 275
 276         return pages;
 277 }
 278
 279 int
 280 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 281         unsigned long start, unsigned long end, unsigned long newflags)
 282 {
 283         struct mm_struct *mm = vma->vm_mm;
 284         unsigned long oldflags = vma->vm_flags;
 285         long nrpages = (end - start) >> PAGE_SHIFT;
 286         unsigned long charged = 0;
 287         pgoff_t pgoff;
 288         int error;
 289         int dirty_accountable = 0;
 290
 291         if (newflags == oldflags) {
 292                 *pprev = vma;
 293                 return 0;
 294         }
 295
 296         /*
 297          * If we make a private mapping writable we increase our commit;
 298          * but (without finer accounting) cannot reduce our commit if we
 299          * make it unwritable again. hugetlb mapping were accounted for
 300          * even if read-only so there is no need to account for them here
 301          */
 302         if (newflags & VM_WRITE) {
 303                 /* Check space limits when area turns into data. */
 304                 if (!may_expand_vm(mm, newflags, nrpages) &&
 305                                 may_expand_vm(mm, oldflags, nrpages))
 306                         return -ENOMEM;
 307                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 308                                                 VM_SHARED|VM_NORESERVE))) {
 309                         charged = nrpages;
 310                         if (security_vm_enough_memory_mm(mm, charged))
 311                                 return -ENOMEM;
 312                         newflags |= VM_ACCOUNT;
 313                 }
 314         }
 315
 316         /*
 317          * First try to merge with previous and/or next vma.
 318          */
 319         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 320         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 321                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 322                            vma->vm_userfaultfd_ctx);
 323         if (*pprev) {
 324                 vma = *pprev;
 325                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 326                 goto success;
 327         }
 328
 329         *pprev = vma;
 330
 331         if (start != vma->vm_start) {
 332                 error = split_vma(mm, vma, start, 1);
 333                 if (error)
 334                         goto fail;
 335         }
 336
 337         if (end != vma->vm_end) {
 338                 error = split_vma(mm, vma, end, 0);
 339                 if (error)
 340                         goto fail;
 341         }
 342
 343 success:
 344         /*
 345          * vm_flags and vm_page_prot are protected by the mmap_sem
 346          * held in write mode.
 347          */
 348         vma->vm_flags = newflags;
 349         dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
 350         vma_set_page_prot(vma);
 351
 352         change_protection(vma, start, end, vma->vm_page_prot,
 353                           dirty_accountable, 0);
 354
 355         /*
 356          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 357          * fault on access.
 358          */
 359         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 360                         (newflags & VM_WRITE)) {
 361                 populate_vma_page_range(vma, start, end, NULL);
 362         }
 363
 364         vm_stat_account(mm, oldflags, -nrpages);
 365         vm_stat_account(mm, newflags, nrpages);
 366         perf_event_mmap(vma);
 367         return 0;
 368
 369 fail:
 370         vm_unacct_memory(charged);
 371         return error;
 372 }
 373
 374 /*
 375  * pkey==-1 when doing a legacy mprotect()
 376  */
 377 static int do_mprotect_pkey(unsigned long start, size_t len,
 378                 unsigned long prot, int pkey)
 379 {
 380         unsigned long nstart, end, tmp, reqprot;
 381         struct vm_area_struct *vma, *prev;
 382         int error = -EINVAL;
 383         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 384         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 385                                 (prot & PROT_READ);
 386
 387         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 388         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 389                 return -EINVAL;
 390
 391         if (start & ~PAGE_MASK)
 392                 return -EINVAL;
 393         if (!len)
 394                 return 0;
 395         len = PAGE_ALIGN(len);
 396         end = start + len;
 397         if (end <= start)
 398                 return -ENOMEM;
 399         if (!arch_validate_prot(prot))
 400                 return -EINVAL;
 401
 402         reqprot = prot;
 403
 404         if (down_write_killable(&current->mm->mmap_sem))
 405                 return -EINTR;
 406
 407         /*
 408          * If userspace did not allocate the pkey, do not let
 409          * them use it here.
 410          */
 411         error = -EINVAL;
 412         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 413                 goto out;
 414
 415         vma = find_vma(current->mm, start);
 416         error = -ENOMEM;
 417         if (!vma)
 418                 goto out;
 419         prev = vma->vm_prev;
 420         if (unlikely(grows & PROT_GROWSDOWN)) {
 421                 if (vma->vm_start >= end)
 422                         goto out;
 423                 start = vma->vm_start;
 424                 error = -EINVAL;
 425                 if (!(vma->vm_flags & VM_GROWSDOWN))
 426                         goto out;
 427         } else {
 428                 if (vma->vm_start > start)
 429                         goto out;
 430                 if (unlikely(grows & PROT_GROWSUP)) {
 431                         end = vma->vm_end;
 432                         error = -EINVAL;
 433                         if (!(vma->vm_flags & VM_GROWSUP))
 434                                 goto out;
 435                 }
 436         }
 437         if (start > vma->vm_start)
 438                 prev = vma;
 439
 440         for (nstart = start ; ; ) {
 441                 unsigned long mask_off_old_flags;
 442                 unsigned long newflags;
 443                 int new_vma_pkey;
 444
 445                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 446
 447                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 448                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 449                         prot |= PROT_EXEC;
 450
 451                 /*
 452                  * Each mprotect() call explicitly passes r/w/x permissions.
 453                  * If a permission is not passed to mprotect(), it must be
 454                  * cleared from the VMA.
 455                  */
 456                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 457                                         ARCH_VM_PKEY_FLAGS;
 458
 459                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 460                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 461                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 462
 463                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 464                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 465                         error = -EACCES;
 466                         goto out;
 467                 }
 468
 469                 error = security_file_mprotect(vma, reqprot, prot);
 470                 if (error)
 471                         goto out;
 472
 473                 tmp = vma->vm_end;
 474                 if (tmp > end)
 475                         tmp = end;
 476                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 477                 if (error)
 478                         goto out;
 479                 nstart = tmp;
 480
 481                 if (nstart < prev->vm_end)
 482                         nstart = prev->vm_end;
 483                 if (nstart >= end)
 484                         goto out;
 485
 486                 vma = prev->vm_next;
 487                 if (!vma || vma->vm_start != nstart) {
 488                         error = -ENOMEM;
 489                         goto out;
 490                 }
 491                 prot = reqprot;
 492         }
 493 out:
 494         up_write(&current->mm->mmap_sem);
 495         return error;
 496 }
 497
 498 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 499                 unsigned long, prot)
 500 {
 501         return do_mprotect_pkey(start, len, prot, -1);
 502 }
 503
 504 #ifdef CONFIG_ARCH_HAS_PKEYS
 505
 506 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 507                 unsigned long, prot, int, pkey)
 508 {
 509         return do_mprotect_pkey(start, len, prot, pkey);
 510 }
 511
 512 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 513 {
 514         int pkey;
 515         int ret;
 516
 517         /* No flags supported yet. */
 518         if (flags)
 519                 return -EINVAL;
 520         /* check for unsupported init values */
 521         if (init_val & ~PKEY_ACCESS_MASK)
 522                 return -EINVAL;
 523
 524         down_write(&current->mm->mmap_sem);
 525         pkey = mm_pkey_alloc(current->mm);
 526
 527         ret = -ENOSPC;
 528         if (pkey == -1)
 529                 goto out;
 530
 531         ret = arch_set_user_pkey_access(current, pkey, init_val);
 532         if (ret) {
 533                 mm_pkey_free(current->mm, pkey);
 534                 goto out;
 535         }
 536         ret = pkey;
 537 out:
 538         up_write(&current->mm->mmap_sem);
 539         return ret;
 540 }
 541
 542 SYSCALL_DEFINE1(pkey_free, int, pkey)
 543 {
 544         int ret;
 545
 546         down_write(&current->mm->mmap_sem);
 547         ret = mm_pkey_free(current->mm, pkey);
 548         up_write(&current->mm->mmap_sem);
 549
 550         /*
 551          * We could provie warnings or errors if any VMA still
 552          * has the pkey set here.
 553          */
 554         return ret;
 555 }
 556
 557 #endif /* CONFIG_ARCH_HAS_PKEYS */