mm/mprotect.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  mm/mprotect.c
   4  *
   5  *  (C) Copyright 1994 Linus Torvalds
   6  *  (C) Copyright 2002 Christoph Hellwig
   7  *
   8  *  Address space accounting code       <alan@lxorguk.ukuu.org.uk>
   9  *  (C) Copyright 2002 Red Hat Inc, All Rights Reserved
  10  */
  11
  12 #include <linux/mm.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/shm.h>
  15 #include <linux/mman.h>
  16 #include <linux/fs.h>
  17 #include <linux/highmem.h>
  18 #include <linux/security.h>
  19 #include <linux/mempolicy.h>
  20 #include <linux/personality.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/swap.h>
  23 #include <linux/swapops.h>
  24 #include <linux/mmu_notifier.h>
  25 #include <linux/migrate.h>
  26 #include <linux/perf_event.h>
  27 #include <linux/pkeys.h>
  28 #include <linux/ksm.h>
  29 #include <linux/uaccess.h>
  30 #include <asm/pgtable.h>
  31 #include <asm/cacheflush.h>
  32 #include <asm/mmu_context.h>
  33 #include <asm/tlbflush.h>
  34
  35 #include "internal.h"
  36
  37 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  38                 unsigned long addr, unsigned long end, pgprot_t newprot,
  39                 int dirty_accountable, int prot_numa)
  40 {
  41         struct mm_struct *mm = vma->vm_mm;
  42         pte_t *pte, oldpte;
  43         spinlock_t *ptl;
  44         unsigned long pages = 0;
  45         int target_node = NUMA_NO_NODE;
  46
  47         /*
  48          * Can be called with only the mmap_sem for reading by
  49          * prot_numa so we must check the pmd isn't constantly
  50          * changing from under us from pmd_none to pmd_trans_huge
  51          * and/or the other way around.
  52          */
  53         if (pmd_trans_unstable(pmd))
  54                 return 0;
  55
  56         /*
  57          * The pmd points to a regular pte so the pmd can't change
  58          * from under us even if the mmap_sem is only hold for
  59          * reading.
  60          */
  61         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
  62
  63         /* Get target node for single threaded private VMAs */
  64         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
  65             atomic_read(&vma->vm_mm->mm_users) == 1)
  66                 target_node = numa_node_id();
  67
  68         flush_tlb_batched_pending(vma->vm_mm);
  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                         if (is_write_device_private_entry(entry)) {
 131                                 pte_t newpte;
 132
 133                                 /*
 134                                  * We do not preserve soft-dirtiness. See
 135                                  * copy_one_pte() for explanation.
 136                                  */
 137                                 make_device_private_entry_read(&entry);
 138                                 newpte = swp_entry_to_pte(entry);
 139                                 set_pte_at(mm, addr, pte, newpte);
 140
 141                                 pages++;
 142                         }
 143                 }
 144         } while (pte++, addr += PAGE_SIZE, addr != end);
 145         arch_leave_lazy_mmu_mode();
 146         pte_unmap_unlock(pte - 1, ptl);
 147
 148         return pages;
 149 }
 150
 151 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 152                 pud_t *pud, unsigned long addr, unsigned long end,
 153                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 154 {
 155         pmd_t *pmd;
 156         struct mm_struct *mm = vma->vm_mm;
 157         unsigned long next;
 158         unsigned long pages = 0;
 159         unsigned long nr_huge_updates = 0;
 160         unsigned long mni_start = 0;
 161
 162         pmd = pmd_offset(pud, addr);
 163         do {
 164                 unsigned long this_pages;
 165
 166                 next = pmd_addr_end(addr, end);
 167                 if (!is_swap_pmd(*pmd) && !pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
 168                                 && pmd_none_or_clear_bad(pmd))
 169                         continue;
 170
 171                 /* invoke the mmu notifier if the pmd is populated */
 172                 if (!mni_start) {
 173                         mni_start = addr;
 174                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 175                 }
 176
 177                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 178                         if (next - addr != HPAGE_PMD_SIZE) {
 179                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 180                         } else {
 181                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 182                                                 newprot, prot_numa);
 183
 184                                 if (nr_ptes) {
 185                                         if (nr_ptes == HPAGE_PMD_NR) {
 186                                                 pages += HPAGE_PMD_NR;
 187                                                 nr_huge_updates++;
 188                                         }
 189
 190                                         /* huge pmd was handled */
 191                                         continue;
 192                                 }
 193                         }
 194                         /* fall through, the trans huge pmd just split */
 195                 }
 196                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 197                                  dirty_accountable, prot_numa);
 198                 pages += this_pages;
 199         } while (pmd++, addr = next, addr != end);
 200
 201         if (mni_start)
 202                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 203
 204         if (nr_huge_updates)
 205                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 206         return pages;
 207 }
 208
 209 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 210                 p4d_t *p4d, unsigned long addr, unsigned long end,
 211                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 212 {
 213         pud_t *pud;
 214         unsigned long next;
 215         unsigned long pages = 0;
 216
 217         pud = pud_offset(p4d, addr);
 218         do {
 219                 next = pud_addr_end(addr, end);
 220                 if (pud_none_or_clear_bad(pud))
 221                         continue;
 222                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 223                                  dirty_accountable, prot_numa);
 224         } while (pud++, addr = next, addr != end);
 225
 226         return pages;
 227 }
 228
 229 static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
 230                 pgd_t *pgd, unsigned long addr, unsigned long end,
 231                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 232 {
 233         p4d_t *p4d;
 234         unsigned long next;
 235         unsigned long pages = 0;
 236
 237         p4d = p4d_offset(pgd, addr);
 238         do {
 239                 next = p4d_addr_end(addr, end);
 240                 if (p4d_none_or_clear_bad(p4d))
 241                         continue;
 242                 pages += change_pud_range(vma, p4d, addr, next, newprot,
 243                                  dirty_accountable, prot_numa);
 244         } while (p4d++, addr = next, addr != end);
 245
 246         return pages;
 247 }
 248
 249 static unsigned long change_protection_range(struct vm_area_struct *vma,
 250                 unsigned long addr, unsigned long end, pgprot_t newprot,
 251                 int dirty_accountable, int prot_numa)
 252 {
 253         struct mm_struct *mm = vma->vm_mm;
 254         pgd_t *pgd;
 255         unsigned long next;
 256         unsigned long start = addr;
 257         unsigned long pages = 0;
 258
 259         BUG_ON(addr >= end);
 260         pgd = pgd_offset(mm, addr);
 261         flush_cache_range(vma, addr, end);
 262         inc_tlb_flush_pending(mm);
 263         do {
 264                 next = pgd_addr_end(addr, end);
 265                 if (pgd_none_or_clear_bad(pgd))
 266                         continue;
 267                 pages += change_p4d_range(vma, pgd, addr, next, newprot,
 268                                  dirty_accountable, prot_numa);
 269         } while (pgd++, addr = next, addr != end);
 270
 271         /* Only flush the TLB if we actually modified any entries: */
 272         if (pages)
 273                 flush_tlb_range(vma, start, end);
 274         dec_tlb_flush_pending(mm);
 275
 276         return pages;
 277 }
 278
 279 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 280                        unsigned long end, pgprot_t newprot,
 281                        int dirty_accountable, int prot_numa)
 282 {
 283         unsigned long pages;
 284
 285         if (is_vm_hugetlb_page(vma))
 286                 pages = hugetlb_change_protection(vma, start, end, newprot);
 287         else
 288                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 289
 290         return pages;
 291 }
 292
 293 int
 294 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 295         unsigned long start, unsigned long end, unsigned long newflags)
 296 {
 297         struct mm_struct *mm = vma->vm_mm;
 298         unsigned long oldflags = vma->vm_flags;
 299         long nrpages = (end - start) >> PAGE_SHIFT;
 300         unsigned long charged = 0;
 301         pgoff_t pgoff;
 302         int error;
 303         int dirty_accountable = 0;
 304
 305         if (newflags == oldflags) {
 306                 *pprev = vma;
 307                 return 0;
 308         }
 309
 310         /*
 311          * If we make a private mapping writable we increase our commit;
 312          * but (without finer accounting) cannot reduce our commit if we
 313          * make it unwritable again. hugetlb mapping were accounted for
 314          * even if read-only so there is no need to account for them here
 315          */
 316         if (newflags & VM_WRITE) {
 317                 /* Check space limits when area turns into data. */
 318                 if (!may_expand_vm(mm, newflags, nrpages) &&
 319                                 may_expand_vm(mm, oldflags, nrpages))
 320                         return -ENOMEM;
 321                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 322                                                 VM_SHARED|VM_NORESERVE))) {
 323                         charged = nrpages;
 324                         if (security_vm_enough_memory_mm(mm, charged))
 325                                 return -ENOMEM;
 326                         newflags |= VM_ACCOUNT;
 327                 }
 328         }
 329
 330         /*
 331          * First try to merge with previous and/or next vma.
 332          */
 333         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 334         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 335                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 336                            vma->vm_userfaultfd_ctx);
 337         if (*pprev) {
 338                 vma = *pprev;
 339                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 340                 goto success;
 341         }
 342
 343         *pprev = vma;
 344
 345         if (start != vma->vm_start) {
 346                 error = split_vma(mm, vma, start, 1);
 347                 if (error)
 348                         goto fail;
 349         }
 350
 351         if (end != vma->vm_end) {
 352                 error = split_vma(mm, vma, end, 0);
 353                 if (error)
 354                         goto fail;
 355         }
 356
 357 success:
 358         /*
 359          * vm_flags and vm_page_prot are protected by the mmap_sem
 360          * held in write mode.
 361          */
 362         vma->vm_flags = newflags;
 363         dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
 364         vma_set_page_prot(vma);
 365
 366         change_protection(vma, start, end, vma->vm_page_prot,
 367                           dirty_accountable, 0);
 368
 369         /*
 370          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 371          * fault on access.
 372          */
 373         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 374                         (newflags & VM_WRITE)) {
 375                 populate_vma_page_range(vma, start, end, NULL);
 376         }
 377
 378         vm_stat_account(mm, oldflags, -nrpages);
 379         vm_stat_account(mm, newflags, nrpages);
 380         perf_event_mmap(vma);
 381         return 0;
 382
 383 fail:
 384         vm_unacct_memory(charged);
 385         return error;
 386 }
 387
 388 /*
 389  * pkey==-1 when doing a legacy mprotect()
 390  */
 391 static int do_mprotect_pkey(unsigned long start, size_t len,
 392                 unsigned long prot, int pkey)
 393 {
 394         unsigned long nstart, end, tmp, reqprot;
 395         struct vm_area_struct *vma, *prev;
 396         int error = -EINVAL;
 397         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 398         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 399                                 (prot & PROT_READ);
 400
 401         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 402         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 403                 return -EINVAL;
 404
 405         if (start & ~PAGE_MASK)
 406                 return -EINVAL;
 407         if (!len)
 408                 return 0;
 409         len = PAGE_ALIGN(len);
 410         end = start + len;
 411         if (end <= start)
 412                 return -ENOMEM;
 413         if (!arch_validate_prot(prot))
 414                 return -EINVAL;
 415
 416         reqprot = prot;
 417
 418         if (down_write_killable(&current->mm->mmap_sem))
 419                 return -EINTR;
 420
 421         /*
 422          * If userspace did not allocate the pkey, do not let
 423          * them use it here.
 424          */
 425         error = -EINVAL;
 426         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 427                 goto out;
 428
 429         vma = find_vma(current->mm, start);
 430         error = -ENOMEM;
 431         if (!vma)
 432                 goto out;
 433         prev = vma->vm_prev;
 434         if (unlikely(grows & PROT_GROWSDOWN)) {
 435                 if (vma->vm_start >= end)
 436                         goto out;
 437                 start = vma->vm_start;
 438                 error = -EINVAL;
 439                 if (!(vma->vm_flags & VM_GROWSDOWN))
 440                         goto out;
 441         } else {
 442                 if (vma->vm_start > start)
 443                         goto out;
 444                 if (unlikely(grows & PROT_GROWSUP)) {
 445                         end = vma->vm_end;
 446                         error = -EINVAL;
 447                         if (!(vma->vm_flags & VM_GROWSUP))
 448                                 goto out;
 449                 }
 450         }
 451         if (start > vma->vm_start)
 452                 prev = vma;
 453
 454         for (nstart = start ; ; ) {
 455                 unsigned long mask_off_old_flags;
 456                 unsigned long newflags;
 457                 int new_vma_pkey;
 458
 459                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 460
 461                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 462                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 463                         prot |= PROT_EXEC;
 464
 465                 /*
 466                  * Each mprotect() call explicitly passes r/w/x permissions.
 467                  * If a permission is not passed to mprotect(), it must be
 468                  * cleared from the VMA.
 469                  */
 470                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 471                                         ARCH_VM_PKEY_FLAGS;
 472
 473                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 474                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 475                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 476
 477                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 478                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 479                         error = -EACCES;
 480                         goto out;
 481                 }
 482
 483                 error = security_file_mprotect(vma, reqprot, prot);
 484                 if (error)
 485                         goto out;
 486
 487                 tmp = vma->vm_end;
 488                 if (tmp > end)
 489                         tmp = end;
 490                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 491                 if (error)
 492                         goto out;
 493                 nstart = tmp;
 494
 495                 if (nstart < prev->vm_end)
 496                         nstart = prev->vm_end;
 497                 if (nstart >= end)
 498                         goto out;
 499
 500                 vma = prev->vm_next;
 501                 if (!vma || vma->vm_start != nstart) {
 502                         error = -ENOMEM;
 503                         goto out;
 504                 }
 505                 prot = reqprot;
 506         }
 507 out:
 508         up_write(&current->mm->mmap_sem);
 509         return error;
 510 }
 511
 512 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 513                 unsigned long, prot)
 514 {
 515         return do_mprotect_pkey(start, len, prot, -1);
 516 }
 517
 518 #ifdef CONFIG_ARCH_HAS_PKEYS
 519
 520 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 521                 unsigned long, prot, int, pkey)
 522 {
 523         return do_mprotect_pkey(start, len, prot, pkey);
 524 }
 525
 526 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 527 {
 528         int pkey;
 529         int ret;
 530
 531         /* No flags supported yet. */
 532         if (flags)
 533                 return -EINVAL;
 534         /* check for unsupported init values */
 535         if (init_val & ~PKEY_ACCESS_MASK)
 536                 return -EINVAL;
 537
 538         down_write(&current->mm->mmap_sem);
 539         pkey = mm_pkey_alloc(current->mm);
 540
 541         ret = -ENOSPC;
 542         if (pkey == -1)
 543                 goto out;
 544
 545         ret = arch_set_user_pkey_access(current, pkey, init_val);
 546         if (ret) {
 547                 mm_pkey_free(current->mm, pkey);
 548                 goto out;
 549         }
 550         ret = pkey;
 551 out:
 552         up_write(&current->mm->mmap_sem);
 553         return ret;
 554 }
 555
 556 SYSCALL_DEFINE1(pkey_free, int, pkey)
 557 {
 558         int ret;
 559
 560         down_write(&current->mm->mmap_sem);
 561         ret = mm_pkey_free(current->mm, pkey);
 562         up_write(&current->mm->mmap_sem);
 563
 564         /*
 565          * We could provie warnings or errors if any VMA still
 566          * has the pkey set here.
 567          */
 568         return ret;
 569 }
 570
 571 #endif /* CONFIG_ARCH_HAS_PKEYS */