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       <[email protected]>
   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                                 /* Also skip shared copy-on-write pages */
  88                                 if (is_cow_mapping(vma->vm_flags) &&
  89                                     page_mapcount(page) != 1)
  90                                         continue;
  91
  92                                 /* Avoid TLB flush if possible */
  93                                 if (pte_protnone(oldpte))
  94                                         continue;
  95
  96                                 /*
  97                                  * Don't mess with PTEs if page is already on the node
  98                                  * a single-threaded process is running on.
  99                                  */
 100                                 if (target_node == page_to_nid(page))
 101                                         continue;
 102                         }
 103
 104                         ptent = ptep_modify_prot_start(mm, addr, pte);
 105                         ptent = pte_modify(ptent, newprot);
 106                         if (preserve_write)
 107                                 ptent = pte_mk_savedwrite(ptent);
 108
 109                         /* Avoid taking write faults for known dirty pages */
 110                         if (dirty_accountable && pte_dirty(ptent) &&
 111                                         (pte_soft_dirty(ptent) ||
 112                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 113                                 ptent = pte_mkwrite(ptent);
 114                         }
 115                         ptep_modify_prot_commit(mm, addr, pte, ptent);
 116                         pages++;
 117                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 118                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 119
 120                         if (is_write_migration_entry(entry)) {
 121                                 pte_t newpte;
 122                                 /*
 123                                  * A protection check is difficult so
 124                                  * just be safe and disable write
 125                                  */
 126                                 make_migration_entry_read(&entry);
 127                                 newpte = swp_entry_to_pte(entry);
 128                                 if (pte_swp_soft_dirty(oldpte))
 129                                         newpte = pte_swp_mksoft_dirty(newpte);
 130                                 set_pte_at(mm, addr, pte, newpte);
 131
 132                                 pages++;
 133                         }
 134
 135                         if (is_write_device_private_entry(entry)) {
 136                                 pte_t newpte;
 137
 138                                 /*
 139                                  * We do not preserve soft-dirtiness. See
 140                                  * copy_one_pte() for explanation.
 141                                  */
 142                                 make_device_private_entry_read(&entry);
 143                                 newpte = swp_entry_to_pte(entry);
 144                                 set_pte_at(mm, addr, pte, newpte);
 145
 146                                 pages++;
 147                         }
 148                 }
 149         } while (pte++, addr += PAGE_SIZE, addr != end);
 150         arch_leave_lazy_mmu_mode();
 151         pte_unmap_unlock(pte - 1, ptl);
 152
 153         return pages;
 154 }
 155
 156 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 157                 pud_t *pud, unsigned long addr, unsigned long end,
 158                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 159 {
 160         pmd_t *pmd;
 161         struct mm_struct *mm = vma->vm_mm;
 162         unsigned long next;
 163         unsigned long pages = 0;
 164         unsigned long nr_huge_updates = 0;
 165         unsigned long mni_start = 0;
 166
 167         pmd = pmd_offset(pud, addr);
 168         do {
 169                 unsigned long this_pages;
 170
 171                 next = pmd_addr_end(addr, end);
 172                 if (!is_swap_pmd(*pmd) && !pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
 173                                 && pmd_none_or_clear_bad(pmd))
 174                         goto next;
 175
 176                 /* invoke the mmu notifier if the pmd is populated */
 177                 if (!mni_start) {
 178                         mni_start = addr;
 179                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 180                 }
 181
 182                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 183                         if (next - addr != HPAGE_PMD_SIZE) {
 184                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 185                         } else {
 186                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 187                                                 newprot, prot_numa);
 188
 189                                 if (nr_ptes) {
 190                                         if (nr_ptes == HPAGE_PMD_NR) {
 191                                                 pages += HPAGE_PMD_NR;
 192                                                 nr_huge_updates++;
 193                                         }
 194
 195                                         /* huge pmd was handled */
 196                                         goto next;
 197                                 }
 198                         }
 199                         /* fall through, the trans huge pmd just split */
 200                 }
 201                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 202                                  dirty_accountable, prot_numa);
 203                 pages += this_pages;
 204 next:
 205                 cond_resched();
 206         } while (pmd++, addr = next, addr != end);
 207
 208         if (mni_start)
 209                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 210
 211         if (nr_huge_updates)
 212                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 213         return pages;
 214 }
 215
 216 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 217                 p4d_t *p4d, unsigned long addr, unsigned long end,
 218                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 219 {
 220         pud_t *pud;
 221         unsigned long next;
 222         unsigned long pages = 0;
 223
 224         pud = pud_offset(p4d, addr);
 225         do {
 226                 next = pud_addr_end(addr, end);
 227                 if (pud_none_or_clear_bad(pud))
 228                         continue;
 229                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 230                                  dirty_accountable, prot_numa);
 231         } while (pud++, addr = next, addr != end);
 232
 233         return pages;
 234 }
 235
 236 static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
 237                 pgd_t *pgd, unsigned long addr, unsigned long end,
 238                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 239 {
 240         p4d_t *p4d;
 241         unsigned long next;
 242         unsigned long pages = 0;
 243
 244         p4d = p4d_offset(pgd, addr);
 245         do {
 246                 next = p4d_addr_end(addr, end);
 247                 if (p4d_none_or_clear_bad(p4d))
 248                         continue;
 249                 pages += change_pud_range(vma, p4d, addr, next, newprot,
 250                                  dirty_accountable, prot_numa);
 251         } while (p4d++, addr = next, addr != end);
 252
 253         return pages;
 254 }
 255
 256 static unsigned long change_protection_range(struct vm_area_struct *vma,
 257                 unsigned long addr, unsigned long end, pgprot_t newprot,
 258                 int dirty_accountable, int prot_numa)
 259 {
 260         struct mm_struct *mm = vma->vm_mm;
 261         pgd_t *pgd;
 262         unsigned long next;
 263         unsigned long start = addr;
 264         unsigned long pages = 0;
 265
 266         BUG_ON(addr >= end);
 267         pgd = pgd_offset(mm, addr);
 268         flush_cache_range(vma, addr, end);
 269         inc_tlb_flush_pending(mm);
 270         do {
 271                 next = pgd_addr_end(addr, end);
 272                 if (pgd_none_or_clear_bad(pgd))
 273                         continue;
 274                 pages += change_p4d_range(vma, pgd, addr, next, newprot,
 275                                  dirty_accountable, prot_numa);
 276         } while (pgd++, addr = next, addr != end);
 277
 278         /* Only flush the TLB if we actually modified any entries: */
 279         if (pages)
 280                 flush_tlb_range(vma, start, end);
 281         dec_tlb_flush_pending(mm);
 282
 283         return pages;
 284 }
 285
 286 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 287                        unsigned long end, pgprot_t newprot,
 288                        int dirty_accountable, int prot_numa)
 289 {
 290         unsigned long pages;
 291
 292         if (is_vm_hugetlb_page(vma))
 293                 pages = hugetlb_change_protection(vma, start, end, newprot);
 294         else
 295                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 296
 297         return pages;
 298 }
 299
 300 int
 301 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 302         unsigned long start, unsigned long end, unsigned long newflags)
 303 {
 304         struct mm_struct *mm = vma->vm_mm;
 305         unsigned long oldflags = vma->vm_flags;
 306         long nrpages = (end - start) >> PAGE_SHIFT;
 307         unsigned long charged = 0;
 308         pgoff_t pgoff;
 309         int error;
 310         int dirty_accountable = 0;
 311
 312         if (newflags == oldflags) {
 313                 *pprev = vma;
 314                 return 0;
 315         }
 316
 317         /*
 318          * If we make a private mapping writable we increase our commit;
 319          * but (without finer accounting) cannot reduce our commit if we
 320          * make it unwritable again. hugetlb mapping were accounted for
 321          * even if read-only so there is no need to account for them here
 322          */
 323         if (newflags & VM_WRITE) {
 324                 /* Check space limits when area turns into data. */
 325                 if (!may_expand_vm(mm, newflags, nrpages) &&
 326                                 may_expand_vm(mm, oldflags, nrpages))
 327                         return -ENOMEM;
 328                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 329                                                 VM_SHARED|VM_NORESERVE))) {
 330                         charged = nrpages;
 331                         if (security_vm_enough_memory_mm(mm, charged))
 332                                 return -ENOMEM;
 333                         newflags |= VM_ACCOUNT;
 334                 }
 335         }
 336
 337         /*
 338          * First try to merge with previous and/or next vma.
 339          */
 340         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 341         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 342                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 343                            vma->vm_userfaultfd_ctx);
 344         if (*pprev) {
 345                 vma = *pprev;
 346                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 347                 goto success;
 348         }
 349
 350         *pprev = vma;
 351
 352         if (start != vma->vm_start) {
 353                 error = split_vma(mm, vma, start, 1);
 354                 if (error)
 355                         goto fail;
 356         }
 357
 358         if (end != vma->vm_end) {
 359                 error = split_vma(mm, vma, end, 0);
 360                 if (error)
 361                         goto fail;
 362         }
 363
 364 success:
 365         /*
 366          * vm_flags and vm_page_prot are protected by the mmap_sem
 367          * held in write mode.
 368          */
 369         vma->vm_flags = newflags;
 370         dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
 371         vma_set_page_prot(vma);
 372
 373         change_protection(vma, start, end, vma->vm_page_prot,
 374                           dirty_accountable, 0);
 375
 376         /*
 377          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 378          * fault on access.
 379          */
 380         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 381                         (newflags & VM_WRITE)) {
 382                 populate_vma_page_range(vma, start, end, NULL);
 383         }
 384
 385         vm_stat_account(mm, oldflags, -nrpages);
 386         vm_stat_account(mm, newflags, nrpages);
 387         perf_event_mmap(vma);
 388         return 0;
 389
 390 fail:
 391         vm_unacct_memory(charged);
 392         return error;
 393 }
 394
 395 /*
 396  * pkey==-1 when doing a legacy mprotect()
 397  */
 398 static int do_mprotect_pkey(unsigned long start, size_t len,
 399                 unsigned long prot, int pkey)
 400 {
 401         unsigned long nstart, end, tmp, reqprot;
 402         struct vm_area_struct *vma, *prev;
 403         int error = -EINVAL;
 404         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 405         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 406                                 (prot & PROT_READ);
 407
 408         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 409         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 410                 return -EINVAL;
 411
 412         if (start & ~PAGE_MASK)
 413                 return -EINVAL;
 414         if (!len)
 415                 return 0;
 416         len = PAGE_ALIGN(len);
 417         end = start + len;
 418         if (end <= start)
 419                 return -ENOMEM;
 420         if (!arch_validate_prot(prot))
 421                 return -EINVAL;
 422
 423         reqprot = prot;
 424
 425         if (down_write_killable(&current->mm->mmap_sem))
 426                 return -EINTR;
 427
 428         /*
 429          * If userspace did not allocate the pkey, do not let
 430          * them use it here.
 431          */
 432         error = -EINVAL;
 433         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 434                 goto out;
 435
 436         vma = find_vma(current->mm, start);
 437         error = -ENOMEM;
 438         if (!vma)
 439                 goto out;
 440         prev = vma->vm_prev;
 441         if (unlikely(grows & PROT_GROWSDOWN)) {
 442                 if (vma->vm_start >= end)
 443                         goto out;
 444                 start = vma->vm_start;
 445                 error = -EINVAL;
 446                 if (!(vma->vm_flags & VM_GROWSDOWN))
 447                         goto out;
 448         } else {
 449                 if (vma->vm_start > start)
 450                         goto out;
 451                 if (unlikely(grows & PROT_GROWSUP)) {
 452                         end = vma->vm_end;
 453                         error = -EINVAL;
 454                         if (!(vma->vm_flags & VM_GROWSUP))
 455                                 goto out;
 456                 }
 457         }
 458         if (start > vma->vm_start)
 459                 prev = vma;
 460
 461         for (nstart = start ; ; ) {
 462                 unsigned long mask_off_old_flags;
 463                 unsigned long newflags;
 464                 int new_vma_pkey;
 465
 466                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 467
 468                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 469                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 470                         prot |= PROT_EXEC;
 471
 472                 /*
 473                  * Each mprotect() call explicitly passes r/w/x permissions.
 474                  * If a permission is not passed to mprotect(), it must be
 475                  * cleared from the VMA.
 476                  */
 477                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 478                                         ARCH_VM_PKEY_FLAGS;
 479
 480                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 481                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 482                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 483
 484                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 485                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 486                         error = -EACCES;
 487                         goto out;
 488                 }
 489
 490                 error = security_file_mprotect(vma, reqprot, prot);
 491                 if (error)
 492                         goto out;
 493
 494                 tmp = vma->vm_end;
 495                 if (tmp > end)
 496                         tmp = end;
 497                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 498                 if (error)
 499                         goto out;
 500                 nstart = tmp;
 501
 502                 if (nstart < prev->vm_end)
 503                         nstart = prev->vm_end;
 504                 if (nstart >= end)
 505                         goto out;
 506
 507                 vma = prev->vm_next;
 508                 if (!vma || vma->vm_start != nstart) {
 509                         error = -ENOMEM;
 510                         goto out;
 511                 }
 512                 prot = reqprot;
 513         }
 514 out:
 515         up_write(&current->mm->mmap_sem);
 516         return error;
 517 }
 518
 519 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 520                 unsigned long, prot)
 521 {
 522         return do_mprotect_pkey(start, len, prot, -1);
 523 }
 524
 525 #ifdef CONFIG_ARCH_HAS_PKEYS
 526
 527 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 528                 unsigned long, prot, int, pkey)
 529 {
 530         return do_mprotect_pkey(start, len, prot, pkey);
 531 }
 532
 533 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 534 {
 535         int pkey;
 536         int ret;
 537
 538         /* No flags supported yet. */
 539         if (flags)
 540                 return -EINVAL;
 541         /* check for unsupported init values */
 542         if (init_val & ~PKEY_ACCESS_MASK)
 543                 return -EINVAL;
 544
 545         down_write(&current->mm->mmap_sem);
 546         pkey = mm_pkey_alloc(current->mm);
 547
 548         ret = -ENOSPC;
 549         if (pkey == -1)
 550                 goto out;
 551
 552         ret = arch_set_user_pkey_access(current, pkey, init_val);
 553         if (ret) {
 554                 mm_pkey_free(current->mm, pkey);
 555                 goto out;
 556         }
 557         ret = pkey;
 558 out:
 559         up_write(&current->mm->mmap_sem);
 560         return ret;
 561 }
 562
 563 SYSCALL_DEFINE1(pkey_free, int, pkey)
 564 {
 565         int ret;
 566
 567         down_write(&current->mm->mmap_sem);
 568         ret = mm_pkey_free(current->mm, pkey);
 569         up_write(&current->mm->mmap_sem);
 570
 571         /*
 572          * We could provie warnings or errors if any VMA still
 573          * has the pkey set here.
 574          */
 575         return ret;
 576 }
 577
 578 #endif /* CONFIG_ARCH_HAS_PKEYS */