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 <asm/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 /*
  37  * For a prot_numa update we only hold mmap_sem for read so there is a
  38  * potential race with faulting where a pmd was temporarily none. This
  39  * function checks for a transhuge pmd under the appropriate lock. It
  40  * returns a pte if it was successfully locked or NULL if it raced with
  41  * a transhuge insertion.
  42  */
  43 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
  44                         unsigned long addr, int prot_numa, spinlock_t **ptl)
  45 {
  46         pte_t *pte;
  47         spinlock_t *pmdl;
  48
  49         /* !prot_numa is protected by mmap_sem held for write */
  50         if (!prot_numa)
  51                 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  52
  53         pmdl = pmd_lock(vma->vm_mm, pmd);
  54         if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
  55                 spin_unlock(pmdl);
  56                 return NULL;
  57         }
  58
  59         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  60         spin_unlock(pmdl);
  61         return pte;
  62 }
  63
  64 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  65                 unsigned long addr, unsigned long end, pgprot_t newprot,
  66                 int dirty_accountable, int prot_numa)
  67 {
  68         struct mm_struct *mm = vma->vm_mm;
  69         pte_t *pte, oldpte;
  70         spinlock_t *ptl;
  71         unsigned long pages = 0;
  72
  73         pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
  74         if (!pte)
  75                 return 0;
  76
  77         flush_tlb_batched_pending(vma->vm_mm);
  78         arch_enter_lazy_mmu_mode();
  79         do {
  80                 oldpte = *pte;
  81                 if (pte_present(oldpte)) {
  82                         pte_t ptent;
  83                         bool preserve_write = prot_numa && pte_write(oldpte);
  84
  85                         /*
  86                          * Avoid trapping faults against the zero or KSM
  87                          * pages. See similar comment in change_huge_pmd.
  88                          */
  89                         if (prot_numa) {
  90                                 struct page *page;
  91
  92                                 page = vm_normal_page(vma, addr, oldpte);
  93                                 if (!page || PageKsm(page))
  94                                         continue;
  95
  96                                 /* Avoid TLB flush if possible */
  97                                 if (pte_protnone(oldpte))
  98                                         continue;
  99                         }
 100
 101                         ptent = ptep_modify_prot_start(mm, addr, pte);
 102                         ptent = pte_modify(ptent, newprot);
 103                         if (preserve_write)
 104                                 ptent = pte_mkwrite(ptent);
 105
 106                         /* Avoid taking write faults for known dirty pages */
 107                         if (dirty_accountable && pte_dirty(ptent) &&
 108                                         (pte_soft_dirty(ptent) ||
 109                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 110                                 ptent = pte_mkwrite(ptent);
 111                         }
 112                         ptep_modify_prot_commit(mm, addr, pte, ptent);
 113                         pages++;
 114                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 115                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 116
 117                         if (is_write_migration_entry(entry)) {
 118                                 pte_t newpte;
 119                                 /*
 120                                  * A protection check is difficult so
 121                                  * just be safe and disable write
 122                                  */
 123                                 make_migration_entry_read(&entry);
 124                                 newpte = swp_entry_to_pte(entry);
 125                                 if (pte_swp_soft_dirty(oldpte))
 126                                         newpte = pte_swp_mksoft_dirty(newpte);
 127                                 set_pte_at(mm, addr, pte, newpte);
 128
 129                                 pages++;
 130                         }
 131                 }
 132         } while (pte++, addr += PAGE_SIZE, addr != end);
 133         arch_leave_lazy_mmu_mode();
 134         pte_unmap_unlock(pte - 1, ptl);
 135
 136         return pages;
 137 }
 138
 139 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 140                 pud_t *pud, unsigned long addr, unsigned long end,
 141                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 142 {
 143         pmd_t *pmd;
 144         struct mm_struct *mm = vma->vm_mm;
 145         unsigned long next;
 146         unsigned long pages = 0;
 147         unsigned long nr_huge_updates = 0;
 148         unsigned long mni_start = 0;
 149
 150         pmd = pmd_offset(pud, addr);
 151         do {
 152                 unsigned long this_pages;
 153
 154                 next = pmd_addr_end(addr, end);
 155                 if (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
 156                                 && pmd_none_or_clear_bad(pmd))
 157                         continue;
 158
 159                 /* invoke the mmu notifier if the pmd is populated */
 160                 if (!mni_start) {
 161                         mni_start = addr;
 162                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 163                 }
 164
 165                 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 166                         if (next - addr != HPAGE_PMD_SIZE) {
 167                                 split_huge_pmd(vma, pmd, addr);
 168                                 if (pmd_trans_unstable(pmd))
 169                                         continue;
 170                         } else {
 171                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 172                                                 newprot, prot_numa);
 173
 174                                 if (nr_ptes) {
 175                                         if (nr_ptes == HPAGE_PMD_NR) {
 176                                                 pages += HPAGE_PMD_NR;
 177                                                 nr_huge_updates++;
 178                                         }
 179
 180                                         /* huge pmd was handled */
 181                                         continue;
 182                                 }
 183                         }
 184                         /* fall through, the trans huge pmd just split */
 185                 }
 186                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 187                                  dirty_accountable, prot_numa);
 188                 pages += this_pages;
 189         } while (pmd++, addr = next, addr != end);
 190
 191         if (mni_start)
 192                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 193
 194         if (nr_huge_updates)
 195                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 196         return pages;
 197 }
 198
 199 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 200                 pgd_t *pgd, unsigned long addr, unsigned long end,
 201                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 202 {
 203         pud_t *pud;
 204         unsigned long next;
 205         unsigned long pages = 0;
 206
 207         pud = pud_offset(pgd, addr);
 208         do {
 209                 next = pud_addr_end(addr, end);
 210                 if (pud_none_or_clear_bad(pud))
 211                         continue;
 212                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 213                                  dirty_accountable, prot_numa);
 214         } while (pud++, addr = next, addr != end);
 215
 216         return pages;
 217 }
 218
 219 static unsigned long change_protection_range(struct vm_area_struct *vma,
 220                 unsigned long addr, unsigned long end, pgprot_t newprot,
 221                 int dirty_accountable, int prot_numa)
 222 {
 223         struct mm_struct *mm = vma->vm_mm;
 224         pgd_t *pgd;
 225         unsigned long next;
 226         unsigned long start = addr;
 227         unsigned long pages = 0;
 228
 229         BUG_ON(addr >= end);
 230         pgd = pgd_offset(mm, addr);
 231         flush_cache_range(vma, addr, end);
 232         set_tlb_flush_pending(mm);
 233         do {
 234                 next = pgd_addr_end(addr, end);
 235                 if (pgd_none_or_clear_bad(pgd))
 236                         continue;
 237                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 238                                  dirty_accountable, prot_numa);
 239         } while (pgd++, addr = next, addr != end);
 240
 241         /* Only flush the TLB if we actually modified any entries: */
 242         if (pages)
 243                 flush_tlb_range(vma, start, end);
 244         clear_tlb_flush_pending(mm);
 245
 246         return pages;
 247 }
 248
 249 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 250                        unsigned long end, pgprot_t newprot,
 251                        int dirty_accountable, int prot_numa)
 252 {
 253         unsigned long pages;
 254
 255         if (is_vm_hugetlb_page(vma))
 256                 pages = hugetlb_change_protection(vma, start, end, newprot);
 257         else
 258                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 259
 260         return pages;
 261 }
 262
 263 int
 264 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 265         unsigned long start, unsigned long end, unsigned long newflags)
 266 {
 267         struct mm_struct *mm = vma->vm_mm;
 268         unsigned long oldflags = vma->vm_flags;
 269         long nrpages = (end - start) >> PAGE_SHIFT;
 270         unsigned long charged = 0;
 271         pgoff_t pgoff;
 272         int error;
 273         int dirty_accountable = 0;
 274
 275         if (newflags == oldflags) {
 276                 *pprev = vma;
 277                 return 0;
 278         }
 279
 280         /*
 281          * If we make a private mapping writable we increase our commit;
 282          * but (without finer accounting) cannot reduce our commit if we
 283          * make it unwritable again. hugetlb mapping were accounted for
 284          * even if read-only so there is no need to account for them here
 285          */
 286         if (newflags & VM_WRITE) {
 287                 /* Check space limits when area turns into data. */
 288                 if (!may_expand_vm(mm, newflags, nrpages) &&
 289                                 may_expand_vm(mm, oldflags, nrpages))
 290                         return -ENOMEM;
 291                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 292                                                 VM_SHARED|VM_NORESERVE))) {
 293                         charged = nrpages;
 294                         if (security_vm_enough_memory_mm(mm, charged))
 295                                 return -ENOMEM;
 296                         newflags |= VM_ACCOUNT;
 297                 }
 298         }
 299
 300         /*
 301          * First try to merge with previous and/or next vma.
 302          */
 303         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 304         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 305                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 306                            vma->vm_userfaultfd_ctx);
 307         if (*pprev) {
 308                 vma = *pprev;
 309                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 310                 goto success;
 311         }
 312
 313         *pprev = vma;
 314
 315         if (start != vma->vm_start) {
 316                 error = split_vma(mm, vma, start, 1);
 317                 if (error)
 318                         goto fail;
 319         }
 320
 321         if (end != vma->vm_end) {
 322                 error = split_vma(mm, vma, end, 0);
 323                 if (error)
 324                         goto fail;
 325         }
 326
 327 success:
 328         /*
 329          * vm_flags and vm_page_prot are protected by the mmap_sem
 330          * held in write mode.
 331          */
 332         vma->vm_flags = newflags;
 333         dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
 334         vma_set_page_prot(vma);
 335
 336         change_protection(vma, start, end, vma->vm_page_prot,
 337                           dirty_accountable, 0);
 338
 339         /*
 340          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 341          * fault on access.
 342          */
 343         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 344                         (newflags & VM_WRITE)) {
 345                 populate_vma_page_range(vma, start, end, NULL);
 346         }
 347
 348         vm_stat_account(mm, oldflags, -nrpages);
 349         vm_stat_account(mm, newflags, nrpages);
 350         perf_event_mmap(vma);
 351         return 0;
 352
 353 fail:
 354         vm_unacct_memory(charged);
 355         return error;
 356 }
 357
 358 /*
 359  * pkey==-1 when doing a legacy mprotect()
 360  */
 361 static int do_mprotect_pkey(unsigned long start, size_t len,
 362                 unsigned long prot, int pkey)
 363 {
 364         unsigned long nstart, end, tmp, reqprot;
 365         struct vm_area_struct *vma, *prev;
 366         int error = -EINVAL;
 367         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 368         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 369                                 (prot & PROT_READ);
 370
 371         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 372         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 373                 return -EINVAL;
 374
 375         if (start & ~PAGE_MASK)
 376                 return -EINVAL;
 377         if (!len)
 378                 return 0;
 379         len = PAGE_ALIGN(len);
 380         end = start + len;
 381         if (end <= start)
 382                 return -ENOMEM;
 383         if (!arch_validate_prot(prot))
 384                 return -EINVAL;
 385
 386         reqprot = prot;
 387
 388         if (down_write_killable(&current->mm->mmap_sem))
 389                 return -EINTR;
 390
 391         /*
 392          * If userspace did not allocate the pkey, do not let
 393          * them use it here.
 394          */
 395         error = -EINVAL;
 396         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 397                 goto out;
 398
 399         vma = find_vma(current->mm, start);
 400         error = -ENOMEM;
 401         if (!vma)
 402                 goto out;
 403         prev = vma->vm_prev;
 404         if (unlikely(grows & PROT_GROWSDOWN)) {
 405                 if (vma->vm_start >= end)
 406                         goto out;
 407                 start = vma->vm_start;
 408                 error = -EINVAL;
 409                 if (!(vma->vm_flags & VM_GROWSDOWN))
 410                         goto out;
 411         } else {
 412                 if (vma->vm_start > start)
 413                         goto out;
 414                 if (unlikely(grows & PROT_GROWSUP)) {
 415                         end = vma->vm_end;
 416                         error = -EINVAL;
 417                         if (!(vma->vm_flags & VM_GROWSUP))
 418                                 goto out;
 419                 }
 420         }
 421         if (start > vma->vm_start)
 422                 prev = vma;
 423
 424         for (nstart = start ; ; ) {
 425                 unsigned long mask_off_old_flags;
 426                 unsigned long newflags;
 427                 int new_vma_pkey;
 428
 429                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 430
 431                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 432                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 433                         prot |= PROT_EXEC;
 434
 435                 /*
 436                  * Each mprotect() call explicitly passes r/w/x permissions.
 437                  * If a permission is not passed to mprotect(), it must be
 438                  * cleared from the VMA.
 439                  */
 440                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 441                                         ARCH_VM_PKEY_FLAGS;
 442
 443                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 444                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 445                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 446
 447                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 448                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 449                         error = -EACCES;
 450                         goto out;
 451                 }
 452
 453                 error = security_file_mprotect(vma, reqprot, prot);
 454                 if (error)
 455                         goto out;
 456
 457                 tmp = vma->vm_end;
 458                 if (tmp > end)
 459                         tmp = end;
 460                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 461                 if (error)
 462                         goto out;
 463                 nstart = tmp;
 464
 465                 if (nstart < prev->vm_end)
 466                         nstart = prev->vm_end;
 467                 if (nstart >= end)
 468                         goto out;
 469
 470                 vma = prev->vm_next;
 471                 if (!vma || vma->vm_start != nstart) {
 472                         error = -ENOMEM;
 473                         goto out;
 474                 }
 475                 prot = reqprot;
 476         }
 477 out:
 478         up_write(&current->mm->mmap_sem);
 479         return error;
 480 }
 481
 482 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 483                 unsigned long, prot)
 484 {
 485         return do_mprotect_pkey(start, len, prot, -1);
 486 }
 487
 488 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 489                 unsigned long, prot, int, pkey)
 490 {
 491         return do_mprotect_pkey(start, len, prot, pkey);
 492 }
 493
 494 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 495 {
 496         int pkey;
 497         int ret;
 498
 499         /* No flags supported yet. */
 500         if (flags)
 501                 return -EINVAL;
 502         /* check for unsupported init values */
 503         if (init_val & ~PKEY_ACCESS_MASK)
 504                 return -EINVAL;
 505
 506         down_write(&current->mm->mmap_sem);
 507         pkey = mm_pkey_alloc(current->mm);
 508
 509         ret = -ENOSPC;
 510         if (pkey == -1)
 511                 goto out;
 512
 513         ret = arch_set_user_pkey_access(current, pkey, init_val);
 514         if (ret) {
 515                 mm_pkey_free(current->mm, pkey);
 516                 goto out;
 517         }
 518         ret = pkey;
 519 out:
 520         up_write(&current->mm->mmap_sem);
 521         return ret;
 522 }
 523
 524 SYSCALL_DEFINE1(pkey_free, int, pkey)
 525 {
 526         int ret;
 527
 528         down_write(&current->mm->mmap_sem);
 529         ret = mm_pkey_free(current->mm, pkey);
 530         up_write(&current->mm->mmap_sem);
 531
 532         /*
 533          * We could provie warnings or errors if any VMA still
 534          * has the pkey set here.
 535          */
 536         return ret;
 537 }