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