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/ksm.h>
  27 #include <linux/pkeys.h>
  28 #include <asm/uaccess.h>
  29 #include <asm/pgtable.h>
  30 #include <asm/cacheflush.h>
  31 #include <asm/tlbflush.h>
  32
  33 #include "internal.h"
  34
  35 /*
  36  * For a prot_numa update we only hold mmap_sem for read so there is a
  37  * potential race with faulting where a pmd was temporarily none. This
  38  * function checks for a transhuge pmd under the appropriate lock. It
  39  * returns a pte if it was successfully locked or NULL if it raced with
  40  * a transhuge insertion.
  41  */
  42 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
  43                         unsigned long addr, int prot_numa, spinlock_t **ptl)
  44 {
  45         pte_t *pte;
  46         spinlock_t *pmdl;
  47
  48         /* !prot_numa is protected by mmap_sem held for write */
  49         if (!prot_numa)
  50                 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  51
  52         pmdl = pmd_lock(vma->vm_mm, pmd);
  53         if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
  54                 spin_unlock(pmdl);
  55                 return NULL;
  56         }
  57
  58         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  59         spin_unlock(pmdl);
  60         return pte;
  61 }
  62
  63 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  64                 unsigned long addr, unsigned long end, pgprot_t newprot,
  65                 int dirty_accountable, int prot_numa)
  66 {
  67         struct mm_struct *mm = vma->vm_mm;
  68         pte_t *pte, oldpte;
  69         spinlock_t *ptl;
  70         unsigned long pages = 0;
  71
  72         pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
  73         if (!pte)
  74                 return 0;
  75
  76         arch_enter_lazy_mmu_mode();
  77         do {
  78                 oldpte = *pte;
  79                 if (pte_present(oldpte)) {
  80                         pte_t ptent;
  81                         bool preserve_write = prot_numa && pte_write(oldpte);
  82
  83                         /*
  84                          * Avoid trapping faults against the zero or KSM
  85                          * pages. See similar comment in change_huge_pmd.
  86                          */
  87                         if (prot_numa) {
  88                                 struct page *page;
  89
  90                                 page = vm_normal_page(vma, addr, oldpte);
  91                                 if (!page || PageKsm(page))
  92                                         continue;
  93
  94                                 /* Avoid TLB flush if possible */
  95                                 if (pte_protnone(oldpte))
  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_mkwrite(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);
 166                                 if (pmd_none(*pmd))
 167                                         continue;
 168                         } else {
 169                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 170                                                 newprot, prot_numa);
 171
 172                                 if (nr_ptes) {
 173                                         if (nr_ptes == HPAGE_PMD_NR) {
 174                                                 pages += HPAGE_PMD_NR;
 175                                                 nr_huge_updates++;
 176                                         }
 177
 178                                         /* huge pmd was handled */
 179                                         continue;
 180                                 }
 181                         }
 182                         /* fall through, the trans huge pmd just split */
 183                 }
 184                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 185                                  dirty_accountable, prot_numa);
 186                 pages += this_pages;
 187         } while (pmd++, addr = next, addr != end);
 188
 189         if (mni_start)
 190                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 191
 192         if (nr_huge_updates)
 193                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 194         return pages;
 195 }
 196
 197 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 198                 pgd_t *pgd, unsigned long addr, unsigned long end,
 199                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 200 {
 201         pud_t *pud;
 202         unsigned long next;
 203         unsigned long pages = 0;
 204
 205         pud = pud_offset(pgd, addr);
 206         do {
 207                 next = pud_addr_end(addr, end);
 208                 if (pud_none_or_clear_bad(pud))
 209                         continue;
 210                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 211                                  dirty_accountable, prot_numa);
 212         } while (pud++, addr = next, addr != end);
 213
 214         return pages;
 215 }
 216
 217 static unsigned long change_protection_range(struct vm_area_struct *vma,
 218                 unsigned long addr, unsigned long end, pgprot_t newprot,
 219                 int dirty_accountable, int prot_numa)
 220 {
 221         struct mm_struct *mm = vma->vm_mm;
 222         pgd_t *pgd;
 223         unsigned long next;
 224         unsigned long start = addr;
 225         unsigned long pages = 0;
 226
 227         BUG_ON(addr >= end);
 228         pgd = pgd_offset(mm, addr);
 229         flush_cache_range(vma, addr, end);
 230         set_tlb_flush_pending(mm);
 231         do {
 232                 next = pgd_addr_end(addr, end);
 233                 if (pgd_none_or_clear_bad(pgd))
 234                         continue;
 235                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 236                                  dirty_accountable, prot_numa);
 237         } while (pgd++, addr = next, addr != end);
 238
 239         /* Only flush the TLB if we actually modified any entries: */
 240         if (pages)
 241                 flush_tlb_range(vma, start, end);
 242         clear_tlb_flush_pending(mm);
 243
 244         return pages;
 245 }
 246
 247 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 248                        unsigned long end, pgprot_t newprot,
 249                        int dirty_accountable, int prot_numa)
 250 {
 251         unsigned long pages;
 252
 253         if (is_vm_hugetlb_page(vma))
 254                 pages = hugetlb_change_protection(vma, start, end, newprot);
 255         else
 256                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 257
 258         return pages;
 259 }
 260
 261 int
 262 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 263         unsigned long start, unsigned long end, unsigned long newflags)
 264 {
 265         struct mm_struct *mm = vma->vm_mm;
 266         unsigned long oldflags = vma->vm_flags;
 267         long nrpages = (end - start) >> PAGE_SHIFT;
 268         unsigned long charged = 0;
 269         pgoff_t pgoff;
 270         int error;
 271         int dirty_accountable = 0;
 272
 273         if (newflags == oldflags) {
 274                 *pprev = vma;
 275                 return 0;
 276         }
 277
 278         /*
 279          * If we make a private mapping writable we increase our commit;
 280          * but (without finer accounting) cannot reduce our commit if we
 281          * make it unwritable again. hugetlb mapping were accounted for
 282          * even if read-only so there is no need to account for them here
 283          */
 284         if (newflags & VM_WRITE) {
 285                 /* Check space limits when area turns into data. */
 286                 if (!may_expand_vm(mm, newflags, nrpages) &&
 287                                 may_expand_vm(mm, oldflags, nrpages))
 288                         return -ENOMEM;
 289                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 290                                                 VM_SHARED|VM_NORESERVE))) {
 291                         charged = nrpages;
 292                         if (security_vm_enough_memory_mm(mm, charged))
 293                                 return -ENOMEM;
 294                         newflags |= VM_ACCOUNT;
 295                 }
 296         }
 297
 298         /*
 299          * First try to merge with previous and/or next vma.
 300          */
 301         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 302         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 303                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 304                            vma->vm_userfaultfd_ctx);
 305         if (*pprev) {
 306                 vma = *pprev;
 307                 goto success;
 308         }
 309
 310         *pprev = vma;
 311
 312         if (start != vma->vm_start) {
 313                 error = split_vma(mm, vma, start, 1);
 314                 if (error)
 315                         goto fail;
 316         }
 317
 318         if (end != vma->vm_end) {
 319                 error = split_vma(mm, vma, end, 0);
 320                 if (error)
 321                         goto fail;
 322         }
 323
 324 success:
 325         /*
 326          * vm_flags and vm_page_prot are protected by the mmap_sem
 327          * held in write mode.
 328          */
 329         vma->vm_flags = newflags;
 330         dirty_accountable = vma_wants_writenotify(vma);
 331         vma_set_page_prot(vma);
 332
 333         change_protection(vma, start, end, vma->vm_page_prot,
 334                           dirty_accountable, 0);
 335
 336         /*
 337          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 338          * fault on access.
 339          */
 340         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 341                         (newflags & VM_WRITE)) {
 342                 populate_vma_page_range(vma, start, end, NULL);
 343         }
 344
 345         vm_stat_account(mm, oldflags, -nrpages);
 346         vm_stat_account(mm, newflags, nrpages);
 347         perf_event_mmap(vma);
 348         return 0;
 349
 350 fail:
 351         vm_unacct_memory(charged);
 352         return error;
 353 }
 354
 355 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 356                 unsigned long, prot)
 357 {
 358         unsigned long nstart, end, tmp, reqprot;
 359         struct vm_area_struct *vma, *prev;
 360         int error = -EINVAL;
 361         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 362         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 363                                 (prot & PROT_READ);
 364
 365         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 366         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 367                 return -EINVAL;
 368
 369         if (start & ~PAGE_MASK)
 370                 return -EINVAL;
 371         if (!len)
 372                 return 0;
 373         len = PAGE_ALIGN(len);
 374         end = start + len;
 375         if (end <= start)
 376                 return -ENOMEM;
 377         if (!arch_validate_prot(prot))
 378                 return -EINVAL;
 379
 380         reqprot = prot;
 381
 382         down_write(&current->mm->mmap_sem);
 383
 384         vma = find_vma(current->mm, start);
 385         error = -ENOMEM;
 386         if (!vma)
 387                 goto out;
 388         prev = vma->vm_prev;
 389         if (unlikely(grows & PROT_GROWSDOWN)) {
 390                 if (vma->vm_start >= end)
 391                         goto out;
 392                 start = vma->vm_start;
 393                 error = -EINVAL;
 394                 if (!(vma->vm_flags & VM_GROWSDOWN))
 395                         goto out;
 396         } else {
 397                 if (vma->vm_start > start)
 398                         goto out;
 399                 if (unlikely(grows & PROT_GROWSUP)) {
 400                         end = vma->vm_end;
 401                         error = -EINVAL;
 402                         if (!(vma->vm_flags & VM_GROWSUP))
 403                                 goto out;
 404                 }
 405         }
 406         if (start > vma->vm_start)
 407                 prev = vma;
 408
 409         for (nstart = start ; ; ) {
 410                 unsigned long newflags;
 411                 int pkey = arch_override_mprotect_pkey(vma, prot, -1);
 412
 413                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 414
 415                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 416                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 417                         prot |= PROT_EXEC;
 418
 419                 newflags = calc_vm_prot_bits(prot, pkey);
 420                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 421
 422                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 423                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 424                         error = -EACCES;
 425                         goto out;
 426                 }
 427
 428                 error = security_file_mprotect(vma, reqprot, prot);
 429                 if (error)
 430                         goto out;
 431
 432                 tmp = vma->vm_end;
 433                 if (tmp > end)
 434                         tmp = end;
 435                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 436                 if (error)
 437                         goto out;
 438                 nstart = tmp;
 439
 440                 if (nstart < prev->vm_end)
 441                         nstart = prev->vm_end;
 442                 if (nstart >= end)
 443                         goto out;
 444
 445                 vma = prev->vm_next;
 446                 if (!vma || vma->vm_start != nstart) {
 447                         error = -ENOMEM;
 448                         goto out;
 449                 }
 450                 prot = reqprot;
 451         }
 452 out:
 453         up_write(&current->mm->mmap_sem);
 454         return error;
 455 }