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 <asm/uaccess.h>
  28 #include <asm/pgtable.h>
  29 #include <asm/cacheflush.h>
  30 #include <asm/tlbflush.h>
  31
  32 #ifndef pgprot_modify
  33 static inline pgprot_t pgprot_modify(pgprot_t oldprot, pgprot_t newprot)
  34 {
  35         return newprot;
  36 }
  37 #endif
  38
  39 /*
  40  * For a prot_numa update we only hold mmap_sem for read so there is a
  41  * potential race with faulting where a pmd was temporarily none. This
  42  * function checks for a transhuge pmd under the appropriate lock. It
  43  * returns a pte if it was successfully locked or NULL if it raced with
  44  * a transhuge insertion.
  45  */
  46 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
  47                         unsigned long addr, int prot_numa, spinlock_t **ptl)
  48 {
  49         pte_t *pte;
  50         spinlock_t *pmdl;
  51
  52         /* !prot_numa is protected by mmap_sem held for write */
  53         if (!prot_numa)
  54                 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  55
  56         pmdl = pmd_lock(vma->vm_mm, pmd);
  57         if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
  58                 spin_unlock(pmdl);
  59                 return NULL;
  60         }
  61
  62         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  63         spin_unlock(pmdl);
  64         return pte;
  65 }
  66
  67 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  68                 unsigned long addr, unsigned long end, pgprot_t newprot,
  69                 int dirty_accountable, int prot_numa)
  70 {
  71         struct mm_struct *mm = vma->vm_mm;
  72         pte_t *pte, oldpte;
  73         spinlock_t *ptl;
  74         unsigned long pages = 0;
  75
  76         pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
  77         if (!pte)
  78                 return 0;
  79
  80         arch_enter_lazy_mmu_mode();
  81         do {
  82                 oldpte = *pte;
  83                 if (pte_present(oldpte)) {
  84                         pte_t ptent;
  85                         bool updated = false;
  86
  87                         if (!prot_numa) {
  88                                 ptent = ptep_modify_prot_start(mm, addr, pte);
  89                                 if (pte_numa(ptent))
  90                                         ptent = pte_mknonnuma(ptent);
  91                                 ptent = pte_modify(ptent, newprot);
  92                                 /*
  93                                  * Avoid taking write faults for pages we
  94                                  * know to be dirty.
  95                                  */
  96                                 if (dirty_accountable && pte_dirty(ptent))
  97                                         ptent = pte_mkwrite(ptent);
  98                                 ptep_modify_prot_commit(mm, addr, pte, ptent);
  99                                 updated = true;
 100                         } else {
 101                                 struct page *page;
 102
 103                                 page = vm_normal_page(vma, addr, oldpte);
 104                                 if (page && !PageKsm(page)) {
 105                                         if (!pte_numa(oldpte)) {
 106                                                 ptep_set_numa(mm, addr, pte);
 107                                                 updated = true;
 108                                         }
 109                                 }
 110                         }
 111                         if (updated)
 112                                 pages++;
 113                 } else if (IS_ENABLED(CONFIG_MIGRATION) && !pte_file(oldpte)) {
 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         unsigned long next;
 144         unsigned long pages = 0;
 145         unsigned long nr_huge_updates = 0;
 146
 147         pmd = pmd_offset(pud, addr);
 148         do {
 149                 unsigned long this_pages;
 150
 151                 next = pmd_addr_end(addr, end);
 152                 if (!pmd_trans_huge(*pmd) && pmd_none_or_clear_bad(pmd))
 153                         continue;
 154                 if (pmd_trans_huge(*pmd)) {
 155                         if (next - addr != HPAGE_PMD_SIZE)
 156                                 split_huge_page_pmd(vma, addr, pmd);
 157                         else {
 158                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 159                                                 newprot, prot_numa);
 160
 161                                 if (nr_ptes) {
 162                                         if (nr_ptes == HPAGE_PMD_NR) {
 163                                                 pages += HPAGE_PMD_NR;
 164                                                 nr_huge_updates++;
 165                                         }
 166
 167                                         /* huge pmd was handled */
 168                                         continue;
 169                                 }
 170                         }
 171                         /* fall through, the trans huge pmd just split */
 172                 }
 173                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 174                                  dirty_accountable, prot_numa);
 175                 pages += this_pages;
 176         } while (pmd++, addr = next, addr != end);
 177
 178         if (nr_huge_updates)
 179                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 180         return pages;
 181 }
 182
 183 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 184                 pgd_t *pgd, unsigned long addr, unsigned long end,
 185                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 186 {
 187         pud_t *pud;
 188         unsigned long next;
 189         unsigned long pages = 0;
 190
 191         pud = pud_offset(pgd, addr);
 192         do {
 193                 next = pud_addr_end(addr, end);
 194                 if (pud_none_or_clear_bad(pud))
 195                         continue;
 196                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 197                                  dirty_accountable, prot_numa);
 198         } while (pud++, addr = next, addr != end);
 199
 200         return pages;
 201 }
 202
 203 static unsigned long change_protection_range(struct vm_area_struct *vma,
 204                 unsigned long addr, unsigned long end, pgprot_t newprot,
 205                 int dirty_accountable, int prot_numa)
 206 {
 207         struct mm_struct *mm = vma->vm_mm;
 208         pgd_t *pgd;
 209         unsigned long next;
 210         unsigned long start = addr;
 211         unsigned long pages = 0;
 212
 213         BUG_ON(addr >= end);
 214         pgd = pgd_offset(mm, addr);
 215         flush_cache_range(vma, addr, end);
 216         set_tlb_flush_pending(mm);
 217         do {
 218                 next = pgd_addr_end(addr, end);
 219                 if (pgd_none_or_clear_bad(pgd))
 220                         continue;
 221                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 222                                  dirty_accountable, prot_numa);
 223         } while (pgd++, addr = next, addr != end);
 224
 225         /* Only flush the TLB if we actually modified any entries: */
 226         if (pages)
 227                 flush_tlb_range(vma, start, end);
 228         clear_tlb_flush_pending(mm);
 229
 230         return pages;
 231 }
 232
 233 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 234                        unsigned long end, pgprot_t newprot,
 235                        int dirty_accountable, int prot_numa)
 236 {
 237         struct mm_struct *mm = vma->vm_mm;
 238         unsigned long pages;
 239
 240         mmu_notifier_invalidate_range_start(mm, start, end);
 241         if (is_vm_hugetlb_page(vma))
 242                 pages = hugetlb_change_protection(vma, start, end, newprot);
 243         else
 244                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 245         mmu_notifier_invalidate_range_end(mm, start, end);
 246
 247         return pages;
 248 }
 249
 250 int
 251 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 252         unsigned long start, unsigned long end, unsigned long newflags)
 253 {
 254         struct mm_struct *mm = vma->vm_mm;
 255         unsigned long oldflags = vma->vm_flags;
 256         long nrpages = (end - start) >> PAGE_SHIFT;
 257         unsigned long charged = 0;
 258         pgoff_t pgoff;
 259         int error;
 260         int dirty_accountable = 0;
 261
 262         if (newflags == oldflags) {
 263                 *pprev = vma;
 264                 return 0;
 265         }
 266
 267         /*
 268          * If we make a private mapping writable we increase our commit;
 269          * but (without finer accounting) cannot reduce our commit if we
 270          * make it unwritable again. hugetlb mapping were accounted for
 271          * even if read-only so there is no need to account for them here
 272          */
 273         if (newflags & VM_WRITE) {
 274                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 275                                                 VM_SHARED|VM_NORESERVE))) {
 276                         charged = nrpages;
 277                         if (security_vm_enough_memory_mm(mm, charged))
 278                                 return -ENOMEM;
 279                         newflags |= VM_ACCOUNT;
 280                 }
 281         }
 282
 283         /*
 284          * First try to merge with previous and/or next vma.
 285          */
 286         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 287         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 288                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
 289         if (*pprev) {
 290                 vma = *pprev;
 291                 goto success;
 292         }
 293
 294         *pprev = vma;
 295
 296         if (start != vma->vm_start) {
 297                 error = split_vma(mm, vma, start, 1);
 298                 if (error)
 299                         goto fail;
 300         }
 301
 302         if (end != vma->vm_end) {
 303                 error = split_vma(mm, vma, end, 0);
 304                 if (error)
 305                         goto fail;
 306         }
 307
 308 success:
 309         /*
 310          * vm_flags and vm_page_prot are protected by the mmap_sem
 311          * held in write mode.
 312          */
 313         vma->vm_flags = newflags;
 314         vma->vm_page_prot = pgprot_modify(vma->vm_page_prot,
 315                                           vm_get_page_prot(newflags));
 316
 317         if (vma_wants_writenotify(vma)) {
 318                 vma->vm_page_prot = vm_get_page_prot(newflags & ~VM_SHARED);
 319                 dirty_accountable = 1;
 320         }
 321
 322         change_protection(vma, start, end, vma->vm_page_prot,
 323                           dirty_accountable, 0);
 324
 325         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
 326         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
 327         perf_event_mmap(vma);
 328         return 0;
 329
 330 fail:
 331         vm_unacct_memory(charged);
 332         return error;
 333 }
 334
 335 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 336                 unsigned long, prot)
 337 {
 338         unsigned long vm_flags, nstart, end, tmp, reqprot;
 339         struct vm_area_struct *vma, *prev;
 340         int error = -EINVAL;
 341         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 342         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 343         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 344                 return -EINVAL;
 345
 346         if (start & ~PAGE_MASK)
 347                 return -EINVAL;
 348         if (!len)
 349                 return 0;
 350         len = PAGE_ALIGN(len);
 351         end = start + len;
 352         if (end <= start)
 353                 return -ENOMEM;
 354         if (!arch_validate_prot(prot))
 355                 return -EINVAL;
 356
 357         reqprot = prot;
 358         /*
 359          * Does the application expect PROT_READ to imply PROT_EXEC:
 360          */
 361         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 362                 prot |= PROT_EXEC;
 363
 364         vm_flags = calc_vm_prot_bits(prot);
 365
 366         down_write(&current->mm->mmap_sem);
 367
 368         vma = find_vma(current->mm, start);
 369         error = -ENOMEM;
 370         if (!vma)
 371                 goto out;
 372         prev = vma->vm_prev;
 373         if (unlikely(grows & PROT_GROWSDOWN)) {
 374                 if (vma->vm_start >= end)
 375                         goto out;
 376                 start = vma->vm_start;
 377                 error = -EINVAL;
 378                 if (!(vma->vm_flags & VM_GROWSDOWN))
 379                         goto out;
 380         } else {
 381                 if (vma->vm_start > start)
 382                         goto out;
 383                 if (unlikely(grows & PROT_GROWSUP)) {
 384                         end = vma->vm_end;
 385                         error = -EINVAL;
 386                         if (!(vma->vm_flags & VM_GROWSUP))
 387                                 goto out;
 388                 }
 389         }
 390         if (start > vma->vm_start)
 391                 prev = vma;
 392
 393         for (nstart = start ; ; ) {
 394                 unsigned long newflags;
 395
 396                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 397
 398                 newflags = vm_flags;
 399                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 400
 401                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 402                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 403                         error = -EACCES;
 404                         goto out;
 405                 }
 406
 407                 error = security_file_mprotect(vma, reqprot, prot);
 408                 if (error)
 409                         goto out;
 410
 411                 tmp = vma->vm_end;
 412                 if (tmp > end)
 413                         tmp = end;
 414                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 415                 if (error)
 416                         goto out;
 417                 nstart = tmp;
 418
 419                 if (nstart < prev->vm_end)
 420                         nstart = prev->vm_end;
 421                 if (nstart >= end)
 422                         goto out;
 423
 424                 vma = prev->vm_next;
 425                 if (!vma || vma->vm_start != nstart) {
 426                         error = -ENOMEM;
 427                         goto out;
 428                 }
 429         }
 430 out:
 431         up_write(&current->mm->mmap_sem);
 432         return error;
 433 }