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 /*
  33  * For a prot_numa update we only hold mmap_sem for read so there is a
  34  * potential race with faulting where a pmd was temporarily none. This
  35  * function checks for a transhuge pmd under the appropriate lock. It
  36  * returns a pte if it was successfully locked or NULL if it raced with
  37  * a transhuge insertion.
  38  */
  39 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
  40                         unsigned long addr, int prot_numa, spinlock_t **ptl)
  41 {
  42         pte_t *pte;
  43         spinlock_t *pmdl;
  44
  45         /* !prot_numa is protected by mmap_sem held for write */
  46         if (!prot_numa)
  47                 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  48
  49         pmdl = pmd_lock(vma->vm_mm, pmd);
  50         if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
  51                 spin_unlock(pmdl);
  52                 return NULL;
  53         }
  54
  55         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  56         spin_unlock(pmdl);
  57         return pte;
  58 }
  59
  60 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  61                 unsigned long addr, unsigned long end, pgprot_t newprot,
  62                 int dirty_accountable, int prot_numa)
  63 {
  64         struct mm_struct *mm = vma->vm_mm;
  65         pte_t *pte, oldpte;
  66         spinlock_t *ptl;
  67         unsigned long pages = 0;
  68
  69         pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
  70         if (!pte)
  71                 return 0;
  72
  73         arch_enter_lazy_mmu_mode();
  74         do {
  75                 oldpte = *pte;
  76                 if (pte_present(oldpte)) {
  77                         pte_t ptent;
  78                         bool updated = false;
  79
  80                         if (!prot_numa) {
  81                                 ptent = ptep_modify_prot_start(mm, addr, pte);
  82                                 if (pte_numa(ptent))
  83                                         ptent = pte_mknonnuma(ptent);
  84                                 ptent = pte_modify(ptent, newprot);
  85                                 /*
  86                                  * Avoid taking write faults for pages we
  87                                  * know to be dirty.
  88                                  */
  89                                 if (dirty_accountable && pte_dirty(ptent) &&
  90                                     (pte_soft_dirty(ptent) ||
  91                                      !(vma->vm_flags & VM_SOFTDIRTY)))
  92                                         ptent = pte_mkwrite(ptent);
  93                                 ptep_modify_prot_commit(mm, addr, pte, ptent);
  94                                 updated = true;
  95                         } else {
  96                                 struct page *page;
  97
  98                                 page = vm_normal_page(vma, addr, oldpte);
  99                                 if (page && !PageKsm(page)) {
 100                                         if (!pte_numa(oldpte)) {
 101                                                 ptep_set_numa(mm, addr, pte);
 102                                                 updated = true;
 103                                         }
 104                                 }
 105                         }
 106                         if (updated)
 107                                 pages++;
 108                 } else if (IS_ENABLED(CONFIG_MIGRATION) && !pte_file(oldpte)) {
 109                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 110
 111                         if (is_write_migration_entry(entry)) {
 112                                 pte_t newpte;
 113                                 /*
 114                                  * A protection check is difficult so
 115                                  * just be safe and disable write
 116                                  */
 117                                 make_migration_entry_read(&entry);
 118                                 newpte = swp_entry_to_pte(entry);
 119                                 if (pte_swp_soft_dirty(oldpte))
 120                                         newpte = pte_swp_mksoft_dirty(newpte);
 121                                 set_pte_at(mm, addr, pte, newpte);
 122
 123                                 pages++;
 124                         }
 125                 }
 126         } while (pte++, addr += PAGE_SIZE, addr != end);
 127         arch_leave_lazy_mmu_mode();
 128         pte_unmap_unlock(pte - 1, ptl);
 129
 130         return pages;
 131 }
 132
 133 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 134                 pud_t *pud, unsigned long addr, unsigned long end,
 135                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 136 {
 137         pmd_t *pmd;
 138         struct mm_struct *mm = vma->vm_mm;
 139         unsigned long next;
 140         unsigned long pages = 0;
 141         unsigned long nr_huge_updates = 0;
 142         unsigned long mni_start = 0;
 143
 144         pmd = pmd_offset(pud, addr);
 145         do {
 146                 unsigned long this_pages;
 147
 148                 next = pmd_addr_end(addr, end);
 149                 if (!pmd_trans_huge(*pmd) && pmd_none_or_clear_bad(pmd))
 150                         continue;
 151
 152                 /* invoke the mmu notifier if the pmd is populated */
 153                 if (!mni_start) {
 154                         mni_start = addr;
 155                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 156                 }
 157
 158                 if (pmd_trans_huge(*pmd)) {
 159                         if (next - addr != HPAGE_PMD_SIZE)
 160                                 split_huge_page_pmd(vma, addr, pmd);
 161                         else {
 162                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 163                                                 newprot, prot_numa);
 164
 165                                 if (nr_ptes) {
 166                                         if (nr_ptes == HPAGE_PMD_NR) {
 167                                                 pages += HPAGE_PMD_NR;
 168                                                 nr_huge_updates++;
 169                                         }
 170
 171                                         /* huge pmd was handled */
 172                                         continue;
 173                                 }
 174                         }
 175                         /* fall through, the trans huge pmd just split */
 176                 }
 177                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 178                                  dirty_accountable, prot_numa);
 179                 pages += this_pages;
 180         } while (pmd++, addr = next, addr != end);
 181
 182         if (mni_start)
 183                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 184
 185         if (nr_huge_updates)
 186                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 187         return pages;
 188 }
 189
 190 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 191                 pgd_t *pgd, unsigned long addr, unsigned long end,
 192                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 193 {
 194         pud_t *pud;
 195         unsigned long next;
 196         unsigned long pages = 0;
 197
 198         pud = pud_offset(pgd, addr);
 199         do {
 200                 next = pud_addr_end(addr, end);
 201                 if (pud_none_or_clear_bad(pud))
 202                         continue;
 203                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 204                                  dirty_accountable, prot_numa);
 205         } while (pud++, addr = next, addr != end);
 206
 207         return pages;
 208 }
 209
 210 static unsigned long change_protection_range(struct vm_area_struct *vma,
 211                 unsigned long addr, unsigned long end, pgprot_t newprot,
 212                 int dirty_accountable, int prot_numa)
 213 {
 214         struct mm_struct *mm = vma->vm_mm;
 215         pgd_t *pgd;
 216         unsigned long next;
 217         unsigned long start = addr;
 218         unsigned long pages = 0;
 219
 220         BUG_ON(addr >= end);
 221         pgd = pgd_offset(mm, addr);
 222         flush_cache_range(vma, addr, end);
 223         set_tlb_flush_pending(mm);
 224         do {
 225                 next = pgd_addr_end(addr, end);
 226                 if (pgd_none_or_clear_bad(pgd))
 227                         continue;
 228                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 229                                  dirty_accountable, prot_numa);
 230         } while (pgd++, addr = next, addr != end);
 231
 232         /* Only flush the TLB if we actually modified any entries: */
 233         if (pages)
 234                 flush_tlb_range(vma, start, end);
 235         clear_tlb_flush_pending(mm);
 236
 237         return pages;
 238 }
 239
 240 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 241                        unsigned long end, pgprot_t newprot,
 242                        int dirty_accountable, int prot_numa)
 243 {
 244         unsigned long pages;
 245
 246         if (is_vm_hugetlb_page(vma))
 247                 pages = hugetlb_change_protection(vma, start, end, newprot);
 248         else
 249                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 250
 251         return pages;
 252 }
 253
 254 int
 255 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 256         unsigned long start, unsigned long end, unsigned long newflags)
 257 {
 258         struct mm_struct *mm = vma->vm_mm;
 259         unsigned long oldflags = vma->vm_flags;
 260         long nrpages = (end - start) >> PAGE_SHIFT;
 261         unsigned long charged = 0;
 262         pgoff_t pgoff;
 263         int error;
 264         int dirty_accountable = 0;
 265
 266         if (newflags == oldflags) {
 267                 *pprev = vma;
 268                 return 0;
 269         }
 270
 271         /*
 272          * If we make a private mapping writable we increase our commit;
 273          * but (without finer accounting) cannot reduce our commit if we
 274          * make it unwritable again. hugetlb mapping were accounted for
 275          * even if read-only so there is no need to account for them here
 276          */
 277         if (newflags & VM_WRITE) {
 278                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 279                                                 VM_SHARED|VM_NORESERVE))) {
 280                         charged = nrpages;
 281                         if (security_vm_enough_memory_mm(mm, charged))
 282                                 return -ENOMEM;
 283                         newflags |= VM_ACCOUNT;
 284                 }
 285         }
 286
 287         /*
 288          * First try to merge with previous and/or next vma.
 289          */
 290         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 291         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 292                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
 293         if (*pprev) {
 294                 vma = *pprev;
 295                 goto success;
 296         }
 297
 298         *pprev = vma;
 299
 300         if (start != vma->vm_start) {
 301                 error = split_vma(mm, vma, start, 1);
 302                 if (error)
 303                         goto fail;
 304         }
 305
 306         if (end != vma->vm_end) {
 307                 error = split_vma(mm, vma, end, 0);
 308                 if (error)
 309                         goto fail;
 310         }
 311
 312 success:
 313         /*
 314          * vm_flags and vm_page_prot are protected by the mmap_sem
 315          * held in write mode.
 316          */
 317         vma->vm_flags = newflags;
 318         dirty_accountable = vma_wants_writenotify(vma);
 319         vma_set_page_prot(vma);
 320
 321         change_protection(vma, start, end, vma->vm_page_prot,
 322                           dirty_accountable, 0);
 323
 324         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
 325         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
 326         perf_event_mmap(vma);
 327         return 0;
 328
 329 fail:
 330         vm_unacct_memory(charged);
 331         return error;
 332 }
 333
 334 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 335                 unsigned long, prot)
 336 {
 337         unsigned long vm_flags, nstart, end, tmp, reqprot;
 338         struct vm_area_struct *vma, *prev;
 339         int error = -EINVAL;
 340         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 341         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 342         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 343                 return -EINVAL;
 344
 345         if (start & ~PAGE_MASK)
 346                 return -EINVAL;
 347         if (!len)
 348                 return 0;
 349         len = PAGE_ALIGN(len);
 350         end = start + len;
 351         if (end <= start)
 352                 return -ENOMEM;
 353         if (!arch_validate_prot(prot))
 354                 return -EINVAL;
 355
 356         reqprot = prot;
 357         /*
 358          * Does the application expect PROT_READ to imply PROT_EXEC:
 359          */
 360         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 361                 prot |= PROT_EXEC;
 362
 363         vm_flags = calc_vm_prot_bits(prot);
 364
 365         down_write(&current->mm->mmap_sem);
 366
 367         vma = find_vma(current->mm, start);
 368         error = -ENOMEM;
 369         if (!vma)
 370                 goto out;
 371         prev = vma->vm_prev;
 372         if (unlikely(grows & PROT_GROWSDOWN)) {
 373                 if (vma->vm_start >= end)
 374                         goto out;
 375                 start = vma->vm_start;
 376                 error = -EINVAL;
 377                 if (!(vma->vm_flags & VM_GROWSDOWN))
 378                         goto out;
 379         } else {
 380                 if (vma->vm_start > start)
 381                         goto out;
 382                 if (unlikely(grows & PROT_GROWSUP)) {
 383                         end = vma->vm_end;
 384                         error = -EINVAL;
 385                         if (!(vma->vm_flags & VM_GROWSUP))
 386                                 goto out;
 387                 }
 388         }
 389         if (start > vma->vm_start)
 390                 prev = vma;
 391
 392         for (nstart = start ; ; ) {
 393                 unsigned long newflags;
 394
 395                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 396
 397                 newflags = vm_flags;
 398                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 399
 400                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 401                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 402                         error = -EACCES;
 403                         goto out;
 404                 }
 405
 406                 error = security_file_mprotect(vma, reqprot, prot);
 407                 if (error)
 408                         goto out;
 409
 410                 tmp = vma->vm_end;
 411                 if (tmp > end)
 412                         tmp = end;
 413                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 414                 if (error)
 415                         goto out;
 416                 nstart = tmp;
 417
 418                 if (nstart < prev->vm_end)
 419                         nstart = prev->vm_end;
 420                 if (nstart >= end)
 421                         goto out;
 422
 423                 vma = prev->vm_next;
 424                 if (!vma || vma->vm_start != nstart) {
 425                         error = -ENOMEM;
 426                         goto out;
 427                 }
 428         }
 429 out:
 430         up_write(&current->mm->mmap_sem);
 431         return error;
 432 }