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