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 <asm/uaccess.h>
  27 #include <asm/pgtable.h>
  28 #include <asm/cacheflush.h>
  29 #include <asm/tlbflush.h>
  30
  31 #ifndef pgprot_modify
  32 static inline pgprot_t pgprot_modify(pgprot_t oldprot, pgprot_t newprot)
  33 {
  34         return newprot;
  35 }
  36 #endif
  37
  38 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  39                 unsigned long addr, unsigned long end, pgprot_t newprot,
  40                 int dirty_accountable, int prot_numa)
  41 {
  42         struct mm_struct *mm = vma->vm_mm;
  43         pte_t *pte, oldpte;
  44         spinlock_t *ptl;
  45         unsigned long pages = 0;
  46
  47         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
  48         arch_enter_lazy_mmu_mode();
  49         do {
  50                 oldpte = *pte;
  51                 if (pte_present(oldpte)) {
  52                         pte_t ptent;
  53                         bool updated = false;
  54
  55                         if (!prot_numa) {
  56                                 ptent = ptep_modify_prot_start(mm, addr, pte);
  57                                 ptent = pte_modify(ptent, newprot);
  58                                 updated = true;
  59                         } else {
  60                                 struct page *page;
  61
  62                                 ptent = *pte;
  63                                 page = vm_normal_page(vma, addr, oldpte);
  64                                 if (page) {
  65                                         if (!pte_numa(oldpte)) {
  66                                                 ptent = pte_mknuma(ptent);
  67                                                 set_pte_at(mm, addr, pte, ptent);
  68                                                 updated = true;
  69                                         }
  70                                 }
  71                         }
  72
  73                         /*
  74                          * Avoid taking write faults for pages we know to be
  75                          * dirty.
  76                          */
  77                         if (dirty_accountable && pte_dirty(ptent)) {
  78                                 ptent = pte_mkwrite(ptent);
  79                                 updated = true;
  80                         }
  81
  82                         if (updated)
  83                                 pages++;
  84
  85                         /* Only !prot_numa always clears the pte */
  86                         if (!prot_numa)
  87                                 ptep_modify_prot_commit(mm, addr, pte, ptent);
  88                 } else if (IS_ENABLED(CONFIG_MIGRATION) && !pte_file(oldpte)) {
  89                         swp_entry_t entry = pte_to_swp_entry(oldpte);
  90
  91                         if (is_write_migration_entry(entry)) {
  92                                 pte_t newpte;
  93                                 /*
  94                                  * A protection check is difficult so
  95                                  * just be safe and disable write
  96                                  */
  97                                 make_migration_entry_read(&entry);
  98                                 newpte = swp_entry_to_pte(entry);
  99                                 if (pte_swp_soft_dirty(oldpte))
 100                                         newpte = pte_swp_mksoft_dirty(newpte);
 101                                 set_pte_at(mm, addr, pte, newpte);
 102
 103                                 pages++;
 104                         }
 105                 }
 106         } while (pte++, addr += PAGE_SIZE, addr != end);
 107         arch_leave_lazy_mmu_mode();
 108         pte_unmap_unlock(pte - 1, ptl);
 109
 110         return pages;
 111 }
 112
 113 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 114                 pud_t *pud, unsigned long addr, unsigned long end,
 115                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 116 {
 117         pmd_t *pmd;
 118         unsigned long next;
 119         unsigned long pages = 0;
 120         unsigned long nr_huge_updates = 0;
 121
 122         pmd = pmd_offset(pud, addr);
 123         do {
 124                 unsigned long this_pages;
 125
 126                 next = pmd_addr_end(addr, end);
 127                 if (pmd_trans_huge(*pmd)) {
 128                         if (next - addr != HPAGE_PMD_SIZE)
 129                                 split_huge_page_pmd(vma, addr, pmd);
 130                         else {
 131                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 132                                                 newprot, prot_numa);
 133
 134                                 if (nr_ptes) {
 135                                         if (nr_ptes == HPAGE_PMD_NR) {
 136                                                 pages += HPAGE_PMD_NR;
 137                                                 nr_huge_updates++;
 138                                         }
 139                                         continue;
 140                                 }
 141                         }
 142                         /* fall through */
 143                 }
 144                 if (pmd_none_or_clear_bad(pmd))
 145                         continue;
 146                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 147                                  dirty_accountable, prot_numa);
 148                 pages += this_pages;
 149         } while (pmd++, addr = next, addr != end);
 150
 151         if (nr_huge_updates)
 152                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 153         return pages;
 154 }
 155
 156 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 157                 pgd_t *pgd, unsigned long addr, unsigned long end,
 158                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 159 {
 160         pud_t *pud;
 161         unsigned long next;
 162         unsigned long pages = 0;
 163
 164         pud = pud_offset(pgd, addr);
 165         do {
 166                 next = pud_addr_end(addr, end);
 167                 if (pud_none_or_clear_bad(pud))
 168                         continue;
 169                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 170                                  dirty_accountable, prot_numa);
 171         } while (pud++, addr = next, addr != end);
 172
 173         return pages;
 174 }
 175
 176 static unsigned long change_protection_range(struct vm_area_struct *vma,
 177                 unsigned long addr, unsigned long end, pgprot_t newprot,
 178                 int dirty_accountable, int prot_numa)
 179 {
 180         struct mm_struct *mm = vma->vm_mm;
 181         pgd_t *pgd;
 182         unsigned long next;
 183         unsigned long start = addr;
 184         unsigned long pages = 0;
 185
 186         BUG_ON(addr >= end);
 187         pgd = pgd_offset(mm, addr);
 188         flush_cache_range(vma, addr, end);
 189         do {
 190                 next = pgd_addr_end(addr, end);
 191                 if (pgd_none_or_clear_bad(pgd))
 192                         continue;
 193                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 194                                  dirty_accountable, prot_numa);
 195         } while (pgd++, addr = next, addr != end);
 196
 197         /* Only flush the TLB if we actually modified any entries: */
 198         if (pages)
 199                 flush_tlb_range(vma, start, end);
 200
 201         return pages;
 202 }
 203
 204 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 205                        unsigned long end, pgprot_t newprot,
 206                        int dirty_accountable, int prot_numa)
 207 {
 208         struct mm_struct *mm = vma->vm_mm;
 209         unsigned long pages;
 210
 211         mmu_notifier_invalidate_range_start(mm, start, end);
 212         if (is_vm_hugetlb_page(vma))
 213                 pages = hugetlb_change_protection(vma, start, end, newprot);
 214         else
 215                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 216         mmu_notifier_invalidate_range_end(mm, start, end);
 217
 218         return pages;
 219 }
 220
 221 int
 222 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 223         unsigned long start, unsigned long end, unsigned long newflags)
 224 {
 225         struct mm_struct *mm = vma->vm_mm;
 226         unsigned long oldflags = vma->vm_flags;
 227         long nrpages = (end - start) >> PAGE_SHIFT;
 228         unsigned long charged = 0;
 229         pgoff_t pgoff;
 230         int error;
 231         int dirty_accountable = 0;
 232
 233         if (newflags == oldflags) {
 234                 *pprev = vma;
 235                 return 0;
 236         }
 237
 238         /*
 239          * If we make a private mapping writable we increase our commit;
 240          * but (without finer accounting) cannot reduce our commit if we
 241          * make it unwritable again. hugetlb mapping were accounted for
 242          * even if read-only so there is no need to account for them here
 243          */
 244         if (newflags & VM_WRITE) {
 245                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 246                                                 VM_SHARED|VM_NORESERVE))) {
 247                         charged = nrpages;
 248                         if (security_vm_enough_memory_mm(mm, charged))
 249                                 return -ENOMEM;
 250                         newflags |= VM_ACCOUNT;
 251                 }
 252         }
 253
 254         /*
 255          * First try to merge with previous and/or next vma.
 256          */
 257         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 258         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 259                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
 260         if (*pprev) {
 261                 vma = *pprev;
 262                 goto success;
 263         }
 264
 265         *pprev = vma;
 266
 267         if (start != vma->vm_start) {
 268                 error = split_vma(mm, vma, start, 1);
 269                 if (error)
 270                         goto fail;
 271         }
 272
 273         if (end != vma->vm_end) {
 274                 error = split_vma(mm, vma, end, 0);
 275                 if (error)
 276                         goto fail;
 277         }
 278
 279 success:
 280         /*
 281          * vm_flags and vm_page_prot are protected by the mmap_sem
 282          * held in write mode.
 283          */
 284         vma->vm_flags = newflags;
 285         vma->vm_page_prot = pgprot_modify(vma->vm_page_prot,
 286                                           vm_get_page_prot(newflags));
 287
 288         if (vma_wants_writenotify(vma)) {
 289                 vma->vm_page_prot = vm_get_page_prot(newflags & ~VM_SHARED);
 290                 dirty_accountable = 1;
 291         }
 292
 293         change_protection(vma, start, end, vma->vm_page_prot,
 294                           dirty_accountable, 0);
 295
 296         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
 297         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
 298         perf_event_mmap(vma);
 299         return 0;
 300
 301 fail:
 302         vm_unacct_memory(charged);
 303         return error;
 304 }
 305
 306 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 307                 unsigned long, prot)
 308 {
 309         unsigned long vm_flags, nstart, end, tmp, reqprot;
 310         struct vm_area_struct *vma, *prev;
 311         int error = -EINVAL;
 312         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 313         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 314         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 315                 return -EINVAL;
 316
 317         if (start & ~PAGE_MASK)
 318                 return -EINVAL;
 319         if (!len)
 320                 return 0;
 321         len = PAGE_ALIGN(len);
 322         end = start + len;
 323         if (end <= start)
 324                 return -ENOMEM;
 325         if (!arch_validate_prot(prot))
 326                 return -EINVAL;
 327
 328         reqprot = prot;
 329         /*
 330          * Does the application expect PROT_READ to imply PROT_EXEC:
 331          */
 332         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 333                 prot |= PROT_EXEC;
 334
 335         vm_flags = calc_vm_prot_bits(prot);
 336
 337         down_write(&current->mm->mmap_sem);
 338
 339         vma = find_vma(current->mm, start);
 340         error = -ENOMEM;
 341         if (!vma)
 342                 goto out;
 343         prev = vma->vm_prev;
 344         if (unlikely(grows & PROT_GROWSDOWN)) {
 345                 if (vma->vm_start >= end)
 346                         goto out;
 347                 start = vma->vm_start;
 348                 error = -EINVAL;
 349                 if (!(vma->vm_flags & VM_GROWSDOWN))
 350                         goto out;
 351         } else {
 352                 if (vma->vm_start > start)
 353                         goto out;
 354                 if (unlikely(grows & PROT_GROWSUP)) {
 355                         end = vma->vm_end;
 356                         error = -EINVAL;
 357                         if (!(vma->vm_flags & VM_GROWSUP))
 358                                 goto out;
 359                 }
 360         }
 361         if (start > vma->vm_start)
 362                 prev = vma;
 363
 364         for (nstart = start ; ; ) {
 365                 unsigned long newflags;
 366
 367                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 368
 369                 newflags = vm_flags;
 370                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 371
 372                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 373                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 374                         error = -EACCES;
 375                         goto out;
 376                 }
 377
 378                 error = security_file_mprotect(vma, reqprot, prot);
 379                 if (error)
 380                         goto out;
 381
 382                 tmp = vma->vm_end;
 383                 if (tmp > end)
 384                         tmp = end;
 385                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 386                 if (error)
 387                         goto out;
 388                 nstart = tmp;
 389
 390                 if (nstart < prev->vm_end)
 391                         nstart = prev->vm_end;
 392                 if (nstart >= end)
 393                         goto out;
 394
 395                 vma = prev->vm_next;
 396                 if (!vma || vma->vm_start != nstart) {
 397                         error = -ENOMEM;
 398                         goto out;
 399                 }
 400         }
 401 out:
 402         up_write(&current->mm->mmap_sem);
 403         return error;
 404 }