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, bool *ret_all_same_node)
  41 {
  42         struct mm_struct *mm = vma->vm_mm;
  43         pte_t *pte, oldpte;
  44         spinlock_t *ptl;
  45         unsigned long pages = 0;
  46         bool all_same_node = true;
  47         int last_nid = -1;
  48
  49         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
  50         arch_enter_lazy_mmu_mode();
  51         do {
  52                 oldpte = *pte;
  53                 if (pte_present(oldpte)) {
  54                         pte_t ptent;
  55                         bool updated = false;
  56
  57                         if (!prot_numa) {
  58                                 ptent = ptep_modify_prot_start(mm, addr, pte);
  59                                 if (pte_numa(ptent))
  60                                         ptent = pte_mknonnuma(ptent);
  61                                 ptent = pte_modify(ptent, newprot);
  62                                 updated = true;
  63                         } else {
  64                                 struct page *page;
  65
  66                                 ptent = *pte;
  67                                 page = vm_normal_page(vma, addr, oldpte);
  68                                 if (page) {
  69                                         int this_nid = page_to_nid(page);
  70                                         if (last_nid == -1)
  71                                                 last_nid = this_nid;
  72                                         if (last_nid != this_nid)
  73                                                 all_same_node = false;
  74
  75                                         /* only check non-shared pages */
  76                                         if (!pte_numa(oldpte) &&
  77                                             page_mapcount(page) == 1) {
  78                                                 ptent = pte_mknuma(ptent);
  79                                                 set_pte_at(mm, addr, pte, ptent);
  80                                                 updated = true;
  81                                         }
  82                                 }
  83                         }
  84
  85                         /*
  86                          * Avoid taking write faults for pages we know to be
  87                          * dirty.
  88                          */
  89                         if (dirty_accountable && pte_dirty(ptent)) {
  90                                 ptent = pte_mkwrite(ptent);
  91                                 updated = true;
  92                         }
  93
  94                         if (updated)
  95                                 pages++;
  96
  97                         /* Only !prot_numa always clears the pte */
  98                         if (!prot_numa)
  99                                 ptep_modify_prot_commit(mm, addr, pte, ptent);
 100                 } else if (IS_ENABLED(CONFIG_MIGRATION) && !pte_file(oldpte)) {
 101                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 102
 103                         if (is_write_migration_entry(entry)) {
 104                                 pte_t newpte;
 105                                 /*
 106                                  * A protection check is difficult so
 107                                  * just be safe and disable write
 108                                  */
 109                                 make_migration_entry_read(&entry);
 110                                 newpte = swp_entry_to_pte(entry);
 111                                 if (pte_swp_soft_dirty(oldpte))
 112                                         newpte = pte_swp_mksoft_dirty(newpte);
 113                                 set_pte_at(mm, addr, pte, newpte);
 114                         }
 115                         pages++;
 116                 }
 117         } while (pte++, addr += PAGE_SIZE, addr != end);
 118         arch_leave_lazy_mmu_mode();
 119         pte_unmap_unlock(pte - 1, ptl);
 120
 121         *ret_all_same_node = all_same_node;
 122         return pages;
 123 }
 124
 125 #ifdef CONFIG_NUMA_BALANCING
 126 static inline void change_pmd_protnuma(struct mm_struct *mm, unsigned long addr,
 127                                        pmd_t *pmd)
 128 {
 129         spin_lock(&mm->page_table_lock);
 130         set_pmd_at(mm, addr & PMD_MASK, pmd, pmd_mknuma(*pmd));
 131         spin_unlock(&mm->page_table_lock);
 132 }
 133 #else
 134 static inline void change_pmd_protnuma(struct mm_struct *mm, unsigned long addr,
 135                                        pmd_t *pmd)
 136 {
 137         BUG();
 138 }
 139 #endif /* CONFIG_NUMA_BALANCING */
 140
 141 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 142                 pud_t *pud, unsigned long addr, unsigned long end,
 143                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 144 {
 145         pmd_t *pmd;
 146         unsigned long next;
 147         unsigned long pages = 0;
 148         unsigned long nr_huge_updates = 0;
 149         bool all_same_node;
 150
 151         pmd = pmd_offset(pud, addr);
 152         do {
 153                 next = pmd_addr_end(addr, end);
 154                 if (pmd_trans_huge(*pmd)) {
 155                         if (next - addr != HPAGE_PMD_SIZE)
 156                                 split_huge_page_pmd(vma, addr, pmd);
 157                         else if (change_huge_pmd(vma, pmd, addr, newprot,
 158                                                  prot_numa)) {
 159                                 pages += HPAGE_PMD_NR;
 160                                 nr_huge_updates++;
 161                                 continue;
 162                         }
 163                         /* fall through */
 164                 }
 165                 if (pmd_none_or_clear_bad(pmd))
 166                         continue;
 167                 pages += change_pte_range(vma, pmd, addr, next, newprot,
 168                                  dirty_accountable, prot_numa, &all_same_node);
 169
 170                 /*
 171                  * If we are changing protections for NUMA hinting faults then
 172                  * set pmd_numa if the examined pages were all on the same
 173                  * node. This allows a regular PMD to be handled as one fault
 174                  * and effectively batches the taking of the PTL
 175                  */
 176                 if (prot_numa && all_same_node)
 177                         change_pmd_protnuma(vma->vm_mm, addr, pmd);
 178         } while (pmd++, addr = next, addr != end);
 179
 180         if (nr_huge_updates)
 181                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 182
 183         return pages;
 184 }
 185
 186 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 187                 pgd_t *pgd, unsigned long addr, unsigned long end,
 188                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 189 {
 190         pud_t *pud;
 191         unsigned long next;
 192         unsigned long pages = 0;
 193
 194         pud = pud_offset(pgd, addr);
 195         do {
 196                 next = pud_addr_end(addr, end);
 197                 if (pud_none_or_clear_bad(pud))
 198                         continue;
 199                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 200                                  dirty_accountable, prot_numa);
 201         } while (pud++, addr = next, addr != end);
 202
 203         return pages;
 204 }
 205
 206 static unsigned long change_protection_range(struct vm_area_struct *vma,
 207                 unsigned long addr, unsigned long end, pgprot_t newprot,
 208                 int dirty_accountable, int prot_numa)
 209 {
 210         struct mm_struct *mm = vma->vm_mm;
 211         pgd_t *pgd;
 212         unsigned long next;
 213         unsigned long start = addr;
 214         unsigned long pages = 0;
 215
 216         BUG_ON(addr >= end);
 217         pgd = pgd_offset(mm, addr);
 218         flush_cache_range(vma, addr, end);
 219         set_tlb_flush_pending(mm);
 220         do {
 221                 next = pgd_addr_end(addr, end);
 222                 if (pgd_none_or_clear_bad(pgd))
 223                         continue;
 224                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 225                                  dirty_accountable, prot_numa);
 226         } while (pgd++, addr = next, addr != end);
 227
 228         /* Only flush the TLB if we actually modified any entries: */
 229         if (pages)
 230                 flush_tlb_range(vma, start, end);
 231         clear_tlb_flush_pending(mm);
 232
 233         return pages;
 234 }
 235
 236 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 237                        unsigned long end, pgprot_t newprot,
 238                        int dirty_accountable, int prot_numa)
 239 {
 240         struct mm_struct *mm = vma->vm_mm;
 241         unsigned long pages;
 242
 243         mmu_notifier_invalidate_range_start(mm, start, end);
 244         if (is_vm_hugetlb_page(vma))
 245                 pages = hugetlb_change_protection(vma, start, end, newprot);
 246         else
 247                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 248         mmu_notifier_invalidate_range_end(mm, start, end);
 249
 250         return pages;
 251 }
 252
 253 int
 254 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 255         unsigned long start, unsigned long end, unsigned long newflags)
 256 {
 257         struct mm_struct *mm = vma->vm_mm;
 258         unsigned long oldflags = vma->vm_flags;
 259         long nrpages = (end - start) >> PAGE_SHIFT;
 260         unsigned long charged = 0;
 261         pgoff_t pgoff;
 262         int error;
 263         int dirty_accountable = 0;
 264
 265         if (newflags == oldflags) {
 266                 *pprev = vma;
 267                 return 0;
 268         }
 269
 270         /*
 271          * If we make a private mapping writable we increase our commit;
 272          * but (without finer accounting) cannot reduce our commit if we
 273          * make it unwritable again. hugetlb mapping were accounted for
 274          * even if read-only so there is no need to account for them here
 275          */
 276         if (newflags & VM_WRITE) {
 277                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 278                                                 VM_SHARED|VM_NORESERVE))) {
 279                         charged = nrpages;
 280                         if (security_vm_enough_memory_mm(mm, charged))
 281                                 return -ENOMEM;
 282                         newflags |= VM_ACCOUNT;
 283                 }
 284         }
 285
 286         /*
 287          * First try to merge with previous and/or next vma.
 288          */
 289         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 290         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 291                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
 292         if (*pprev) {
 293                 vma = *pprev;
 294                 goto success;
 295         }
 296
 297         *pprev = vma;
 298
 299         if (start != vma->vm_start) {
 300                 error = split_vma(mm, vma, start, 1);
 301                 if (error)
 302                         goto fail;
 303         }
 304
 305         if (end != vma->vm_end) {
 306                 error = split_vma(mm, vma, end, 0);
 307                 if (error)
 308                         goto fail;
 309         }
 310
 311 success:
 312         /*
 313          * vm_flags and vm_page_prot are protected by the mmap_sem
 314          * held in write mode.
 315          */
 316         vma->vm_flags = newflags;
 317         vma->vm_page_prot = pgprot_modify(vma->vm_page_prot,
 318                                           vm_get_page_prot(newflags));
 319
 320         if (vma_wants_writenotify(vma)) {
 321                 vma->vm_page_prot = vm_get_page_prot(newflags & ~VM_SHARED);
 322                 dirty_accountable = 1;
 323         }
 324
 325         change_protection(vma, start, end, vma->vm_page_prot,
 326                           dirty_accountable, 0);
 327
 328         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
 329         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
 330         perf_event_mmap(vma);
 331         return 0;
 332
 333 fail:
 334         vm_unacct_memory(charged);
 335         return error;
 336 }
 337
 338 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 339                 unsigned long, prot)
 340 {
 341         unsigned long vm_flags, nstart, end, tmp, reqprot;
 342         struct vm_area_struct *vma, *prev;
 343         int error = -EINVAL;
 344         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 345         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 346         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 347                 return -EINVAL;
 348
 349         if (start & ~PAGE_MASK)
 350                 return -EINVAL;
 351         if (!len)
 352                 return 0;
 353         len = PAGE_ALIGN(len);
 354         end = start + len;
 355         if (end <= start)
 356                 return -ENOMEM;
 357         if (!arch_validate_prot(prot))
 358                 return -EINVAL;
 359
 360         reqprot = prot;
 361         /*
 362          * Does the application expect PROT_READ to imply PROT_EXEC:
 363          */
 364         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 365                 prot |= PROT_EXEC;
 366
 367         vm_flags = calc_vm_prot_bits(prot);
 368
 369         down_write(&current->mm->mmap_sem);
 370
 371         vma = find_vma(current->mm, start);
 372         error = -ENOMEM;
 373         if (!vma)
 374                 goto out;
 375         prev = vma->vm_prev;
 376         if (unlikely(grows & PROT_GROWSDOWN)) {
 377                 if (vma->vm_start >= end)
 378                         goto out;
 379                 start = vma->vm_start;
 380                 error = -EINVAL;
 381                 if (!(vma->vm_flags & VM_GROWSDOWN))
 382                         goto out;
 383         } else {
 384                 if (vma->vm_start > start)
 385                         goto out;
 386                 if (unlikely(grows & PROT_GROWSUP)) {
 387                         end = vma->vm_end;
 388                         error = -EINVAL;
 389                         if (!(vma->vm_flags & VM_GROWSUP))
 390                                 goto out;
 391                 }
 392         }
 393         if (start > vma->vm_start)
 394                 prev = vma;
 395
 396         for (nstart = start ; ; ) {
 397                 unsigned long newflags;
 398
 399                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 400
 401                 newflags = vm_flags;
 402                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 403
 404                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 405                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 406                         error = -EACCES;
 407                         goto out;
 408                 }
 409
 410                 error = security_file_mprotect(vma, reqprot, prot);
 411                 if (error)
 412                         goto out;
 413
 414                 tmp = vma->vm_end;
 415                 if (tmp > end)
 416                         tmp = end;
 417                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 418                 if (error)
 419                         goto out;
 420                 nstart = tmp;
 421
 422                 if (nstart < prev->vm_end)
 423                         nstart = prev->vm_end;
 424                 if (nstart >= end)
 425                         goto out;
 426
 427                 vma = prev->vm_next;
 428                 if (!vma || vma->vm_start != nstart) {
 429                         error = -ENOMEM;
 430                         goto out;
 431                 }
 432         }
 433 out:
 434         up_write(&current->mm->mmap_sem);
 435         return error;
 436 }