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