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