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