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