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