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                 next = pmd_addr_end(addr, end);
 150                 if (pmd_trans_huge(*pmd)) {
 151                         if (next - addr != HPAGE_PMD_SIZE)
 152                                 split_huge_page_pmd(vma, addr, pmd);
 153                         else {
 154                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 155                                                 newprot, prot_numa);
 156
 157                                 if (nr_ptes) {
 158                                         if (nr_ptes == HPAGE_PMD_NR)
 159                                                 pages++;
 160
 161                                         continue;
 162                                 }
 163                         }
 164                         /* fall through */
 165                 }
 166                 if (pmd_none_or_clear_bad(pmd))
 167                         continue;
 168                 pages += change_pte_range(vma, pmd, addr, next, newprot,
 169                                  dirty_accountable, prot_numa, &all_same_nidpid);
 170
 171                 /*
 172                  * If we are changing protections for NUMA hinting faults then
 173                  * set pmd_numa if the examined pages were all on the same
 174                  * node. This allows a regular PMD to be handled as one fault
 175                  * and effectively batches the taking of the PTL
 176                  */
 177                 if (prot_numa && all_same_nidpid)
 178                         change_pmd_protnuma(vma->vm_mm, addr, pmd);
 179         } while (pmd++, addr = next, addr != end);
 180
 181         return pages;
 182 }
 183
 184 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 185                 pgd_t *pgd, unsigned long addr, unsigned long end,
 186                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 187 {
 188         pud_t *pud;
 189         unsigned long next;
 190         unsigned long pages = 0;
 191
 192         pud = pud_offset(pgd, addr);
 193         do {
 194                 next = pud_addr_end(addr, end);
 195                 if (pud_none_or_clear_bad(pud))
 196                         continue;
 197                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 198                                  dirty_accountable, prot_numa);
 199         } while (pud++, addr = next, addr != end);
 200
 201         return pages;
 202 }
 203
 204 static unsigned long change_protection_range(struct vm_area_struct *vma,
 205                 unsigned long addr, unsigned long end, pgprot_t newprot,
 206                 int dirty_accountable, int prot_numa)
 207 {
 208         struct mm_struct *mm = vma->vm_mm;
 209         pgd_t *pgd;
 210         unsigned long next;
 211         unsigned long start = addr;
 212         unsigned long pages = 0;
 213
 214         BUG_ON(addr >= end);
 215         pgd = pgd_offset(mm, addr);
 216         flush_cache_range(vma, addr, end);
 217         do {
 218                 next = pgd_addr_end(addr, end);
 219                 if (pgd_none_or_clear_bad(pgd))
 220                         continue;
 221                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 222                                  dirty_accountable, prot_numa);
 223         } while (pgd++, addr = next, addr != end);
 224
 225         /* Only flush the TLB if we actually modified any entries: */
 226         if (pages)
 227                 flush_tlb_range(vma, start, end);
 228
 229         return pages;
 230 }
 231
 232 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 233                        unsigned long end, pgprot_t newprot,
 234                        int dirty_accountable, int prot_numa)
 235 {
 236         struct mm_struct *mm = vma->vm_mm;
 237         unsigned long pages;
 238
 239         mmu_notifier_invalidate_range_start(mm, start, end);
 240         if (is_vm_hugetlb_page(vma))
 241                 pages = hugetlb_change_protection(vma, start, end, newprot);
 242         else
 243                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 244         mmu_notifier_invalidate_range_end(mm, start, end);
 245
 246         return pages;
 247 }
 248
 249 int
 250 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 251         unsigned long start, unsigned long end, unsigned long newflags)
 252 {
 253         struct mm_struct *mm = vma->vm_mm;
 254         unsigned long oldflags = vma->vm_flags;
 255         long nrpages = (end - start) >> PAGE_SHIFT;
 256         unsigned long charged = 0;
 257         pgoff_t pgoff;
 258         int error;
 259         int dirty_accountable = 0;
 260
 261         if (newflags == oldflags) {
 262                 *pprev = vma;
 263                 return 0;
 264         }
 265
 266         /*
 267          * If we make a private mapping writable we increase our commit;
 268          * but (without finer accounting) cannot reduce our commit if we
 269          * make it unwritable again. hugetlb mapping were accounted for
 270          * even if read-only so there is no need to account for them here
 271          */
 272         if (newflags & VM_WRITE) {
 273                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 274                                                 VM_SHARED|VM_NORESERVE))) {
 275                         charged = nrpages;
 276                         if (security_vm_enough_memory_mm(mm, charged))
 277                                 return -ENOMEM;
 278                         newflags |= VM_ACCOUNT;
 279                 }
 280         }
 281
 282         /*
 283          * First try to merge with previous and/or next vma.
 284          */
 285         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 286         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 287                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
 288         if (*pprev) {
 289                 vma = *pprev;
 290                 goto success;
 291         }
 292
 293         *pprev = vma;
 294
 295         if (start != vma->vm_start) {
 296                 error = split_vma(mm, vma, start, 1);
 297                 if (error)
 298                         goto fail;
 299         }
 300
 301         if (end != vma->vm_end) {
 302                 error = split_vma(mm, vma, end, 0);
 303                 if (error)
 304                         goto fail;
 305         }
 306
 307 success:
 308         /*
 309          * vm_flags and vm_page_prot are protected by the mmap_sem
 310          * held in write mode.
 311          */
 312         vma->vm_flags = newflags;
 313         vma->vm_page_prot = pgprot_modify(vma->vm_page_prot,
 314                                           vm_get_page_prot(newflags));
 315
 316         if (vma_wants_writenotify(vma)) {
 317                 vma->vm_page_prot = vm_get_page_prot(newflags & ~VM_SHARED);
 318                 dirty_accountable = 1;
 319         }
 320
 321         change_protection(vma, start, end, vma->vm_page_prot,
 322                           dirty_accountable, 0);
 323
 324         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
 325         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
 326         perf_event_mmap(vma);
 327         return 0;
 328
 329 fail:
 330         vm_unacct_memory(charged);
 331         return error;
 332 }
 333
 334 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 335                 unsigned long, prot)
 336 {
 337         unsigned long vm_flags, nstart, end, tmp, reqprot;
 338         struct vm_area_struct *vma, *prev;
 339         int error = -EINVAL;
 340         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 341         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 342         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 343                 return -EINVAL;
 344
 345         if (start & ~PAGE_MASK)
 346                 return -EINVAL;
 347         if (!len)
 348                 return 0;
 349         len = PAGE_ALIGN(len);
 350         end = start + len;
 351         if (end <= start)
 352                 return -ENOMEM;
 353         if (!arch_validate_prot(prot))
 354                 return -EINVAL;
 355
 356         reqprot = prot;
 357         /*
 358          * Does the application expect PROT_READ to imply PROT_EXEC:
 359          */
 360         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 361                 prot |= PROT_EXEC;
 362
 363         vm_flags = calc_vm_prot_bits(prot);
 364
 365         down_write(&current->mm->mmap_sem);
 366
 367         vma = find_vma(current->mm, start);
 368         error = -ENOMEM;
 369         if (!vma)
 370                 goto out;
 371         prev = vma->vm_prev;
 372         if (unlikely(grows & PROT_GROWSDOWN)) {
 373                 if (vma->vm_start >= end)
 374                         goto out;
 375                 start = vma->vm_start;
 376                 error = -EINVAL;
 377                 if (!(vma->vm_flags & VM_GROWSDOWN))
 378                         goto out;
 379         } else {
 380                 if (vma->vm_start > start)
 381                         goto out;
 382                 if (unlikely(grows & PROT_GROWSUP)) {
 383                         end = vma->vm_end;
 384                         error = -EINVAL;
 385                         if (!(vma->vm_flags & VM_GROWSUP))
 386                                 goto out;
 387                 }
 388         }
 389         if (start > vma->vm_start)
 390                 prev = vma;
 391
 392         for (nstart = start ; ; ) {
 393                 unsigned long newflags;
 394
 395                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 396
 397                 newflags = vm_flags;
 398                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 399
 400                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 401                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 402                         error = -EACCES;
 403                         goto out;
 404                 }
 405
 406                 error = security_file_mprotect(vma, reqprot, prot);
 407                 if (error)
 408                         goto out;
 409
 410                 tmp = vma->vm_end;
 411                 if (tmp > end)
 412                         tmp = end;
 413                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 414                 if (error)
 415                         goto out;
 416                 nstart = tmp;
 417
 418                 if (nstart < prev->vm_end)
 419                         nstart = prev->vm_end;
 420                 if (nstart >= end)
 421                         goto out;
 422
 423                 vma = prev->vm_next;
 424                 if (!vma || vma->vm_start != nstart) {
 425                         error = -ENOMEM;
 426                         goto out;
 427                 }
 428         }
 429 out:
 430         up_write(&current->mm->mmap_sem);
 431         return error;
 432 }