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 #include "internal.h"
  33
  34 /*
  35  * For a prot_numa update we only hold mmap_sem for read so there is a
  36  * potential race with faulting where a pmd was temporarily none. This
  37  * function checks for a transhuge pmd under the appropriate lock. It
  38  * returns a pte if it was successfully locked or NULL if it raced with
  39  * a transhuge insertion.
  40  */
  41 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
  42                         unsigned long addr, int prot_numa, spinlock_t **ptl)
  43 {
  44         pte_t *pte;
  45         spinlock_t *pmdl;
  46
  47         /* !prot_numa is protected by mmap_sem held for write */
  48         if (!prot_numa)
  49                 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  50
  51         pmdl = pmd_lock(vma->vm_mm, pmd);
  52         if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
  53                 spin_unlock(pmdl);
  54                 return NULL;
  55         }
  56
  57         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  58         spin_unlock(pmdl);
  59         return pte;
  60 }
  61
  62 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  63                 unsigned long addr, unsigned long end, pgprot_t newprot,
  64                 int dirty_accountable, int prot_numa)
  65 {
  66         struct mm_struct *mm = vma->vm_mm;
  67         pte_t *pte, oldpte;
  68         spinlock_t *ptl;
  69         unsigned long pages = 0;
  70
  71         pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
  72         if (!pte)
  73                 return 0;
  74
  75         arch_enter_lazy_mmu_mode();
  76         do {
  77                 oldpte = *pte;
  78                 if (pte_present(oldpte)) {
  79                         pte_t ptent;
  80                         bool preserve_write = prot_numa && pte_write(oldpte);
  81
  82                         /*
  83                          * Avoid trapping faults against the zero or KSM
  84                          * pages. See similar comment in change_huge_pmd.
  85                          */
  86                         if (prot_numa) {
  87                                 struct page *page;
  88
  89                                 page = vm_normal_page(vma, addr, oldpte);
  90                                 if (!page || PageKsm(page))
  91                                         continue;
  92
  93                                 /* Avoid TLB flush if possible */
  94                                 if (pte_protnone(oldpte))
  95                                         continue;
  96                         }
  97
  98                         ptent = ptep_modify_prot_start(mm, addr, pte);
  99                         ptent = pte_modify(ptent, newprot);
 100                         if (preserve_write)
 101                                 ptent = pte_mkwrite(ptent);
 102
 103                         /* Avoid taking write faults for known dirty pages */
 104                         if (dirty_accountable && pte_dirty(ptent) &&
 105                                         (pte_soft_dirty(ptent) ||
 106                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 107                                 ptent = pte_mkwrite(ptent);
 108                         }
 109                         ptep_modify_prot_commit(mm, addr, pte, ptent);
 110                         pages++;
 111                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 112                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 113
 114                         if (is_write_migration_entry(entry)) {
 115                                 pte_t newpte;
 116                                 /*
 117                                  * A protection check is difficult so
 118                                  * just be safe and disable write
 119                                  */
 120                                 make_migration_entry_read(&entry);
 121                                 newpte = swp_entry_to_pte(entry);
 122                                 if (pte_swp_soft_dirty(oldpte))
 123                                         newpte = pte_swp_mksoft_dirty(newpte);
 124                                 set_pte_at(mm, addr, pte, newpte);
 125
 126                                 pages++;
 127                         }
 128                 }
 129         } while (pte++, addr += PAGE_SIZE, addr != end);
 130         arch_leave_lazy_mmu_mode();
 131         pte_unmap_unlock(pte - 1, ptl);
 132
 133         return pages;
 134 }
 135
 136 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 137                 pud_t *pud, unsigned long addr, unsigned long end,
 138                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 139 {
 140         pmd_t *pmd;
 141         struct mm_struct *mm = vma->vm_mm;
 142         unsigned long next;
 143         unsigned long pages = 0;
 144         unsigned long nr_huge_updates = 0;
 145         unsigned long mni_start = 0;
 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) && pmd_none_or_clear_bad(pmd))
 153                         continue;
 154
 155                 /* invoke the mmu notifier if the pmd is populated */
 156                 if (!mni_start) {
 157                         mni_start = addr;
 158                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 159                 }
 160
 161                 if (pmd_trans_huge(*pmd)) {
 162                         if (next - addr != HPAGE_PMD_SIZE)
 163                                 split_huge_page_pmd(vma, addr, pmd);
 164                         else {
 165                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 166                                                 newprot, prot_numa);
 167
 168                                 if (nr_ptes) {
 169                                         if (nr_ptes == HPAGE_PMD_NR) {
 170                                                 pages += HPAGE_PMD_NR;
 171                                                 nr_huge_updates++;
 172                                         }
 173
 174                                         /* huge pmd was handled */
 175                                         continue;
 176                                 }
 177                         }
 178                         /* fall through, the trans huge pmd just split */
 179                 }
 180                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 181                                  dirty_accountable, prot_numa);
 182                 pages += this_pages;
 183         } while (pmd++, addr = next, addr != end);
 184
 185         if (mni_start)
 186                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 187
 188         if (nr_huge_updates)
 189                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 190         return pages;
 191 }
 192
 193 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 194                 pgd_t *pgd, unsigned long addr, unsigned long end,
 195                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 196 {
 197         pud_t *pud;
 198         unsigned long next;
 199         unsigned long pages = 0;
 200
 201         pud = pud_offset(pgd, addr);
 202         do {
 203                 next = pud_addr_end(addr, end);
 204                 if (pud_none_or_clear_bad(pud))
 205                         continue;
 206                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 207                                  dirty_accountable, prot_numa);
 208         } while (pud++, addr = next, addr != end);
 209
 210         return pages;
 211 }
 212
 213 static unsigned long change_protection_range(struct vm_area_struct *vma,
 214                 unsigned long addr, unsigned long end, pgprot_t newprot,
 215                 int dirty_accountable, int prot_numa)
 216 {
 217         struct mm_struct *mm = vma->vm_mm;
 218         pgd_t *pgd;
 219         unsigned long next;
 220         unsigned long start = addr;
 221         unsigned long pages = 0;
 222
 223         BUG_ON(addr >= end);
 224         pgd = pgd_offset(mm, addr);
 225         flush_cache_range(vma, addr, end);
 226         set_tlb_flush_pending(mm);
 227         do {
 228                 next = pgd_addr_end(addr, end);
 229                 if (pgd_none_or_clear_bad(pgd))
 230                         continue;
 231                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 232                                  dirty_accountable, prot_numa);
 233         } while (pgd++, addr = next, addr != end);
 234
 235         /* Only flush the TLB if we actually modified any entries: */
 236         if (pages)
 237                 flush_tlb_range(vma, start, end);
 238         clear_tlb_flush_pending(mm);
 239
 240         return pages;
 241 }
 242
 243 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 244                        unsigned long end, pgprot_t newprot,
 245                        int dirty_accountable, int prot_numa)
 246 {
 247         unsigned long pages;
 248
 249         if (is_vm_hugetlb_page(vma))
 250                 pages = hugetlb_change_protection(vma, start, end, newprot);
 251         else
 252                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 253
 254         return pages;
 255 }
 256
 257 int
 258 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 259         unsigned long start, unsigned long end, unsigned long newflags)
 260 {
 261         struct mm_struct *mm = vma->vm_mm;
 262         unsigned long oldflags = vma->vm_flags;
 263         long nrpages = (end - start) >> PAGE_SHIFT;
 264         unsigned long charged = 0;
 265         pgoff_t pgoff;
 266         int error;
 267         int dirty_accountable = 0;
 268
 269         if (newflags == oldflags) {
 270                 *pprev = vma;
 271                 return 0;
 272         }
 273
 274         /*
 275          * If we make a private mapping writable we increase our commit;
 276          * but (without finer accounting) cannot reduce our commit if we
 277          * make it unwritable again. hugetlb mapping were accounted for
 278          * even if read-only so there is no need to account for them here
 279          */
 280         if (newflags & VM_WRITE) {
 281                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 282                                                 VM_SHARED|VM_NORESERVE))) {
 283                         charged = nrpages;
 284                         if (security_vm_enough_memory_mm(mm, charged))
 285                                 return -ENOMEM;
 286                         newflags |= VM_ACCOUNT;
 287                 }
 288         }
 289
 290         /*
 291          * First try to merge with previous and/or next vma.
 292          */
 293         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 294         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 295                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 296                            vma->vm_userfaultfd_ctx);
 297         if (*pprev) {
 298                 vma = *pprev;
 299                 goto success;
 300         }
 301
 302         *pprev = vma;
 303
 304         if (start != vma->vm_start) {
 305                 error = split_vma(mm, vma, start, 1);
 306                 if (error)
 307                         goto fail;
 308         }
 309
 310         if (end != vma->vm_end) {
 311                 error = split_vma(mm, vma, end, 0);
 312                 if (error)
 313                         goto fail;
 314         }
 315
 316 success:
 317         /*
 318          * vm_flags and vm_page_prot are protected by the mmap_sem
 319          * held in write mode.
 320          */
 321         vma->vm_flags = newflags;
 322         dirty_accountable = vma_wants_writenotify(vma);
 323         vma_set_page_prot(vma);
 324
 325         change_protection(vma, start, end, vma->vm_page_prot,
 326                           dirty_accountable, 0);
 327
 328         /*
 329          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 330          * fault on access.
 331          */
 332         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 333                         (newflags & VM_WRITE)) {
 334                 populate_vma_page_range(vma, start, end, NULL);
 335         }
 336
 337         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
 338         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
 339         perf_event_mmap(vma);
 340         return 0;
 341
 342 fail:
 343         vm_unacct_memory(charged);
 344         return error;
 345 }
 346
 347 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 348                 unsigned long, prot)
 349 {
 350         unsigned long vm_flags, nstart, end, tmp, reqprot;
 351         struct vm_area_struct *vma, *prev;
 352         int error = -EINVAL;
 353         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 354         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 355         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 356                 return -EINVAL;
 357
 358         if (start & ~PAGE_MASK)
 359                 return -EINVAL;
 360         if (!len)
 361                 return 0;
 362         len = PAGE_ALIGN(len);
 363         end = start + len;
 364         if (end <= start)
 365                 return -ENOMEM;
 366         if (!arch_validate_prot(prot))
 367                 return -EINVAL;
 368
 369         reqprot = prot;
 370         /*
 371          * Does the application expect PROT_READ to imply PROT_EXEC:
 372          */
 373         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 374                 prot |= PROT_EXEC;
 375
 376         vm_flags = calc_vm_prot_bits(prot);
 377
 378         down_write(&current->mm->mmap_sem);
 379
 380         vma = find_vma(current->mm, start);
 381         error = -ENOMEM;
 382         if (!vma)
 383                 goto out;
 384         prev = vma->vm_prev;
 385         if (unlikely(grows & PROT_GROWSDOWN)) {
 386                 if (vma->vm_start >= end)
 387                         goto out;
 388                 start = vma->vm_start;
 389                 error = -EINVAL;
 390                 if (!(vma->vm_flags & VM_GROWSDOWN))
 391                         goto out;
 392         } else {
 393                 if (vma->vm_start > start)
 394                         goto out;
 395                 if (unlikely(grows & PROT_GROWSUP)) {
 396                         end = vma->vm_end;
 397                         error = -EINVAL;
 398                         if (!(vma->vm_flags & VM_GROWSUP))
 399                                 goto out;
 400                 }
 401         }
 402         if (start > vma->vm_start)
 403                 prev = vma;
 404
 405         for (nstart = start ; ; ) {
 406                 unsigned long newflags;
 407
 408                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 409
 410                 newflags = vm_flags;
 411                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 412
 413                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 414                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 415                         error = -EACCES;
 416                         goto out;
 417                 }
 418
 419                 error = security_file_mprotect(vma, reqprot, prot);
 420                 if (error)
 421                         goto out;
 422
 423                 tmp = vma->vm_end;
 424                 if (tmp > end)
 425                         tmp = end;
 426                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 427                 if (error)
 428                         goto out;
 429                 nstart = tmp;
 430
 431                 if (nstart < prev->vm_end)
 432                         nstart = prev->vm_end;
 433                 if (nstart >= end)
 434                         goto out;
 435
 436                 vma = prev->vm_next;
 437                 if (!vma || vma->vm_start != nstart) {
 438                         error = -ENOMEM;
 439                         goto out;
 440                 }
 441         }
 442 out:
 443         up_write(&current->mm->mmap_sem);
 444         return error;
 445 }