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 /*
  33  * For a prot_numa update we only hold mmap_sem for read so there is a
  34  * potential race with faulting where a pmd was temporarily none. This
  35  * function checks for a transhuge pmd under the appropriate lock. It
  36  * returns a pte if it was successfully locked or NULL if it raced with
  37  * a transhuge insertion.
  38  */
  39 static pte_t *lock_pte_protection(struct vm_area_struct *vma, pmd_t *pmd,
  40                         unsigned long addr, int prot_numa, spinlock_t **ptl)
  41 {
  42         pte_t *pte;
  43         spinlock_t *pmdl;
  44
  45         /* !prot_numa is protected by mmap_sem held for write */
  46         if (!prot_numa)
  47                 return pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  48
  49         pmdl = pmd_lock(vma->vm_mm, pmd);
  50         if (unlikely(pmd_trans_huge(*pmd) || pmd_none(*pmd))) {
  51                 spin_unlock(pmdl);
  52                 return NULL;
  53         }
  54
  55         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, ptl);
  56         spin_unlock(pmdl);
  57         return pte;
  58 }
  59
  60 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  61                 unsigned long addr, unsigned long end, pgprot_t newprot,
  62                 int dirty_accountable, int prot_numa)
  63 {
  64         struct mm_struct *mm = vma->vm_mm;
  65         pte_t *pte, oldpte;
  66         spinlock_t *ptl;
  67         unsigned long pages = 0;
  68
  69         pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
  70         if (!pte)
  71                 return 0;
  72
  73         arch_enter_lazy_mmu_mode();
  74         do {
  75                 oldpte = *pte;
  76                 if (pte_present(oldpte)) {
  77                         pte_t ptent;
  78                         bool preserve_write = prot_numa && pte_write(oldpte);
  79
  80                         /*
  81                          * Avoid trapping faults against the zero or KSM
  82                          * pages. See similar comment in change_huge_pmd.
  83                          */
  84                         if (prot_numa) {
  85                                 struct page *page;
  86
  87                                 page = vm_normal_page(vma, addr, oldpte);
  88                                 if (!page || PageKsm(page))
  89                                         continue;
  90
  91                                 /* Avoid TLB flush if possible */
  92                                 if (pte_protnone(oldpte))
  93                                         continue;
  94                         }
  95
  96                         ptent = ptep_modify_prot_start(mm, addr, pte);
  97                         ptent = pte_modify(ptent, newprot);
  98                         if (preserve_write)
  99                                 ptent = pte_mkwrite(ptent);
 100
 101                         /* Avoid taking write faults for known dirty pages */
 102                         if (dirty_accountable && pte_dirty(ptent) &&
 103                                         (pte_soft_dirty(ptent) ||
 104                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 105                                 ptent = pte_mkwrite(ptent);
 106                         }
 107                         ptep_modify_prot_commit(mm, addr, pte, ptent);
 108                         pages++;
 109                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 110                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 111
 112                         if (is_write_migration_entry(entry)) {
 113                                 pte_t newpte;
 114                                 /*
 115                                  * A protection check is difficult so
 116                                  * just be safe and disable write
 117                                  */
 118                                 make_migration_entry_read(&entry);
 119                                 newpte = swp_entry_to_pte(entry);
 120                                 if (pte_swp_soft_dirty(oldpte))
 121                                         newpte = pte_swp_mksoft_dirty(newpte);
 122                                 set_pte_at(mm, addr, pte, newpte);
 123
 124                                 pages++;
 125                         }
 126                 }
 127         } while (pte++, addr += PAGE_SIZE, addr != end);
 128         arch_leave_lazy_mmu_mode();
 129         pte_unmap_unlock(pte - 1, ptl);
 130
 131         return pages;
 132 }
 133
 134 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 135                 pud_t *pud, unsigned long addr, unsigned long end,
 136                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 137 {
 138         pmd_t *pmd;
 139         struct mm_struct *mm = vma->vm_mm;
 140         unsigned long next;
 141         unsigned long pages = 0;
 142         unsigned long nr_huge_updates = 0;
 143         unsigned long mni_start = 0;
 144
 145         pmd = pmd_offset(pud, addr);
 146         do {
 147                 unsigned long this_pages;
 148
 149                 next = pmd_addr_end(addr, end);
 150                 if (!pmd_trans_huge(*pmd) && pmd_none_or_clear_bad(pmd))
 151                         continue;
 152
 153                 /* invoke the mmu notifier if the pmd is populated */
 154                 if (!mni_start) {
 155                         mni_start = addr;
 156                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 157                 }
 158
 159                 if (pmd_trans_huge(*pmd)) {
 160                         if (next - addr != HPAGE_PMD_SIZE)
 161                                 split_huge_page_pmd(vma, addr, pmd);
 162                         else {
 163                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 164                                                 newprot, prot_numa);
 165
 166                                 if (nr_ptes) {
 167                                         if (nr_ptes == HPAGE_PMD_NR) {
 168                                                 pages += HPAGE_PMD_NR;
 169                                                 nr_huge_updates++;
 170                                         }
 171
 172                                         /* huge pmd was handled */
 173                                         continue;
 174                                 }
 175                         }
 176                         /* fall through, the trans huge pmd just split */
 177                 }
 178                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 179                                  dirty_accountable, prot_numa);
 180                 pages += this_pages;
 181         } while (pmd++, addr = next, addr != end);
 182
 183         if (mni_start)
 184                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 185
 186         if (nr_huge_updates)
 187                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 188         return pages;
 189 }
 190
 191 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 192                 pgd_t *pgd, unsigned long addr, unsigned long end,
 193                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 194 {
 195         pud_t *pud;
 196         unsigned long next;
 197         unsigned long pages = 0;
 198
 199         pud = pud_offset(pgd, addr);
 200         do {
 201                 next = pud_addr_end(addr, end);
 202                 if (pud_none_or_clear_bad(pud))
 203                         continue;
 204                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 205                                  dirty_accountable, prot_numa);
 206         } while (pud++, addr = next, addr != end);
 207
 208         return pages;
 209 }
 210
 211 static unsigned long change_protection_range(struct vm_area_struct *vma,
 212                 unsigned long addr, unsigned long end, pgprot_t newprot,
 213                 int dirty_accountable, int prot_numa)
 214 {
 215         struct mm_struct *mm = vma->vm_mm;
 216         pgd_t *pgd;
 217         unsigned long next;
 218         unsigned long start = addr;
 219         unsigned long pages = 0;
 220
 221         BUG_ON(addr >= end);
 222         pgd = pgd_offset(mm, addr);
 223         flush_cache_range(vma, addr, end);
 224         set_tlb_flush_pending(mm);
 225         do {
 226                 next = pgd_addr_end(addr, end);
 227                 if (pgd_none_or_clear_bad(pgd))
 228                         continue;
 229                 pages += change_pud_range(vma, pgd, addr, next, newprot,
 230                                  dirty_accountable, prot_numa);
 231         } while (pgd++, addr = next, addr != end);
 232
 233         /* Only flush the TLB if we actually modified any entries: */
 234         if (pages)
 235                 flush_tlb_range(vma, start, end);
 236         clear_tlb_flush_pending(mm);
 237
 238         return pages;
 239 }
 240
 241 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 242                        unsigned long end, pgprot_t newprot,
 243                        int dirty_accountable, int prot_numa)
 244 {
 245         unsigned long pages;
 246
 247         if (is_vm_hugetlb_page(vma))
 248                 pages = hugetlb_change_protection(vma, start, end, newprot);
 249         else
 250                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 251
 252         return pages;
 253 }
 254
 255 int
 256 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 257         unsigned long start, unsigned long end, unsigned long newflags)
 258 {
 259         struct mm_struct *mm = vma->vm_mm;
 260         unsigned long oldflags = vma->vm_flags;
 261         long nrpages = (end - start) >> PAGE_SHIFT;
 262         unsigned long charged = 0;
 263         pgoff_t pgoff;
 264         int error;
 265         int dirty_accountable = 0;
 266
 267         if (newflags == oldflags) {
 268                 *pprev = vma;
 269                 return 0;
 270         }
 271
 272         /*
 273          * If we make a private mapping writable we increase our commit;
 274          * but (without finer accounting) cannot reduce our commit if we
 275          * make it unwritable again. hugetlb mapping were accounted for
 276          * even if read-only so there is no need to account for them here
 277          */
 278         if (newflags & VM_WRITE) {
 279                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 280                                                 VM_SHARED|VM_NORESERVE))) {
 281                         charged = nrpages;
 282                         if (security_vm_enough_memory_mm(mm, charged))
 283                                 return -ENOMEM;
 284                         newflags |= VM_ACCOUNT;
 285                 }
 286         }
 287
 288         /*
 289          * First try to merge with previous and/or next vma.
 290          */
 291         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 292         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 293                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
 294         if (*pprev) {
 295                 vma = *pprev;
 296                 goto success;
 297         }
 298
 299         *pprev = vma;
 300
 301         if (start != vma->vm_start) {
 302                 error = split_vma(mm, vma, start, 1);
 303                 if (error)
 304                         goto fail;
 305         }
 306
 307         if (end != vma->vm_end) {
 308                 error = split_vma(mm, vma, end, 0);
 309                 if (error)
 310                         goto fail;
 311         }
 312
 313 success:
 314         /*
 315          * vm_flags and vm_page_prot are protected by the mmap_sem
 316          * held in write mode.
 317          */
 318         vma->vm_flags = newflags;
 319         dirty_accountable = vma_wants_writenotify(vma);
 320         vma_set_page_prot(vma);
 321
 322         change_protection(vma, start, end, vma->vm_page_prot,
 323                           dirty_accountable, 0);
 324
 325         vm_stat_account(mm, oldflags, vma->vm_file, -nrpages);
 326         vm_stat_account(mm, newflags, vma->vm_file, nrpages);
 327         perf_event_mmap(vma);
 328         return 0;
 329
 330 fail:
 331         vm_unacct_memory(charged);
 332         return error;
 333 }
 334
 335 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 336                 unsigned long, prot)
 337 {
 338         unsigned long vm_flags, nstart, end, tmp, reqprot;
 339         struct vm_area_struct *vma, *prev;
 340         int error = -EINVAL;
 341         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 342         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 343         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 344                 return -EINVAL;
 345
 346         if (start & ~PAGE_MASK)
 347                 return -EINVAL;
 348         if (!len)
 349                 return 0;
 350         len = PAGE_ALIGN(len);
 351         end = start + len;
 352         if (end <= start)
 353                 return -ENOMEM;
 354         if (!arch_validate_prot(prot))
 355                 return -EINVAL;
 356
 357         reqprot = prot;
 358         /*
 359          * Does the application expect PROT_READ to imply PROT_EXEC:
 360          */
 361         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
 362                 prot |= PROT_EXEC;
 363
 364         vm_flags = calc_vm_prot_bits(prot);
 365
 366         down_write(&current->mm->mmap_sem);
 367
 368         vma = find_vma(current->mm, start);
 369         error = -ENOMEM;
 370         if (!vma)
 371                 goto out;
 372         prev = vma->vm_prev;
 373         if (unlikely(grows & PROT_GROWSDOWN)) {
 374                 if (vma->vm_start >= end)
 375                         goto out;
 376                 start = vma->vm_start;
 377                 error = -EINVAL;
 378                 if (!(vma->vm_flags & VM_GROWSDOWN))
 379                         goto out;
 380         } else {
 381                 if (vma->vm_start > start)
 382                         goto out;
 383                 if (unlikely(grows & PROT_GROWSUP)) {
 384                         end = vma->vm_end;
 385                         error = -EINVAL;
 386                         if (!(vma->vm_flags & VM_GROWSUP))
 387                                 goto out;
 388                 }
 389         }
 390         if (start > vma->vm_start)
 391                 prev = vma;
 392
 393         for (nstart = start ; ; ) {
 394                 unsigned long newflags;
 395
 396                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 397
 398                 newflags = vm_flags;
 399                 newflags |= (vma->vm_flags & ~(VM_READ | VM_WRITE | VM_EXEC));
 400
 401                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 402                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 403                         error = -EACCES;
 404                         goto out;
 405                 }
 406
 407                 error = security_file_mprotect(vma, reqprot, prot);
 408                 if (error)
 409                         goto out;
 410
 411                 tmp = vma->vm_end;
 412                 if (tmp > end)
 413                         tmp = end;
 414                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 415                 if (error)
 416                         goto out;
 417                 nstart = tmp;
 418
 419                 if (nstart < prev->vm_end)
 420                         nstart = prev->vm_end;
 421                 if (nstart >= end)
 422                         goto out;
 423
 424                 vma = prev->vm_next;
 425                 if (!vma || vma->vm_start != nstart) {
 426                         error = -ENOMEM;
 427                         goto out;
 428                 }
 429         }
 430 out:
 431         up_write(&current->mm->mmap_sem);
 432         return error;
 433 }