mm/mremap.c

   1 /*
   2  *      mm/mremap.c
   3  *
   4  *      (C) Copyright 1996 Linus Torvalds
   5  *
   6  *      Address space accounting code   <alan@lxorguk.ukuu.org.uk>
   7  *      (C) Copyright 2002 Red Hat Inc, All Rights Reserved
   8  */
   9
  10 #include <linux/mm.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/shm.h>
  13 #include <linux/ksm.h>
  14 #include <linux/mman.h>
  15 #include <linux/swap.h>
  16 #include <linux/capability.h>
  17 #include <linux/fs.h>
  18 #include <linux/highmem.h>
  19 #include <linux/security.h>
  20 #include <linux/syscalls.h>
  21 #include <linux/mmu_notifier.h>
  22 #include <linux/sched/sysctl.h>
  23
  24 #include <asm/uaccess.h>
  25 #include <asm/cacheflush.h>
  26 #include <asm/tlbflush.h>
  27
  28 #include "internal.h"
  29
  30 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
  31 {
  32         pgd_t *pgd;
  33         pud_t *pud;
  34         pmd_t *pmd;
  35
  36         pgd = pgd_offset(mm, addr);
  37         if (pgd_none_or_clear_bad(pgd))
  38                 return NULL;
  39
  40         pud = pud_offset(pgd, addr);
  41         if (pud_none_or_clear_bad(pud))
  42                 return NULL;
  43
  44         pmd = pmd_offset(pud, addr);
  45         if (pmd_none(*pmd))
  46                 return NULL;
  47
  48         return pmd;
  49 }
  50
  51 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
  52                             unsigned long addr)
  53 {
  54         pgd_t *pgd;
  55         pud_t *pud;
  56         pmd_t *pmd;
  57
  58         pgd = pgd_offset(mm, addr);
  59         pud = pud_alloc(mm, pgd, addr);
  60         if (!pud)
  61                 return NULL;
  62
  63         pmd = pmd_alloc(mm, pud, addr);
  64         if (!pmd)
  65                 return NULL;
  66
  67         VM_BUG_ON(pmd_trans_huge(*pmd));
  68
  69         return pmd;
  70 }
  71
  72 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
  73                 unsigned long old_addr, unsigned long old_end,
  74                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
  75                 unsigned long new_addr, bool need_rmap_locks)
  76 {
  77         struct address_space *mapping = NULL;
  78         struct anon_vma *anon_vma = NULL;
  79         struct mm_struct *mm = vma->vm_mm;
  80         pte_t *old_pte, *new_pte, pte;
  81         spinlock_t *old_ptl, *new_ptl;
  82
  83         /*
  84          * When need_rmap_locks is true, we take the i_mmap_mutex and anon_vma
  85          * locks to ensure that rmap will always observe either the old or the
  86          * new ptes. This is the easiest way to avoid races with
  87          * truncate_pagecache(), page migration, etc...
  88          *
  89          * When need_rmap_locks is false, we use other ways to avoid
  90          * such races:
  91          *
  92          * - During exec() shift_arg_pages(), we use a specially tagged vma
  93          *   which rmap call sites look for using is_vma_temporary_stack().
  94          *
  95          * - During mremap(), new_vma is often known to be placed after vma
  96          *   in rmap traversal order. This ensures rmap will always observe
  97          *   either the old pte, or the new pte, or both (the page table locks
  98          *   serialize access to individual ptes, but only rmap traversal
  99          *   order guarantees that we won't miss both the old and new ptes).
 100          */
 101         if (need_rmap_locks) {
 102                 if (vma->vm_file) {
 103                         mapping = vma->vm_file->f_mapping;
 104                         mutex_lock(&mapping->i_mmap_mutex);
 105                 }
 106                 if (vma->anon_vma) {
 107                         anon_vma = vma->anon_vma;
 108                         anon_vma_lock_write(anon_vma);
 109                 }
 110         }
 111
 112         /*
 113          * We don't have to worry about the ordering of src and dst
 114          * pte locks because exclusive mmap_sem prevents deadlock.
 115          */
 116         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 117         new_pte = pte_offset_map(new_pmd, new_addr);
 118         new_ptl = pte_lockptr(mm, new_pmd);
 119         if (new_ptl != old_ptl)
 120                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 121         arch_enter_lazy_mmu_mode();
 122
 123         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 124                                    new_pte++, new_addr += PAGE_SIZE) {
 125                 if (pte_none(*old_pte))
 126                         continue;
 127                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
 128                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 129                 set_pte_at(mm, new_addr, new_pte, pte);
 130         }
 131
 132         arch_leave_lazy_mmu_mode();
 133         if (new_ptl != old_ptl)
 134                 spin_unlock(new_ptl);
 135         pte_unmap(new_pte - 1);
 136         pte_unmap_unlock(old_pte - 1, old_ptl);
 137         if (anon_vma)
 138                 anon_vma_unlock_write(anon_vma);
 139         if (mapping)
 140                 mutex_unlock(&mapping->i_mmap_mutex);
 141 }
 142
 143 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 144
 145 unsigned long move_page_tables(struct vm_area_struct *vma,
 146                 unsigned long old_addr, struct vm_area_struct *new_vma,
 147                 unsigned long new_addr, unsigned long len,
 148                 bool need_rmap_locks)
 149 {
 150         unsigned long extent, next, old_end;
 151         pmd_t *old_pmd, *new_pmd;
 152         bool need_flush = false;
 153         unsigned long mmun_start;       /* For mmu_notifiers */
 154         unsigned long mmun_end;         /* For mmu_notifiers */
 155
 156         old_end = old_addr + len;
 157         flush_cache_range(vma, old_addr, old_end);
 158
 159         mmun_start = old_addr;
 160         mmun_end   = old_end;
 161         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 162
 163         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 164                 cond_resched();
 165                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 166                 /* even if next overflowed, extent below will be ok */
 167                 extent = next - old_addr;
 168                 if (extent > old_end - old_addr)
 169                         extent = old_end - old_addr;
 170                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 171                 if (!old_pmd)
 172                         continue;
 173                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 174                 if (!new_pmd)
 175                         break;
 176                 if (pmd_trans_huge(*old_pmd)) {
 177                         int err = 0;
 178                         if (extent == HPAGE_PMD_SIZE) {
 179                                 VM_BUG_ON(vma->vm_file || !vma->anon_vma);
 180                                 /* See comment in move_ptes() */
 181                                 if (need_rmap_locks)
 182                                         anon_vma_lock_write(vma->anon_vma);
 183                                 err = move_huge_pmd(vma, new_vma, old_addr,
 184                                                     new_addr, old_end,
 185                                                     old_pmd, new_pmd);
 186                                 if (need_rmap_locks)
 187                                         anon_vma_unlock_write(vma->anon_vma);
 188                         }
 189                         if (err > 0) {
 190                                 need_flush = true;
 191                                 continue;
 192                         } else if (!err) {
 193                                 split_huge_page_pmd(vma, old_addr, old_pmd);
 194                         }
 195                         VM_BUG_ON(pmd_trans_huge(*old_pmd));
 196                 }
 197                 if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
 198                                                       new_pmd, new_addr))
 199                         break;
 200                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 201                 if (extent > next - new_addr)
 202                         extent = next - new_addr;
 203                 if (extent > LATENCY_LIMIT)
 204                         extent = LATENCY_LIMIT;
 205                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 206                           new_vma, new_pmd, new_addr, need_rmap_locks);
 207                 need_flush = true;
 208         }
 209         if (likely(need_flush))
 210                 flush_tlb_range(vma, old_end-len, old_addr);
 211
 212         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 213
 214         return len + old_addr - old_end;        /* how much done */
 215 }
 216
 217 static unsigned long move_vma(struct vm_area_struct *vma,
 218                 unsigned long old_addr, unsigned long old_len,
 219                 unsigned long new_len, unsigned long new_addr, bool *locked)
 220 {
 221         struct mm_struct *mm = vma->vm_mm;
 222         struct vm_area_struct *new_vma;
 223         unsigned long vm_flags = vma->vm_flags;
 224         unsigned long new_pgoff;
 225         unsigned long moved_len;
 226         unsigned long excess = 0;
 227         unsigned long hiwater_vm;
 228         int split = 0;
 229         int err;
 230         bool need_rmap_locks;
 231
 232         /*
 233          * We'd prefer to avoid failure later on in do_munmap:
 234          * which may split one vma into three before unmapping.
 235          */
 236         if (mm->map_count >= sysctl_max_map_count - 3)
 237                 return -ENOMEM;
 238
 239         /*
 240          * Advise KSM to break any KSM pages in the area to be moved:
 241          * it would be confusing if they were to turn up at the new
 242          * location, where they happen to coincide with different KSM
 243          * pages recently unmapped.  But leave vma->vm_flags as it was,
 244          * so KSM can come around to merge on vma and new_vma afterwards.
 245          */
 246         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 247                                                 MADV_UNMERGEABLE, &vm_flags);
 248         if (err)
 249                 return err;
 250
 251         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 252         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 253                            &need_rmap_locks);
 254         if (!new_vma)
 255                 return -ENOMEM;
 256
 257         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 258                                      need_rmap_locks);
 259         if (moved_len < old_len) {
 260                 /*
 261                  * On error, move entries back from new area to old,
 262                  * which will succeed since page tables still there,
 263                  * and then proceed to unmap new area instead of old.
 264                  */
 265                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 266                                  true);
 267                 vma = new_vma;
 268                 old_len = new_len;
 269                 old_addr = new_addr;
 270                 new_addr = -ENOMEM;
 271         }
 272
 273         /* Conceal VM_ACCOUNT so old reservation is not undone */
 274         if (vm_flags & VM_ACCOUNT) {
 275                 vma->vm_flags &= ~VM_ACCOUNT;
 276                 excess = vma->vm_end - vma->vm_start - old_len;
 277                 if (old_addr > vma->vm_start &&
 278                     old_addr + old_len < vma->vm_end)
 279                         split = 1;
 280         }
 281
 282         /*
 283          * If we failed to move page tables we still do total_vm increment
 284          * since do_munmap() will decrement it by old_len == new_len.
 285          *
 286          * Since total_vm is about to be raised artificially high for a
 287          * moment, we need to restore high watermark afterwards: if stats
 288          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 289          * If this were a serious issue, we'd add a flag to do_munmap().
 290          */
 291         hiwater_vm = mm->hiwater_vm;
 292         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 293
 294         if (do_munmap(mm, old_addr, old_len) < 0) {
 295                 /* OOM: unable to split vma, just get accounts right */
 296                 vm_unacct_memory(excess >> PAGE_SHIFT);
 297                 excess = 0;
 298         }
 299         mm->hiwater_vm = hiwater_vm;
 300
 301         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 302         if (excess) {
 303                 vma->vm_flags |= VM_ACCOUNT;
 304                 if (split)
 305                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 306         }
 307
 308         if (vm_flags & VM_LOCKED) {
 309                 mm->locked_vm += new_len >> PAGE_SHIFT;
 310                 *locked = true;
 311         }
 312
 313         return new_addr;
 314 }
 315
 316 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 317         unsigned long old_len, unsigned long new_len, unsigned long *p)
 318 {
 319         struct mm_struct *mm = current->mm;
 320         struct vm_area_struct *vma = find_vma(mm, addr);
 321
 322         if (!vma || vma->vm_start > addr)
 323                 goto Efault;
 324
 325         if (is_vm_hugetlb_page(vma))
 326                 goto Einval;
 327
 328         /* We can't remap across vm area boundaries */
 329         if (old_len > vma->vm_end - addr)
 330                 goto Efault;
 331
 332         /* Need to be careful about a growing mapping */
 333         if (new_len > old_len) {
 334                 unsigned long pgoff;
 335
 336                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 337                         goto Efault;
 338                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 339                 pgoff += vma->vm_pgoff;
 340                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 341                         goto Einval;
 342         }
 343
 344         if (vma->vm_flags & VM_LOCKED) {
 345                 unsigned long locked, lock_limit;
 346                 locked = mm->locked_vm << PAGE_SHIFT;
 347                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 348                 locked += new_len - old_len;
 349                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 350                         goto Eagain;
 351         }
 352
 353         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
 354                 goto Enomem;
 355
 356         if (vma->vm_flags & VM_ACCOUNT) {
 357                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 358                 if (security_vm_enough_memory_mm(mm, charged))
 359                         goto Efault;
 360                 *p = charged;
 361         }
 362
 363         return vma;
 364
 365 Efault: /* very odd choice for most of the cases, but... */
 366         return ERR_PTR(-EFAULT);
 367 Einval:
 368         return ERR_PTR(-EINVAL);
 369 Enomem:
 370         return ERR_PTR(-ENOMEM);
 371 Eagain:
 372         return ERR_PTR(-EAGAIN);
 373 }
 374
 375 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 376                 unsigned long new_addr, unsigned long new_len, bool *locked)
 377 {
 378         struct mm_struct *mm = current->mm;
 379         struct vm_area_struct *vma;
 380         unsigned long ret = -EINVAL;
 381         unsigned long charged = 0;
 382         unsigned long map_flags;
 383
 384         if (new_addr & ~PAGE_MASK)
 385                 goto out;
 386
 387         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 388                 goto out;
 389
 390         /* Check if the location we're moving into overlaps the
 391          * old location at all, and fail if it does.
 392          */
 393         if ((new_addr <= addr) && (new_addr+new_len) > addr)
 394                 goto out;
 395
 396         if ((addr <= new_addr) && (addr+old_len) > new_addr)
 397                 goto out;
 398
 399         ret = do_munmap(mm, new_addr, new_len);
 400         if (ret)
 401                 goto out;
 402
 403         if (old_len >= new_len) {
 404                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 405                 if (ret && old_len != new_len)
 406                         goto out;
 407                 old_len = new_len;
 408         }
 409
 410         vma = vma_to_resize(addr, old_len, new_len, &charged);
 411         if (IS_ERR(vma)) {
 412                 ret = PTR_ERR(vma);
 413                 goto out;
 414         }
 415
 416         map_flags = MAP_FIXED;
 417         if (vma->vm_flags & VM_MAYSHARE)
 418                 map_flags |= MAP_SHARED;
 419
 420         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 421                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 422                                 map_flags);
 423         if (ret & ~PAGE_MASK)
 424                 goto out1;
 425
 426         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
 427         if (!(ret & ~PAGE_MASK))
 428                 goto out;
 429 out1:
 430         vm_unacct_memory(charged);
 431
 432 out:
 433         return ret;
 434 }
 435
 436 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 437 {
 438         unsigned long end = vma->vm_end + delta;
 439         if (end < vma->vm_end) /* overflow */
 440                 return 0;
 441         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 442                 return 0;
 443         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 444                               0, MAP_FIXED) & ~PAGE_MASK)
 445                 return 0;
 446         return 1;
 447 }
 448
 449 /*
 450  * Expand (or shrink) an existing mapping, potentially moving it at the
 451  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 452  *
 453  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 454  * This option implies MREMAP_MAYMOVE.
 455  */
 456 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 457                 unsigned long, new_len, unsigned long, flags,
 458                 unsigned long, new_addr)
 459 {
 460         struct mm_struct *mm = current->mm;
 461         struct vm_area_struct *vma;
 462         unsigned long ret = -EINVAL;
 463         unsigned long charged = 0;
 464         bool locked = false;
 465
 466         down_write(&current->mm->mmap_sem);
 467
 468         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 469                 goto out;
 470
 471         if (addr & ~PAGE_MASK)
 472                 goto out;
 473
 474         old_len = PAGE_ALIGN(old_len);
 475         new_len = PAGE_ALIGN(new_len);
 476
 477         /*
 478          * We allow a zero old-len as a special case
 479          * for DOS-emu "duplicate shm area" thing. But
 480          * a zero new-len is nonsensical.
 481          */
 482         if (!new_len)
 483                 goto out;
 484
 485         if (flags & MREMAP_FIXED) {
 486                 if (flags & MREMAP_MAYMOVE)
 487                         ret = mremap_to(addr, old_len, new_addr, new_len,
 488                                         &locked);
 489                 goto out;
 490         }
 491
 492         /*
 493          * Always allow a shrinking remap: that just unmaps
 494          * the unnecessary pages..
 495          * do_munmap does all the needed commit accounting
 496          */
 497         if (old_len >= new_len) {
 498                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 499                 if (ret && old_len != new_len)
 500                         goto out;
 501                 ret = addr;
 502                 goto out;
 503         }
 504
 505         /*
 506          * Ok, we need to grow..
 507          */
 508         vma = vma_to_resize(addr, old_len, new_len, &charged);
 509         if (IS_ERR(vma)) {
 510                 ret = PTR_ERR(vma);
 511                 goto out;
 512         }
 513
 514         /* old_len exactly to the end of the area..
 515          */
 516         if (old_len == vma->vm_end - addr) {
 517                 /* can we just expand the current mapping? */
 518                 if (vma_expandable(vma, new_len - old_len)) {
 519                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 520
 521                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 522                                        vma->vm_pgoff, NULL)) {
 523                                 ret = -ENOMEM;
 524                                 goto out;
 525                         }
 526
 527                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
 528                         if (vma->vm_flags & VM_LOCKED) {
 529                                 mm->locked_vm += pages;
 530                                 locked = true;
 531                                 new_addr = addr;
 532                         }
 533                         ret = addr;
 534                         goto out;
 535                 }
 536         }
 537
 538         /*
 539          * We weren't able to just expand or shrink the area,
 540          * we need to create a new one and move it..
 541          */
 542         ret = -ENOMEM;
 543         if (flags & MREMAP_MAYMOVE) {
 544                 unsigned long map_flags = 0;
 545                 if (vma->vm_flags & VM_MAYSHARE)
 546                         map_flags |= MAP_SHARED;
 547
 548                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 549                                         vma->vm_pgoff +
 550                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 551                                         map_flags);
 552                 if (new_addr & ~PAGE_MASK) {
 553                         ret = new_addr;
 554                         goto out;
 555                 }
 556
 557                 ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
 558         }
 559 out:
 560         if (ret & ~PAGE_MASK)
 561                 vm_unacct_memory(charged);
 562         up_write(&current->mm->mmap_sem);
 563         if (locked && new_len > old_len)
 564                 mm_populate(new_addr + old_len, new_len - old_len);
 565         return ret;
 566 }