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/swapops.h>
  19 #include <linux/highmem.h>
  20 #include <linux/security.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/mmu_notifier.h>
  23 #include <linux/sched/sysctl.h>
  24 #include <linux/uaccess.h>
  25
  26 #include <asm/cacheflush.h>
  27 #include <asm/tlbflush.h>
  28
  29 #include "internal.h"
  30
  31 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
  32 {
  33         pgd_t *pgd;
  34         pud_t *pud;
  35         pmd_t *pmd;
  36
  37         pgd = pgd_offset(mm, addr);
  38         if (pgd_none_or_clear_bad(pgd))
  39                 return NULL;
  40
  41         pud = pud_offset(pgd, addr);
  42         if (pud_none_or_clear_bad(pud))
  43                 return NULL;
  44
  45         pmd = pmd_offset(pud, addr);
  46         if (pmd_none(*pmd))
  47                 return NULL;
  48
  49         return pmd;
  50 }
  51
  52 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
  53                             unsigned long addr)
  54 {
  55         pgd_t *pgd;
  56         pud_t *pud;
  57         pmd_t *pmd;
  58
  59         pgd = pgd_offset(mm, addr);
  60         pud = pud_alloc(mm, pgd, addr);
  61         if (!pud)
  62                 return NULL;
  63
  64         pmd = pmd_alloc(mm, pud, addr);
  65         if (!pmd)
  66                 return NULL;
  67
  68         VM_BUG_ON(pmd_trans_huge(*pmd));
  69
  70         return pmd;
  71 }
  72
  73 static pte_t move_soft_dirty_pte(pte_t pte)
  74 {
  75         /*
  76          * Set soft dirty bit so we can notice
  77          * in userspace the ptes were moved.
  78          */
  79 #ifdef CONFIG_MEM_SOFT_DIRTY
  80         if (pte_present(pte))
  81                 pte = pte_mksoft_dirty(pte);
  82         else if (is_swap_pte(pte))
  83                 pte = pte_swp_mksoft_dirty(pte);
  84         else if (pte_file(pte))
  85                 pte = pte_file_mksoft_dirty(pte);
  86 #endif
  87         return pte;
  88 }
  89
  90 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
  91                 unsigned long old_addr, unsigned long old_end,
  92                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
  93                 unsigned long new_addr, bool need_rmap_locks)
  94 {
  95         struct address_space *mapping = NULL;
  96         struct anon_vma *anon_vma = NULL;
  97         struct mm_struct *mm = vma->vm_mm;
  98         pte_t *old_pte, *new_pte, pte;
  99         spinlock_t *old_ptl, *new_ptl;
 100
 101         /*
 102          * When need_rmap_locks is true, we take the i_mmap_mutex and anon_vma
 103          * locks to ensure that rmap will always observe either the old or the
 104          * new ptes. This is the easiest way to avoid races with
 105          * truncate_pagecache(), page migration, etc...
 106          *
 107          * When need_rmap_locks is false, we use other ways to avoid
 108          * such races:
 109          *
 110          * - During exec() shift_arg_pages(), we use a specially tagged vma
 111          *   which rmap call sites look for using is_vma_temporary_stack().
 112          *
 113          * - During mremap(), new_vma is often known to be placed after vma
 114          *   in rmap traversal order. This ensures rmap will always observe
 115          *   either the old pte, or the new pte, or both (the page table locks
 116          *   serialize access to individual ptes, but only rmap traversal
 117          *   order guarantees that we won't miss both the old and new ptes).
 118          */
 119         if (need_rmap_locks) {
 120                 if (vma->vm_file) {
 121                         mapping = vma->vm_file->f_mapping;
 122                         mutex_lock(&mapping->i_mmap_mutex);
 123                 }
 124                 if (vma->anon_vma) {
 125                         anon_vma = vma->anon_vma;
 126                         anon_vma_lock_write(anon_vma);
 127                 }
 128         }
 129
 130         /*
 131          * We don't have to worry about the ordering of src and dst
 132          * pte locks because exclusive mmap_sem prevents deadlock.
 133          */
 134         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 135         new_pte = pte_offset_map(new_pmd, new_addr);
 136         new_ptl = pte_lockptr(mm, new_pmd);
 137         if (new_ptl != old_ptl)
 138                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 139         arch_enter_lazy_mmu_mode();
 140
 141         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 142                                    new_pte++, new_addr += PAGE_SIZE) {
 143                 if (pte_none(*old_pte))
 144                         continue;
 145                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
 146                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 147                 pte = move_soft_dirty_pte(pte);
 148                 set_pte_at(mm, new_addr, new_pte, pte);
 149         }
 150
 151         arch_leave_lazy_mmu_mode();
 152         if (new_ptl != old_ptl)
 153                 spin_unlock(new_ptl);
 154         pte_unmap(new_pte - 1);
 155         pte_unmap_unlock(old_pte - 1, old_ptl);
 156         if (anon_vma)
 157                 anon_vma_unlock_write(anon_vma);
 158         if (mapping)
 159                 mutex_unlock(&mapping->i_mmap_mutex);
 160 }
 161
 162 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 163
 164 unsigned long move_page_tables(struct vm_area_struct *vma,
 165                 unsigned long old_addr, struct vm_area_struct *new_vma,
 166                 unsigned long new_addr, unsigned long len,
 167                 bool need_rmap_locks)
 168 {
 169         unsigned long extent, next, old_end;
 170         pmd_t *old_pmd, *new_pmd;
 171         bool need_flush = false;
 172         unsigned long mmun_start;       /* For mmu_notifiers */
 173         unsigned long mmun_end;         /* For mmu_notifiers */
 174
 175         old_end = old_addr + len;
 176         flush_cache_range(vma, old_addr, old_end);
 177
 178         mmun_start = old_addr;
 179         mmun_end   = old_end;
 180         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 181
 182         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 183                 cond_resched();
 184                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 185                 /* even if next overflowed, extent below will be ok */
 186                 extent = next - old_addr;
 187                 if (extent > old_end - old_addr)
 188                         extent = old_end - old_addr;
 189                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 190                 if (!old_pmd)
 191                         continue;
 192                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 193                 if (!new_pmd)
 194                         break;
 195                 if (pmd_trans_huge(*old_pmd)) {
 196                         int err = 0;
 197                         if (extent == HPAGE_PMD_SIZE) {
 198                                 VM_BUG_ON_VMA(vma->vm_file || !vma->anon_vma,
 199                                               vma);
 200                                 /* See comment in move_ptes() */
 201                                 if (need_rmap_locks)
 202                                         anon_vma_lock_write(vma->anon_vma);
 203                                 err = move_huge_pmd(vma, new_vma, old_addr,
 204                                                     new_addr, old_end,
 205                                                     old_pmd, new_pmd);
 206                                 if (need_rmap_locks)
 207                                         anon_vma_unlock_write(vma->anon_vma);
 208                         }
 209                         if (err > 0) {
 210                                 need_flush = true;
 211                                 continue;
 212                         } else if (!err) {
 213                                 split_huge_page_pmd(vma, old_addr, old_pmd);
 214                         }
 215                         VM_BUG_ON(pmd_trans_huge(*old_pmd));
 216                 }
 217                 if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
 218                                                       new_pmd, new_addr))
 219                         break;
 220                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 221                 if (extent > next - new_addr)
 222                         extent = next - new_addr;
 223                 if (extent > LATENCY_LIMIT)
 224                         extent = LATENCY_LIMIT;
 225                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 226                           new_vma, new_pmd, new_addr, need_rmap_locks);
 227                 need_flush = true;
 228         }
 229         if (likely(need_flush))
 230                 flush_tlb_range(vma, old_end-len, old_addr);
 231
 232         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 233
 234         return len + old_addr - old_end;        /* how much done */
 235 }
 236
 237 static unsigned long move_vma(struct vm_area_struct *vma,
 238                 unsigned long old_addr, unsigned long old_len,
 239                 unsigned long new_len, unsigned long new_addr, bool *locked)
 240 {
 241         struct mm_struct *mm = vma->vm_mm;
 242         struct vm_area_struct *new_vma;
 243         unsigned long vm_flags = vma->vm_flags;
 244         unsigned long new_pgoff;
 245         unsigned long moved_len;
 246         unsigned long excess = 0;
 247         unsigned long hiwater_vm;
 248         int split = 0;
 249         int err;
 250         bool need_rmap_locks;
 251
 252         /*
 253          * We'd prefer to avoid failure later on in do_munmap:
 254          * which may split one vma into three before unmapping.
 255          */
 256         if (mm->map_count >= sysctl_max_map_count - 3)
 257                 return -ENOMEM;
 258
 259         /*
 260          * Advise KSM to break any KSM pages in the area to be moved:
 261          * it would be confusing if they were to turn up at the new
 262          * location, where they happen to coincide with different KSM
 263          * pages recently unmapped.  But leave vma->vm_flags as it was,
 264          * so KSM can come around to merge on vma and new_vma afterwards.
 265          */
 266         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 267                                                 MADV_UNMERGEABLE, &vm_flags);
 268         if (err)
 269                 return err;
 270
 271         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 272         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 273                            &need_rmap_locks);
 274         if (!new_vma)
 275                 return -ENOMEM;
 276
 277         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 278                                      need_rmap_locks);
 279         if (moved_len < old_len) {
 280                 /*
 281                  * On error, move entries back from new area to old,
 282                  * which will succeed since page tables still there,
 283                  * and then proceed to unmap new area instead of old.
 284                  */
 285                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 286                                  true);
 287                 vma = new_vma;
 288                 old_len = new_len;
 289                 old_addr = new_addr;
 290                 new_addr = -ENOMEM;
 291         }
 292
 293         /* Conceal VM_ACCOUNT so old reservation is not undone */
 294         if (vm_flags & VM_ACCOUNT) {
 295                 vma->vm_flags &= ~VM_ACCOUNT;
 296                 excess = vma->vm_end - vma->vm_start - old_len;
 297                 if (old_addr > vma->vm_start &&
 298                     old_addr + old_len < vma->vm_end)
 299                         split = 1;
 300         }
 301
 302         /*
 303          * If we failed to move page tables we still do total_vm increment
 304          * since do_munmap() will decrement it by old_len == new_len.
 305          *
 306          * Since total_vm is about to be raised artificially high for a
 307          * moment, we need to restore high watermark afterwards: if stats
 308          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 309          * If this were a serious issue, we'd add a flag to do_munmap().
 310          */
 311         hiwater_vm = mm->hiwater_vm;
 312         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 313
 314         if (do_munmap(mm, old_addr, old_len) < 0) {
 315                 /* OOM: unable to split vma, just get accounts right */
 316                 vm_unacct_memory(excess >> PAGE_SHIFT);
 317                 excess = 0;
 318         }
 319         mm->hiwater_vm = hiwater_vm;
 320
 321         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 322         if (excess) {
 323                 vma->vm_flags |= VM_ACCOUNT;
 324                 if (split)
 325                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 326         }
 327
 328         if (vm_flags & VM_LOCKED) {
 329                 mm->locked_vm += new_len >> PAGE_SHIFT;
 330                 *locked = true;
 331         }
 332
 333         return new_addr;
 334 }
 335
 336 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 337         unsigned long old_len, unsigned long new_len, unsigned long *p)
 338 {
 339         struct mm_struct *mm = current->mm;
 340         struct vm_area_struct *vma = find_vma(mm, addr);
 341
 342         if (!vma || vma->vm_start > addr)
 343                 goto Efault;
 344
 345         if (is_vm_hugetlb_page(vma))
 346                 goto Einval;
 347
 348         /* We can't remap across vm area boundaries */
 349         if (old_len > vma->vm_end - addr)
 350                 goto Efault;
 351
 352         /* Need to be careful about a growing mapping */
 353         if (new_len > old_len) {
 354                 unsigned long pgoff;
 355
 356                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 357                         goto Efault;
 358                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 359                 pgoff += vma->vm_pgoff;
 360                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 361                         goto Einval;
 362         }
 363
 364         if (vma->vm_flags & VM_LOCKED) {
 365                 unsigned long locked, lock_limit;
 366                 locked = mm->locked_vm << PAGE_SHIFT;
 367                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 368                 locked += new_len - old_len;
 369                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 370                         goto Eagain;
 371         }
 372
 373         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
 374                 goto Enomem;
 375
 376         if (vma->vm_flags & VM_ACCOUNT) {
 377                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 378                 if (security_vm_enough_memory_mm(mm, charged))
 379                         goto Efault;
 380                 *p = charged;
 381         }
 382
 383         return vma;
 384
 385 Efault: /* very odd choice for most of the cases, but... */
 386         return ERR_PTR(-EFAULT);
 387 Einval:
 388         return ERR_PTR(-EINVAL);
 389 Enomem:
 390         return ERR_PTR(-ENOMEM);
 391 Eagain:
 392         return ERR_PTR(-EAGAIN);
 393 }
 394
 395 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 396                 unsigned long new_addr, unsigned long new_len, bool *locked)
 397 {
 398         struct mm_struct *mm = current->mm;
 399         struct vm_area_struct *vma;
 400         unsigned long ret = -EINVAL;
 401         unsigned long charged = 0;
 402         unsigned long map_flags;
 403
 404         if (new_addr & ~PAGE_MASK)
 405                 goto out;
 406
 407         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 408                 goto out;
 409
 410         /* Check if the location we're moving into overlaps the
 411          * old location at all, and fail if it does.
 412          */
 413         if ((new_addr <= addr) && (new_addr+new_len) > addr)
 414                 goto out;
 415
 416         if ((addr <= new_addr) && (addr+old_len) > new_addr)
 417                 goto out;
 418
 419         ret = do_munmap(mm, new_addr, new_len);
 420         if (ret)
 421                 goto out;
 422
 423         if (old_len >= new_len) {
 424                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 425                 if (ret && old_len != new_len)
 426                         goto out;
 427                 old_len = new_len;
 428         }
 429
 430         vma = vma_to_resize(addr, old_len, new_len, &charged);
 431         if (IS_ERR(vma)) {
 432                 ret = PTR_ERR(vma);
 433                 goto out;
 434         }
 435
 436         map_flags = MAP_FIXED;
 437         if (vma->vm_flags & VM_MAYSHARE)
 438                 map_flags |= MAP_SHARED;
 439
 440         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 441                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 442                                 map_flags);
 443         if (ret & ~PAGE_MASK)
 444                 goto out1;
 445
 446         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
 447         if (!(ret & ~PAGE_MASK))
 448                 goto out;
 449 out1:
 450         vm_unacct_memory(charged);
 451
 452 out:
 453         return ret;
 454 }
 455
 456 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 457 {
 458         unsigned long end = vma->vm_end + delta;
 459         if (end < vma->vm_end) /* overflow */
 460                 return 0;
 461         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 462                 return 0;
 463         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 464                               0, MAP_FIXED) & ~PAGE_MASK)
 465                 return 0;
 466         return 1;
 467 }
 468
 469 /*
 470  * Expand (or shrink) an existing mapping, potentially moving it at the
 471  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 472  *
 473  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 474  * This option implies MREMAP_MAYMOVE.
 475  */
 476 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 477                 unsigned long, new_len, unsigned long, flags,
 478                 unsigned long, new_addr)
 479 {
 480         struct mm_struct *mm = current->mm;
 481         struct vm_area_struct *vma;
 482         unsigned long ret = -EINVAL;
 483         unsigned long charged = 0;
 484         bool locked = false;
 485
 486         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 487                 return ret;
 488
 489         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 490                 return ret;
 491
 492         if (addr & ~PAGE_MASK)
 493                 return ret;
 494
 495         old_len = PAGE_ALIGN(old_len);
 496         new_len = PAGE_ALIGN(new_len);
 497
 498         /*
 499          * We allow a zero old-len as a special case
 500          * for DOS-emu "duplicate shm area" thing. But
 501          * a zero new-len is nonsensical.
 502          */
 503         if (!new_len)
 504                 return ret;
 505
 506         down_write(&current->mm->mmap_sem);
 507
 508         if (flags & MREMAP_FIXED) {
 509                 ret = mremap_to(addr, old_len, new_addr, new_len,
 510                                 &locked);
 511                 goto out;
 512         }
 513
 514         /*
 515          * Always allow a shrinking remap: that just unmaps
 516          * the unnecessary pages..
 517          * do_munmap does all the needed commit accounting
 518          */
 519         if (old_len >= new_len) {
 520                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 521                 if (ret && old_len != new_len)
 522                         goto out;
 523                 ret = addr;
 524                 goto out;
 525         }
 526
 527         /*
 528          * Ok, we need to grow..
 529          */
 530         vma = vma_to_resize(addr, old_len, new_len, &charged);
 531         if (IS_ERR(vma)) {
 532                 ret = PTR_ERR(vma);
 533                 goto out;
 534         }
 535
 536         /* old_len exactly to the end of the area..
 537          */
 538         if (old_len == vma->vm_end - addr) {
 539                 /* can we just expand the current mapping? */
 540                 if (vma_expandable(vma, new_len - old_len)) {
 541                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 542
 543                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 544                                        vma->vm_pgoff, NULL)) {
 545                                 ret = -ENOMEM;
 546                                 goto out;
 547                         }
 548
 549                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
 550                         if (vma->vm_flags & VM_LOCKED) {
 551                                 mm->locked_vm += pages;
 552                                 locked = true;
 553                                 new_addr = addr;
 554                         }
 555                         ret = addr;
 556                         goto out;
 557                 }
 558         }
 559
 560         /*
 561          * We weren't able to just expand or shrink the area,
 562          * we need to create a new one and move it..
 563          */
 564         ret = -ENOMEM;
 565         if (flags & MREMAP_MAYMOVE) {
 566                 unsigned long map_flags = 0;
 567                 if (vma->vm_flags & VM_MAYSHARE)
 568                         map_flags |= MAP_SHARED;
 569
 570                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 571                                         vma->vm_pgoff +
 572                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 573                                         map_flags);
 574                 if (new_addr & ~PAGE_MASK) {
 575                         ret = new_addr;
 576                         goto out;
 577                 }
 578
 579                 ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
 580         }
 581 out:
 582         if (ret & ~PAGE_MASK)
 583                 vm_unacct_memory(charged);
 584         up_write(&current->mm->mmap_sem);
 585         if (locked && new_len > old_len)
 586                 mm_populate(new_addr + old_len, new_len - old_len);
 587         return ret;
 588 }