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 #include <linux/mm-arch-hooks.h>
  26
  27 #include <asm/cacheflush.h>
  28 #include <asm/tlbflush.h>
  29
  30 #include "internal.h"
  31
  32 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
  33 {
  34         pgd_t *pgd;
  35         pud_t *pud;
  36         pmd_t *pmd;
  37
  38         pgd = pgd_offset(mm, addr);
  39         if (pgd_none_or_clear_bad(pgd))
  40                 return NULL;
  41
  42         pud = pud_offset(pgd, addr);
  43         if (pud_none_or_clear_bad(pud))
  44                 return NULL;
  45
  46         pmd = pmd_offset(pud, addr);
  47         if (pmd_none(*pmd))
  48                 return NULL;
  49
  50         return pmd;
  51 }
  52
  53 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
  54                             unsigned long addr)
  55 {
  56         pgd_t *pgd;
  57         pud_t *pud;
  58         pmd_t *pmd;
  59
  60         pgd = pgd_offset(mm, addr);
  61         pud = pud_alloc(mm, pgd, addr);
  62         if (!pud)
  63                 return NULL;
  64
  65         pmd = pmd_alloc(mm, pud, addr);
  66         if (!pmd)
  67                 return NULL;
  68
  69         VM_BUG_ON(pmd_trans_huge(*pmd));
  70
  71         return pmd;
  72 }
  73
  74 static pte_t move_soft_dirty_pte(pte_t pte)
  75 {
  76         /*
  77          * Set soft dirty bit so we can notice
  78          * in userspace the ptes were moved.
  79          */
  80 #ifdef CONFIG_MEM_SOFT_DIRTY
  81         if (pte_present(pte))
  82                 pte = pte_mksoft_dirty(pte);
  83         else if (is_swap_pte(pte))
  84                 pte = pte_swp_mksoft_dirty(pte);
  85 #endif
  86         return pte;
  87 }
  88
  89 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
  90                 unsigned long old_addr, unsigned long old_end,
  91                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
  92                 unsigned long new_addr, bool need_rmap_locks)
  93 {
  94         struct address_space *mapping = NULL;
  95         struct anon_vma *anon_vma = NULL;
  96         struct mm_struct *mm = vma->vm_mm;
  97         pte_t *old_pte, *new_pte, pte;
  98         spinlock_t *old_ptl, *new_ptl;
  99
 100         /*
 101          * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
 102          * locks to ensure that rmap will always observe either the old or the
 103          * new ptes. This is the easiest way to avoid races with
 104          * truncate_pagecache(), page migration, etc...
 105          *
 106          * When need_rmap_locks is false, we use other ways to avoid
 107          * such races:
 108          *
 109          * - During exec() shift_arg_pages(), we use a specially tagged vma
 110          *   which rmap call sites look for using is_vma_temporary_stack().
 111          *
 112          * - During mremap(), new_vma is often known to be placed after vma
 113          *   in rmap traversal order. This ensures rmap will always observe
 114          *   either the old pte, or the new pte, or both (the page table locks
 115          *   serialize access to individual ptes, but only rmap traversal
 116          *   order guarantees that we won't miss both the old and new ptes).
 117          */
 118         if (need_rmap_locks) {
 119                 if (vma->vm_file) {
 120                         mapping = vma->vm_file->f_mapping;
 121                         i_mmap_lock_write(mapping);
 122                 }
 123                 if (vma->anon_vma) {
 124                         anon_vma = vma->anon_vma;
 125                         anon_vma_lock_write(anon_vma);
 126                 }
 127         }
 128
 129         /*
 130          * We don't have to worry about the ordering of src and dst
 131          * pte locks because exclusive mmap_sem prevents deadlock.
 132          */
 133         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 134         new_pte = pte_offset_map(new_pmd, new_addr);
 135         new_ptl = pte_lockptr(mm, new_pmd);
 136         if (new_ptl != old_ptl)
 137                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 138         arch_enter_lazy_mmu_mode();
 139
 140         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 141                                    new_pte++, new_addr += PAGE_SIZE) {
 142                 if (pte_none(*old_pte))
 143                         continue;
 144                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
 145                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 146                 pte = move_soft_dirty_pte(pte);
 147                 set_pte_at(mm, new_addr, new_pte, pte);
 148         }
 149
 150         arch_leave_lazy_mmu_mode();
 151         if (new_ptl != old_ptl)
 152                 spin_unlock(new_ptl);
 153         pte_unmap(new_pte - 1);
 154         pte_unmap_unlock(old_pte - 1, old_ptl);
 155         if (anon_vma)
 156                 anon_vma_unlock_write(anon_vma);
 157         if (mapping)
 158                 i_mmap_unlock_write(mapping);
 159 }
 160
 161 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 162
 163 unsigned long move_page_tables(struct vm_area_struct *vma,
 164                 unsigned long old_addr, struct vm_area_struct *new_vma,
 165                 unsigned long new_addr, unsigned long len,
 166                 bool need_rmap_locks)
 167 {
 168         unsigned long extent, next, old_end;
 169         pmd_t *old_pmd, *new_pmd;
 170         bool need_flush = false;
 171         unsigned long mmun_start;       /* For mmu_notifiers */
 172         unsigned long mmun_end;         /* For mmu_notifiers */
 173
 174         old_end = old_addr + len;
 175         flush_cache_range(vma, old_addr, old_end);
 176
 177         mmun_start = old_addr;
 178         mmun_end   = old_end;
 179         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 180
 181         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 182                 cond_resched();
 183                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 184                 /* even if next overflowed, extent below will be ok */
 185                 extent = next - old_addr;
 186                 if (extent > old_end - old_addr)
 187                         extent = old_end - old_addr;
 188                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 189                 if (!old_pmd)
 190                         continue;
 191                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 192                 if (!new_pmd)
 193                         break;
 194                 if (pmd_trans_huge(*old_pmd)) {
 195                         int err = 0;
 196                         if (extent == HPAGE_PMD_SIZE) {
 197                                 VM_BUG_ON_VMA(vma->vm_file || !vma->anon_vma,
 198                                               vma);
 199                                 /* See comment in move_ptes() */
 200                                 if (need_rmap_locks)
 201                                         anon_vma_lock_write(vma->anon_vma);
 202                                 err = move_huge_pmd(vma, new_vma, old_addr,
 203                                                     new_addr, old_end,
 204                                                     old_pmd, new_pmd);
 205                                 if (need_rmap_locks)
 206                                         anon_vma_unlock_write(vma->anon_vma);
 207                         }
 208                         if (err > 0) {
 209                                 need_flush = true;
 210                                 continue;
 211                         } else if (!err) {
 212                                 split_huge_page_pmd(vma, old_addr, old_pmd);
 213                         }
 214                         VM_BUG_ON(pmd_trans_huge(*old_pmd));
 215                 }
 216                 if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
 217                                                       new_pmd, new_addr))
 218                         break;
 219                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 220                 if (extent > next - new_addr)
 221                         extent = next - new_addr;
 222                 if (extent > LATENCY_LIMIT)
 223                         extent = LATENCY_LIMIT;
 224                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 225                           new_vma, new_pmd, new_addr, need_rmap_locks);
 226                 need_flush = true;
 227         }
 228         if (likely(need_flush))
 229                 flush_tlb_range(vma, old_end-len, old_addr);
 230
 231         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 232
 233         return len + old_addr - old_end;        /* how much done */
 234 }
 235
 236 static unsigned long move_vma(struct vm_area_struct *vma,
 237                 unsigned long old_addr, unsigned long old_len,
 238                 unsigned long new_len, unsigned long new_addr, bool *locked)
 239 {
 240         struct mm_struct *mm = vma->vm_mm;
 241         struct vm_area_struct *new_vma;
 242         unsigned long vm_flags = vma->vm_flags;
 243         unsigned long new_pgoff;
 244         unsigned long moved_len;
 245         unsigned long excess = 0;
 246         unsigned long hiwater_vm;
 247         int split = 0;
 248         int err;
 249         bool need_rmap_locks;
 250
 251         /*
 252          * We'd prefer to avoid failure later on in do_munmap:
 253          * which may split one vma into three before unmapping.
 254          */
 255         if (mm->map_count >= sysctl_max_map_count - 3)
 256                 return -ENOMEM;
 257
 258         /*
 259          * Advise KSM to break any KSM pages in the area to be moved:
 260          * it would be confusing if they were to turn up at the new
 261          * location, where they happen to coincide with different KSM
 262          * pages recently unmapped.  But leave vma->vm_flags as it was,
 263          * so KSM can come around to merge on vma and new_vma afterwards.
 264          */
 265         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 266                                                 MADV_UNMERGEABLE, &vm_flags);
 267         if (err)
 268                 return err;
 269
 270         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 271         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 272                            &need_rmap_locks);
 273         if (!new_vma)
 274                 return -ENOMEM;
 275
 276         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 277                                      need_rmap_locks);
 278         if (moved_len < old_len) {
 279                 /*
 280                  * On error, move entries back from new area to old,
 281                  * which will succeed since page tables still there,
 282                  * and then proceed to unmap new area instead of old.
 283                  */
 284                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 285                                  true);
 286                 vma = new_vma;
 287                 old_len = new_len;
 288                 old_addr = new_addr;
 289                 new_addr = -ENOMEM;
 290         } else {
 291                 if (vma->vm_file && vma->vm_file->f_op->mremap) {
 292                         err = vma->vm_file->f_op->mremap(vma->vm_file, new_vma);
 293                         if (err < 0) {
 294                                 move_page_tables(new_vma, new_addr, vma,
 295                                                  old_addr, moved_len, true);
 296                                 return err;
 297                         }
 298                 }
 299                 arch_remap(mm, old_addr, old_addr + old_len,
 300                            new_addr, new_addr + new_len);
 301         }
 302
 303         /* Conceal VM_ACCOUNT so old reservation is not undone */
 304         if (vm_flags & VM_ACCOUNT) {
 305                 vma->vm_flags &= ~VM_ACCOUNT;
 306                 excess = vma->vm_end - vma->vm_start - old_len;
 307                 if (old_addr > vma->vm_start &&
 308                     old_addr + old_len < vma->vm_end)
 309                         split = 1;
 310         }
 311
 312         /*
 313          * If we failed to move page tables we still do total_vm increment
 314          * since do_munmap() will decrement it by old_len == new_len.
 315          *
 316          * Since total_vm is about to be raised artificially high for a
 317          * moment, we need to restore high watermark afterwards: if stats
 318          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 319          * If this were a serious issue, we'd add a flag to do_munmap().
 320          */
 321         hiwater_vm = mm->hiwater_vm;
 322         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 323
 324         if (do_munmap(mm, old_addr, old_len) < 0) {
 325                 /* OOM: unable to split vma, just get accounts right */
 326                 vm_unacct_memory(excess >> PAGE_SHIFT);
 327                 excess = 0;
 328         }
 329         mm->hiwater_vm = hiwater_vm;
 330
 331         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 332         if (excess) {
 333                 vma->vm_flags |= VM_ACCOUNT;
 334                 if (split)
 335                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 336         }
 337
 338         if (vm_flags & VM_LOCKED) {
 339                 mm->locked_vm += new_len >> PAGE_SHIFT;
 340                 *locked = true;
 341         }
 342
 343         return new_addr;
 344 }
 345
 346 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 347         unsigned long old_len, unsigned long new_len, unsigned long *p)
 348 {
 349         struct mm_struct *mm = current->mm;
 350         struct vm_area_struct *vma = find_vma(mm, addr);
 351
 352         if (!vma || vma->vm_start > addr)
 353                 return ERR_PTR(-EFAULT);
 354
 355         if (is_vm_hugetlb_page(vma))
 356                 return ERR_PTR(-EINVAL);
 357
 358         /* We can't remap across vm area boundaries */
 359         if (old_len > vma->vm_end - addr)
 360                 return ERR_PTR(-EFAULT);
 361
 362         /* Need to be careful about a growing mapping */
 363         if (new_len > old_len) {
 364                 unsigned long pgoff;
 365
 366                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 367                         return ERR_PTR(-EFAULT);
 368                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 369                 pgoff += vma->vm_pgoff;
 370                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 371                         return ERR_PTR(-EINVAL);
 372         }
 373
 374         if (vma->vm_flags & VM_LOCKED) {
 375                 unsigned long locked, lock_limit;
 376                 locked = mm->locked_vm << PAGE_SHIFT;
 377                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 378                 locked += new_len - old_len;
 379                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 380                         return ERR_PTR(-EAGAIN);
 381         }
 382
 383         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
 384                 return ERR_PTR(-ENOMEM);
 385
 386         if (vma->vm_flags & VM_ACCOUNT) {
 387                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 388                 if (security_vm_enough_memory_mm(mm, charged))
 389                         return ERR_PTR(-ENOMEM);
 390                 *p = charged;
 391         }
 392
 393         return vma;
 394 }
 395
 396 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 397                 unsigned long new_addr, unsigned long new_len, bool *locked)
 398 {
 399         struct mm_struct *mm = current->mm;
 400         struct vm_area_struct *vma;
 401         unsigned long ret = -EINVAL;
 402         unsigned long charged = 0;
 403         unsigned long map_flags;
 404
 405         if (new_addr & ~PAGE_MASK)
 406                 goto out;
 407
 408         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 409                 goto out;
 410
 411         /* Check if the location we're moving into overlaps the
 412          * old location at all, and fail if it does.
 413          */
 414         if ((new_addr <= addr) && (new_addr+new_len) > addr)
 415                 goto out;
 416
 417         if ((addr <= new_addr) && (addr+old_len) > new_addr)
 418                 goto out;
 419
 420         ret = do_munmap(mm, new_addr, new_len);
 421         if (ret)
 422                 goto out;
 423
 424         if (old_len >= new_len) {
 425                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 426                 if (ret && old_len != new_len)
 427                         goto out;
 428                 old_len = new_len;
 429         }
 430
 431         vma = vma_to_resize(addr, old_len, new_len, &charged);
 432         if (IS_ERR(vma)) {
 433                 ret = PTR_ERR(vma);
 434                 goto out;
 435         }
 436
 437         map_flags = MAP_FIXED;
 438         if (vma->vm_flags & VM_MAYSHARE)
 439                 map_flags |= MAP_SHARED;
 440
 441         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 442                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 443                                 map_flags);
 444         if (ret & ~PAGE_MASK)
 445                 goto out1;
 446
 447         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
 448         if (!(ret & ~PAGE_MASK))
 449                 goto out;
 450 out1:
 451         vm_unacct_memory(charged);
 452
 453 out:
 454         return ret;
 455 }
 456
 457 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 458 {
 459         unsigned long end = vma->vm_end + delta;
 460         if (end < vma->vm_end) /* overflow */
 461                 return 0;
 462         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 463                 return 0;
 464         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 465                               0, MAP_FIXED) & ~PAGE_MASK)
 466                 return 0;
 467         return 1;
 468 }
 469
 470 /*
 471  * Expand (or shrink) an existing mapping, potentially moving it at the
 472  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 473  *
 474  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 475  * This option implies MREMAP_MAYMOVE.
 476  */
 477 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 478                 unsigned long, new_len, unsigned long, flags,
 479                 unsigned long, new_addr)
 480 {
 481         struct mm_struct *mm = current->mm;
 482         struct vm_area_struct *vma;
 483         unsigned long ret = -EINVAL;
 484         unsigned long charged = 0;
 485         bool locked = false;
 486
 487         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 488                 return ret;
 489
 490         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 491                 return ret;
 492
 493         if (addr & ~PAGE_MASK)
 494                 return ret;
 495
 496         old_len = PAGE_ALIGN(old_len);
 497         new_len = PAGE_ALIGN(new_len);
 498
 499         /*
 500          * We allow a zero old-len as a special case
 501          * for DOS-emu "duplicate shm area" thing. But
 502          * a zero new-len is nonsensical.
 503          */
 504         if (!new_len)
 505                 return ret;
 506
 507         down_write(&current->mm->mmap_sem);
 508
 509         if (flags & MREMAP_FIXED) {
 510                 ret = mremap_to(addr, old_len, new_addr, new_len,
 511                                 &locked);
 512                 goto out;
 513         }
 514
 515         /*
 516          * Always allow a shrinking remap: that just unmaps
 517          * the unnecessary pages..
 518          * do_munmap does all the needed commit accounting
 519          */
 520         if (old_len >= new_len) {
 521                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 522                 if (ret && old_len != new_len)
 523                         goto out;
 524                 ret = addr;
 525                 goto out;
 526         }
 527
 528         /*
 529          * Ok, we need to grow..
 530          */
 531         vma = vma_to_resize(addr, old_len, new_len, &charged);
 532         if (IS_ERR(vma)) {
 533                 ret = PTR_ERR(vma);
 534                 goto out;
 535         }
 536
 537         /* old_len exactly to the end of the area..
 538          */
 539         if (old_len == vma->vm_end - addr) {
 540                 /* can we just expand the current mapping? */
 541                 if (vma_expandable(vma, new_len - old_len)) {
 542                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 543
 544                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 545                                        vma->vm_pgoff, NULL)) {
 546                                 ret = -ENOMEM;
 547                                 goto out;
 548                         }
 549
 550                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
 551                         if (vma->vm_flags & VM_LOCKED) {
 552                                 mm->locked_vm += pages;
 553                                 locked = true;
 554                                 new_addr = addr;
 555                         }
 556                         ret = addr;
 557                         goto out;
 558                 }
 559         }
 560
 561         /*
 562          * We weren't able to just expand or shrink the area,
 563          * we need to create a new one and move it..
 564          */
 565         ret = -ENOMEM;
 566         if (flags & MREMAP_MAYMOVE) {
 567                 unsigned long map_flags = 0;
 568                 if (vma->vm_flags & VM_MAYSHARE)
 569                         map_flags |= MAP_SHARED;
 570
 571                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 572                                         vma->vm_pgoff +
 573                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 574                                         map_flags);
 575                 if (new_addr & ~PAGE_MASK) {
 576                         ret = new_addr;
 577                         goto out;
 578                 }
 579
 580                 ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
 581         }
 582 out:
 583         if (ret & ~PAGE_MASK)
 584                 vm_unacct_memory(charged);
 585         up_write(&current->mm->mmap_sem);
 586         if (locked && new_len > old_len)
 587                 mm_populate(new_addr + old_len, new_len - old_len);
 588         return ret;
 589 }