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