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                                 err = move_huge_pmd(vma, new_vma, old_addr,
 180                                                     new_addr, old_end,
 181                                                     old_pmd, new_pmd);
 182                         if (err > 0) {
 183                                 need_flush = true;
 184                                 continue;
 185                         } else if (!err) {
 186                                 split_huge_page_pmd(vma, old_addr, old_pmd);
 187                         }
 188                         VM_BUG_ON(pmd_trans_huge(*old_pmd));
 189                 }
 190                 if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
 191                                                       new_pmd, new_addr))
 192                         break;
 193                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 194                 if (extent > next - new_addr)
 195                         extent = next - new_addr;
 196                 if (extent > LATENCY_LIMIT)
 197                         extent = LATENCY_LIMIT;
 198                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 199                           new_vma, new_pmd, new_addr, need_rmap_locks);
 200                 need_flush = true;
 201         }
 202         if (likely(need_flush))
 203                 flush_tlb_range(vma, old_end-len, old_addr);
 204
 205         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 206
 207         return len + old_addr - old_end;        /* how much done */
 208 }
 209
 210 static unsigned long move_vma(struct vm_area_struct *vma,
 211                 unsigned long old_addr, unsigned long old_len,
 212                 unsigned long new_len, unsigned long new_addr, bool *locked)
 213 {
 214         struct mm_struct *mm = vma->vm_mm;
 215         struct vm_area_struct *new_vma;
 216         unsigned long vm_flags = vma->vm_flags;
 217         unsigned long new_pgoff;
 218         unsigned long moved_len;
 219         unsigned long excess = 0;
 220         unsigned long hiwater_vm;
 221         int split = 0;
 222         int err;
 223         bool need_rmap_locks;
 224
 225         /*
 226          * We'd prefer to avoid failure later on in do_munmap:
 227          * which may split one vma into three before unmapping.
 228          */
 229         if (mm->map_count >= sysctl_max_map_count - 3)
 230                 return -ENOMEM;
 231
 232         /*
 233          * Advise KSM to break any KSM pages in the area to be moved:
 234          * it would be confusing if they were to turn up at the new
 235          * location, where they happen to coincide with different KSM
 236          * pages recently unmapped.  But leave vma->vm_flags as it was,
 237          * so KSM can come around to merge on vma and new_vma afterwards.
 238          */
 239         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 240                                                 MADV_UNMERGEABLE, &vm_flags);
 241         if (err)
 242                 return err;
 243
 244         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 245         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 246                            &need_rmap_locks);
 247         if (!new_vma)
 248                 return -ENOMEM;
 249
 250         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 251                                      need_rmap_locks);
 252         if (moved_len < old_len) {
 253                 /*
 254                  * On error, move entries back from new area to old,
 255                  * which will succeed since page tables still there,
 256                  * and then proceed to unmap new area instead of old.
 257                  */
 258                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 259                                  true);
 260                 vma = new_vma;
 261                 old_len = new_len;
 262                 old_addr = new_addr;
 263                 new_addr = -ENOMEM;
 264         }
 265
 266         /* Conceal VM_ACCOUNT so old reservation is not undone */
 267         if (vm_flags & VM_ACCOUNT) {
 268                 vma->vm_flags &= ~VM_ACCOUNT;
 269                 excess = vma->vm_end - vma->vm_start - old_len;
 270                 if (old_addr > vma->vm_start &&
 271                     old_addr + old_len < vma->vm_end)
 272                         split = 1;
 273         }
 274
 275         /*
 276          * If we failed to move page tables we still do total_vm increment
 277          * since do_munmap() will decrement it by old_len == new_len.
 278          *
 279          * Since total_vm is about to be raised artificially high for a
 280          * moment, we need to restore high watermark afterwards: if stats
 281          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 282          * If this were a serious issue, we'd add a flag to do_munmap().
 283          */
 284         hiwater_vm = mm->hiwater_vm;
 285         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 286
 287         if (do_munmap(mm, old_addr, old_len) < 0) {
 288                 /* OOM: unable to split vma, just get accounts right */
 289                 vm_unacct_memory(excess >> PAGE_SHIFT);
 290                 excess = 0;
 291         }
 292         mm->hiwater_vm = hiwater_vm;
 293
 294         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 295         if (excess) {
 296                 vma->vm_flags |= VM_ACCOUNT;
 297                 if (split)
 298                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 299         }
 300
 301         if (vm_flags & VM_LOCKED) {
 302                 mm->locked_vm += new_len >> PAGE_SHIFT;
 303                 *locked = true;
 304         }
 305
 306         return new_addr;
 307 }
 308
 309 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 310         unsigned long old_len, unsigned long new_len, unsigned long *p)
 311 {
 312         struct mm_struct *mm = current->mm;
 313         struct vm_area_struct *vma = find_vma(mm, addr);
 314
 315         if (!vma || vma->vm_start > addr)
 316                 goto Efault;
 317
 318         if (is_vm_hugetlb_page(vma))
 319                 goto Einval;
 320
 321         /* We can't remap across vm area boundaries */
 322         if (old_len > vma->vm_end - addr)
 323                 goto Efault;
 324
 325         /* Need to be careful about a growing mapping */
 326         if (new_len > old_len) {
 327                 unsigned long pgoff;
 328
 329                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 330                         goto Efault;
 331                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 332                 pgoff += vma->vm_pgoff;
 333                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 334                         goto Einval;
 335         }
 336
 337         if (vma->vm_flags & VM_LOCKED) {
 338                 unsigned long locked, lock_limit;
 339                 locked = mm->locked_vm << PAGE_SHIFT;
 340                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 341                 locked += new_len - old_len;
 342                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 343                         goto Eagain;
 344         }
 345
 346         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
 347                 goto Enomem;
 348
 349         if (vma->vm_flags & VM_ACCOUNT) {
 350                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 351                 if (security_vm_enough_memory_mm(mm, charged))
 352                         goto Efault;
 353                 *p = charged;
 354         }
 355
 356         return vma;
 357
 358 Efault: /* very odd choice for most of the cases, but... */
 359         return ERR_PTR(-EFAULT);
 360 Einval:
 361         return ERR_PTR(-EINVAL);
 362 Enomem:
 363         return ERR_PTR(-ENOMEM);
 364 Eagain:
 365         return ERR_PTR(-EAGAIN);
 366 }
 367
 368 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 369                 unsigned long new_addr, unsigned long new_len, bool *locked)
 370 {
 371         struct mm_struct *mm = current->mm;
 372         struct vm_area_struct *vma;
 373         unsigned long ret = -EINVAL;
 374         unsigned long charged = 0;
 375         unsigned long map_flags;
 376
 377         if (new_addr & ~PAGE_MASK)
 378                 goto out;
 379
 380         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 381                 goto out;
 382
 383         /* Check if the location we're moving into overlaps the
 384          * old location at all, and fail if it does.
 385          */
 386         if ((new_addr <= addr) && (new_addr+new_len) > addr)
 387                 goto out;
 388
 389         if ((addr <= new_addr) && (addr+old_len) > new_addr)
 390                 goto out;
 391
 392         ret = do_munmap(mm, new_addr, new_len);
 393         if (ret)
 394                 goto out;
 395
 396         if (old_len >= new_len) {
 397                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 398                 if (ret && old_len != new_len)
 399                         goto out;
 400                 old_len = new_len;
 401         }
 402
 403         vma = vma_to_resize(addr, old_len, new_len, &charged);
 404         if (IS_ERR(vma)) {
 405                 ret = PTR_ERR(vma);
 406                 goto out;
 407         }
 408
 409         map_flags = MAP_FIXED;
 410         if (vma->vm_flags & VM_MAYSHARE)
 411                 map_flags |= MAP_SHARED;
 412
 413         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 414                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 415                                 map_flags);
 416         if (ret & ~PAGE_MASK)
 417                 goto out1;
 418
 419         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
 420         if (!(ret & ~PAGE_MASK))
 421                 goto out;
 422 out1:
 423         vm_unacct_memory(charged);
 424
 425 out:
 426         return ret;
 427 }
 428
 429 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 430 {
 431         unsigned long end = vma->vm_end + delta;
 432         if (end < vma->vm_end) /* overflow */
 433                 return 0;
 434         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 435                 return 0;
 436         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 437                               0, MAP_FIXED) & ~PAGE_MASK)
 438                 return 0;
 439         return 1;
 440 }
 441
 442 /*
 443  * Expand (or shrink) an existing mapping, potentially moving it at the
 444  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 445  *
 446  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 447  * This option implies MREMAP_MAYMOVE.
 448  */
 449 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 450                 unsigned long, new_len, unsigned long, flags,
 451                 unsigned long, new_addr)
 452 {
 453         struct mm_struct *mm = current->mm;
 454         struct vm_area_struct *vma;
 455         unsigned long ret = -EINVAL;
 456         unsigned long charged = 0;
 457         bool locked = false;
 458
 459         down_write(&current->mm->mmap_sem);
 460
 461         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 462                 goto out;
 463
 464         if (addr & ~PAGE_MASK)
 465                 goto out;
 466
 467         old_len = PAGE_ALIGN(old_len);
 468         new_len = PAGE_ALIGN(new_len);
 469
 470         /*
 471          * We allow a zero old-len as a special case
 472          * for DOS-emu "duplicate shm area" thing. But
 473          * a zero new-len is nonsensical.
 474          */
 475         if (!new_len)
 476                 goto out;
 477
 478         if (flags & MREMAP_FIXED) {
 479                 if (flags & MREMAP_MAYMOVE)
 480                         ret = mremap_to(addr, old_len, new_addr, new_len,
 481                                         &locked);
 482                 goto out;
 483         }
 484
 485         /*
 486          * Always allow a shrinking remap: that just unmaps
 487          * the unnecessary pages..
 488          * do_munmap does all the needed commit accounting
 489          */
 490         if (old_len >= new_len) {
 491                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 492                 if (ret && old_len != new_len)
 493                         goto out;
 494                 ret = addr;
 495                 goto out;
 496         }
 497
 498         /*
 499          * Ok, we need to grow..
 500          */
 501         vma = vma_to_resize(addr, old_len, new_len, &charged);
 502         if (IS_ERR(vma)) {
 503                 ret = PTR_ERR(vma);
 504                 goto out;
 505         }
 506
 507         /* old_len exactly to the end of the area..
 508          */
 509         if (old_len == vma->vm_end - addr) {
 510                 /* can we just expand the current mapping? */
 511                 if (vma_expandable(vma, new_len - old_len)) {
 512                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 513
 514                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 515                                        vma->vm_pgoff, NULL)) {
 516                                 ret = -ENOMEM;
 517                                 goto out;
 518                         }
 519
 520                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
 521                         if (vma->vm_flags & VM_LOCKED) {
 522                                 mm->locked_vm += pages;
 523                                 locked = true;
 524                                 new_addr = addr;
 525                         }
 526                         ret = addr;
 527                         goto out;
 528                 }
 529         }
 530
 531         /*
 532          * We weren't able to just expand or shrink the area,
 533          * we need to create a new one and move it..
 534          */
 535         ret = -ENOMEM;
 536         if (flags & MREMAP_MAYMOVE) {
 537                 unsigned long map_flags = 0;
 538                 if (vma->vm_flags & VM_MAYSHARE)
 539                         map_flags |= MAP_SHARED;
 540
 541                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 542                                         vma->vm_pgoff +
 543                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 544                                         map_flags);
 545                 if (new_addr & ~PAGE_MASK) {
 546                         ret = new_addr;
 547                         goto out;
 548                 }
 549
 550                 ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
 551         }
 552 out:
 553         if (ret & ~PAGE_MASK)
 554                 vm_unacct_memory(charged);
 555         up_write(&current->mm->mmap_sem);
 556         if (locked && new_len > old_len)
 557                 mm_populate(new_addr + old_len, new_len - old_len);
 558         return ret;
 559 }