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