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