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