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