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/slab.h>
  13 #include <linux/shm.h>
  14 #include <linux/ksm.h>
  15 #include <linux/mman.h>
  16 #include <linux/swap.h>
  17 #include <linux/capability.h>
  18 #include <linux/fs.h>
  19 #include <linux/highmem.h>
  20 #include <linux/security.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/mmu_notifier.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_or_clear_bad(pmd))
  46                 return NULL;
  47
  48         return pmd;
  49 }
  50
  51 static pmd_t *alloc_new_pmd(struct mm_struct *mm, 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         if (!pmd_present(*pmd) && __pte_alloc(mm, pmd, addr))
  67                 return NULL;
  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)
  76 {
  77         struct address_space *mapping = 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         unsigned long old_start;
  82
  83         old_start = old_addr;
  84         mmu_notifier_invalidate_range_start(vma->vm_mm,
  85                                             old_start, old_end);
  86         if (vma->vm_file) {
  87                 /*
  88                  * Subtle point from Rajesh Venkatasubramanian: before
  89                  * moving file-based ptes, we must lock truncate_pagecache
  90                  * out, since it might clean the dst vma before the src vma,
  91                  * and we propagate stale pages into the dst afterward.
  92                  */
  93                 mapping = vma->vm_file->f_mapping;
  94                 spin_lock(&mapping->i_mmap_lock);
  95                 new_vma->vm_truncate_count = 0;
  96         }
  97
  98         /*
  99          * We don't have to worry about the ordering of src and dst
 100          * pte locks because exclusive mmap_sem prevents deadlock.
 101          */
 102         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 103         new_pte = pte_offset_map_nested(new_pmd, new_addr);
 104         new_ptl = pte_lockptr(mm, new_pmd);
 105         if (new_ptl != old_ptl)
 106                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 107         arch_enter_lazy_mmu_mode();
 108
 109         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 110                                    new_pte++, new_addr += PAGE_SIZE) {
 111                 if (pte_none(*old_pte))
 112                         continue;
 113                 pte = ptep_clear_flush(vma, old_addr, old_pte);
 114                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 115                 set_pte_at(mm, new_addr, new_pte, pte);
 116         }
 117
 118         arch_leave_lazy_mmu_mode();
 119         if (new_ptl != old_ptl)
 120                 spin_unlock(new_ptl);
 121         pte_unmap_nested(new_pte - 1);
 122         pte_unmap_unlock(old_pte - 1, old_ptl);
 123         if (mapping)
 124                 spin_unlock(&mapping->i_mmap_lock);
 125         mmu_notifier_invalidate_range_end(vma->vm_mm, old_start, old_end);
 126 }
 127
 128 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 129
 130 unsigned long move_page_tables(struct vm_area_struct *vma,
 131                 unsigned long old_addr, struct vm_area_struct *new_vma,
 132                 unsigned long new_addr, unsigned long len)
 133 {
 134         unsigned long extent, next, old_end;
 135         pmd_t *old_pmd, *new_pmd;
 136
 137         old_end = old_addr + len;
 138         flush_cache_range(vma, old_addr, old_end);
 139
 140         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 141                 cond_resched();
 142                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 143                 if (next - 1 > old_end)
 144                         next = old_end;
 145                 extent = next - old_addr;
 146                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 147                 if (!old_pmd)
 148                         continue;
 149                 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
 150                 if (!new_pmd)
 151                         break;
 152                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 153                 if (extent > next - new_addr)
 154                         extent = next - new_addr;
 155                 if (extent > LATENCY_LIMIT)
 156                         extent = LATENCY_LIMIT;
 157                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 158                                 new_vma, new_pmd, new_addr);
 159         }
 160
 161         return len + old_addr - old_end;        /* how much done */
 162 }
 163
 164 static unsigned long move_vma(struct vm_area_struct *vma,
 165                 unsigned long old_addr, unsigned long old_len,
 166                 unsigned long new_len, unsigned long new_addr)
 167 {
 168         struct mm_struct *mm = vma->vm_mm;
 169         struct vm_area_struct *new_vma;
 170         unsigned long vm_flags = vma->vm_flags;
 171         unsigned long new_pgoff;
 172         unsigned long moved_len;
 173         unsigned long excess = 0;
 174         unsigned long hiwater_vm;
 175         int split = 0;
 176         int err;
 177
 178         /*
 179          * We'd prefer to avoid failure later on in do_munmap:
 180          * which may split one vma into three before unmapping.
 181          */
 182         if (mm->map_count >= sysctl_max_map_count - 3)
 183                 return -ENOMEM;
 184
 185         /*
 186          * Advise KSM to break any KSM pages in the area to be moved:
 187          * it would be confusing if they were to turn up at the new
 188          * location, where they happen to coincide with different KSM
 189          * pages recently unmapped.  But leave vma->vm_flags as it was,
 190          * so KSM can come around to merge on vma and new_vma afterwards.
 191          */
 192         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 193                                                 MADV_UNMERGEABLE, &vm_flags);
 194         if (err)
 195                 return err;
 196
 197         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 198         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 199         if (!new_vma)
 200                 return -ENOMEM;
 201
 202         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
 203         if (moved_len < old_len) {
 204                 /*
 205                  * On error, move entries back from new area to old,
 206                  * which will succeed since page tables still there,
 207                  * and then proceed to unmap new area instead of old.
 208                  */
 209                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
 210                 vma = new_vma;
 211                 old_len = new_len;
 212                 old_addr = new_addr;
 213                 new_addr = -ENOMEM;
 214         }
 215
 216         /* Conceal VM_ACCOUNT so old reservation is not undone */
 217         if (vm_flags & VM_ACCOUNT) {
 218                 vma->vm_flags &= ~VM_ACCOUNT;
 219                 excess = vma->vm_end - vma->vm_start - old_len;
 220                 if (old_addr > vma->vm_start &&
 221                     old_addr + old_len < vma->vm_end)
 222                         split = 1;
 223         }
 224
 225         /*
 226          * If we failed to move page tables we still do total_vm increment
 227          * since do_munmap() will decrement it by old_len == new_len.
 228          *
 229          * Since total_vm is about to be raised artificially high for a
 230          * moment, we need to restore high watermark afterwards: if stats
 231          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 232          * If this were a serious issue, we'd add a flag to do_munmap().
 233          */
 234         hiwater_vm = mm->hiwater_vm;
 235         mm->total_vm += new_len >> PAGE_SHIFT;
 236         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 237
 238         if (do_munmap(mm, old_addr, old_len) < 0) {
 239                 /* OOM: unable to split vma, just get accounts right */
 240                 vm_unacct_memory(excess >> PAGE_SHIFT);
 241                 excess = 0;
 242         }
 243         mm->hiwater_vm = hiwater_vm;
 244
 245         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 246         if (excess) {
 247                 vma->vm_flags |= VM_ACCOUNT;
 248                 if (split)
 249                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 250         }
 251
 252         if (vm_flags & VM_LOCKED) {
 253                 mm->locked_vm += new_len >> PAGE_SHIFT;
 254                 if (new_len > old_len)
 255                         mlock_vma_pages_range(new_vma, new_addr + old_len,
 256                                                        new_addr + new_len);
 257         }
 258
 259         return new_addr;
 260 }
 261
 262 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 263         unsigned long old_len, unsigned long new_len, unsigned long *p)
 264 {
 265         struct mm_struct *mm = current->mm;
 266         struct vm_area_struct *vma = find_vma(mm, addr);
 267
 268         if (!vma || vma->vm_start > addr)
 269                 goto Efault;
 270
 271         if (is_vm_hugetlb_page(vma))
 272                 goto Einval;
 273
 274         /* We can't remap across vm area boundaries */
 275         if (old_len > vma->vm_end - addr)
 276                 goto Efault;
 277
 278         /* Need to be careful about a growing mapping */
 279         if (new_len > old_len) {
 280                 unsigned long pgoff;
 281
 282                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 283                         goto Efault;
 284                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 285                 pgoff += vma->vm_pgoff;
 286                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 287                         goto Einval;
 288         }
 289
 290         if (vma->vm_flags & VM_LOCKED) {
 291                 unsigned long locked, lock_limit;
 292                 locked = mm->locked_vm << PAGE_SHIFT;
 293                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 294                 locked += new_len - old_len;
 295                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 296                         goto Eagain;
 297         }
 298
 299         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
 300                 goto Enomem;
 301
 302         if (vma->vm_flags & VM_ACCOUNT) {
 303                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 304                 if (security_vm_enough_memory(charged))
 305                         goto Efault;
 306                 *p = charged;
 307         }
 308
 309         return vma;
 310
 311 Efault: /* very odd choice for most of the cases, but... */
 312         return ERR_PTR(-EFAULT);
 313 Einval:
 314         return ERR_PTR(-EINVAL);
 315 Enomem:
 316         return ERR_PTR(-ENOMEM);
 317 Eagain:
 318         return ERR_PTR(-EAGAIN);
 319 }
 320
 321 static unsigned long mremap_to(unsigned long addr,
 322         unsigned long old_len, unsigned long new_addr,
 323         unsigned long new_len)
 324 {
 325         struct mm_struct *mm = current->mm;
 326         struct vm_area_struct *vma;
 327         unsigned long ret = -EINVAL;
 328         unsigned long charged = 0;
 329         unsigned long map_flags;
 330
 331         if (new_addr & ~PAGE_MASK)
 332                 goto out;
 333
 334         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 335                 goto out;
 336
 337         /* Check if the location we're moving into overlaps the
 338          * old location at all, and fail if it does.
 339          */
 340         if ((new_addr <= addr) && (new_addr+new_len) > addr)
 341                 goto out;
 342
 343         if ((addr <= new_addr) && (addr+old_len) > new_addr)
 344                 goto out;
 345
 346         ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
 347         if (ret)
 348                 goto out;
 349
 350         ret = do_munmap(mm, new_addr, new_len);
 351         if (ret)
 352                 goto out;
 353
 354         if (old_len >= new_len) {
 355                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 356                 if (ret && old_len != new_len)
 357                         goto out;
 358                 old_len = new_len;
 359         }
 360
 361         vma = vma_to_resize(addr, old_len, new_len, &charged);
 362         if (IS_ERR(vma)) {
 363                 ret = PTR_ERR(vma);
 364                 goto out;
 365         }
 366
 367         map_flags = MAP_FIXED;
 368         if (vma->vm_flags & VM_MAYSHARE)
 369                 map_flags |= MAP_SHARED;
 370
 371         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 372                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 373                                 map_flags);
 374         if (ret & ~PAGE_MASK)
 375                 goto out1;
 376
 377         ret = move_vma(vma, addr, old_len, new_len, new_addr);
 378         if (!(ret & ~PAGE_MASK))
 379                 goto out;
 380 out1:
 381         vm_unacct_memory(charged);
 382
 383 out:
 384         return ret;
 385 }
 386
 387 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 388 {
 389         unsigned long end = vma->vm_end + delta;
 390         if (end < vma->vm_end) /* overflow */
 391                 return 0;
 392         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 393                 return 0;
 394         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 395                               0, MAP_FIXED) & ~PAGE_MASK)
 396                 return 0;
 397         return 1;
 398 }
 399
 400 /*
 401  * Expand (or shrink) an existing mapping, potentially moving it at the
 402  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 403  *
 404  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 405  * This option implies MREMAP_MAYMOVE.
 406  */
 407 unsigned long do_mremap(unsigned long addr,
 408         unsigned long old_len, unsigned long new_len,
 409         unsigned long flags, unsigned long new_addr)
 410 {
 411         struct mm_struct *mm = current->mm;
 412         struct vm_area_struct *vma;
 413         unsigned long ret = -EINVAL;
 414         unsigned long charged = 0;
 415
 416         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 417                 goto out;
 418
 419         if (addr & ~PAGE_MASK)
 420                 goto out;
 421
 422         old_len = PAGE_ALIGN(old_len);
 423         new_len = PAGE_ALIGN(new_len);
 424
 425         /*
 426          * We allow a zero old-len as a special case
 427          * for DOS-emu "duplicate shm area" thing. But
 428          * a zero new-len is nonsensical.
 429          */
 430         if (!new_len)
 431                 goto out;
 432
 433         if (flags & MREMAP_FIXED) {
 434                 if (flags & MREMAP_MAYMOVE)
 435                         ret = mremap_to(addr, old_len, new_addr, new_len);
 436                 goto out;
 437         }
 438
 439         /*
 440          * Always allow a shrinking remap: that just unmaps
 441          * the unnecessary pages..
 442          * do_munmap does all the needed commit accounting
 443          */
 444         if (old_len >= new_len) {
 445                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 446                 if (ret && old_len != new_len)
 447                         goto out;
 448                 ret = addr;
 449                 goto out;
 450         }
 451
 452         /*
 453          * Ok, we need to grow..
 454          */
 455         vma = vma_to_resize(addr, old_len, new_len, &charged);
 456         if (IS_ERR(vma)) {
 457                 ret = PTR_ERR(vma);
 458                 goto out;
 459         }
 460
 461         /* old_len exactly to the end of the area..
 462          */
 463         if (old_len == vma->vm_end - addr) {
 464                 /* can we just expand the current mapping? */
 465                 if (vma_expandable(vma, new_len - old_len)) {
 466                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 467
 468                         vma_adjust(vma, vma->vm_start,
 469                                 addr + new_len, vma->vm_pgoff, NULL);
 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 }