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, unsigned long addr)
  51 {
  52         pgd_t *pgd;
  53         pud_t *pud;
  54         pmd_t *pmd;
  55
  56         pgd = pgd_offset(mm, addr);
  57         pud = pud_alloc(mm, pgd, addr);
  58         if (!pud)
  59                 return NULL;
  60
  61         pmd = pmd_alloc(mm, pud, addr);
  62         if (!pmd)
  63                 return NULL;
  64
  65         if (!pmd_present(*pmd) && __pte_alloc(mm, pmd, addr))
  66                 return NULL;
  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 mm_struct *mm = vma->vm_mm;
  78         pte_t *old_pte, *new_pte, pte;
  79         spinlock_t *old_ptl, *new_ptl;
  80         unsigned long old_start;
  81
  82         old_start = old_addr;
  83         mmu_notifier_invalidate_range_start(vma->vm_mm,
  84                                             old_start, old_end);
  85         if (vma->vm_file) {
  86                 /*
  87                  * Subtle point from Rajesh Venkatasubramanian: before
  88                  * moving file-based ptes, we must lock truncate_pagecache
  89                  * out, since it might clean the dst vma before the src vma,
  90                  * and we propagate stale pages into the dst afterward.
  91                  */
  92                 mapping = vma->vm_file->f_mapping;
  93                 spin_lock(&mapping->i_mmap_lock);
  94                 new_vma->vm_truncate_count = 0;
  95         }
  96
  97         /*
  98          * We don't have to worry about the ordering of src and dst
  99          * pte locks because exclusive mmap_sem prevents deadlock.
 100          */
 101         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 102         new_pte = pte_offset_map_nested(new_pmd, new_addr);
 103         new_ptl = pte_lockptr(mm, new_pmd);
 104         if (new_ptl != old_ptl)
 105                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 106         arch_enter_lazy_mmu_mode();
 107
 108         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 109                                    new_pte++, new_addr += PAGE_SIZE) {
 110                 if (pte_none(*old_pte))
 111                         continue;
 112                 pte = ptep_clear_flush(vma, old_addr, old_pte);
 113                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 114                 set_pte_at(mm, new_addr, new_pte, pte);
 115         }
 116
 117         arch_leave_lazy_mmu_mode();
 118         if (new_ptl != old_ptl)
 119                 spin_unlock(new_ptl);
 120         pte_unmap_nested(new_pte - 1);
 121         pte_unmap_unlock(old_pte - 1, old_ptl);
 122         if (mapping)
 123                 spin_unlock(&mapping->i_mmap_lock);
 124         mmu_notifier_invalidate_range_end(vma->vm_mm, old_start, old_end);
 125 }
 126
 127 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 128
 129 unsigned long move_page_tables(struct vm_area_struct *vma,
 130                 unsigned long old_addr, struct vm_area_struct *new_vma,
 131                 unsigned long new_addr, unsigned long len)
 132 {
 133         unsigned long extent, next, old_end;
 134         pmd_t *old_pmd, *new_pmd;
 135
 136         old_end = old_addr + len;
 137         flush_cache_range(vma, old_addr, old_end);
 138
 139         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 140                 cond_resched();
 141                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 142                 if (next - 1 > old_end)
 143                         next = old_end;
 144                 extent = next - old_addr;
 145                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 146                 if (!old_pmd)
 147                         continue;
 148                 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
 149                 if (!new_pmd)
 150                         break;
 151                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 152                 if (extent > next - new_addr)
 153                         extent = next - new_addr;
 154                 if (extent > LATENCY_LIMIT)
 155                         extent = LATENCY_LIMIT;
 156                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 157                                 new_vma, new_pmd, new_addr);
 158         }
 159
 160         return len + old_addr - old_end;        /* how much done */
 161 }
 162
 163 static unsigned long move_vma(struct vm_area_struct *vma,
 164                 unsigned long old_addr, unsigned long old_len,
 165                 unsigned long new_len, unsigned long new_addr)
 166 {
 167         struct mm_struct *mm = vma->vm_mm;
 168         struct vm_area_struct *new_vma;
 169         unsigned long vm_flags = vma->vm_flags;
 170         unsigned long new_pgoff;
 171         unsigned long moved_len;
 172         unsigned long excess = 0;
 173         unsigned long hiwater_vm;
 174         int split = 0;
 175         int err;
 176
 177         /*
 178          * We'd prefer to avoid failure later on in do_munmap:
 179          * which may split one vma into three before unmapping.
 180          */
 181         if (mm->map_count >= sysctl_max_map_count - 3)
 182                 return -ENOMEM;
 183
 184         /*
 185          * Advise KSM to break any KSM pages in the area to be moved:
 186          * it would be confusing if they were to turn up at the new
 187          * location, where they happen to coincide with different KSM
 188          * pages recently unmapped.  But leave vma->vm_flags as it was,
 189          * so KSM can come around to merge on vma and new_vma afterwards.
 190          */
 191         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 192                                                 MADV_UNMERGEABLE, &vm_flags);
 193         if (err)
 194                 return err;
 195
 196         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 197         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 198         if (!new_vma)
 199                 return -ENOMEM;
 200
 201         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
 202         if (moved_len < old_len) {
 203                 /*
 204                  * On error, move entries back from new area to old,
 205                  * which will succeed since page tables still there,
 206                  * and then proceed to unmap new area instead of old.
 207                  */
 208                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
 209                 vma = new_vma;
 210                 old_len = new_len;
 211                 old_addr = new_addr;
 212                 new_addr = -ENOMEM;
 213         }
 214
 215         /* Conceal VM_ACCOUNT so old reservation is not undone */
 216         if (vm_flags & VM_ACCOUNT) {
 217                 vma->vm_flags &= ~VM_ACCOUNT;
 218                 excess = vma->vm_end - vma->vm_start - old_len;
 219                 if (old_addr > vma->vm_start &&
 220                     old_addr + old_len < vma->vm_end)
 221                         split = 1;
 222         }
 223
 224         /*
 225          * If we failed to move page tables we still do total_vm increment
 226          * since do_munmap() will decrement it by old_len == new_len.
 227          *
 228          * Since total_vm is about to be raised artificially high for a
 229          * moment, we need to restore high watermark afterwards: if stats
 230          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 231          * If this were a serious issue, we'd add a flag to do_munmap().
 232          */
 233         hiwater_vm = mm->hiwater_vm;
 234         mm->total_vm += new_len >> PAGE_SHIFT;
 235         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 236
 237         if (do_munmap(mm, old_addr, old_len) < 0) {
 238                 /* OOM: unable to split vma, just get accounts right */
 239                 vm_unacct_memory(excess >> PAGE_SHIFT);
 240                 excess = 0;
 241         }
 242         mm->hiwater_vm = hiwater_vm;
 243
 244         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 245         if (excess) {
 246                 vma->vm_flags |= VM_ACCOUNT;
 247                 if (split)
 248                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 249         }
 250
 251         if (vm_flags & VM_LOCKED) {
 252                 mm->locked_vm += new_len >> PAGE_SHIFT;
 253                 if (new_len > old_len)
 254                         mlock_vma_pages_range(new_vma, new_addr + old_len,
 255                                                        new_addr + new_len);
 256         }
 257
 258         return new_addr;
 259 }
 260
 261 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 262         unsigned long old_len, unsigned long new_len, unsigned long *p)
 263 {
 264         struct mm_struct *mm = current->mm;
 265         struct vm_area_struct *vma = find_vma(mm, addr);
 266
 267         if (!vma || vma->vm_start > addr)
 268                 goto Efault;
 269
 270         if (is_vm_hugetlb_page(vma))
 271                 goto Einval;
 272
 273         /* We can't remap across vm area boundaries */
 274         if (old_len > vma->vm_end - addr)
 275                 goto Efault;
 276
 277         /* Need to be careful about a growing mapping */
 278         if (new_len > old_len) {
 279                 unsigned long pgoff;
 280
 281                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 282                         goto Efault;
 283                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 284                 pgoff += vma->vm_pgoff;
 285                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 286                         goto Einval;
 287         }
 288
 289         if (vma->vm_flags & VM_LOCKED) {
 290                 unsigned long locked, lock_limit;
 291                 locked = mm->locked_vm << PAGE_SHIFT;
 292                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 293                 locked += new_len - old_len;
 294                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 295                         goto Eagain;
 296         }
 297
 298         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
 299                 goto Enomem;
 300
 301         if (vma->vm_flags & VM_ACCOUNT) {
 302                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 303                 if (security_vm_enough_memory(charged))
 304                         goto Efault;
 305                 *p = charged;
 306         }
 307
 308         return vma;
 309
 310 Efault: /* very odd choice for most of the cases, but... */
 311         return ERR_PTR(-EFAULT);
 312 Einval:
 313         return ERR_PTR(-EINVAL);
 314 Enomem:
 315         return ERR_PTR(-ENOMEM);
 316 Eagain:
 317         return ERR_PTR(-EAGAIN);
 318 }
 319
 320 static unsigned long mremap_to(unsigned long addr,
 321         unsigned long old_len, unsigned long new_addr,
 322         unsigned long new_len)
 323 {
 324         struct mm_struct *mm = current->mm;
 325         struct vm_area_struct *vma;
 326         unsigned long ret = -EINVAL;
 327         unsigned long charged = 0;
 328         unsigned long map_flags;
 329
 330         if (new_addr & ~PAGE_MASK)
 331                 goto out;
 332
 333         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 334                 goto out;
 335
 336         /* Check if the location we're moving into overlaps the
 337          * old location at all, and fail if it does.
 338          */
 339         if ((new_addr <= addr) && (new_addr+new_len) > addr)
 340                 goto out;
 341
 342         if ((addr <= new_addr) && (addr+old_len) > new_addr)
 343                 goto out;
 344
 345         ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
 346         if (ret)
 347                 goto out;
 348
 349         ret = do_munmap(mm, new_addr, new_len);
 350         if (ret)
 351                 goto out;
 352
 353         if (old_len >= new_len) {
 354                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 355                 if (ret && old_len != new_len)
 356                         goto out;
 357                 old_len = new_len;
 358         }
 359
 360         vma = vma_to_resize(addr, old_len, new_len, &charged);
 361         if (IS_ERR(vma)) {
 362                 ret = PTR_ERR(vma);
 363                 goto out;
 364         }
 365
 366         map_flags = MAP_FIXED;
 367         if (vma->vm_flags & VM_MAYSHARE)
 368                 map_flags |= MAP_SHARED;
 369
 370         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 371                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 372                                 map_flags);
 373         if (ret & ~PAGE_MASK)
 374                 goto out1;
 375
 376         ret = move_vma(vma, addr, old_len, new_len, new_addr);
 377         if (!(ret & ~PAGE_MASK))
 378                 goto out;
 379 out1:
 380         vm_unacct_memory(charged);
 381
 382 out:
 383         return ret;
 384 }
 385
 386 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 387 {
 388         unsigned long end = vma->vm_end + delta;
 389         if (end < vma->vm_end) /* overflow */
 390                 return 0;
 391         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 392                 return 0;
 393         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 394                               0, MAP_FIXED) & ~PAGE_MASK)
 395                 return 0;
 396         return 1;
 397 }
 398
 399 /*
 400  * Expand (or shrink) an existing mapping, potentially moving it at the
 401  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 402  *
 403  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 404  * This option implies MREMAP_MAYMOVE.
 405  */
 406 unsigned long do_mremap(unsigned long addr,
 407         unsigned long old_len, unsigned long new_len,
 408         unsigned long flags, unsigned long new_addr)
 409 {
 410         struct mm_struct *mm = current->mm;
 411         struct vm_area_struct *vma;
 412         unsigned long ret = -EINVAL;
 413         unsigned long charged = 0;
 414
 415         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 416                 goto out;
 417
 418         if (addr & ~PAGE_MASK)
 419                 goto out;
 420
 421         old_len = PAGE_ALIGN(old_len);
 422         new_len = PAGE_ALIGN(new_len);
 423
 424         /*
 425          * We allow a zero old-len as a special case
 426          * for DOS-emu "duplicate shm area" thing. But
 427          * a zero new-len is nonsensical.
 428          */
 429         if (!new_len)
 430                 goto out;
 431
 432         if (flags & MREMAP_FIXED) {
 433                 if (flags & MREMAP_MAYMOVE)
 434                         ret = mremap_to(addr, old_len, new_addr, new_len);
 435                 goto out;
 436         }
 437
 438         /*
 439          * Always allow a shrinking remap: that just unmaps
 440          * the unnecessary pages..
 441          * do_munmap does all the needed commit accounting
 442          */
 443         if (old_len >= new_len) {
 444                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 445                 if (ret && old_len != new_len)
 446                         goto out;
 447                 ret = addr;
 448                 goto out;
 449         }
 450
 451         /*
 452          * Ok, we need to grow..
 453          */
 454         vma = vma_to_resize(addr, old_len, new_len, &charged);
 455         if (IS_ERR(vma)) {
 456                 ret = PTR_ERR(vma);
 457                 goto out;
 458         }
 459
 460         /* old_len exactly to the end of the area..
 461          */
 462         if (old_len == vma->vm_end - addr) {
 463                 /* can we just expand the current mapping? */
 464                 if (vma_expandable(vma, new_len - old_len)) {
 465                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 466
 467                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 468                                        vma->vm_pgoff, NULL)) {
 469                                 ret = -ENOMEM;
 470                                 goto out;
 471                         }
 472
 473                         mm->total_vm += pages;
 474                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
 475                         if (vma->vm_flags & VM_LOCKED) {
 476                                 mm->locked_vm += pages;
 477                                 mlock_vma_pages_range(vma, addr + old_len,
 478                                                    addr + new_len);
 479                         }
 480                         ret = addr;
 481                         goto out;
 482                 }
 483         }
 484
 485         /*
 486          * We weren't able to just expand or shrink the area,
 487          * we need to create a new one and move it..
 488          */
 489         ret = -ENOMEM;
 490         if (flags & MREMAP_MAYMOVE) {
 491                 unsigned long map_flags = 0;
 492                 if (vma->vm_flags & VM_MAYSHARE)
 493                         map_flags |= MAP_SHARED;
 494
 495                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 496                                         vma->vm_pgoff +
 497                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 498                                         map_flags);
 499                 if (new_addr & ~PAGE_MASK) {
 500                         ret = new_addr;
 501                         goto out;
 502                 }
 503
 504                 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
 505                 if (ret)
 506                         goto out;
 507                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 508         }
 509 out:
 510         if (ret & ~PAGE_MASK)
 511                 vm_unacct_memory(charged);
 512         return ret;
 513 }
 514
 515 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 516                 unsigned long, new_len, unsigned long, flags,
 517                 unsigned long, new_addr)
 518 {
 519         unsigned long ret;
 520
 521         down_write(&current->mm->mmap_sem);
 522         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 523         up_write(&current->mm->mmap_sem);
 524         return ret;
 525 }