mm/mremap.c

   1 /*
   2  *      mm/mremap.c
   3  *
   4  *      (C) Copyright 1996 Linus Torvalds
   5  *
   6  *      Address space accounting code   <alan@redhat.com>
   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/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 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
  28 {
  29         pgd_t *pgd;
  30         pud_t *pud;
  31         pmd_t *pmd;
  32
  33         pgd = pgd_offset(mm, addr);
  34         if (pgd_none_or_clear_bad(pgd))
  35                 return NULL;
  36
  37         pud = pud_offset(pgd, addr);
  38         if (pud_none_or_clear_bad(pud))
  39                 return NULL;
  40
  41         pmd = pmd_offset(pud, addr);
  42         if (pmd_none_or_clear_bad(pmd))
  43                 return NULL;
  44
  45         return pmd;
  46 }
  47
  48 static pmd_t *alloc_new_pmd(struct mm_struct *mm, unsigned long addr)
  49 {
  50         pgd_t *pgd;
  51         pud_t *pud;
  52         pmd_t *pmd;
  53
  54         pgd = pgd_offset(mm, addr);
  55         pud = pud_alloc(mm, pgd, addr);
  56         if (!pud)
  57                 return NULL;
  58
  59         pmd = pmd_alloc(mm, pud, addr);
  60         if (!pmd)
  61                 return NULL;
  62
  63         if (!pmd_present(*pmd) && __pte_alloc(mm, pmd, addr))
  64                 return NULL;
  65
  66         return pmd;
  67 }
  68
  69 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
  70                 unsigned long old_addr, unsigned long old_end,
  71                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
  72                 unsigned long new_addr)
  73 {
  74         struct address_space *mapping = NULL;
  75         struct mm_struct *mm = vma->vm_mm;
  76         pte_t *old_pte, *new_pte, pte;
  77         spinlock_t *old_ptl, *new_ptl;
  78         unsigned long old_start;
  79
  80         old_start = old_addr;
  81         mmu_notifier_invalidate_range_start(vma->vm_mm,
  82                                             old_start, old_end);
  83         if (vma->vm_file) {
  84                 /*
  85                  * Subtle point from Rajesh Venkatasubramanian: before
  86                  * moving file-based ptes, we must lock vmtruncate out,
  87                  * since it might clean the dst vma before the src vma,
  88                  * and we propagate stale pages into the dst afterward.
  89                  */
  90                 mapping = vma->vm_file->f_mapping;
  91                 spin_lock(&mapping->i_mmap_lock);
  92                 new_vma->vm_truncate_count = 0;
  93         }
  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_nested(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_clear_flush(vma, 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_nested(new_pte - 1);
 119         pte_unmap_unlock(old_pte - 1, old_ptl);
 120         if (mapping)
 121                 spin_unlock(&mapping->i_mmap_lock);
 122         mmu_notifier_invalidate_range_end(vma->vm_mm, old_start, old_end);
 123 }
 124
 125 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 126
 127 unsigned long move_page_tables(struct vm_area_struct *vma,
 128                 unsigned long old_addr, struct vm_area_struct *new_vma,
 129                 unsigned long new_addr, unsigned long len)
 130 {
 131         unsigned long extent, next, old_end;
 132         pmd_t *old_pmd, *new_pmd;
 133
 134         old_end = old_addr + len;
 135         flush_cache_range(vma, old_addr, old_end);
 136
 137         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 138                 cond_resched();
 139                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 140                 if (next - 1 > old_end)
 141                         next = old_end;
 142                 extent = next - old_addr;
 143                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 144                 if (!old_pmd)
 145                         continue;
 146                 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
 147                 if (!new_pmd)
 148                         break;
 149                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 150                 if (extent > next - new_addr)
 151                         extent = next - new_addr;
 152                 if (extent > LATENCY_LIMIT)
 153                         extent = LATENCY_LIMIT;
 154                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 155                                 new_vma, new_pmd, new_addr);
 156         }
 157
 158         return len + old_addr - old_end;        /* how much done */
 159 }
 160
 161 static unsigned long move_vma(struct vm_area_struct *vma,
 162                 unsigned long old_addr, unsigned long old_len,
 163                 unsigned long new_len, unsigned long new_addr)
 164 {
 165         struct mm_struct *mm = vma->vm_mm;
 166         struct vm_area_struct *new_vma;
 167         unsigned long vm_flags = vma->vm_flags;
 168         unsigned long new_pgoff;
 169         unsigned long moved_len;
 170         unsigned long excess = 0;
 171         unsigned long hiwater_vm;
 172         int split = 0;
 173
 174         /*
 175          * We'd prefer to avoid failure later on in do_munmap:
 176          * which may split one vma into three before unmapping.
 177          */
 178         if (mm->map_count >= sysctl_max_map_count - 3)
 179                 return -ENOMEM;
 180
 181         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 182         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 183         if (!new_vma)
 184                 return -ENOMEM;
 185
 186         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
 187         if (moved_len < old_len) {
 188                 /*
 189                  * On error, move entries back from new area to old,
 190                  * which will succeed since page tables still there,
 191                  * and then proceed to unmap new area instead of old.
 192                  */
 193                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
 194                 vma = new_vma;
 195                 old_len = new_len;
 196                 old_addr = new_addr;
 197                 new_addr = -ENOMEM;
 198         }
 199
 200         /* Conceal VM_ACCOUNT so old reservation is not undone */
 201         if (vm_flags & VM_ACCOUNT) {
 202                 vma->vm_flags &= ~VM_ACCOUNT;
 203                 excess = vma->vm_end - vma->vm_start - old_len;
 204                 if (old_addr > vma->vm_start &&
 205                     old_addr + old_len < vma->vm_end)
 206                         split = 1;
 207         }
 208
 209         /*
 210          * If we failed to move page tables we still do total_vm increment
 211          * since do_munmap() will decrement it by old_len == new_len.
 212          *
 213          * Since total_vm is about to be raised artificially high for a
 214          * moment, we need to restore high watermark afterwards: if stats
 215          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 216          * If this were a serious issue, we'd add a flag to do_munmap().
 217          */
 218         hiwater_vm = mm->hiwater_vm;
 219         mm->total_vm += new_len >> PAGE_SHIFT;
 220         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 221
 222         if (do_munmap(mm, old_addr, old_len) < 0) {
 223                 /* OOM: unable to split vma, just get accounts right */
 224                 vm_unacct_memory(excess >> PAGE_SHIFT);
 225                 excess = 0;
 226         }
 227         mm->hiwater_vm = hiwater_vm;
 228
 229         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 230         if (excess) {
 231                 vma->vm_flags |= VM_ACCOUNT;
 232                 if (split)
 233                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 234         }
 235
 236         if (vm_flags & VM_LOCKED) {
 237                 mm->locked_vm += new_len >> PAGE_SHIFT;
 238                 if (new_len > old_len)
 239                         make_pages_present(new_addr + old_len,
 240                                            new_addr + new_len);
 241         }
 242
 243         return new_addr;
 244 }
 245
 246 /*
 247  * Expand (or shrink) an existing mapping, potentially moving it at the
 248  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 249  *
 250  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 251  * This option implies MREMAP_MAYMOVE.
 252  */
 253 unsigned long do_mremap(unsigned long addr,
 254         unsigned long old_len, unsigned long new_len,
 255         unsigned long flags, unsigned long new_addr)
 256 {
 257         struct mm_struct *mm = current->mm;
 258         struct vm_area_struct *vma;
 259         unsigned long ret = -EINVAL;
 260         unsigned long charged = 0;
 261
 262         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 263                 goto out;
 264
 265         if (addr & ~PAGE_MASK)
 266                 goto out;
 267
 268         old_len = PAGE_ALIGN(old_len);
 269         new_len = PAGE_ALIGN(new_len);
 270
 271         /*
 272          * We allow a zero old-len as a special case
 273          * for DOS-emu "duplicate shm area" thing. But
 274          * a zero new-len is nonsensical.
 275          */
 276         if (!new_len)
 277                 goto out;
 278
 279         /* new_addr is only valid if MREMAP_FIXED is specified */
 280         if (flags & MREMAP_FIXED) {
 281                 if (new_addr & ~PAGE_MASK)
 282                         goto out;
 283                 if (!(flags & MREMAP_MAYMOVE))
 284                         goto out;
 285
 286                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 287                         goto out;
 288
 289                 /* Check if the location we're moving into overlaps the
 290                  * old location at all, and fail if it does.
 291                  */
 292                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 293                         goto out;
 294
 295                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 296                         goto out;
 297
 298                 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
 299                 if (ret)
 300                         goto out;
 301
 302                 ret = do_munmap(mm, new_addr, new_len);
 303                 if (ret)
 304                         goto out;
 305         }
 306
 307         /*
 308          * Always allow a shrinking remap: that just unmaps
 309          * the unnecessary pages..
 310          * do_munmap does all the needed commit accounting
 311          */
 312         if (old_len >= new_len) {
 313                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 314                 if (ret && old_len != new_len)
 315                         goto out;
 316                 ret = addr;
 317                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 318                         goto out;
 319                 old_len = new_len;
 320         }
 321
 322         /*
 323          * Ok, we need to grow..  or relocate.
 324          */
 325         ret = -EFAULT;
 326         vma = find_vma(mm, addr);
 327         if (!vma || vma->vm_start > addr)
 328                 goto out;
 329         if (is_vm_hugetlb_page(vma)) {
 330                 ret = -EINVAL;
 331                 goto out;
 332         }
 333         /* We can't remap across vm area boundaries */
 334         if (old_len > vma->vm_end - addr)
 335                 goto out;
 336
 337         /* Need to be careful about a growing mapping */
 338         if (new_len > old_len) {
 339                 unsigned long pgoff;
 340
 341                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 342                         goto out;
 343                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 344                 pgoff += vma->vm_pgoff;
 345                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff) {
 346                         ret = -EINVAL;
 347                         goto out;
 348                 }
 349         }
 350
 351         if (vma->vm_flags & VM_LOCKED) {
 352                 unsigned long locked, lock_limit;
 353                 locked = mm->locked_vm << PAGE_SHIFT;
 354                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 355                 locked += new_len - old_len;
 356                 ret = -EAGAIN;
 357                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 358                         goto out;
 359         }
 360         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT)) {
 361                 ret = -ENOMEM;
 362                 goto out;
 363         }
 364
 365         if (vma->vm_flags & VM_ACCOUNT) {
 366                 charged = (new_len - old_len) >> PAGE_SHIFT;
 367                 if (security_vm_enough_memory(charged))
 368                         goto out_nc;
 369         }
 370
 371         /* old_len exactly to the end of the area..
 372          * And we're not relocating the area.
 373          */
 374         if (old_len == vma->vm_end - addr &&
 375             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 376             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 377                 unsigned long max_addr = TASK_SIZE;
 378                 if (vma->vm_next)
 379                         max_addr = vma->vm_next->vm_start;
 380                 /* can we just expand the current mapping? */
 381                 if (max_addr - addr >= new_len) {
 382                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 383
 384                         vma_adjust(vma, vma->vm_start,
 385                                 addr + new_len, vma->vm_pgoff, NULL);
 386
 387                         mm->total_vm += pages;
 388                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
 389                         if (vma->vm_flags & VM_LOCKED) {
 390                                 mm->locked_vm += pages;
 391                                 make_pages_present(addr + old_len,
 392                                                    addr + new_len);
 393                         }
 394                         ret = addr;
 395                         goto out;
 396                 }
 397         }
 398
 399         /*
 400          * We weren't able to just expand or shrink the area,
 401          * we need to create a new one and move it..
 402          */
 403         ret = -ENOMEM;
 404         if (flags & MREMAP_MAYMOVE) {
 405                 if (!(flags & MREMAP_FIXED)) {
 406                         unsigned long map_flags = 0;
 407                         if (vma->vm_flags & VM_MAYSHARE)
 408                                 map_flags |= MAP_SHARED;
 409
 410                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 411                                                 vma->vm_pgoff, map_flags);
 412                         if (new_addr & ~PAGE_MASK) {
 413                                 ret = new_addr;
 414                                 goto out;
 415                         }
 416
 417                         ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
 418                         if (ret)
 419                                 goto out;
 420                 }
 421                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 422         }
 423 out:
 424         if (ret & ~PAGE_MASK)
 425                 vm_unacct_memory(charged);
 426 out_nc:
 427         return ret;
 428 }
 429
 430 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 431                 unsigned long, new_len, unsigned long, flags,
 432                 unsigned long, new_addr)
 433 {
 434         unsigned long ret;
 435
 436         down_write(&current->mm->mmap_sem);
 437         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 438         up_write(&current->mm->mmap_sem);
 439         return ret;
 440 }