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/vs_base.h>
  22 #include <linux/vs_memory.h>
  23
  24 #include <asm/uaccess.h>
  25 #include <asm/cacheflush.h>
  26 #include <asm/tlbflush.h>
  27
  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
  81         if (vma->vm_file) {
  82                 /*
  83                  * Subtle point from Rajesh Venkatasubramanian: before
  84                  * moving file-based ptes, we must lock vmtruncate out,
  85                  * since it might clean the dst vma before the src vma,
  86                  * and we propagate stale pages into the dst afterward.
  87                  */
  88                 mapping = vma->vm_file->f_mapping;
  89                 spin_lock(&mapping->i_mmap_lock);
  90                 if (new_vma->vm_truncate_count &&
  91                     new_vma->vm_truncate_count != vma->vm_truncate_count)
  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(new_ptl);
 104
 105         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 106                                    new_pte++, new_addr += PAGE_SIZE) {
 107                 if (pte_none(*old_pte))
 108                         continue;
 109                 pte = ptep_clear_flush(vma, old_addr, old_pte);
 110                 /* ZERO_PAGE can be dependant on virtual addr */
 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         if (new_ptl != old_ptl)
 116                 spin_unlock(new_ptl);
 117         pte_unmap_nested(new_pte - 1);
 118         pte_unmap_unlock(old_pte - 1, old_ptl);
 119         if (mapping)
 120                 spin_unlock(&mapping->i_mmap_lock);
 121 }
 122
 123 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 124
 125 static unsigned long move_page_tables(struct vm_area_struct *vma,
 126                 unsigned long old_addr, struct vm_area_struct *new_vma,
 127                 unsigned long new_addr, unsigned long len)
 128 {
 129         unsigned long extent, next, old_end;
 130         pmd_t *old_pmd, *new_pmd;
 131
 132         old_end = old_addr + len;
 133         flush_cache_range(vma, old_addr, old_end);
 134
 135         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 136                 cond_resched();
 137                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 138                 if (next - 1 > old_end)
 139                         next = old_end;
 140                 extent = next - old_addr;
 141                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 142                 if (!old_pmd)
 143                         continue;
 144                 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
 145                 if (!new_pmd)
 146                         break;
 147                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 148                 if (extent > next - new_addr)
 149                         extent = next - new_addr;
 150                 if (extent > LATENCY_LIMIT)
 151                         extent = LATENCY_LIMIT;
 152                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 153                                 new_vma, new_pmd, new_addr);
 154         }
 155
 156         return len + old_addr - old_end;        /* how much done */
 157 }
 158
 159 static unsigned long move_vma(struct vm_area_struct *vma,
 160                 unsigned long old_addr, unsigned long old_len,
 161                 unsigned long new_len, unsigned long new_addr)
 162 {
 163         struct mm_struct *mm = vma->vm_mm;
 164         struct vm_area_struct *new_vma;
 165         unsigned long vm_flags = vma->vm_flags;
 166         unsigned long new_pgoff;
 167         unsigned long moved_len;
 168         unsigned long excess = 0;
 169         unsigned long hiwater_vm;
 170         int split = 0;
 171
 172         /*
 173          * We'd prefer to avoid failure later on in do_munmap:
 174          * which may split one vma into three before unmapping.
 175          */
 176         if (mm->map_count >= sysctl_max_map_count - 3)
 177                 return -ENOMEM;
 178
 179         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 180         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 181         if (!new_vma)
 182                 return -ENOMEM;
 183
 184         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
 185         if (moved_len < old_len) {
 186                 /*
 187                  * On error, move entries back from new area to old,
 188                  * which will succeed since page tables still there,
 189                  * and then proceed to unmap new area instead of old.
 190                  */
 191                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
 192                 vma = new_vma;
 193                 old_len = new_len;
 194                 old_addr = new_addr;
 195                 new_addr = -ENOMEM;
 196         }
 197
 198         /* Conceal VM_ACCOUNT so old reservation is not undone */
 199         if (vm_flags & VM_ACCOUNT) {
 200                 vma->vm_flags &= ~VM_ACCOUNT;
 201                 excess = vma->vm_end - vma->vm_start - old_len;
 202                 if (old_addr > vma->vm_start &&
 203                     old_addr + old_len < vma->vm_end)
 204                         split = 1;
 205         }
 206
 207         /*
 208          * If we failed to move page tables we still do total_vm increment
 209          * since do_munmap() will decrement it by old_len == new_len.
 210          *
 211          * Since total_vm is about to be raised artificially high for a
 212          * moment, we need to restore high watermark afterwards: if stats
 213          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 214          * If this were a serious issue, we'd add a flag to do_munmap().
 215          */
 216         hiwater_vm = mm->hiwater_vm;
 217         vx_vmpages_add(mm, new_len >> PAGE_SHIFT);
 218         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 219
 220         if (do_munmap(mm, old_addr, old_len) < 0) {
 221                 /* OOM: unable to split vma, just get accounts right */
 222                 vm_unacct_memory(excess >> PAGE_SHIFT);
 223                 excess = 0;
 224         }
 225         mm->hiwater_vm = hiwater_vm;
 226
 227         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 228         if (excess) {
 229                 vma->vm_flags |= VM_ACCOUNT;
 230                 if (split)
 231                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 232         }
 233
 234         if (vm_flags & VM_LOCKED) {
 235                 vx_vmlocked_add(mm, new_len >> PAGE_SHIFT);
 236                 if (new_len > old_len)
 237                         make_pages_present(new_addr + old_len,
 238                                            new_addr + new_len);
 239         }
 240
 241         return new_addr;
 242 }
 243
 244 /*
 245  * Expand (or shrink) an existing mapping, potentially moving it at the
 246  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 247  *
 248  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 249  * This option implies MREMAP_MAYMOVE.
 250  */
 251 unsigned long do_mremap(unsigned long addr,
 252         unsigned long old_len, unsigned long new_len,
 253         unsigned long flags, unsigned long new_addr)
 254 {
 255         struct mm_struct *mm = current->mm;
 256         struct vm_area_struct *vma;
 257         unsigned long ret = -EINVAL;
 258         unsigned long charged = 0;
 259
 260         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 261                 goto out;
 262
 263         if (addr & ~PAGE_MASK)
 264                 goto out;
 265
 266         old_len = PAGE_ALIGN(old_len);
 267         new_len = PAGE_ALIGN(new_len);
 268
 269         /*
 270          * We allow a zero old-len as a special case
 271          * for DOS-emu "duplicate shm area" thing. But
 272          * a zero new-len is nonsensical.
 273          */
 274         if (!new_len)
 275                 goto out;
 276
 277         /* new_addr is only valid if MREMAP_FIXED is specified */
 278         if (flags & MREMAP_FIXED) {
 279                 if (new_addr & ~PAGE_MASK)
 280                         goto out;
 281                 if (!(flags & MREMAP_MAYMOVE))
 282                         goto out;
 283
 284                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 285                         goto out;
 286
 287                 /* Check if the location we're moving into overlaps the
 288                  * old location at all, and fail if it does.
 289                  */
 290                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 291                         goto out;
 292
 293                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 294                         goto out;
 295
 296                 ret = do_munmap(mm, new_addr, new_len);
 297                 if (ret)
 298                         goto out;
 299         }
 300
 301         /*
 302          * Always allow a shrinking remap: that just unmaps
 303          * the unnecessary pages..
 304          * do_munmap does all the needed commit accounting
 305          */
 306         if (old_len >= new_len) {
 307                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 308                 if (ret && old_len != new_len)
 309                         goto out;
 310                 ret = addr;
 311                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 312                         goto out;
 313                 old_len = new_len;
 314         }
 315
 316         /*
 317          * Ok, we need to grow..  or relocate.
 318          */
 319         ret = -EFAULT;
 320         vma = find_vma(mm, addr);
 321         if (!vma || vma->vm_start > addr)
 322                 goto out;
 323         if (is_vm_hugetlb_page(vma)) {
 324                 ret = -EINVAL;
 325                 goto out;
 326         }
 327         /* We can't remap across vm area boundaries */
 328         if (old_len > vma->vm_end - addr)
 329                 goto out;
 330         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)) {
 331                 if (new_len > old_len)
 332                         goto out;
 333         }
 334         if (vma->vm_flags & VM_LOCKED) {
 335                 unsigned long locked, lock_limit;
 336                 locked = mm->locked_vm << PAGE_SHIFT;
 337                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 338                 locked += new_len - old_len;
 339                 ret = -EAGAIN;
 340                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 341                         goto out;
 342                 if (!vx_vmlocked_avail(current->mm,
 343                         (new_len - old_len) >> PAGE_SHIFT))
 344                         goto out;
 345         }
 346         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT)) {
 347                 ret = -ENOMEM;
 348                 goto out;
 349         }
 350
 351         if (vma->vm_flags & VM_ACCOUNT) {
 352                 charged = (new_len - old_len) >> PAGE_SHIFT;
 353                 if (security_vm_enough_memory(charged))
 354                         goto out_nc;
 355         }
 356
 357         /* old_len exactly to the end of the area..
 358          * And we're not relocating the area.
 359          */
 360         if (old_len == vma->vm_end - addr &&
 361             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 362             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 363                 unsigned long max_addr = TASK_SIZE;
 364                 if (vma->vm_next)
 365                         max_addr = vma->vm_next->vm_start;
 366                 /* can we just expand the current mapping? */
 367                 if (max_addr - addr >= new_len) {
 368                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 369
 370                         vma_adjust(vma, vma->vm_start,
 371                                 addr + new_len, vma->vm_pgoff, NULL);
 372
 373                         vx_vmpages_add(mm, pages);
 374                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
 375                         if (vma->vm_flags & VM_LOCKED) {
 376                                 vx_vmlocked_add(mm, pages);
 377                                 make_pages_present(addr + old_len,
 378                                                    addr + new_len);
 379                         }
 380                         ret = addr;
 381                         goto out;
 382                 }
 383         }
 384
 385         /*
 386          * We weren't able to just expand or shrink the area,
 387          * we need to create a new one and move it..
 388          */
 389         ret = -ENOMEM;
 390         if (flags & MREMAP_MAYMOVE) {
 391                 if (!(flags & MREMAP_FIXED)) {
 392                         unsigned long map_flags = 0;
 393                         if (vma->vm_flags & VM_MAYSHARE)
 394                                 map_flags |= MAP_SHARED;
 395
 396                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 397                                                 vma->vm_pgoff, map_flags);
 398                         ret = new_addr;
 399                         if (new_addr & ~PAGE_MASK)
 400                                 goto out;
 401                 }
 402                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 403         }
 404 out:
 405         if (ret & ~PAGE_MASK)
 406                 vm_unacct_memory(charged);
 407 out_nc:
 408         return ret;
 409 }
 410
 411 asmlinkage unsigned long sys_mremap(unsigned long addr,
 412         unsigned long old_len, unsigned long new_len,
 413         unsigned long flags, unsigned long new_addr)
 414 {
 415         unsigned long ret;
 416
 417         down_write(&current->mm->mmap_sem);
 418         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 419         up_write(&current->mm->mmap_sem);
 420         return ret;
 421 }