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 #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 vmtruncate out,
  89                  * 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                 if (new_vma->vm_truncate_count &&
  95                     new_vma->vm_truncate_count != vma->vm_truncate_count)
  96                         new_vma->vm_truncate_count = 0;
  97         }
  98
  99         /*
 100          * We don't have to worry about the ordering of src and dst
 101          * pte locks because exclusive mmap_sem prevents deadlock.
 102          */
 103         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 104         new_pte = pte_offset_map_nested(new_pmd, new_addr);
 105         new_ptl = pte_lockptr(mm, new_pmd);
 106         if (new_ptl != old_ptl)
 107                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 108         arch_enter_lazy_mmu_mode();
 109
 110         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 111                                    new_pte++, new_addr += PAGE_SIZE) {
 112                 if (pte_none(*old_pte))
 113                         continue;
 114                 pte = ptep_clear_flush(vma, old_addr, old_pte);
 115                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 116                 set_pte_at(mm, new_addr, new_pte, pte);
 117         }
 118
 119         arch_leave_lazy_mmu_mode();
 120         if (new_ptl != old_ptl)
 121                 spin_unlock(new_ptl);
 122         pte_unmap_nested(new_pte - 1);
 123         pte_unmap_unlock(old_pte - 1, old_ptl);
 124         if (mapping)
 125                 spin_unlock(&mapping->i_mmap_lock);
 126         mmu_notifier_invalidate_range_end(vma->vm_mm, old_start, old_end);
 127 }
 128
 129 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 130
 131 unsigned long move_page_tables(struct vm_area_struct *vma,
 132                 unsigned long old_addr, struct vm_area_struct *new_vma,
 133                 unsigned long new_addr, unsigned long len)
 134 {
 135         unsigned long extent, next, old_end;
 136         pmd_t *old_pmd, *new_pmd;
 137
 138         old_end = old_addr + len;
 139         flush_cache_range(vma, old_addr, old_end);
 140
 141         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 142                 cond_resched();
 143                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 144                 if (next - 1 > old_end)
 145                         next = old_end;
 146                 extent = next - old_addr;
 147                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 148                 if (!old_pmd)
 149                         continue;
 150                 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
 151                 if (!new_pmd)
 152                         break;
 153                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 154                 if (extent > next - new_addr)
 155                         extent = next - new_addr;
 156                 if (extent > LATENCY_LIMIT)
 157                         extent = LATENCY_LIMIT;
 158                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 159                                 new_vma, new_pmd, new_addr);
 160         }
 161
 162         return len + old_addr - old_end;        /* how much done */
 163 }
 164
 165 static unsigned long move_vma(struct vm_area_struct *vma,
 166                 unsigned long old_addr, unsigned long old_len,
 167                 unsigned long new_len, unsigned long new_addr)
 168 {
 169         struct mm_struct *mm = vma->vm_mm;
 170         struct vm_area_struct *new_vma;
 171         unsigned long vm_flags = vma->vm_flags;
 172         unsigned long new_pgoff;
 173         unsigned long moved_len;
 174         unsigned long excess = 0;
 175         unsigned long hiwater_vm;
 176         int split = 0;
 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         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 186         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 187         if (!new_vma)
 188                 return -ENOMEM;
 189
 190         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
 191         if (moved_len < old_len) {
 192                 /*
 193                  * On error, move entries back from new area to old,
 194                  * which will succeed since page tables still there,
 195                  * and then proceed to unmap new area instead of old.
 196                  */
 197                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
 198                 vma = new_vma;
 199                 old_len = new_len;
 200                 old_addr = new_addr;
 201                 new_addr = -ENOMEM;
 202         }
 203
 204         /* Conceal VM_ACCOUNT so old reservation is not undone */
 205         if (vm_flags & VM_ACCOUNT) {
 206                 vma->vm_flags &= ~VM_ACCOUNT;
 207                 excess = vma->vm_end - vma->vm_start - old_len;
 208                 if (old_addr > vma->vm_start &&
 209                     old_addr + old_len < vma->vm_end)
 210                         split = 1;
 211         }
 212
 213         /*
 214          * If we failed to move page tables we still do total_vm increment
 215          * since do_munmap() will decrement it by old_len == new_len.
 216          *
 217          * Since total_vm is about to be raised artificially high for a
 218          * moment, we need to restore high watermark afterwards: if stats
 219          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 220          * If this were a serious issue, we'd add a flag to do_munmap().
 221          */
 222         hiwater_vm = mm->hiwater_vm;
 223         mm->total_vm += new_len >> PAGE_SHIFT;
 224         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 225
 226         if (do_munmap(mm, old_addr, old_len) < 0) {
 227                 /* OOM: unable to split vma, just get accounts right */
 228                 vm_unacct_memory(excess >> PAGE_SHIFT);
 229                 excess = 0;
 230         }
 231         mm->hiwater_vm = hiwater_vm;
 232
 233         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 234         if (excess) {
 235                 vma->vm_flags |= VM_ACCOUNT;
 236                 if (split)
 237                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 238         }
 239
 240         if (vm_flags & VM_LOCKED) {
 241                 mm->locked_vm += new_len >> PAGE_SHIFT;
 242                 if (new_len > old_len)
 243                         mlock_vma_pages_range(new_vma, new_addr + old_len,
 244                                                        new_addr + new_len);
 245         }
 246
 247         return new_addr;
 248 }
 249
 250 /*
 251  * Expand (or shrink) an existing mapping, potentially moving it at the
 252  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 253  *
 254  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 255  * This option implies MREMAP_MAYMOVE.
 256  */
 257 unsigned long do_mremap(unsigned long addr,
 258         unsigned long old_len, unsigned long new_len,
 259         unsigned long flags, unsigned long new_addr)
 260 {
 261         struct mm_struct *mm = current->mm;
 262         struct vm_area_struct *vma;
 263         unsigned long ret = -EINVAL;
 264         unsigned long charged = 0;
 265
 266         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 267                 goto out;
 268
 269         if (addr & ~PAGE_MASK)
 270                 goto out;
 271
 272         old_len = PAGE_ALIGN(old_len);
 273         new_len = PAGE_ALIGN(new_len);
 274
 275         /*
 276          * We allow a zero old-len as a special case
 277          * for DOS-emu "duplicate shm area" thing. But
 278          * a zero new-len is nonsensical.
 279          */
 280         if (!new_len)
 281                 goto out;
 282
 283         /* new_addr is only valid if MREMAP_FIXED is specified */
 284         if (flags & MREMAP_FIXED) {
 285                 if (new_addr & ~PAGE_MASK)
 286                         goto out;
 287                 if (!(flags & MREMAP_MAYMOVE))
 288                         goto out;
 289
 290                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 291                         goto out;
 292
 293                 /* Check if the location we're moving into overlaps the
 294                  * old location at all, and fail if it does.
 295                  */
 296                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 297                         goto out;
 298
 299                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 300                         goto out;
 301
 302                 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
 303                 if (ret)
 304                         goto out;
 305
 306                 ret = do_munmap(mm, new_addr, new_len);
 307                 if (ret)
 308                         goto out;
 309         }
 310
 311         /*
 312          * Always allow a shrinking remap: that just unmaps
 313          * the unnecessary pages..
 314          * do_munmap does all the needed commit accounting
 315          */
 316         if (old_len >= new_len) {
 317                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 318                 if (ret && old_len != new_len)
 319                         goto out;
 320                 ret = addr;
 321                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 322                         goto out;
 323                 old_len = new_len;
 324         }
 325
 326         /*
 327          * Ok, we need to grow..  or relocate.
 328          */
 329         ret = -EFAULT;
 330         vma = find_vma(mm, addr);
 331         if (!vma || vma->vm_start > addr)
 332                 goto out;
 333         if (is_vm_hugetlb_page(vma)) {
 334                 ret = -EINVAL;
 335                 goto out;
 336         }
 337         /* We can't remap across vm area boundaries */
 338         if (old_len > vma->vm_end - addr)
 339                 goto out;
 340         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)) {
 341                 if (new_len > old_len)
 342                         goto out;
 343         }
 344         if (vma->vm_flags & VM_LOCKED) {
 345                 unsigned long locked, lock_limit;
 346                 locked = mm->locked_vm << PAGE_SHIFT;
 347                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 348                 locked += new_len - old_len;
 349                 ret = -EAGAIN;
 350                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 351                         goto out;
 352         }
 353         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT)) {
 354                 ret = -ENOMEM;
 355                 goto out;
 356         }
 357
 358         if (vma->vm_flags & VM_ACCOUNT) {
 359                 charged = (new_len - old_len) >> PAGE_SHIFT;
 360                 if (security_vm_enough_memory(charged))
 361                         goto out_nc;
 362         }
 363
 364         /* old_len exactly to the end of the area..
 365          * And we're not relocating the area.
 366          */
 367         if (old_len == vma->vm_end - addr &&
 368             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 369             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 370                 unsigned long max_addr = TASK_SIZE;
 371                 if (vma->vm_next)
 372                         max_addr = vma->vm_next->vm_start;
 373                 /* can we just expand the current mapping? */
 374                 if (max_addr - addr >= new_len) {
 375                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 376
 377                         vma_adjust(vma, vma->vm_start,
 378                                 addr + new_len, vma->vm_pgoff, NULL);
 379
 380                         mm->total_vm += pages;
 381                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
 382                         if (vma->vm_flags & VM_LOCKED) {
 383                                 mm->locked_vm += pages;
 384                                 mlock_vma_pages_range(vma, addr + old_len,
 385                                                    addr + new_len);
 386                         }
 387                         ret = addr;
 388                         goto out;
 389                 }
 390         }
 391
 392         /*
 393          * We weren't able to just expand or shrink the area,
 394          * we need to create a new one and move it..
 395          */
 396         ret = -ENOMEM;
 397         if (flags & MREMAP_MAYMOVE) {
 398                 if (!(flags & MREMAP_FIXED)) {
 399                         unsigned long map_flags = 0;
 400                         if (vma->vm_flags & VM_MAYSHARE)
 401                                 map_flags |= MAP_SHARED;
 402
 403                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 404                                                 vma->vm_pgoff, map_flags);
 405                         if (new_addr & ~PAGE_MASK) {
 406                                 ret = new_addr;
 407                                 goto out;
 408                         }
 409
 410                         ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
 411                         if (ret)
 412                                 goto out;
 413                 }
 414                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 415         }
 416 out:
 417         if (ret & ~PAGE_MASK)
 418                 vm_unacct_memory(charged);
 419 out_nc:
 420         return ret;
 421 }
 422
 423 asmlinkage unsigned long sys_mremap(unsigned long addr,
 424         unsigned long old_len, unsigned long new_len,
 425         unsigned long flags, unsigned long new_addr)
 426 {
 427         unsigned long ret;
 428
 429         down_write(&current->mm->mmap_sem);
 430         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 431         up_write(&current->mm->mmap_sem);
 432         return ret;
 433 }