LXR linux-bk/mm/mremap.c

   1/*
   2 *      mm/remap.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/slab.h>
  12#include <linux/shm.h>
  13#include <linux/mman.h>
  14#include <linux/swap.h>
  15#include <linux/fs.h>
  16#include <linux/highmem.h>
  17
  18#include <asm/uaccess.h>
  19#include <asm/pgalloc.h>
  20#include <asm/cacheflush.h>
  21#include <asm/tlbflush.h>
  22
  23static inline pte_t *get_one_pte_map_nested(struct mm_struct *mm, unsigned long addr)
  24{
  25        pgd_t * pgd;
  26        pmd_t * pmd;
  27        pte_t * pte = NULL;
  28
  29        pgd = pgd_offset(mm, addr);
  30        if (pgd_none(*pgd))
  31                goto end;
  32        if (pgd_bad(*pgd)) {
  33                pgd_ERROR(*pgd);
  34                pgd_clear(pgd);
  35                goto end;
  36        }
  37
  38        pmd = pmd_offset(pgd, addr);
  39        if (pmd_none(*pmd))
  40                goto end;
  41        if (pmd_bad(*pmd)) {
  42                pmd_ERROR(*pmd);
  43                pmd_clear(pmd);
  44                goto end;
  45        }
  46
  47        pte = pte_offset_map_nested(pmd, addr);
  48        if (pte_none(*pte)) {
  49                pte_unmap_nested(pte);
  50                pte = NULL;
  51        }
  52end:
  53        return pte;
  54}
  55
  56static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr)
  57{
  58        pmd_t * pmd;
  59        pte_t * pte = NULL;
  60
  61        pmd = pmd_alloc(mm, pgd_offset(mm, addr), addr);
  62        if (pmd)
  63                pte = pte_alloc_map(mm, pmd, addr);
  64        return pte;
  65}
  66
  67static inline int copy_one_pte(struct mm_struct *mm, pte_t * src, pte_t * dst)
  68{
  69        int error = 0;
  70        pte_t pte;
  71        struct page * page = NULL;
  72
  73        if (pte_present(*src))
  74                page = pte_page(*src);
  75
  76        if (!pte_none(*src)) {
  77                if (page)
  78                        page_remove_rmap(page, src);
  79                pte = ptep_get_and_clear(src);
  80                if (!dst) {
  81                        /* No dest?  We must put it back. */
  82                        dst = src;
  83                        error++;
  84                }
  85                set_pte(dst, pte);
  86                if (page)
  87                        page_add_rmap(page, dst);
  88        }
  89        return error;
  90}
  91
  92static int move_one_page(struct vm_area_struct *vma, unsigned long old_addr, unsigned long new_addr)
  93{
  94        struct mm_struct *mm = vma->vm_mm;
  95        int error = 0;
  96        pte_t *src, *dst;
  97
  98        spin_lock(&mm->page_table_lock);
  99        src = get_one_pte_map_nested(mm, old_addr);
 100        if (src) {
 101                dst = alloc_one_pte_map(mm, new_addr);
 102                error = copy_one_pte(mm, src, dst);
 103                pte_unmap_nested(src);
 104                pte_unmap(dst);
 105        }
 106        flush_tlb_page(vma, old_addr);
 107        spin_unlock(&mm->page_table_lock);
 108        return error;
 109}
 110
 111static int move_page_tables(struct vm_area_struct *vma,
 112        unsigned long new_addr, unsigned long old_addr, unsigned long len)
 113{
 114        unsigned long offset = len;
 115
 116        flush_cache_range(vma, old_addr, old_addr + len);
 117
 118        /*
 119         * This is not the clever way to do this, but we're taking the
 120         * easy way out on the assumption that most remappings will be
 121         * only a few pages.. This also makes error recovery easier.
 122         */
 123        while (offset) {
 124                offset -= PAGE_SIZE;
 125                if (move_one_page(vma, old_addr + offset, new_addr + offset))
 126                        goto oops_we_failed;
 127        }
 128        return 0;
 129
 130        /*
 131         * Ok, the move failed because we didn't have enough pages for
 132         * the new page table tree. This is unlikely, but we have to
 133         * take the possibility into account. In that case we just move
 134         * all the pages back (this will work, because we still have
 135         * the old page tables)
 136         */
 137oops_we_failed:
 138        flush_cache_range(vma, new_addr, new_addr + len);
 139        while ((offset += PAGE_SIZE) < len)
 140                move_one_page(vma, new_addr + offset, old_addr + offset);
 141        zap_page_range(vma, new_addr, len);
 142        return -1;
 143}
 144
 145static inline unsigned long move_vma(struct vm_area_struct * vma,
 146        unsigned long addr, unsigned long old_len, unsigned long new_len,
 147        unsigned long new_addr)
 148{
 149        struct mm_struct * mm = vma->vm_mm;
 150        struct vm_area_struct * new_vma, * next, * prev;
 151        int allocated_vma;
 152        int split = 0;
 153
 154        new_vma = NULL;
 155        next = find_vma_prev(mm, new_addr, &prev);
 156        if (next) {
 157                if (prev && prev->vm_end == new_addr &&
 158                    can_vma_merge(prev, vma->vm_flags) && !vma->vm_file && !(vma->vm_flags & VM_SHARED)) {
 159                        spin_lock(&mm->page_table_lock);
 160                        prev->vm_end = new_addr + new_len;
 161                        spin_unlock(&mm->page_table_lock);
 162                        new_vma = prev;
 163                        if (next != prev->vm_next)
 164                                BUG();
 165                        if (prev->vm_end == next->vm_start && can_vma_merge(next, prev->vm_flags)) {
 166                                spin_lock(&mm->page_table_lock);
 167                                prev->vm_end = next->vm_end;
 168                                __vma_unlink(mm, next, prev);
 169                                spin_unlock(&mm->page_table_lock);
 170
 171                                mm->map_count--;
 172                                kmem_cache_free(vm_area_cachep, next);
 173                        }
 174                } else if (next->vm_start == new_addr + new_len &&
 175                           can_vma_merge(next, vma->vm_flags) && !vma->vm_file && !(vma->vm_flags & VM_SHARED)) {
 176                        spin_lock(&mm->page_table_lock);
 177                        next->vm_start = new_addr;
 178                        spin_unlock(&mm->page_table_lock);
 179                        new_vma = next;
 180                }
 181        } else {
 182                prev = find_vma(mm, new_addr-1);
 183                if (prev && prev->vm_end == new_addr &&
 184                    can_vma_merge(prev, vma->vm_flags) && !vma->vm_file && !(vma->vm_flags & VM_SHARED)) {
 185                        spin_lock(&mm->page_table_lock);
 186                        prev->vm_end = new_addr + new_len;
 187                        spin_unlock(&mm->page_table_lock);
 188                        new_vma = prev;
 189                }
 190        }
 191
 192        allocated_vma = 0;
 193        if (!new_vma) {
 194                new_vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
 195                if (!new_vma)
 196                        goto out;
 197                allocated_vma = 1;
 198        }
 199
 200        if (!move_page_tables(vma, new_addr, addr, old_len)) {
 201                if (allocated_vma) {
 202                        *new_vma = *vma;
 203                        new_vma->vm_start = new_addr;
 204                        new_vma->vm_end = new_addr+new_len;
 205                        new_vma->vm_pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
 206                        new_vma->vm_raend = 0;
 207                        if (new_vma->vm_file)
 208                                get_file(new_vma->vm_file);
 209                        if (new_vma->vm_ops && new_vma->vm_ops->open)
 210                                new_vma->vm_ops->open(new_vma);
 211                        insert_vm_struct(current->mm, new_vma);
 212                }
 213
 214                /* Conceal VM_ACCOUNT so old reservation is not undone */
 215                if (vma->vm_flags & VM_ACCOUNT) {
 216                        vma->vm_flags &= ~VM_ACCOUNT;
 217                        if (addr > vma->vm_start) {
 218                                if (addr + old_len < vma->vm_end)
 219                                        split = 1;
 220                        } else if (addr + old_len == vma->vm_end)
 221                                vma = NULL;     /* it will be removed */
 222                } else
 223                        vma = NULL;             /* nothing more to do */
 224
 225                do_munmap(current->mm, addr, old_len);
 226
 227                /* Restore VM_ACCOUNT if one or two pieces of vma left */
 228                if (vma) {
 229                        vma->vm_flags |= VM_ACCOUNT;
 230                        if (split)
 231                                vma->vm_next->vm_flags |= VM_ACCOUNT;
 232                }
 233                current->mm->total_vm += new_len >> PAGE_SHIFT;
 234                if (new_vma->vm_flags & VM_LOCKED) {
 235                        current->mm->locked_vm += new_len >> PAGE_SHIFT;
 236                        make_pages_present(new_vma->vm_start,
 237                                           new_vma->vm_end);
 238                }
 239                return new_addr;
 240        }
 241        if (allocated_vma)
 242                kmem_cache_free(vm_area_cachep, new_vma);
 243 out:
 244        return -ENOMEM;
 245}
 246
 247/*
 248 * Expand (or shrink) an existing mapping, potentially moving it at the
 249 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 250 *
 251 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 252 * This option implies MREMAP_MAYMOVE.
 253 */
 254unsigned long do_mremap(unsigned long addr,
 255        unsigned long old_len, unsigned long new_len,
 256        unsigned long flags, unsigned long new_addr)
 257{
 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        /* new_addr is only valid if MREMAP_FIXED is specified */
 272        if (flags & MREMAP_FIXED) {
 273                if (new_addr & ~PAGE_MASK)
 274                        goto out;
 275                if (!(flags & MREMAP_MAYMOVE))
 276                        goto out;
 277
 278                if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 279                        goto out;
 280
 281                /* Check if the location we're moving into overlaps the
 282                 * old location at all, and fail if it does.
 283                 */
 284                if ((new_addr <= addr) && (new_addr+new_len) > addr)
 285                        goto out;
 286
 287                if ((addr <= new_addr) && (addr+old_len) > new_addr)
 288                        goto out;
 289
 290                do_munmap(current->mm, new_addr, new_len);
 291        }
 292
 293        /*
 294         * Always allow a shrinking remap: that just unmaps
 295         * the unnecessary pages..
 296         * do_munmap does all the needed commit accounting
 297         */
 298        ret = addr;
 299        if (old_len >= new_len) {
 300                do_munmap(current->mm, addr+new_len, old_len - new_len);
 301                if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 302                        goto out;
 303                old_len = new_len;
 304        }
 305
 306        /*
 307         * Ok, we need to grow..  or relocate.
 308         */
 309        ret = -EFAULT;
 310        vma = find_vma(current->mm, addr);
 311        if (!vma || vma->vm_start > addr)
 312                goto out;
 313        if (is_vm_hugetlb_page(vma)) {
 314                ret = -EINVAL;
 315                goto out;
 316        }
 317        /* We can't remap across vm area boundaries */
 318        if (old_len > vma->vm_end - addr)
 319                goto out;
 320        if (vma->vm_flags & VM_DONTEXPAND) {
 321                if (new_len > old_len)
 322                        goto out;
 323        }
 324        if (vma->vm_flags & VM_LOCKED) {
 325                unsigned long locked = current->mm->locked_vm << PAGE_SHIFT;
 326                locked += new_len - old_len;
 327                ret = -EAGAIN;
 328                if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur)
 329                        goto out;
 330        }
 331        ret = -ENOMEM;
 332        if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
 333            > current->rlim[RLIMIT_AS].rlim_cur)
 334                goto out;
 335
 336        if (vma->vm_flags & VM_ACCOUNT) {
 337                charged = (new_len - old_len) >> PAGE_SHIFT;
 338                if (!vm_enough_memory(charged))
 339                        goto out_nc;
 340        }
 341
 342        /* old_len exactly to the end of the area..
 343         * And we're not relocating the area.
 344         */
 345        if (old_len == vma->vm_end - addr &&
 346            !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 347            (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 348                unsigned long max_addr = TASK_SIZE;
 349                if (vma->vm_next)
 350                        max_addr = vma->vm_next->vm_start;
 351                /* can we just expand the current mapping? */
 352                if (max_addr - addr >= new_len) {
 353                        int pages = (new_len - old_len) >> PAGE_SHIFT;
 354                        spin_lock(&vma->vm_mm->page_table_lock);
 355                        vma->vm_end = addr + new_len;
 356                        spin_unlock(&vma->vm_mm->page_table_lock);
 357                        current->mm->total_vm += pages;
 358                        if (vma->vm_flags & VM_LOCKED) {
 359                                current->mm->locked_vm += pages;
 360                                make_pages_present(addr + old_len,
 361                                                   addr + new_len);
 362                        }
 363                        ret = addr;
 364                        goto out;
 365                }
 366        }
 367
 368        /*
 369         * We weren't able to just expand or shrink the area,
 370         * we need to create a new one and move it..
 371         */
 372        ret = -ENOMEM;
 373        if (flags & MREMAP_MAYMOVE) {
 374                if (!(flags & MREMAP_FIXED)) {
 375                        unsigned long map_flags = 0;
 376                        if (vma->vm_flags & VM_SHARED)
 377                                map_flags |= MAP_SHARED;
 378
 379                        new_addr = get_unmapped_area(vma->vm_file, 0, new_len, vma->vm_pgoff, map_flags);
 380                        ret = new_addr;
 381                        if (new_addr & ~PAGE_MASK)
 382                                goto out;
 383                }
 384                ret = move_vma(vma, addr, old_len, new_len, new_addr);
 385        }
 386out:
 387        if (ret & ~PAGE_MASK)
 388                vm_unacct_memory(charged);
 389out_nc:
 390        return ret;
 391}
 392
 393asmlinkage unsigned long sys_mremap(unsigned long addr,
 394        unsigned long old_len, unsigned long new_len,
 395        unsigned long flags, unsigned long new_addr)
 396{
 397        unsigned long ret;
 398
 399        down_write(&current->mm->mmap_sem);
 400        ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 401        up_write(&current->mm->mmap_sem);
 402        return ret;
 403}
 404