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/fs.h>
  17 #include <linux/highmem.h>
  18 #include <linux/security.h>
  19 #include <linux/vs_memory.h>
  20 #include <linux/syscalls.h>
  21 #include <linux/vs_memory.h>
  22
  23 #include <asm/uaccess.h>
  24 #include <asm/cacheflush.h>
  25 #include <asm/tlbflush.h>
  26
  27 static pte_t *get_one_pte_map_nested(struct mm_struct *mm, unsigned long addr)
  28 {
  29         pgd_t *pgd;
  30         pmd_t *pmd;
  31         pte_t *pte = NULL;
  32
  33         pgd = pgd_offset(mm, addr);
  34         if (pgd_none(*pgd))
  35                 goto end;
  36         if (pgd_bad(*pgd)) {
  37                 pgd_ERROR(*pgd);
  38                 pgd_clear(pgd);
  39                 goto end;
  40         }
  41
  42         pmd = pmd_offset(pgd, addr);
  43         if (pmd_none(*pmd))
  44                 goto end;
  45         if (pmd_bad(*pmd)) {
  46                 pmd_ERROR(*pmd);
  47                 pmd_clear(pmd);
  48                 goto end;
  49         }
  50
  51         pte = pte_offset_map_nested(pmd, addr);
  52         if (pte_none(*pte)) {
  53                 pte_unmap_nested(pte);
  54                 pte = NULL;
  55         }
  56 end:
  57         return pte;
  58 }
  59
  60 static pte_t *get_one_pte_map(struct mm_struct *mm, unsigned long addr)
  61 {
  62         pgd_t *pgd;
  63         pmd_t *pmd;
  64
  65         pgd = pgd_offset(mm, addr);
  66         if (pgd_none(*pgd))
  67                 return NULL;
  68         pmd = pmd_offset(pgd, addr);
  69         if (!pmd_present(*pmd))
  70                 return NULL;
  71         return pte_offset_map(pmd, addr);
  72 }
  73
  74 static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr)
  75 {
  76         pmd_t *pmd;
  77         pte_t *pte = NULL;
  78
  79         pmd = pmd_alloc(mm, pgd_offset(mm, addr), addr);
  80         if (pmd)
  81                 pte = pte_alloc_map(mm, pmd, addr);
  82         return pte;
  83 }
  84
  85 static int
  86 move_one_page(struct vm_area_struct *vma, unsigned long old_addr,
  87                 unsigned long new_addr)
  88 {
  89         struct address_space *mapping = NULL;
  90         struct mm_struct *mm = vma->vm_mm;
  91         int error = 0;
  92         pte_t *src, *dst;
  93
  94         if (vma->vm_file) {
  95                 /*
  96                  * Subtle point from Rajesh Venkatasubramanian: before
  97                  * moving file-based ptes, we must lock vmtruncate out,
  98                  * since it might clean the dst vma before the src vma,
  99                  * and we propagate stale pages into the dst afterward.
 100                  */
 101                 mapping = vma->vm_file->f_mapping;
 102                 spin_lock(&mapping->i_mmap_lock);
 103         }
 104         spin_lock(&mm->page_table_lock);
 105
 106         src = get_one_pte_map_nested(mm, old_addr);
 107         if (src) {
 108                 /*
 109                  * Look to see whether alloc_one_pte_map needs to perform a
 110                  * memory allocation.  If it does then we need to drop the
 111                  * atomic kmap
 112                  */
 113                 dst = get_one_pte_map(mm, new_addr);
 114                 if (unlikely(!dst)) {
 115                         pte_unmap_nested(src);
 116                         if (mapping)
 117                                 spin_unlock(&mapping->i_mmap_lock);
 118                         dst = alloc_one_pte_map(mm, new_addr);
 119                         if (mapping && !spin_trylock(&mapping->i_mmap_lock)) {
 120                                 spin_unlock(&mm->page_table_lock);
 121                                 spin_lock(&mapping->i_mmap_lock);
 122                                 spin_lock(&mm->page_table_lock);
 123                         }
 124                         src = get_one_pte_map_nested(mm, old_addr);
 125                 }
 126                 /*
 127                  * Since alloc_one_pte_map can drop and re-acquire
 128                  * page_table_lock, we should re-check the src entry...
 129                  */
 130                 if (src) {
 131                         if (dst) {
 132                                 pte_t pte;
 133                                 pte = ptep_clear_flush(vma, old_addr, src);
 134                                 set_pte(dst, pte);
 135                         } else
 136                                 error = -ENOMEM;
 137                         pte_unmap_nested(src);
 138                 }
 139                 if (dst)
 140                         pte_unmap(dst);
 141         }
 142         spin_unlock(&mm->page_table_lock);
 143         if (mapping)
 144                 spin_unlock(&mapping->i_mmap_lock);
 145         return error;
 146 }
 147
 148 static unsigned long move_page_tables(struct vm_area_struct *vma,
 149                 unsigned long new_addr, unsigned long old_addr,
 150                 unsigned long len)
 151 {
 152         unsigned long offset;
 153
 154         flush_cache_range(vma, old_addr, old_addr + len);
 155
 156         /*
 157          * This is not the clever way to do this, but we're taking the
 158          * easy way out on the assumption that most remappings will be
 159          * only a few pages.. This also makes error recovery easier.
 160          */
 161         for (offset = 0; offset < len; offset += PAGE_SIZE) {
 162                 if (move_one_page(vma, old_addr+offset, new_addr+offset) < 0)
 163                         break;
 164                 cond_resched();
 165         }
 166         return offset;
 167 }
 168
 169 static unsigned long move_vma(struct vm_area_struct *vma,
 170                 unsigned long old_addr, unsigned long old_len,
 171                 unsigned long new_len, unsigned long new_addr)
 172 {
 173         struct mm_struct *mm = vma->vm_mm;
 174         struct vm_area_struct *new_vma;
 175         unsigned long vm_flags = vma->vm_flags;
 176         unsigned long new_pgoff;
 177         unsigned long moved_len;
 178         unsigned long excess = 0;
 179         int split = 0;
 180
 181         /*
 182          * We'd prefer to avoid failure later on in do_munmap:
 183          * which may split one vma into three before unmapping.
 184          */
 185         if (mm->map_count >= sysctl_max_map_count - 3)
 186                 return -ENOMEM;
 187
 188         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 189         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 190         if (!new_vma)
 191                 return -ENOMEM;
 192
 193         moved_len = move_page_tables(vma, new_addr, old_addr, old_len);
 194         if (moved_len < old_len) {
 195                 /*
 196                  * On error, move entries back from new area to old,
 197                  * which will succeed since page tables still there,
 198                  * and then proceed to unmap new area instead of old.
 199                  */
 200                 move_page_tables(new_vma, old_addr, new_addr, moved_len);
 201                 vma = new_vma;
 202                 old_len = new_len;
 203                 old_addr = new_addr;
 204                 new_addr = -ENOMEM;
 205         }
 206
 207         /* Conceal VM_ACCOUNT so old reservation is not undone */
 208         if (vm_flags & VM_ACCOUNT) {
 209                 vma->vm_flags &= ~VM_ACCOUNT;
 210                 excess = vma->vm_end - vma->vm_start - old_len;
 211                 if (old_addr > vma->vm_start &&
 212                     old_addr + old_len < vma->vm_end)
 213                         split = 1;
 214         }
 215
 216         if (do_munmap(mm, old_addr, old_len) < 0) {
 217                 /* OOM: unable to split vma, just get accounts right */
 218                 vm_unacct_memory(excess >> PAGE_SHIFT);
 219                 excess = 0;
 220         }
 221
 222         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 223         if (excess) {
 224                 vma->vm_flags |= VM_ACCOUNT;
 225                 if (split)
 226                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 227         }
 228
 229         // mm->total_vm += new_len >> PAGE_SHIFT;
 230         vx_vmpages_add(mm, new_len >> PAGE_SHIFT);
 231         __vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 232         if (vm_flags & VM_LOCKED) {
 233                 // mm->locked_vm += new_len >> PAGE_SHIFT;
 234                 vx_vmlocked_add(mm, new_len >> PAGE_SHIFT);
 235                 if (new_len > old_len)
 236                         make_pages_present(new_addr + old_len,
 237                                            new_addr + new_len);
 238         }
 239
 240         return new_addr;
 241 }
 242
 243 /*
 244  * Expand (or shrink) an existing mapping, potentially moving it at the
 245  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 246  *
 247  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 248  * This option implies MREMAP_MAYMOVE.
 249  */
 250 unsigned long do_mremap(unsigned long addr,
 251         unsigned long old_len, unsigned long new_len,
 252         unsigned long flags, unsigned long new_addr)
 253 {
 254         struct vm_area_struct *vma;
 255         unsigned long ret = -EINVAL;
 256         unsigned long charged = 0;
 257
 258         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 259                 goto out;
 260
 261         if (addr & ~PAGE_MASK)
 262                 goto out;
 263
 264         old_len = PAGE_ALIGN(old_len);
 265         new_len = PAGE_ALIGN(new_len);
 266
 267         /*
 268          * We allow a zero old-len as a special case
 269          * for DOS-emu "duplicate shm area" thing. But
 270          * a zero new-len is nonsensical.
 271          */
 272         if (!new_len)
 273                 goto out;
 274
 275         /* new_addr is only valid if MREMAP_FIXED is specified */
 276         if (flags & MREMAP_FIXED) {
 277                 if (new_addr & ~PAGE_MASK)
 278                         goto out;
 279                 if (!(flags & MREMAP_MAYMOVE))
 280                         goto out;
 281
 282                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 283                         goto out;
 284
 285                 /* Check if the location we're moving into overlaps the
 286                  * old location at all, and fail if it does.
 287                  */
 288                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 289                         goto out;
 290
 291                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 292                         goto out;
 293
 294                 ret = do_munmap(current->mm, new_addr, new_len);
 295                 if (ret)
 296                         goto out;
 297         }
 298
 299         /*
 300          * Always allow a shrinking remap: that just unmaps
 301          * the unnecessary pages..
 302          * do_munmap does all the needed commit accounting
 303          */
 304         if (old_len >= new_len) {
 305                 ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
 306                 if (ret && old_len != new_len)
 307                         goto out;
 308                 ret = addr;
 309                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 310                         goto out;
 311                 old_len = new_len;
 312         }
 313
 314         /*
 315          * Ok, we need to grow..  or relocate.
 316          */
 317         ret = -EFAULT;
 318         vma = find_vma(current->mm, addr);
 319         if (!vma || vma->vm_start > addr)
 320                 goto out;
 321         if (is_vm_hugetlb_page(vma)) {
 322                 ret = -EINVAL;
 323                 goto out;
 324         }
 325         /* We can't remap across vm area boundaries */
 326         if (old_len > vma->vm_end - addr)
 327                 goto out;
 328         if (vma->vm_flags & VM_DONTEXPAND) {
 329                 if (new_len > old_len)
 330                         goto out;
 331         }
 332         if (vma->vm_flags & VM_LOCKED) {
 333                 unsigned long locked, lock_limit;
 334                 locked = current->mm->locked_vm << PAGE_SHIFT;
 335                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 336                 locked += new_len - old_len;
 337                 ret = -EAGAIN;
 338                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 339                         goto out;
 340                 ret = -ENOMEM;
 341                 if (!vx_vmlocked_avail(current->mm,
 342                         (new_len - old_len) >> PAGE_SHIFT))
 343                         goto out;
 344         }
 345         ret = -ENOMEM;
 346         if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
 347             > current->signal->rlim[RLIMIT_AS].rlim_cur)
 348                 goto out;
 349         /* check context space, maybe only Private writable mapping? */
 350         if (!vx_vmpages_avail(current->mm, (new_len - old_len) >> PAGE_SHIFT))
 351                 goto out;
 352
 353         if (vma->vm_flags & VM_ACCOUNT) {
 354                 charged = (new_len - old_len) >> PAGE_SHIFT;
 355                 if (security_vm_enough_memory(charged))
 356                         goto out_nc;
 357         }
 358
 359         /* old_len exactly to the end of the area..
 360          * And we're not relocating the area.
 361          */
 362         if (old_len == vma->vm_end - addr &&
 363             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 364             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 365                 unsigned long max_addr = TASK_SIZE;
 366                 if (vma->vm_next)
 367                         max_addr = vma->vm_next->vm_start;
 368                 /* can we just expand the current mapping? */
 369                 if (max_addr - addr >= new_len) {
 370                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 371
 372                         vma_adjust(vma, vma->vm_start,
 373                                 addr + new_len, vma->vm_pgoff, NULL);
 374
 375                         // current->mm->total_vm += pages;
 376                         vx_vmpages_add(current->mm, pages);
 377                         __vm_stat_account(vma->vm_mm, vma->vm_flags,
 378                                                         vma->vm_file, pages);
 379                         if (vma->vm_flags & VM_LOCKED) {
 380                                 // current->mm->locked_vm += pages;
 381                                 vx_vmlocked_add(vma->vm_mm, pages);
 382                                 make_pages_present(addr + old_len,
 383                                                    addr + new_len);
 384                         }
 385                         ret = addr;
 386                         goto out;
 387                 }
 388         }
 389
 390         /*
 391          * We weren't able to just expand or shrink the area,
 392          * we need to create a new one and move it..
 393          */
 394         ret = -ENOMEM;
 395         if (flags & MREMAP_MAYMOVE) {
 396                 if (!(flags & MREMAP_FIXED)) {
 397                         unsigned long map_flags = 0;
 398                         if (vma->vm_flags & VM_MAYSHARE)
 399                                 map_flags |= MAP_SHARED;
 400
 401                         new_addr = get_unmapped_area_prot(vma->vm_file, 0, new_len,
 402                                 vma->vm_pgoff, map_flags, vma->vm_flags & VM_EXEC);
 403                         ret = new_addr;
 404                         if (new_addr & ~PAGE_MASK)
 405                                 goto out;
 406                 }
 407                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 408         }
 409 out:
 410         if (ret & ~PAGE_MASK)
 411                 vm_unacct_memory(charged);
 412 out_nc:
 413         return ret;
 414 }
 415
 416 asmlinkage unsigned long sys_mremap(unsigned long addr,
 417         unsigned long old_len, unsigned long new_len,
 418         unsigned long flags, unsigned long new_addr)
 419 {
 420         unsigned long ret;
 421
 422         down_write(&current->mm->mmap_sem);
 423         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 424         up_write(&current->mm->mmap_sem);
 425         return ret;
 426 }