mm/mremap.c

   1 /*
   2  *      mm/mremap.c
   3  *
   4  *      (C) Copyright 1996 Linus Torvalds
   5  *
   6  *      Address space accounting code   <alan@lxorguk.ukuu.org.uk>
   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/shm.h>
  13 #include <linux/ksm.h>
  14 #include <linux/mman.h>
  15 #include <linux/swap.h>
  16 #include <linux/capability.h>
  17 #include <linux/fs.h>
  18 #include <linux/swapops.h>
  19 #include <linux/highmem.h>
  20 #include <linux/security.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/mmu_notifier.h>
  23 #include <linux/sched/sysctl.h>
  24 #include <linux/uaccess.h>
  25
  26 #include <asm/cacheflush.h>
  27 #include <asm/tlbflush.h>
  28
  29 #include "internal.h"
  30
  31 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
  32 {
  33         pgd_t *pgd;
  34         pud_t *pud;
  35         pmd_t *pmd;
  36
  37         pgd = pgd_offset(mm, addr);
  38         if (pgd_none_or_clear_bad(pgd))
  39                 return NULL;
  40
  41         pud = pud_offset(pgd, addr);
  42         if (pud_none_or_clear_bad(pud))
  43                 return NULL;
  44
  45         pmd = pmd_offset(pud, addr);
  46         if (pmd_none(*pmd))
  47                 return NULL;
  48
  49         return pmd;
  50 }
  51
  52 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
  53                             unsigned long addr)
  54 {
  55         pgd_t *pgd;
  56         pud_t *pud;
  57         pmd_t *pmd;
  58
  59         pgd = pgd_offset(mm, addr);
  60         pud = pud_alloc(mm, pgd, addr);
  61         if (!pud)
  62                 return NULL;
  63
  64         pmd = pmd_alloc(mm, pud, addr);
  65         if (!pmd)
  66                 return NULL;
  67
  68         VM_BUG_ON(pmd_trans_huge(*pmd));
  69
  70         return pmd;
  71 }
  72
  73 static pte_t move_soft_dirty_pte(pte_t pte)
  74 {
  75         /*
  76          * Set soft dirty bit so we can notice
  77          * in userspace the ptes were moved.
  78          */
  79 #ifdef CONFIG_MEM_SOFT_DIRTY
  80         if (pte_present(pte))
  81                 pte = pte_mksoft_dirty(pte);
  82         else if (is_swap_pte(pte))
  83                 pte = pte_swp_mksoft_dirty(pte);
  84 #endif
  85         return pte;
  86 }
  87
  88 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
  89                 unsigned long old_addr, unsigned long old_end,
  90                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
  91                 unsigned long new_addr, bool need_rmap_locks)
  92 {
  93         struct address_space *mapping = NULL;
  94         struct anon_vma *anon_vma = NULL;
  95         struct mm_struct *mm = vma->vm_mm;
  96         pte_t *old_pte, *new_pte, pte;
  97         spinlock_t *old_ptl, *new_ptl;
  98
  99         /*
 100          * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
 101          * locks to ensure that rmap will always observe either the old or the
 102          * new ptes. This is the easiest way to avoid races with
 103          * truncate_pagecache(), page migration, etc...
 104          *
 105          * When need_rmap_locks is false, we use other ways to avoid
 106          * such races:
 107          *
 108          * - During exec() shift_arg_pages(), we use a specially tagged vma
 109          *   which rmap call sites look for using is_vma_temporary_stack().
 110          *
 111          * - During mremap(), new_vma is often known to be placed after vma
 112          *   in rmap traversal order. This ensures rmap will always observe
 113          *   either the old pte, or the new pte, or both (the page table locks
 114          *   serialize access to individual ptes, but only rmap traversal
 115          *   order guarantees that we won't miss both the old and new ptes).
 116          */
 117         if (need_rmap_locks) {
 118                 if (vma->vm_file) {
 119                         mapping = vma->vm_file->f_mapping;
 120                         i_mmap_lock_write(mapping);
 121                 }
 122                 if (vma->anon_vma) {
 123                         anon_vma = vma->anon_vma;
 124                         anon_vma_lock_write(anon_vma);
 125                 }
 126         }
 127
 128         /*
 129          * We don't have to worry about the ordering of src and dst
 130          * pte locks because exclusive mmap_sem prevents deadlock.
 131          */
 132         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 133         new_pte = pte_offset_map(new_pmd, new_addr);
 134         new_ptl = pte_lockptr(mm, new_pmd);
 135         if (new_ptl != old_ptl)
 136                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 137         arch_enter_lazy_mmu_mode();
 138
 139         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 140                                    new_pte++, new_addr += PAGE_SIZE) {
 141                 if (pte_none(*old_pte))
 142                         continue;
 143                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
 144                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 145                 pte = move_soft_dirty_pte(pte);
 146                 set_pte_at(mm, new_addr, new_pte, pte);
 147         }
 148
 149         arch_leave_lazy_mmu_mode();
 150         if (new_ptl != old_ptl)
 151                 spin_unlock(new_ptl);
 152         pte_unmap(new_pte - 1);
 153         pte_unmap_unlock(old_pte - 1, old_ptl);
 154         if (anon_vma)
 155                 anon_vma_unlock_write(anon_vma);
 156         if (mapping)
 157                 i_mmap_unlock_write(mapping);
 158 }
 159
 160 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
 161
 162 unsigned long move_page_tables(struct vm_area_struct *vma,
 163                 unsigned long old_addr, struct vm_area_struct *new_vma,
 164                 unsigned long new_addr, unsigned long len,
 165                 bool need_rmap_locks)
 166 {
 167         unsigned long extent, next, old_end;
 168         pmd_t *old_pmd, *new_pmd;
 169         bool need_flush = false;
 170         unsigned long mmun_start;       /* For mmu_notifiers */
 171         unsigned long mmun_end;         /* For mmu_notifiers */
 172
 173         old_end = old_addr + len;
 174         flush_cache_range(vma, old_addr, old_end);
 175
 176         mmun_start = old_addr;
 177         mmun_end   = old_end;
 178         mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
 179
 180         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 181                 cond_resched();
 182                 next = (old_addr + PMD_SIZE) & PMD_MASK;
 183                 /* even if next overflowed, extent below will be ok */
 184                 extent = next - old_addr;
 185                 if (extent > old_end - old_addr)
 186                         extent = old_end - old_addr;
 187                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 188                 if (!old_pmd)
 189                         continue;
 190                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 191                 if (!new_pmd)
 192                         break;
 193                 if (pmd_trans_huge(*old_pmd)) {
 194                         int err = 0;
 195                         if (extent == HPAGE_PMD_SIZE) {
 196                                 VM_BUG_ON_VMA(vma->vm_file || !vma->anon_vma,
 197                                               vma);
 198                                 /* See comment in move_ptes() */
 199                                 if (need_rmap_locks)
 200                                         anon_vma_lock_write(vma->anon_vma);
 201                                 err = move_huge_pmd(vma, new_vma, old_addr,
 202                                                     new_addr, old_end,
 203                                                     old_pmd, new_pmd);
 204                                 if (need_rmap_locks)
 205                                         anon_vma_unlock_write(vma->anon_vma);
 206                         }
 207                         if (err > 0) {
 208                                 need_flush = true;
 209                                 continue;
 210                         } else if (!err) {
 211                                 split_huge_page_pmd(vma, old_addr, old_pmd);
 212                         }
 213                         VM_BUG_ON(pmd_trans_huge(*old_pmd));
 214                 }
 215                 if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
 216                                                       new_pmd, new_addr))
 217                         break;
 218                 next = (new_addr + PMD_SIZE) & PMD_MASK;
 219                 if (extent > next - new_addr)
 220                         extent = next - new_addr;
 221                 if (extent > LATENCY_LIMIT)
 222                         extent = LATENCY_LIMIT;
 223                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
 224                           new_vma, new_pmd, new_addr, need_rmap_locks);
 225                 need_flush = true;
 226         }
 227         if (likely(need_flush))
 228                 flush_tlb_range(vma, old_end-len, old_addr);
 229
 230         mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
 231
 232         return len + old_addr - old_end;        /* how much done */
 233 }
 234
 235 static unsigned long move_vma(struct vm_area_struct *vma,
 236                 unsigned long old_addr, unsigned long old_len,
 237                 unsigned long new_len, unsigned long new_addr, bool *locked)
 238 {
 239         struct mm_struct *mm = vma->vm_mm;
 240         struct vm_area_struct *new_vma;
 241         unsigned long vm_flags = vma->vm_flags;
 242         unsigned long new_pgoff;
 243         unsigned long moved_len;
 244         unsigned long excess = 0;
 245         unsigned long hiwater_vm;
 246         int split = 0;
 247         int err;
 248         bool need_rmap_locks;
 249
 250         /*
 251          * We'd prefer to avoid failure later on in do_munmap:
 252          * which may split one vma into three before unmapping.
 253          */
 254         if (mm->map_count >= sysctl_max_map_count - 3)
 255                 return -ENOMEM;
 256
 257         /*
 258          * Advise KSM to break any KSM pages in the area to be moved:
 259          * it would be confusing if they were to turn up at the new
 260          * location, where they happen to coincide with different KSM
 261          * pages recently unmapped.  But leave vma->vm_flags as it was,
 262          * so KSM can come around to merge on vma and new_vma afterwards.
 263          */
 264         err = ksm_madvise(vma, old_addr, old_addr + old_len,
 265                                                 MADV_UNMERGEABLE, &vm_flags);
 266         if (err)
 267                 return err;
 268
 269         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 270         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 271                            &need_rmap_locks);
 272         if (!new_vma)
 273                 return -ENOMEM;
 274
 275         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 276                                      need_rmap_locks);
 277         if (moved_len < old_len) {
 278                 /*
 279                  * On error, move entries back from new area to old,
 280                  * which will succeed since page tables still there,
 281                  * and then proceed to unmap new area instead of old.
 282                  */
 283                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 284                                  true);
 285                 vma = new_vma;
 286                 old_len = new_len;
 287                 old_addr = new_addr;
 288                 new_addr = -ENOMEM;
 289         } else if (vma->vm_file && vma->vm_file->f_op->mremap) {
 290                 err = vma->vm_file->f_op->mremap(vma->vm_file, new_vma);
 291                 if (err < 0) {
 292                         move_page_tables(new_vma, new_addr, vma, old_addr,
 293                                          moved_len, true);
 294                         return err;
 295                 }
 296         }
 297
 298         /* Conceal VM_ACCOUNT so old reservation is not undone */
 299         if (vm_flags & VM_ACCOUNT) {
 300                 vma->vm_flags &= ~VM_ACCOUNT;
 301                 excess = vma->vm_end - vma->vm_start - old_len;
 302                 if (old_addr > vma->vm_start &&
 303                     old_addr + old_len < vma->vm_end)
 304                         split = 1;
 305         }
 306
 307         /*
 308          * If we failed to move page tables we still do total_vm increment
 309          * since do_munmap() will decrement it by old_len == new_len.
 310          *
 311          * Since total_vm is about to be raised artificially high for a
 312          * moment, we need to restore high watermark afterwards: if stats
 313          * are taken meanwhile, total_vm and hiwater_vm appear too high.
 314          * If this were a serious issue, we'd add a flag to do_munmap().
 315          */
 316         hiwater_vm = mm->hiwater_vm;
 317         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 318
 319         if (do_munmap(mm, old_addr, old_len) < 0) {
 320                 /* OOM: unable to split vma, just get accounts right */
 321                 vm_unacct_memory(excess >> PAGE_SHIFT);
 322                 excess = 0;
 323         }
 324         mm->hiwater_vm = hiwater_vm;
 325
 326         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 327         if (excess) {
 328                 vma->vm_flags |= VM_ACCOUNT;
 329                 if (split)
 330                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 331         }
 332
 333         if (vm_flags & VM_LOCKED) {
 334                 mm->locked_vm += new_len >> PAGE_SHIFT;
 335                 *locked = true;
 336         }
 337
 338         return new_addr;
 339 }
 340
 341 static struct vm_area_struct *vma_to_resize(unsigned long addr,
 342         unsigned long old_len, unsigned long new_len, unsigned long *p)
 343 {
 344         struct mm_struct *mm = current->mm;
 345         struct vm_area_struct *vma = find_vma(mm, addr);
 346
 347         if (!vma || vma->vm_start > addr)
 348                 goto Efault;
 349
 350         if (is_vm_hugetlb_page(vma))
 351                 goto Einval;
 352
 353         /* We can't remap across vm area boundaries */
 354         if (old_len > vma->vm_end - addr)
 355                 goto Efault;
 356
 357         /* Need to be careful about a growing mapping */
 358         if (new_len > old_len) {
 359                 unsigned long pgoff;
 360
 361                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 362                         goto Efault;
 363                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 364                 pgoff += vma->vm_pgoff;
 365                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 366                         goto Einval;
 367         }
 368
 369         if (vma->vm_flags & VM_LOCKED) {
 370                 unsigned long locked, lock_limit;
 371                 locked = mm->locked_vm << PAGE_SHIFT;
 372                 lock_limit = rlimit(RLIMIT_MEMLOCK);
 373                 locked += new_len - old_len;
 374                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 375                         goto Eagain;
 376         }
 377
 378         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
 379                 goto Enomem;
 380
 381         if (vma->vm_flags & VM_ACCOUNT) {
 382                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 383                 if (security_vm_enough_memory_mm(mm, charged))
 384                         goto Efault;
 385                 *p = charged;
 386         }
 387
 388         return vma;
 389
 390 Efault: /* very odd choice for most of the cases, but... */
 391         return ERR_PTR(-EFAULT);
 392 Einval:
 393         return ERR_PTR(-EINVAL);
 394 Enomem:
 395         return ERR_PTR(-ENOMEM);
 396 Eagain:
 397         return ERR_PTR(-EAGAIN);
 398 }
 399
 400 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 401                 unsigned long new_addr, unsigned long new_len, bool *locked)
 402 {
 403         struct mm_struct *mm = current->mm;
 404         struct vm_area_struct *vma;
 405         unsigned long ret = -EINVAL;
 406         unsigned long charged = 0;
 407         unsigned long map_flags;
 408
 409         if (new_addr & ~PAGE_MASK)
 410                 goto out;
 411
 412         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 413                 goto out;
 414
 415         /* Check if the location we're moving into overlaps the
 416          * old location at all, and fail if it does.
 417          */
 418         if ((new_addr <= addr) && (new_addr+new_len) > addr)
 419                 goto out;
 420
 421         if ((addr <= new_addr) && (addr+old_len) > new_addr)
 422                 goto out;
 423
 424         ret = do_munmap(mm, new_addr, new_len);
 425         if (ret)
 426                 goto out;
 427
 428         if (old_len >= new_len) {
 429                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 430                 if (ret && old_len != new_len)
 431                         goto out;
 432                 old_len = new_len;
 433         }
 434
 435         vma = vma_to_resize(addr, old_len, new_len, &charged);
 436         if (IS_ERR(vma)) {
 437                 ret = PTR_ERR(vma);
 438                 goto out;
 439         }
 440
 441         map_flags = MAP_FIXED;
 442         if (vma->vm_flags & VM_MAYSHARE)
 443                 map_flags |= MAP_SHARED;
 444
 445         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 446                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
 447                                 map_flags);
 448         if (ret & ~PAGE_MASK)
 449                 goto out1;
 450
 451         ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
 452         if (!(ret & ~PAGE_MASK))
 453                 goto out;
 454 out1:
 455         vm_unacct_memory(charged);
 456
 457 out:
 458         return ret;
 459 }
 460
 461 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 462 {
 463         unsigned long end = vma->vm_end + delta;
 464         if (end < vma->vm_end) /* overflow */
 465                 return 0;
 466         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 467                 return 0;
 468         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 469                               0, MAP_FIXED) & ~PAGE_MASK)
 470                 return 0;
 471         return 1;
 472 }
 473
 474 /*
 475  * Expand (or shrink) an existing mapping, potentially moving it at the
 476  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 477  *
 478  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 479  * This option implies MREMAP_MAYMOVE.
 480  */
 481 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 482                 unsigned long, new_len, unsigned long, flags,
 483                 unsigned long, new_addr)
 484 {
 485         struct mm_struct *mm = current->mm;
 486         struct vm_area_struct *vma;
 487         unsigned long ret = -EINVAL;
 488         unsigned long charged = 0;
 489         bool locked = false;
 490
 491         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 492                 return ret;
 493
 494         if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 495                 return ret;
 496
 497         if (addr & ~PAGE_MASK)
 498                 return ret;
 499
 500         old_len = PAGE_ALIGN(old_len);
 501         new_len = PAGE_ALIGN(new_len);
 502
 503         /*
 504          * We allow a zero old-len as a special case
 505          * for DOS-emu "duplicate shm area" thing. But
 506          * a zero new-len is nonsensical.
 507          */
 508         if (!new_len)
 509                 return ret;
 510
 511         down_write(&current->mm->mmap_sem);
 512
 513         if (flags & MREMAP_FIXED) {
 514                 ret = mremap_to(addr, old_len, new_addr, new_len,
 515                                 &locked);
 516                 goto out;
 517         }
 518
 519         /*
 520          * Always allow a shrinking remap: that just unmaps
 521          * the unnecessary pages..
 522          * do_munmap does all the needed commit accounting
 523          */
 524         if (old_len >= new_len) {
 525                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
 526                 if (ret && old_len != new_len)
 527                         goto out;
 528                 ret = addr;
 529                 goto out;
 530         }
 531
 532         /*
 533          * Ok, we need to grow..
 534          */
 535         vma = vma_to_resize(addr, old_len, new_len, &charged);
 536         if (IS_ERR(vma)) {
 537                 ret = PTR_ERR(vma);
 538                 goto out;
 539         }
 540
 541         /* old_len exactly to the end of the area..
 542          */
 543         if (old_len == vma->vm_end - addr) {
 544                 /* can we just expand the current mapping? */
 545                 if (vma_expandable(vma, new_len - old_len)) {
 546                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 547
 548                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
 549                                        vma->vm_pgoff, NULL)) {
 550                                 ret = -ENOMEM;
 551                                 goto out;
 552                         }
 553
 554                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
 555                         if (vma->vm_flags & VM_LOCKED) {
 556                                 mm->locked_vm += pages;
 557                                 locked = true;
 558                                 new_addr = addr;
 559                         }
 560                         ret = addr;
 561                         goto out;
 562                 }
 563         }
 564
 565         /*
 566          * We weren't able to just expand or shrink the area,
 567          * we need to create a new one and move it..
 568          */
 569         ret = -ENOMEM;
 570         if (flags & MREMAP_MAYMOVE) {
 571                 unsigned long map_flags = 0;
 572                 if (vma->vm_flags & VM_MAYSHARE)
 573                         map_flags |= MAP_SHARED;
 574
 575                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 576                                         vma->vm_pgoff +
 577                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
 578                                         map_flags);
 579                 if (new_addr & ~PAGE_MASK) {
 580                         ret = new_addr;
 581                         goto out;
 582                 }
 583
 584                 ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
 585         }
 586 out:
 587         if (ret & ~PAGE_MASK)
 588                 vm_unacct_memory(charged);
 589         up_write(&current->mm->mmap_sem);
 590         if (locked && new_len > old_len)
 591                 mm_populate(new_addr + old_len, new_len - old_len);
 592         return ret;
 593 }