Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * mm/mremap.c
4 *
5 * (C) Copyright 1996 Linus Torvalds
6 *
7 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
8 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
9 */
10
11#include <linux/mm.h>
12#include <linux/hugetlb.h>
13#include <linux/shm.h>
14#include <linux/ksm.h>
15#include <linux/mman.h>
16#include <linux/swap.h>
17#include <linux/capability.h>
18#include <linux/fs.h>
19#include <linux/swapops.h>
20#include <linux/highmem.h>
21#include <linux/security.h>
22#include <linux/syscalls.h>
23#include <linux/mmu_notifier.h>
24#include <linux/uaccess.h>
25#include <linux/mm-arch-hooks.h>
26#include <linux/userfaultfd_k.h>
27
28#include <asm/cacheflush.h>
29#include <asm/tlbflush.h>
30
31#include "internal.h"
32
33static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
34{
35 pgd_t *pgd;
36 p4d_t *p4d;
37 pud_t *pud;
38 pmd_t *pmd;
39
40 pgd = pgd_offset(mm, addr);
41 if (pgd_none_or_clear_bad(pgd))
42 return NULL;
43
44 p4d = p4d_offset(pgd, addr);
45 if (p4d_none_or_clear_bad(p4d))
46 return NULL;
47
48 pud = pud_offset(p4d, addr);
49 if (pud_none_or_clear_bad(pud))
50 return NULL;
51
52 pmd = pmd_offset(pud, addr);
53 if (pmd_none(*pmd))
54 return NULL;
55
56 return pmd;
57}
58
59static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
60 unsigned long addr)
61{
62 pgd_t *pgd;
63 p4d_t *p4d;
64 pud_t *pud;
65 pmd_t *pmd;
66
67 pgd = pgd_offset(mm, addr);
68 p4d = p4d_alloc(mm, pgd, addr);
69 if (!p4d)
70 return NULL;
71 pud = pud_alloc(mm, p4d, addr);
72 if (!pud)
73 return NULL;
74
75 pmd = pmd_alloc(mm, pud, addr);
76 if (!pmd)
77 return NULL;
78
79 VM_BUG_ON(pmd_trans_huge(*pmd));
80
81 return pmd;
82}
83
84static void take_rmap_locks(struct vm_area_struct *vma)
85{
86 if (vma->vm_file)
87 i_mmap_lock_write(vma->vm_file->f_mapping);
88 if (vma->anon_vma)
89 anon_vma_lock_write(vma->anon_vma);
90}
91
92static void drop_rmap_locks(struct vm_area_struct *vma)
93{
94 if (vma->anon_vma)
95 anon_vma_unlock_write(vma->anon_vma);
96 if (vma->vm_file)
97 i_mmap_unlock_write(vma->vm_file->f_mapping);
98}
99
100static pte_t move_soft_dirty_pte(pte_t pte)
101{
102 /*
103 * Set soft dirty bit so we can notice
104 * in userspace the ptes were moved.
105 */
106#ifdef CONFIG_MEM_SOFT_DIRTY
107 if (pte_present(pte))
108 pte = pte_mksoft_dirty(pte);
109 else if (is_swap_pte(pte))
110 pte = pte_swp_mksoft_dirty(pte);
111#endif
112 return pte;
113}
114
115static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
116 unsigned long old_addr, unsigned long old_end,
117 struct vm_area_struct *new_vma, pmd_t *new_pmd,
118 unsigned long new_addr, bool need_rmap_locks)
119{
120 struct mm_struct *mm = vma->vm_mm;
121 pte_t *old_pte, *new_pte, pte;
122 spinlock_t *old_ptl, *new_ptl;
123 bool force_flush = false;
124 unsigned long len = old_end - old_addr;
125
126 /*
127 * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
128 * locks to ensure that rmap will always observe either the old or the
129 * new ptes. This is the easiest way to avoid races with
130 * truncate_pagecache(), page migration, etc...
131 *
132 * When need_rmap_locks is false, we use other ways to avoid
133 * such races:
134 *
135 * - During exec() shift_arg_pages(), we use a specially tagged vma
136 * which rmap call sites look for using is_vma_temporary_stack().
137 *
138 * - During mremap(), new_vma is often known to be placed after vma
139 * in rmap traversal order. This ensures rmap will always observe
140 * either the old pte, or the new pte, or both (the page table locks
141 * serialize access to individual ptes, but only rmap traversal
142 * order guarantees that we won't miss both the old and new ptes).
143 */
144 if (need_rmap_locks)
145 take_rmap_locks(vma);
146
147 /*
148 * We don't have to worry about the ordering of src and dst
149 * pte locks because exclusive mmap_sem prevents deadlock.
150 */
151 old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
152 new_pte = pte_offset_map(new_pmd, new_addr);
153 new_ptl = pte_lockptr(mm, new_pmd);
154 if (new_ptl != old_ptl)
155 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
156 flush_tlb_batched_pending(vma->vm_mm);
157 arch_enter_lazy_mmu_mode();
158
159 for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
160 new_pte++, new_addr += PAGE_SIZE) {
161 if (pte_none(*old_pte))
162 continue;
163
164 pte = ptep_get_and_clear(mm, old_addr, old_pte);
165 /*
166 * If we are remapping a valid PTE, make sure
167 * to flush TLB before we drop the PTL for the
168 * PTE.
169 *
170 * NOTE! Both old and new PTL matter: the old one
171 * for racing with page_mkclean(), the new one to
172 * make sure the physical page stays valid until
173 * the TLB entry for the old mapping has been
174 * flushed.
175 */
176 if (pte_present(pte))
177 force_flush = true;
178 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
179 pte = move_soft_dirty_pte(pte);
180 set_pte_at(mm, new_addr, new_pte, pte);
181 }
182
183 arch_leave_lazy_mmu_mode();
184 if (force_flush)
185 flush_tlb_range(vma, old_end - len, old_end);
186 if (new_ptl != old_ptl)
187 spin_unlock(new_ptl);
188 pte_unmap(new_pte - 1);
189 pte_unmap_unlock(old_pte - 1, old_ptl);
190 if (need_rmap_locks)
191 drop_rmap_locks(vma);
192}
193
194#ifdef CONFIG_HAVE_MOVE_PMD
195static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr,
196 unsigned long new_addr, unsigned long old_end,
197 pmd_t *old_pmd, pmd_t *new_pmd)
198{
199 spinlock_t *old_ptl, *new_ptl;
200 struct mm_struct *mm = vma->vm_mm;
201 pmd_t pmd;
202
203 if ((old_addr & ~PMD_MASK) || (new_addr & ~PMD_MASK)
204 || old_end - old_addr < PMD_SIZE)
205 return false;
206
207 /*
208 * The destination pmd shouldn't be established, free_pgtables()
209 * should have release it.
210 */
211 if (WARN_ON(!pmd_none(*new_pmd)))
212 return false;
213
214 /*
215 * We don't have to worry about the ordering of src and dst
216 * ptlocks because exclusive mmap_sem prevents deadlock.
217 */
218 old_ptl = pmd_lock(vma->vm_mm, old_pmd);
219 new_ptl = pmd_lockptr(mm, new_pmd);
220 if (new_ptl != old_ptl)
221 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
222
223 /* Clear the pmd */
224 pmd = *old_pmd;
225 pmd_clear(old_pmd);
226
227 VM_BUG_ON(!pmd_none(*new_pmd));
228
229 /* Set the new pmd */
230 set_pmd_at(mm, new_addr, new_pmd, pmd);
231 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
232 if (new_ptl != old_ptl)
233 spin_unlock(new_ptl);
234 spin_unlock(old_ptl);
235
236 return true;
237}
238#endif
239
240unsigned long move_page_tables(struct vm_area_struct *vma,
241 unsigned long old_addr, struct vm_area_struct *new_vma,
242 unsigned long new_addr, unsigned long len,
243 bool need_rmap_locks)
244{
245 unsigned long extent, next, old_end;
246 struct mmu_notifier_range range;
247 pmd_t *old_pmd, *new_pmd;
248
249 old_end = old_addr + len;
250 flush_cache_range(vma, old_addr, old_end);
251
252 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
253 old_addr, old_end);
254 mmu_notifier_invalidate_range_start(&range);
255
256 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
257 cond_resched();
258 next = (old_addr + PMD_SIZE) & PMD_MASK;
259 /* even if next overflowed, extent below will be ok */
260 extent = next - old_addr;
261 if (extent > old_end - old_addr)
262 extent = old_end - old_addr;
263 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
264 if (!old_pmd)
265 continue;
266 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
267 if (!new_pmd)
268 break;
269 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd)) {
270 if (extent == HPAGE_PMD_SIZE) {
271 bool moved;
272 /* See comment in move_ptes() */
273 if (need_rmap_locks)
274 take_rmap_locks(vma);
275 moved = move_huge_pmd(vma, old_addr, new_addr,
276 old_end, old_pmd, new_pmd);
277 if (need_rmap_locks)
278 drop_rmap_locks(vma);
279 if (moved)
280 continue;
281 }
282 split_huge_pmd(vma, old_pmd, old_addr);
283 if (pmd_trans_unstable(old_pmd))
284 continue;
285 } else if (extent == PMD_SIZE) {
286#ifdef CONFIG_HAVE_MOVE_PMD
287 /*
288 * If the extent is PMD-sized, try to speed the move by
289 * moving at the PMD level if possible.
290 */
291 bool moved;
292
293 if (need_rmap_locks)
294 take_rmap_locks(vma);
295 moved = move_normal_pmd(vma, old_addr, new_addr,
296 old_end, old_pmd, new_pmd);
297 if (need_rmap_locks)
298 drop_rmap_locks(vma);
299 if (moved)
300 continue;
301#endif
302 }
303
304 if (pte_alloc(new_vma->vm_mm, new_pmd))
305 break;
306 next = (new_addr + PMD_SIZE) & PMD_MASK;
307 if (extent > next - new_addr)
308 extent = next - new_addr;
309 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
310 new_pmd, new_addr, need_rmap_locks);
311 }
312
313 mmu_notifier_invalidate_range_end(&range);
314
315 return len + old_addr - old_end; /* how much done */
316}
317
318static unsigned long move_vma(struct vm_area_struct *vma,
319 unsigned long old_addr, unsigned long old_len,
320 unsigned long new_len, unsigned long new_addr,
321 bool *locked, struct vm_userfaultfd_ctx *uf,
322 struct list_head *uf_unmap)
323{
324 struct mm_struct *mm = vma->vm_mm;
325 struct vm_area_struct *new_vma;
326 unsigned long vm_flags = vma->vm_flags;
327 unsigned long new_pgoff;
328 unsigned long moved_len;
329 unsigned long excess = 0;
330 unsigned long hiwater_vm;
331 int split = 0;
332 int err;
333 bool need_rmap_locks;
334
335 /*
336 * We'd prefer to avoid failure later on in do_munmap:
337 * which may split one vma into three before unmapping.
338 */
339 if (mm->map_count >= sysctl_max_map_count - 3)
340 return -ENOMEM;
341
342 /*
343 * Advise KSM to break any KSM pages in the area to be moved:
344 * it would be confusing if they were to turn up at the new
345 * location, where they happen to coincide with different KSM
346 * pages recently unmapped. But leave vma->vm_flags as it was,
347 * so KSM can come around to merge on vma and new_vma afterwards.
348 */
349 err = ksm_madvise(vma, old_addr, old_addr + old_len,
350 MADV_UNMERGEABLE, &vm_flags);
351 if (err)
352 return err;
353
354 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
355 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
356 &need_rmap_locks);
357 if (!new_vma)
358 return -ENOMEM;
359
360 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
361 need_rmap_locks);
362 if (moved_len < old_len) {
363 err = -ENOMEM;
364 } else if (vma->vm_ops && vma->vm_ops->mremap) {
365 err = vma->vm_ops->mremap(new_vma);
366 }
367
368 if (unlikely(err)) {
369 /*
370 * On error, move entries back from new area to old,
371 * which will succeed since page tables still there,
372 * and then proceed to unmap new area instead of old.
373 */
374 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
375 true);
376 vma = new_vma;
377 old_len = new_len;
378 old_addr = new_addr;
379 new_addr = err;
380 } else {
381 mremap_userfaultfd_prep(new_vma, uf);
382 arch_remap(mm, old_addr, old_addr + old_len,
383 new_addr, new_addr + new_len);
384 }
385
386 /* Conceal VM_ACCOUNT so old reservation is not undone */
387 if (vm_flags & VM_ACCOUNT) {
388 vma->vm_flags &= ~VM_ACCOUNT;
389 excess = vma->vm_end - vma->vm_start - old_len;
390 if (old_addr > vma->vm_start &&
391 old_addr + old_len < vma->vm_end)
392 split = 1;
393 }
394
395 /*
396 * If we failed to move page tables we still do total_vm increment
397 * since do_munmap() will decrement it by old_len == new_len.
398 *
399 * Since total_vm is about to be raised artificially high for a
400 * moment, we need to restore high watermark afterwards: if stats
401 * are taken meanwhile, total_vm and hiwater_vm appear too high.
402 * If this were a serious issue, we'd add a flag to do_munmap().
403 */
404 hiwater_vm = mm->hiwater_vm;
405 vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
406
407 /* Tell pfnmap has moved from this vma */
408 if (unlikely(vma->vm_flags & VM_PFNMAP))
409 untrack_pfn_moved(vma);
410
411 if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
412 /* OOM: unable to split vma, just get accounts right */
413 vm_unacct_memory(excess >> PAGE_SHIFT);
414 excess = 0;
415 }
416 mm->hiwater_vm = hiwater_vm;
417
418 /* Restore VM_ACCOUNT if one or two pieces of vma left */
419 if (excess) {
420 vma->vm_flags |= VM_ACCOUNT;
421 if (split)
422 vma->vm_next->vm_flags |= VM_ACCOUNT;
423 }
424
425 if (vm_flags & VM_LOCKED) {
426 mm->locked_vm += new_len >> PAGE_SHIFT;
427 *locked = true;
428 }
429
430 return new_addr;
431}
432
433static struct vm_area_struct *vma_to_resize(unsigned long addr,
434 unsigned long old_len, unsigned long new_len, unsigned long *p)
435{
436 struct mm_struct *mm = current->mm;
437 struct vm_area_struct *vma = find_vma(mm, addr);
438 unsigned long pgoff;
439
440 if (!vma || vma->vm_start > addr)
441 return ERR_PTR(-EFAULT);
442
443 /*
444 * !old_len is a special case where an attempt is made to 'duplicate'
445 * a mapping. This makes no sense for private mappings as it will
446 * instead create a fresh/new mapping unrelated to the original. This
447 * is contrary to the basic idea of mremap which creates new mappings
448 * based on the original. There are no known use cases for this
449 * behavior. As a result, fail such attempts.
450 */
451 if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
452 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap. This is not supported.\n", current->comm, current->pid);
453 return ERR_PTR(-EINVAL);
454 }
455
456 if (is_vm_hugetlb_page(vma))
457 return ERR_PTR(-EINVAL);
458
459 /* We can't remap across vm area boundaries */
460 if (old_len > vma->vm_end - addr)
461 return ERR_PTR(-EFAULT);
462
463 if (new_len == old_len)
464 return vma;
465
466 /* Need to be careful about a growing mapping */
467 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
468 pgoff += vma->vm_pgoff;
469 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
470 return ERR_PTR(-EINVAL);
471
472 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
473 return ERR_PTR(-EFAULT);
474
475 if (vma->vm_flags & VM_LOCKED) {
476 unsigned long locked, lock_limit;
477 locked = mm->locked_vm << PAGE_SHIFT;
478 lock_limit = rlimit(RLIMIT_MEMLOCK);
479 locked += new_len - old_len;
480 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
481 return ERR_PTR(-EAGAIN);
482 }
483
484 if (!may_expand_vm(mm, vma->vm_flags,
485 (new_len - old_len) >> PAGE_SHIFT))
486 return ERR_PTR(-ENOMEM);
487
488 if (vma->vm_flags & VM_ACCOUNT) {
489 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
490 if (security_vm_enough_memory_mm(mm, charged))
491 return ERR_PTR(-ENOMEM);
492 *p = charged;
493 }
494
495 return vma;
496}
497
498static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
499 unsigned long new_addr, unsigned long new_len, bool *locked,
500 struct vm_userfaultfd_ctx *uf,
501 struct list_head *uf_unmap_early,
502 struct list_head *uf_unmap)
503{
504 struct mm_struct *mm = current->mm;
505 struct vm_area_struct *vma;
506 unsigned long ret = -EINVAL;
507 unsigned long charged = 0;
508 unsigned long map_flags;
509
510 if (offset_in_page(new_addr))
511 goto out;
512
513 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
514 goto out;
515
516 /* Ensure the old/new locations do not overlap */
517 if (addr + old_len > new_addr && new_addr + new_len > addr)
518 goto out;
519
520 /*
521 * move_vma() need us to stay 4 maps below the threshold, otherwise
522 * it will bail out at the very beginning.
523 * That is a problem if we have already unmaped the regions here
524 * (new_addr, and old_addr), because userspace will not know the
525 * state of the vma's after it gets -ENOMEM.
526 * So, to avoid such scenario we can pre-compute if the whole
527 * operation has high chances to success map-wise.
528 * Worst-scenario case is when both vma's (new_addr and old_addr) get
529 * split in 3 before unmaping it.
530 * That means 2 more maps (1 for each) to the ones we already hold.
531 * Check whether current map count plus 2 still leads us to 4 maps below
532 * the threshold, otherwise return -ENOMEM here to be more safe.
533 */
534 if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
535 return -ENOMEM;
536
537 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
538 if (ret)
539 goto out;
540
541 if (old_len >= new_len) {
542 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
543 if (ret && old_len != new_len)
544 goto out;
545 old_len = new_len;
546 }
547
548 vma = vma_to_resize(addr, old_len, new_len, &charged);
549 if (IS_ERR(vma)) {
550 ret = PTR_ERR(vma);
551 goto out;
552 }
553
554 map_flags = MAP_FIXED;
555 if (vma->vm_flags & VM_MAYSHARE)
556 map_flags |= MAP_SHARED;
557
558 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
559 ((addr - vma->vm_start) >> PAGE_SHIFT),
560 map_flags);
561 if (offset_in_page(ret))
562 goto out1;
563
564 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, uf,
565 uf_unmap);
566 if (!(offset_in_page(ret)))
567 goto out;
568out1:
569 vm_unacct_memory(charged);
570
571out:
572 return ret;
573}
574
575static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
576{
577 unsigned long end = vma->vm_end + delta;
578 if (end < vma->vm_end) /* overflow */
579 return 0;
580 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
581 return 0;
582 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
583 0, MAP_FIXED) & ~PAGE_MASK)
584 return 0;
585 return 1;
586}
587
588/*
589 * Expand (or shrink) an existing mapping, potentially moving it at the
590 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
591 *
592 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
593 * This option implies MREMAP_MAYMOVE.
594 */
595SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
596 unsigned long, new_len, unsigned long, flags,
597 unsigned long, new_addr)
598{
599 struct mm_struct *mm = current->mm;
600 struct vm_area_struct *vma;
601 unsigned long ret = -EINVAL;
602 unsigned long charged = 0;
603 bool locked = false;
604 bool downgraded = false;
605 struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
606 LIST_HEAD(uf_unmap_early);
607 LIST_HEAD(uf_unmap);
608
609 addr = untagged_addr(addr);
610 new_addr = untagged_addr(new_addr);
611
612 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
613 return ret;
614
615 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
616 return ret;
617
618 if (offset_in_page(addr))
619 return ret;
620
621 old_len = PAGE_ALIGN(old_len);
622 new_len = PAGE_ALIGN(new_len);
623
624 /*
625 * We allow a zero old-len as a special case
626 * for DOS-emu "duplicate shm area" thing. But
627 * a zero new-len is nonsensical.
628 */
629 if (!new_len)
630 return ret;
631
632 if (down_write_killable(¤t->mm->mmap_sem))
633 return -EINTR;
634
635 if (flags & MREMAP_FIXED) {
636 ret = mremap_to(addr, old_len, new_addr, new_len,
637 &locked, &uf, &uf_unmap_early, &uf_unmap);
638 goto out;
639 }
640
641 /*
642 * Always allow a shrinking remap: that just unmaps
643 * the unnecessary pages..
644 * __do_munmap does all the needed commit accounting, and
645 * downgrades mmap_sem to read if so directed.
646 */
647 if (old_len >= new_len) {
648 int retval;
649
650 retval = __do_munmap(mm, addr+new_len, old_len - new_len,
651 &uf_unmap, true);
652 if (retval < 0 && old_len != new_len) {
653 ret = retval;
654 goto out;
655 /* Returning 1 indicates mmap_sem is downgraded to read. */
656 } else if (retval == 1)
657 downgraded = true;
658 ret = addr;
659 goto out;
660 }
661
662 /*
663 * Ok, we need to grow..
664 */
665 vma = vma_to_resize(addr, old_len, new_len, &charged);
666 if (IS_ERR(vma)) {
667 ret = PTR_ERR(vma);
668 goto out;
669 }
670
671 /* old_len exactly to the end of the area..
672 */
673 if (old_len == vma->vm_end - addr) {
674 /* can we just expand the current mapping? */
675 if (vma_expandable(vma, new_len - old_len)) {
676 int pages = (new_len - old_len) >> PAGE_SHIFT;
677
678 if (vma_adjust(vma, vma->vm_start, addr + new_len,
679 vma->vm_pgoff, NULL)) {
680 ret = -ENOMEM;
681 goto out;
682 }
683
684 vm_stat_account(mm, vma->vm_flags, pages);
685 if (vma->vm_flags & VM_LOCKED) {
686 mm->locked_vm += pages;
687 locked = true;
688 new_addr = addr;
689 }
690 ret = addr;
691 goto out;
692 }
693 }
694
695 /*
696 * We weren't able to just expand or shrink the area,
697 * we need to create a new one and move it..
698 */
699 ret = -ENOMEM;
700 if (flags & MREMAP_MAYMOVE) {
701 unsigned long map_flags = 0;
702 if (vma->vm_flags & VM_MAYSHARE)
703 map_flags |= MAP_SHARED;
704
705 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
706 vma->vm_pgoff +
707 ((addr - vma->vm_start) >> PAGE_SHIFT),
708 map_flags);
709 if (offset_in_page(new_addr)) {
710 ret = new_addr;
711 goto out;
712 }
713
714 ret = move_vma(vma, addr, old_len, new_len, new_addr,
715 &locked, &uf, &uf_unmap);
716 }
717out:
718 if (offset_in_page(ret)) {
719 vm_unacct_memory(charged);
720 locked = 0;
721 }
722 if (downgraded)
723 up_read(¤t->mm->mmap_sem);
724 else
725 up_write(¤t->mm->mmap_sem);
726 if (locked && new_len > old_len)
727 mm_populate(new_addr + old_len, new_len - old_len);
728 userfaultfd_unmap_complete(mm, &uf_unmap_early);
729 mremap_userfaultfd_complete(&uf, addr, new_addr, old_len);
730 userfaultfd_unmap_complete(mm, &uf_unmap);
731 return ret;
732}
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/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
29static 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(*pmd))
45 return NULL;
46
47 return pmd;
48}
49
50static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
51 unsigned long addr)
52{
53 pgd_t *pgd;
54 pud_t *pud;
55 pmd_t *pmd;
56
57 pgd = pgd_offset(mm, addr);
58 pud = pud_alloc(mm, pgd, addr);
59 if (!pud)
60 return NULL;
61
62 pmd = pmd_alloc(mm, pud, addr);
63 if (!pmd)
64 return NULL;
65
66 VM_BUG_ON(pmd_trans_huge(*pmd));
67
68 return pmd;
69}
70
71static 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 truncate_pagecache
85 * out, 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 mutex_lock(&mapping->i_mmap_mutex);
90 }
91
92 /*
93 * We don't have to worry about the ordering of src and dst
94 * pte locks because exclusive mmap_sem prevents deadlock.
95 */
96 old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
97 new_pte = pte_offset_map(new_pmd, new_addr);
98 new_ptl = pte_lockptr(mm, new_pmd);
99 if (new_ptl != old_ptl)
100 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
101 arch_enter_lazy_mmu_mode();
102
103 for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
104 new_pte++, new_addr += PAGE_SIZE) {
105 if (pte_none(*old_pte))
106 continue;
107 pte = ptep_get_and_clear(mm, old_addr, old_pte);
108 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
109 set_pte_at(mm, new_addr, new_pte, pte);
110 }
111
112 arch_leave_lazy_mmu_mode();
113 if (new_ptl != old_ptl)
114 spin_unlock(new_ptl);
115 pte_unmap(new_pte - 1);
116 pte_unmap_unlock(old_pte - 1, old_ptl);
117 if (mapping)
118 mutex_unlock(&mapping->i_mmap_mutex);
119}
120
121#define LATENCY_LIMIT (64 * PAGE_SIZE)
122
123unsigned long move_page_tables(struct vm_area_struct *vma,
124 unsigned long old_addr, struct vm_area_struct *new_vma,
125 unsigned long new_addr, unsigned long len)
126{
127 unsigned long extent, next, old_end;
128 pmd_t *old_pmd, *new_pmd;
129 bool need_flush = false;
130
131 old_end = old_addr + len;
132 flush_cache_range(vma, old_addr, old_end);
133
134 mmu_notifier_invalidate_range_start(vma->vm_mm, old_addr, old_end);
135
136 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
137 cond_resched();
138 next = (old_addr + PMD_SIZE) & PMD_MASK;
139 /* even if next overflowed, extent below will be ok */
140 extent = next - old_addr;
141 if (extent > old_end - old_addr)
142 extent = old_end - old_addr;
143 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
144 if (!old_pmd)
145 continue;
146 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
147 if (!new_pmd)
148 break;
149 if (pmd_trans_huge(*old_pmd)) {
150 int err = 0;
151 if (extent == HPAGE_PMD_SIZE)
152 err = move_huge_pmd(vma, new_vma, old_addr,
153 new_addr, old_end,
154 old_pmd, new_pmd);
155 if (err > 0) {
156 need_flush = true;
157 continue;
158 } else if (!err) {
159 split_huge_page_pmd(vma->vm_mm, old_pmd);
160 }
161 VM_BUG_ON(pmd_trans_huge(*old_pmd));
162 }
163 if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
164 new_pmd, new_addr))
165 break;
166 next = (new_addr + PMD_SIZE) & PMD_MASK;
167 if (extent > next - new_addr)
168 extent = next - new_addr;
169 if (extent > LATENCY_LIMIT)
170 extent = LATENCY_LIMIT;
171 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
172 new_vma, new_pmd, new_addr);
173 need_flush = true;
174 }
175 if (likely(need_flush))
176 flush_tlb_range(vma, old_end-len, old_addr);
177
178 mmu_notifier_invalidate_range_end(vma->vm_mm, old_end-len, old_end);
179
180 return len + old_addr - old_end; /* how much done */
181}
182
183static unsigned long move_vma(struct vm_area_struct *vma,
184 unsigned long old_addr, unsigned long old_len,
185 unsigned long new_len, unsigned long new_addr)
186{
187 struct mm_struct *mm = vma->vm_mm;
188 struct vm_area_struct *new_vma;
189 unsigned long vm_flags = vma->vm_flags;
190 unsigned long new_pgoff;
191 unsigned long moved_len;
192 unsigned long excess = 0;
193 unsigned long hiwater_vm;
194 int split = 0;
195 int err;
196
197 /*
198 * We'd prefer to avoid failure later on in do_munmap:
199 * which may split one vma into three before unmapping.
200 */
201 if (mm->map_count >= sysctl_max_map_count - 3)
202 return -ENOMEM;
203
204 /*
205 * Advise KSM to break any KSM pages in the area to be moved:
206 * it would be confusing if they were to turn up at the new
207 * location, where they happen to coincide with different KSM
208 * pages recently unmapped. But leave vma->vm_flags as it was,
209 * so KSM can come around to merge on vma and new_vma afterwards.
210 */
211 err = ksm_madvise(vma, old_addr, old_addr + old_len,
212 MADV_UNMERGEABLE, &vm_flags);
213 if (err)
214 return err;
215
216 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
217 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
218 if (!new_vma)
219 return -ENOMEM;
220
221 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
222 if (moved_len < old_len) {
223 /*
224 * Before moving the page tables from the new vma to
225 * the old vma, we need to be sure the old vma is
226 * queued after new vma in the same_anon_vma list to
227 * prevent SMP races with rmap_walk (that could lead
228 * rmap_walk to miss some page table).
229 */
230 anon_vma_moveto_tail(vma);
231
232 /*
233 * On error, move entries back from new area to old,
234 * which will succeed since page tables still there,
235 * and then proceed to unmap new area instead of old.
236 */
237 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
238 vma = new_vma;
239 old_len = new_len;
240 old_addr = new_addr;
241 new_addr = -ENOMEM;
242 }
243
244 /* Conceal VM_ACCOUNT so old reservation is not undone */
245 if (vm_flags & VM_ACCOUNT) {
246 vma->vm_flags &= ~VM_ACCOUNT;
247 excess = vma->vm_end - vma->vm_start - old_len;
248 if (old_addr > vma->vm_start &&
249 old_addr + old_len < vma->vm_end)
250 split = 1;
251 }
252
253 /*
254 * If we failed to move page tables we still do total_vm increment
255 * since do_munmap() will decrement it by old_len == new_len.
256 *
257 * Since total_vm is about to be raised artificially high for a
258 * moment, we need to restore high watermark afterwards: if stats
259 * are taken meanwhile, total_vm and hiwater_vm appear too high.
260 * If this were a serious issue, we'd add a flag to do_munmap().
261 */
262 hiwater_vm = mm->hiwater_vm;
263 mm->total_vm += new_len >> PAGE_SHIFT;
264 vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
265
266 if (do_munmap(mm, old_addr, old_len) < 0) {
267 /* OOM: unable to split vma, just get accounts right */
268 vm_unacct_memory(excess >> PAGE_SHIFT);
269 excess = 0;
270 }
271 mm->hiwater_vm = hiwater_vm;
272
273 /* Restore VM_ACCOUNT if one or two pieces of vma left */
274 if (excess) {
275 vma->vm_flags |= VM_ACCOUNT;
276 if (split)
277 vma->vm_next->vm_flags |= VM_ACCOUNT;
278 }
279
280 if (vm_flags & VM_LOCKED) {
281 mm->locked_vm += new_len >> PAGE_SHIFT;
282 if (new_len > old_len)
283 mlock_vma_pages_range(new_vma, new_addr + old_len,
284 new_addr + new_len);
285 }
286
287 return new_addr;
288}
289
290static struct vm_area_struct *vma_to_resize(unsigned long addr,
291 unsigned long old_len, unsigned long new_len, unsigned long *p)
292{
293 struct mm_struct *mm = current->mm;
294 struct vm_area_struct *vma = find_vma(mm, addr);
295
296 if (!vma || vma->vm_start > addr)
297 goto Efault;
298
299 if (is_vm_hugetlb_page(vma))
300 goto Einval;
301
302 /* We can't remap across vm area boundaries */
303 if (old_len > vma->vm_end - addr)
304 goto Efault;
305
306 /* Need to be careful about a growing mapping */
307 if (new_len > old_len) {
308 unsigned long pgoff;
309
310 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
311 goto Efault;
312 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
313 pgoff += vma->vm_pgoff;
314 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
315 goto Einval;
316 }
317
318 if (vma->vm_flags & VM_LOCKED) {
319 unsigned long locked, lock_limit;
320 locked = mm->locked_vm << PAGE_SHIFT;
321 lock_limit = rlimit(RLIMIT_MEMLOCK);
322 locked += new_len - old_len;
323 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
324 goto Eagain;
325 }
326
327 if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
328 goto Enomem;
329
330 if (vma->vm_flags & VM_ACCOUNT) {
331 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
332 if (security_vm_enough_memory_mm(mm, charged))
333 goto Efault;
334 *p = charged;
335 }
336
337 return vma;
338
339Efault: /* very odd choice for most of the cases, but... */
340 return ERR_PTR(-EFAULT);
341Einval:
342 return ERR_PTR(-EINVAL);
343Enomem:
344 return ERR_PTR(-ENOMEM);
345Eagain:
346 return ERR_PTR(-EAGAIN);
347}
348
349static unsigned long mremap_to(unsigned long addr,
350 unsigned long old_len, unsigned long new_addr,
351 unsigned long new_len)
352{
353 struct mm_struct *mm = current->mm;
354 struct vm_area_struct *vma;
355 unsigned long ret = -EINVAL;
356 unsigned long charged = 0;
357 unsigned long map_flags;
358
359 if (new_addr & ~PAGE_MASK)
360 goto out;
361
362 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
363 goto out;
364
365 /* Check if the location we're moving into overlaps the
366 * old location at all, and fail if it does.
367 */
368 if ((new_addr <= addr) && (new_addr+new_len) > addr)
369 goto out;
370
371 if ((addr <= new_addr) && (addr+old_len) > new_addr)
372 goto out;
373
374 ret = do_munmap(mm, new_addr, new_len);
375 if (ret)
376 goto out;
377
378 if (old_len >= new_len) {
379 ret = do_munmap(mm, addr+new_len, old_len - new_len);
380 if (ret && old_len != new_len)
381 goto out;
382 old_len = new_len;
383 }
384
385 vma = vma_to_resize(addr, old_len, new_len, &charged);
386 if (IS_ERR(vma)) {
387 ret = PTR_ERR(vma);
388 goto out;
389 }
390
391 map_flags = MAP_FIXED;
392 if (vma->vm_flags & VM_MAYSHARE)
393 map_flags |= MAP_SHARED;
394
395 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
396 ((addr - vma->vm_start) >> PAGE_SHIFT),
397 map_flags);
398 if (ret & ~PAGE_MASK)
399 goto out1;
400
401 ret = move_vma(vma, addr, old_len, new_len, new_addr);
402 if (!(ret & ~PAGE_MASK))
403 goto out;
404out1:
405 vm_unacct_memory(charged);
406
407out:
408 return ret;
409}
410
411static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
412{
413 unsigned long end = vma->vm_end + delta;
414 if (end < vma->vm_end) /* overflow */
415 return 0;
416 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
417 return 0;
418 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
419 0, MAP_FIXED) & ~PAGE_MASK)
420 return 0;
421 return 1;
422}
423
424/*
425 * Expand (or shrink) an existing mapping, potentially moving it at the
426 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
427 *
428 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
429 * This option implies MREMAP_MAYMOVE.
430 */
431SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
432 unsigned long, new_len, unsigned long, flags,
433 unsigned long, new_addr)
434{
435 struct mm_struct *mm = current->mm;
436 struct vm_area_struct *vma;
437 unsigned long ret = -EINVAL;
438 unsigned long charged = 0;
439
440 down_write(¤t->mm->mmap_sem);
441
442 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
443 goto out;
444
445 if (addr & ~PAGE_MASK)
446 goto out;
447
448 old_len = PAGE_ALIGN(old_len);
449 new_len = PAGE_ALIGN(new_len);
450
451 /*
452 * We allow a zero old-len as a special case
453 * for DOS-emu "duplicate shm area" thing. But
454 * a zero new-len is nonsensical.
455 */
456 if (!new_len)
457 goto out;
458
459 if (flags & MREMAP_FIXED) {
460 if (flags & MREMAP_MAYMOVE)
461 ret = mremap_to(addr, old_len, new_addr, new_len);
462 goto out;
463 }
464
465 /*
466 * Always allow a shrinking remap: that just unmaps
467 * the unnecessary pages..
468 * do_munmap does all the needed commit accounting
469 */
470 if (old_len >= new_len) {
471 ret = do_munmap(mm, addr+new_len, old_len - new_len);
472 if (ret && old_len != new_len)
473 goto out;
474 ret = addr;
475 goto out;
476 }
477
478 /*
479 * Ok, we need to grow..
480 */
481 vma = vma_to_resize(addr, old_len, new_len, &charged);
482 if (IS_ERR(vma)) {
483 ret = PTR_ERR(vma);
484 goto out;
485 }
486
487 /* old_len exactly to the end of the area..
488 */
489 if (old_len == vma->vm_end - addr) {
490 /* can we just expand the current mapping? */
491 if (vma_expandable(vma, new_len - old_len)) {
492 int pages = (new_len - old_len) >> PAGE_SHIFT;
493
494 if (vma_adjust(vma, vma->vm_start, addr + new_len,
495 vma->vm_pgoff, NULL)) {
496 ret = -ENOMEM;
497 goto out;
498 }
499
500 mm->total_vm += pages;
501 vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
502 if (vma->vm_flags & VM_LOCKED) {
503 mm->locked_vm += pages;
504 mlock_vma_pages_range(vma, addr + old_len,
505 addr + new_len);
506 }
507 ret = addr;
508 goto out;
509 }
510 }
511
512 /*
513 * We weren't able to just expand or shrink the area,
514 * we need to create a new one and move it..
515 */
516 ret = -ENOMEM;
517 if (flags & MREMAP_MAYMOVE) {
518 unsigned long map_flags = 0;
519 if (vma->vm_flags & VM_MAYSHARE)
520 map_flags |= MAP_SHARED;
521
522 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
523 vma->vm_pgoff +
524 ((addr - vma->vm_start) >> PAGE_SHIFT),
525 map_flags);
526 if (new_addr & ~PAGE_MASK) {
527 ret = new_addr;
528 goto out;
529 }
530
531 ret = move_vma(vma, addr, old_len, new_len, new_addr);
532 }
533out:
534 if (ret & ~PAGE_MASK)
535 vm_unacct_memory(charged);
536 up_write(¤t->mm->mmap_sem);
537 return ret;
538}