1
2
3
4
5
6
7
8
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
104
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, bool *need_flush)
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
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144 if (need_rmap_locks)
145 take_rmap_locks(vma);
146
147
148
149
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
167
168
169
170
171
172
173
174 if (pte_present(pte) && pte_dirty(pte))
175 force_flush = true;
176 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
177 pte = move_soft_dirty_pte(pte);
178 set_pte_at(mm, new_addr, new_pte, pte);
179 }
180
181 arch_leave_lazy_mmu_mode();
182 if (new_ptl != old_ptl)
183 spin_unlock(new_ptl);
184 pte_unmap(new_pte - 1);
185 if (force_flush)
186 flush_tlb_range(vma, old_end - len, old_end);
187 else
188 *need_flush = true;
189 pte_unmap_unlock(old_pte - 1, old_ptl);
190 if (need_rmap_locks)
191 drop_rmap_locks(vma);
192}
193
194#define LATENCY_LIMIT (64 * PAGE_SIZE)
195
196unsigned long move_page_tables(struct vm_area_struct *vma,
197 unsigned long old_addr, struct vm_area_struct *new_vma,
198 unsigned long new_addr, unsigned long len,
199 bool need_rmap_locks)
200{
201 unsigned long extent, next, old_end;
202 pmd_t *old_pmd, *new_pmd;
203 bool need_flush = false;
204 unsigned long mmun_start;
205 unsigned long mmun_end;
206
207 old_end = old_addr + len;
208 flush_cache_range(vma, old_addr, old_end);
209
210 mmun_start = old_addr;
211 mmun_end = old_end;
212 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
213
214 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
215 cond_resched();
216 next = (old_addr + PMD_SIZE) & PMD_MASK;
217
218 extent = next - old_addr;
219 if (extent > old_end - old_addr)
220 extent = old_end - old_addr;
221 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
222 if (!old_pmd)
223 continue;
224 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
225 if (!new_pmd)
226 break;
227 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd)) {
228 if (extent == HPAGE_PMD_SIZE) {
229 bool moved;
230
231 if (need_rmap_locks)
232 take_rmap_locks(vma);
233 moved = move_huge_pmd(vma, old_addr, new_addr,
234 old_end, old_pmd, new_pmd,
235 &need_flush);
236 if (need_rmap_locks)
237 drop_rmap_locks(vma);
238 if (moved)
239 continue;
240 }
241 split_huge_pmd(vma, old_pmd, old_addr);
242 if (pmd_trans_unstable(old_pmd))
243 continue;
244 }
245 if (pte_alloc(new_vma->vm_mm, new_pmd, new_addr))
246 break;
247 next = (new_addr + PMD_SIZE) & PMD_MASK;
248 if (extent > next - new_addr)
249 extent = next - new_addr;
250 if (extent > LATENCY_LIMIT)
251 extent = LATENCY_LIMIT;
252 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
253 new_pmd, new_addr, need_rmap_locks, &need_flush);
254 }
255 if (need_flush)
256 flush_tlb_range(vma, old_end-len, old_addr);
257
258 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
259
260 return len + old_addr - old_end;
261}
262
263static unsigned long move_vma(struct vm_area_struct *vma,
264 unsigned long old_addr, unsigned long old_len,
265 unsigned long new_len, unsigned long new_addr,
266 bool *locked, struct vm_userfaultfd_ctx *uf,
267 struct list_head *uf_unmap)
268{
269 struct mm_struct *mm = vma->vm_mm;
270 struct vm_area_struct *new_vma;
271 unsigned long vm_flags = vma->vm_flags;
272 unsigned long new_pgoff;
273 unsigned long moved_len;
274 unsigned long excess = 0;
275 unsigned long hiwater_vm;
276 int split = 0;
277 int err;
278 bool need_rmap_locks;
279
280
281
282
283
284 if (mm->map_count >= sysctl_max_map_count - 3)
285 return -ENOMEM;
286
287
288
289
290
291
292
293
294 err = ksm_madvise(vma, old_addr, old_addr + old_len,
295 MADV_UNMERGEABLE, &vm_flags);
296 if (err)
297 return err;
298
299 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
300 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
301 &need_rmap_locks);
302 if (!new_vma)
303 return -ENOMEM;
304
305 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
306 need_rmap_locks);
307 if (moved_len < old_len) {
308 err = -ENOMEM;
309 } else if (vma->vm_ops && vma->vm_ops->mremap) {
310 err = vma->vm_ops->mremap(new_vma);
311 }
312
313 if (unlikely(err)) {
314
315
316
317
318
319 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
320 true);
321 vma = new_vma;
322 old_len = new_len;
323 old_addr = new_addr;
324 new_addr = err;
325 } else {
326 mremap_userfaultfd_prep(new_vma, uf);
327 arch_remap(mm, old_addr, old_addr + old_len,
328 new_addr, new_addr + new_len);
329 }
330
331
332 if (vm_flags & VM_ACCOUNT) {
333 vma->vm_flags &= ~VM_ACCOUNT;
334 excess = vma->vm_end - vma->vm_start - old_len;
335 if (old_addr > vma->vm_start &&
336 old_addr + old_len < vma->vm_end)
337 split = 1;
338 }
339
340
341
342
343
344
345
346
347
348
349 hiwater_vm = mm->hiwater_vm;
350 vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
351
352
353 if (unlikely(vma->vm_flags & VM_PFNMAP))
354 untrack_pfn_moved(vma);
355
356 if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
357
358 vm_unacct_memory(excess >> PAGE_SHIFT);
359 excess = 0;
360 }
361 mm->hiwater_vm = hiwater_vm;
362
363
364 if (excess) {
365 vma->vm_flags |= VM_ACCOUNT;
366 if (split)
367 vma->vm_next->vm_flags |= VM_ACCOUNT;
368 }
369
370 if (vm_flags & VM_LOCKED) {
371 mm->locked_vm += new_len >> PAGE_SHIFT;
372 *locked = true;
373 }
374
375 return new_addr;
376}
377
378static struct vm_area_struct *vma_to_resize(unsigned long addr,
379 unsigned long old_len, unsigned long new_len, unsigned long *p)
380{
381 struct mm_struct *mm = current->mm;
382 struct vm_area_struct *vma = find_vma(mm, addr);
383 unsigned long pgoff;
384
385 if (!vma || vma->vm_start > addr)
386 return ERR_PTR(-EFAULT);
387
388
389
390
391
392
393
394
395
396 if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
397 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap. This is not supported.\n", current->comm, current->pid);
398 return ERR_PTR(-EINVAL);
399 }
400
401 if (is_vm_hugetlb_page(vma))
402 return ERR_PTR(-EINVAL);
403
404
405 if (old_len > vma->vm_end - addr)
406 return ERR_PTR(-EFAULT);
407
408 if (new_len == old_len)
409 return vma;
410
411
412 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
413 pgoff += vma->vm_pgoff;
414 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
415 return ERR_PTR(-EINVAL);
416
417 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
418 return ERR_PTR(-EFAULT);
419
420 if (vma->vm_flags & VM_LOCKED) {
421 unsigned long locked, lock_limit;
422 locked = mm->locked_vm << PAGE_SHIFT;
423 lock_limit = rlimit(RLIMIT_MEMLOCK);
424 locked += new_len - old_len;
425 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
426 return ERR_PTR(-EAGAIN);
427 }
428
429 if (!may_expand_vm(mm, vma->vm_flags,
430 (new_len - old_len) >> PAGE_SHIFT))
431 return ERR_PTR(-ENOMEM);
432
433 if (vma->vm_flags & VM_ACCOUNT) {
434 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
435 if (security_vm_enough_memory_mm(mm, charged))
436 return ERR_PTR(-ENOMEM);
437 *p = charged;
438 }
439
440 return vma;
441}
442
443static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
444 unsigned long new_addr, unsigned long new_len, bool *locked,
445 struct vm_userfaultfd_ctx *uf,
446 struct list_head *uf_unmap_early,
447 struct list_head *uf_unmap)
448{
449 struct mm_struct *mm = current->mm;
450 struct vm_area_struct *vma;
451 unsigned long ret = -EINVAL;
452 unsigned long charged = 0;
453 unsigned long map_flags;
454
455 if (offset_in_page(new_addr))
456 goto out;
457
458 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
459 goto out;
460
461
462 if (addr + old_len > new_addr && new_addr + new_len > addr)
463 goto out;
464
465 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
466 if (ret)
467 goto out;
468
469 if (old_len >= new_len) {
470 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
471 if (ret && old_len != new_len)
472 goto out;
473 old_len = new_len;
474 }
475
476 vma = vma_to_resize(addr, old_len, new_len, &charged);
477 if (IS_ERR(vma)) {
478 ret = PTR_ERR(vma);
479 goto out;
480 }
481
482 map_flags = MAP_FIXED;
483 if (vma->vm_flags & VM_MAYSHARE)
484 map_flags |= MAP_SHARED;
485
486 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
487 ((addr - vma->vm_start) >> PAGE_SHIFT),
488 map_flags);
489 if (offset_in_page(ret))
490 goto out1;
491
492 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, uf,
493 uf_unmap);
494 if (!(offset_in_page(ret)))
495 goto out;
496out1:
497 vm_unacct_memory(charged);
498
499out:
500 return ret;
501}
502
503static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
504{
505 unsigned long end = vma->vm_end + delta;
506 if (end < vma->vm_end)
507 return 0;
508 if (vma->vm_next && vma->vm_next->vm_start < end)
509 return 0;
510 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
511 0, MAP_FIXED) & ~PAGE_MASK)
512 return 0;
513 return 1;
514}
515
516
517
518
519
520
521
522
523SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
524 unsigned long, new_len, unsigned long, flags,
525 unsigned long, new_addr)
526{
527 struct mm_struct *mm = current->mm;
528 struct vm_area_struct *vma;
529 unsigned long ret = -EINVAL;
530 unsigned long charged = 0;
531 bool locked = false;
532 struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
533 LIST_HEAD(uf_unmap_early);
534 LIST_HEAD(uf_unmap);
535
536 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
537 return ret;
538
539 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
540 return ret;
541
542 if (offset_in_page(addr))
543 return ret;
544
545 old_len = PAGE_ALIGN(old_len);
546 new_len = PAGE_ALIGN(new_len);
547
548
549
550
551
552
553 if (!new_len)
554 return ret;
555
556 if (down_write_killable(¤t->mm->mmap_sem))
557 return -EINTR;
558
559 if (flags & MREMAP_FIXED) {
560 ret = mremap_to(addr, old_len, new_addr, new_len,
561 &locked, &uf, &uf_unmap_early, &uf_unmap);
562 goto out;
563 }
564
565
566
567
568
569
570 if (old_len >= new_len) {
571 ret = do_munmap(mm, addr+new_len, old_len - new_len, &uf_unmap);
572 if (ret && old_len != new_len)
573 goto out;
574 ret = addr;
575 goto out;
576 }
577
578
579
580
581 vma = vma_to_resize(addr, old_len, new_len, &charged);
582 if (IS_ERR(vma)) {
583 ret = PTR_ERR(vma);
584 goto out;
585 }
586
587
588
589 if (old_len == vma->vm_end - addr) {
590
591 if (vma_expandable(vma, new_len - old_len)) {
592 int pages = (new_len - old_len) >> PAGE_SHIFT;
593
594 if (vma_adjust(vma, vma->vm_start, addr + new_len,
595 vma->vm_pgoff, NULL)) {
596 ret = -ENOMEM;
597 goto out;
598 }
599
600 vm_stat_account(mm, vma->vm_flags, pages);
601 if (vma->vm_flags & VM_LOCKED) {
602 mm->locked_vm += pages;
603 locked = true;
604 new_addr = addr;
605 }
606 ret = addr;
607 goto out;
608 }
609 }
610
611
612
613
614
615 ret = -ENOMEM;
616 if (flags & MREMAP_MAYMOVE) {
617 unsigned long map_flags = 0;
618 if (vma->vm_flags & VM_MAYSHARE)
619 map_flags |= MAP_SHARED;
620
621 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
622 vma->vm_pgoff +
623 ((addr - vma->vm_start) >> PAGE_SHIFT),
624 map_flags);
625 if (offset_in_page(new_addr)) {
626 ret = new_addr;
627 goto out;
628 }
629
630 ret = move_vma(vma, addr, old_len, new_len, new_addr,
631 &locked, &uf, &uf_unmap);
632 }
633out:
634 if (offset_in_page(ret)) {
635 vm_unacct_memory(charged);
636 locked = 0;
637 }
638 up_write(¤t->mm->mmap_sem);
639 if (locked && new_len > old_len)
640 mm_populate(new_addr + old_len, new_len - old_len);
641 userfaultfd_unmap_complete(mm, &uf_unmap_early);
642 mremap_userfaultfd_complete(&uf, addr, new_addr, old_len);
643 userfaultfd_unmap_complete(mm, &uf_unmap);
644 return ret;
645}
646