Loading...
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 split_huge_page_pmd(mm, pmd);
45 if (pmd_none_or_clear_bad(pmd))
46 return NULL;
47
48 return pmd;
49}
50
51static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
52 unsigned long addr)
53{
54 pgd_t *pgd;
55 pud_t *pud;
56 pmd_t *pmd;
57
58 pgd = pgd_offset(mm, addr);
59 pud = pud_alloc(mm, pgd, addr);
60 if (!pud)
61 return NULL;
62
63 pmd = pmd_alloc(mm, pud, addr);
64 if (!pmd)
65 return NULL;
66
67 VM_BUG_ON(pmd_trans_huge(*pmd));
68 if (pmd_none(*pmd) && __pte_alloc(mm, vma, pmd, addr))
69 return NULL;
70
71 return pmd;
72}
73
74static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
75 unsigned long old_addr, unsigned long old_end,
76 struct vm_area_struct *new_vma, pmd_t *new_pmd,
77 unsigned long new_addr)
78{
79 struct address_space *mapping = NULL;
80 struct mm_struct *mm = vma->vm_mm;
81 pte_t *old_pte, *new_pte, pte;
82 spinlock_t *old_ptl, *new_ptl;
83 unsigned long old_start;
84
85 old_start = old_addr;
86 mmu_notifier_invalidate_range_start(vma->vm_mm,
87 old_start, old_end);
88 if (vma->vm_file) {
89 /*
90 * Subtle point from Rajesh Venkatasubramanian: before
91 * moving file-based ptes, we must lock truncate_pagecache
92 * out, since it might clean the dst vma before the src vma,
93 * and we propagate stale pages into the dst afterward.
94 */
95 mapping = vma->vm_file->f_mapping;
96 mutex_lock(&mapping->i_mmap_mutex);
97 }
98
99 /*
100 * We don't have to worry about the ordering of src and dst
101 * pte locks because exclusive mmap_sem prevents deadlock.
102 */
103 old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
104 new_pte = pte_offset_map(new_pmd, new_addr);
105 new_ptl = pte_lockptr(mm, new_pmd);
106 if (new_ptl != old_ptl)
107 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
108 arch_enter_lazy_mmu_mode();
109
110 for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
111 new_pte++, new_addr += PAGE_SIZE) {
112 if (pte_none(*old_pte))
113 continue;
114 pte = ptep_clear_flush(vma, old_addr, old_pte);
115 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
116 set_pte_at(mm, new_addr, new_pte, pte);
117 }
118
119 arch_leave_lazy_mmu_mode();
120 if (new_ptl != old_ptl)
121 spin_unlock(new_ptl);
122 pte_unmap(new_pte - 1);
123 pte_unmap_unlock(old_pte - 1, old_ptl);
124 if (mapping)
125 mutex_unlock(&mapping->i_mmap_mutex);
126 mmu_notifier_invalidate_range_end(vma->vm_mm, old_start, old_end);
127}
128
129#define LATENCY_LIMIT (64 * PAGE_SIZE)
130
131unsigned long move_page_tables(struct vm_area_struct *vma,
132 unsigned long old_addr, struct vm_area_struct *new_vma,
133 unsigned long new_addr, unsigned long len)
134{
135 unsigned long extent, next, old_end;
136 pmd_t *old_pmd, *new_pmd;
137
138 old_end = old_addr + len;
139 flush_cache_range(vma, old_addr, old_end);
140
141 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
142 cond_resched();
143 next = (old_addr + PMD_SIZE) & PMD_MASK;
144 if (next - 1 > old_end)
145 next = old_end;
146 extent = next - old_addr;
147 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
148 if (!old_pmd)
149 continue;
150 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
151 if (!new_pmd)
152 break;
153 next = (new_addr + PMD_SIZE) & PMD_MASK;
154 if (extent > next - new_addr)
155 extent = next - new_addr;
156 if (extent > LATENCY_LIMIT)
157 extent = LATENCY_LIMIT;
158 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
159 new_vma, new_pmd, new_addr);
160 }
161
162 return len + old_addr - old_end; /* how much done */
163}
164
165static unsigned long move_vma(struct vm_area_struct *vma,
166 unsigned long old_addr, unsigned long old_len,
167 unsigned long new_len, unsigned long new_addr)
168{
169 struct mm_struct *mm = vma->vm_mm;
170 struct vm_area_struct *new_vma;
171 unsigned long vm_flags = vma->vm_flags;
172 unsigned long new_pgoff;
173 unsigned long moved_len;
174 unsigned long excess = 0;
175 unsigned long hiwater_vm;
176 int split = 0;
177 int err;
178
179 /*
180 * We'd prefer to avoid failure later on in do_munmap:
181 * which may split one vma into three before unmapping.
182 */
183 if (mm->map_count >= sysctl_max_map_count - 3)
184 return -ENOMEM;
185
186 /*
187 * Advise KSM to break any KSM pages in the area to be moved:
188 * it would be confusing if they were to turn up at the new
189 * location, where they happen to coincide with different KSM
190 * pages recently unmapped. But leave vma->vm_flags as it was,
191 * so KSM can come around to merge on vma and new_vma afterwards.
192 */
193 err = ksm_madvise(vma, old_addr, old_addr + old_len,
194 MADV_UNMERGEABLE, &vm_flags);
195 if (err)
196 return err;
197
198 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
199 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
200 if (!new_vma)
201 return -ENOMEM;
202
203 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
204 if (moved_len < old_len) {
205 /*
206 * On error, move entries back from new area to old,
207 * which will succeed since page tables still there,
208 * and then proceed to unmap new area instead of old.
209 */
210 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
211 vma = new_vma;
212 old_len = new_len;
213 old_addr = new_addr;
214 new_addr = -ENOMEM;
215 }
216
217 /* Conceal VM_ACCOUNT so old reservation is not undone */
218 if (vm_flags & VM_ACCOUNT) {
219 vma->vm_flags &= ~VM_ACCOUNT;
220 excess = vma->vm_end - vma->vm_start - old_len;
221 if (old_addr > vma->vm_start &&
222 old_addr + old_len < vma->vm_end)
223 split = 1;
224 }
225
226 /*
227 * If we failed to move page tables we still do total_vm increment
228 * since do_munmap() will decrement it by old_len == new_len.
229 *
230 * Since total_vm is about to be raised artificially high for a
231 * moment, we need to restore high watermark afterwards: if stats
232 * are taken meanwhile, total_vm and hiwater_vm appear too high.
233 * If this were a serious issue, we'd add a flag to do_munmap().
234 */
235 hiwater_vm = mm->hiwater_vm;
236 mm->total_vm += new_len >> PAGE_SHIFT;
237 vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
238
239 if (do_munmap(mm, old_addr, old_len) < 0) {
240 /* OOM: unable to split vma, just get accounts right */
241 vm_unacct_memory(excess >> PAGE_SHIFT);
242 excess = 0;
243 }
244 mm->hiwater_vm = hiwater_vm;
245
246 /* Restore VM_ACCOUNT if one or two pieces of vma left */
247 if (excess) {
248 vma->vm_flags |= VM_ACCOUNT;
249 if (split)
250 vma->vm_next->vm_flags |= VM_ACCOUNT;
251 }
252
253 if (vm_flags & VM_LOCKED) {
254 mm->locked_vm += new_len >> PAGE_SHIFT;
255 if (new_len > old_len)
256 mlock_vma_pages_range(new_vma, new_addr + old_len,
257 new_addr + new_len);
258 }
259
260 return new_addr;
261}
262
263static struct vm_area_struct *vma_to_resize(unsigned long addr,
264 unsigned long old_len, unsigned long new_len, unsigned long *p)
265{
266 struct mm_struct *mm = current->mm;
267 struct vm_area_struct *vma = find_vma(mm, addr);
268
269 if (!vma || vma->vm_start > addr)
270 goto Efault;
271
272 if (is_vm_hugetlb_page(vma))
273 goto Einval;
274
275 /* We can't remap across vm area boundaries */
276 if (old_len > vma->vm_end - addr)
277 goto Efault;
278
279 /* Need to be careful about a growing mapping */
280 if (new_len > old_len) {
281 unsigned long pgoff;
282
283 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
284 goto Efault;
285 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
286 pgoff += vma->vm_pgoff;
287 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
288 goto Einval;
289 }
290
291 if (vma->vm_flags & VM_LOCKED) {
292 unsigned long locked, lock_limit;
293 locked = mm->locked_vm << PAGE_SHIFT;
294 lock_limit = rlimit(RLIMIT_MEMLOCK);
295 locked += new_len - old_len;
296 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
297 goto Eagain;
298 }
299
300 if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
301 goto Enomem;
302
303 if (vma->vm_flags & VM_ACCOUNT) {
304 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
305 if (security_vm_enough_memory(charged))
306 goto Efault;
307 *p = charged;
308 }
309
310 return vma;
311
312Efault: /* very odd choice for most of the cases, but... */
313 return ERR_PTR(-EFAULT);
314Einval:
315 return ERR_PTR(-EINVAL);
316Enomem:
317 return ERR_PTR(-ENOMEM);
318Eagain:
319 return ERR_PTR(-EAGAIN);
320}
321
322static unsigned long mremap_to(unsigned long addr,
323 unsigned long old_len, unsigned long new_addr,
324 unsigned long new_len)
325{
326 struct mm_struct *mm = current->mm;
327 struct vm_area_struct *vma;
328 unsigned long ret = -EINVAL;
329 unsigned long charged = 0;
330 unsigned long map_flags;
331
332 if (new_addr & ~PAGE_MASK)
333 goto out;
334
335 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
336 goto out;
337
338 /* Check if the location we're moving into overlaps the
339 * old location at all, and fail if it does.
340 */
341 if ((new_addr <= addr) && (new_addr+new_len) > addr)
342 goto out;
343
344 if ((addr <= new_addr) && (addr+old_len) > new_addr)
345 goto out;
346
347 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
348 if (ret)
349 goto out;
350
351 ret = do_munmap(mm, new_addr, new_len);
352 if (ret)
353 goto out;
354
355 if (old_len >= new_len) {
356 ret = do_munmap(mm, addr+new_len, old_len - new_len);
357 if (ret && old_len != new_len)
358 goto out;
359 old_len = new_len;
360 }
361
362 vma = vma_to_resize(addr, old_len, new_len, &charged);
363 if (IS_ERR(vma)) {
364 ret = PTR_ERR(vma);
365 goto out;
366 }
367
368 map_flags = MAP_FIXED;
369 if (vma->vm_flags & VM_MAYSHARE)
370 map_flags |= MAP_SHARED;
371
372 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
373 ((addr - vma->vm_start) >> PAGE_SHIFT),
374 map_flags);
375 if (ret & ~PAGE_MASK)
376 goto out1;
377
378 ret = move_vma(vma, addr, old_len, new_len, new_addr);
379 if (!(ret & ~PAGE_MASK))
380 goto out;
381out1:
382 vm_unacct_memory(charged);
383
384out:
385 return ret;
386}
387
388static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
389{
390 unsigned long end = vma->vm_end + delta;
391 if (end < vma->vm_end) /* overflow */
392 return 0;
393 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
394 return 0;
395 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
396 0, MAP_FIXED) & ~PAGE_MASK)
397 return 0;
398 return 1;
399}
400
401/*
402 * Expand (or shrink) an existing mapping, potentially moving it at the
403 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
404 *
405 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
406 * This option implies MREMAP_MAYMOVE.
407 */
408unsigned long do_mremap(unsigned long addr,
409 unsigned long old_len, unsigned long new_len,
410 unsigned long flags, unsigned long new_addr)
411{
412 struct mm_struct *mm = current->mm;
413 struct vm_area_struct *vma;
414 unsigned long ret = -EINVAL;
415 unsigned long charged = 0;
416
417 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
418 goto out;
419
420 if (addr & ~PAGE_MASK)
421 goto out;
422
423 old_len = PAGE_ALIGN(old_len);
424 new_len = PAGE_ALIGN(new_len);
425
426 /*
427 * We allow a zero old-len as a special case
428 * for DOS-emu "duplicate shm area" thing. But
429 * a zero new-len is nonsensical.
430 */
431 if (!new_len)
432 goto out;
433
434 if (flags & MREMAP_FIXED) {
435 if (flags & MREMAP_MAYMOVE)
436 ret = mremap_to(addr, old_len, new_addr, new_len);
437 goto out;
438 }
439
440 /*
441 * Always allow a shrinking remap: that just unmaps
442 * the unnecessary pages..
443 * do_munmap does all the needed commit accounting
444 */
445 if (old_len >= new_len) {
446 ret = do_munmap(mm, addr+new_len, old_len - new_len);
447 if (ret && old_len != new_len)
448 goto out;
449 ret = addr;
450 goto out;
451 }
452
453 /*
454 * Ok, we need to grow..
455 */
456 vma = vma_to_resize(addr, old_len, new_len, &charged);
457 if (IS_ERR(vma)) {
458 ret = PTR_ERR(vma);
459 goto out;
460 }
461
462 /* old_len exactly to the end of the area..
463 */
464 if (old_len == vma->vm_end - addr) {
465 /* can we just expand the current mapping? */
466 if (vma_expandable(vma, new_len - old_len)) {
467 int pages = (new_len - old_len) >> PAGE_SHIFT;
468
469 if (vma_adjust(vma, vma->vm_start, addr + new_len,
470 vma->vm_pgoff, NULL)) {
471 ret = -ENOMEM;
472 goto out;
473 }
474
475 mm->total_vm += pages;
476 vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
477 if (vma->vm_flags & VM_LOCKED) {
478 mm->locked_vm += pages;
479 mlock_vma_pages_range(vma, addr + old_len,
480 addr + new_len);
481 }
482 ret = addr;
483 goto out;
484 }
485 }
486
487 /*
488 * We weren't able to just expand or shrink the area,
489 * we need to create a new one and move it..
490 */
491 ret = -ENOMEM;
492 if (flags & MREMAP_MAYMOVE) {
493 unsigned long map_flags = 0;
494 if (vma->vm_flags & VM_MAYSHARE)
495 map_flags |= MAP_SHARED;
496
497 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
498 vma->vm_pgoff +
499 ((addr - vma->vm_start) >> PAGE_SHIFT),
500 map_flags);
501 if (new_addr & ~PAGE_MASK) {
502 ret = new_addr;
503 goto out;
504 }
505
506 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
507 if (ret)
508 goto out;
509 ret = move_vma(vma, addr, old_len, new_len, new_addr);
510 }
511out:
512 if (ret & ~PAGE_MASK)
513 vm_unacct_memory(charged);
514 return ret;
515}
516
517SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
518 unsigned long, new_len, unsigned long, flags,
519 unsigned long, new_addr)
520{
521 unsigned long ret;
522
523 down_write(¤t->mm->mmap_sem);
524 ret = do_mremap(addr, old_len, new_len, flags, new_addr);
525 up_write(¤t->mm->mmap_sem);
526 return ret;
527}
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 vma_is_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_lock 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, pmd_t *old_pmd, pmd_t *new_pmd)
197{
198 spinlock_t *old_ptl, *new_ptl;
199 struct mm_struct *mm = vma->vm_mm;
200 pmd_t pmd;
201
202 /*
203 * The destination pmd shouldn't be established, free_pgtables()
204 * should have released it.
205 *
206 * However, there's a case during execve() where we use mremap
207 * to move the initial stack, and in that case the target area
208 * may overlap the source area (always moving down).
209 *
210 * If everything is PMD-aligned, that works fine, as moving
211 * each pmd down will clear the source pmd. But if we first
212 * have a few 4kB-only pages that get moved down, and then
213 * hit the "now the rest is PMD-aligned, let's do everything
214 * one pmd at a time", we will still have the old (now empty
215 * of any 4kB pages, but still there) PMD in the page table
216 * tree.
217 *
218 * Warn on it once - because we really should try to figure
219 * out how to do this better - but then say "I won't move
220 * this pmd".
221 *
222 * One alternative might be to just unmap the target pmd at
223 * this point, and verify that it really is empty. We'll see.
224 */
225 if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
226 return false;
227
228 /*
229 * We don't have to worry about the ordering of src and dst
230 * ptlocks because exclusive mmap_lock prevents deadlock.
231 */
232 old_ptl = pmd_lock(vma->vm_mm, old_pmd);
233 new_ptl = pmd_lockptr(mm, new_pmd);
234 if (new_ptl != old_ptl)
235 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
236
237 /* Clear the pmd */
238 pmd = *old_pmd;
239 pmd_clear(old_pmd);
240
241 VM_BUG_ON(!pmd_none(*new_pmd));
242
243 /* Set the new pmd */
244 set_pmd_at(mm, new_addr, new_pmd, pmd);
245 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
246 if (new_ptl != old_ptl)
247 spin_unlock(new_ptl);
248 spin_unlock(old_ptl);
249
250 return true;
251}
252#endif
253
254unsigned long move_page_tables(struct vm_area_struct *vma,
255 unsigned long old_addr, struct vm_area_struct *new_vma,
256 unsigned long new_addr, unsigned long len,
257 bool need_rmap_locks)
258{
259 unsigned long extent, next, old_end;
260 struct mmu_notifier_range range;
261 pmd_t *old_pmd, *new_pmd;
262
263 old_end = old_addr + len;
264 flush_cache_range(vma, old_addr, old_end);
265
266 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
267 old_addr, old_end);
268 mmu_notifier_invalidate_range_start(&range);
269
270 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
271 cond_resched();
272 next = (old_addr + PMD_SIZE) & PMD_MASK;
273 /* even if next overflowed, extent below will be ok */
274 extent = next - old_addr;
275 if (extent > old_end - old_addr)
276 extent = old_end - old_addr;
277 next = (new_addr + PMD_SIZE) & PMD_MASK;
278 if (extent > next - new_addr)
279 extent = next - new_addr;
280 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
281 if (!old_pmd)
282 continue;
283 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
284 if (!new_pmd)
285 break;
286 if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) || pmd_devmap(*old_pmd)) {
287 if (extent == HPAGE_PMD_SIZE) {
288 bool moved;
289 /* See comment in move_ptes() */
290 if (need_rmap_locks)
291 take_rmap_locks(vma);
292 moved = move_huge_pmd(vma, old_addr, new_addr,
293 old_pmd, new_pmd);
294 if (need_rmap_locks)
295 drop_rmap_locks(vma);
296 if (moved)
297 continue;
298 }
299 split_huge_pmd(vma, old_pmd, old_addr);
300 if (pmd_trans_unstable(old_pmd))
301 continue;
302 } else if (extent == PMD_SIZE) {
303#ifdef CONFIG_HAVE_MOVE_PMD
304 /*
305 * If the extent is PMD-sized, try to speed the move by
306 * moving at the PMD level if possible.
307 */
308 bool moved;
309
310 if (need_rmap_locks)
311 take_rmap_locks(vma);
312 moved = move_normal_pmd(vma, old_addr, new_addr,
313 old_pmd, new_pmd);
314 if (need_rmap_locks)
315 drop_rmap_locks(vma);
316 if (moved)
317 continue;
318#endif
319 }
320
321 if (pte_alloc(new_vma->vm_mm, new_pmd))
322 break;
323 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
324 new_pmd, new_addr, need_rmap_locks);
325 }
326
327 mmu_notifier_invalidate_range_end(&range);
328
329 return len + old_addr - old_end; /* how much done */
330}
331
332static unsigned long move_vma(struct vm_area_struct *vma,
333 unsigned long old_addr, unsigned long old_len,
334 unsigned long new_len, unsigned long new_addr,
335 bool *locked, unsigned long flags,
336 struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
337{
338 struct mm_struct *mm = vma->vm_mm;
339 struct vm_area_struct *new_vma;
340 unsigned long vm_flags = vma->vm_flags;
341 unsigned long new_pgoff;
342 unsigned long moved_len;
343 unsigned long excess = 0;
344 unsigned long hiwater_vm;
345 int split = 0;
346 int err;
347 bool need_rmap_locks;
348
349 /*
350 * We'd prefer to avoid failure later on in do_munmap:
351 * which may split one vma into three before unmapping.
352 */
353 if (mm->map_count >= sysctl_max_map_count - 3)
354 return -ENOMEM;
355
356 /*
357 * Advise KSM to break any KSM pages in the area to be moved:
358 * it would be confusing if they were to turn up at the new
359 * location, where they happen to coincide with different KSM
360 * pages recently unmapped. But leave vma->vm_flags as it was,
361 * so KSM can come around to merge on vma and new_vma afterwards.
362 */
363 err = ksm_madvise(vma, old_addr, old_addr + old_len,
364 MADV_UNMERGEABLE, &vm_flags);
365 if (err)
366 return err;
367
368 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
369 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
370 &need_rmap_locks);
371 if (!new_vma)
372 return -ENOMEM;
373
374 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
375 need_rmap_locks);
376 if (moved_len < old_len) {
377 err = -ENOMEM;
378 } else if (vma->vm_ops && vma->vm_ops->mremap) {
379 err = vma->vm_ops->mremap(new_vma);
380 }
381
382 if (unlikely(err)) {
383 /*
384 * On error, move entries back from new area to old,
385 * which will succeed since page tables still there,
386 * and then proceed to unmap new area instead of old.
387 */
388 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
389 true);
390 vma = new_vma;
391 old_len = new_len;
392 old_addr = new_addr;
393 new_addr = err;
394 } else {
395 mremap_userfaultfd_prep(new_vma, uf);
396 arch_remap(mm, old_addr, old_addr + old_len,
397 new_addr, new_addr + new_len);
398 }
399
400 /* Conceal VM_ACCOUNT so old reservation is not undone */
401 if (vm_flags & VM_ACCOUNT) {
402 vma->vm_flags &= ~VM_ACCOUNT;
403 excess = vma->vm_end - vma->vm_start - old_len;
404 if (old_addr > vma->vm_start &&
405 old_addr + old_len < vma->vm_end)
406 split = 1;
407 }
408
409 /*
410 * If we failed to move page tables we still do total_vm increment
411 * since do_munmap() will decrement it by old_len == new_len.
412 *
413 * Since total_vm is about to be raised artificially high for a
414 * moment, we need to restore high watermark afterwards: if stats
415 * are taken meanwhile, total_vm and hiwater_vm appear too high.
416 * If this were a serious issue, we'd add a flag to do_munmap().
417 */
418 hiwater_vm = mm->hiwater_vm;
419 vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
420
421 /* Tell pfnmap has moved from this vma */
422 if (unlikely(vma->vm_flags & VM_PFNMAP))
423 untrack_pfn_moved(vma);
424
425 if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
426 if (vm_flags & VM_ACCOUNT) {
427 /* Always put back VM_ACCOUNT since we won't unmap */
428 vma->vm_flags |= VM_ACCOUNT;
429
430 vm_acct_memory(new_len >> PAGE_SHIFT);
431 }
432
433 /*
434 * VMAs can actually be merged back together in copy_vma
435 * calling merge_vma. This can happen with anonymous vmas
436 * which have not yet been faulted, so if we were to consider
437 * this VMA split we'll end up adding VM_ACCOUNT on the
438 * next VMA, which is completely unrelated if this VMA
439 * was re-merged.
440 */
441 if (split && new_vma == vma)
442 split = 0;
443
444 /* We always clear VM_LOCKED[ONFAULT] on the old vma */
445 vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
446
447 /* Because we won't unmap we don't need to touch locked_vm */
448 goto out;
449 }
450
451 if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
452 /* OOM: unable to split vma, just get accounts right */
453 vm_unacct_memory(excess >> PAGE_SHIFT);
454 excess = 0;
455 }
456
457 if (vm_flags & VM_LOCKED) {
458 mm->locked_vm += new_len >> PAGE_SHIFT;
459 *locked = true;
460 }
461out:
462 mm->hiwater_vm = hiwater_vm;
463
464 /* Restore VM_ACCOUNT if one or two pieces of vma left */
465 if (excess) {
466 vma->vm_flags |= VM_ACCOUNT;
467 if (split)
468 vma->vm_next->vm_flags |= VM_ACCOUNT;
469 }
470
471 return new_addr;
472}
473
474static struct vm_area_struct *vma_to_resize(unsigned long addr,
475 unsigned long old_len, unsigned long new_len, unsigned long flags,
476 unsigned long *p)
477{
478 struct mm_struct *mm = current->mm;
479 struct vm_area_struct *vma = find_vma(mm, addr);
480 unsigned long pgoff;
481
482 if (!vma || vma->vm_start > addr)
483 return ERR_PTR(-EFAULT);
484
485 /*
486 * !old_len is a special case where an attempt is made to 'duplicate'
487 * a mapping. This makes no sense for private mappings as it will
488 * instead create a fresh/new mapping unrelated to the original. This
489 * is contrary to the basic idea of mremap which creates new mappings
490 * based on the original. There are no known use cases for this
491 * behavior. As a result, fail such attempts.
492 */
493 if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
494 pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap. This is not supported.\n", current->comm, current->pid);
495 return ERR_PTR(-EINVAL);
496 }
497
498 if (flags & MREMAP_DONTUNMAP && (!vma_is_anonymous(vma) ||
499 vma->vm_flags & VM_SHARED))
500 return ERR_PTR(-EINVAL);
501
502 if (is_vm_hugetlb_page(vma))
503 return ERR_PTR(-EINVAL);
504
505 /* We can't remap across vm area boundaries */
506 if (old_len > vma->vm_end - addr)
507 return ERR_PTR(-EFAULT);
508
509 if (new_len == old_len)
510 return vma;
511
512 /* Need to be careful about a growing mapping */
513 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
514 pgoff += vma->vm_pgoff;
515 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
516 return ERR_PTR(-EINVAL);
517
518 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
519 return ERR_PTR(-EFAULT);
520
521 if (vma->vm_flags & VM_LOCKED) {
522 unsigned long locked, lock_limit;
523 locked = mm->locked_vm << PAGE_SHIFT;
524 lock_limit = rlimit(RLIMIT_MEMLOCK);
525 locked += new_len - old_len;
526 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
527 return ERR_PTR(-EAGAIN);
528 }
529
530 if (!may_expand_vm(mm, vma->vm_flags,
531 (new_len - old_len) >> PAGE_SHIFT))
532 return ERR_PTR(-ENOMEM);
533
534 if (vma->vm_flags & VM_ACCOUNT) {
535 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
536 if (security_vm_enough_memory_mm(mm, charged))
537 return ERR_PTR(-ENOMEM);
538 *p = charged;
539 }
540
541 return vma;
542}
543
544static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
545 unsigned long new_addr, unsigned long new_len, bool *locked,
546 unsigned long flags, struct vm_userfaultfd_ctx *uf,
547 struct list_head *uf_unmap_early,
548 struct list_head *uf_unmap)
549{
550 struct mm_struct *mm = current->mm;
551 struct vm_area_struct *vma;
552 unsigned long ret = -EINVAL;
553 unsigned long charged = 0;
554 unsigned long map_flags = 0;
555
556 if (offset_in_page(new_addr))
557 goto out;
558
559 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
560 goto out;
561
562 /* Ensure the old/new locations do not overlap */
563 if (addr + old_len > new_addr && new_addr + new_len > addr)
564 goto out;
565
566 /*
567 * move_vma() need us to stay 4 maps below the threshold, otherwise
568 * it will bail out at the very beginning.
569 * That is a problem if we have already unmaped the regions here
570 * (new_addr, and old_addr), because userspace will not know the
571 * state of the vma's after it gets -ENOMEM.
572 * So, to avoid such scenario we can pre-compute if the whole
573 * operation has high chances to success map-wise.
574 * Worst-scenario case is when both vma's (new_addr and old_addr) get
575 * split in 3 before unmaping it.
576 * That means 2 more maps (1 for each) to the ones we already hold.
577 * Check whether current map count plus 2 still leads us to 4 maps below
578 * the threshold, otherwise return -ENOMEM here to be more safe.
579 */
580 if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
581 return -ENOMEM;
582
583 if (flags & MREMAP_FIXED) {
584 ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
585 if (ret)
586 goto out;
587 }
588
589 if (old_len >= new_len) {
590 ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
591 if (ret && old_len != new_len)
592 goto out;
593 old_len = new_len;
594 }
595
596 vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
597 if (IS_ERR(vma)) {
598 ret = PTR_ERR(vma);
599 goto out;
600 }
601
602 /* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
603 if (flags & MREMAP_DONTUNMAP &&
604 !may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
605 ret = -ENOMEM;
606 goto out;
607 }
608
609 if (flags & MREMAP_FIXED)
610 map_flags |= MAP_FIXED;
611
612 if (vma->vm_flags & VM_MAYSHARE)
613 map_flags |= MAP_SHARED;
614
615 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
616 ((addr - vma->vm_start) >> PAGE_SHIFT),
617 map_flags);
618 if (IS_ERR_VALUE(ret))
619 goto out1;
620
621 /* We got a new mapping */
622 if (!(flags & MREMAP_FIXED))
623 new_addr = ret;
624
625 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
626 uf_unmap);
627
628 if (!(offset_in_page(ret)))
629 goto out;
630
631out1:
632 vm_unacct_memory(charged);
633
634out:
635 return ret;
636}
637
638static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
639{
640 unsigned long end = vma->vm_end + delta;
641 if (end < vma->vm_end) /* overflow */
642 return 0;
643 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
644 return 0;
645 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
646 0, MAP_FIXED) & ~PAGE_MASK)
647 return 0;
648 return 1;
649}
650
651/*
652 * Expand (or shrink) an existing mapping, potentially moving it at the
653 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
654 *
655 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
656 * This option implies MREMAP_MAYMOVE.
657 */
658SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
659 unsigned long, new_len, unsigned long, flags,
660 unsigned long, new_addr)
661{
662 struct mm_struct *mm = current->mm;
663 struct vm_area_struct *vma;
664 unsigned long ret = -EINVAL;
665 unsigned long charged = 0;
666 bool locked = false;
667 bool downgraded = false;
668 struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
669 LIST_HEAD(uf_unmap_early);
670 LIST_HEAD(uf_unmap);
671
672 /*
673 * There is a deliberate asymmetry here: we strip the pointer tag
674 * from the old address but leave the new address alone. This is
675 * for consistency with mmap(), where we prevent the creation of
676 * aliasing mappings in userspace by leaving the tag bits of the
677 * mapping address intact. A non-zero tag will cause the subsequent
678 * range checks to reject the address as invalid.
679 *
680 * See Documentation/arm64/tagged-address-abi.rst for more information.
681 */
682 addr = untagged_addr(addr);
683
684 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
685 return ret;
686
687 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
688 return ret;
689
690 /*
691 * MREMAP_DONTUNMAP is always a move and it does not allow resizing
692 * in the process.
693 */
694 if (flags & MREMAP_DONTUNMAP &&
695 (!(flags & MREMAP_MAYMOVE) || old_len != new_len))
696 return ret;
697
698
699 if (offset_in_page(addr))
700 return ret;
701
702 old_len = PAGE_ALIGN(old_len);
703 new_len = PAGE_ALIGN(new_len);
704
705 /*
706 * We allow a zero old-len as a special case
707 * for DOS-emu "duplicate shm area" thing. But
708 * a zero new-len is nonsensical.
709 */
710 if (!new_len)
711 return ret;
712
713 if (mmap_write_lock_killable(current->mm))
714 return -EINTR;
715
716 if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
717 ret = mremap_to(addr, old_len, new_addr, new_len,
718 &locked, flags, &uf, &uf_unmap_early,
719 &uf_unmap);
720 goto out;
721 }
722
723 /*
724 * Always allow a shrinking remap: that just unmaps
725 * the unnecessary pages..
726 * __do_munmap does all the needed commit accounting, and
727 * downgrades mmap_lock to read if so directed.
728 */
729 if (old_len >= new_len) {
730 int retval;
731
732 retval = __do_munmap(mm, addr+new_len, old_len - new_len,
733 &uf_unmap, true);
734 if (retval < 0 && old_len != new_len) {
735 ret = retval;
736 goto out;
737 /* Returning 1 indicates mmap_lock is downgraded to read. */
738 } else if (retval == 1)
739 downgraded = true;
740 ret = addr;
741 goto out;
742 }
743
744 /*
745 * Ok, we need to grow..
746 */
747 vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
748 if (IS_ERR(vma)) {
749 ret = PTR_ERR(vma);
750 goto out;
751 }
752
753 /* old_len exactly to the end of the area..
754 */
755 if (old_len == vma->vm_end - addr) {
756 /* can we just expand the current mapping? */
757 if (vma_expandable(vma, new_len - old_len)) {
758 int pages = (new_len - old_len) >> PAGE_SHIFT;
759
760 if (vma_adjust(vma, vma->vm_start, addr + new_len,
761 vma->vm_pgoff, NULL)) {
762 ret = -ENOMEM;
763 goto out;
764 }
765
766 vm_stat_account(mm, vma->vm_flags, pages);
767 if (vma->vm_flags & VM_LOCKED) {
768 mm->locked_vm += pages;
769 locked = true;
770 new_addr = addr;
771 }
772 ret = addr;
773 goto out;
774 }
775 }
776
777 /*
778 * We weren't able to just expand or shrink the area,
779 * we need to create a new one and move it..
780 */
781 ret = -ENOMEM;
782 if (flags & MREMAP_MAYMOVE) {
783 unsigned long map_flags = 0;
784 if (vma->vm_flags & VM_MAYSHARE)
785 map_flags |= MAP_SHARED;
786
787 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
788 vma->vm_pgoff +
789 ((addr - vma->vm_start) >> PAGE_SHIFT),
790 map_flags);
791 if (IS_ERR_VALUE(new_addr)) {
792 ret = new_addr;
793 goto out;
794 }
795
796 ret = move_vma(vma, addr, old_len, new_len, new_addr,
797 &locked, flags, &uf, &uf_unmap);
798 }
799out:
800 if (offset_in_page(ret)) {
801 vm_unacct_memory(charged);
802 locked = false;
803 }
804 if (downgraded)
805 mmap_read_unlock(current->mm);
806 else
807 mmap_write_unlock(current->mm);
808 if (locked && new_len > old_len)
809 mm_populate(new_addr + old_len, new_len - old_len);
810 userfaultfd_unmap_complete(mm, &uf_unmap_early);
811 mremap_userfaultfd_complete(&uf, addr, ret, old_len);
812 userfaultfd_unmap_complete(mm, &uf_unmap);
813 return ret;
814}