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/uaccess.h>
24 #include <linux/mm-arch-hooks.h>
25
26 #include <asm/cacheflush.h>
27 #include <asm/tlbflush.h>
28
29 #include "internal.h"
30
get_old_pmd(struct mm_struct * mm,unsigned long addr)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
alloc_new_pmd(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr)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
take_rmap_locks(struct vm_area_struct * vma)73 static void take_rmap_locks(struct vm_area_struct *vma)
74 {
75 if (vma->vm_file)
76 i_mmap_lock_write(vma->vm_file->f_mapping);
77 if (vma->anon_vma)
78 anon_vma_lock_write(vma->anon_vma);
79 }
80
drop_rmap_locks(struct vm_area_struct * vma)81 static void drop_rmap_locks(struct vm_area_struct *vma)
82 {
83 if (vma->anon_vma)
84 anon_vma_unlock_write(vma->anon_vma);
85 if (vma->vm_file)
86 i_mmap_unlock_write(vma->vm_file->f_mapping);
87 }
88
move_soft_dirty_pte(pte_t pte)89 static pte_t move_soft_dirty_pte(pte_t pte)
90 {
91 /*
92 * Set soft dirty bit so we can notice
93 * in userspace the ptes were moved.
94 */
95 #ifdef CONFIG_MEM_SOFT_DIRTY
96 if (pte_present(pte))
97 pte = pte_mksoft_dirty(pte);
98 else if (is_swap_pte(pte))
99 pte = pte_swp_mksoft_dirty(pte);
100 #endif
101 return pte;
102 }
103
move_ptes(struct vm_area_struct * vma,pmd_t * old_pmd,unsigned long old_addr,unsigned long old_end,struct vm_area_struct * new_vma,pmd_t * new_pmd,unsigned long new_addr,bool need_rmap_locks,bool * need_flush)104 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
105 unsigned long old_addr, unsigned long old_end,
106 struct vm_area_struct *new_vma, pmd_t *new_pmd,
107 unsigned long new_addr, bool need_rmap_locks, bool *need_flush)
108 {
109 struct mm_struct *mm = vma->vm_mm;
110 pte_t *old_pte, *new_pte, pte;
111 spinlock_t *old_ptl, *new_ptl;
112 bool force_flush = false;
113 unsigned long len = old_end - old_addr;
114
115 /*
116 * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
117 * locks to ensure that rmap will always observe either the old or the
118 * new ptes. This is the easiest way to avoid races with
119 * truncate_pagecache(), page migration, etc...
120 *
121 * When need_rmap_locks is false, we use other ways to avoid
122 * such races:
123 *
124 * - During exec() shift_arg_pages(), we use a specially tagged vma
125 * which rmap call sites look for using is_vma_temporary_stack().
126 *
127 * - During mremap(), new_vma is often known to be placed after vma
128 * in rmap traversal order. This ensures rmap will always observe
129 * either the old pte, or the new pte, or both (the page table locks
130 * serialize access to individual ptes, but only rmap traversal
131 * order guarantees that we won't miss both the old and new ptes).
132 */
133 if (need_rmap_locks)
134 take_rmap_locks(vma);
135
136 /*
137 * We don't have to worry about the ordering of src and dst
138 * pte locks because exclusive mmap_sem prevents deadlock.
139 */
140 old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
141 new_pte = pte_offset_map(new_pmd, new_addr);
142 new_ptl = pte_lockptr(mm, new_pmd);
143 if (new_ptl != old_ptl)
144 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
145 flush_tlb_batched_pending(vma->vm_mm);
146 arch_enter_lazy_mmu_mode();
147
148 for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
149 new_pte++, new_addr += PAGE_SIZE) {
150 if (pte_none(*old_pte))
151 continue;
152
153 pte = ptep_get_and_clear(mm, old_addr, old_pte);
154 /*
155 * If we are remapping a dirty PTE, make sure
156 * to flush TLB before we drop the PTL for the
157 * old PTE or we may race with page_mkclean().
158 *
159 * This check has to be done after we removed the
160 * old PTE from page tables or another thread may
161 * dirty it after the check and before the removal.
162 */
163 if (pte_present(pte) && pte_dirty(pte))
164 force_flush = true;
165 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
166 pte = move_soft_dirty_pte(pte);
167 set_pte_at(mm, new_addr, new_pte, pte);
168 }
169
170 arch_leave_lazy_mmu_mode();
171 if (new_ptl != old_ptl)
172 spin_unlock(new_ptl);
173 pte_unmap(new_pte - 1);
174 if (force_flush)
175 flush_tlb_range(vma, old_end - len, old_end);
176 else
177 *need_flush = true;
178 pte_unmap_unlock(old_pte - 1, old_ptl);
179 if (need_rmap_locks)
180 drop_rmap_locks(vma);
181 }
182
183 #define LATENCY_LIMIT (64 * PAGE_SIZE)
184
move_page_tables(struct vm_area_struct * vma,unsigned long old_addr,struct vm_area_struct * new_vma,unsigned long new_addr,unsigned long len,bool need_rmap_locks)185 unsigned long move_page_tables(struct vm_area_struct *vma,
186 unsigned long old_addr, struct vm_area_struct *new_vma,
187 unsigned long new_addr, unsigned long len,
188 bool need_rmap_locks)
189 {
190 unsigned long extent, next, old_end;
191 pmd_t *old_pmd, *new_pmd;
192 bool need_flush = false;
193 unsigned long mmun_start; /* For mmu_notifiers */
194 unsigned long mmun_end; /* For mmu_notifiers */
195
196 old_end = old_addr + len;
197 flush_cache_range(vma, old_addr, old_end);
198
199 mmun_start = old_addr;
200 mmun_end = old_end;
201 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
202
203 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
204 cond_resched();
205 next = (old_addr + PMD_SIZE) & PMD_MASK;
206 /* even if next overflowed, extent below will be ok */
207 extent = next - old_addr;
208 if (extent > old_end - old_addr)
209 extent = old_end - old_addr;
210 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
211 if (!old_pmd)
212 continue;
213 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
214 if (!new_pmd)
215 break;
216 if (pmd_trans_huge(*old_pmd)) {
217 if (extent == HPAGE_PMD_SIZE) {
218 bool moved;
219 /* See comment in move_ptes() */
220 if (need_rmap_locks)
221 take_rmap_locks(vma);
222 moved = move_huge_pmd(vma, old_addr, new_addr,
223 old_end, old_pmd, new_pmd,
224 &need_flush);
225 if (need_rmap_locks)
226 drop_rmap_locks(vma);
227 if (moved)
228 continue;
229 }
230 split_huge_pmd(vma, old_pmd, old_addr);
231 if (pmd_trans_unstable(old_pmd))
232 continue;
233 }
234 if (pte_alloc(new_vma->vm_mm, new_pmd, new_addr))
235 break;
236 next = (new_addr + PMD_SIZE) & PMD_MASK;
237 if (extent > next - new_addr)
238 extent = next - new_addr;
239 if (extent > LATENCY_LIMIT)
240 extent = LATENCY_LIMIT;
241 move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
242 new_pmd, new_addr, need_rmap_locks, &need_flush);
243 }
244 if (need_flush)
245 flush_tlb_range(vma, old_end-len, old_addr);
246
247 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
248
249 return len + old_addr - old_end; /* how much done */
250 }
251
move_vma(struct vm_area_struct * vma,unsigned long old_addr,unsigned long old_len,unsigned long new_len,unsigned long new_addr,bool * locked)252 static unsigned long move_vma(struct vm_area_struct *vma,
253 unsigned long old_addr, unsigned long old_len,
254 unsigned long new_len, unsigned long new_addr, bool *locked)
255 {
256 struct mm_struct *mm = vma->vm_mm;
257 struct vm_area_struct *new_vma;
258 unsigned long vm_flags = vma->vm_flags;
259 unsigned long new_pgoff;
260 unsigned long moved_len;
261 unsigned long excess = 0;
262 unsigned long hiwater_vm;
263 int split = 0;
264 int err;
265 bool need_rmap_locks;
266
267 /*
268 * We'd prefer to avoid failure later on in do_munmap:
269 * which may split one vma into three before unmapping.
270 */
271 if (mm->map_count >= sysctl_max_map_count - 3)
272 return -ENOMEM;
273
274 /*
275 * Advise KSM to break any KSM pages in the area to be moved:
276 * it would be confusing if they were to turn up at the new
277 * location, where they happen to coincide with different KSM
278 * pages recently unmapped. But leave vma->vm_flags as it was,
279 * so KSM can come around to merge on vma and new_vma afterwards.
280 */
281 err = ksm_madvise(vma, old_addr, old_addr + old_len,
282 MADV_UNMERGEABLE, &vm_flags);
283 if (err)
284 return err;
285
286 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
287 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
288 &need_rmap_locks);
289 if (!new_vma)
290 return -ENOMEM;
291
292 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
293 need_rmap_locks);
294 if (moved_len < old_len) {
295 err = -ENOMEM;
296 } else if (vma->vm_ops && vma->vm_ops->mremap) {
297 err = vma->vm_ops->mremap(new_vma);
298 }
299
300 if (unlikely(err)) {
301 /*
302 * On error, move entries back from new area to old,
303 * which will succeed since page tables still there,
304 * and then proceed to unmap new area instead of old.
305 */
306 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
307 true);
308 vma = new_vma;
309 old_len = new_len;
310 old_addr = new_addr;
311 new_addr = err;
312 } else {
313 arch_remap(mm, old_addr, old_addr + old_len,
314 new_addr, new_addr + new_len);
315 }
316
317 /* Conceal VM_ACCOUNT so old reservation is not undone */
318 if (vm_flags & VM_ACCOUNT) {
319 vma->vm_flags &= ~VM_ACCOUNT;
320 excess = vma->vm_end - vma->vm_start - old_len;
321 if (old_addr > vma->vm_start &&
322 old_addr + old_len < vma->vm_end)
323 split = 1;
324 }
325
326 /*
327 * If we failed to move page tables we still do total_vm increment
328 * since do_munmap() will decrement it by old_len == new_len.
329 *
330 * Since total_vm is about to be raised artificially high for a
331 * moment, we need to restore high watermark afterwards: if stats
332 * are taken meanwhile, total_vm and hiwater_vm appear too high.
333 * If this were a serious issue, we'd add a flag to do_munmap().
334 */
335 hiwater_vm = mm->hiwater_vm;
336 vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
337
338 /* Tell pfnmap has moved from this vma */
339 if (unlikely(vma->vm_flags & VM_PFNMAP))
340 untrack_pfn_moved(vma);
341
342 if (do_munmap(mm, old_addr, old_len) < 0) {
343 /* OOM: unable to split vma, just get accounts right */
344 vm_unacct_memory(excess >> PAGE_SHIFT);
345 excess = 0;
346 }
347 mm->hiwater_vm = hiwater_vm;
348
349 /* Restore VM_ACCOUNT if one or two pieces of vma left */
350 if (excess) {
351 vma->vm_flags |= VM_ACCOUNT;
352 if (split)
353 vma->vm_next->vm_flags |= VM_ACCOUNT;
354 }
355
356 if (vm_flags & VM_LOCKED) {
357 mm->locked_vm += new_len >> PAGE_SHIFT;
358 *locked = true;
359 }
360
361 return new_addr;
362 }
363
vma_to_resize(unsigned long addr,unsigned long old_len,unsigned long new_len,unsigned long * p)364 static struct vm_area_struct *vma_to_resize(unsigned long addr,
365 unsigned long old_len, unsigned long new_len, unsigned long *p)
366 {
367 struct mm_struct *mm = current->mm;
368 struct vm_area_struct *vma = find_vma(mm, addr);
369 unsigned long pgoff;
370
371 if (!vma || vma->vm_start > addr)
372 return ERR_PTR(-EFAULT);
373
374 if (is_vm_hugetlb_page(vma))
375 return ERR_PTR(-EINVAL);
376
377 /* We can't remap across vm area boundaries */
378 if (old_len > vma->vm_end - addr)
379 return ERR_PTR(-EFAULT);
380
381 if (new_len == old_len)
382 return vma;
383
384 /* Need to be careful about a growing mapping */
385 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
386 pgoff += vma->vm_pgoff;
387 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
388 return ERR_PTR(-EINVAL);
389
390 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
391 return ERR_PTR(-EFAULT);
392
393 if (vma->vm_flags & VM_LOCKED) {
394 unsigned long locked, lock_limit;
395 locked = mm->locked_vm << PAGE_SHIFT;
396 lock_limit = rlimit(RLIMIT_MEMLOCK);
397 locked += new_len - old_len;
398 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
399 return ERR_PTR(-EAGAIN);
400 }
401
402 if (!may_expand_vm(mm, vma->vm_flags,
403 (new_len - old_len) >> PAGE_SHIFT))
404 return ERR_PTR(-ENOMEM);
405
406 if (vma->vm_flags & VM_ACCOUNT) {
407 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
408 if (security_vm_enough_memory_mm(mm, charged))
409 return ERR_PTR(-ENOMEM);
410 *p = charged;
411 }
412
413 return vma;
414 }
415
mremap_to(unsigned long addr,unsigned long old_len,unsigned long new_addr,unsigned long new_len,bool * locked)416 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
417 unsigned long new_addr, unsigned long new_len, bool *locked)
418 {
419 struct mm_struct *mm = current->mm;
420 struct vm_area_struct *vma;
421 unsigned long ret = -EINVAL;
422 unsigned long charged = 0;
423 unsigned long map_flags;
424
425 if (offset_in_page(new_addr))
426 goto out;
427
428 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
429 goto out;
430
431 /* Ensure the old/new locations do not overlap */
432 if (addr + old_len > new_addr && new_addr + new_len > addr)
433 goto out;
434
435 ret = do_munmap(mm, new_addr, new_len);
436 if (ret)
437 goto out;
438
439 if (old_len >= new_len) {
440 ret = do_munmap(mm, addr+new_len, old_len - new_len);
441 if (ret && old_len != new_len)
442 goto out;
443 old_len = new_len;
444 }
445
446 vma = vma_to_resize(addr, old_len, new_len, &charged);
447 if (IS_ERR(vma)) {
448 ret = PTR_ERR(vma);
449 goto out;
450 }
451
452 map_flags = MAP_FIXED;
453 if (vma->vm_flags & VM_MAYSHARE)
454 map_flags |= MAP_SHARED;
455
456 ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
457 ((addr - vma->vm_start) >> PAGE_SHIFT),
458 map_flags);
459 if (offset_in_page(ret))
460 goto out1;
461
462 ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
463 if (!(offset_in_page(ret)))
464 goto out;
465 out1:
466 vm_unacct_memory(charged);
467
468 out:
469 return ret;
470 }
471
vma_expandable(struct vm_area_struct * vma,unsigned long delta)472 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
473 {
474 unsigned long end = vma->vm_end + delta;
475 if (end < vma->vm_end) /* overflow */
476 return 0;
477 if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
478 return 0;
479 if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
480 0, MAP_FIXED) & ~PAGE_MASK)
481 return 0;
482 return 1;
483 }
484
485 /*
486 * Expand (or shrink) an existing mapping, potentially moving it at the
487 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
488 *
489 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
490 * This option implies MREMAP_MAYMOVE.
491 */
SYSCALL_DEFINE5(mremap,unsigned long,addr,unsigned long,old_len,unsigned long,new_len,unsigned long,flags,unsigned long,new_addr)492 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
493 unsigned long, new_len, unsigned long, flags,
494 unsigned long, new_addr)
495 {
496 struct mm_struct *mm = current->mm;
497 struct vm_area_struct *vma;
498 unsigned long ret = -EINVAL;
499 unsigned long charged = 0;
500 bool locked = false;
501
502 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
503 return ret;
504
505 if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
506 return ret;
507
508 if (offset_in_page(addr))
509 return ret;
510
511 old_len = PAGE_ALIGN(old_len);
512 new_len = PAGE_ALIGN(new_len);
513
514 /*
515 * We allow a zero old-len as a special case
516 * for DOS-emu "duplicate shm area" thing. But
517 * a zero new-len is nonsensical.
518 */
519 if (!new_len)
520 return ret;
521
522 if (down_write_killable(¤t->mm->mmap_sem))
523 return -EINTR;
524
525 if (flags & MREMAP_FIXED) {
526 ret = mremap_to(addr, old_len, new_addr, new_len,
527 &locked);
528 goto out;
529 }
530
531 /*
532 * Always allow a shrinking remap: that just unmaps
533 * the unnecessary pages..
534 * do_munmap does all the needed commit accounting
535 */
536 if (old_len >= new_len) {
537 ret = do_munmap(mm, addr+new_len, old_len - new_len);
538 if (ret && old_len != new_len)
539 goto out;
540 ret = addr;
541 goto out;
542 }
543
544 /*
545 * Ok, we need to grow..
546 */
547 vma = vma_to_resize(addr, old_len, new_len, &charged);
548 if (IS_ERR(vma)) {
549 ret = PTR_ERR(vma);
550 goto out;
551 }
552
553 /* old_len exactly to the end of the area..
554 */
555 if (old_len == vma->vm_end - addr) {
556 /* can we just expand the current mapping? */
557 if (vma_expandable(vma, new_len - old_len)) {
558 int pages = (new_len - old_len) >> PAGE_SHIFT;
559
560 if (vma_adjust(vma, vma->vm_start, addr + new_len,
561 vma->vm_pgoff, NULL)) {
562 ret = -ENOMEM;
563 goto out;
564 }
565
566 vm_stat_account(mm, vma->vm_flags, pages);
567 if (vma->vm_flags & VM_LOCKED) {
568 mm->locked_vm += pages;
569 locked = true;
570 new_addr = addr;
571 }
572 ret = addr;
573 goto out;
574 }
575 }
576
577 /*
578 * We weren't able to just expand or shrink the area,
579 * we need to create a new one and move it..
580 */
581 ret = -ENOMEM;
582 if (flags & MREMAP_MAYMOVE) {
583 unsigned long map_flags = 0;
584 if (vma->vm_flags & VM_MAYSHARE)
585 map_flags |= MAP_SHARED;
586
587 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
588 vma->vm_pgoff +
589 ((addr - vma->vm_start) >> PAGE_SHIFT),
590 map_flags);
591 if (offset_in_page(new_addr)) {
592 ret = new_addr;
593 goto out;
594 }
595
596 ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
597 }
598 out:
599 if (offset_in_page(ret)) {
600 vm_unacct_memory(charged);
601 locked = 0;
602 }
603 up_write(¤t->mm->mmap_sem);
604 if (locked && new_len > old_len)
605 mm_populate(new_addr + old_len, new_len - old_len);
606 return ret;
607 }
608