xref: /linux/mm/mremap.c (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
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 #include <linux/userfaultfd_k.h>
26 
27 #include <asm/cacheflush.h>
28 #include <asm/tlbflush.h>
29 
30 #include "internal.h"
31 
32 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
33 {
34 	pgd_t *pgd;
35 	p4d_t *p4d;
36 	pud_t *pud;
37 	pmd_t *pmd;
38 
39 	pgd = pgd_offset(mm, addr);
40 	if (pgd_none_or_clear_bad(pgd))
41 		return NULL;
42 
43 	p4d = p4d_offset(pgd, addr);
44 	if (p4d_none_or_clear_bad(p4d))
45 		return NULL;
46 
47 	pud = pud_offset(p4d, addr);
48 	if (pud_none_or_clear_bad(pud))
49 		return NULL;
50 
51 	pmd = pmd_offset(pud, addr);
52 	if (pmd_none(*pmd))
53 		return NULL;
54 
55 	return pmd;
56 }
57 
58 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
59 			    unsigned long addr)
60 {
61 	pgd_t *pgd;
62 	p4d_t *p4d;
63 	pud_t *pud;
64 	pmd_t *pmd;
65 
66 	pgd = pgd_offset(mm, addr);
67 	p4d = p4d_alloc(mm, pgd, addr);
68 	if (!p4d)
69 		return NULL;
70 	pud = pud_alloc(mm, p4d, addr);
71 	if (!pud)
72 		return NULL;
73 
74 	pmd = pmd_alloc(mm, pud, addr);
75 	if (!pmd)
76 		return NULL;
77 
78 	VM_BUG_ON(pmd_trans_huge(*pmd));
79 
80 	return pmd;
81 }
82 
83 static void take_rmap_locks(struct vm_area_struct *vma)
84 {
85 	if (vma->vm_file)
86 		i_mmap_lock_write(vma->vm_file->f_mapping);
87 	if (vma->anon_vma)
88 		anon_vma_lock_write(vma->anon_vma);
89 }
90 
91 static void drop_rmap_locks(struct vm_area_struct *vma)
92 {
93 	if (vma->anon_vma)
94 		anon_vma_unlock_write(vma->anon_vma);
95 	if (vma->vm_file)
96 		i_mmap_unlock_write(vma->vm_file->f_mapping);
97 }
98 
99 static pte_t move_soft_dirty_pte(pte_t pte)
100 {
101 	/*
102 	 * Set soft dirty bit so we can notice
103 	 * in userspace the ptes were moved.
104 	 */
105 #ifdef CONFIG_MEM_SOFT_DIRTY
106 	if (pte_present(pte))
107 		pte = pte_mksoft_dirty(pte);
108 	else if (is_swap_pte(pte))
109 		pte = pte_swp_mksoft_dirty(pte);
110 #endif
111 	return pte;
112 }
113 
114 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
115 		unsigned long old_addr, unsigned long old_end,
116 		struct vm_area_struct *new_vma, pmd_t *new_pmd,
117 		unsigned long new_addr, bool need_rmap_locks, bool *need_flush)
118 {
119 	struct mm_struct *mm = vma->vm_mm;
120 	pte_t *old_pte, *new_pte, pte;
121 	spinlock_t *old_ptl, *new_ptl;
122 	bool force_flush = false;
123 	unsigned long len = old_end - old_addr;
124 
125 	/*
126 	 * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
127 	 * locks to ensure that rmap will always observe either the old or the
128 	 * new ptes. This is the easiest way to avoid races with
129 	 * truncate_pagecache(), page migration, etc...
130 	 *
131 	 * When need_rmap_locks is false, we use other ways to avoid
132 	 * such races:
133 	 *
134 	 * - During exec() shift_arg_pages(), we use a specially tagged vma
135 	 *   which rmap call sites look for using is_vma_temporary_stack().
136 	 *
137 	 * - During mremap(), new_vma is often known to be placed after vma
138 	 *   in rmap traversal order. This ensures rmap will always observe
139 	 *   either the old pte, or the new pte, or both (the page table locks
140 	 *   serialize access to individual ptes, but only rmap traversal
141 	 *   order guarantees that we won't miss both the old and new ptes).
142 	 */
143 	if (need_rmap_locks)
144 		take_rmap_locks(vma);
145 
146 	/*
147 	 * We don't have to worry about the ordering of src and dst
148 	 * pte locks because exclusive mmap_sem prevents deadlock.
149 	 */
150 	old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
151 	new_pte = pte_offset_map(new_pmd, new_addr);
152 	new_ptl = pte_lockptr(mm, new_pmd);
153 	if (new_ptl != old_ptl)
154 		spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
155 	arch_enter_lazy_mmu_mode();
156 
157 	for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
158 				   new_pte++, new_addr += PAGE_SIZE) {
159 		if (pte_none(*old_pte))
160 			continue;
161 
162 		pte = ptep_get_and_clear(mm, old_addr, old_pte);
163 		/*
164 		 * If we are remapping a dirty PTE, make sure
165 		 * to flush TLB before we drop the PTL for the
166 		 * old PTE or we may race with page_mkclean().
167 		 *
168 		 * This check has to be done after we removed the
169 		 * old PTE from page tables or another thread may
170 		 * dirty it after the check and before the removal.
171 		 */
172 		if (pte_present(pte) && pte_dirty(pte))
173 			force_flush = true;
174 		pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
175 		pte = move_soft_dirty_pte(pte);
176 		set_pte_at(mm, new_addr, new_pte, pte);
177 	}
178 
179 	arch_leave_lazy_mmu_mode();
180 	if (new_ptl != old_ptl)
181 		spin_unlock(new_ptl);
182 	pte_unmap(new_pte - 1);
183 	if (force_flush)
184 		flush_tlb_range(vma, old_end - len, old_end);
185 	else
186 		*need_flush = true;
187 	pte_unmap_unlock(old_pte - 1, old_ptl);
188 	if (need_rmap_locks)
189 		drop_rmap_locks(vma);
190 }
191 
192 #define LATENCY_LIMIT	(64 * PAGE_SIZE)
193 
194 unsigned long move_page_tables(struct vm_area_struct *vma,
195 		unsigned long old_addr, struct vm_area_struct *new_vma,
196 		unsigned long new_addr, unsigned long len,
197 		bool need_rmap_locks)
198 {
199 	unsigned long extent, next, old_end;
200 	pmd_t *old_pmd, *new_pmd;
201 	bool need_flush = false;
202 	unsigned long mmun_start;	/* For mmu_notifiers */
203 	unsigned long mmun_end;		/* For mmu_notifiers */
204 
205 	old_end = old_addr + len;
206 	flush_cache_range(vma, old_addr, old_end);
207 
208 	mmun_start = old_addr;
209 	mmun_end   = old_end;
210 	mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
211 
212 	for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
213 		cond_resched();
214 		next = (old_addr + PMD_SIZE) & PMD_MASK;
215 		/* even if next overflowed, extent below will be ok */
216 		extent = next - old_addr;
217 		if (extent > old_end - old_addr)
218 			extent = old_end - old_addr;
219 		old_pmd = get_old_pmd(vma->vm_mm, old_addr);
220 		if (!old_pmd)
221 			continue;
222 		new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
223 		if (!new_pmd)
224 			break;
225 		if (pmd_trans_huge(*old_pmd)) {
226 			if (extent == HPAGE_PMD_SIZE) {
227 				bool moved;
228 				/* See comment in move_ptes() */
229 				if (need_rmap_locks)
230 					take_rmap_locks(vma);
231 				moved = move_huge_pmd(vma, old_addr, new_addr,
232 						    old_end, old_pmd, new_pmd,
233 						    &need_flush);
234 				if (need_rmap_locks)
235 					drop_rmap_locks(vma);
236 				if (moved)
237 					continue;
238 			}
239 			split_huge_pmd(vma, old_pmd, old_addr);
240 			if (pmd_trans_unstable(old_pmd))
241 				continue;
242 		}
243 		if (pte_alloc(new_vma->vm_mm, new_pmd, new_addr))
244 			break;
245 		next = (new_addr + PMD_SIZE) & PMD_MASK;
246 		if (extent > next - new_addr)
247 			extent = next - new_addr;
248 		if (extent > LATENCY_LIMIT)
249 			extent = LATENCY_LIMIT;
250 		move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
251 			  new_pmd, new_addr, need_rmap_locks, &need_flush);
252 	}
253 	if (need_flush)
254 		flush_tlb_range(vma, old_end-len, old_addr);
255 
256 	mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
257 
258 	return len + old_addr - old_end;	/* how much done */
259 }
260 
261 static unsigned long move_vma(struct vm_area_struct *vma,
262 		unsigned long old_addr, unsigned long old_len,
263 		unsigned long new_len, unsigned long new_addr,
264 		bool *locked, struct vm_userfaultfd_ctx *uf,
265 		struct list_head *uf_unmap)
266 {
267 	struct mm_struct *mm = vma->vm_mm;
268 	struct vm_area_struct *new_vma;
269 	unsigned long vm_flags = vma->vm_flags;
270 	unsigned long new_pgoff;
271 	unsigned long moved_len;
272 	unsigned long excess = 0;
273 	unsigned long hiwater_vm;
274 	int split = 0;
275 	int err;
276 	bool need_rmap_locks;
277 
278 	/*
279 	 * We'd prefer to avoid failure later on in do_munmap:
280 	 * which may split one vma into three before unmapping.
281 	 */
282 	if (mm->map_count >= sysctl_max_map_count - 3)
283 		return -ENOMEM;
284 
285 	/*
286 	 * Advise KSM to break any KSM pages in the area to be moved:
287 	 * it would be confusing if they were to turn up at the new
288 	 * location, where they happen to coincide with different KSM
289 	 * pages recently unmapped.  But leave vma->vm_flags as it was,
290 	 * so KSM can come around to merge on vma and new_vma afterwards.
291 	 */
292 	err = ksm_madvise(vma, old_addr, old_addr + old_len,
293 						MADV_UNMERGEABLE, &vm_flags);
294 	if (err)
295 		return err;
296 
297 	new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
298 	new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
299 			   &need_rmap_locks);
300 	if (!new_vma)
301 		return -ENOMEM;
302 
303 	moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
304 				     need_rmap_locks);
305 	if (moved_len < old_len) {
306 		err = -ENOMEM;
307 	} else if (vma->vm_ops && vma->vm_ops->mremap) {
308 		err = vma->vm_ops->mremap(new_vma);
309 	}
310 
311 	if (unlikely(err)) {
312 		/*
313 		 * On error, move entries back from new area to old,
314 		 * which will succeed since page tables still there,
315 		 * and then proceed to unmap new area instead of old.
316 		 */
317 		move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
318 				 true);
319 		vma = new_vma;
320 		old_len = new_len;
321 		old_addr = new_addr;
322 		new_addr = err;
323 	} else {
324 		mremap_userfaultfd_prep(new_vma, uf);
325 		arch_remap(mm, old_addr, old_addr + old_len,
326 			   new_addr, new_addr + new_len);
327 	}
328 
329 	/* Conceal VM_ACCOUNT so old reservation is not undone */
330 	if (vm_flags & VM_ACCOUNT) {
331 		vma->vm_flags &= ~VM_ACCOUNT;
332 		excess = vma->vm_end - vma->vm_start - old_len;
333 		if (old_addr > vma->vm_start &&
334 		    old_addr + old_len < vma->vm_end)
335 			split = 1;
336 	}
337 
338 	/*
339 	 * If we failed to move page tables we still do total_vm increment
340 	 * since do_munmap() will decrement it by old_len == new_len.
341 	 *
342 	 * Since total_vm is about to be raised artificially high for a
343 	 * moment, we need to restore high watermark afterwards: if stats
344 	 * are taken meanwhile, total_vm and hiwater_vm appear too high.
345 	 * If this were a serious issue, we'd add a flag to do_munmap().
346 	 */
347 	hiwater_vm = mm->hiwater_vm;
348 	vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
349 
350 	/* Tell pfnmap has moved from this vma */
351 	if (unlikely(vma->vm_flags & VM_PFNMAP))
352 		untrack_pfn_moved(vma);
353 
354 	if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
355 		/* OOM: unable to split vma, just get accounts right */
356 		vm_unacct_memory(excess >> PAGE_SHIFT);
357 		excess = 0;
358 	}
359 	mm->hiwater_vm = hiwater_vm;
360 
361 	/* Restore VM_ACCOUNT if one or two pieces of vma left */
362 	if (excess) {
363 		vma->vm_flags |= VM_ACCOUNT;
364 		if (split)
365 			vma->vm_next->vm_flags |= VM_ACCOUNT;
366 	}
367 
368 	if (vm_flags & VM_LOCKED) {
369 		mm->locked_vm += new_len >> PAGE_SHIFT;
370 		*locked = true;
371 	}
372 
373 	return new_addr;
374 }
375 
376 static struct vm_area_struct *vma_to_resize(unsigned long addr,
377 	unsigned long old_len, unsigned long new_len, unsigned long *p)
378 {
379 	struct mm_struct *mm = current->mm;
380 	struct vm_area_struct *vma = find_vma(mm, addr);
381 	unsigned long pgoff;
382 
383 	if (!vma || vma->vm_start > addr)
384 		return ERR_PTR(-EFAULT);
385 
386 	if (is_vm_hugetlb_page(vma))
387 		return ERR_PTR(-EINVAL);
388 
389 	/* We can't remap across vm area boundaries */
390 	if (old_len > vma->vm_end - addr)
391 		return ERR_PTR(-EFAULT);
392 
393 	if (new_len == old_len)
394 		return vma;
395 
396 	/* Need to be careful about a growing mapping */
397 	pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
398 	pgoff += vma->vm_pgoff;
399 	if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
400 		return ERR_PTR(-EINVAL);
401 
402 	if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
403 		return ERR_PTR(-EFAULT);
404 
405 	if (vma->vm_flags & VM_LOCKED) {
406 		unsigned long locked, lock_limit;
407 		locked = mm->locked_vm << PAGE_SHIFT;
408 		lock_limit = rlimit(RLIMIT_MEMLOCK);
409 		locked += new_len - old_len;
410 		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
411 			return ERR_PTR(-EAGAIN);
412 	}
413 
414 	if (!may_expand_vm(mm, vma->vm_flags,
415 				(new_len - old_len) >> PAGE_SHIFT))
416 		return ERR_PTR(-ENOMEM);
417 
418 	if (vma->vm_flags & VM_ACCOUNT) {
419 		unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
420 		if (security_vm_enough_memory_mm(mm, charged))
421 			return ERR_PTR(-ENOMEM);
422 		*p = charged;
423 	}
424 
425 	return vma;
426 }
427 
428 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
429 		unsigned long new_addr, unsigned long new_len, bool *locked,
430 		struct vm_userfaultfd_ctx *uf,
431 		struct list_head *uf_unmap)
432 {
433 	struct mm_struct *mm = current->mm;
434 	struct vm_area_struct *vma;
435 	unsigned long ret = -EINVAL;
436 	unsigned long charged = 0;
437 	unsigned long map_flags;
438 
439 	if (offset_in_page(new_addr))
440 		goto out;
441 
442 	if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
443 		goto out;
444 
445 	/* Ensure the old/new locations do not overlap */
446 	if (addr + old_len > new_addr && new_addr + new_len > addr)
447 		goto out;
448 
449 	ret = do_munmap(mm, new_addr, new_len, NULL);
450 	if (ret)
451 		goto out;
452 
453 	if (old_len >= new_len) {
454 		ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
455 		if (ret && old_len != new_len)
456 			goto out;
457 		old_len = new_len;
458 	}
459 
460 	vma = vma_to_resize(addr, old_len, new_len, &charged);
461 	if (IS_ERR(vma)) {
462 		ret = PTR_ERR(vma);
463 		goto out;
464 	}
465 
466 	map_flags = MAP_FIXED;
467 	if (vma->vm_flags & VM_MAYSHARE)
468 		map_flags |= MAP_SHARED;
469 
470 	ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
471 				((addr - vma->vm_start) >> PAGE_SHIFT),
472 				map_flags);
473 	if (offset_in_page(ret))
474 		goto out1;
475 
476 	ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, uf,
477 		       uf_unmap);
478 	if (!(offset_in_page(ret)))
479 		goto out;
480 out1:
481 	vm_unacct_memory(charged);
482 
483 out:
484 	return ret;
485 }
486 
487 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
488 {
489 	unsigned long end = vma->vm_end + delta;
490 	if (end < vma->vm_end) /* overflow */
491 		return 0;
492 	if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
493 		return 0;
494 	if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
495 			      0, MAP_FIXED) & ~PAGE_MASK)
496 		return 0;
497 	return 1;
498 }
499 
500 /*
501  * Expand (or shrink) an existing mapping, potentially moving it at the
502  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
503  *
504  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
505  * This option implies MREMAP_MAYMOVE.
506  */
507 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
508 		unsigned long, new_len, unsigned long, flags,
509 		unsigned long, new_addr)
510 {
511 	struct mm_struct *mm = current->mm;
512 	struct vm_area_struct *vma;
513 	unsigned long ret = -EINVAL;
514 	unsigned long charged = 0;
515 	bool locked = false;
516 	struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
517 	LIST_HEAD(uf_unmap);
518 
519 	if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
520 		return ret;
521 
522 	if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
523 		return ret;
524 
525 	if (offset_in_page(addr))
526 		return ret;
527 
528 	old_len = PAGE_ALIGN(old_len);
529 	new_len = PAGE_ALIGN(new_len);
530 
531 	/*
532 	 * We allow a zero old-len as a special case
533 	 * for DOS-emu "duplicate shm area" thing. But
534 	 * a zero new-len is nonsensical.
535 	 */
536 	if (!new_len)
537 		return ret;
538 
539 	if (down_write_killable(&current->mm->mmap_sem))
540 		return -EINTR;
541 
542 	if (flags & MREMAP_FIXED) {
543 		ret = mremap_to(addr, old_len, new_addr, new_len,
544 				&locked, &uf, &uf_unmap);
545 		goto out;
546 	}
547 
548 	/*
549 	 * Always allow a shrinking remap: that just unmaps
550 	 * the unnecessary pages..
551 	 * do_munmap does all the needed commit accounting
552 	 */
553 	if (old_len >= new_len) {
554 		ret = do_munmap(mm, addr+new_len, old_len - new_len, &uf_unmap);
555 		if (ret && old_len != new_len)
556 			goto out;
557 		ret = addr;
558 		goto out;
559 	}
560 
561 	/*
562 	 * Ok, we need to grow..
563 	 */
564 	vma = vma_to_resize(addr, old_len, new_len, &charged);
565 	if (IS_ERR(vma)) {
566 		ret = PTR_ERR(vma);
567 		goto out;
568 	}
569 
570 	/* old_len exactly to the end of the area..
571 	 */
572 	if (old_len == vma->vm_end - addr) {
573 		/* can we just expand the current mapping? */
574 		if (vma_expandable(vma, new_len - old_len)) {
575 			int pages = (new_len - old_len) >> PAGE_SHIFT;
576 
577 			if (vma_adjust(vma, vma->vm_start, addr + new_len,
578 				       vma->vm_pgoff, NULL)) {
579 				ret = -ENOMEM;
580 				goto out;
581 			}
582 
583 			vm_stat_account(mm, vma->vm_flags, pages);
584 			if (vma->vm_flags & VM_LOCKED) {
585 				mm->locked_vm += pages;
586 				locked = true;
587 				new_addr = addr;
588 			}
589 			ret = addr;
590 			goto out;
591 		}
592 	}
593 
594 	/*
595 	 * We weren't able to just expand or shrink the area,
596 	 * we need to create a new one and move it..
597 	 */
598 	ret = -ENOMEM;
599 	if (flags & MREMAP_MAYMOVE) {
600 		unsigned long map_flags = 0;
601 		if (vma->vm_flags & VM_MAYSHARE)
602 			map_flags |= MAP_SHARED;
603 
604 		new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
605 					vma->vm_pgoff +
606 					((addr - vma->vm_start) >> PAGE_SHIFT),
607 					map_flags);
608 		if (offset_in_page(new_addr)) {
609 			ret = new_addr;
610 			goto out;
611 		}
612 
613 		ret = move_vma(vma, addr, old_len, new_len, new_addr,
614 			       &locked, &uf, &uf_unmap);
615 	}
616 out:
617 	if (offset_in_page(ret)) {
618 		vm_unacct_memory(charged);
619 		locked = 0;
620 	}
621 	up_write(&current->mm->mmap_sem);
622 	if (locked && new_len > old_len)
623 		mm_populate(new_addr + old_len, new_len - old_len);
624 	mremap_userfaultfd_complete(&uf, addr, new_addr, old_len);
625 	userfaultfd_unmap_complete(mm, &uf_unmap);
626 	return ret;
627 }
628