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/mm_inline.h>
13	#include <linux/hugetlb.h>
14	#include <linux/shm.h>
15	#include <linux/ksm.h>
16	#include <linux/mman.h>
17	#include <linux/swap.h>
18	#include <linux/capability.h>
19	#include <linux/fs.h>
20	#include <linux/swapops.h>
21	#include <linux/highmem.h>
22	#include <linux/security.h>
23	#include <linux/syscalls.h>
24	#include <linux/mmu_notifier.h>
25	#include <linux/uaccess.h>
26	#include <linux/userfaultfd_k.h>
27	#include <linux/mempolicy.h>
28
29	#include <asm/cacheflush.h>
30	#include <asm/tlb.h>
31	#include <asm/pgalloc.h>
32
33	#include "internal.h"
34
35	/* Classify the kind of remap operation being performed. */
36	enum mremap_type {
37	MREMAP_INVALID, /* Initial state. */
38	MREMAP_NO_RESIZE, /* old_len == new_len, if not moved, do nothing. */
39	MREMAP_SHRINK, /* old_len > new_len. */
40	MREMAP_EXPAND, /* old_len < new_len. */
41	};
42
43	/*
44	* Describes a VMA mremap() operation and is threaded throughout it.
45	*
46	* Any of the fields may be mutated by the operation, however these values will
47	* always accurately reflect the remap (for instance, we may adjust lengths and
48	* delta to account for hugetlb alignment).
49	*/
50	struct vma_remap_struct {
51	/* User-provided state. */
52	unsigned long addr; /* User-specified address from which we remap. */
53	unsigned long old_len; /* Length of range being remapped. */
54	unsigned long new_len; /* Desired new length of mapping. */
55	unsigned long flags; /* user-specified MREMAP_* flags. */
56	unsigned long new_addr; /* Optionally, desired new address. */
57
58	/* uffd state. */
59	struct vm_userfaultfd_ctx *uf;
60	struct list_head *uf_unmap_early;
61	struct list_head *uf_unmap;
62
63	/* VMA state, determined in do_mremap(). */
64	struct vm_area_struct *vma;
65
66	/* Internal state, determined in do_mremap(). */
67	unsigned long delta; /* Absolute delta of old_len,new_len. */
68	bool mlocked; /* Was the VMA mlock()'d? */
69	enum mremap_type remap_type; /* expand, shrink, etc. */
70	bool mmap_locked; /* Is mm currently write-locked? */
71	unsigned long charged; /* If VM_ACCOUNT, # pages to account. */
72	};
73
74	static pud_t *get_old_pud(struct mm_struct *mm, unsigned long addr)
75	{
76	pgd_t *pgd;
77	p4d_t *p4d;
78	pud_t *pud;
79
80	pgd = pgd_offset(mm, addr);
81	if (pgd_none_or_clear_bad(pgd))
82	return NULL;
83
84	p4d = p4d_offset(pgd, address: addr);
85	if (p4d_none_or_clear_bad(p4d))
86	return NULL;
87
88	pud = pud_offset(p4d, address: addr);
89	if (pud_none_or_clear_bad(pud))
90	return NULL;
91
92	return pud;
93	}
94
95	static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
96	{
97	pud_t *pud;
98	pmd_t *pmd;
99
100	pud = get_old_pud(mm, addr);
101	if (!pud)
102	return NULL;
103
104	pmd = pmd_offset(pud, address: addr);
105	if (pmd_none(pmd: *pmd))
106	return NULL;
107
108	return pmd;
109	}
110
111	static pud_t *alloc_new_pud(struct mm_struct *mm, unsigned long addr)
112	{
113	pgd_t *pgd;
114	p4d_t *p4d;
115
116	pgd = pgd_offset(mm, addr);
117	p4d = p4d_alloc(mm, pgd, address: addr);
118	if (!p4d)
119	return NULL;
120
121	return pud_alloc(mm, p4d, address: addr);
122	}
123
124	static pmd_t *alloc_new_pmd(struct mm_struct *mm, unsigned long addr)
125	{
126	pud_t *pud;
127	pmd_t *pmd;
128
129	pud = alloc_new_pud(mm, addr);
130	if (!pud)
131	return NULL;
132
133	pmd = pmd_alloc(mm, pud, address: addr);
134	if (!pmd)
135	return NULL;
136
137	VM_BUG_ON(pmd_trans_huge(*pmd));
138
139	return pmd;
140	}
141
142	static void take_rmap_locks(struct vm_area_struct *vma)
143	{
144	if (vma->vm_file)
145	i_mmap_lock_write(mapping: vma->vm_file->f_mapping);
146	if (vma->anon_vma)
147	anon_vma_lock_write(anon_vma: vma->anon_vma);
148	}
149
150	static void drop_rmap_locks(struct vm_area_struct *vma)
151	{
152	if (vma->anon_vma)
153	anon_vma_unlock_write(anon_vma: vma->anon_vma);
154	if (vma->vm_file)
155	i_mmap_unlock_write(mapping: vma->vm_file->f_mapping);
156	}
157
158	static pte_t move_soft_dirty_pte(pte_t pte)
159	{
160	/*
161	* Set soft dirty bit so we can notice
162	* in userspace the ptes were moved.
163	*/
164	#ifdef CONFIG_MEM_SOFT_DIRTY
165	if (pte_present(a: pte))
166	pte = pte_mksoft_dirty(pte);
167	else if (is_swap_pte(pte))
168	pte = pte_swp_mksoft_dirty(pte);
169	#endif
170	return pte;
171	}
172
173	static int move_ptes(struct pagetable_move_control *pmc,
174	unsigned long extent, pmd_t *old_pmd, pmd_t *new_pmd)
175	{
176	struct vm_area_struct *vma = pmc->old;
177	bool need_clear_uffd_wp = vma_has_uffd_without_event_remap(vma);
178	struct mm_struct *mm = vma->vm_mm;
179	pte_t *old_pte, *new_pte, pte;
180	pmd_t dummy_pmdval;
181	spinlock_t *old_ptl, *new_ptl;
182	bool force_flush = false;
183	unsigned long old_addr = pmc->old_addr;
184	unsigned long new_addr = pmc->new_addr;
185	unsigned long old_end = old_addr + extent;
186	unsigned long len = old_end - old_addr;
187	int err = 0;
188
189	/*
190	* When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
191	* locks to ensure that rmap will always observe either the old or the
192	* new ptes. This is the easiest way to avoid races with
193	* truncate_pagecache(), page migration, etc...
194	*
195	* When need_rmap_locks is false, we use other ways to avoid
196	* such races:
197	*
198	* - During exec() shift_arg_pages(), we use a specially tagged vma
199	* which rmap call sites look for using vma_is_temporary_stack().
200	*
201	* - During mremap(), new_vma is often known to be placed after vma
202	* in rmap traversal order. This ensures rmap will always observe
203	* either the old pte, or the new pte, or both (the page table locks
204	* serialize access to individual ptes, but only rmap traversal
205	* order guarantees that we won't miss both the old and new ptes).
206	*/
207	if (pmc->need_rmap_locks)
208	take_rmap_locks(vma);
209
210	/*
211	* We don't have to worry about the ordering of src and dst
212	* pte locks because exclusive mmap_lock prevents deadlock.
213	*/
214	old_pte = pte_offset_map_lock(mm, pmd: old_pmd, addr: old_addr, ptlp: &old_ptl);
215	if (!old_pte) {
216	err = -EAGAIN;
217	goto out;
218	}
219	/*
220	* Now new_pte is none, so hpage_collapse_scan_file() path can not find
221	* this by traversing file->f_mapping, so there is no concurrency with
222	* retract_page_tables(). In addition, we already hold the exclusive
223	* mmap_lock, so this new_pte page is stable, so there is no need to get
224	* pmdval and do pmd_same() check.
225	*/
226	new_pte = pte_offset_map_rw_nolock(mm, pmd: new_pmd, addr: new_addr, pmdvalp: &dummy_pmdval,
227	ptlp: &new_ptl);
228	if (!new_pte) {
229	pte_unmap_unlock(old_pte, old_ptl);
230	err = -EAGAIN;
231	goto out;
232	}
233	if (new_ptl != old_ptl)
234	spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
235	flush_tlb_batched_pending(mm: vma->vm_mm);
236	arch_enter_lazy_mmu_mode();
237
238	for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
239	new_pte++, new_addr += PAGE_SIZE) {
240	VM_WARN_ON_ONCE(!pte_none(*new_pte));
241
242	if (pte_none(pte: ptep_get(ptep: old_pte)))
243	continue;
244
245	pte = ptep_get_and_clear(mm, addr: old_addr, ptep: old_pte);
246	/*
247	* If we are remapping a valid PTE, make sure
248	* to flush TLB before we drop the PTL for the
249	* PTE.
250	*
251	* NOTE! Both old and new PTL matter: the old one
252	* for racing with folio_mkclean(), the new one to
253	* make sure the physical page stays valid until
254	* the TLB entry for the old mapping has been
255	* flushed.
256	*/
257	if (pte_present(a: pte))
258	force_flush = true;
259	pte = move_pte(pte, old_addr, new_addr);
260	pte = move_soft_dirty_pte(pte);
261
262	if (need_clear_uffd_wp && pte_marker_uffd_wp(pte))
263	pte_clear(mm, addr: new_addr, ptep: new_pte);
264	else {
265	if (need_clear_uffd_wp) {
266	if (pte_present(a: pte))
267	pte = pte_clear_uffd_wp(pte);
268	else if (is_swap_pte(pte))
269	pte = pte_swp_clear_uffd_wp(pte);
270	}
271	set_pte_at(mm, new_addr, new_pte, pte);
272	}
273	}
274
275	arch_leave_lazy_mmu_mode();
276	if (force_flush)
277	flush_tlb_range(vma, old_end - len, old_end);
278	if (new_ptl != old_ptl)
279	spin_unlock(lock: new_ptl);
280	pte_unmap(pte: new_pte - 1);
281	pte_unmap_unlock(old_pte - 1, old_ptl);
282	out:
283	if (pmc->need_rmap_locks)
284	drop_rmap_locks(vma);
285	return err;
286	}
287
288	#ifndef arch_supports_page_table_move
289	#define arch_supports_page_table_move arch_supports_page_table_move
290	static inline bool arch_supports_page_table_move(void)
291	{
292	return IS_ENABLED(CONFIG_HAVE_MOVE_PMD) ||
293	IS_ENABLED(CONFIG_HAVE_MOVE_PUD);
294	}
295	#endif
296
297	#ifdef CONFIG_HAVE_MOVE_PMD
298	static bool move_normal_pmd(struct pagetable_move_control *pmc,
299	pmd_t *old_pmd, pmd_t *new_pmd)
300	{
301	spinlock_t *old_ptl, *new_ptl;
302	struct vm_area_struct *vma = pmc->old;
303	struct mm_struct *mm = vma->vm_mm;
304	bool res = false;
305	pmd_t pmd;
306
307	if (!arch_supports_page_table_move())
308	return false;
309	/*
310	* The destination pmd shouldn't be established, free_pgtables()
311	* should have released it.
312	*
313	* However, there's a case during execve() where we use mremap
314	* to move the initial stack, and in that case the target area
315	* may overlap the source area (always moving down).
316	*
317	* If everything is PMD-aligned, that works fine, as moving
318	* each pmd down will clear the source pmd. But if we first
319	* have a few 4kB-only pages that get moved down, and then
320	* hit the "now the rest is PMD-aligned, let's do everything
321	* one pmd at a time", we will still have the old (now empty
322	* of any 4kB pages, but still there) PMD in the page table
323	* tree.
324	*
325	* Warn on it once - because we really should try to figure
326	* out how to do this better - but then say "I won't move
327	* this pmd".
328	*
329	* One alternative might be to just unmap the target pmd at
330	* this point, and verify that it really is empty. We'll see.
331	*/
332	if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
333	return false;
334
335	/* If this pmd belongs to a uffd vma with remap events disabled, we need
336	* to ensure that the uffd-wp state is cleared from all pgtables. This
337	* means recursing into lower page tables in move_page_tables(), and we
338	* can reuse the existing code if we simply treat the entry as "not
339	* moved".
340	*/
341	if (vma_has_uffd_without_event_remap(vma))
342	return false;
343
344	/*
345	* We don't have to worry about the ordering of src and dst
346	* ptlocks because exclusive mmap_lock prevents deadlock.
347	*/
348	old_ptl = pmd_lock(mm, pmd: old_pmd);
349	new_ptl = pmd_lockptr(mm, pmd: new_pmd);
350	if (new_ptl != old_ptl)
351	spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
352
353	pmd = *old_pmd;
354
355	/* Racing with collapse? */
356	if (unlikely(!pmd_present(pmd) || pmd_leaf(pmd)))
357	goto out_unlock;
358	/* Clear the pmd */
359	pmd_clear(pmdp: old_pmd);
360	res = true;
361
362	VM_BUG_ON(!pmd_none(*new_pmd));
363
364	pmd_populate(mm, pmd: new_pmd, pmd_pgtable(pmd));
365	flush_tlb_range(vma, pmc->old_addr, pmc->old_addr + PMD_SIZE);
366	out_unlock:
367	if (new_ptl != old_ptl)
368	spin_unlock(lock: new_ptl);
369	spin_unlock(lock: old_ptl);
370
371	return res;
372	}
373	#else
374	static inline bool move_normal_pmd(struct pagetable_move_control *pmc,
375	pmd_t *old_pmd, pmd_t *new_pmd)
376	{
377	return false;
378	}
379	#endif
380
381	#if CONFIG_PGTABLE_LEVELS > 2 && defined(CONFIG_HAVE_MOVE_PUD)
382	static bool move_normal_pud(struct pagetable_move_control *pmc,
383	pud_t *old_pud, pud_t *new_pud)
384	{
385	spinlock_t *old_ptl, *new_ptl;
386	struct vm_area_struct *vma = pmc->old;
387	struct mm_struct *mm = vma->vm_mm;
388	pud_t pud;
389
390	if (!arch_supports_page_table_move())
391	return false;
392	/*
393	* The destination pud shouldn't be established, free_pgtables()
394	* should have released it.
395	*/
396	if (WARN_ON_ONCE(!pud_none(*new_pud)))
397	return false;
398
399	/* If this pud belongs to a uffd vma with remap events disabled, we need
400	* to ensure that the uffd-wp state is cleared from all pgtables. This
401	* means recursing into lower page tables in move_page_tables(), and we
402	* can reuse the existing code if we simply treat the entry as "not
403	* moved".
404	*/
405	if (vma_has_uffd_without_event_remap(vma))
406	return false;
407
408	/*
409	* We don't have to worry about the ordering of src and dst
410	* ptlocks because exclusive mmap_lock prevents deadlock.
411	*/
412	old_ptl = pud_lock(mm, pud: old_pud);
413	new_ptl = pud_lockptr(mm, pud: new_pud);
414	if (new_ptl != old_ptl)
415	spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
416
417	/* Clear the pud */
418	pud = *old_pud;
419	pud_clear(pudp: old_pud);
420
421	VM_BUG_ON(!pud_none(*new_pud));
422
423	pud_populate(mm, pud: new_pud, pmd: pud_pgtable(pud));
424	flush_tlb_range(vma, pmc->old_addr, pmc->old_addr + PUD_SIZE);
425	if (new_ptl != old_ptl)
426	spin_unlock(lock: new_ptl);
427	spin_unlock(lock: old_ptl);
428
429	return true;
430	}
431	#else
432	static inline bool move_normal_pud(struct pagetable_move_control *pmc,
433	pud_t *old_pud, pud_t *new_pud)
434	{
435	return false;
436	}
437	#endif
438
439	#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
440	static bool move_huge_pud(struct pagetable_move_control *pmc,
441	pud_t *old_pud, pud_t *new_pud)
442	{
443	spinlock_t *old_ptl, *new_ptl;
444	struct vm_area_struct *vma = pmc->old;
445	struct mm_struct *mm = vma->vm_mm;
446	pud_t pud;
447
448	/*
449	* The destination pud shouldn't be established, free_pgtables()
450	* should have released it.
451	*/
452	if (WARN_ON_ONCE(!pud_none(*new_pud)))
453	return false;
454
455	/*
456	* We don't have to worry about the ordering of src and dst
457	* ptlocks because exclusive mmap_lock prevents deadlock.
458	*/
459	old_ptl = pud_lock(mm, pud: old_pud);
460	new_ptl = pud_lockptr(mm, pud: new_pud);
461	if (new_ptl != old_ptl)
462	spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
463
464	/* Clear the pud */
465	pud = *old_pud;
466	pud_clear(pudp: old_pud);
467
468	VM_BUG_ON(!pud_none(*new_pud));
469
470	/* Set the new pud */
471	/* mark soft_ditry when we add pud level soft dirty support */
472	set_pud_at(mm, addr: pmc->new_addr, pudp: new_pud, pud);
473	flush_pud_tlb_range(vma, pmc->old_addr, pmc->old_addr + HPAGE_PUD_SIZE);
474	if (new_ptl != old_ptl)
475	spin_unlock(lock: new_ptl);
476	spin_unlock(lock: old_ptl);
477
478	return true;
479	}
480	#else
481	static bool move_huge_pud(struct pagetable_move_control *pmc,
482	pud_t *old_pud, pud_t *new_pud)
483
484	{
485	WARN_ON_ONCE(1);
486	return false;
487
488	}
489	#endif
490
491	enum pgt_entry {
492	NORMAL_PMD,
493	HPAGE_PMD,
494	NORMAL_PUD,
495	HPAGE_PUD,
496	};
497
498	/*
499	* Returns an extent of the corresponding size for the pgt_entry specified if
500	* valid. Else returns a smaller extent bounded by the end of the source and
501	* destination pgt_entry.
502	*/
503	static __always_inline unsigned long get_extent(enum pgt_entry entry,
504	struct pagetable_move_control *pmc)
505	{
506	unsigned long next, extent, mask, size;
507	unsigned long old_addr = pmc->old_addr;
508	unsigned long old_end = pmc->old_end;
509	unsigned long new_addr = pmc->new_addr;
510
511	switch (entry) {
512	case HPAGE_PMD:
513	case NORMAL_PMD:
514	mask = PMD_MASK;
515	size = PMD_SIZE;
516	break;
517	case HPAGE_PUD:
518	case NORMAL_PUD:
519	mask = PUD_MASK;
520	size = PUD_SIZE;
521	break;
522	default:
523	BUILD_BUG();
524	break;
525	}
526
527	next = (old_addr + size) & mask;
528	/* even if next overflowed, extent below will be ok */
529	extent = next - old_addr;
530	if (extent > old_end - old_addr)
531	extent = old_end - old_addr;
532	next = (new_addr + size) & mask;
533	if (extent > next - new_addr)
534	extent = next - new_addr;
535	return extent;
536	}
537
538	/*
539	* Should move_pgt_entry() acquire the rmap locks? This is either expressed in
540	* the PMC, or overridden in the case of normal, larger page tables.
541	*/
542	static bool should_take_rmap_locks(struct pagetable_move_control *pmc,
543	enum pgt_entry entry)
544	{
545	switch (entry) {
546	case NORMAL_PMD:
547	case NORMAL_PUD:
548	return true;
549	default:
550	return pmc->need_rmap_locks;
551	}
552	}
553
554	/*
555	* Attempts to speedup the move by moving entry at the level corresponding to
556	* pgt_entry. Returns true if the move was successful, else false.
557	*/
558	static bool move_pgt_entry(struct pagetable_move_control *pmc,
559	enum pgt_entry entry, void *old_entry, void *new_entry)
560	{
561	bool moved = false;
562	bool need_rmap_locks = should_take_rmap_locks(pmc, entry);
563
564	/* See comment in move_ptes() */
565	if (need_rmap_locks)
566	take_rmap_locks(vma: pmc->old);
567
568	switch (entry) {
569	case NORMAL_PMD:
570	moved = move_normal_pmd(pmc, old_pmd: old_entry, new_pmd: new_entry);
571	break;
572	case NORMAL_PUD:
573	moved = move_normal_pud(pmc, old_pud: old_entry, new_pud: new_entry);
574	break;
575	case HPAGE_PMD:
576	moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
577	move_huge_pmd(vma: pmc->old, old_addr: pmc->old_addr, new_addr: pmc->new_addr, old_pmd: old_entry,
578	new_pmd: new_entry);
579	break;
580	case HPAGE_PUD:
581	moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
582	move_huge_pud(pmc, old_pud: old_entry, new_pud: new_entry);
583	break;
584
585	default:
586	WARN_ON_ONCE(1);
587	break;
588	}
589
590	if (need_rmap_locks)
591	drop_rmap_locks(vma: pmc->old);
592
593	return moved;
594	}
595
596	/*
597	* A helper to check if aligning down is OK. The aligned address should fall
598	* on *no mapping*. For the stack moving down, that's a special move within
599	* the VMA that is created to span the source and destination of the move,
600	* so we make an exception for it.
601	*/
602	static bool can_align_down(struct pagetable_move_control *pmc,
603	struct vm_area_struct *vma, unsigned long addr_to_align,
604	unsigned long mask)
605	{
606	unsigned long addr_masked = addr_to_align & mask;
607
608	/*
609	* If @addr_to_align of either source or destination is not the beginning
610	* of the corresponding VMA, we can't align down or we will destroy part
611	* of the current mapping.
612	*/
613	if (!pmc->for_stack && vma->vm_start != addr_to_align)
614	return false;
615
616	/* In the stack case we explicitly permit in-VMA alignment. */
617	if (pmc->for_stack && addr_masked >= vma->vm_start)
618	return true;
619
620	/*
621	* Make sure the realignment doesn't cause the address to fall on an
622	* existing mapping.
623	*/
624	return find_vma_intersection(mm: vma->vm_mm, start_addr: addr_masked, end_addr: vma->vm_start) == NULL;
625	}
626
627	/*
628	* Determine if are in fact able to realign for efficiency to a higher page
629	* table boundary.
630	*/
631	static bool can_realign_addr(struct pagetable_move_control *pmc,
632	unsigned long pagetable_mask)
633	{
634	unsigned long align_mask = ~pagetable_mask;
635	unsigned long old_align = pmc->old_addr & align_mask;
636	unsigned long new_align = pmc->new_addr & align_mask;
637	unsigned long pagetable_size = align_mask + 1;
638	unsigned long old_align_next = pagetable_size - old_align;
639
640	/*
641	* We don't want to have to go hunting for VMAs from the end of the old
642	* VMA to the next page table boundary, also we want to make sure the
643	* operation is wortwhile.
644	*
645	* So ensure that we only perform this realignment if the end of the
646	* range being copied reaches or crosses the page table boundary.
647	*
648	* boundary boundary
649	* .<- old_align -> .
650	* . |----------------.-----------|
651	* . | vma . |
652	* . |----------------.-----------|
653	* . <----------------.----------->
654	* . len_in
655	* <------------------------------->
656	* . pagetable_size .
657	* . <---------------->
658	* . old_align_next .
659	*/
660	if (pmc->len_in < old_align_next)
661	return false;
662
663	/* Skip if the addresses are already aligned. */
664	if (old_align == 0)
665	return false;
666
667	/* Only realign if the new and old addresses are mutually aligned. */
668	if (old_align != new_align)
669	return false;
670
671	/* Ensure realignment doesn't cause overlap with existing mappings. */
672	if (!can_align_down(pmc, vma: pmc->old, addr_to_align: pmc->old_addr, mask: pagetable_mask) ||
673	!can_align_down(pmc, vma: pmc->new, addr_to_align: pmc->new_addr, mask: pagetable_mask))
674	return false;
675
676	return true;
677	}
678
679	/*
680	* Opportunistically realign to specified boundary for faster copy.
681	*
682	* Consider an mremap() of a VMA with page table boundaries as below, and no
683	* preceding VMAs from the lower page table boundary to the start of the VMA,
684	* with the end of the range reaching or crossing the page table boundary.
685	*
686	* boundary boundary
687	* . |----------------.-----------|
688	* . | vma . |
689	* . |----------------.-----------|
690	* . pmc->old_addr . pmc->old_end
691	* . <---------------------------->
692	* . move these page tables
693	*
694	* If we proceed with moving page tables in this scenario, we will have a lot of
695	* work to do traversing old page tables and establishing new ones in the
696	* destination across multiple lower level page tables.
697	*
698	* The idea here is simply to align pmc->old_addr, pmc->new_addr down to the
699	* page table boundary, so we can simply copy a single page table entry for the
700	* aligned portion of the VMA instead:
701	*
702	* boundary boundary
703	* . |----------------.-----------|
704	* . | vma . |
705	* . |----------------.-----------|
706	* pmc->old_addr . pmc->old_end
707	* <------------------------------------------->
708	* . move these page tables
709	*/
710	static void try_realign_addr(struct pagetable_move_control *pmc,
711	unsigned long pagetable_mask)
712	{
713
714	if (!can_realign_addr(pmc, pagetable_mask))
715	return;
716
717	/*
718	* Simply align to page table boundaries. Note that we do NOT update the
719	* pmc->old_end value, and since the move_page_tables() operation spans
720	* from [old_addr, old_end) (offsetting new_addr as it is performed),
721	* this simply changes the start of the copy, not the end.
722	*/
723	pmc->old_addr &= pagetable_mask;
724	pmc->new_addr &= pagetable_mask;
725	}
726
727	/* Is the page table move operation done? */
728	static bool pmc_done(struct pagetable_move_control *pmc)
729	{
730	return pmc->old_addr >= pmc->old_end;
731	}
732
733	/* Advance to the next page table, offset by extent bytes. */
734	static void pmc_next(struct pagetable_move_control *pmc, unsigned long extent)
735	{
736	pmc->old_addr += extent;
737	pmc->new_addr += extent;
738	}
739
740	/*
741	* Determine how many bytes in the specified input range have had their page
742	* tables moved so far.
743	*/
744	static unsigned long pmc_progress(struct pagetable_move_control *pmc)
745	{
746	unsigned long orig_old_addr = pmc->old_end - pmc->len_in;
747	unsigned long old_addr = pmc->old_addr;
748
749	/*
750	* Prevent negative return values when {old,new}_addr was realigned but
751	* we broke out of the loop in move_page_tables() for the first PMD
752	* itself.
753	*/
754	return old_addr < orig_old_addr ? 0 : old_addr - orig_old_addr;
755	}
756
757	unsigned long move_page_tables(struct pagetable_move_control *pmc)
758	{
759	unsigned long extent;
760	struct mmu_notifier_range range;
761	pmd_t *old_pmd, *new_pmd;
762	pud_t *old_pud, *new_pud;
763	struct mm_struct *mm = pmc->old->vm_mm;
764
765	if (!pmc->len_in)
766	return 0;
767
768	if (is_vm_hugetlb_page(vma: pmc->old))
769	return move_hugetlb_page_tables(vma: pmc->old, new_vma: pmc->new, old_addr: pmc->old_addr,
770	new_addr: pmc->new_addr, len: pmc->len_in);
771
772	/*
773	* If possible, realign addresses to PMD boundary for faster copy.
774	* Only realign if the mremap copying hits a PMD boundary.
775	*/
776	try_realign_addr(pmc, PMD_MASK);
777
778	flush_cache_range(vma: pmc->old, start: pmc->old_addr, end: pmc->old_end);
779	mmu_notifier_range_init(range: &range, event: MMU_NOTIFY_UNMAP, flags: 0, mm,
780	start: pmc->old_addr, end: pmc->old_end);
781	mmu_notifier_invalidate_range_start(range: &range);
782
783	for (; !pmc_done(pmc); pmc_next(pmc, extent)) {
784	cond_resched();
785	/*
786	* If extent is PUD-sized try to speed up the move by moving at the
787	* PUD level if possible.
788	*/
789	extent = get_extent(entry: NORMAL_PUD, pmc);
790
791	old_pud = get_old_pud(mm, addr: pmc->old_addr);
792	if (!old_pud)
793	continue;
794	new_pud = alloc_new_pud(mm, addr: pmc->new_addr);
795	if (!new_pud)
796	break;
797	if (pud_trans_huge(pud: *old_pud) || pud_devmap(pud: *old_pud)) {
798	if (extent == HPAGE_PUD_SIZE) {
799	move_pgt_entry(pmc, entry: HPAGE_PUD, old_entry: old_pud, new_entry: new_pud);
800	/* We ignore and continue on error? */
801	continue;
802	}
803	} else if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) {
804	if (move_pgt_entry(pmc, entry: NORMAL_PUD, old_entry: old_pud, new_entry: new_pud))
805	continue;
806	}
807
808	extent = get_extent(entry: NORMAL_PMD, pmc);
809	old_pmd = get_old_pmd(mm, addr: pmc->old_addr);
810	if (!old_pmd)
811	continue;
812	new_pmd = alloc_new_pmd(mm, addr: pmc->new_addr);
813	if (!new_pmd)
814	break;
815	again:
816	if (is_swap_pmd(pmd: *old_pmd) || pmd_trans_huge(pmd: *old_pmd) ||
817	pmd_devmap(pmd: *old_pmd)) {
818	if (extent == HPAGE_PMD_SIZE &&
819	move_pgt_entry(pmc, entry: HPAGE_PMD, old_entry: old_pmd, new_entry: new_pmd))
820	continue;
821	split_huge_pmd(pmc->old, old_pmd, pmc->old_addr);
822	} else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) &&
823	extent == PMD_SIZE) {
824	/*
825	* If the extent is PMD-sized, try to speed the move by
826	* moving at the PMD level if possible.
827	*/
828	if (move_pgt_entry(pmc, entry: NORMAL_PMD, old_entry: old_pmd, new_entry: new_pmd))
829	continue;
830	}
831	if (pmd_none(pmd: *old_pmd))
832	continue;
833	if (pte_alloc(pmc->new->vm_mm, new_pmd))
834	break;
835	if (move_ptes(pmc, extent, old_pmd, new_pmd) < 0)
836	goto again;
837	}
838
839	mmu_notifier_invalidate_range_end(range: &range);
840
841	return pmc_progress(pmc);
842	}
843
844	/* Set vrm->delta to the difference in VMA size specified by user. */
845	static void vrm_set_delta(struct vma_remap_struct *vrm)
846	{
847	vrm->delta = abs_diff(vrm->old_len, vrm->new_len);
848	}
849
850	/* Determine what kind of remap this is - shrink, expand or no resize at all. */
851	static enum mremap_type vrm_remap_type(struct vma_remap_struct *vrm)
852	{
853	if (vrm->delta == 0)
854	return MREMAP_NO_RESIZE;
855
856	if (vrm->old_len > vrm->new_len)
857	return MREMAP_SHRINK;
858
859	return MREMAP_EXPAND;
860	}
861
862	/*
863	* When moving a VMA to vrm->new_adr, does this result in the new and old VMAs
864	* overlapping?
865	*/
866	static bool vrm_overlaps(struct vma_remap_struct *vrm)
867	{
868	unsigned long start_old = vrm->addr;
869	unsigned long start_new = vrm->new_addr;
870	unsigned long end_old = vrm->addr + vrm->old_len;
871	unsigned long end_new = vrm->new_addr + vrm->new_len;
872
873	/*
874	* start_old end_old
875	* |-----------|
876	* | |
877	* |-----------|
878	* |-------------|
879	* | |
880	* |-------------|
881	* start_new end_new
882	*/
883	if (end_old > start_new && end_new > start_old)
884	return true;
885
886	return false;
887	}
888
889	/* Do the mremap() flags require that the new_addr parameter be specified? */
890	static bool vrm_implies_new_addr(struct vma_remap_struct *vrm)
891	{
892	return vrm->flags & (MREMAP_FIXED | MREMAP_DONTUNMAP);
893	}
894
895	/*
896	* Find an unmapped area for the requested vrm->new_addr.
897	*
898	* If MREMAP_FIXED then this is equivalent to a MAP_FIXED mmap() call. If only
899	* MREMAP_DONTUNMAP is set, then this is equivalent to providing a hint to
900	* mmap(), otherwise this is equivalent to mmap() specifying a NULL address.
901	*
902	* Returns 0 on success (with vrm->new_addr updated), or an error code upon
903	* failure.
904	*/
905	static unsigned long vrm_set_new_addr(struct vma_remap_struct *vrm)
906	{
907	struct vm_area_struct *vma = vrm->vma;
908	unsigned long map_flags = 0;
909	/* Page Offset _into_ the VMA. */
910	pgoff_t internal_pgoff = (vrm->addr - vma->vm_start) >> PAGE_SHIFT;
911	pgoff_t pgoff = vma->vm_pgoff + internal_pgoff;
912	unsigned long new_addr = vrm_implies_new_addr(vrm) ? vrm->new_addr : 0;
913	unsigned long res;
914
915	if (vrm->flags & MREMAP_FIXED)
916	map_flags |= MAP_FIXED;
917	if (vma->vm_flags & VM_MAYSHARE)
918	map_flags |= MAP_SHARED;
919
920	res = get_unmapped_area(file: vma->vm_file, addr: new_addr, len: vrm->new_len, pgoff,
921	flags: map_flags);
922	if (IS_ERR_VALUE(res))
923	return res;
924
925	vrm->new_addr = res;
926	return 0;
927	}
928
929	/*
930	* Keep track of pages which have been added to the memory mapping. If the VMA
931	* is accounted, also check to see if there is sufficient memory.
932	*
933	* Returns true on success, false if insufficient memory to charge.
934	*/
935	static bool vrm_charge(struct vma_remap_struct *vrm)
936	{
937	unsigned long charged;
938
939	if (!(vrm->vma->vm_flags & VM_ACCOUNT))
940	return true;
941
942	/*
943	* If we don't unmap the old mapping, then we account the entirety of
944	* the length of the new one. Otherwise it's just the delta in size.
945	*/
946	if (vrm->flags & MREMAP_DONTUNMAP)
947	charged = vrm->new_len >> PAGE_SHIFT;
948	else
949	charged = vrm->delta >> PAGE_SHIFT;
950
951
952	/* This accounts 'charged' pages of memory. */
953	if (security_vm_enough_memory_mm(current->mm, pages: charged))
954	return false;
955
956	vrm->charged = charged;
957	return true;
958	}
959
960	/*
961	* an error has occurred so we will not be using vrm->charged memory. Unaccount
962	* this memory if the VMA is accounted.
963	*/
964	static void vrm_uncharge(struct vma_remap_struct *vrm)
965	{
966	if (!(vrm->vma->vm_flags & VM_ACCOUNT))
967	return;
968
969	vm_unacct_memory(pages: vrm->charged);
970	vrm->charged = 0;
971	}
972
973	/*
974	* Update mm exec_vm, stack_vm, data_vm, and locked_vm fields as needed to
975	* account for 'bytes' memory used, and if locked, indicate this in the VRM so
976	* we can handle this correctly later.
977	*/
978	static void vrm_stat_account(struct vma_remap_struct *vrm,
979	unsigned long bytes)
980	{
981	unsigned long pages = bytes >> PAGE_SHIFT;
982	struct mm_struct *mm = current->mm;
983	struct vm_area_struct *vma = vrm->vma;
984
985	vm_stat_account(mm, vma->vm_flags, npages: pages);
986	if (vma->vm_flags & VM_LOCKED) {
987	mm->locked_vm += pages;
988	vrm->mlocked = true;
989	}
990	}
991
992	/*
993	* Perform checks before attempting to write a VMA prior to it being
994	* moved.
995	*/
996	static unsigned long prep_move_vma(struct vma_remap_struct *vrm)
997	{
998	unsigned long err = 0;
999	struct vm_area_struct *vma = vrm->vma;
1000	unsigned long old_addr = vrm->addr;
1001	unsigned long old_len = vrm->old_len;
1002	unsigned long dummy = vma->vm_flags;
1003
1004	/*
1005	* We'd prefer to avoid failure later on in do_munmap:
1006	* which may split one vma into three before unmapping.
1007	*/
1008	if (current->mm->map_count >= sysctl_max_map_count - 3)
1009	return -ENOMEM;
1010
1011	if (vma->vm_ops && vma->vm_ops->may_split) {
1012	if (vma->vm_start != old_addr)
1013	err = vma->vm_ops->may_split(vma, old_addr);
1014	if (!err && vma->vm_end != old_addr + old_len)
1015	err = vma->vm_ops->may_split(vma, old_addr + old_len);
1016	if (err)
1017	return err;
1018	}
1019
1020	/*
1021	* Advise KSM to break any KSM pages in the area to be moved:
1022	* it would be confusing if they were to turn up at the new
1023	* location, where they happen to coincide with different KSM
1024	* pages recently unmapped. But leave vma->vm_flags as it was,
1025	* so KSM can come around to merge on vma and new_vma afterwards.
1026	*/
1027	err = ksm_madvise(vma, start: old_addr, end: old_addr + old_len,
1028	MADV_UNMERGEABLE, vm_flags: &dummy);
1029	if (err)
1030	return err;
1031
1032	return 0;
1033	}
1034
1035	/*
1036	* Unmap source VMA for VMA move, turning it from a copy to a move, being
1037	* careful to ensure we do not underflow memory account while doing so if an
1038	* accountable move.
1039	*
1040	* This is best effort, if we fail to unmap then we simply try to correct
1041	* accounting and exit.
1042	*/
1043	static void unmap_source_vma(struct vma_remap_struct *vrm)
1044	{
1045	struct mm_struct *mm = current->mm;
1046	unsigned long addr = vrm->addr;
1047	unsigned long len = vrm->old_len;
1048	struct vm_area_struct *vma = vrm->vma;
1049	VMA_ITERATOR(vmi, mm, addr);
1050	int err;
1051	unsigned long vm_start;
1052	unsigned long vm_end;
1053	/*
1054	* It might seem odd that we check for MREMAP_DONTUNMAP here, given this
1055	* function implies that we unmap the original VMA, which seems
1056	* contradictory.
1057	*
1058	* However, this occurs when this operation was attempted and an error
1059	* arose, in which case we _do_ wish to unmap the _new_ VMA, which means
1060	* we actually _do_ want it be unaccounted.
1061	*/
1062	bool accountable_move = (vma->vm_flags & VM_ACCOUNT) &&
1063	!(vrm->flags & MREMAP_DONTUNMAP);
1064
1065	/*
1066	* So we perform a trick here to prevent incorrect accounting. Any merge
1067	* or new VMA allocation performed in copy_vma() does not adjust
1068	* accounting, it is expected that callers handle this.
1069	*
1070	* And indeed we already have, accounting appropriately in the case of
1071	* both in vrm_charge().
1072	*
1073	* However, when we unmap the existing VMA (to effect the move), this
1074	* code will, if the VMA has VM_ACCOUNT set, attempt to unaccount
1075	* removed pages.
1076	*
1077	* To avoid this we temporarily clear this flag, reinstating on any
1078	* portions of the original VMA that remain.
1079	*/
1080	if (accountable_move) {
1081	vm_flags_clear(vma, VM_ACCOUNT);
1082	/* We are about to split vma, so store the start/end. */
1083	vm_start = vma->vm_start;
1084	vm_end = vma->vm_end;
1085	}
1086
1087	err = do_vmi_munmap(vmi: &vmi, mm, start: addr, len, uf: vrm->uf_unmap, /* unlock= */false);
1088	vrm->vma = NULL; /* Invalidated. */
1089	if (err) {
1090	/* OOM: unable to split vma, just get accounts right */
1091	vm_acct_memory(pages: len >> PAGE_SHIFT);
1092	return;
1093	}
1094
1095	/*
1096	* If we mremap() from a VMA like this:
1097	*
1098	* addr end
1099	* | |
1100	* v v
1101	* |-------------|
1102	* | |
1103	* |-------------|
1104	*
1105	* Having cleared VM_ACCOUNT from the whole VMA, after we unmap above
1106	* we'll end up with:
1107	*
1108	* addr end
1109	* | |
1110	* v v
1111	* |---| |---|
1112	* | A | | B |
1113	* |---| |---|
1114	*
1115	* The VMI is still pointing at addr, so vma_prev() will give us A, and
1116	* a subsequent or lone vma_next() will give as B.
1117	*
1118	* do_vmi_munmap() will have restored the VMI back to addr.
1119	*/
1120	if (accountable_move) {
1121	unsigned long end = addr + len;
1122
1123	if (vm_start < addr) {
1124	struct vm_area_struct *prev = vma_prev(vmi: &vmi);
1125
1126	vm_flags_set(vma: prev, VM_ACCOUNT); /* Acquires VMA lock. */
1127	}
1128
1129	if (vm_end > end) {
1130	struct vm_area_struct *next = vma_next(vmi: &vmi);
1131
1132	vm_flags_set(vma: next, VM_ACCOUNT); /* Acquires VMA lock. */
1133	}
1134	}
1135	}
1136
1137	/*
1138	* Copy vrm->vma over to vrm->new_addr possibly adjusting size as part of the
1139	* process. Additionally handle an error occurring on moving of page tables,
1140	* where we reset vrm state to cause unmapping of the new VMA.
1141	*
1142	* Outputs the newly installed VMA to new_vma_ptr. Returns 0 on success or an
1143	* error code.
1144	*/
1145	static int copy_vma_and_data(struct vma_remap_struct *vrm,
1146	struct vm_area_struct **new_vma_ptr)
1147	{
1148	unsigned long internal_offset = vrm->addr - vrm->vma->vm_start;
1149	unsigned long internal_pgoff = internal_offset >> PAGE_SHIFT;
1150	unsigned long new_pgoff = vrm->vma->vm_pgoff + internal_pgoff;
1151	unsigned long moved_len;
1152	struct vm_area_struct *vma = vrm->vma;
1153	struct vm_area_struct *new_vma;
1154	int err = 0;
1155	PAGETABLE_MOVE(pmc, NULL, NULL, vrm->addr, vrm->new_addr, vrm->old_len);
1156
1157	new_vma = copy_vma(vmap: &vma, addr: vrm->new_addr, len: vrm->new_len, pgoff: new_pgoff,
1158	need_rmap_locks: &pmc.need_rmap_locks);
1159	if (!new_vma) {
1160	vrm_uncharge(vrm);
1161	*new_vma_ptr = NULL;
1162	return -ENOMEM;
1163	}
1164	vrm->vma = vma;
1165	pmc.old = vma;
1166	pmc.new = new_vma;
1167
1168	moved_len = move_page_tables(pmc: &pmc);
1169	if (moved_len < vrm->old_len)
1170	err = -ENOMEM;
1171	else if (vma->vm_ops && vma->vm_ops->mremap)
1172	err = vma->vm_ops->mremap(new_vma);
1173
1174	if (unlikely(err)) {
1175	PAGETABLE_MOVE(pmc_revert, new_vma, vma, vrm->new_addr,
1176	vrm->addr, moved_len);
1177
1178	/*
1179	* On error, move entries back from new area to old,
1180	* which will succeed since page tables still there,
1181	* and then proceed to unmap new area instead of old.
1182	*/
1183	pmc_revert.need_rmap_locks = true;
1184	move_page_tables(pmc: &pmc_revert);
1185
1186	vrm->vma = new_vma;
1187	vrm->old_len = vrm->new_len;
1188	vrm->addr = vrm->new_addr;
1189	} else {
1190	mremap_userfaultfd_prep(new_vma, vrm->uf);
1191	}
1192
1193	fixup_hugetlb_reservations(vma);
1194
1195	*new_vma_ptr = new_vma;
1196	return err;
1197	}
1198
1199	/*
1200	* Perform final tasks for MADV_DONTUNMAP operation, clearing mlock() and
1201	* account flags on remaining VMA by convention (it cannot be mlock()'d any
1202	* longer, as pages in range are no longer mapped), and removing anon_vma_chain
1203	* links from it (if the entire VMA was copied over).
1204	*/
1205	static void dontunmap_complete(struct vma_remap_struct *vrm,
1206	struct vm_area_struct *new_vma)
1207	{
1208	unsigned long start = vrm->addr;
1209	unsigned long end = vrm->addr + vrm->old_len;
1210	unsigned long old_start = vrm->vma->vm_start;
1211	unsigned long old_end = vrm->vma->vm_end;
1212
1213	/*
1214	* We always clear VM_LOCKED[ONFAULT] | VM_ACCOUNT on the old
1215	* vma.
1216	*/
1217	vm_flags_clear(vma: vrm->vma, VM_LOCKED_MASK | VM_ACCOUNT);
1218
1219	/*
1220	* anon_vma links of the old vma is no longer needed after its page
1221	* table has been moved.
1222	*/
1223	if (new_vma != vrm->vma && start == old_start && end == old_end)
1224	unlink_anon_vmas(vrm->vma);
1225
1226	/* Because we won't unmap we don't need to touch locked_vm. */
1227	}
1228
1229	static unsigned long move_vma(struct vma_remap_struct *vrm)
1230	{
1231	struct mm_struct *mm = current->mm;
1232	struct vm_area_struct *new_vma;
1233	unsigned long hiwater_vm;
1234	int err;
1235
1236	err = prep_move_vma(vrm);
1237	if (err)
1238	return err;
1239
1240	/* If accounted, charge the number of bytes the operation will use. */
1241	if (!vrm_charge(vrm))
1242	return -ENOMEM;
1243
1244	/* We don't want racing faults. */
1245	vma_start_write(vma: vrm->vma);
1246
1247	/* Perform copy step. */
1248	err = copy_vma_and_data(vrm, new_vma_ptr: &new_vma);
1249	/*
1250	* If we established the copied-to VMA, we attempt to recover from the
1251	* error by setting the destination VMA to the source VMA and unmapping
1252	* it below.
1253	*/
1254	if (err && !new_vma)
1255	return err;
1256
1257	/*
1258	* If we failed to move page tables we still do total_vm increment
1259	* since do_munmap() will decrement it by old_len == new_len.
1260	*
1261	* Since total_vm is about to be raised artificially high for a
1262	* moment, we need to restore high watermark afterwards: if stats
1263	* are taken meanwhile, total_vm and hiwater_vm appear too high.
1264	* If this were a serious issue, we'd add a flag to do_munmap().
1265	*/
1266	hiwater_vm = mm->hiwater_vm;
1267
1268	vrm_stat_account(vrm, bytes: vrm->new_len);
1269	if (unlikely(!err && (vrm->flags & MREMAP_DONTUNMAP)))
1270	dontunmap_complete(vrm, new_vma);
1271	else
1272	unmap_source_vma(vrm);
1273
1274	mm->hiwater_vm = hiwater_vm;
1275
1276	return err ? (unsigned long)err : vrm->new_addr;
1277	}
1278
1279	/*
1280	* resize_is_valid() - Ensure the vma can be resized to the new length at the give
1281	* address.
1282	*
1283	* Return 0 on success, error otherwise.
1284	*/
1285	static int resize_is_valid(struct vma_remap_struct *vrm)
1286	{
1287	struct mm_struct *mm = current->mm;
1288	struct vm_area_struct *vma = vrm->vma;
1289	unsigned long addr = vrm->addr;
1290	unsigned long old_len = vrm->old_len;
1291	unsigned long new_len = vrm->new_len;
1292	unsigned long pgoff;
1293
1294	/*
1295	* !old_len is a special case where an attempt is made to 'duplicate'
1296	* a mapping. This makes no sense for private mappings as it will
1297	* instead create a fresh/new mapping unrelated to the original. This
1298	* is contrary to the basic idea of mremap which creates new mappings
1299	* based on the original. There are no known use cases for this
1300	* behavior. As a result, fail such attempts.
1301	*/
1302	if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
1303	pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap. This is not supported.\n",
1304	current->comm, current->pid);
1305	return -EINVAL;
1306	}
1307
1308	if ((vrm->flags & MREMAP_DONTUNMAP) &&
1309	(vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)))
1310	return -EINVAL;
1311
1312	/* We can't remap across vm area boundaries */
1313	if (old_len > vma->vm_end - addr)
1314	return -EFAULT;
1315
1316	if (new_len == old_len)
1317	return 0;
1318
1319	/* Need to be careful about a growing mapping */
1320	pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
1321	pgoff += vma->vm_pgoff;
1322	if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
1323	return -EINVAL;
1324
1325	if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
1326	return -EFAULT;
1327
1328	if (!mlock_future_ok(mm, flags: vma->vm_flags, bytes: vrm->delta))
1329	return -EAGAIN;
1330
1331	if (!may_expand_vm(mm, vma->vm_flags, npages: vrm->delta >> PAGE_SHIFT))
1332	return -ENOMEM;
1333
1334	return 0;
1335	}
1336
1337	/*
1338	* The user has requested that the VMA be shrunk (i.e., old_len > new_len), so
1339	* execute this, optionally dropping the mmap lock when we do so.
1340	*
1341	* In both cases this invalidates the VMA, however if we don't drop the lock,
1342	* then load the correct VMA into vrm->vma afterwards.
1343	*/
1344	static unsigned long shrink_vma(struct vma_remap_struct *vrm,
1345	bool drop_lock)
1346	{
1347	struct mm_struct *mm = current->mm;
1348	unsigned long unmap_start = vrm->addr + vrm->new_len;
1349	unsigned long unmap_bytes = vrm->delta;
1350	unsigned long res;
1351	VMA_ITERATOR(vmi, mm, unmap_start);
1352
1353	VM_BUG_ON(vrm->remap_type != MREMAP_SHRINK);
1354
1355	res = do_vmi_munmap(vmi: &vmi, mm, start: unmap_start, len: unmap_bytes,
1356	uf: vrm->uf_unmap, unlock: drop_lock);
1357	vrm->vma = NULL; /* Invalidated. */
1358	if (res)
1359	return res;
1360
1361	/*
1362	* If we've not dropped the lock, then we should reload the VMA to
1363	* replace the invalidated VMA with the one that may have now been
1364	* split.
1365	*/
1366	if (drop_lock) {
1367	vrm->mmap_locked = false;
1368	} else {
1369	vrm->vma = vma_lookup(mm, addr: vrm->addr);
1370	if (!vrm->vma)
1371	return -EFAULT;
1372	}
1373
1374	return 0;
1375	}
1376
1377	/*
1378	* mremap_to() - remap a vma to a new location.
1379	* Returns: The new address of the vma or an error.
1380	*/
1381	static unsigned long mremap_to(struct vma_remap_struct *vrm)
1382	{
1383	struct mm_struct *mm = current->mm;
1384	unsigned long err;
1385
1386	/* Is the new length or address silly? */
1387	if (vrm->new_len > TASK_SIZE ||
1388	vrm->new_addr > TASK_SIZE - vrm->new_len)
1389	return -EINVAL;
1390
1391	if (vrm_overlaps(vrm))
1392	return -EINVAL;
1393
1394	if (vrm->flags & MREMAP_FIXED) {
1395	/*
1396	* In mremap_to().
1397	* VMA is moved to dst address, and munmap dst first.
1398	* do_munmap will check if dst is sealed.
1399	*/
1400	err = do_munmap(mm, vrm->new_addr, vrm->new_len,
1401	uf: vrm->uf_unmap_early);
1402	vrm->vma = NULL; /* Invalidated. */
1403	if (err)
1404	return err;
1405
1406	/*
1407	* If we remap a portion of a VMA elsewhere in the same VMA,
1408	* this can invalidate the old VMA. Reset.
1409	*/
1410	vrm->vma = vma_lookup(mm, addr: vrm->addr);
1411	if (!vrm->vma)
1412	return -EFAULT;
1413	}
1414
1415	if (vrm->remap_type == MREMAP_SHRINK) {
1416	err = shrink_vma(vrm, /* drop_lock= */false);
1417	if (err)
1418	return err;
1419
1420	/* Set up for the move now shrink has been executed. */
1421	vrm->old_len = vrm->new_len;
1422	}
1423
1424	err = resize_is_valid(vrm);
1425	if (err)
1426	return err;
1427
1428	/* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
1429	if (vrm->flags & MREMAP_DONTUNMAP) {
1430	vm_flags_t vm_flags = vrm->vma->vm_flags;
1431	unsigned long pages = vrm->old_len >> PAGE_SHIFT;
1432
1433	if (!may_expand_vm(mm, vm_flags, npages: pages))
1434	return -ENOMEM;
1435	}
1436
1437	err = vrm_set_new_addr(vrm);
1438	if (err)
1439	return err;
1440
1441	return move_vma(vrm);
1442	}
1443
1444	static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
1445	{
1446	unsigned long end = vma->vm_end + delta;
1447
1448	if (end < vma->vm_end) /* overflow */
1449	return 0;
1450	if (find_vma_intersection(mm: vma->vm_mm, start_addr: vma->vm_end, end_addr: end))
1451	return 0;
1452	if (get_unmapped_area(NULL, addr: vma->vm_start, len: end - vma->vm_start,
1453	pgoff: 0, MAP_FIXED) & ~PAGE_MASK)
1454	return 0;
1455	return 1;
1456	}
1457
1458	/* Determine whether we are actually able to execute an in-place expansion. */
1459	static bool vrm_can_expand_in_place(struct vma_remap_struct *vrm)
1460	{
1461	/* Number of bytes from vrm->addr to end of VMA. */
1462	unsigned long suffix_bytes = vrm->vma->vm_end - vrm->addr;
1463
1464	/* If end of range aligns to end of VMA, we can just expand in-place. */
1465	if (suffix_bytes != vrm->old_len)
1466	return false;
1467
1468	/* Check whether this is feasible. */
1469	if (!vma_expandable(vma: vrm->vma, delta: vrm->delta))
1470	return false;
1471
1472	return true;
1473	}
1474
1475	/*
1476	* Are the parameters passed to mremap() valid? If so return 0, otherwise return
1477	* error.
1478	*/
1479	static unsigned long check_mremap_params(struct vma_remap_struct *vrm)
1480
1481	{
1482	unsigned long addr = vrm->addr;
1483	unsigned long flags = vrm->flags;
1484
1485	/* Ensure no unexpected flag values. */
1486	if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
1487	return -EINVAL;
1488
1489	/* Start address must be page-aligned. */
1490	if (offset_in_page(addr))
1491	return -EINVAL;
1492
1493	/*
1494	* We allow a zero old-len as a special case
1495	* for DOS-emu "duplicate shm area" thing. But
1496	* a zero new-len is nonsensical.
1497	*/
1498	if (!PAGE_ALIGN(vrm->new_len))
1499	return -EINVAL;
1500
1501	/* Remainder of checks are for cases with specific new_addr. */
1502	if (!vrm_implies_new_addr(vrm))
1503	return 0;
1504
1505	/* The new address must be page-aligned. */
1506	if (offset_in_page(vrm->new_addr))
1507	return -EINVAL;
1508
1509	/* A fixed address implies a move. */
1510	if (!(flags & MREMAP_MAYMOVE))
1511	return -EINVAL;
1512
1513	/* MREMAP_DONTUNMAP does not allow resizing in the process. */
1514	if (flags & MREMAP_DONTUNMAP && vrm->old_len != vrm->new_len)
1515	return -EINVAL;
1516
1517	/*
1518	* move_vma() need us to stay 4 maps below the threshold, otherwise
1519	* it will bail out at the very beginning.
1520	* That is a problem if we have already unmaped the regions here
1521	* (new_addr, and old_addr), because userspace will not know the
1522	* state of the vma's after it gets -ENOMEM.
1523	* So, to avoid such scenario we can pre-compute if the whole
1524	* operation has high chances to success map-wise.
1525	* Worst-scenario case is when both vma's (new_addr and old_addr) get
1526	* split in 3 before unmapping it.
1527	* That means 2 more maps (1 for each) to the ones we already hold.
1528	* Check whether current map count plus 2 still leads us to 4 maps below
1529	* the threshold, otherwise return -ENOMEM here to be more safe.
1530	*/
1531	if ((current->mm->map_count + 2) >= sysctl_max_map_count - 3)
1532	return -ENOMEM;
1533
1534	return 0;
1535	}
1536
1537	/*
1538	* We know we can expand the VMA in-place by delta pages, so do so.
1539	*
1540	* If we discover the VMA is locked, update mm_struct statistics accordingly and
1541	* indicate so to the caller.
1542	*/
1543	static unsigned long expand_vma_in_place(struct vma_remap_struct *vrm)
1544	{
1545	struct mm_struct *mm = current->mm;
1546	struct vm_area_struct *vma = vrm->vma;
1547	VMA_ITERATOR(vmi, mm, vma->vm_end);
1548
1549	if (!vrm_charge(vrm))
1550	return -ENOMEM;
1551
1552	/*
1553	* Function vma_merge_extend() is called on the
1554	* extension we are adding to the already existing vma,
1555	* vma_merge_extend() will merge this extension with the
1556	* already existing vma (expand operation itself) and
1557	* possibly also with the next vma if it becomes
1558	* adjacent to the expanded vma and otherwise
1559	* compatible.
1560	*/
1561	vma = vma_merge_extend(vmi: &vmi, vma, delta: vrm->delta);
1562	if (!vma) {
1563	vrm_uncharge(vrm);
1564	return -ENOMEM;
1565	}
1566	vrm->vma = vma;
1567
1568	vrm_stat_account(vrm, bytes: vrm->delta);
1569
1570	return 0;
1571	}
1572
1573	static bool align_hugetlb(struct vma_remap_struct *vrm)
1574	{
1575	struct hstate *h __maybe_unused = hstate_vma(vma: vrm->vma);
1576
1577	vrm->old_len = ALIGN(vrm->old_len, huge_page_size(h));
1578	vrm->new_len = ALIGN(vrm->new_len, huge_page_size(h));
1579
1580	/* addrs must be huge page aligned */
1581	if (vrm->addr & ~huge_page_mask(h))
1582	return false;
1583	if (vrm->new_addr & ~huge_page_mask(h))
1584	return false;
1585
1586	/*
1587	* Don't allow remap expansion, because the underlying hugetlb
1588	* reservation is not yet capable to handle split reservation.
1589	*/
1590	if (vrm->new_len > vrm->old_len)
1591	return false;
1592
1593	vrm_set_delta(vrm);
1594
1595	return true;
1596	}
1597
1598	/*
1599	* We are mremap()'ing without specifying a fixed address to move to, but are
1600	* requesting that the VMA's size be increased.
1601	*
1602	* Try to do so in-place, if this fails, then move the VMA to a new location to
1603	* action the change.
1604	*/
1605	static unsigned long expand_vma(struct vma_remap_struct *vrm)
1606	{
1607	unsigned long err;
1608	unsigned long addr = vrm->addr;
1609
1610	err = resize_is_valid(vrm);
1611	if (err)
1612	return err;
1613
1614	/*
1615	* [addr, old_len) spans precisely to the end of the VMA, so try to
1616	* expand it in-place.
1617	*/
1618	if (vrm_can_expand_in_place(vrm)) {
1619	err = expand_vma_in_place(vrm);
1620	if (err)
1621	return err;
1622
1623	/*
1624	* We want to populate the newly expanded portion of the VMA to
1625	* satisfy the expectation that mlock()'ing a VMA maintains all
1626	* of its pages in memory.
1627	*/
1628	if (vrm->mlocked)
1629	vrm->new_addr = addr;
1630
1631	/* OK we're done! */
1632	return addr;
1633	}
1634
1635	/*
1636	* We weren't able to just expand or shrink the area,
1637	* we need to create a new one and move it.
1638	*/
1639
1640	/* We're not allowed to move the VMA, so error out. */
1641	if (!(vrm->flags & MREMAP_MAYMOVE))
1642	return -ENOMEM;
1643
1644	/* Find a new location to move the VMA to. */
1645	err = vrm_set_new_addr(vrm);
1646	if (err)
1647	return err;
1648
1649	return move_vma(vrm);
1650	}
1651
1652	/*
1653	* Attempt to resize the VMA in-place, if we cannot, then move the VMA to the
1654	* first available address to perform the operation.
1655	*/
1656	static unsigned long mremap_at(struct vma_remap_struct *vrm)
1657	{
1658	unsigned long res;
1659
1660	switch (vrm->remap_type) {
1661	case MREMAP_INVALID:
1662	break;
1663	case MREMAP_NO_RESIZE:
1664	/* NO-OP CASE - resizing to the same size. */
1665	return vrm->addr;
1666	case MREMAP_SHRINK:
1667	/*
1668	* SHRINK CASE. Can always be done in-place.
1669	*
1670	* Simply unmap the shrunken portion of the VMA. This does all
1671	* the needed commit accounting, and we indicate that the mmap
1672	* lock should be dropped.
1673	*/
1674	res = shrink_vma(vrm, /* drop_lock= */true);
1675	if (res)
1676	return res;
1677
1678	return vrm->addr;
1679	case MREMAP_EXPAND:
1680	return expand_vma(vrm);
1681	}
1682
1683	BUG();
1684	}
1685
1686	static unsigned long do_mremap(struct vma_remap_struct *vrm)
1687	{
1688	struct mm_struct *mm = current->mm;
1689	struct vm_area_struct *vma;
1690	unsigned long ret;
1691
1692	ret = check_mremap_params(vrm);
1693	if (ret)
1694	return ret;
1695
1696	vrm->old_len = PAGE_ALIGN(vrm->old_len);
1697	vrm->new_len = PAGE_ALIGN(vrm->new_len);
1698	vrm_set_delta(vrm);
1699
1700	if (mmap_write_lock_killable(mm))
1701	return -EINTR;
1702	vrm->mmap_locked = true;
1703
1704	vma = vrm->vma = vma_lookup(mm, addr: vrm->addr);
1705	if (!vma) {
1706	ret = -EFAULT;
1707	goto out;
1708	}
1709
1710	/* If mseal()'d, mremap() is prohibited. */
1711	if (!can_modify_vma(vma)) {
1712	ret = -EPERM;
1713	goto out;
1714	}
1715
1716	/* Align to hugetlb page size, if required. */
1717	if (is_vm_hugetlb_page(vma) && !align_hugetlb(vrm)) {
1718	ret = -EINVAL;
1719	goto out;
1720	}
1721
1722	vrm->remap_type = vrm_remap_type(vrm);
1723
1724	/* Actually execute mremap. */
1725	ret = vrm_implies_new_addr(vrm) ? mremap_to(vrm) : mremap_at(vrm);
1726
1727	out:
1728	if (vrm->mmap_locked) {
1729	mmap_write_unlock(mm);
1730	vrm->mmap_locked = false;
1731
1732	if (!offset_in_page(ret) && vrm->mlocked && vrm->new_len > vrm->old_len)
1733	mm_populate(addr: vrm->new_addr + vrm->old_len, len: vrm->delta);
1734	}
1735
1736	userfaultfd_unmap_complete(mm, uf: vrm->uf_unmap_early);
1737	mremap_userfaultfd_complete(vrm->uf, from: vrm->addr, to: ret, len: vrm->old_len);
1738	userfaultfd_unmap_complete(mm, uf: vrm->uf_unmap);
1739
1740	return ret;
1741	}
1742
1743	/*
1744	* Expand (or shrink) an existing mapping, potentially moving it at the
1745	* same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
1746	*
1747	* MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
1748	* This option implies MREMAP_MAYMOVE.
1749	*/
1750	SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
1751	unsigned long, new_len, unsigned long, flags,
1752	unsigned long, new_addr)
1753	{
1754	struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
1755	LIST_HEAD(uf_unmap_early);
1756	LIST_HEAD(uf_unmap);
1757	/*
1758	* There is a deliberate asymmetry here: we strip the pointer tag
1759	* from the old address but leave the new address alone. This is
1760	* for consistency with mmap(), where we prevent the creation of
1761	* aliasing mappings in userspace by leaving the tag bits of the
1762	* mapping address intact. A non-zero tag will cause the subsequent
1763	* range checks to reject the address as invalid.
1764	*
1765	* See Documentation/arch/arm64/tagged-address-abi.rst for more
1766	* information.
1767	*/
1768	struct vma_remap_struct vrm = {
1769	.addr = untagged_addr(addr),
1770	.old_len = old_len,
1771	.new_len = new_len,
1772	.flags = flags,
1773	.new_addr = new_addr,
1774
1775	.uf = &uf,
1776	.uf_unmap_early = &uf_unmap_early,
1777	.uf_unmap = &uf_unmap,
1778
1779	.remap_type = MREMAP_INVALID, /* We set later. */
1780	};
1781
1782	return do_mremap(vrm: &vrm);
1783	}
1784