• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *	mm/mremap.c
4  *
5  *	(C) Copyright 1996 Linus Torvalds
6  *
7  *	Address space accounting code	<alan@lxorguk.ukuu.org.uk>
8  *	(C) Copyright 2002 Red Hat Inc, All Rights Reserved
9  */
10 
11 #include <linux/mm.h>
12 #include <linux/hugetlb.h>
13 #include <linux/shm.h>
14 #include <linux/ksm.h>
15 #include <linux/mman.h>
16 #include <linux/swap.h>
17 #include <linux/capability.h>
18 #include <linux/fs.h>
19 #include <linux/swapops.h>
20 #include <linux/highmem.h>
21 #include <linux/security.h>
22 #include <linux/syscalls.h>
23 #include <linux/mmu_notifier.h>
24 #include <linux/uaccess.h>
25 #include <linux/mm-arch-hooks.h>
26 #include <linux/userfaultfd_k.h>
27 
28 #include <asm/cacheflush.h>
29 #include <asm/tlbflush.h>
30 
31 #include "internal.h"
32 
get_old_pmd(struct mm_struct * mm,unsigned long addr)33 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
34 {
35 	pgd_t *pgd;
36 	p4d_t *p4d;
37 	pud_t *pud;
38 	pmd_t *pmd;
39 
40 	pgd = pgd_offset(mm, addr);
41 	if (pgd_none_or_clear_bad(pgd))
42 		return NULL;
43 
44 	p4d = p4d_offset(pgd, addr);
45 	if (p4d_none_or_clear_bad(p4d))
46 		return NULL;
47 
48 	pud = pud_offset(p4d, addr);
49 	if (pud_none_or_clear_bad(pud))
50 		return NULL;
51 
52 	pmd = pmd_offset(pud, addr);
53 	if (pmd_none(*pmd))
54 		return NULL;
55 
56 	return pmd;
57 }
58 
alloc_new_pmd(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr)59 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
60 			    unsigned long addr)
61 {
62 	pgd_t *pgd;
63 	p4d_t *p4d;
64 	pud_t *pud;
65 	pmd_t *pmd;
66 
67 	pgd = pgd_offset(mm, addr);
68 	p4d = p4d_alloc(mm, pgd, addr);
69 	if (!p4d)
70 		return NULL;
71 	pud = pud_alloc(mm, p4d, addr);
72 	if (!pud)
73 		return NULL;
74 
75 	pmd = pmd_alloc(mm, pud, addr);
76 	if (!pmd)
77 		return NULL;
78 
79 	VM_BUG_ON(pmd_trans_huge(*pmd));
80 
81 	return pmd;
82 }
83 
take_rmap_locks(struct vm_area_struct * vma)84 static void take_rmap_locks(struct vm_area_struct *vma)
85 {
86 	if (vma->vm_file)
87 		i_mmap_lock_write(vma->vm_file->f_mapping);
88 	if (vma->anon_vma)
89 		anon_vma_lock_write(vma->anon_vma);
90 }
91 
drop_rmap_locks(struct vm_area_struct * vma)92 static void drop_rmap_locks(struct vm_area_struct *vma)
93 {
94 	if (vma->anon_vma)
95 		anon_vma_unlock_write(vma->anon_vma);
96 	if (vma->vm_file)
97 		i_mmap_unlock_write(vma->vm_file->f_mapping);
98 }
99 
move_soft_dirty_pte(pte_t pte)100 static pte_t move_soft_dirty_pte(pte_t pte)
101 {
102 	/*
103 	 * Set soft dirty bit so we can notice
104 	 * in userspace the ptes were moved.
105 	 */
106 #ifdef CONFIG_MEM_SOFT_DIRTY
107 	if (pte_present(pte))
108 		pte = pte_mksoft_dirty(pte);
109 	else if (is_swap_pte(pte))
110 		pte = pte_swp_mksoft_dirty(pte);
111 #endif
112 	return pte;
113 }
114 
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)115 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
116 		unsigned long old_addr, unsigned long old_end,
117 		struct vm_area_struct *new_vma, pmd_t *new_pmd,
118 		unsigned long new_addr, bool need_rmap_locks)
119 {
120 	struct mm_struct *mm = vma->vm_mm;
121 	pte_t *old_pte, *new_pte, pte;
122 	spinlock_t *old_ptl, *new_ptl;
123 	bool force_flush = false;
124 	unsigned long len = old_end - old_addr;
125 
126 	/*
127 	 * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
128 	 * locks to ensure that rmap will always observe either the old or the
129 	 * new ptes. This is the easiest way to avoid races with
130 	 * truncate_pagecache(), page migration, etc...
131 	 *
132 	 * When need_rmap_locks is false, we use other ways to avoid
133 	 * such races:
134 	 *
135 	 * - During exec() shift_arg_pages(), we use a specially tagged vma
136 	 *   which rmap call sites look for using vma_is_temporary_stack().
137 	 *
138 	 * - During mremap(), new_vma is often known to be placed after vma
139 	 *   in rmap traversal order. This ensures rmap will always observe
140 	 *   either the old pte, or the new pte, or both (the page table locks
141 	 *   serialize access to individual ptes, but only rmap traversal
142 	 *   order guarantees that we won't miss both the old and new ptes).
143 	 */
144 	if (need_rmap_locks)
145 		take_rmap_locks(vma);
146 
147 	/*
148 	 * We don't have to worry about the ordering of src and dst
149 	 * pte locks because exclusive mmap_lock prevents deadlock.
150 	 */
151 	old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
152 	new_pte = pte_offset_map(new_pmd, new_addr);
153 	new_ptl = pte_lockptr(mm, new_pmd);
154 	if (new_ptl != old_ptl)
155 		spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
156 	flush_tlb_batched_pending(vma->vm_mm);
157 	arch_enter_lazy_mmu_mode();
158 
159 	for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
160 				   new_pte++, new_addr += PAGE_SIZE) {
161 		if (pte_none(*old_pte))
162 			continue;
163 
164 		pte = ptep_get_and_clear(mm, old_addr, old_pte);
165 		/*
166 		 * If we are remapping a valid PTE, make sure
167 		 * to flush TLB before we drop the PTL for the
168 		 * PTE.
169 		 *
170 		 * NOTE! Both old and new PTL matter: the old one
171 		 * for racing with page_mkclean(), the new one to
172 		 * make sure the physical page stays valid until
173 		 * the TLB entry for the old mapping has been
174 		 * flushed.
175 		 */
176 		if (pte_present(pte))
177 			force_flush = true;
178 		pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
179 		pte = move_soft_dirty_pte(pte);
180 		set_pte_at(mm, new_addr, new_pte, pte);
181 	}
182 
183 	arch_leave_lazy_mmu_mode();
184 	if (force_flush)
185 		flush_tlb_range(vma, old_end - len, old_end);
186 	if (new_ptl != old_ptl)
187 		spin_unlock(new_ptl);
188 	pte_unmap(new_pte - 1);
189 	pte_unmap_unlock(old_pte - 1, old_ptl);
190 	if (need_rmap_locks)
191 		drop_rmap_locks(vma);
192 }
193 
194 #ifdef CONFIG_HAVE_MOVE_PMD
move_normal_pmd(struct vm_area_struct * vma,unsigned long old_addr,unsigned long new_addr,pmd_t * old_pmd,pmd_t * new_pmd)195 static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr,
196 		  unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
197 {
198 	spinlock_t *old_ptl, *new_ptl;
199 	struct mm_struct *mm = vma->vm_mm;
200 	pmd_t pmd;
201 
202 	/*
203 	 * The destination pmd shouldn't be established, free_pgtables()
204 	 * should have released it.
205 	 *
206 	 * However, there's a case during execve() where we use mremap
207 	 * to move the initial stack, and in that case the target area
208 	 * may overlap the source area (always moving down).
209 	 *
210 	 * If everything is PMD-aligned, that works fine, as moving
211 	 * each pmd down will clear the source pmd. But if we first
212 	 * have a few 4kB-only pages that get moved down, and then
213 	 * hit the "now the rest is PMD-aligned, let's do everything
214 	 * one pmd at a time", we will still have the old (now empty
215 	 * of any 4kB pages, but still there) PMD in the page table
216 	 * tree.
217 	 *
218 	 * Warn on it once - because we really should try to figure
219 	 * out how to do this better - but then say "I won't move
220 	 * this pmd".
221 	 *
222 	 * One alternative might be to just unmap the target pmd at
223 	 * this point, and verify that it really is empty. We'll see.
224 	 */
225 	if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
226 		return false;
227 
228 	/*
229 	 * We don't have to worry about the ordering of src and dst
230 	 * ptlocks because exclusive mmap_lock prevents deadlock.
231 	 */
232 	old_ptl = pmd_lock(vma->vm_mm, old_pmd);
233 	new_ptl = pmd_lockptr(mm, new_pmd);
234 	if (new_ptl != old_ptl)
235 		spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
236 
237 	/* Clear the pmd */
238 	pmd = *old_pmd;
239 	pmd_clear(old_pmd);
240 
241 	VM_BUG_ON(!pmd_none(*new_pmd));
242 
243 	/* Set the new pmd */
244 	set_pmd_at(mm, new_addr, new_pmd, pmd);
245 	flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
246 	if (new_ptl != old_ptl)
247 		spin_unlock(new_ptl);
248 	spin_unlock(old_ptl);
249 
250 	return true;
251 }
252 #endif
253 
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)254 unsigned long move_page_tables(struct vm_area_struct *vma,
255 		unsigned long old_addr, struct vm_area_struct *new_vma,
256 		unsigned long new_addr, unsigned long len,
257 		bool need_rmap_locks)
258 {
259 	unsigned long extent, next, old_end;
260 	struct mmu_notifier_range range;
261 	pmd_t *old_pmd, *new_pmd;
262 
263 	if (!len)
264 		return 0;
265 
266 	old_end = old_addr + len;
267 	flush_cache_range(vma, old_addr, old_end);
268 
269 	mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
270 				old_addr, old_end);
271 	mmu_notifier_invalidate_range_start(&range);
272 
273 	for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
274 		cond_resched();
275 		next = (old_addr + PMD_SIZE) & PMD_MASK;
276 		/* even if next overflowed, extent below will be ok */
277 		extent = next - old_addr;
278 		if (extent > old_end - old_addr)
279 			extent = old_end - old_addr;
280 		next = (new_addr + PMD_SIZE) & PMD_MASK;
281 		if (extent > next - new_addr)
282 			extent = next - new_addr;
283 		old_pmd = get_old_pmd(vma->vm_mm, old_addr);
284 		if (!old_pmd)
285 			continue;
286 		new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
287 		if (!new_pmd)
288 			break;
289 		if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) || pmd_devmap(*old_pmd)) {
290 			if (extent == HPAGE_PMD_SIZE) {
291 				bool moved;
292 				/* See comment in move_ptes() */
293 				if (need_rmap_locks)
294 					take_rmap_locks(vma);
295 				moved = move_huge_pmd(vma, old_addr, new_addr,
296 						      old_pmd, new_pmd);
297 				if (need_rmap_locks)
298 					drop_rmap_locks(vma);
299 				if (moved)
300 					continue;
301 			}
302 			split_huge_pmd(vma, old_pmd, old_addr);
303 			if (pmd_trans_unstable(old_pmd))
304 				continue;
305 		} else if (extent == PMD_SIZE) {
306 #ifdef CONFIG_HAVE_MOVE_PMD
307 			/*
308 			 * If the extent is PMD-sized, try to speed the move by
309 			 * moving at the PMD level if possible.
310 			 */
311 			bool moved;
312 
313 			take_rmap_locks(vma);
314 			moved = move_normal_pmd(vma, old_addr, new_addr,
315 						old_pmd, new_pmd);
316 			drop_rmap_locks(vma);
317 			if (moved)
318 				continue;
319 #endif
320 		}
321 
322 		if (pte_alloc(new_vma->vm_mm, new_pmd))
323 			break;
324 		move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
325 			  new_pmd, new_addr, need_rmap_locks);
326 	}
327 
328 	mmu_notifier_invalidate_range_end(&range);
329 
330 	return len + old_addr - old_end;	/* how much done */
331 }
332 
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,unsigned long flags,struct vm_userfaultfd_ctx * uf,struct list_head * uf_unmap)333 static unsigned long move_vma(struct vm_area_struct *vma,
334 		unsigned long old_addr, unsigned long old_len,
335 		unsigned long new_len, unsigned long new_addr,
336 		bool *locked, unsigned long flags,
337 		struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
338 {
339 	struct mm_struct *mm = vma->vm_mm;
340 	struct vm_area_struct *new_vma;
341 	unsigned long vm_flags = vma->vm_flags;
342 	unsigned long new_pgoff;
343 	unsigned long moved_len;
344 	unsigned long excess = 0;
345 	unsigned long hiwater_vm;
346 	int split = 0;
347 	int err;
348 	bool need_rmap_locks;
349 
350 	/*
351 	 * We'd prefer to avoid failure later on in do_munmap:
352 	 * which may split one vma into three before unmapping.
353 	 */
354 	if (mm->map_count >= sysctl_max_map_count - 3)
355 		return -ENOMEM;
356 
357 	/*
358 	 * Advise KSM to break any KSM pages in the area to be moved:
359 	 * it would be confusing if they were to turn up at the new
360 	 * location, where they happen to coincide with different KSM
361 	 * pages recently unmapped.  But leave vma->vm_flags as it was,
362 	 * so KSM can come around to merge on vma and new_vma afterwards.
363 	 */
364 	err = ksm_madvise(vma, old_addr, old_addr + old_len,
365 						MADV_UNMERGEABLE, &vm_flags);
366 	if (err)
367 		return err;
368 
369 	new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
370 	new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
371 			   &need_rmap_locks);
372 	if (!new_vma)
373 		return -ENOMEM;
374 
375 	moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
376 				     need_rmap_locks);
377 	if (moved_len < old_len) {
378 		err = -ENOMEM;
379 	} else if (vma->vm_ops && vma->vm_ops->mremap) {
380 		err = vma->vm_ops->mremap(new_vma);
381 	}
382 
383 	if (unlikely(err)) {
384 		/*
385 		 * On error, move entries back from new area to old,
386 		 * which will succeed since page tables still there,
387 		 * and then proceed to unmap new area instead of old.
388 		 */
389 		move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
390 				 true);
391 		vma = new_vma;
392 		old_len = new_len;
393 		old_addr = new_addr;
394 		new_addr = err;
395 	} else {
396 		mremap_userfaultfd_prep(new_vma, uf);
397 		arch_remap(mm, old_addr, old_addr + old_len,
398 			   new_addr, new_addr + new_len);
399 	}
400 
401 	/* Conceal VM_ACCOUNT so old reservation is not undone */
402 	if (vm_flags & VM_ACCOUNT) {
403 		vma->vm_flags &= ~VM_ACCOUNT;
404 		excess = vma->vm_end - vma->vm_start - old_len;
405 		if (old_addr > vma->vm_start &&
406 		    old_addr + old_len < vma->vm_end)
407 			split = 1;
408 	}
409 
410 	/*
411 	 * If we failed to move page tables we still do total_vm increment
412 	 * since do_munmap() will decrement it by old_len == new_len.
413 	 *
414 	 * Since total_vm is about to be raised artificially high for a
415 	 * moment, we need to restore high watermark afterwards: if stats
416 	 * are taken meanwhile, total_vm and hiwater_vm appear too high.
417 	 * If this were a serious issue, we'd add a flag to do_munmap().
418 	 */
419 	hiwater_vm = mm->hiwater_vm;
420 	vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
421 
422 	/* Tell pfnmap has moved from this vma */
423 	if (unlikely(vma->vm_flags & VM_PFNMAP))
424 		untrack_pfn_moved(vma);
425 
426 	if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
427 		if (vm_flags & VM_ACCOUNT) {
428 			/* Always put back VM_ACCOUNT since we won't unmap */
429 			vma->vm_flags |= VM_ACCOUNT;
430 
431 			vm_acct_memory(new_len >> PAGE_SHIFT);
432 		}
433 
434 		/*
435 		 * VMAs can actually be merged back together in copy_vma
436 		 * calling merge_vma. This can happen with anonymous vmas
437 		 * which have not yet been faulted, so if we were to consider
438 		 * this VMA split we'll end up adding VM_ACCOUNT on the
439 		 * next VMA, which is completely unrelated if this VMA
440 		 * was re-merged.
441 		 */
442 		if (split && new_vma == vma)
443 			split = 0;
444 
445 		/* We always clear VM_LOCKED[ONFAULT] on the old vma */
446 		vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
447 
448 		/* Because we won't unmap we don't need to touch locked_vm */
449 		goto out;
450 	}
451 
452 	if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
453 		/* OOM: unable to split vma, just get accounts right */
454 		vm_unacct_memory(excess >> PAGE_SHIFT);
455 		excess = 0;
456 	}
457 
458 	if (vm_flags & VM_LOCKED) {
459 		mm->locked_vm += new_len >> PAGE_SHIFT;
460 		*locked = true;
461 	}
462 out:
463 	mm->hiwater_vm = hiwater_vm;
464 
465 	/* Restore VM_ACCOUNT if one or two pieces of vma left */
466 	if (excess) {
467 		vma->vm_flags |= VM_ACCOUNT;
468 		if (split)
469 			vma->vm_next->vm_flags |= VM_ACCOUNT;
470 	}
471 
472 	return new_addr;
473 }
474 
vma_to_resize(unsigned long addr,unsigned long old_len,unsigned long new_len,unsigned long flags,unsigned long * p)475 static struct vm_area_struct *vma_to_resize(unsigned long addr,
476 	unsigned long old_len, unsigned long new_len, unsigned long flags,
477 	unsigned long *p)
478 {
479 	struct mm_struct *mm = current->mm;
480 	struct vm_area_struct *vma = find_vma(mm, addr);
481 	unsigned long pgoff;
482 
483 	if (!vma || vma->vm_start > addr)
484 		return ERR_PTR(-EFAULT);
485 
486 	/*
487 	 * !old_len is a special case where an attempt is made to 'duplicate'
488 	 * a mapping.  This makes no sense for private mappings as it will
489 	 * instead create a fresh/new mapping unrelated to the original.  This
490 	 * is contrary to the basic idea of mremap which creates new mappings
491 	 * based on the original.  There are no known use cases for this
492 	 * behavior.  As a result, fail such attempts.
493 	 */
494 	if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
495 		pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap.  This is not supported.\n", current->comm, current->pid);
496 		return ERR_PTR(-EINVAL);
497 	}
498 
499 	if (flags & MREMAP_DONTUNMAP && (!vma_is_anonymous(vma) ||
500 			vma->vm_flags & VM_SHARED))
501 		return ERR_PTR(-EINVAL);
502 
503 	if (is_vm_hugetlb_page(vma))
504 		return ERR_PTR(-EINVAL);
505 
506 	/* We can't remap across vm area boundaries */
507 	if (old_len > vma->vm_end - addr)
508 		return ERR_PTR(-EFAULT);
509 
510 	if (new_len == old_len)
511 		return vma;
512 
513 	/* Need to be careful about a growing mapping */
514 	pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
515 	pgoff += vma->vm_pgoff;
516 	if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
517 		return ERR_PTR(-EINVAL);
518 
519 	if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
520 		return ERR_PTR(-EFAULT);
521 
522 	if (vma->vm_flags & VM_LOCKED) {
523 		unsigned long locked, lock_limit;
524 		locked = mm->locked_vm << PAGE_SHIFT;
525 		lock_limit = rlimit(RLIMIT_MEMLOCK);
526 		locked += new_len - old_len;
527 		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
528 			return ERR_PTR(-EAGAIN);
529 	}
530 
531 	if (!may_expand_vm(mm, vma->vm_flags,
532 				(new_len - old_len) >> PAGE_SHIFT))
533 		return ERR_PTR(-ENOMEM);
534 
535 	if (vma->vm_flags & VM_ACCOUNT) {
536 		unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
537 		if (security_vm_enough_memory_mm(mm, charged))
538 			return ERR_PTR(-ENOMEM);
539 		*p = charged;
540 	}
541 
542 	return vma;
543 }
544 
mremap_to(unsigned long addr,unsigned long old_len,unsigned long new_addr,unsigned long new_len,bool * locked,unsigned long flags,struct vm_userfaultfd_ctx * uf,struct list_head * uf_unmap_early,struct list_head * uf_unmap)545 static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
546 		unsigned long new_addr, unsigned long new_len, bool *locked,
547 		unsigned long flags, struct vm_userfaultfd_ctx *uf,
548 		struct list_head *uf_unmap_early,
549 		struct list_head *uf_unmap)
550 {
551 	struct mm_struct *mm = current->mm;
552 	struct vm_area_struct *vma;
553 	unsigned long ret = -EINVAL;
554 	unsigned long charged = 0;
555 	unsigned long map_flags = 0;
556 
557 	if (offset_in_page(new_addr))
558 		goto out;
559 
560 	if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
561 		goto out;
562 
563 	/* Ensure the old/new locations do not overlap */
564 	if (addr + old_len > new_addr && new_addr + new_len > addr)
565 		goto out;
566 
567 	/*
568 	 * move_vma() need us to stay 4 maps below the threshold, otherwise
569 	 * it will bail out at the very beginning.
570 	 * That is a problem if we have already unmaped the regions here
571 	 * (new_addr, and old_addr), because userspace will not know the
572 	 * state of the vma's after it gets -ENOMEM.
573 	 * So, to avoid such scenario we can pre-compute if the whole
574 	 * operation has high chances to success map-wise.
575 	 * Worst-scenario case is when both vma's (new_addr and old_addr) get
576 	 * split in 3 before unmaping it.
577 	 * That means 2 more maps (1 for each) to the ones we already hold.
578 	 * Check whether current map count plus 2 still leads us to 4 maps below
579 	 * the threshold, otherwise return -ENOMEM here to be more safe.
580 	 */
581 	if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
582 		return -ENOMEM;
583 
584 	if (flags & MREMAP_FIXED) {
585 		ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
586 		if (ret)
587 			goto out;
588 	}
589 
590 	if (old_len >= new_len) {
591 		ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
592 		if (ret && old_len != new_len)
593 			goto out;
594 		old_len = new_len;
595 	}
596 
597 	vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
598 	if (IS_ERR(vma)) {
599 		ret = PTR_ERR(vma);
600 		goto out;
601 	}
602 
603 	/* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
604 	if (flags & MREMAP_DONTUNMAP &&
605 		!may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
606 		ret = -ENOMEM;
607 		goto out;
608 	}
609 
610 	if (flags & MREMAP_FIXED)
611 		map_flags |= MAP_FIXED;
612 
613 	if (vma->vm_flags & VM_MAYSHARE)
614 		map_flags |= MAP_SHARED;
615 
616 	ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
617 				((addr - vma->vm_start) >> PAGE_SHIFT),
618 				map_flags);
619 	if (IS_ERR_VALUE(ret))
620 		goto out1;
621 
622 	/* We got a new mapping */
623 	if (!(flags & MREMAP_FIXED))
624 		new_addr = ret;
625 
626 	ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
627 		       uf_unmap);
628 
629 	if (!(offset_in_page(ret)))
630 		goto out;
631 
632 out1:
633 	vm_unacct_memory(charged);
634 
635 out:
636 	return ret;
637 }
638 
vma_expandable(struct vm_area_struct * vma,unsigned long delta)639 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
640 {
641 	unsigned long end = vma->vm_end + delta;
642 	if (end < vma->vm_end) /* overflow */
643 		return 0;
644 	if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
645 		return 0;
646 	if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
647 			      0, MAP_FIXED) & ~PAGE_MASK)
648 		return 0;
649 	return 1;
650 }
651 
652 /*
653  * Expand (or shrink) an existing mapping, potentially moving it at the
654  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
655  *
656  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
657  * This option implies MREMAP_MAYMOVE.
658  */
SYSCALL_DEFINE5(mremap,unsigned long,addr,unsigned long,old_len,unsigned long,new_len,unsigned long,flags,unsigned long,new_addr)659 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
660 		unsigned long, new_len, unsigned long, flags,
661 		unsigned long, new_addr)
662 {
663 	struct mm_struct *mm = current->mm;
664 	struct vm_area_struct *vma;
665 	unsigned long ret = -EINVAL;
666 	unsigned long charged = 0;
667 	bool locked = false;
668 	bool downgraded = false;
669 	struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
670 	LIST_HEAD(uf_unmap_early);
671 	LIST_HEAD(uf_unmap);
672 
673 	/*
674 	 * There is a deliberate asymmetry here: we strip the pointer tag
675 	 * from the old address but leave the new address alone. This is
676 	 * for consistency with mmap(), where we prevent the creation of
677 	 * aliasing mappings in userspace by leaving the tag bits of the
678 	 * mapping address intact. A non-zero tag will cause the subsequent
679 	 * range checks to reject the address as invalid.
680 	 *
681 	 * See Documentation/arm64/tagged-address-abi.rst for more information.
682 	 */
683 	addr = untagged_addr(addr);
684 
685 	if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
686 		return ret;
687 
688 	if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
689 		return ret;
690 
691 	/*
692 	 * MREMAP_DONTUNMAP is always a move and it does not allow resizing
693 	 * in the process.
694 	 */
695 	if (flags & MREMAP_DONTUNMAP &&
696 			(!(flags & MREMAP_MAYMOVE) || old_len != new_len))
697 		return ret;
698 
699 
700 	if (offset_in_page(addr))
701 		return ret;
702 
703 	old_len = PAGE_ALIGN(old_len);
704 	new_len = PAGE_ALIGN(new_len);
705 
706 	/*
707 	 * We allow a zero old-len as a special case
708 	 * for DOS-emu "duplicate shm area" thing. But
709 	 * a zero new-len is nonsensical.
710 	 */
711 	if (!new_len)
712 		return ret;
713 
714 	if (mmap_write_lock_killable(current->mm))
715 		return -EINTR;
716 
717 	if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
718 		ret = mremap_to(addr, old_len, new_addr, new_len,
719 				&locked, flags, &uf, &uf_unmap_early,
720 				&uf_unmap);
721 		goto out;
722 	}
723 
724 	/*
725 	 * Always allow a shrinking remap: that just unmaps
726 	 * the unnecessary pages..
727 	 * __do_munmap does all the needed commit accounting, and
728 	 * downgrades mmap_lock to read if so directed.
729 	 */
730 	if (old_len >= new_len) {
731 		int retval;
732 
733 		retval = __do_munmap(mm, addr+new_len, old_len - new_len,
734 				  &uf_unmap, true);
735 		if (retval < 0 && old_len != new_len) {
736 			ret = retval;
737 			goto out;
738 		/* Returning 1 indicates mmap_lock is downgraded to read. */
739 		} else if (retval == 1)
740 			downgraded = true;
741 		ret = addr;
742 		goto out;
743 	}
744 
745 	/*
746 	 * Ok, we need to grow..
747 	 */
748 	vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
749 	if (IS_ERR(vma)) {
750 		ret = PTR_ERR(vma);
751 		goto out;
752 	}
753 
754 	/* old_len exactly to the end of the area..
755 	 */
756 	if (old_len == vma->vm_end - addr) {
757 		/* can we just expand the current mapping? */
758 		if (vma_expandable(vma, new_len - old_len)) {
759 			int pages = (new_len - old_len) >> PAGE_SHIFT;
760 
761 			if (vma_adjust(vma, vma->vm_start, addr + new_len,
762 				       vma->vm_pgoff, NULL)) {
763 				ret = -ENOMEM;
764 				goto out;
765 			}
766 
767 			vm_stat_account(mm, vma->vm_flags, pages);
768 			if (vma->vm_flags & VM_LOCKED) {
769 				mm->locked_vm += pages;
770 				locked = true;
771 				new_addr = addr;
772 			}
773 			ret = addr;
774 			goto out;
775 		}
776 	}
777 
778 	/*
779 	 * We weren't able to just expand or shrink the area,
780 	 * we need to create a new one and move it..
781 	 */
782 	ret = -ENOMEM;
783 	if (flags & MREMAP_MAYMOVE) {
784 		unsigned long map_flags = 0;
785 		if (vma->vm_flags & VM_MAYSHARE)
786 			map_flags |= MAP_SHARED;
787 
788 		new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
789 					vma->vm_pgoff +
790 					((addr - vma->vm_start) >> PAGE_SHIFT),
791 					map_flags);
792 		if (IS_ERR_VALUE(new_addr)) {
793 			ret = new_addr;
794 			goto out;
795 		}
796 
797 		ret = move_vma(vma, addr, old_len, new_len, new_addr,
798 			       &locked, flags, &uf, &uf_unmap);
799 	}
800 out:
801 	if (offset_in_page(ret)) {
802 		vm_unacct_memory(charged);
803 		locked = false;
804 	}
805 	if (downgraded)
806 		mmap_read_unlock(current->mm);
807 	else
808 		mmap_write_unlock(current->mm);
809 	if (locked && new_len > old_len)
810 		mm_populate(new_addr + old_len, new_len - old_len);
811 	userfaultfd_unmap_complete(mm, &uf_unmap_early);
812 	mremap_userfaultfd_complete(&uf, addr, ret, old_len);
813 	userfaultfd_unmap_complete(mm, &uf_unmap);
814 	return ret;
815 }
816