xref: /linux/mm/vma.c (revision 7e7b2370ed0551d2fa8e4bd6e4bbd603fef3cdcd)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 
3 /*
4  * VMA-specific functions.
5  */
6 
7 #include "vma_internal.h"
8 #include "vma.h"
9 
10 /*
11  * If the vma has a ->close operation then the driver probably needs to release
12  * per-vma resources, so we don't attempt to merge those if the caller indicates
13  * the current vma may be removed as part of the merge.
14  */
15 static inline bool is_mergeable_vma(struct vm_area_struct *vma,
16 		struct file *file, unsigned long vm_flags,
17 		struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
18 		struct anon_vma_name *anon_name, bool may_remove_vma)
19 {
20 	/*
21 	 * VM_SOFTDIRTY should not prevent from VMA merging, if we
22 	 * match the flags but dirty bit -- the caller should mark
23 	 * merged VMA as dirty. If dirty bit won't be excluded from
24 	 * comparison, we increase pressure on the memory system forcing
25 	 * the kernel to generate new VMAs when old one could be
26 	 * extended instead.
27 	 */
28 	if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
29 		return false;
30 	if (vma->vm_file != file)
31 		return false;
32 	if (may_remove_vma && vma->vm_ops && vma->vm_ops->close)
33 		return false;
34 	if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
35 		return false;
36 	if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
37 		return false;
38 	return true;
39 }
40 
41 static inline bool is_mergeable_anon_vma(struct anon_vma *anon_vma1,
42 		 struct anon_vma *anon_vma2, struct vm_area_struct *vma)
43 {
44 	/*
45 	 * The list_is_singular() test is to avoid merging VMA cloned from
46 	 * parents. This can improve scalability caused by anon_vma lock.
47 	 */
48 	if ((!anon_vma1 || !anon_vma2) && (!vma ||
49 		list_is_singular(&vma->anon_vma_chain)))
50 		return true;
51 	return anon_vma1 == anon_vma2;
52 }
53 
54 /*
55  * init_multi_vma_prep() - Initializer for struct vma_prepare
56  * @vp: The vma_prepare struct
57  * @vma: The vma that will be altered once locked
58  * @next: The next vma if it is to be adjusted
59  * @remove: The first vma to be removed
60  * @remove2: The second vma to be removed
61  */
62 static void init_multi_vma_prep(struct vma_prepare *vp,
63 				struct vm_area_struct *vma,
64 				struct vm_area_struct *next,
65 				struct vm_area_struct *remove,
66 				struct vm_area_struct *remove2)
67 {
68 	memset(vp, 0, sizeof(struct vma_prepare));
69 	vp->vma = vma;
70 	vp->anon_vma = vma->anon_vma;
71 	vp->remove = remove;
72 	vp->remove2 = remove2;
73 	vp->adj_next = next;
74 	if (!vp->anon_vma && next)
75 		vp->anon_vma = next->anon_vma;
76 
77 	vp->file = vma->vm_file;
78 	if (vp->file)
79 		vp->mapping = vma->vm_file->f_mapping;
80 
81 }
82 
83 /*
84  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
85  * in front of (at a lower virtual address and file offset than) the vma.
86  *
87  * We cannot merge two vmas if they have differently assigned (non-NULL)
88  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
89  *
90  * We don't check here for the merged mmap wrapping around the end of pagecache
91  * indices (16TB on ia32) because do_mmap() does not permit mmap's which
92  * wrap, nor mmaps which cover the final page at index -1UL.
93  *
94  * We assume the vma may be removed as part of the merge.
95  */
96 bool
97 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
98 		struct anon_vma *anon_vma, struct file *file,
99 		pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
100 		struct anon_vma_name *anon_name)
101 {
102 	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, true) &&
103 	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
104 		if (vma->vm_pgoff == vm_pgoff)
105 			return true;
106 	}
107 	return false;
108 }
109 
110 /*
111  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
112  * beyond (at a higher virtual address and file offset than) the vma.
113  *
114  * We cannot merge two vmas if they have differently assigned (non-NULL)
115  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
116  *
117  * We assume that vma is not removed as part of the merge.
118  */
119 bool
120 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
121 		struct anon_vma *anon_vma, struct file *file,
122 		pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
123 		struct anon_vma_name *anon_name)
124 {
125 	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, false) &&
126 	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
127 		pgoff_t vm_pglen;
128 
129 		vm_pglen = vma_pages(vma);
130 		if (vma->vm_pgoff + vm_pglen == vm_pgoff)
131 			return true;
132 	}
133 	return false;
134 }
135 
136 /*
137  * Close a vm structure and free it.
138  */
139 void remove_vma(struct vm_area_struct *vma, bool unreachable)
140 {
141 	might_sleep();
142 	if (vma->vm_ops && vma->vm_ops->close)
143 		vma->vm_ops->close(vma);
144 	if (vma->vm_file)
145 		fput(vma->vm_file);
146 	mpol_put(vma_policy(vma));
147 	if (unreachable)
148 		__vm_area_free(vma);
149 	else
150 		vm_area_free(vma);
151 }
152 
153 /*
154  * Get rid of page table information in the indicated region.
155  *
156  * Called with the mm semaphore held.
157  */
158 void unmap_region(struct mm_struct *mm, struct ma_state *mas,
159 		struct vm_area_struct *vma, struct vm_area_struct *prev,
160 		struct vm_area_struct *next, unsigned long start,
161 		unsigned long end, unsigned long tree_end, bool mm_wr_locked)
162 {
163 	struct mmu_gather tlb;
164 	unsigned long mt_start = mas->index;
165 
166 	lru_add_drain();
167 	tlb_gather_mmu(&tlb, mm);
168 	update_hiwater_rss(mm);
169 	unmap_vmas(&tlb, mas, vma, start, end, tree_end, mm_wr_locked);
170 	mas_set(mas, mt_start);
171 	free_pgtables(&tlb, mas, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
172 				 next ? next->vm_start : USER_PGTABLES_CEILING,
173 				 mm_wr_locked);
174 	tlb_finish_mmu(&tlb);
175 }
176 
177 /*
178  * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
179  * has already been checked or doesn't make sense to fail.
180  * VMA Iterator will point to the original VMA.
181  */
182 static int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
183 		       unsigned long addr, int new_below)
184 {
185 	struct vma_prepare vp;
186 	struct vm_area_struct *new;
187 	int err;
188 
189 	WARN_ON(vma->vm_start >= addr);
190 	WARN_ON(vma->vm_end <= addr);
191 
192 	if (vma->vm_ops && vma->vm_ops->may_split) {
193 		err = vma->vm_ops->may_split(vma, addr);
194 		if (err)
195 			return err;
196 	}
197 
198 	new = vm_area_dup(vma);
199 	if (!new)
200 		return -ENOMEM;
201 
202 	if (new_below) {
203 		new->vm_end = addr;
204 	} else {
205 		new->vm_start = addr;
206 		new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
207 	}
208 
209 	err = -ENOMEM;
210 	vma_iter_config(vmi, new->vm_start, new->vm_end);
211 	if (vma_iter_prealloc(vmi, new))
212 		goto out_free_vma;
213 
214 	err = vma_dup_policy(vma, new);
215 	if (err)
216 		goto out_free_vmi;
217 
218 	err = anon_vma_clone(new, vma);
219 	if (err)
220 		goto out_free_mpol;
221 
222 	if (new->vm_file)
223 		get_file(new->vm_file);
224 
225 	if (new->vm_ops && new->vm_ops->open)
226 		new->vm_ops->open(new);
227 
228 	vma_start_write(vma);
229 	vma_start_write(new);
230 
231 	init_vma_prep(&vp, vma);
232 	vp.insert = new;
233 	vma_prepare(&vp);
234 	vma_adjust_trans_huge(vma, vma->vm_start, addr, 0);
235 
236 	if (new_below) {
237 		vma->vm_start = addr;
238 		vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT;
239 	} else {
240 		vma->vm_end = addr;
241 	}
242 
243 	/* vma_complete stores the new vma */
244 	vma_complete(&vp, vmi, vma->vm_mm);
245 
246 	/* Success. */
247 	if (new_below)
248 		vma_next(vmi);
249 	else
250 		vma_prev(vmi);
251 
252 	return 0;
253 
254 out_free_mpol:
255 	mpol_put(vma_policy(new));
256 out_free_vmi:
257 	vma_iter_free(vmi);
258 out_free_vma:
259 	vm_area_free(new);
260 	return err;
261 }
262 
263 /*
264  * Split a vma into two pieces at address 'addr', a new vma is allocated
265  * either for the first part or the tail.
266  */
267 static int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
268 		     unsigned long addr, int new_below)
269 {
270 	if (vma->vm_mm->map_count >= sysctl_max_map_count)
271 		return -ENOMEM;
272 
273 	return __split_vma(vmi, vma, addr, new_below);
274 }
275 
276 /*
277  * Ok - we have the memory areas we should free on a maple tree so release them,
278  * and do the vma updates.
279  *
280  * Called with the mm semaphore held.
281  */
282 static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
283 {
284 	unsigned long nr_accounted = 0;
285 	struct vm_area_struct *vma;
286 
287 	/* Update high watermark before we lower total_vm */
288 	update_hiwater_vm(mm);
289 	mas_for_each(mas, vma, ULONG_MAX) {
290 		long nrpages = vma_pages(vma);
291 
292 		if (vma->vm_flags & VM_ACCOUNT)
293 			nr_accounted += nrpages;
294 		vm_stat_account(mm, vma->vm_flags, -nrpages);
295 		remove_vma(vma, false);
296 	}
297 	vm_unacct_memory(nr_accounted);
298 }
299 
300 /*
301  * init_vma_prep() - Initializer wrapper for vma_prepare struct
302  * @vp: The vma_prepare struct
303  * @vma: The vma that will be altered once locked
304  */
305 void init_vma_prep(struct vma_prepare *vp,
306 		   struct vm_area_struct *vma)
307 {
308 	init_multi_vma_prep(vp, vma, NULL, NULL, NULL);
309 }
310 
311 /*
312  * Requires inode->i_mapping->i_mmap_rwsem
313  */
314 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
315 				      struct address_space *mapping)
316 {
317 	if (vma_is_shared_maywrite(vma))
318 		mapping_unmap_writable(mapping);
319 
320 	flush_dcache_mmap_lock(mapping);
321 	vma_interval_tree_remove(vma, &mapping->i_mmap);
322 	flush_dcache_mmap_unlock(mapping);
323 }
324 
325 /*
326  * vma has some anon_vma assigned, and is already inserted on that
327  * anon_vma's interval trees.
328  *
329  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
330  * vma must be removed from the anon_vma's interval trees using
331  * anon_vma_interval_tree_pre_update_vma().
332  *
333  * After the update, the vma will be reinserted using
334  * anon_vma_interval_tree_post_update_vma().
335  *
336  * The entire update must be protected by exclusive mmap_lock and by
337  * the root anon_vma's mutex.
338  */
339 void
340 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
341 {
342 	struct anon_vma_chain *avc;
343 
344 	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
345 		anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
346 }
347 
348 void
349 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
350 {
351 	struct anon_vma_chain *avc;
352 
353 	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
354 		anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
355 }
356 
357 static void __vma_link_file(struct vm_area_struct *vma,
358 			    struct address_space *mapping)
359 {
360 	if (vma_is_shared_maywrite(vma))
361 		mapping_allow_writable(mapping);
362 
363 	flush_dcache_mmap_lock(mapping);
364 	vma_interval_tree_insert(vma, &mapping->i_mmap);
365 	flush_dcache_mmap_unlock(mapping);
366 }
367 
368 /*
369  * vma_prepare() - Helper function for handling locking VMAs prior to altering
370  * @vp: The initialized vma_prepare struct
371  */
372 void vma_prepare(struct vma_prepare *vp)
373 {
374 	if (vp->file) {
375 		uprobe_munmap(vp->vma, vp->vma->vm_start, vp->vma->vm_end);
376 
377 		if (vp->adj_next)
378 			uprobe_munmap(vp->adj_next, vp->adj_next->vm_start,
379 				      vp->adj_next->vm_end);
380 
381 		i_mmap_lock_write(vp->mapping);
382 		if (vp->insert && vp->insert->vm_file) {
383 			/*
384 			 * Put into interval tree now, so instantiated pages
385 			 * are visible to arm/parisc __flush_dcache_page
386 			 * throughout; but we cannot insert into address
387 			 * space until vma start or end is updated.
388 			 */
389 			__vma_link_file(vp->insert,
390 					vp->insert->vm_file->f_mapping);
391 		}
392 	}
393 
394 	if (vp->anon_vma) {
395 		anon_vma_lock_write(vp->anon_vma);
396 		anon_vma_interval_tree_pre_update_vma(vp->vma);
397 		if (vp->adj_next)
398 			anon_vma_interval_tree_pre_update_vma(vp->adj_next);
399 	}
400 
401 	if (vp->file) {
402 		flush_dcache_mmap_lock(vp->mapping);
403 		vma_interval_tree_remove(vp->vma, &vp->mapping->i_mmap);
404 		if (vp->adj_next)
405 			vma_interval_tree_remove(vp->adj_next,
406 						 &vp->mapping->i_mmap);
407 	}
408 
409 }
410 
411 /*
412  * dup_anon_vma() - Helper function to duplicate anon_vma
413  * @dst: The destination VMA
414  * @src: The source VMA
415  * @dup: Pointer to the destination VMA when successful.
416  *
417  * Returns: 0 on success.
418  */
419 static int dup_anon_vma(struct vm_area_struct *dst,
420 			struct vm_area_struct *src, struct vm_area_struct **dup)
421 {
422 	/*
423 	 * Easily overlooked: when mprotect shifts the boundary, make sure the
424 	 * expanding vma has anon_vma set if the shrinking vma had, to cover any
425 	 * anon pages imported.
426 	 */
427 	if (src->anon_vma && !dst->anon_vma) {
428 		int ret;
429 
430 		vma_assert_write_locked(dst);
431 		dst->anon_vma = src->anon_vma;
432 		ret = anon_vma_clone(dst, src);
433 		if (ret)
434 			return ret;
435 
436 		*dup = dst;
437 	}
438 
439 	return 0;
440 }
441 
442 #ifdef CONFIG_DEBUG_VM_MAPLE_TREE
443 void validate_mm(struct mm_struct *mm)
444 {
445 	int bug = 0;
446 	int i = 0;
447 	struct vm_area_struct *vma;
448 	VMA_ITERATOR(vmi, mm, 0);
449 
450 	mt_validate(&mm->mm_mt);
451 	for_each_vma(vmi, vma) {
452 #ifdef CONFIG_DEBUG_VM_RB
453 		struct anon_vma *anon_vma = vma->anon_vma;
454 		struct anon_vma_chain *avc;
455 #endif
456 		unsigned long vmi_start, vmi_end;
457 		bool warn = 0;
458 
459 		vmi_start = vma_iter_addr(&vmi);
460 		vmi_end = vma_iter_end(&vmi);
461 		if (VM_WARN_ON_ONCE_MM(vma->vm_end != vmi_end, mm))
462 			warn = 1;
463 
464 		if (VM_WARN_ON_ONCE_MM(vma->vm_start != vmi_start, mm))
465 			warn = 1;
466 
467 		if (warn) {
468 			pr_emerg("issue in %s\n", current->comm);
469 			dump_stack();
470 			dump_vma(vma);
471 			pr_emerg("tree range: %px start %lx end %lx\n", vma,
472 				 vmi_start, vmi_end - 1);
473 			vma_iter_dump_tree(&vmi);
474 		}
475 
476 #ifdef CONFIG_DEBUG_VM_RB
477 		if (anon_vma) {
478 			anon_vma_lock_read(anon_vma);
479 			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
480 				anon_vma_interval_tree_verify(avc);
481 			anon_vma_unlock_read(anon_vma);
482 		}
483 #endif
484 		i++;
485 	}
486 	if (i != mm->map_count) {
487 		pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);
488 		bug = 1;
489 	}
490 	VM_BUG_ON_MM(bug, mm);
491 }
492 #endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
493 
494 /*
495  * vma_expand - Expand an existing VMA
496  *
497  * @vmi: The vma iterator
498  * @vma: The vma to expand
499  * @start: The start of the vma
500  * @end: The exclusive end of the vma
501  * @pgoff: The page offset of vma
502  * @next: The current of next vma.
503  *
504  * Expand @vma to @start and @end.  Can expand off the start and end.  Will
505  * expand over @next if it's different from @vma and @end == @next->vm_end.
506  * Checking if the @vma can expand and merge with @next needs to be handled by
507  * the caller.
508  *
509  * Returns: 0 on success
510  */
511 int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma,
512 	       unsigned long start, unsigned long end, pgoff_t pgoff,
513 	       struct vm_area_struct *next)
514 {
515 	struct vm_area_struct *anon_dup = NULL;
516 	bool remove_next = false;
517 	struct vma_prepare vp;
518 
519 	vma_start_write(vma);
520 	if (next && (vma != next) && (end == next->vm_end)) {
521 		int ret;
522 
523 		remove_next = true;
524 		vma_start_write(next);
525 		ret = dup_anon_vma(vma, next, &anon_dup);
526 		if (ret)
527 			return ret;
528 	}
529 
530 	init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL);
531 	/* Not merging but overwriting any part of next is not handled. */
532 	VM_WARN_ON(next && !vp.remove &&
533 		  next != vma && end > next->vm_start);
534 	/* Only handles expanding */
535 	VM_WARN_ON(vma->vm_start < start || vma->vm_end > end);
536 
537 	/* Note: vma iterator must be pointing to 'start' */
538 	vma_iter_config(vmi, start, end);
539 	if (vma_iter_prealloc(vmi, vma))
540 		goto nomem;
541 
542 	vma_prepare(&vp);
543 	vma_adjust_trans_huge(vma, start, end, 0);
544 	vma_set_range(vma, start, end, pgoff);
545 	vma_iter_store(vmi, vma);
546 
547 	vma_complete(&vp, vmi, vma->vm_mm);
548 	return 0;
549 
550 nomem:
551 	if (anon_dup)
552 		unlink_anon_vmas(anon_dup);
553 	return -ENOMEM;
554 }
555 
556 /*
557  * vma_shrink() - Reduce an existing VMAs memory area
558  * @vmi: The vma iterator
559  * @vma: The VMA to modify
560  * @start: The new start
561  * @end: The new end
562  *
563  * Returns: 0 on success, -ENOMEM otherwise
564  */
565 int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
566 	       unsigned long start, unsigned long end, pgoff_t pgoff)
567 {
568 	struct vma_prepare vp;
569 
570 	WARN_ON((vma->vm_start != start) && (vma->vm_end != end));
571 
572 	if (vma->vm_start < start)
573 		vma_iter_config(vmi, vma->vm_start, start);
574 	else
575 		vma_iter_config(vmi, end, vma->vm_end);
576 
577 	if (vma_iter_prealloc(vmi, NULL))
578 		return -ENOMEM;
579 
580 	vma_start_write(vma);
581 
582 	init_vma_prep(&vp, vma);
583 	vma_prepare(&vp);
584 	vma_adjust_trans_huge(vma, start, end, 0);
585 
586 	vma_iter_clear(vmi);
587 	vma_set_range(vma, start, end, pgoff);
588 	vma_complete(&vp, vmi, vma->vm_mm);
589 	return 0;
590 }
591 
592 /*
593  * vma_complete- Helper function for handling the unlocking after altering VMAs,
594  * or for inserting a VMA.
595  *
596  * @vp: The vma_prepare struct
597  * @vmi: The vma iterator
598  * @mm: The mm_struct
599  */
600 void vma_complete(struct vma_prepare *vp,
601 		  struct vma_iterator *vmi, struct mm_struct *mm)
602 {
603 	if (vp->file) {
604 		if (vp->adj_next)
605 			vma_interval_tree_insert(vp->adj_next,
606 						 &vp->mapping->i_mmap);
607 		vma_interval_tree_insert(vp->vma, &vp->mapping->i_mmap);
608 		flush_dcache_mmap_unlock(vp->mapping);
609 	}
610 
611 	if (vp->remove && vp->file) {
612 		__remove_shared_vm_struct(vp->remove, vp->mapping);
613 		if (vp->remove2)
614 			__remove_shared_vm_struct(vp->remove2, vp->mapping);
615 	} else if (vp->insert) {
616 		/*
617 		 * split_vma has split insert from vma, and needs
618 		 * us to insert it before dropping the locks
619 		 * (it may either follow vma or precede it).
620 		 */
621 		vma_iter_store(vmi, vp->insert);
622 		mm->map_count++;
623 	}
624 
625 	if (vp->anon_vma) {
626 		anon_vma_interval_tree_post_update_vma(vp->vma);
627 		if (vp->adj_next)
628 			anon_vma_interval_tree_post_update_vma(vp->adj_next);
629 		anon_vma_unlock_write(vp->anon_vma);
630 	}
631 
632 	if (vp->file) {
633 		i_mmap_unlock_write(vp->mapping);
634 		uprobe_mmap(vp->vma);
635 
636 		if (vp->adj_next)
637 			uprobe_mmap(vp->adj_next);
638 	}
639 
640 	if (vp->remove) {
641 again:
642 		vma_mark_detached(vp->remove, true);
643 		if (vp->file) {
644 			uprobe_munmap(vp->remove, vp->remove->vm_start,
645 				      vp->remove->vm_end);
646 			fput(vp->file);
647 		}
648 		if (vp->remove->anon_vma)
649 			anon_vma_merge(vp->vma, vp->remove);
650 		mm->map_count--;
651 		mpol_put(vma_policy(vp->remove));
652 		if (!vp->remove2)
653 			WARN_ON_ONCE(vp->vma->vm_end < vp->remove->vm_end);
654 		vm_area_free(vp->remove);
655 
656 		/*
657 		 * In mprotect's case 6 (see comments on vma_merge),
658 		 * we are removing both mid and next vmas
659 		 */
660 		if (vp->remove2) {
661 			vp->remove = vp->remove2;
662 			vp->remove2 = NULL;
663 			goto again;
664 		}
665 	}
666 	if (vp->insert && vp->file)
667 		uprobe_mmap(vp->insert);
668 	validate_mm(mm);
669 }
670 
671 /*
672  * abort_munmap_vmas - Undo any munmap work and free resources
673  *
674  * Reattach any detached vmas and free up the maple tree used to track the vmas.
675  */
676 static inline void abort_munmap_vmas(struct ma_state *mas_detach)
677 {
678 	struct vm_area_struct *vma;
679 
680 	mas_set(mas_detach, 0);
681 	mas_for_each(mas_detach, vma, ULONG_MAX)
682 		vma_mark_detached(vma, false);
683 
684 	__mt_destroy(mas_detach->tree);
685 }
686 
687 /*
688  * do_vmi_align_munmap() - munmap the aligned region from @start to @end.
689  * @vmi: The vma iterator
690  * @vma: The starting vm_area_struct
691  * @mm: The mm_struct
692  * @start: The aligned start address to munmap.
693  * @end: The aligned end address to munmap.
694  * @uf: The userfaultfd list_head
695  * @unlock: Set to true to drop the mmap_lock.  unlocking only happens on
696  * success.
697  *
698  * Return: 0 on success and drops the lock if so directed, error and leaves the
699  * lock held otherwise.
700  */
701 int
702 do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
703 		    struct mm_struct *mm, unsigned long start,
704 		    unsigned long end, struct list_head *uf, bool unlock)
705 {
706 	struct vm_area_struct *prev, *next = NULL;
707 	struct maple_tree mt_detach;
708 	int count = 0;
709 	int error = -ENOMEM;
710 	unsigned long locked_vm = 0;
711 	MA_STATE(mas_detach, &mt_detach, 0, 0);
712 	mt_init_flags(&mt_detach, vmi->mas.tree->ma_flags & MT_FLAGS_LOCK_MASK);
713 	mt_on_stack(mt_detach);
714 
715 	/*
716 	 * If we need to split any vma, do it now to save pain later.
717 	 *
718 	 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
719 	 * unmapped vm_area_struct will remain in use: so lower split_vma
720 	 * places tmp vma above, and higher split_vma places tmp vma below.
721 	 */
722 
723 	/* Does it split the first one? */
724 	if (start > vma->vm_start) {
725 
726 		/*
727 		 * Make sure that map_count on return from munmap() will
728 		 * not exceed its limit; but let map_count go just above
729 		 * its limit temporarily, to help free resources as expected.
730 		 */
731 		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
732 			goto map_count_exceeded;
733 
734 		/* Don't bother splitting the VMA if we can't unmap it anyway */
735 		if (!can_modify_vma(vma)) {
736 			error = -EPERM;
737 			goto start_split_failed;
738 		}
739 
740 		error = __split_vma(vmi, vma, start, 1);
741 		if (error)
742 			goto start_split_failed;
743 	}
744 
745 	/*
746 	 * Detach a range of VMAs from the mm. Using next as a temp variable as
747 	 * it is always overwritten.
748 	 */
749 	next = vma;
750 	do {
751 		if (!can_modify_vma(next)) {
752 			error = -EPERM;
753 			goto modify_vma_failed;
754 		}
755 
756 		/* Does it split the end? */
757 		if (next->vm_end > end) {
758 			error = __split_vma(vmi, next, end, 0);
759 			if (error)
760 				goto end_split_failed;
761 		}
762 		vma_start_write(next);
763 		mas_set(&mas_detach, count);
764 		error = mas_store_gfp(&mas_detach, next, GFP_KERNEL);
765 		if (error)
766 			goto munmap_gather_failed;
767 		vma_mark_detached(next, true);
768 		if (next->vm_flags & VM_LOCKED)
769 			locked_vm += vma_pages(next);
770 
771 		count++;
772 		if (unlikely(uf)) {
773 			/*
774 			 * If userfaultfd_unmap_prep returns an error the vmas
775 			 * will remain split, but userland will get a
776 			 * highly unexpected error anyway. This is no
777 			 * different than the case where the first of the two
778 			 * __split_vma fails, but we don't undo the first
779 			 * split, despite we could. This is unlikely enough
780 			 * failure that it's not worth optimizing it for.
781 			 */
782 			error = userfaultfd_unmap_prep(next, start, end, uf);
783 
784 			if (error)
785 				goto userfaultfd_error;
786 		}
787 #ifdef CONFIG_DEBUG_VM_MAPLE_TREE
788 		BUG_ON(next->vm_start < start);
789 		BUG_ON(next->vm_start > end);
790 #endif
791 	} for_each_vma_range(*vmi, next, end);
792 
793 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
794 	/* Make sure no VMAs are about to be lost. */
795 	{
796 		MA_STATE(test, &mt_detach, 0, 0);
797 		struct vm_area_struct *vma_mas, *vma_test;
798 		int test_count = 0;
799 
800 		vma_iter_set(vmi, start);
801 		rcu_read_lock();
802 		vma_test = mas_find(&test, count - 1);
803 		for_each_vma_range(*vmi, vma_mas, end) {
804 			BUG_ON(vma_mas != vma_test);
805 			test_count++;
806 			vma_test = mas_next(&test, count - 1);
807 		}
808 		rcu_read_unlock();
809 		BUG_ON(count != test_count);
810 	}
811 #endif
812 
813 	while (vma_iter_addr(vmi) > start)
814 		vma_iter_prev_range(vmi);
815 
816 	error = vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL);
817 	if (error)
818 		goto clear_tree_failed;
819 
820 	/* Point of no return */
821 	mm->locked_vm -= locked_vm;
822 	mm->map_count -= count;
823 	if (unlock)
824 		mmap_write_downgrade(mm);
825 
826 	prev = vma_iter_prev_range(vmi);
827 	next = vma_next(vmi);
828 	if (next)
829 		vma_iter_prev_range(vmi);
830 
831 	/*
832 	 * We can free page tables without write-locking mmap_lock because VMAs
833 	 * were isolated before we downgraded mmap_lock.
834 	 */
835 	mas_set(&mas_detach, 1);
836 	unmap_region(mm, &mas_detach, vma, prev, next, start, end, count,
837 		     !unlock);
838 	/* Statistics and freeing VMAs */
839 	mas_set(&mas_detach, 0);
840 	remove_mt(mm, &mas_detach);
841 	validate_mm(mm);
842 	if (unlock)
843 		mmap_read_unlock(mm);
844 
845 	__mt_destroy(&mt_detach);
846 	return 0;
847 
848 modify_vma_failed:
849 clear_tree_failed:
850 userfaultfd_error:
851 munmap_gather_failed:
852 end_split_failed:
853 	abort_munmap_vmas(&mas_detach);
854 start_split_failed:
855 map_count_exceeded:
856 	validate_mm(mm);
857 	return error;
858 }
859 
860 /*
861  * do_vmi_munmap() - munmap a given range.
862  * @vmi: The vma iterator
863  * @mm: The mm_struct
864  * @start: The start address to munmap
865  * @len: The length of the range to munmap
866  * @uf: The userfaultfd list_head
867  * @unlock: set to true if the user wants to drop the mmap_lock on success
868  *
869  * This function takes a @mas that is either pointing to the previous VMA or set
870  * to MA_START and sets it up to remove the mapping(s).  The @len will be
871  * aligned.
872  *
873  * Return: 0 on success and drops the lock if so directed, error and leaves the
874  * lock held otherwise.
875  */
876 int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm,
877 		  unsigned long start, size_t len, struct list_head *uf,
878 		  bool unlock)
879 {
880 	unsigned long end;
881 	struct vm_area_struct *vma;
882 
883 	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
884 		return -EINVAL;
885 
886 	end = start + PAGE_ALIGN(len);
887 	if (end == start)
888 		return -EINVAL;
889 
890 	/* Find the first overlapping VMA */
891 	vma = vma_find(vmi, end);
892 	if (!vma) {
893 		if (unlock)
894 			mmap_write_unlock(mm);
895 		return 0;
896 	}
897 
898 	return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
899 }
900 
901 /*
902  * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
903  * figure out whether that can be merged with its predecessor or its
904  * successor.  Or both (it neatly fills a hole).
905  *
906  * In most cases - when called for mmap, brk or mremap - [addr,end) is
907  * certain not to be mapped by the time vma_merge is called; but when
908  * called for mprotect, it is certain to be already mapped (either at
909  * an offset within prev, or at the start of next), and the flags of
910  * this area are about to be changed to vm_flags - and the no-change
911  * case has already been eliminated.
912  *
913  * The following mprotect cases have to be considered, where **** is
914  * the area passed down from mprotect_fixup, never extending beyond one
915  * vma, PPPP is the previous vma, CCCC is a concurrent vma that starts
916  * at the same address as **** and is of the same or larger span, and
917  * NNNN the next vma after ****:
918  *
919  *     ****             ****                   ****
920  *    PPPPPPNNNNNN    PPPPPPNNNNNN       PPPPPPCCCCCC
921  *    cannot merge    might become       might become
922  *                    PPNNNNNNNNNN       PPPPPPPPPPCC
923  *    mmap, brk or    case 4 below       case 5 below
924  *    mremap move:
925  *                        ****               ****
926  *                    PPPP    NNNN       PPPPCCCCNNNN
927  *                    might become       might become
928  *                    PPPPPPPPPPPP 1 or  PPPPPPPPPPPP 6 or
929  *                    PPPPPPPPNNNN 2 or  PPPPPPPPNNNN 7 or
930  *                    PPPPNNNNNNNN 3     PPPPNNNNNNNN 8
931  *
932  * It is important for case 8 that the vma CCCC overlapping the
933  * region **** is never going to extended over NNNN. Instead NNNN must
934  * be extended in region **** and CCCC must be removed. This way in
935  * all cases where vma_merge succeeds, the moment vma_merge drops the
936  * rmap_locks, the properties of the merged vma will be already
937  * correct for the whole merged range. Some of those properties like
938  * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
939  * be correct for the whole merged range immediately after the
940  * rmap_locks are released. Otherwise if NNNN would be removed and
941  * CCCC would be extended over the NNNN range, remove_migration_ptes
942  * or other rmap walkers (if working on addresses beyond the "end"
943  * parameter) may establish ptes with the wrong permissions of CCCC
944  * instead of the right permissions of NNNN.
945  *
946  * In the code below:
947  * PPPP is represented by *prev
948  * CCCC is represented by *curr or not represented at all (NULL)
949  * NNNN is represented by *next or not represented at all (NULL)
950  * **** is not represented - it will be merged and the vma containing the
951  *      area is returned, or the function will return NULL
952  */
953 static struct vm_area_struct
954 *vma_merge(struct vma_iterator *vmi, struct vm_area_struct *prev,
955 	   struct vm_area_struct *src, unsigned long addr, unsigned long end,
956 	   unsigned long vm_flags, pgoff_t pgoff, struct mempolicy *policy,
957 	   struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
958 	   struct anon_vma_name *anon_name)
959 {
960 	struct mm_struct *mm = src->vm_mm;
961 	struct anon_vma *anon_vma = src->anon_vma;
962 	struct file *file = src->vm_file;
963 	struct vm_area_struct *curr, *next, *res;
964 	struct vm_area_struct *vma, *adjust, *remove, *remove2;
965 	struct vm_area_struct *anon_dup = NULL;
966 	struct vma_prepare vp;
967 	pgoff_t vma_pgoff;
968 	int err = 0;
969 	bool merge_prev = false;
970 	bool merge_next = false;
971 	bool vma_expanded = false;
972 	unsigned long vma_start = addr;
973 	unsigned long vma_end = end;
974 	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
975 	long adj_start = 0;
976 
977 	/*
978 	 * We later require that vma->vm_flags == vm_flags,
979 	 * so this tests vma->vm_flags & VM_SPECIAL, too.
980 	 */
981 	if (vm_flags & VM_SPECIAL)
982 		return NULL;
983 
984 	/* Does the input range span an existing VMA? (cases 5 - 8) */
985 	curr = find_vma_intersection(mm, prev ? prev->vm_end : 0, end);
986 
987 	if (!curr ||			/* cases 1 - 4 */
988 	    end == curr->vm_end)	/* cases 6 - 8, adjacent VMA */
989 		next = vma_lookup(mm, end);
990 	else
991 		next = NULL;		/* case 5 */
992 
993 	if (prev) {
994 		vma_start = prev->vm_start;
995 		vma_pgoff = prev->vm_pgoff;
996 
997 		/* Can we merge the predecessor? */
998 		if (addr == prev->vm_end && mpol_equal(vma_policy(prev), policy)
999 		    && can_vma_merge_after(prev, vm_flags, anon_vma, file,
1000 					   pgoff, vm_userfaultfd_ctx, anon_name)) {
1001 			merge_prev = true;
1002 			vma_prev(vmi);
1003 		}
1004 	}
1005 
1006 	/* Can we merge the successor? */
1007 	if (next && mpol_equal(policy, vma_policy(next)) &&
1008 	    can_vma_merge_before(next, vm_flags, anon_vma, file, pgoff+pglen,
1009 				 vm_userfaultfd_ctx, anon_name)) {
1010 		merge_next = true;
1011 	}
1012 
1013 	/* Verify some invariant that must be enforced by the caller. */
1014 	VM_WARN_ON(prev && addr <= prev->vm_start);
1015 	VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end));
1016 	VM_WARN_ON(addr >= end);
1017 
1018 	if (!merge_prev && !merge_next)
1019 		return NULL; /* Not mergeable. */
1020 
1021 	if (merge_prev)
1022 		vma_start_write(prev);
1023 
1024 	res = vma = prev;
1025 	remove = remove2 = adjust = NULL;
1026 
1027 	/* Can we merge both the predecessor and the successor? */
1028 	if (merge_prev && merge_next &&
1029 	    is_mergeable_anon_vma(prev->anon_vma, next->anon_vma, NULL)) {
1030 		vma_start_write(next);
1031 		remove = next;				/* case 1 */
1032 		vma_end = next->vm_end;
1033 		err = dup_anon_vma(prev, next, &anon_dup);
1034 		if (curr) {				/* case 6 */
1035 			vma_start_write(curr);
1036 			remove = curr;
1037 			remove2 = next;
1038 			/*
1039 			 * Note that the dup_anon_vma below cannot overwrite err
1040 			 * since the first caller would do nothing unless next
1041 			 * has an anon_vma.
1042 			 */
1043 			if (!next->anon_vma)
1044 				err = dup_anon_vma(prev, curr, &anon_dup);
1045 		}
1046 	} else if (merge_prev) {			/* case 2 */
1047 		if (curr) {
1048 			vma_start_write(curr);
1049 			if (end == curr->vm_end) {	/* case 7 */
1050 				/*
1051 				 * can_vma_merge_after() assumed we would not be
1052 				 * removing prev vma, so it skipped the check
1053 				 * for vm_ops->close, but we are removing curr
1054 				 */
1055 				if (curr->vm_ops && curr->vm_ops->close)
1056 					err = -EINVAL;
1057 				remove = curr;
1058 			} else {			/* case 5 */
1059 				adjust = curr;
1060 				adj_start = (end - curr->vm_start);
1061 			}
1062 			if (!err)
1063 				err = dup_anon_vma(prev, curr, &anon_dup);
1064 		}
1065 	} else { /* merge_next */
1066 		vma_start_write(next);
1067 		res = next;
1068 		if (prev && addr < prev->vm_end) {	/* case 4 */
1069 			vma_start_write(prev);
1070 			vma_end = addr;
1071 			adjust = next;
1072 			adj_start = -(prev->vm_end - addr);
1073 			err = dup_anon_vma(next, prev, &anon_dup);
1074 		} else {
1075 			/*
1076 			 * Note that cases 3 and 8 are the ONLY ones where prev
1077 			 * is permitted to be (but is not necessarily) NULL.
1078 			 */
1079 			vma = next;			/* case 3 */
1080 			vma_start = addr;
1081 			vma_end = next->vm_end;
1082 			vma_pgoff = next->vm_pgoff - pglen;
1083 			if (curr) {			/* case 8 */
1084 				vma_pgoff = curr->vm_pgoff;
1085 				vma_start_write(curr);
1086 				remove = curr;
1087 				err = dup_anon_vma(next, curr, &anon_dup);
1088 			}
1089 		}
1090 	}
1091 
1092 	/* Error in anon_vma clone. */
1093 	if (err)
1094 		goto anon_vma_fail;
1095 
1096 	if (vma_start < vma->vm_start || vma_end > vma->vm_end)
1097 		vma_expanded = true;
1098 
1099 	if (vma_expanded) {
1100 		vma_iter_config(vmi, vma_start, vma_end);
1101 	} else {
1102 		vma_iter_config(vmi, adjust->vm_start + adj_start,
1103 				adjust->vm_end);
1104 	}
1105 
1106 	if (vma_iter_prealloc(vmi, vma))
1107 		goto prealloc_fail;
1108 
1109 	init_multi_vma_prep(&vp, vma, adjust, remove, remove2);
1110 	VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
1111 		   vp.anon_vma != adjust->anon_vma);
1112 
1113 	vma_prepare(&vp);
1114 	vma_adjust_trans_huge(vma, vma_start, vma_end, adj_start);
1115 	vma_set_range(vma, vma_start, vma_end, vma_pgoff);
1116 
1117 	if (vma_expanded)
1118 		vma_iter_store(vmi, vma);
1119 
1120 	if (adj_start) {
1121 		adjust->vm_start += adj_start;
1122 		adjust->vm_pgoff += adj_start >> PAGE_SHIFT;
1123 		if (adj_start < 0) {
1124 			WARN_ON(vma_expanded);
1125 			vma_iter_store(vmi, next);
1126 		}
1127 	}
1128 
1129 	vma_complete(&vp, vmi, mm);
1130 	khugepaged_enter_vma(res, vm_flags);
1131 	return res;
1132 
1133 prealloc_fail:
1134 	if (anon_dup)
1135 		unlink_anon_vmas(anon_dup);
1136 
1137 anon_vma_fail:
1138 	vma_iter_set(vmi, addr);
1139 	vma_iter_load(vmi);
1140 	return NULL;
1141 }
1142 
1143 /*
1144  * We are about to modify one or multiple of a VMA's flags, policy, userfaultfd
1145  * context and anonymous VMA name within the range [start, end).
1146  *
1147  * As a result, we might be able to merge the newly modified VMA range with an
1148  * adjacent VMA with identical properties.
1149  *
1150  * If no merge is possible and the range does not span the entirety of the VMA,
1151  * we then need to split the VMA to accommodate the change.
1152  *
1153  * The function returns either the merged VMA, the original VMA if a split was
1154  * required instead, or an error if the split failed.
1155  */
1156 struct vm_area_struct *vma_modify(struct vma_iterator *vmi,
1157 				  struct vm_area_struct *prev,
1158 				  struct vm_area_struct *vma,
1159 				  unsigned long start, unsigned long end,
1160 				  unsigned long vm_flags,
1161 				  struct mempolicy *policy,
1162 				  struct vm_userfaultfd_ctx uffd_ctx,
1163 				  struct anon_vma_name *anon_name)
1164 {
1165 	pgoff_t pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
1166 	struct vm_area_struct *merged;
1167 
1168 	merged = vma_merge(vmi, prev, vma, start, end, vm_flags,
1169 			   pgoff, policy, uffd_ctx, anon_name);
1170 	if (merged)
1171 		return merged;
1172 
1173 	if (vma->vm_start < start) {
1174 		int err = split_vma(vmi, vma, start, 1);
1175 
1176 		if (err)
1177 			return ERR_PTR(err);
1178 	}
1179 
1180 	if (vma->vm_end > end) {
1181 		int err = split_vma(vmi, vma, end, 0);
1182 
1183 		if (err)
1184 			return ERR_PTR(err);
1185 	}
1186 
1187 	return vma;
1188 }
1189 
1190 /*
1191  * Attempt to merge a newly mapped VMA with those adjacent to it. The caller
1192  * must ensure that [start, end) does not overlap any existing VMA.
1193  */
1194 struct vm_area_struct
1195 *vma_merge_new_vma(struct vma_iterator *vmi, struct vm_area_struct *prev,
1196 		   struct vm_area_struct *vma, unsigned long start,
1197 		   unsigned long end, pgoff_t pgoff)
1198 {
1199 	return vma_merge(vmi, prev, vma, start, end, vma->vm_flags, pgoff,
1200 			 vma_policy(vma), vma->vm_userfaultfd_ctx, anon_vma_name(vma));
1201 }
1202 
1203 /*
1204  * Expand vma by delta bytes, potentially merging with an immediately adjacent
1205  * VMA with identical properties.
1206  */
1207 struct vm_area_struct *vma_merge_extend(struct vma_iterator *vmi,
1208 					struct vm_area_struct *vma,
1209 					unsigned long delta)
1210 {
1211 	pgoff_t pgoff = vma->vm_pgoff + vma_pages(vma);
1212 
1213 	/* vma is specified as prev, so case 1 or 2 will apply. */
1214 	return vma_merge(vmi, vma, vma, vma->vm_end, vma->vm_end + delta,
1215 			 vma->vm_flags, pgoff, vma_policy(vma),
1216 			 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
1217 }
1218 
1219 void unlink_file_vma_batch_init(struct unlink_vma_file_batch *vb)
1220 {
1221 	vb->count = 0;
1222 }
1223 
1224 static void unlink_file_vma_batch_process(struct unlink_vma_file_batch *vb)
1225 {
1226 	struct address_space *mapping;
1227 	int i;
1228 
1229 	mapping = vb->vmas[0]->vm_file->f_mapping;
1230 	i_mmap_lock_write(mapping);
1231 	for (i = 0; i < vb->count; i++) {
1232 		VM_WARN_ON_ONCE(vb->vmas[i]->vm_file->f_mapping != mapping);
1233 		__remove_shared_vm_struct(vb->vmas[i], mapping);
1234 	}
1235 	i_mmap_unlock_write(mapping);
1236 
1237 	unlink_file_vma_batch_init(vb);
1238 }
1239 
1240 void unlink_file_vma_batch_add(struct unlink_vma_file_batch *vb,
1241 			       struct vm_area_struct *vma)
1242 {
1243 	if (vma->vm_file == NULL)
1244 		return;
1245 
1246 	if ((vb->count > 0 && vb->vmas[0]->vm_file != vma->vm_file) ||
1247 	    vb->count == ARRAY_SIZE(vb->vmas))
1248 		unlink_file_vma_batch_process(vb);
1249 
1250 	vb->vmas[vb->count] = vma;
1251 	vb->count++;
1252 }
1253 
1254 void unlink_file_vma_batch_final(struct unlink_vma_file_batch *vb)
1255 {
1256 	if (vb->count > 0)
1257 		unlink_file_vma_batch_process(vb);
1258 }
1259 
1260 /*
1261  * Unlink a file-based vm structure from its interval tree, to hide
1262  * vma from rmap and vmtruncate before freeing its page tables.
1263  */
1264 void unlink_file_vma(struct vm_area_struct *vma)
1265 {
1266 	struct file *file = vma->vm_file;
1267 
1268 	if (file) {
1269 		struct address_space *mapping = file->f_mapping;
1270 
1271 		i_mmap_lock_write(mapping);
1272 		__remove_shared_vm_struct(vma, mapping);
1273 		i_mmap_unlock_write(mapping);
1274 	}
1275 }
1276 
1277 void vma_link_file(struct vm_area_struct *vma)
1278 {
1279 	struct file *file = vma->vm_file;
1280 	struct address_space *mapping;
1281 
1282 	if (file) {
1283 		mapping = file->f_mapping;
1284 		i_mmap_lock_write(mapping);
1285 		__vma_link_file(vma, mapping);
1286 		i_mmap_unlock_write(mapping);
1287 	}
1288 }
1289 
1290 int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
1291 {
1292 	VMA_ITERATOR(vmi, mm, 0);
1293 
1294 	vma_iter_config(&vmi, vma->vm_start, vma->vm_end);
1295 	if (vma_iter_prealloc(&vmi, vma))
1296 		return -ENOMEM;
1297 
1298 	vma_start_write(vma);
1299 	vma_iter_store(&vmi, vma);
1300 	vma_link_file(vma);
1301 	mm->map_count++;
1302 	validate_mm(mm);
1303 	return 0;
1304 }
1305 
1306 /*
1307  * Copy the vma structure to a new location in the same mm,
1308  * prior to moving page table entries, to effect an mremap move.
1309  */
1310 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
1311 	unsigned long addr, unsigned long len, pgoff_t pgoff,
1312 	bool *need_rmap_locks)
1313 {
1314 	struct vm_area_struct *vma = *vmap;
1315 	unsigned long vma_start = vma->vm_start;
1316 	struct mm_struct *mm = vma->vm_mm;
1317 	struct vm_area_struct *new_vma, *prev;
1318 	bool faulted_in_anon_vma = true;
1319 	VMA_ITERATOR(vmi, mm, addr);
1320 
1321 	/*
1322 	 * If anonymous vma has not yet been faulted, update new pgoff
1323 	 * to match new location, to increase its chance of merging.
1324 	 */
1325 	if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
1326 		pgoff = addr >> PAGE_SHIFT;
1327 		faulted_in_anon_vma = false;
1328 	}
1329 
1330 	new_vma = find_vma_prev(mm, addr, &prev);
1331 	if (new_vma && new_vma->vm_start < addr + len)
1332 		return NULL;	/* should never get here */
1333 
1334 	new_vma = vma_merge_new_vma(&vmi, prev, vma, addr, addr + len, pgoff);
1335 	if (new_vma) {
1336 		/*
1337 		 * Source vma may have been merged into new_vma
1338 		 */
1339 		if (unlikely(vma_start >= new_vma->vm_start &&
1340 			     vma_start < new_vma->vm_end)) {
1341 			/*
1342 			 * The only way we can get a vma_merge with
1343 			 * self during an mremap is if the vma hasn't
1344 			 * been faulted in yet and we were allowed to
1345 			 * reset the dst vma->vm_pgoff to the
1346 			 * destination address of the mremap to allow
1347 			 * the merge to happen. mremap must change the
1348 			 * vm_pgoff linearity between src and dst vmas
1349 			 * (in turn preventing a vma_merge) to be
1350 			 * safe. It is only safe to keep the vm_pgoff
1351 			 * linear if there are no pages mapped yet.
1352 			 */
1353 			VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
1354 			*vmap = vma = new_vma;
1355 		}
1356 		*need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
1357 	} else {
1358 		new_vma = vm_area_dup(vma);
1359 		if (!new_vma)
1360 			goto out;
1361 		vma_set_range(new_vma, addr, addr + len, pgoff);
1362 		if (vma_dup_policy(vma, new_vma))
1363 			goto out_free_vma;
1364 		if (anon_vma_clone(new_vma, vma))
1365 			goto out_free_mempol;
1366 		if (new_vma->vm_file)
1367 			get_file(new_vma->vm_file);
1368 		if (new_vma->vm_ops && new_vma->vm_ops->open)
1369 			new_vma->vm_ops->open(new_vma);
1370 		if (vma_link(mm, new_vma))
1371 			goto out_vma_link;
1372 		*need_rmap_locks = false;
1373 	}
1374 	return new_vma;
1375 
1376 out_vma_link:
1377 	if (new_vma->vm_ops && new_vma->vm_ops->close)
1378 		new_vma->vm_ops->close(new_vma);
1379 
1380 	if (new_vma->vm_file)
1381 		fput(new_vma->vm_file);
1382 
1383 	unlink_anon_vmas(new_vma);
1384 out_free_mempol:
1385 	mpol_put(vma_policy(new_vma));
1386 out_free_vma:
1387 	vm_area_free(new_vma);
1388 out:
1389 	return NULL;
1390 }
1391 
1392 /*
1393  * Rough compatibility check to quickly see if it's even worth looking
1394  * at sharing an anon_vma.
1395  *
1396  * They need to have the same vm_file, and the flags can only differ
1397  * in things that mprotect may change.
1398  *
1399  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1400  * we can merge the two vma's. For example, we refuse to merge a vma if
1401  * there is a vm_ops->close() function, because that indicates that the
1402  * driver is doing some kind of reference counting. But that doesn't
1403  * really matter for the anon_vma sharing case.
1404  */
1405 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1406 {
1407 	return a->vm_end == b->vm_start &&
1408 		mpol_equal(vma_policy(a), vma_policy(b)) &&
1409 		a->vm_file == b->vm_file &&
1410 		!((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1411 		b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1412 }
1413 
1414 /*
1415  * Do some basic sanity checking to see if we can re-use the anon_vma
1416  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1417  * the same as 'old', the other will be the new one that is trying
1418  * to share the anon_vma.
1419  *
1420  * NOTE! This runs with mmap_lock held for reading, so it is possible that
1421  * the anon_vma of 'old' is concurrently in the process of being set up
1422  * by another page fault trying to merge _that_. But that's ok: if it
1423  * is being set up, that automatically means that it will be a singleton
1424  * acceptable for merging, so we can do all of this optimistically. But
1425  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1426  *
1427  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1428  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1429  * is to return an anon_vma that is "complex" due to having gone through
1430  * a fork).
1431  *
1432  * We also make sure that the two vma's are compatible (adjacent,
1433  * and with the same memory policies). That's all stable, even with just
1434  * a read lock on the mmap_lock.
1435  */
1436 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old,
1437 					  struct vm_area_struct *a,
1438 					  struct vm_area_struct *b)
1439 {
1440 	if (anon_vma_compatible(a, b)) {
1441 		struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1442 
1443 		if (anon_vma && list_is_singular(&old->anon_vma_chain))
1444 			return anon_vma;
1445 	}
1446 	return NULL;
1447 }
1448 
1449 /*
1450  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1451  * neighbouring vmas for a suitable anon_vma, before it goes off
1452  * to allocate a new anon_vma.  It checks because a repetitive
1453  * sequence of mprotects and faults may otherwise lead to distinct
1454  * anon_vmas being allocated, preventing vma merge in subsequent
1455  * mprotect.
1456  */
1457 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1458 {
1459 	struct anon_vma *anon_vma = NULL;
1460 	struct vm_area_struct *prev, *next;
1461 	VMA_ITERATOR(vmi, vma->vm_mm, vma->vm_end);
1462 
1463 	/* Try next first. */
1464 	next = vma_iter_load(&vmi);
1465 	if (next) {
1466 		anon_vma = reusable_anon_vma(next, vma, next);
1467 		if (anon_vma)
1468 			return anon_vma;
1469 	}
1470 
1471 	prev = vma_prev(&vmi);
1472 	VM_BUG_ON_VMA(prev != vma, vma);
1473 	prev = vma_prev(&vmi);
1474 	/* Try prev next. */
1475 	if (prev)
1476 		anon_vma = reusable_anon_vma(prev, prev, vma);
1477 
1478 	/*
1479 	 * We might reach here with anon_vma == NULL if we can't find
1480 	 * any reusable anon_vma.
1481 	 * There's no absolute need to look only at touching neighbours:
1482 	 * we could search further afield for "compatible" anon_vmas.
1483 	 * But it would probably just be a waste of time searching,
1484 	 * or lead to too many vmas hanging off the same anon_vma.
1485 	 * We're trying to allow mprotect remerging later on,
1486 	 * not trying to minimize memory used for anon_vmas.
1487 	 */
1488 	return anon_vma;
1489 }
1490 
1491 static bool vm_ops_needs_writenotify(const struct vm_operations_struct *vm_ops)
1492 {
1493 	return vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite);
1494 }
1495 
1496 static bool vma_is_shared_writable(struct vm_area_struct *vma)
1497 {
1498 	return (vma->vm_flags & (VM_WRITE | VM_SHARED)) ==
1499 		(VM_WRITE | VM_SHARED);
1500 }
1501 
1502 static bool vma_fs_can_writeback(struct vm_area_struct *vma)
1503 {
1504 	/* No managed pages to writeback. */
1505 	if (vma->vm_flags & VM_PFNMAP)
1506 		return false;
1507 
1508 	return vma->vm_file && vma->vm_file->f_mapping &&
1509 		mapping_can_writeback(vma->vm_file->f_mapping);
1510 }
1511 
1512 /*
1513  * Does this VMA require the underlying folios to have their dirty state
1514  * tracked?
1515  */
1516 bool vma_needs_dirty_tracking(struct vm_area_struct *vma)
1517 {
1518 	/* Only shared, writable VMAs require dirty tracking. */
1519 	if (!vma_is_shared_writable(vma))
1520 		return false;
1521 
1522 	/* Does the filesystem need to be notified? */
1523 	if (vm_ops_needs_writenotify(vma->vm_ops))
1524 		return true;
1525 
1526 	/*
1527 	 * Even if the filesystem doesn't indicate a need for writenotify, if it
1528 	 * can writeback, dirty tracking is still required.
1529 	 */
1530 	return vma_fs_can_writeback(vma);
1531 }
1532 
1533 /*
1534  * Some shared mappings will want the pages marked read-only
1535  * to track write events. If so, we'll downgrade vm_page_prot
1536  * to the private version (using protection_map[] without the
1537  * VM_SHARED bit).
1538  */
1539 bool vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1540 {
1541 	/* If it was private or non-writable, the write bit is already clear */
1542 	if (!vma_is_shared_writable(vma))
1543 		return false;
1544 
1545 	/* The backer wishes to know when pages are first written to? */
1546 	if (vm_ops_needs_writenotify(vma->vm_ops))
1547 		return true;
1548 
1549 	/* The open routine did something to the protections that pgprot_modify
1550 	 * won't preserve? */
1551 	if (pgprot_val(vm_page_prot) !=
1552 	    pgprot_val(vm_pgprot_modify(vm_page_prot, vma->vm_flags)))
1553 		return false;
1554 
1555 	/*
1556 	 * Do we need to track softdirty? hugetlb does not support softdirty
1557 	 * tracking yet.
1558 	 */
1559 	if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1560 		return true;
1561 
1562 	/* Do we need write faults for uffd-wp tracking? */
1563 	if (userfaultfd_wp(vma))
1564 		return true;
1565 
1566 	/* Can the mapping track the dirty pages? */
1567 	return vma_fs_can_writeback(vma);
1568 }
1569 
1570 unsigned long count_vma_pages_range(struct mm_struct *mm,
1571 				    unsigned long addr, unsigned long end)
1572 {
1573 	VMA_ITERATOR(vmi, mm, addr);
1574 	struct vm_area_struct *vma;
1575 	unsigned long nr_pages = 0;
1576 
1577 	for_each_vma_range(vmi, vma, end) {
1578 		unsigned long vm_start = max(addr, vma->vm_start);
1579 		unsigned long vm_end = min(end, vma->vm_end);
1580 
1581 		nr_pages += PHYS_PFN(vm_end - vm_start);
1582 	}
1583 
1584 	return nr_pages;
1585 }
1586 
1587 static DEFINE_MUTEX(mm_all_locks_mutex);
1588 
1589 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
1590 {
1591 	if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
1592 		/*
1593 		 * The LSB of head.next can't change from under us
1594 		 * because we hold the mm_all_locks_mutex.
1595 		 */
1596 		down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
1597 		/*
1598 		 * We can safely modify head.next after taking the
1599 		 * anon_vma->root->rwsem. If some other vma in this mm shares
1600 		 * the same anon_vma we won't take it again.
1601 		 *
1602 		 * No need of atomic instructions here, head.next
1603 		 * can't change from under us thanks to the
1604 		 * anon_vma->root->rwsem.
1605 		 */
1606 		if (__test_and_set_bit(0, (unsigned long *)
1607 				       &anon_vma->root->rb_root.rb_root.rb_node))
1608 			BUG();
1609 	}
1610 }
1611 
1612 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
1613 {
1614 	if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
1615 		/*
1616 		 * AS_MM_ALL_LOCKS can't change from under us because
1617 		 * we hold the mm_all_locks_mutex.
1618 		 *
1619 		 * Operations on ->flags have to be atomic because
1620 		 * even if AS_MM_ALL_LOCKS is stable thanks to the
1621 		 * mm_all_locks_mutex, there may be other cpus
1622 		 * changing other bitflags in parallel to us.
1623 		 */
1624 		if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
1625 			BUG();
1626 		down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
1627 	}
1628 }
1629 
1630 /*
1631  * This operation locks against the VM for all pte/vma/mm related
1632  * operations that could ever happen on a certain mm. This includes
1633  * vmtruncate, try_to_unmap, and all page faults.
1634  *
1635  * The caller must take the mmap_lock in write mode before calling
1636  * mm_take_all_locks(). The caller isn't allowed to release the
1637  * mmap_lock until mm_drop_all_locks() returns.
1638  *
1639  * mmap_lock in write mode is required in order to block all operations
1640  * that could modify pagetables and free pages without need of
1641  * altering the vma layout. It's also needed in write mode to avoid new
1642  * anon_vmas to be associated with existing vmas.
1643  *
1644  * A single task can't take more than one mm_take_all_locks() in a row
1645  * or it would deadlock.
1646  *
1647  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
1648  * mapping->flags avoid to take the same lock twice, if more than one
1649  * vma in this mm is backed by the same anon_vma or address_space.
1650  *
1651  * We take locks in following order, accordingly to comment at beginning
1652  * of mm/rmap.c:
1653  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
1654  *     hugetlb mapping);
1655  *   - all vmas marked locked
1656  *   - all i_mmap_rwsem locks;
1657  *   - all anon_vma->rwseml
1658  *
1659  * We can take all locks within these types randomly because the VM code
1660  * doesn't nest them and we protected from parallel mm_take_all_locks() by
1661  * mm_all_locks_mutex.
1662  *
1663  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
1664  * that may have to take thousand of locks.
1665  *
1666  * mm_take_all_locks() can fail if it's interrupted by signals.
1667  */
1668 int mm_take_all_locks(struct mm_struct *mm)
1669 {
1670 	struct vm_area_struct *vma;
1671 	struct anon_vma_chain *avc;
1672 	VMA_ITERATOR(vmi, mm, 0);
1673 
1674 	mmap_assert_write_locked(mm);
1675 
1676 	mutex_lock(&mm_all_locks_mutex);
1677 
1678 	/*
1679 	 * vma_start_write() does not have a complement in mm_drop_all_locks()
1680 	 * because vma_start_write() is always asymmetrical; it marks a VMA as
1681 	 * being written to until mmap_write_unlock() or mmap_write_downgrade()
1682 	 * is reached.
1683 	 */
1684 	for_each_vma(vmi, vma) {
1685 		if (signal_pending(current))
1686 			goto out_unlock;
1687 		vma_start_write(vma);
1688 	}
1689 
1690 	vma_iter_init(&vmi, mm, 0);
1691 	for_each_vma(vmi, vma) {
1692 		if (signal_pending(current))
1693 			goto out_unlock;
1694 		if (vma->vm_file && vma->vm_file->f_mapping &&
1695 				is_vm_hugetlb_page(vma))
1696 			vm_lock_mapping(mm, vma->vm_file->f_mapping);
1697 	}
1698 
1699 	vma_iter_init(&vmi, mm, 0);
1700 	for_each_vma(vmi, vma) {
1701 		if (signal_pending(current))
1702 			goto out_unlock;
1703 		if (vma->vm_file && vma->vm_file->f_mapping &&
1704 				!is_vm_hugetlb_page(vma))
1705 			vm_lock_mapping(mm, vma->vm_file->f_mapping);
1706 	}
1707 
1708 	vma_iter_init(&vmi, mm, 0);
1709 	for_each_vma(vmi, vma) {
1710 		if (signal_pending(current))
1711 			goto out_unlock;
1712 		if (vma->anon_vma)
1713 			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
1714 				vm_lock_anon_vma(mm, avc->anon_vma);
1715 	}
1716 
1717 	return 0;
1718 
1719 out_unlock:
1720 	mm_drop_all_locks(mm);
1721 	return -EINTR;
1722 }
1723 
1724 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
1725 {
1726 	if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
1727 		/*
1728 		 * The LSB of head.next can't change to 0 from under
1729 		 * us because we hold the mm_all_locks_mutex.
1730 		 *
1731 		 * We must however clear the bitflag before unlocking
1732 		 * the vma so the users using the anon_vma->rb_root will
1733 		 * never see our bitflag.
1734 		 *
1735 		 * No need of atomic instructions here, head.next
1736 		 * can't change from under us until we release the
1737 		 * anon_vma->root->rwsem.
1738 		 */
1739 		if (!__test_and_clear_bit(0, (unsigned long *)
1740 					  &anon_vma->root->rb_root.rb_root.rb_node))
1741 			BUG();
1742 		anon_vma_unlock_write(anon_vma);
1743 	}
1744 }
1745 
1746 static void vm_unlock_mapping(struct address_space *mapping)
1747 {
1748 	if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
1749 		/*
1750 		 * AS_MM_ALL_LOCKS can't change to 0 from under us
1751 		 * because we hold the mm_all_locks_mutex.
1752 		 */
1753 		i_mmap_unlock_write(mapping);
1754 		if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
1755 					&mapping->flags))
1756 			BUG();
1757 	}
1758 }
1759 
1760 /*
1761  * The mmap_lock cannot be released by the caller until
1762  * mm_drop_all_locks() returns.
1763  */
1764 void mm_drop_all_locks(struct mm_struct *mm)
1765 {
1766 	struct vm_area_struct *vma;
1767 	struct anon_vma_chain *avc;
1768 	VMA_ITERATOR(vmi, mm, 0);
1769 
1770 	mmap_assert_write_locked(mm);
1771 	BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
1772 
1773 	for_each_vma(vmi, vma) {
1774 		if (vma->anon_vma)
1775 			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
1776 				vm_unlock_anon_vma(avc->anon_vma);
1777 		if (vma->vm_file && vma->vm_file->f_mapping)
1778 			vm_unlock_mapping(vma->vm_file->f_mapping);
1779 	}
1780 
1781 	mutex_unlock(&mm_all_locks_mutex);
1782 }
1783