xref: /openbmc/linux/mm/rmap.c (revision 3c81195a04e13833196462ab398d8bcf282701f7)
1  /*
2   * mm/rmap.c - physical to virtual reverse mappings
3   *
4   * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5   * Released under the General Public License (GPL).
6   *
7   * Simple, low overhead reverse mapping scheme.
8   * Please try to keep this thing as modular as possible.
9   *
10   * Provides methods for unmapping each kind of mapped page:
11   * the anon methods track anonymous pages, and
12   * the file methods track pages belonging to an inode.
13   *
14   * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15   * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16   * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
17   * Contributions by Hugh Dickins 2003, 2004
18   */
19  
20  /*
21   * Lock ordering in mm:
22   *
23   * inode->i_rwsem	(while writing or truncating, not reading or faulting)
24   *   mm->mmap_lock
25   *     mapping->invalidate_lock (in filemap_fault)
26   *       page->flags PG_locked (lock_page)
27   *         hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share, see hugetlbfs below)
28   *           vma_start_write
29   *             mapping->i_mmap_rwsem
30   *               anon_vma->rwsem
31   *                 mm->page_table_lock or pte_lock
32   *                   swap_lock (in swap_duplicate, swap_info_get)
33   *                     mmlist_lock (in mmput, drain_mmlist and others)
34   *                     mapping->private_lock (in block_dirty_folio)
35   *                       folio_lock_memcg move_lock (in block_dirty_folio)
36   *                         i_pages lock (widely used)
37   *                           lruvec->lru_lock (in folio_lruvec_lock_irq)
38   *                     inode->i_lock (in set_page_dirty's __mark_inode_dirty)
39   *                     bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
40   *                       sb_lock (within inode_lock in fs/fs-writeback.c)
41   *                       i_pages lock (widely used, in set_page_dirty,
42   *                                 in arch-dependent flush_dcache_mmap_lock,
43   *                                 within bdi.wb->list_lock in __sync_single_inode)
44   *
45   * anon_vma->rwsem,mapping->i_mmap_rwsem   (memory_failure, collect_procs_anon)
46   *   ->tasklist_lock
47   *     pte map lock
48   *
49   * hugetlbfs PageHuge() take locks in this order:
50   *   hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
51   *     vma_lock (hugetlb specific lock for pmd_sharing)
52   *       mapping->i_mmap_rwsem (also used for hugetlb pmd sharing)
53   *         page->flags PG_locked (lock_page)
54   */
55  
56  #include <linux/mm.h>
57  #include <linux/sched/mm.h>
58  #include <linux/sched/task.h>
59  #include <linux/pagemap.h>
60  #include <linux/swap.h>
61  #include <linux/swapops.h>
62  #include <linux/slab.h>
63  #include <linux/init.h>
64  #include <linux/ksm.h>
65  #include <linux/rmap.h>
66  #include <linux/rcupdate.h>
67  #include <linux/export.h>
68  #include <linux/memcontrol.h>
69  #include <linux/mmu_notifier.h>
70  #include <linux/migrate.h>
71  #include <linux/hugetlb.h>
72  #include <linux/huge_mm.h>
73  #include <linux/backing-dev.h>
74  #include <linux/page_idle.h>
75  #include <linux/memremap.h>
76  #include <linux/userfaultfd_k.h>
77  #include <linux/mm_inline.h>
78  
79  #include <asm/tlbflush.h>
80  
81  #define CREATE_TRACE_POINTS
82  #include <trace/events/tlb.h>
83  #include <trace/events/migrate.h>
84  
85  #include "internal.h"
86  
87  static struct kmem_cache *anon_vma_cachep;
88  static struct kmem_cache *anon_vma_chain_cachep;
89  
90  static inline struct anon_vma *anon_vma_alloc(void)
91  {
92  	struct anon_vma *anon_vma;
93  
94  	anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
95  	if (anon_vma) {
96  		atomic_set(&anon_vma->refcount, 1);
97  		anon_vma->num_children = 0;
98  		anon_vma->num_active_vmas = 0;
99  		anon_vma->parent = anon_vma;
100  		/*
101  		 * Initialise the anon_vma root to point to itself. If called
102  		 * from fork, the root will be reset to the parents anon_vma.
103  		 */
104  		anon_vma->root = anon_vma;
105  	}
106  
107  	return anon_vma;
108  }
109  
110  static inline void anon_vma_free(struct anon_vma *anon_vma)
111  {
112  	VM_BUG_ON(atomic_read(&anon_vma->refcount));
113  
114  	/*
115  	 * Synchronize against folio_lock_anon_vma_read() such that
116  	 * we can safely hold the lock without the anon_vma getting
117  	 * freed.
118  	 *
119  	 * Relies on the full mb implied by the atomic_dec_and_test() from
120  	 * put_anon_vma() against the acquire barrier implied by
121  	 * down_read_trylock() from folio_lock_anon_vma_read(). This orders:
122  	 *
123  	 * folio_lock_anon_vma_read()	VS	put_anon_vma()
124  	 *   down_read_trylock()		  atomic_dec_and_test()
125  	 *   LOCK				  MB
126  	 *   atomic_read()			  rwsem_is_locked()
127  	 *
128  	 * LOCK should suffice since the actual taking of the lock must
129  	 * happen _before_ what follows.
130  	 */
131  	might_sleep();
132  	if (rwsem_is_locked(&anon_vma->root->rwsem)) {
133  		anon_vma_lock_write(anon_vma);
134  		anon_vma_unlock_write(anon_vma);
135  	}
136  
137  	kmem_cache_free(anon_vma_cachep, anon_vma);
138  }
139  
140  static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
141  {
142  	return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
143  }
144  
145  static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
146  {
147  	kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
148  }
149  
150  static void anon_vma_chain_link(struct vm_area_struct *vma,
151  				struct anon_vma_chain *avc,
152  				struct anon_vma *anon_vma)
153  {
154  	avc->vma = vma;
155  	avc->anon_vma = anon_vma;
156  	list_add(&avc->same_vma, &vma->anon_vma_chain);
157  	anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
158  }
159  
160  /**
161   * __anon_vma_prepare - attach an anon_vma to a memory region
162   * @vma: the memory region in question
163   *
164   * This makes sure the memory mapping described by 'vma' has
165   * an 'anon_vma' attached to it, so that we can associate the
166   * anonymous pages mapped into it with that anon_vma.
167   *
168   * The common case will be that we already have one, which
169   * is handled inline by anon_vma_prepare(). But if
170   * not we either need to find an adjacent mapping that we
171   * can re-use the anon_vma from (very common when the only
172   * reason for splitting a vma has been mprotect()), or we
173   * allocate a new one.
174   *
175   * Anon-vma allocations are very subtle, because we may have
176   * optimistically looked up an anon_vma in folio_lock_anon_vma_read()
177   * and that may actually touch the rwsem even in the newly
178   * allocated vma (it depends on RCU to make sure that the
179   * anon_vma isn't actually destroyed).
180   *
181   * As a result, we need to do proper anon_vma locking even
182   * for the new allocation. At the same time, we do not want
183   * to do any locking for the common case of already having
184   * an anon_vma.
185   *
186   * This must be called with the mmap_lock held for reading.
187   */
188  int __anon_vma_prepare(struct vm_area_struct *vma)
189  {
190  	struct mm_struct *mm = vma->vm_mm;
191  	struct anon_vma *anon_vma, *allocated;
192  	struct anon_vma_chain *avc;
193  
194  	might_sleep();
195  
196  	avc = anon_vma_chain_alloc(GFP_KERNEL);
197  	if (!avc)
198  		goto out_enomem;
199  
200  	anon_vma = find_mergeable_anon_vma(vma);
201  	allocated = NULL;
202  	if (!anon_vma) {
203  		anon_vma = anon_vma_alloc();
204  		if (unlikely(!anon_vma))
205  			goto out_enomem_free_avc;
206  		anon_vma->num_children++; /* self-parent link for new root */
207  		allocated = anon_vma;
208  	}
209  
210  	anon_vma_lock_write(anon_vma);
211  	/* page_table_lock to protect against threads */
212  	spin_lock(&mm->page_table_lock);
213  	if (likely(!vma->anon_vma)) {
214  		vma->anon_vma = anon_vma;
215  		anon_vma_chain_link(vma, avc, anon_vma);
216  		anon_vma->num_active_vmas++;
217  		allocated = NULL;
218  		avc = NULL;
219  	}
220  	spin_unlock(&mm->page_table_lock);
221  	anon_vma_unlock_write(anon_vma);
222  
223  	if (unlikely(allocated))
224  		put_anon_vma(allocated);
225  	if (unlikely(avc))
226  		anon_vma_chain_free(avc);
227  
228  	return 0;
229  
230   out_enomem_free_avc:
231  	anon_vma_chain_free(avc);
232   out_enomem:
233  	return -ENOMEM;
234  }
235  
236  /*
237   * This is a useful helper function for locking the anon_vma root as
238   * we traverse the vma->anon_vma_chain, looping over anon_vma's that
239   * have the same vma.
240   *
241   * Such anon_vma's should have the same root, so you'd expect to see
242   * just a single mutex_lock for the whole traversal.
243   */
244  static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
245  {
246  	struct anon_vma *new_root = anon_vma->root;
247  	if (new_root != root) {
248  		if (WARN_ON_ONCE(root))
249  			up_write(&root->rwsem);
250  		root = new_root;
251  		down_write(&root->rwsem);
252  	}
253  	return root;
254  }
255  
256  static inline void unlock_anon_vma_root(struct anon_vma *root)
257  {
258  	if (root)
259  		up_write(&root->rwsem);
260  }
261  
262  /*
263   * Attach the anon_vmas from src to dst.
264   * Returns 0 on success, -ENOMEM on failure.
265   *
266   * anon_vma_clone() is called by vma_expand(), vma_merge(), __split_vma(),
267   * copy_vma() and anon_vma_fork(). The first four want an exact copy of src,
268   * while the last one, anon_vma_fork(), may try to reuse an existing anon_vma to
269   * prevent endless growth of anon_vma. Since dst->anon_vma is set to NULL before
270   * call, we can identify this case by checking (!dst->anon_vma &&
271   * src->anon_vma).
272   *
273   * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
274   * and reuse existing anon_vma which has no vmas and only one child anon_vma.
275   * This prevents degradation of anon_vma hierarchy to endless linear chain in
276   * case of constantly forking task. On the other hand, an anon_vma with more
277   * than one child isn't reused even if there was no alive vma, thus rmap
278   * walker has a good chance of avoiding scanning the whole hierarchy when it
279   * searches where page is mapped.
280   */
281  int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
282  {
283  	struct anon_vma_chain *avc, *pavc;
284  	struct anon_vma *root = NULL;
285  
286  	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
287  		struct anon_vma *anon_vma;
288  
289  		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
290  		if (unlikely(!avc)) {
291  			unlock_anon_vma_root(root);
292  			root = NULL;
293  			avc = anon_vma_chain_alloc(GFP_KERNEL);
294  			if (!avc)
295  				goto enomem_failure;
296  		}
297  		anon_vma = pavc->anon_vma;
298  		root = lock_anon_vma_root(root, anon_vma);
299  		anon_vma_chain_link(dst, avc, anon_vma);
300  
301  		/*
302  		 * Reuse existing anon_vma if it has no vma and only one
303  		 * anon_vma child.
304  		 *
305  		 * Root anon_vma is never reused:
306  		 * it has self-parent reference and at least one child.
307  		 */
308  		if (!dst->anon_vma && src->anon_vma &&
309  		    anon_vma->num_children < 2 &&
310  		    anon_vma->num_active_vmas == 0)
311  			dst->anon_vma = anon_vma;
312  	}
313  	if (dst->anon_vma)
314  		dst->anon_vma->num_active_vmas++;
315  	unlock_anon_vma_root(root);
316  	return 0;
317  
318   enomem_failure:
319  	/*
320  	 * dst->anon_vma is dropped here otherwise its num_active_vmas can
321  	 * be incorrectly decremented in unlink_anon_vmas().
322  	 * We can safely do this because callers of anon_vma_clone() don't care
323  	 * about dst->anon_vma if anon_vma_clone() failed.
324  	 */
325  	dst->anon_vma = NULL;
326  	unlink_anon_vmas(dst);
327  	return -ENOMEM;
328  }
329  
330  /*
331   * Attach vma to its own anon_vma, as well as to the anon_vmas that
332   * the corresponding VMA in the parent process is attached to.
333   * Returns 0 on success, non-zero on failure.
334   */
335  int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
336  {
337  	struct anon_vma_chain *avc;
338  	struct anon_vma *anon_vma;
339  	int error;
340  
341  	/* Don't bother if the parent process has no anon_vma here. */
342  	if (!pvma->anon_vma)
343  		return 0;
344  
345  	/* Drop inherited anon_vma, we'll reuse existing or allocate new. */
346  	vma->anon_vma = NULL;
347  
348  	/*
349  	 * First, attach the new VMA to the parent VMA's anon_vmas,
350  	 * so rmap can find non-COWed pages in child processes.
351  	 */
352  	error = anon_vma_clone(vma, pvma);
353  	if (error)
354  		return error;
355  
356  	/* An existing anon_vma has been reused, all done then. */
357  	if (vma->anon_vma)
358  		return 0;
359  
360  	/* Then add our own anon_vma. */
361  	anon_vma = anon_vma_alloc();
362  	if (!anon_vma)
363  		goto out_error;
364  	anon_vma->num_active_vmas++;
365  	avc = anon_vma_chain_alloc(GFP_KERNEL);
366  	if (!avc)
367  		goto out_error_free_anon_vma;
368  
369  	/*
370  	 * The root anon_vma's rwsem is the lock actually used when we
371  	 * lock any of the anon_vmas in this anon_vma tree.
372  	 */
373  	anon_vma->root = pvma->anon_vma->root;
374  	anon_vma->parent = pvma->anon_vma;
375  	/*
376  	 * With refcounts, an anon_vma can stay around longer than the
377  	 * process it belongs to. The root anon_vma needs to be pinned until
378  	 * this anon_vma is freed, because the lock lives in the root.
379  	 */
380  	get_anon_vma(anon_vma->root);
381  	/* Mark this anon_vma as the one where our new (COWed) pages go. */
382  	vma->anon_vma = anon_vma;
383  	anon_vma_lock_write(anon_vma);
384  	anon_vma_chain_link(vma, avc, anon_vma);
385  	anon_vma->parent->num_children++;
386  	anon_vma_unlock_write(anon_vma);
387  
388  	return 0;
389  
390   out_error_free_anon_vma:
391  	put_anon_vma(anon_vma);
392   out_error:
393  	unlink_anon_vmas(vma);
394  	return -ENOMEM;
395  }
396  
397  void unlink_anon_vmas(struct vm_area_struct *vma)
398  {
399  	struct anon_vma_chain *avc, *next;
400  	struct anon_vma *root = NULL;
401  
402  	/*
403  	 * Unlink each anon_vma chained to the VMA.  This list is ordered
404  	 * from newest to oldest, ensuring the root anon_vma gets freed last.
405  	 */
406  	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
407  		struct anon_vma *anon_vma = avc->anon_vma;
408  
409  		root = lock_anon_vma_root(root, anon_vma);
410  		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
411  
412  		/*
413  		 * Leave empty anon_vmas on the list - we'll need
414  		 * to free them outside the lock.
415  		 */
416  		if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
417  			anon_vma->parent->num_children--;
418  			continue;
419  		}
420  
421  		list_del(&avc->same_vma);
422  		anon_vma_chain_free(avc);
423  	}
424  	if (vma->anon_vma) {
425  		vma->anon_vma->num_active_vmas--;
426  
427  		/*
428  		 * vma would still be needed after unlink, and anon_vma will be prepared
429  		 * when handle fault.
430  		 */
431  		vma->anon_vma = NULL;
432  	}
433  	unlock_anon_vma_root(root);
434  
435  	/*
436  	 * Iterate the list once more, it now only contains empty and unlinked
437  	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
438  	 * needing to write-acquire the anon_vma->root->rwsem.
439  	 */
440  	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
441  		struct anon_vma *anon_vma = avc->anon_vma;
442  
443  		VM_WARN_ON(anon_vma->num_children);
444  		VM_WARN_ON(anon_vma->num_active_vmas);
445  		put_anon_vma(anon_vma);
446  
447  		list_del(&avc->same_vma);
448  		anon_vma_chain_free(avc);
449  	}
450  }
451  
452  static void anon_vma_ctor(void *data)
453  {
454  	struct anon_vma *anon_vma = data;
455  
456  	init_rwsem(&anon_vma->rwsem);
457  	atomic_set(&anon_vma->refcount, 0);
458  	anon_vma->rb_root = RB_ROOT_CACHED;
459  }
460  
461  void __init anon_vma_init(void)
462  {
463  	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
464  			0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
465  			anon_vma_ctor);
466  	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
467  			SLAB_PANIC|SLAB_ACCOUNT);
468  }
469  
470  /*
471   * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
472   *
473   * Since there is no serialization what so ever against page_remove_rmap()
474   * the best this function can do is return a refcount increased anon_vma
475   * that might have been relevant to this page.
476   *
477   * The page might have been remapped to a different anon_vma or the anon_vma
478   * returned may already be freed (and even reused).
479   *
480   * In case it was remapped to a different anon_vma, the new anon_vma will be a
481   * child of the old anon_vma, and the anon_vma lifetime rules will therefore
482   * ensure that any anon_vma obtained from the page will still be valid for as
483   * long as we observe page_mapped() [ hence all those page_mapped() tests ].
484   *
485   * All users of this function must be very careful when walking the anon_vma
486   * chain and verify that the page in question is indeed mapped in it
487   * [ something equivalent to page_mapped_in_vma() ].
488   *
489   * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
490   * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
491   * if there is a mapcount, we can dereference the anon_vma after observing
492   * those.
493   */
494  struct anon_vma *folio_get_anon_vma(struct folio *folio)
495  {
496  	struct anon_vma *anon_vma = NULL;
497  	unsigned long anon_mapping;
498  
499  	rcu_read_lock();
500  	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
501  	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
502  		goto out;
503  	if (!folio_mapped(folio))
504  		goto out;
505  
506  	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
507  	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
508  		anon_vma = NULL;
509  		goto out;
510  	}
511  
512  	/*
513  	 * If this folio is still mapped, then its anon_vma cannot have been
514  	 * freed.  But if it has been unmapped, we have no security against the
515  	 * anon_vma structure being freed and reused (for another anon_vma:
516  	 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
517  	 * above cannot corrupt).
518  	 */
519  	if (!folio_mapped(folio)) {
520  		rcu_read_unlock();
521  		put_anon_vma(anon_vma);
522  		return NULL;
523  	}
524  out:
525  	rcu_read_unlock();
526  
527  	return anon_vma;
528  }
529  
530  /*
531   * Similar to folio_get_anon_vma() except it locks the anon_vma.
532   *
533   * Its a little more complex as it tries to keep the fast path to a single
534   * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
535   * reference like with folio_get_anon_vma() and then block on the mutex
536   * on !rwc->try_lock case.
537   */
538  struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
539  					  struct rmap_walk_control *rwc)
540  {
541  	struct anon_vma *anon_vma = NULL;
542  	struct anon_vma *root_anon_vma;
543  	unsigned long anon_mapping;
544  
545  	rcu_read_lock();
546  	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
547  	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
548  		goto out;
549  	if (!folio_mapped(folio))
550  		goto out;
551  
552  	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
553  	root_anon_vma = READ_ONCE(anon_vma->root);
554  	if (down_read_trylock(&root_anon_vma->rwsem)) {
555  		/*
556  		 * If the folio is still mapped, then this anon_vma is still
557  		 * its anon_vma, and holding the mutex ensures that it will
558  		 * not go away, see anon_vma_free().
559  		 */
560  		if (!folio_mapped(folio)) {
561  			up_read(&root_anon_vma->rwsem);
562  			anon_vma = NULL;
563  		}
564  		goto out;
565  	}
566  
567  	if (rwc && rwc->try_lock) {
568  		anon_vma = NULL;
569  		rwc->contended = true;
570  		goto out;
571  	}
572  
573  	/* trylock failed, we got to sleep */
574  	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
575  		anon_vma = NULL;
576  		goto out;
577  	}
578  
579  	if (!folio_mapped(folio)) {
580  		rcu_read_unlock();
581  		put_anon_vma(anon_vma);
582  		return NULL;
583  	}
584  
585  	/* we pinned the anon_vma, its safe to sleep */
586  	rcu_read_unlock();
587  	anon_vma_lock_read(anon_vma);
588  
589  	if (atomic_dec_and_test(&anon_vma->refcount)) {
590  		/*
591  		 * Oops, we held the last refcount, release the lock
592  		 * and bail -- can't simply use put_anon_vma() because
593  		 * we'll deadlock on the anon_vma_lock_write() recursion.
594  		 */
595  		anon_vma_unlock_read(anon_vma);
596  		__put_anon_vma(anon_vma);
597  		anon_vma = NULL;
598  	}
599  
600  	return anon_vma;
601  
602  out:
603  	rcu_read_unlock();
604  	return anon_vma;
605  }
606  
607  #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
608  /*
609   * Flush TLB entries for recently unmapped pages from remote CPUs. It is
610   * important if a PTE was dirty when it was unmapped that it's flushed
611   * before any IO is initiated on the page to prevent lost writes. Similarly,
612   * it must be flushed before freeing to prevent data leakage.
613   */
614  void try_to_unmap_flush(void)
615  {
616  	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
617  
618  	if (!tlb_ubc->flush_required)
619  		return;
620  
621  	arch_tlbbatch_flush(&tlb_ubc->arch);
622  	tlb_ubc->flush_required = false;
623  	tlb_ubc->writable = false;
624  }
625  
626  /* Flush iff there are potentially writable TLB entries that can race with IO */
627  void try_to_unmap_flush_dirty(void)
628  {
629  	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
630  
631  	if (tlb_ubc->writable)
632  		try_to_unmap_flush();
633  }
634  
635  /*
636   * Bits 0-14 of mm->tlb_flush_batched record pending generations.
637   * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations.
638   */
639  #define TLB_FLUSH_BATCH_FLUSHED_SHIFT	16
640  #define TLB_FLUSH_BATCH_PENDING_MASK			\
641  	((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1)
642  #define TLB_FLUSH_BATCH_PENDING_LARGE			\
643  	(TLB_FLUSH_BATCH_PENDING_MASK / 2)
644  
645  static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval)
646  {
647  	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
648  	int batch;
649  	bool writable = pte_dirty(pteval);
650  
651  	if (!pte_accessible(mm, pteval))
652  		return;
653  
654  	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
655  	tlb_ubc->flush_required = true;
656  
657  	/*
658  	 * Ensure compiler does not re-order the setting of tlb_flush_batched
659  	 * before the PTE is cleared.
660  	 */
661  	barrier();
662  	batch = atomic_read(&mm->tlb_flush_batched);
663  retry:
664  	if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) {
665  		/*
666  		 * Prevent `pending' from catching up with `flushed' because of
667  		 * overflow.  Reset `pending' and `flushed' to be 1 and 0 if
668  		 * `pending' becomes large.
669  		 */
670  		if (!atomic_try_cmpxchg(&mm->tlb_flush_batched, &batch, 1))
671  			goto retry;
672  	} else {
673  		atomic_inc(&mm->tlb_flush_batched);
674  	}
675  
676  	/*
677  	 * If the PTE was dirty then it's best to assume it's writable. The
678  	 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
679  	 * before the page is queued for IO.
680  	 */
681  	if (writable)
682  		tlb_ubc->writable = true;
683  }
684  
685  /*
686   * Returns true if the TLB flush should be deferred to the end of a batch of
687   * unmap operations to reduce IPIs.
688   */
689  static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
690  {
691  	bool should_defer = false;
692  
693  	if (!(flags & TTU_BATCH_FLUSH))
694  		return false;
695  
696  	/* If remote CPUs need to be flushed then defer batch the flush */
697  	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
698  		should_defer = true;
699  	put_cpu();
700  
701  	return should_defer;
702  }
703  
704  /*
705   * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
706   * releasing the PTL if TLB flushes are batched. It's possible for a parallel
707   * operation such as mprotect or munmap to race between reclaim unmapping
708   * the page and flushing the page. If this race occurs, it potentially allows
709   * access to data via a stale TLB entry. Tracking all mm's that have TLB
710   * batching in flight would be expensive during reclaim so instead track
711   * whether TLB batching occurred in the past and if so then do a flush here
712   * if required. This will cost one additional flush per reclaim cycle paid
713   * by the first operation at risk such as mprotect and mumap.
714   *
715   * This must be called under the PTL so that an access to tlb_flush_batched
716   * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
717   * via the PTL.
718   */
719  void flush_tlb_batched_pending(struct mm_struct *mm)
720  {
721  	int batch = atomic_read(&mm->tlb_flush_batched);
722  	int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK;
723  	int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT;
724  
725  	if (pending != flushed) {
726  		flush_tlb_mm(mm);
727  		/*
728  		 * If the new TLB flushing is pending during flushing, leave
729  		 * mm->tlb_flush_batched as is, to avoid losing flushing.
730  		 */
731  		atomic_cmpxchg(&mm->tlb_flush_batched, batch,
732  			       pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT));
733  	}
734  }
735  #else
736  static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval)
737  {
738  }
739  
740  static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
741  {
742  	return false;
743  }
744  #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
745  
746  /*
747   * At what user virtual address is page expected in vma?
748   * Caller should check the page is actually part of the vma.
749   */
750  unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
751  {
752  	struct folio *folio = page_folio(page);
753  	if (folio_test_anon(folio)) {
754  		struct anon_vma *page__anon_vma = folio_anon_vma(folio);
755  		/*
756  		 * Note: swapoff's unuse_vma() is more efficient with this
757  		 * check, and needs it to match anon_vma when KSM is active.
758  		 */
759  		if (!vma->anon_vma || !page__anon_vma ||
760  		    vma->anon_vma->root != page__anon_vma->root)
761  			return -EFAULT;
762  	} else if (!vma->vm_file) {
763  		return -EFAULT;
764  	} else if (vma->vm_file->f_mapping != folio->mapping) {
765  		return -EFAULT;
766  	}
767  
768  	return vma_address(page, vma);
769  }
770  
771  /*
772   * Returns the actual pmd_t* where we expect 'address' to be mapped from, or
773   * NULL if it doesn't exist.  No guarantees / checks on what the pmd_t*
774   * represents.
775   */
776  pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
777  {
778  	pgd_t *pgd;
779  	p4d_t *p4d;
780  	pud_t *pud;
781  	pmd_t *pmd = NULL;
782  
783  	pgd = pgd_offset(mm, address);
784  	if (!pgd_present(*pgd))
785  		goto out;
786  
787  	p4d = p4d_offset(pgd, address);
788  	if (!p4d_present(*p4d))
789  		goto out;
790  
791  	pud = pud_offset(p4d, address);
792  	if (!pud_present(*pud))
793  		goto out;
794  
795  	pmd = pmd_offset(pud, address);
796  out:
797  	return pmd;
798  }
799  
800  struct folio_referenced_arg {
801  	int mapcount;
802  	int referenced;
803  	unsigned long vm_flags;
804  	struct mem_cgroup *memcg;
805  };
806  /*
807   * arg: folio_referenced_arg will be passed
808   */
809  static bool folio_referenced_one(struct folio *folio,
810  		struct vm_area_struct *vma, unsigned long address, void *arg)
811  {
812  	struct folio_referenced_arg *pra = arg;
813  	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
814  	int referenced = 0;
815  
816  	while (page_vma_mapped_walk(&pvmw)) {
817  		address = pvmw.address;
818  
819  		if ((vma->vm_flags & VM_LOCKED) &&
820  		    (!folio_test_large(folio) || !pvmw.pte)) {
821  			/* Restore the mlock which got missed */
822  			mlock_vma_folio(folio, vma, !pvmw.pte);
823  			page_vma_mapped_walk_done(&pvmw);
824  			pra->vm_flags |= VM_LOCKED;
825  			return false; /* To break the loop */
826  		}
827  
828  		if (pvmw.pte) {
829  			if (lru_gen_enabled() &&
830  			    pte_young(ptep_get(pvmw.pte))) {
831  				lru_gen_look_around(&pvmw);
832  				referenced++;
833  			}
834  
835  			if (ptep_clear_flush_young_notify(vma, address,
836  						pvmw.pte))
837  				referenced++;
838  		} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
839  			if (pmdp_clear_flush_young_notify(vma, address,
840  						pvmw.pmd))
841  				referenced++;
842  		} else {
843  			/* unexpected pmd-mapped folio? */
844  			WARN_ON_ONCE(1);
845  		}
846  
847  		pra->mapcount--;
848  	}
849  
850  	if (referenced)
851  		folio_clear_idle(folio);
852  	if (folio_test_clear_young(folio))
853  		referenced++;
854  
855  	if (referenced) {
856  		pra->referenced++;
857  		pra->vm_flags |= vma->vm_flags & ~VM_LOCKED;
858  	}
859  
860  	if (!pra->mapcount)
861  		return false; /* To break the loop */
862  
863  	return true;
864  }
865  
866  static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg)
867  {
868  	struct folio_referenced_arg *pra = arg;
869  	struct mem_cgroup *memcg = pra->memcg;
870  
871  	/*
872  	 * Ignore references from this mapping if it has no recency. If the
873  	 * folio has been used in another mapping, we will catch it; if this
874  	 * other mapping is already gone, the unmap path will have set the
875  	 * referenced flag or activated the folio in zap_pte_range().
876  	 */
877  	if (!vma_has_recency(vma))
878  		return true;
879  
880  	/*
881  	 * If we are reclaiming on behalf of a cgroup, skip counting on behalf
882  	 * of references from different cgroups.
883  	 */
884  	if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
885  		return true;
886  
887  	return false;
888  }
889  
890  /**
891   * folio_referenced() - Test if the folio was referenced.
892   * @folio: The folio to test.
893   * @is_locked: Caller holds lock on the folio.
894   * @memcg: target memory cgroup
895   * @vm_flags: A combination of all the vma->vm_flags which referenced the folio.
896   *
897   * Quick test_and_clear_referenced for all mappings of a folio,
898   *
899   * Return: The number of mappings which referenced the folio. Return -1 if
900   * the function bailed out due to rmap lock contention.
901   */
902  int folio_referenced(struct folio *folio, int is_locked,
903  		     struct mem_cgroup *memcg, unsigned long *vm_flags)
904  {
905  	int we_locked = 0;
906  	struct folio_referenced_arg pra = {
907  		.mapcount = folio_mapcount(folio),
908  		.memcg = memcg,
909  	};
910  	struct rmap_walk_control rwc = {
911  		.rmap_one = folio_referenced_one,
912  		.arg = (void *)&pra,
913  		.anon_lock = folio_lock_anon_vma_read,
914  		.try_lock = true,
915  		.invalid_vma = invalid_folio_referenced_vma,
916  	};
917  
918  	*vm_flags = 0;
919  	if (!pra.mapcount)
920  		return 0;
921  
922  	if (!folio_raw_mapping(folio))
923  		return 0;
924  
925  	if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) {
926  		we_locked = folio_trylock(folio);
927  		if (!we_locked)
928  			return 1;
929  	}
930  
931  	rmap_walk(folio, &rwc);
932  	*vm_flags = pra.vm_flags;
933  
934  	if (we_locked)
935  		folio_unlock(folio);
936  
937  	return rwc.contended ? -1 : pra.referenced;
938  }
939  
940  static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw)
941  {
942  	int cleaned = 0;
943  	struct vm_area_struct *vma = pvmw->vma;
944  	struct mmu_notifier_range range;
945  	unsigned long address = pvmw->address;
946  
947  	/*
948  	 * We have to assume the worse case ie pmd for invalidation. Note that
949  	 * the folio can not be freed from this function.
950  	 */
951  	mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 0,
952  				vma->vm_mm, address, vma_address_end(pvmw));
953  	mmu_notifier_invalidate_range_start(&range);
954  
955  	while (page_vma_mapped_walk(pvmw)) {
956  		int ret = 0;
957  
958  		address = pvmw->address;
959  		if (pvmw->pte) {
960  			pte_t *pte = pvmw->pte;
961  			pte_t entry = ptep_get(pte);
962  
963  			if (!pte_dirty(entry) && !pte_write(entry))
964  				continue;
965  
966  			flush_cache_page(vma, address, pte_pfn(entry));
967  			entry = ptep_clear_flush(vma, address, pte);
968  			entry = pte_wrprotect(entry);
969  			entry = pte_mkclean(entry);
970  			set_pte_at(vma->vm_mm, address, pte, entry);
971  			ret = 1;
972  		} else {
973  #ifdef CONFIG_TRANSPARENT_HUGEPAGE
974  			pmd_t *pmd = pvmw->pmd;
975  			pmd_t entry;
976  
977  			if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
978  				continue;
979  
980  			flush_cache_range(vma, address,
981  					  address + HPAGE_PMD_SIZE);
982  			entry = pmdp_invalidate(vma, address, pmd);
983  			entry = pmd_wrprotect(entry);
984  			entry = pmd_mkclean(entry);
985  			set_pmd_at(vma->vm_mm, address, pmd, entry);
986  			ret = 1;
987  #else
988  			/* unexpected pmd-mapped folio? */
989  			WARN_ON_ONCE(1);
990  #endif
991  		}
992  
993  		/*
994  		 * No need to call mmu_notifier_invalidate_range() as we are
995  		 * downgrading page table protection not changing it to point
996  		 * to a new page.
997  		 *
998  		 * See Documentation/mm/mmu_notifier.rst
999  		 */
1000  		if (ret)
1001  			cleaned++;
1002  	}
1003  
1004  	mmu_notifier_invalidate_range_end(&range);
1005  
1006  	return cleaned;
1007  }
1008  
1009  static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma,
1010  			     unsigned long address, void *arg)
1011  {
1012  	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC);
1013  	int *cleaned = arg;
1014  
1015  	*cleaned += page_vma_mkclean_one(&pvmw);
1016  
1017  	return true;
1018  }
1019  
1020  static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1021  {
1022  	if (vma->vm_flags & VM_SHARED)
1023  		return false;
1024  
1025  	return true;
1026  }
1027  
1028  int folio_mkclean(struct folio *folio)
1029  {
1030  	int cleaned = 0;
1031  	struct address_space *mapping;
1032  	struct rmap_walk_control rwc = {
1033  		.arg = (void *)&cleaned,
1034  		.rmap_one = page_mkclean_one,
1035  		.invalid_vma = invalid_mkclean_vma,
1036  	};
1037  
1038  	BUG_ON(!folio_test_locked(folio));
1039  
1040  	if (!folio_mapped(folio))
1041  		return 0;
1042  
1043  	mapping = folio_mapping(folio);
1044  	if (!mapping)
1045  		return 0;
1046  
1047  	rmap_walk(folio, &rwc);
1048  
1049  	return cleaned;
1050  }
1051  EXPORT_SYMBOL_GPL(folio_mkclean);
1052  
1053  /**
1054   * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of
1055   *                     [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff)
1056   *                     within the @vma of shared mappings. And since clean PTEs
1057   *                     should also be readonly, write protects them too.
1058   * @pfn: start pfn.
1059   * @nr_pages: number of physically contiguous pages srarting with @pfn.
1060   * @pgoff: page offset that the @pfn mapped with.
1061   * @vma: vma that @pfn mapped within.
1062   *
1063   * Returns the number of cleaned PTEs (including PMDs).
1064   */
1065  int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff,
1066  		      struct vm_area_struct *vma)
1067  {
1068  	struct page_vma_mapped_walk pvmw = {
1069  		.pfn		= pfn,
1070  		.nr_pages	= nr_pages,
1071  		.pgoff		= pgoff,
1072  		.vma		= vma,
1073  		.flags		= PVMW_SYNC,
1074  	};
1075  
1076  	if (invalid_mkclean_vma(vma, NULL))
1077  		return 0;
1078  
1079  	pvmw.address = vma_pgoff_address(pgoff, nr_pages, vma);
1080  	VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma);
1081  
1082  	return page_vma_mkclean_one(&pvmw);
1083  }
1084  
1085  int folio_total_mapcount(struct folio *folio)
1086  {
1087  	int mapcount = folio_entire_mapcount(folio);
1088  	int nr_pages;
1089  	int i;
1090  
1091  	/* In the common case, avoid the loop when no pages mapped by PTE */
1092  	if (folio_nr_pages_mapped(folio) == 0)
1093  		return mapcount;
1094  	/*
1095  	 * Add all the PTE mappings of those pages mapped by PTE.
1096  	 * Limit the loop to folio_nr_pages_mapped()?
1097  	 * Perhaps: given all the raciness, that may be a good or a bad idea.
1098  	 */
1099  	nr_pages = folio_nr_pages(folio);
1100  	for (i = 0; i < nr_pages; i++)
1101  		mapcount += atomic_read(&folio_page(folio, i)->_mapcount);
1102  
1103  	/* But each of those _mapcounts was based on -1 */
1104  	mapcount += nr_pages;
1105  	return mapcount;
1106  }
1107  
1108  /**
1109   * page_move_anon_rmap - move a page to our anon_vma
1110   * @page:	the page to move to our anon_vma
1111   * @vma:	the vma the page belongs to
1112   *
1113   * When a page belongs exclusively to one process after a COW event,
1114   * that page can be moved into the anon_vma that belongs to just that
1115   * process, so the rmap code will not search the parent or sibling
1116   * processes.
1117   */
1118  void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1119  {
1120  	void *anon_vma = vma->anon_vma;
1121  	struct folio *folio = page_folio(page);
1122  
1123  	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1124  	VM_BUG_ON_VMA(!anon_vma, vma);
1125  
1126  	anon_vma += PAGE_MAPPING_ANON;
1127  	/*
1128  	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1129  	 * simultaneously, so a concurrent reader (eg folio_referenced()'s
1130  	 * folio_test_anon()) will not see one without the other.
1131  	 */
1132  	WRITE_ONCE(folio->mapping, anon_vma);
1133  	SetPageAnonExclusive(page);
1134  }
1135  
1136  /**
1137   * __page_set_anon_rmap - set up new anonymous rmap
1138   * @folio:	Folio which contains page.
1139   * @page:	Page to add to rmap.
1140   * @vma:	VM area to add page to.
1141   * @address:	User virtual address of the mapping
1142   * @exclusive:	the page is exclusively owned by the current process
1143   */
1144  static void __page_set_anon_rmap(struct folio *folio, struct page *page,
1145  	struct vm_area_struct *vma, unsigned long address, int exclusive)
1146  {
1147  	struct anon_vma *anon_vma = vma->anon_vma;
1148  
1149  	BUG_ON(!anon_vma);
1150  
1151  	if (folio_test_anon(folio))
1152  		goto out;
1153  
1154  	/*
1155  	 * If the page isn't exclusively mapped into this vma,
1156  	 * we must use the _oldest_ possible anon_vma for the
1157  	 * page mapping!
1158  	 */
1159  	if (!exclusive)
1160  		anon_vma = anon_vma->root;
1161  
1162  	/*
1163  	 * page_idle does a lockless/optimistic rmap scan on folio->mapping.
1164  	 * Make sure the compiler doesn't split the stores of anon_vma and
1165  	 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1166  	 * could mistake the mapping for a struct address_space and crash.
1167  	 */
1168  	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1169  	WRITE_ONCE(folio->mapping, (struct address_space *) anon_vma);
1170  	folio->index = linear_page_index(vma, address);
1171  out:
1172  	if (exclusive)
1173  		SetPageAnonExclusive(page);
1174  }
1175  
1176  /**
1177   * __page_check_anon_rmap - sanity check anonymous rmap addition
1178   * @page:	the page to add the mapping to
1179   * @vma:	the vm area in which the mapping is added
1180   * @address:	the user virtual address mapped
1181   */
1182  static void __page_check_anon_rmap(struct page *page,
1183  	struct vm_area_struct *vma, unsigned long address)
1184  {
1185  	struct folio *folio = page_folio(page);
1186  	/*
1187  	 * The page's anon-rmap details (mapping and index) are guaranteed to
1188  	 * be set up correctly at this point.
1189  	 *
1190  	 * We have exclusion against page_add_anon_rmap because the caller
1191  	 * always holds the page locked.
1192  	 *
1193  	 * We have exclusion against page_add_new_anon_rmap because those pages
1194  	 * are initially only visible via the pagetables, and the pte is locked
1195  	 * over the call to page_add_new_anon_rmap.
1196  	 */
1197  	VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root,
1198  			folio);
1199  	VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1200  		       page);
1201  }
1202  
1203  /**
1204   * page_add_anon_rmap - add pte mapping to an anonymous page
1205   * @page:	the page to add the mapping to
1206   * @vma:	the vm area in which the mapping is added
1207   * @address:	the user virtual address mapped
1208   * @flags:	the rmap flags
1209   *
1210   * The caller needs to hold the pte lock, and the page must be locked in
1211   * the anon_vma case: to serialize mapping,index checking after setting,
1212   * and to ensure that PageAnon is not being upgraded racily to PageKsm
1213   * (but PageKsm is never downgraded to PageAnon).
1214   */
1215  void page_add_anon_rmap(struct page *page, struct vm_area_struct *vma,
1216  		unsigned long address, rmap_t flags)
1217  {
1218  	struct folio *folio = page_folio(page);
1219  	atomic_t *mapped = &folio->_nr_pages_mapped;
1220  	int nr = 0, nr_pmdmapped = 0;
1221  	bool compound = flags & RMAP_COMPOUND;
1222  	bool first = true;
1223  
1224  	/* Is page being mapped by PTE? Is this its first map to be added? */
1225  	if (likely(!compound)) {
1226  		first = atomic_inc_and_test(&page->_mapcount);
1227  		nr = first;
1228  		if (first && folio_test_large(folio)) {
1229  			nr = atomic_inc_return_relaxed(mapped);
1230  			nr = (nr < COMPOUND_MAPPED);
1231  		}
1232  	} else if (folio_test_pmd_mappable(folio)) {
1233  		/* That test is redundant: it's for safety or to optimize out */
1234  
1235  		first = atomic_inc_and_test(&folio->_entire_mapcount);
1236  		if (first) {
1237  			nr = atomic_add_return_relaxed(COMPOUND_MAPPED, mapped);
1238  			if (likely(nr < COMPOUND_MAPPED + COMPOUND_MAPPED)) {
1239  				nr_pmdmapped = folio_nr_pages(folio);
1240  				nr = nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1241  				/* Raced ahead of a remove and another add? */
1242  				if (unlikely(nr < 0))
1243  					nr = 0;
1244  			} else {
1245  				/* Raced ahead of a remove of COMPOUND_MAPPED */
1246  				nr = 0;
1247  			}
1248  		}
1249  	}
1250  
1251  	VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
1252  	VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
1253  
1254  	if (nr_pmdmapped)
1255  		__lruvec_stat_mod_folio(folio, NR_ANON_THPS, nr_pmdmapped);
1256  	if (nr)
1257  		__lruvec_stat_mod_folio(folio, NR_ANON_MAPPED, nr);
1258  
1259  	if (likely(!folio_test_ksm(folio))) {
1260  		/* address might be in next vma when migration races vma_merge */
1261  		if (first)
1262  			__page_set_anon_rmap(folio, page, vma, address,
1263  					     !!(flags & RMAP_EXCLUSIVE));
1264  		else
1265  			__page_check_anon_rmap(page, vma, address);
1266  	}
1267  
1268  	mlock_vma_folio(folio, vma, compound);
1269  }
1270  
1271  /**
1272   * folio_add_new_anon_rmap - Add mapping to a new anonymous folio.
1273   * @folio:	The folio to add the mapping to.
1274   * @vma:	the vm area in which the mapping is added
1275   * @address:	the user virtual address mapped
1276   *
1277   * Like page_add_anon_rmap() but must only be called on *new* folios.
1278   * This means the inc-and-test can be bypassed.
1279   * The folio does not have to be locked.
1280   *
1281   * If the folio is large, it is accounted as a THP.  As the folio
1282   * is new, it's assumed to be mapped exclusively by a single process.
1283   */
1284  void folio_add_new_anon_rmap(struct folio *folio, struct vm_area_struct *vma,
1285  		unsigned long address)
1286  {
1287  	int nr;
1288  
1289  	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1290  	__folio_set_swapbacked(folio);
1291  
1292  	if (likely(!folio_test_pmd_mappable(folio))) {
1293  		/* increment count (starts at -1) */
1294  		atomic_set(&folio->_mapcount, 0);
1295  		nr = 1;
1296  	} else {
1297  		/* increment count (starts at -1) */
1298  		atomic_set(&folio->_entire_mapcount, 0);
1299  		atomic_set(&folio->_nr_pages_mapped, COMPOUND_MAPPED);
1300  		nr = folio_nr_pages(folio);
1301  		__lruvec_stat_mod_folio(folio, NR_ANON_THPS, nr);
1302  	}
1303  
1304  	__lruvec_stat_mod_folio(folio, NR_ANON_MAPPED, nr);
1305  	__page_set_anon_rmap(folio, &folio->page, vma, address, 1);
1306  }
1307  
1308  /**
1309   * page_add_file_rmap - add pte mapping to a file page
1310   * @page:	the page to add the mapping to
1311   * @vma:	the vm area in which the mapping is added
1312   * @compound:	charge the page as compound or small page
1313   *
1314   * The caller needs to hold the pte lock.
1315   */
1316  void page_add_file_rmap(struct page *page, struct vm_area_struct *vma,
1317  		bool compound)
1318  {
1319  	struct folio *folio = page_folio(page);
1320  	atomic_t *mapped = &folio->_nr_pages_mapped;
1321  	int nr = 0, nr_pmdmapped = 0;
1322  	bool first;
1323  
1324  	VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1325  
1326  	/* Is page being mapped by PTE? Is this its first map to be added? */
1327  	if (likely(!compound)) {
1328  		first = atomic_inc_and_test(&page->_mapcount);
1329  		nr = first;
1330  		if (first && folio_test_large(folio)) {
1331  			nr = atomic_inc_return_relaxed(mapped);
1332  			nr = (nr < COMPOUND_MAPPED);
1333  		}
1334  	} else if (folio_test_pmd_mappable(folio)) {
1335  		/* That test is redundant: it's for safety or to optimize out */
1336  
1337  		first = atomic_inc_and_test(&folio->_entire_mapcount);
1338  		if (first) {
1339  			nr = atomic_add_return_relaxed(COMPOUND_MAPPED, mapped);
1340  			if (likely(nr < COMPOUND_MAPPED + COMPOUND_MAPPED)) {
1341  				nr_pmdmapped = folio_nr_pages(folio);
1342  				nr = nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1343  				/* Raced ahead of a remove and another add? */
1344  				if (unlikely(nr < 0))
1345  					nr = 0;
1346  			} else {
1347  				/* Raced ahead of a remove of COMPOUND_MAPPED */
1348  				nr = 0;
1349  			}
1350  		}
1351  	}
1352  
1353  	if (nr_pmdmapped)
1354  		__lruvec_stat_mod_folio(folio, folio_test_swapbacked(folio) ?
1355  			NR_SHMEM_PMDMAPPED : NR_FILE_PMDMAPPED, nr_pmdmapped);
1356  	if (nr)
1357  		__lruvec_stat_mod_folio(folio, NR_FILE_MAPPED, nr);
1358  
1359  	mlock_vma_folio(folio, vma, compound);
1360  }
1361  
1362  /**
1363   * page_remove_rmap - take down pte mapping from a page
1364   * @page:	page to remove mapping from
1365   * @vma:	the vm area from which the mapping is removed
1366   * @compound:	uncharge the page as compound or small page
1367   *
1368   * The caller needs to hold the pte lock.
1369   */
1370  void page_remove_rmap(struct page *page, struct vm_area_struct *vma,
1371  		bool compound)
1372  {
1373  	struct folio *folio = page_folio(page);
1374  	atomic_t *mapped = &folio->_nr_pages_mapped;
1375  	int nr = 0, nr_pmdmapped = 0;
1376  	bool last;
1377  	enum node_stat_item idx;
1378  
1379  	VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1380  
1381  	/* Hugetlb pages are not counted in NR_*MAPPED */
1382  	if (unlikely(folio_test_hugetlb(folio))) {
1383  		/* hugetlb pages are always mapped with pmds */
1384  		atomic_dec(&folio->_entire_mapcount);
1385  		return;
1386  	}
1387  
1388  	/* Is page being unmapped by PTE? Is this its last map to be removed? */
1389  	if (likely(!compound)) {
1390  		last = atomic_add_negative(-1, &page->_mapcount);
1391  		nr = last;
1392  		if (last && folio_test_large(folio)) {
1393  			nr = atomic_dec_return_relaxed(mapped);
1394  			nr = (nr < COMPOUND_MAPPED);
1395  		}
1396  	} else if (folio_test_pmd_mappable(folio)) {
1397  		/* That test is redundant: it's for safety or to optimize out */
1398  
1399  		last = atomic_add_negative(-1, &folio->_entire_mapcount);
1400  		if (last) {
1401  			nr = atomic_sub_return_relaxed(COMPOUND_MAPPED, mapped);
1402  			if (likely(nr < COMPOUND_MAPPED)) {
1403  				nr_pmdmapped = folio_nr_pages(folio);
1404  				nr = nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1405  				/* Raced ahead of another remove and an add? */
1406  				if (unlikely(nr < 0))
1407  					nr = 0;
1408  			} else {
1409  				/* An add of COMPOUND_MAPPED raced ahead */
1410  				nr = 0;
1411  			}
1412  		}
1413  	}
1414  
1415  	if (nr_pmdmapped) {
1416  		if (folio_test_anon(folio))
1417  			idx = NR_ANON_THPS;
1418  		else if (folio_test_swapbacked(folio))
1419  			idx = NR_SHMEM_PMDMAPPED;
1420  		else
1421  			idx = NR_FILE_PMDMAPPED;
1422  		__lruvec_stat_mod_folio(folio, idx, -nr_pmdmapped);
1423  	}
1424  	if (nr) {
1425  		idx = folio_test_anon(folio) ? NR_ANON_MAPPED : NR_FILE_MAPPED;
1426  		__lruvec_stat_mod_folio(folio, idx, -nr);
1427  
1428  		/*
1429  		 * Queue anon THP for deferred split if at least one
1430  		 * page of the folio is unmapped and at least one page
1431  		 * is still mapped.
1432  		 */
1433  		if (folio_test_pmd_mappable(folio) && folio_test_anon(folio))
1434  			if (!compound || nr < nr_pmdmapped)
1435  				deferred_split_folio(folio);
1436  	}
1437  
1438  	/*
1439  	 * It would be tidy to reset folio_test_anon mapping when fully
1440  	 * unmapped, but that might overwrite a racing page_add_anon_rmap
1441  	 * which increments mapcount after us but sets mapping before us:
1442  	 * so leave the reset to free_pages_prepare, and remember that
1443  	 * it's only reliable while mapped.
1444  	 */
1445  
1446  	munlock_vma_folio(folio, vma, compound);
1447  }
1448  
1449  /*
1450   * @arg: enum ttu_flags will be passed to this argument
1451   */
1452  static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
1453  		     unsigned long address, void *arg)
1454  {
1455  	struct mm_struct *mm = vma->vm_mm;
1456  	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1457  	pte_t pteval;
1458  	struct page *subpage;
1459  	bool anon_exclusive, ret = true;
1460  	struct mmu_notifier_range range;
1461  	enum ttu_flags flags = (enum ttu_flags)(long)arg;
1462  	unsigned long pfn;
1463  
1464  	/*
1465  	 * When racing against e.g. zap_pte_range() on another cpu,
1466  	 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1467  	 * try_to_unmap() may return before page_mapped() has become false,
1468  	 * if page table locking is skipped: use TTU_SYNC to wait for that.
1469  	 */
1470  	if (flags & TTU_SYNC)
1471  		pvmw.flags = PVMW_SYNC;
1472  
1473  	if (flags & TTU_SPLIT_HUGE_PMD)
1474  		split_huge_pmd_address(vma, address, false, folio);
1475  
1476  	/*
1477  	 * For THP, we have to assume the worse case ie pmd for invalidation.
1478  	 * For hugetlb, it could be much worse if we need to do pud
1479  	 * invalidation in the case of pmd sharing.
1480  	 *
1481  	 * Note that the folio can not be freed in this function as call of
1482  	 * try_to_unmap() must hold a reference on the folio.
1483  	 */
1484  	range.end = vma_address_end(&pvmw);
1485  	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1486  				address, range.end);
1487  	if (folio_test_hugetlb(folio)) {
1488  		/*
1489  		 * If sharing is possible, start and end will be adjusted
1490  		 * accordingly.
1491  		 */
1492  		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1493  						     &range.end);
1494  	}
1495  	mmu_notifier_invalidate_range_start(&range);
1496  
1497  	while (page_vma_mapped_walk(&pvmw)) {
1498  		/* Unexpected PMD-mapped THP? */
1499  		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1500  
1501  		/*
1502  		 * If the folio is in an mlock()d vma, we must not swap it out.
1503  		 */
1504  		if (!(flags & TTU_IGNORE_MLOCK) &&
1505  		    (vma->vm_flags & VM_LOCKED)) {
1506  			/* Restore the mlock which got missed */
1507  			mlock_vma_folio(folio, vma, false);
1508  			page_vma_mapped_walk_done(&pvmw);
1509  			ret = false;
1510  			break;
1511  		}
1512  
1513  		pfn = pte_pfn(ptep_get(pvmw.pte));
1514  		subpage = folio_page(folio, pfn - folio_pfn(folio));
1515  		address = pvmw.address;
1516  		anon_exclusive = folio_test_anon(folio) &&
1517  				 PageAnonExclusive(subpage);
1518  
1519  		if (folio_test_hugetlb(folio)) {
1520  			bool anon = folio_test_anon(folio);
1521  
1522  			/*
1523  			 * The try_to_unmap() is only passed a hugetlb page
1524  			 * in the case where the hugetlb page is poisoned.
1525  			 */
1526  			VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage);
1527  			/*
1528  			 * huge_pmd_unshare may unmap an entire PMD page.
1529  			 * There is no way of knowing exactly which PMDs may
1530  			 * be cached for this mm, so we must flush them all.
1531  			 * start/end were already adjusted above to cover this
1532  			 * range.
1533  			 */
1534  			flush_cache_range(vma, range.start, range.end);
1535  
1536  			/*
1537  			 * To call huge_pmd_unshare, i_mmap_rwsem must be
1538  			 * held in write mode.  Caller needs to explicitly
1539  			 * do this outside rmap routines.
1540  			 *
1541  			 * We also must hold hugetlb vma_lock in write mode.
1542  			 * Lock order dictates acquiring vma_lock BEFORE
1543  			 * i_mmap_rwsem.  We can only try lock here and fail
1544  			 * if unsuccessful.
1545  			 */
1546  			if (!anon) {
1547  				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1548  				if (!hugetlb_vma_trylock_write(vma)) {
1549  					page_vma_mapped_walk_done(&pvmw);
1550  					ret = false;
1551  					break;
1552  				}
1553  				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1554  					hugetlb_vma_unlock_write(vma);
1555  					flush_tlb_range(vma,
1556  						range.start, range.end);
1557  					mmu_notifier_invalidate_range(mm,
1558  						range.start, range.end);
1559  					/*
1560  					 * The ref count of the PMD page was
1561  					 * dropped which is part of the way map
1562  					 * counting is done for shared PMDs.
1563  					 * Return 'true' here.  When there is
1564  					 * no other sharing, huge_pmd_unshare
1565  					 * returns false and we will unmap the
1566  					 * actual page and drop map count
1567  					 * to zero.
1568  					 */
1569  					page_vma_mapped_walk_done(&pvmw);
1570  					break;
1571  				}
1572  				hugetlb_vma_unlock_write(vma);
1573  			}
1574  			pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1575  		} else {
1576  			flush_cache_page(vma, address, pfn);
1577  			/* Nuke the page table entry. */
1578  			if (should_defer_flush(mm, flags)) {
1579  				/*
1580  				 * We clear the PTE but do not flush so potentially
1581  				 * a remote CPU could still be writing to the folio.
1582  				 * If the entry was previously clean then the
1583  				 * architecture must guarantee that a clear->dirty
1584  				 * transition on a cached TLB entry is written through
1585  				 * and traps if the PTE is unmapped.
1586  				 */
1587  				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1588  
1589  				set_tlb_ubc_flush_pending(mm, pteval);
1590  			} else {
1591  				pteval = ptep_clear_flush(vma, address, pvmw.pte);
1592  			}
1593  		}
1594  
1595  		/*
1596  		 * Now the pte is cleared. If this pte was uffd-wp armed,
1597  		 * we may want to replace a none pte with a marker pte if
1598  		 * it's file-backed, so we don't lose the tracking info.
1599  		 */
1600  		pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval);
1601  
1602  		/* Set the dirty flag on the folio now the pte is gone. */
1603  		if (pte_dirty(pteval))
1604  			folio_mark_dirty(folio);
1605  
1606  		/* Update high watermark before we lower rss */
1607  		update_hiwater_rss(mm);
1608  
1609  		if (PageHWPoison(subpage) && (flags & TTU_HWPOISON)) {
1610  			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1611  			if (folio_test_hugetlb(folio)) {
1612  				hugetlb_count_sub(folio_nr_pages(folio), mm);
1613  				set_huge_pte_at(mm, address, pvmw.pte, pteval);
1614  			} else {
1615  				dec_mm_counter(mm, mm_counter(&folio->page));
1616  				set_pte_at(mm, address, pvmw.pte, pteval);
1617  			}
1618  
1619  		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1620  			/*
1621  			 * The guest indicated that the page content is of no
1622  			 * interest anymore. Simply discard the pte, vmscan
1623  			 * will take care of the rest.
1624  			 * A future reference will then fault in a new zero
1625  			 * page. When userfaultfd is active, we must not drop
1626  			 * this page though, as its main user (postcopy
1627  			 * migration) will not expect userfaults on already
1628  			 * copied pages.
1629  			 */
1630  			dec_mm_counter(mm, mm_counter(&folio->page));
1631  			/* We have to invalidate as we cleared the pte */
1632  			mmu_notifier_invalidate_range(mm, address,
1633  						      address + PAGE_SIZE);
1634  		} else if (folio_test_anon(folio)) {
1635  			swp_entry_t entry = { .val = page_private(subpage) };
1636  			pte_t swp_pte;
1637  			/*
1638  			 * Store the swap location in the pte.
1639  			 * See handle_pte_fault() ...
1640  			 */
1641  			if (unlikely(folio_test_swapbacked(folio) !=
1642  					folio_test_swapcache(folio))) {
1643  				WARN_ON_ONCE(1);
1644  				ret = false;
1645  				/* We have to invalidate as we cleared the pte */
1646  				mmu_notifier_invalidate_range(mm, address,
1647  							address + PAGE_SIZE);
1648  				page_vma_mapped_walk_done(&pvmw);
1649  				break;
1650  			}
1651  
1652  			/* MADV_FREE page check */
1653  			if (!folio_test_swapbacked(folio)) {
1654  				int ref_count, map_count;
1655  
1656  				/*
1657  				 * Synchronize with gup_pte_range():
1658  				 * - clear PTE; barrier; read refcount
1659  				 * - inc refcount; barrier; read PTE
1660  				 */
1661  				smp_mb();
1662  
1663  				ref_count = folio_ref_count(folio);
1664  				map_count = folio_mapcount(folio);
1665  
1666  				/*
1667  				 * Order reads for page refcount and dirty flag
1668  				 * (see comments in __remove_mapping()).
1669  				 */
1670  				smp_rmb();
1671  
1672  				/*
1673  				 * The only page refs must be one from isolation
1674  				 * plus the rmap(s) (dropped by discard:).
1675  				 */
1676  				if (ref_count == 1 + map_count &&
1677  				    !folio_test_dirty(folio)) {
1678  					/* Invalidate as we cleared the pte */
1679  					mmu_notifier_invalidate_range(mm,
1680  						address, address + PAGE_SIZE);
1681  					dec_mm_counter(mm, MM_ANONPAGES);
1682  					goto discard;
1683  				}
1684  
1685  				/*
1686  				 * If the folio was redirtied, it cannot be
1687  				 * discarded. Remap the page to page table.
1688  				 */
1689  				set_pte_at(mm, address, pvmw.pte, pteval);
1690  				folio_set_swapbacked(folio);
1691  				ret = false;
1692  				page_vma_mapped_walk_done(&pvmw);
1693  				break;
1694  			}
1695  
1696  			if (swap_duplicate(entry) < 0) {
1697  				set_pte_at(mm, address, pvmw.pte, pteval);
1698  				ret = false;
1699  				page_vma_mapped_walk_done(&pvmw);
1700  				break;
1701  			}
1702  			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1703  				swap_free(entry);
1704  				set_pte_at(mm, address, pvmw.pte, pteval);
1705  				ret = false;
1706  				page_vma_mapped_walk_done(&pvmw);
1707  				break;
1708  			}
1709  
1710  			/* See page_try_share_anon_rmap(): clear PTE first. */
1711  			if (anon_exclusive &&
1712  			    page_try_share_anon_rmap(subpage)) {
1713  				swap_free(entry);
1714  				set_pte_at(mm, address, pvmw.pte, pteval);
1715  				ret = false;
1716  				page_vma_mapped_walk_done(&pvmw);
1717  				break;
1718  			}
1719  			if (list_empty(&mm->mmlist)) {
1720  				spin_lock(&mmlist_lock);
1721  				if (list_empty(&mm->mmlist))
1722  					list_add(&mm->mmlist, &init_mm.mmlist);
1723  				spin_unlock(&mmlist_lock);
1724  			}
1725  			dec_mm_counter(mm, MM_ANONPAGES);
1726  			inc_mm_counter(mm, MM_SWAPENTS);
1727  			swp_pte = swp_entry_to_pte(entry);
1728  			if (anon_exclusive)
1729  				swp_pte = pte_swp_mkexclusive(swp_pte);
1730  			if (pte_soft_dirty(pteval))
1731  				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1732  			if (pte_uffd_wp(pteval))
1733  				swp_pte = pte_swp_mkuffd_wp(swp_pte);
1734  			set_pte_at(mm, address, pvmw.pte, swp_pte);
1735  			/* Invalidate as we cleared the pte */
1736  			mmu_notifier_invalidate_range(mm, address,
1737  						      address + PAGE_SIZE);
1738  		} else {
1739  			/*
1740  			 * This is a locked file-backed folio,
1741  			 * so it cannot be removed from the page
1742  			 * cache and replaced by a new folio before
1743  			 * mmu_notifier_invalidate_range_end, so no
1744  			 * concurrent thread might update its page table
1745  			 * to point at a new folio while a device is
1746  			 * still using this folio.
1747  			 *
1748  			 * See Documentation/mm/mmu_notifier.rst
1749  			 */
1750  			dec_mm_counter(mm, mm_counter_file(&folio->page));
1751  		}
1752  discard:
1753  		/*
1754  		 * No need to call mmu_notifier_invalidate_range() it has be
1755  		 * done above for all cases requiring it to happen under page
1756  		 * table lock before mmu_notifier_invalidate_range_end()
1757  		 *
1758  		 * See Documentation/mm/mmu_notifier.rst
1759  		 */
1760  		page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
1761  		if (vma->vm_flags & VM_LOCKED)
1762  			mlock_drain_local();
1763  		folio_put(folio);
1764  	}
1765  
1766  	mmu_notifier_invalidate_range_end(&range);
1767  
1768  	return ret;
1769  }
1770  
1771  static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1772  {
1773  	return vma_is_temporary_stack(vma);
1774  }
1775  
1776  static int folio_not_mapped(struct folio *folio)
1777  {
1778  	return !folio_mapped(folio);
1779  }
1780  
1781  /**
1782   * try_to_unmap - Try to remove all page table mappings to a folio.
1783   * @folio: The folio to unmap.
1784   * @flags: action and flags
1785   *
1786   * Tries to remove all the page table entries which are mapping this
1787   * folio.  It is the caller's responsibility to check if the folio is
1788   * still mapped if needed (use TTU_SYNC to prevent accounting races).
1789   *
1790   * Context: Caller must hold the folio lock.
1791   */
1792  void try_to_unmap(struct folio *folio, enum ttu_flags flags)
1793  {
1794  	struct rmap_walk_control rwc = {
1795  		.rmap_one = try_to_unmap_one,
1796  		.arg = (void *)flags,
1797  		.done = folio_not_mapped,
1798  		.anon_lock = folio_lock_anon_vma_read,
1799  	};
1800  
1801  	if (flags & TTU_RMAP_LOCKED)
1802  		rmap_walk_locked(folio, &rwc);
1803  	else
1804  		rmap_walk(folio, &rwc);
1805  }
1806  
1807  /*
1808   * @arg: enum ttu_flags will be passed to this argument.
1809   *
1810   * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
1811   * containing migration entries.
1812   */
1813  static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
1814  		     unsigned long address, void *arg)
1815  {
1816  	struct mm_struct *mm = vma->vm_mm;
1817  	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1818  	pte_t pteval;
1819  	struct page *subpage;
1820  	bool anon_exclusive, ret = true;
1821  	struct mmu_notifier_range range;
1822  	enum ttu_flags flags = (enum ttu_flags)(long)arg;
1823  	unsigned long pfn;
1824  
1825  	/*
1826  	 * When racing against e.g. zap_pte_range() on another cpu,
1827  	 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1828  	 * try_to_migrate() may return before page_mapped() has become false,
1829  	 * if page table locking is skipped: use TTU_SYNC to wait for that.
1830  	 */
1831  	if (flags & TTU_SYNC)
1832  		pvmw.flags = PVMW_SYNC;
1833  
1834  	/*
1835  	 * unmap_page() in mm/huge_memory.c is the only user of migration with
1836  	 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
1837  	 */
1838  	if (flags & TTU_SPLIT_HUGE_PMD)
1839  		split_huge_pmd_address(vma, address, true, folio);
1840  
1841  	/*
1842  	 * For THP, we have to assume the worse case ie pmd for invalidation.
1843  	 * For hugetlb, it could be much worse if we need to do pud
1844  	 * invalidation in the case of pmd sharing.
1845  	 *
1846  	 * Note that the page can not be free in this function as call of
1847  	 * try_to_unmap() must hold a reference on the page.
1848  	 */
1849  	range.end = vma_address_end(&pvmw);
1850  	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1851  				address, range.end);
1852  	if (folio_test_hugetlb(folio)) {
1853  		/*
1854  		 * If sharing is possible, start and end will be adjusted
1855  		 * accordingly.
1856  		 */
1857  		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1858  						     &range.end);
1859  	}
1860  	mmu_notifier_invalidate_range_start(&range);
1861  
1862  	while (page_vma_mapped_walk(&pvmw)) {
1863  #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1864  		/* PMD-mapped THP migration entry */
1865  		if (!pvmw.pte) {
1866  			subpage = folio_page(folio,
1867  				pmd_pfn(*pvmw.pmd) - folio_pfn(folio));
1868  			VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
1869  					!folio_test_pmd_mappable(folio), folio);
1870  
1871  			if (set_pmd_migration_entry(&pvmw, subpage)) {
1872  				ret = false;
1873  				page_vma_mapped_walk_done(&pvmw);
1874  				break;
1875  			}
1876  			continue;
1877  		}
1878  #endif
1879  
1880  		/* Unexpected PMD-mapped THP? */
1881  		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1882  
1883  		pfn = pte_pfn(ptep_get(pvmw.pte));
1884  
1885  		if (folio_is_zone_device(folio)) {
1886  			/*
1887  			 * Our PTE is a non-present device exclusive entry and
1888  			 * calculating the subpage as for the common case would
1889  			 * result in an invalid pointer.
1890  			 *
1891  			 * Since only PAGE_SIZE pages can currently be
1892  			 * migrated, just set it to page. This will need to be
1893  			 * changed when hugepage migrations to device private
1894  			 * memory are supported.
1895  			 */
1896  			VM_BUG_ON_FOLIO(folio_nr_pages(folio) > 1, folio);
1897  			subpage = &folio->page;
1898  		} else {
1899  			subpage = folio_page(folio, pfn - folio_pfn(folio));
1900  		}
1901  		address = pvmw.address;
1902  		anon_exclusive = folio_test_anon(folio) &&
1903  				 PageAnonExclusive(subpage);
1904  
1905  		if (folio_test_hugetlb(folio)) {
1906  			bool anon = folio_test_anon(folio);
1907  
1908  			/*
1909  			 * huge_pmd_unshare may unmap an entire PMD page.
1910  			 * There is no way of knowing exactly which PMDs may
1911  			 * be cached for this mm, so we must flush them all.
1912  			 * start/end were already adjusted above to cover this
1913  			 * range.
1914  			 */
1915  			flush_cache_range(vma, range.start, range.end);
1916  
1917  			/*
1918  			 * To call huge_pmd_unshare, i_mmap_rwsem must be
1919  			 * held in write mode.  Caller needs to explicitly
1920  			 * do this outside rmap routines.
1921  			 *
1922  			 * We also must hold hugetlb vma_lock in write mode.
1923  			 * Lock order dictates acquiring vma_lock BEFORE
1924  			 * i_mmap_rwsem.  We can only try lock here and
1925  			 * fail if unsuccessful.
1926  			 */
1927  			if (!anon) {
1928  				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1929  				if (!hugetlb_vma_trylock_write(vma)) {
1930  					page_vma_mapped_walk_done(&pvmw);
1931  					ret = false;
1932  					break;
1933  				}
1934  				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1935  					hugetlb_vma_unlock_write(vma);
1936  					flush_tlb_range(vma,
1937  						range.start, range.end);
1938  					mmu_notifier_invalidate_range(mm,
1939  						range.start, range.end);
1940  
1941  					/*
1942  					 * The ref count of the PMD page was
1943  					 * dropped which is part of the way map
1944  					 * counting is done for shared PMDs.
1945  					 * Return 'true' here.  When there is
1946  					 * no other sharing, huge_pmd_unshare
1947  					 * returns false and we will unmap the
1948  					 * actual page and drop map count
1949  					 * to zero.
1950  					 */
1951  					page_vma_mapped_walk_done(&pvmw);
1952  					break;
1953  				}
1954  				hugetlb_vma_unlock_write(vma);
1955  			}
1956  			/* Nuke the hugetlb page table entry */
1957  			pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1958  		} else {
1959  			flush_cache_page(vma, address, pfn);
1960  			/* Nuke the page table entry. */
1961  			if (should_defer_flush(mm, flags)) {
1962  				/*
1963  				 * We clear the PTE but do not flush so potentially
1964  				 * a remote CPU could still be writing to the folio.
1965  				 * If the entry was previously clean then the
1966  				 * architecture must guarantee that a clear->dirty
1967  				 * transition on a cached TLB entry is written through
1968  				 * and traps if the PTE is unmapped.
1969  				 */
1970  				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1971  
1972  				set_tlb_ubc_flush_pending(mm, pteval);
1973  			} else {
1974  				pteval = ptep_clear_flush(vma, address, pvmw.pte);
1975  			}
1976  		}
1977  
1978  		/* Set the dirty flag on the folio now the pte is gone. */
1979  		if (pte_dirty(pteval))
1980  			folio_mark_dirty(folio);
1981  
1982  		/* Update high watermark before we lower rss */
1983  		update_hiwater_rss(mm);
1984  
1985  		if (folio_is_device_private(folio)) {
1986  			unsigned long pfn = folio_pfn(folio);
1987  			swp_entry_t entry;
1988  			pte_t swp_pte;
1989  
1990  			if (anon_exclusive)
1991  				BUG_ON(page_try_share_anon_rmap(subpage));
1992  
1993  			/*
1994  			 * Store the pfn of the page in a special migration
1995  			 * pte. do_swap_page() will wait until the migration
1996  			 * pte is removed and then restart fault handling.
1997  			 */
1998  			entry = pte_to_swp_entry(pteval);
1999  			if (is_writable_device_private_entry(entry))
2000  				entry = make_writable_migration_entry(pfn);
2001  			else if (anon_exclusive)
2002  				entry = make_readable_exclusive_migration_entry(pfn);
2003  			else
2004  				entry = make_readable_migration_entry(pfn);
2005  			swp_pte = swp_entry_to_pte(entry);
2006  
2007  			/*
2008  			 * pteval maps a zone device page and is therefore
2009  			 * a swap pte.
2010  			 */
2011  			if (pte_swp_soft_dirty(pteval))
2012  				swp_pte = pte_swp_mksoft_dirty(swp_pte);
2013  			if (pte_swp_uffd_wp(pteval))
2014  				swp_pte = pte_swp_mkuffd_wp(swp_pte);
2015  			set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
2016  			trace_set_migration_pte(pvmw.address, pte_val(swp_pte),
2017  						compound_order(&folio->page));
2018  			/*
2019  			 * No need to invalidate here it will synchronize on
2020  			 * against the special swap migration pte.
2021  			 */
2022  		} else if (PageHWPoison(subpage)) {
2023  			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
2024  			if (folio_test_hugetlb(folio)) {
2025  				hugetlb_count_sub(folio_nr_pages(folio), mm);
2026  				set_huge_pte_at(mm, address, pvmw.pte, pteval);
2027  			} else {
2028  				dec_mm_counter(mm, mm_counter(&folio->page));
2029  				set_pte_at(mm, address, pvmw.pte, pteval);
2030  			}
2031  
2032  		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
2033  			/*
2034  			 * The guest indicated that the page content is of no
2035  			 * interest anymore. Simply discard the pte, vmscan
2036  			 * will take care of the rest.
2037  			 * A future reference will then fault in a new zero
2038  			 * page. When userfaultfd is active, we must not drop
2039  			 * this page though, as its main user (postcopy
2040  			 * migration) will not expect userfaults on already
2041  			 * copied pages.
2042  			 */
2043  			dec_mm_counter(mm, mm_counter(&folio->page));
2044  			/* We have to invalidate as we cleared the pte */
2045  			mmu_notifier_invalidate_range(mm, address,
2046  						      address + PAGE_SIZE);
2047  		} else {
2048  			swp_entry_t entry;
2049  			pte_t swp_pte;
2050  
2051  			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
2052  				if (folio_test_hugetlb(folio))
2053  					set_huge_pte_at(mm, address, pvmw.pte, pteval);
2054  				else
2055  					set_pte_at(mm, address, pvmw.pte, pteval);
2056  				ret = false;
2057  				page_vma_mapped_walk_done(&pvmw);
2058  				break;
2059  			}
2060  			VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) &&
2061  				       !anon_exclusive, subpage);
2062  
2063  			/* See page_try_share_anon_rmap(): clear PTE first. */
2064  			if (anon_exclusive &&
2065  			    page_try_share_anon_rmap(subpage)) {
2066  				if (folio_test_hugetlb(folio))
2067  					set_huge_pte_at(mm, address, pvmw.pte, pteval);
2068  				else
2069  					set_pte_at(mm, address, pvmw.pte, pteval);
2070  				ret = false;
2071  				page_vma_mapped_walk_done(&pvmw);
2072  				break;
2073  			}
2074  
2075  			/*
2076  			 * Store the pfn of the page in a special migration
2077  			 * pte. do_swap_page() will wait until the migration
2078  			 * pte is removed and then restart fault handling.
2079  			 */
2080  			if (pte_write(pteval))
2081  				entry = make_writable_migration_entry(
2082  							page_to_pfn(subpage));
2083  			else if (anon_exclusive)
2084  				entry = make_readable_exclusive_migration_entry(
2085  							page_to_pfn(subpage));
2086  			else
2087  				entry = make_readable_migration_entry(
2088  							page_to_pfn(subpage));
2089  			if (pte_young(pteval))
2090  				entry = make_migration_entry_young(entry);
2091  			if (pte_dirty(pteval))
2092  				entry = make_migration_entry_dirty(entry);
2093  			swp_pte = swp_entry_to_pte(entry);
2094  			if (pte_soft_dirty(pteval))
2095  				swp_pte = pte_swp_mksoft_dirty(swp_pte);
2096  			if (pte_uffd_wp(pteval))
2097  				swp_pte = pte_swp_mkuffd_wp(swp_pte);
2098  			if (folio_test_hugetlb(folio))
2099  				set_huge_pte_at(mm, address, pvmw.pte, swp_pte);
2100  			else
2101  				set_pte_at(mm, address, pvmw.pte, swp_pte);
2102  			trace_set_migration_pte(address, pte_val(swp_pte),
2103  						compound_order(&folio->page));
2104  			/*
2105  			 * No need to invalidate here it will synchronize on
2106  			 * against the special swap migration pte.
2107  			 */
2108  		}
2109  
2110  		/*
2111  		 * No need to call mmu_notifier_invalidate_range() it has be
2112  		 * done above for all cases requiring it to happen under page
2113  		 * table lock before mmu_notifier_invalidate_range_end()
2114  		 *
2115  		 * See Documentation/mm/mmu_notifier.rst
2116  		 */
2117  		page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
2118  		if (vma->vm_flags & VM_LOCKED)
2119  			mlock_drain_local();
2120  		folio_put(folio);
2121  	}
2122  
2123  	mmu_notifier_invalidate_range_end(&range);
2124  
2125  	return ret;
2126  }
2127  
2128  /**
2129   * try_to_migrate - try to replace all page table mappings with swap entries
2130   * @folio: the folio to replace page table entries for
2131   * @flags: action and flags
2132   *
2133   * Tries to remove all the page table entries which are mapping this folio and
2134   * replace them with special swap entries. Caller must hold the folio lock.
2135   */
2136  void try_to_migrate(struct folio *folio, enum ttu_flags flags)
2137  {
2138  	struct rmap_walk_control rwc = {
2139  		.rmap_one = try_to_migrate_one,
2140  		.arg = (void *)flags,
2141  		.done = folio_not_mapped,
2142  		.anon_lock = folio_lock_anon_vma_read,
2143  	};
2144  
2145  	/*
2146  	 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
2147  	 * TTU_SPLIT_HUGE_PMD, TTU_SYNC, and TTU_BATCH_FLUSH flags.
2148  	 */
2149  	if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2150  					TTU_SYNC | TTU_BATCH_FLUSH)))
2151  		return;
2152  
2153  	if (folio_is_zone_device(folio) &&
2154  	    (!folio_is_device_private(folio) && !folio_is_device_coherent(folio)))
2155  		return;
2156  
2157  	/*
2158  	 * During exec, a temporary VMA is setup and later moved.
2159  	 * The VMA is moved under the anon_vma lock but not the
2160  	 * page tables leading to a race where migration cannot
2161  	 * find the migration ptes. Rather than increasing the
2162  	 * locking requirements of exec(), migration skips
2163  	 * temporary VMAs until after exec() completes.
2164  	 */
2165  	if (!folio_test_ksm(folio) && folio_test_anon(folio))
2166  		rwc.invalid_vma = invalid_migration_vma;
2167  
2168  	if (flags & TTU_RMAP_LOCKED)
2169  		rmap_walk_locked(folio, &rwc);
2170  	else
2171  		rmap_walk(folio, &rwc);
2172  }
2173  
2174  #ifdef CONFIG_DEVICE_PRIVATE
2175  struct make_exclusive_args {
2176  	struct mm_struct *mm;
2177  	unsigned long address;
2178  	void *owner;
2179  	bool valid;
2180  };
2181  
2182  static bool page_make_device_exclusive_one(struct folio *folio,
2183  		struct vm_area_struct *vma, unsigned long address, void *priv)
2184  {
2185  	struct mm_struct *mm = vma->vm_mm;
2186  	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
2187  	struct make_exclusive_args *args = priv;
2188  	pte_t pteval;
2189  	struct page *subpage;
2190  	bool ret = true;
2191  	struct mmu_notifier_range range;
2192  	swp_entry_t entry;
2193  	pte_t swp_pte;
2194  	pte_t ptent;
2195  
2196  	mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
2197  				      vma->vm_mm, address, min(vma->vm_end,
2198  				      address + folio_size(folio)),
2199  				      args->owner);
2200  	mmu_notifier_invalidate_range_start(&range);
2201  
2202  	while (page_vma_mapped_walk(&pvmw)) {
2203  		/* Unexpected PMD-mapped THP? */
2204  		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2205  
2206  		ptent = ptep_get(pvmw.pte);
2207  		if (!pte_present(ptent)) {
2208  			ret = false;
2209  			page_vma_mapped_walk_done(&pvmw);
2210  			break;
2211  		}
2212  
2213  		subpage = folio_page(folio,
2214  				pte_pfn(ptent) - folio_pfn(folio));
2215  		address = pvmw.address;
2216  
2217  		/* Nuke the page table entry. */
2218  		flush_cache_page(vma, address, pte_pfn(ptent));
2219  		pteval = ptep_clear_flush(vma, address, pvmw.pte);
2220  
2221  		/* Set the dirty flag on the folio now the pte is gone. */
2222  		if (pte_dirty(pteval))
2223  			folio_mark_dirty(folio);
2224  
2225  		/*
2226  		 * Check that our target page is still mapped at the expected
2227  		 * address.
2228  		 */
2229  		if (args->mm == mm && args->address == address &&
2230  		    pte_write(pteval))
2231  			args->valid = true;
2232  
2233  		/*
2234  		 * Store the pfn of the page in a special migration
2235  		 * pte. do_swap_page() will wait until the migration
2236  		 * pte is removed and then restart fault handling.
2237  		 */
2238  		if (pte_write(pteval))
2239  			entry = make_writable_device_exclusive_entry(
2240  							page_to_pfn(subpage));
2241  		else
2242  			entry = make_readable_device_exclusive_entry(
2243  							page_to_pfn(subpage));
2244  		swp_pte = swp_entry_to_pte(entry);
2245  		if (pte_soft_dirty(pteval))
2246  			swp_pte = pte_swp_mksoft_dirty(swp_pte);
2247  		if (pte_uffd_wp(pteval))
2248  			swp_pte = pte_swp_mkuffd_wp(swp_pte);
2249  
2250  		set_pte_at(mm, address, pvmw.pte, swp_pte);
2251  
2252  		/*
2253  		 * There is a reference on the page for the swap entry which has
2254  		 * been removed, so shouldn't take another.
2255  		 */
2256  		page_remove_rmap(subpage, vma, false);
2257  	}
2258  
2259  	mmu_notifier_invalidate_range_end(&range);
2260  
2261  	return ret;
2262  }
2263  
2264  /**
2265   * folio_make_device_exclusive - Mark the folio exclusively owned by a device.
2266   * @folio: The folio to replace page table entries for.
2267   * @mm: The mm_struct where the folio is expected to be mapped.
2268   * @address: Address where the folio is expected to be mapped.
2269   * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2270   *
2271   * Tries to remove all the page table entries which are mapping this
2272   * folio and replace them with special device exclusive swap entries to
2273   * grant a device exclusive access to the folio.
2274   *
2275   * Context: Caller must hold the folio lock.
2276   * Return: false if the page is still mapped, or if it could not be unmapped
2277   * from the expected address. Otherwise returns true (success).
2278   */
2279  static bool folio_make_device_exclusive(struct folio *folio,
2280  		struct mm_struct *mm, unsigned long address, void *owner)
2281  {
2282  	struct make_exclusive_args args = {
2283  		.mm = mm,
2284  		.address = address,
2285  		.owner = owner,
2286  		.valid = false,
2287  	};
2288  	struct rmap_walk_control rwc = {
2289  		.rmap_one = page_make_device_exclusive_one,
2290  		.done = folio_not_mapped,
2291  		.anon_lock = folio_lock_anon_vma_read,
2292  		.arg = &args,
2293  	};
2294  
2295  	/*
2296  	 * Restrict to anonymous folios for now to avoid potential writeback
2297  	 * issues.
2298  	 */
2299  	if (!folio_test_anon(folio))
2300  		return false;
2301  
2302  	rmap_walk(folio, &rwc);
2303  
2304  	return args.valid && !folio_mapcount(folio);
2305  }
2306  
2307  /**
2308   * make_device_exclusive_range() - Mark a range for exclusive use by a device
2309   * @mm: mm_struct of associated target process
2310   * @start: start of the region to mark for exclusive device access
2311   * @end: end address of region
2312   * @pages: returns the pages which were successfully marked for exclusive access
2313   * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2314   *
2315   * Returns: number of pages found in the range by GUP. A page is marked for
2316   * exclusive access only if the page pointer is non-NULL.
2317   *
2318   * This function finds ptes mapping page(s) to the given address range, locks
2319   * them and replaces mappings with special swap entries preventing userspace CPU
2320   * access. On fault these entries are replaced with the original mapping after
2321   * calling MMU notifiers.
2322   *
2323   * A driver using this to program access from a device must use a mmu notifier
2324   * critical section to hold a device specific lock during programming. Once
2325   * programming is complete it should drop the page lock and reference after
2326   * which point CPU access to the page will revoke the exclusive access.
2327   */
2328  int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2329  				unsigned long end, struct page **pages,
2330  				void *owner)
2331  {
2332  	long npages = (end - start) >> PAGE_SHIFT;
2333  	long i;
2334  
2335  	npages = get_user_pages_remote(mm, start, npages,
2336  				       FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2337  				       pages, NULL);
2338  	if (npages < 0)
2339  		return npages;
2340  
2341  	for (i = 0; i < npages; i++, start += PAGE_SIZE) {
2342  		struct folio *folio = page_folio(pages[i]);
2343  		if (PageTail(pages[i]) || !folio_trylock(folio)) {
2344  			folio_put(folio);
2345  			pages[i] = NULL;
2346  			continue;
2347  		}
2348  
2349  		if (!folio_make_device_exclusive(folio, mm, start, owner)) {
2350  			folio_unlock(folio);
2351  			folio_put(folio);
2352  			pages[i] = NULL;
2353  		}
2354  	}
2355  
2356  	return npages;
2357  }
2358  EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2359  #endif
2360  
2361  void __put_anon_vma(struct anon_vma *anon_vma)
2362  {
2363  	struct anon_vma *root = anon_vma->root;
2364  
2365  	anon_vma_free(anon_vma);
2366  	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2367  		anon_vma_free(root);
2368  }
2369  
2370  static struct anon_vma *rmap_walk_anon_lock(struct folio *folio,
2371  					    struct rmap_walk_control *rwc)
2372  {
2373  	struct anon_vma *anon_vma;
2374  
2375  	if (rwc->anon_lock)
2376  		return rwc->anon_lock(folio, rwc);
2377  
2378  	/*
2379  	 * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read()
2380  	 * because that depends on page_mapped(); but not all its usages
2381  	 * are holding mmap_lock. Users without mmap_lock are required to
2382  	 * take a reference count to prevent the anon_vma disappearing
2383  	 */
2384  	anon_vma = folio_anon_vma(folio);
2385  	if (!anon_vma)
2386  		return NULL;
2387  
2388  	if (anon_vma_trylock_read(anon_vma))
2389  		goto out;
2390  
2391  	if (rwc->try_lock) {
2392  		anon_vma = NULL;
2393  		rwc->contended = true;
2394  		goto out;
2395  	}
2396  
2397  	anon_vma_lock_read(anon_vma);
2398  out:
2399  	return anon_vma;
2400  }
2401  
2402  /*
2403   * rmap_walk_anon - do something to anonymous page using the object-based
2404   * rmap method
2405   * @page: the page to be handled
2406   * @rwc: control variable according to each walk type
2407   *
2408   * Find all the mappings of a page using the mapping pointer and the vma chains
2409   * contained in the anon_vma struct it points to.
2410   */
2411  static void rmap_walk_anon(struct folio *folio,
2412  		struct rmap_walk_control *rwc, bool locked)
2413  {
2414  	struct anon_vma *anon_vma;
2415  	pgoff_t pgoff_start, pgoff_end;
2416  	struct anon_vma_chain *avc;
2417  
2418  	if (locked) {
2419  		anon_vma = folio_anon_vma(folio);
2420  		/* anon_vma disappear under us? */
2421  		VM_BUG_ON_FOLIO(!anon_vma, folio);
2422  	} else {
2423  		anon_vma = rmap_walk_anon_lock(folio, rwc);
2424  	}
2425  	if (!anon_vma)
2426  		return;
2427  
2428  	pgoff_start = folio_pgoff(folio);
2429  	pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2430  	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2431  			pgoff_start, pgoff_end) {
2432  		struct vm_area_struct *vma = avc->vma;
2433  		unsigned long address = vma_address(&folio->page, vma);
2434  
2435  		VM_BUG_ON_VMA(address == -EFAULT, vma);
2436  		cond_resched();
2437  
2438  		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2439  			continue;
2440  
2441  		if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2442  			break;
2443  		if (rwc->done && rwc->done(folio))
2444  			break;
2445  	}
2446  
2447  	if (!locked)
2448  		anon_vma_unlock_read(anon_vma);
2449  }
2450  
2451  /*
2452   * rmap_walk_file - do something to file page using the object-based rmap method
2453   * @page: the page to be handled
2454   * @rwc: control variable according to each walk type
2455   *
2456   * Find all the mappings of a page using the mapping pointer and the vma chains
2457   * contained in the address_space struct it points to.
2458   */
2459  static void rmap_walk_file(struct folio *folio,
2460  		struct rmap_walk_control *rwc, bool locked)
2461  {
2462  	struct address_space *mapping = folio_mapping(folio);
2463  	pgoff_t pgoff_start, pgoff_end;
2464  	struct vm_area_struct *vma;
2465  
2466  	/*
2467  	 * The page lock not only makes sure that page->mapping cannot
2468  	 * suddenly be NULLified by truncation, it makes sure that the
2469  	 * structure at mapping cannot be freed and reused yet,
2470  	 * so we can safely take mapping->i_mmap_rwsem.
2471  	 */
2472  	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2473  
2474  	if (!mapping)
2475  		return;
2476  
2477  	pgoff_start = folio_pgoff(folio);
2478  	pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2479  	if (!locked) {
2480  		if (i_mmap_trylock_read(mapping))
2481  			goto lookup;
2482  
2483  		if (rwc->try_lock) {
2484  			rwc->contended = true;
2485  			return;
2486  		}
2487  
2488  		i_mmap_lock_read(mapping);
2489  	}
2490  lookup:
2491  	vma_interval_tree_foreach(vma, &mapping->i_mmap,
2492  			pgoff_start, pgoff_end) {
2493  		unsigned long address = vma_address(&folio->page, vma);
2494  
2495  		VM_BUG_ON_VMA(address == -EFAULT, vma);
2496  		cond_resched();
2497  
2498  		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2499  			continue;
2500  
2501  		if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2502  			goto done;
2503  		if (rwc->done && rwc->done(folio))
2504  			goto done;
2505  	}
2506  
2507  done:
2508  	if (!locked)
2509  		i_mmap_unlock_read(mapping);
2510  }
2511  
2512  void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc)
2513  {
2514  	if (unlikely(folio_test_ksm(folio)))
2515  		rmap_walk_ksm(folio, rwc);
2516  	else if (folio_test_anon(folio))
2517  		rmap_walk_anon(folio, rwc, false);
2518  	else
2519  		rmap_walk_file(folio, rwc, false);
2520  }
2521  
2522  /* Like rmap_walk, but caller holds relevant rmap lock */
2523  void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc)
2524  {
2525  	/* no ksm support for now */
2526  	VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio);
2527  	if (folio_test_anon(folio))
2528  		rmap_walk_anon(folio, rwc, true);
2529  	else
2530  		rmap_walk_file(folio, rwc, true);
2531  }
2532  
2533  #ifdef CONFIG_HUGETLB_PAGE
2534  /*
2535   * The following two functions are for anonymous (private mapped) hugepages.
2536   * Unlike common anonymous pages, anonymous hugepages have no accounting code
2537   * and no lru code, because we handle hugepages differently from common pages.
2538   *
2539   * RMAP_COMPOUND is ignored.
2540   */
2541  void hugepage_add_anon_rmap(struct page *page, struct vm_area_struct *vma,
2542  			    unsigned long address, rmap_t flags)
2543  {
2544  	struct folio *folio = page_folio(page);
2545  	struct anon_vma *anon_vma = vma->anon_vma;
2546  	int first;
2547  
2548  	BUG_ON(!folio_test_locked(folio));
2549  	BUG_ON(!anon_vma);
2550  	/* address might be in next vma when migration races vma_merge */
2551  	first = atomic_inc_and_test(&folio->_entire_mapcount);
2552  	VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
2553  	VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
2554  	if (first)
2555  		__page_set_anon_rmap(folio, page, vma, address,
2556  				     !!(flags & RMAP_EXCLUSIVE));
2557  }
2558  
2559  void hugepage_add_new_anon_rmap(struct folio *folio,
2560  			struct vm_area_struct *vma, unsigned long address)
2561  {
2562  	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2563  	/* increment count (starts at -1) */
2564  	atomic_set(&folio->_entire_mapcount, 0);
2565  	folio_clear_hugetlb_restore_reserve(folio);
2566  	__page_set_anon_rmap(folio, &folio->page, vma, address, 1);
2567  }
2568  #endif /* CONFIG_HUGETLB_PAGE */
2569