xref: /openbmc/linux/mm/rmap.c (revision 8a0a6c9af053fb93c0edf4581518c77fd131803e)
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   *           mapping->i_mmap_rwsem
29   *             anon_vma->rwsem
30   *               mm->page_table_lock or pte_lock
31   *                 swap_lock (in swap_duplicate, swap_info_get)
32   *                   mmlist_lock (in mmput, drain_mmlist and others)
33   *                   mapping->private_lock (in block_dirty_folio)
34   *                     folio_lock_memcg move_lock (in block_dirty_folio)
35   *                       i_pages lock (widely used)
36   *                         lruvec->lru_lock (in folio_lruvec_lock_irq)
37   *                   inode->i_lock (in set_page_dirty's __mark_inode_dirty)
38   *                   bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
39   *                     sb_lock (within inode_lock in fs/fs-writeback.c)
40   *                     i_pages lock (widely used, in set_page_dirty,
41   *                               in arch-dependent flush_dcache_mmap_lock,
42   *                               within bdi.wb->list_lock in __sync_single_inode)
43   *
44   * anon_vma->rwsem,mapping->i_mmap_rwsem   (memory_failure, collect_procs_anon)
45   *   ->tasklist_lock
46   *     pte map lock
47   *
48   * hugetlbfs PageHuge() take locks in this order:
49   *   hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
50   *     vma_lock (hugetlb specific lock for pmd_sharing)
51   *       mapping->i_mmap_rwsem (also used for hugetlb pmd sharing)
52   *         page->flags PG_locked (lock_page)
53   */
54  
55  #include <linux/mm.h>
56  #include <linux/sched/mm.h>
57  #include <linux/sched/task.h>
58  #include <linux/pagemap.h>
59  #include <linux/swap.h>
60  #include <linux/swapops.h>
61  #include <linux/slab.h>
62  #include <linux/init.h>
63  #include <linux/ksm.h>
64  #include <linux/rmap.h>
65  #include <linux/rcupdate.h>
66  #include <linux/export.h>
67  #include <linux/memcontrol.h>
68  #include <linux/mmu_notifier.h>
69  #include <linux/migrate.h>
70  #include <linux/hugetlb.h>
71  #include <linux/huge_mm.h>
72  #include <linux/backing-dev.h>
73  #include <linux/page_idle.h>
74  #include <linux/memremap.h>
75  #include <linux/userfaultfd_k.h>
76  #include <linux/mm_inline.h>
77  
78  #include <asm/tlbflush.h>
79  
80  #define CREATE_TRACE_POINTS
81  #include <trace/events/tlb.h>
82  #include <trace/events/migrate.h>
83  
84  #include "internal.h"
85  
86  static struct kmem_cache *anon_vma_cachep;
87  static struct kmem_cache *anon_vma_chain_cachep;
88  
89  static inline struct anon_vma *anon_vma_alloc(void)
90  {
91  	struct anon_vma *anon_vma;
92  
93  	anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
94  	if (anon_vma) {
95  		atomic_set(&anon_vma->refcount, 1);
96  		anon_vma->num_children = 0;
97  		anon_vma->num_active_vmas = 0;
98  		anon_vma->parent = anon_vma;
99  		/*
100  		 * Initialise the anon_vma root to point to itself. If called
101  		 * from fork, the root will be reset to the parents anon_vma.
102  		 */
103  		anon_vma->root = anon_vma;
104  	}
105  
106  	return anon_vma;
107  }
108  
109  static inline void anon_vma_free(struct anon_vma *anon_vma)
110  {
111  	VM_BUG_ON(atomic_read(&anon_vma->refcount));
112  
113  	/*
114  	 * Synchronize against folio_lock_anon_vma_read() such that
115  	 * we can safely hold the lock without the anon_vma getting
116  	 * freed.
117  	 *
118  	 * Relies on the full mb implied by the atomic_dec_and_test() from
119  	 * put_anon_vma() against the acquire barrier implied by
120  	 * down_read_trylock() from folio_lock_anon_vma_read(). This orders:
121  	 *
122  	 * folio_lock_anon_vma_read()	VS	put_anon_vma()
123  	 *   down_read_trylock()		  atomic_dec_and_test()
124  	 *   LOCK				  MB
125  	 *   atomic_read()			  rwsem_is_locked()
126  	 *
127  	 * LOCK should suffice since the actual taking of the lock must
128  	 * happen _before_ what follows.
129  	 */
130  	might_sleep();
131  	if (rwsem_is_locked(&anon_vma->root->rwsem)) {
132  		anon_vma_lock_write(anon_vma);
133  		anon_vma_unlock_write(anon_vma);
134  	}
135  
136  	kmem_cache_free(anon_vma_cachep, anon_vma);
137  }
138  
139  static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
140  {
141  	return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
142  }
143  
144  static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
145  {
146  	kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
147  }
148  
149  static void anon_vma_chain_link(struct vm_area_struct *vma,
150  				struct anon_vma_chain *avc,
151  				struct anon_vma *anon_vma)
152  {
153  	avc->vma = vma;
154  	avc->anon_vma = anon_vma;
155  	list_add(&avc->same_vma, &vma->anon_vma_chain);
156  	anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
157  }
158  
159  /**
160   * __anon_vma_prepare - attach an anon_vma to a memory region
161   * @vma: the memory region in question
162   *
163   * This makes sure the memory mapping described by 'vma' has
164   * an 'anon_vma' attached to it, so that we can associate the
165   * anonymous pages mapped into it with that anon_vma.
166   *
167   * The common case will be that we already have one, which
168   * is handled inline by anon_vma_prepare(). But if
169   * not we either need to find an adjacent mapping that we
170   * can re-use the anon_vma from (very common when the only
171   * reason for splitting a vma has been mprotect()), or we
172   * allocate a new one.
173   *
174   * Anon-vma allocations are very subtle, because we may have
175   * optimistically looked up an anon_vma in folio_lock_anon_vma_read()
176   * and that may actually touch the rwsem even in the newly
177   * allocated vma (it depends on RCU to make sure that the
178   * anon_vma isn't actually destroyed).
179   *
180   * As a result, we need to do proper anon_vma locking even
181   * for the new allocation. At the same time, we do not want
182   * to do any locking for the common case of already having
183   * an anon_vma.
184   *
185   * This must be called with the mmap_lock held for reading.
186   */
187  int __anon_vma_prepare(struct vm_area_struct *vma)
188  {
189  	struct mm_struct *mm = vma->vm_mm;
190  	struct anon_vma *anon_vma, *allocated;
191  	struct anon_vma_chain *avc;
192  
193  	might_sleep();
194  
195  	avc = anon_vma_chain_alloc(GFP_KERNEL);
196  	if (!avc)
197  		goto out_enomem;
198  
199  	anon_vma = find_mergeable_anon_vma(vma);
200  	allocated = NULL;
201  	if (!anon_vma) {
202  		anon_vma = anon_vma_alloc();
203  		if (unlikely(!anon_vma))
204  			goto out_enomem_free_avc;
205  		anon_vma->num_children++; /* self-parent link for new root */
206  		allocated = anon_vma;
207  	}
208  
209  	anon_vma_lock_write(anon_vma);
210  	/* page_table_lock to protect against threads */
211  	spin_lock(&mm->page_table_lock);
212  	if (likely(!vma->anon_vma)) {
213  		vma->anon_vma = anon_vma;
214  		anon_vma_chain_link(vma, avc, anon_vma);
215  		anon_vma->num_active_vmas++;
216  		allocated = NULL;
217  		avc = NULL;
218  	}
219  	spin_unlock(&mm->page_table_lock);
220  	anon_vma_unlock_write(anon_vma);
221  
222  	if (unlikely(allocated))
223  		put_anon_vma(allocated);
224  	if (unlikely(avc))
225  		anon_vma_chain_free(avc);
226  
227  	return 0;
228  
229   out_enomem_free_avc:
230  	anon_vma_chain_free(avc);
231   out_enomem:
232  	return -ENOMEM;
233  }
234  
235  /*
236   * This is a useful helper function for locking the anon_vma root as
237   * we traverse the vma->anon_vma_chain, looping over anon_vma's that
238   * have the same vma.
239   *
240   * Such anon_vma's should have the same root, so you'd expect to see
241   * just a single mutex_lock for the whole traversal.
242   */
243  static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
244  {
245  	struct anon_vma *new_root = anon_vma->root;
246  	if (new_root != root) {
247  		if (WARN_ON_ONCE(root))
248  			up_write(&root->rwsem);
249  		root = new_root;
250  		down_write(&root->rwsem);
251  	}
252  	return root;
253  }
254  
255  static inline void unlock_anon_vma_root(struct anon_vma *root)
256  {
257  	if (root)
258  		up_write(&root->rwsem);
259  }
260  
261  /*
262   * Attach the anon_vmas from src to dst.
263   * Returns 0 on success, -ENOMEM on failure.
264   *
265   * anon_vma_clone() is called by vma_expand(), vma_merge(), __split_vma(),
266   * copy_vma() and anon_vma_fork(). The first four want an exact copy of src,
267   * while the last one, anon_vma_fork(), may try to reuse an existing anon_vma to
268   * prevent endless growth of anon_vma. Since dst->anon_vma is set to NULL before
269   * call, we can identify this case by checking (!dst->anon_vma &&
270   * src->anon_vma).
271   *
272   * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
273   * and reuse existing anon_vma which has no vmas and only one child anon_vma.
274   * This prevents degradation of anon_vma hierarchy to endless linear chain in
275   * case of constantly forking task. On the other hand, an anon_vma with more
276   * than one child isn't reused even if there was no alive vma, thus rmap
277   * walker has a good chance of avoiding scanning the whole hierarchy when it
278   * searches where page is mapped.
279   */
280  int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
281  {
282  	struct anon_vma_chain *avc, *pavc;
283  	struct anon_vma *root = NULL;
284  
285  	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
286  		struct anon_vma *anon_vma;
287  
288  		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
289  		if (unlikely(!avc)) {
290  			unlock_anon_vma_root(root);
291  			root = NULL;
292  			avc = anon_vma_chain_alloc(GFP_KERNEL);
293  			if (!avc)
294  				goto enomem_failure;
295  		}
296  		anon_vma = pavc->anon_vma;
297  		root = lock_anon_vma_root(root, anon_vma);
298  		anon_vma_chain_link(dst, avc, anon_vma);
299  
300  		/*
301  		 * Reuse existing anon_vma if it has no vma and only one
302  		 * anon_vma child.
303  		 *
304  		 * Root anon_vma is never reused:
305  		 * it has self-parent reference and at least one child.
306  		 */
307  		if (!dst->anon_vma && src->anon_vma &&
308  		    anon_vma->num_children < 2 &&
309  		    anon_vma->num_active_vmas == 0)
310  			dst->anon_vma = anon_vma;
311  	}
312  	if (dst->anon_vma)
313  		dst->anon_vma->num_active_vmas++;
314  	unlock_anon_vma_root(root);
315  	return 0;
316  
317   enomem_failure:
318  	/*
319  	 * dst->anon_vma is dropped here otherwise its num_active_vmas can
320  	 * be incorrectly decremented in unlink_anon_vmas().
321  	 * We can safely do this because callers of anon_vma_clone() don't care
322  	 * about dst->anon_vma if anon_vma_clone() failed.
323  	 */
324  	dst->anon_vma = NULL;
325  	unlink_anon_vmas(dst);
326  	return -ENOMEM;
327  }
328  
329  /*
330   * Attach vma to its own anon_vma, as well as to the anon_vmas that
331   * the corresponding VMA in the parent process is attached to.
332   * Returns 0 on success, non-zero on failure.
333   */
334  int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
335  {
336  	struct anon_vma_chain *avc;
337  	struct anon_vma *anon_vma;
338  	int error;
339  
340  	/* Don't bother if the parent process has no anon_vma here. */
341  	if (!pvma->anon_vma)
342  		return 0;
343  
344  	/* Drop inherited anon_vma, we'll reuse existing or allocate new. */
345  	vma->anon_vma = NULL;
346  
347  	/*
348  	 * First, attach the new VMA to the parent VMA's anon_vmas,
349  	 * so rmap can find non-COWed pages in child processes.
350  	 */
351  	error = anon_vma_clone(vma, pvma);
352  	if (error)
353  		return error;
354  
355  	/* An existing anon_vma has been reused, all done then. */
356  	if (vma->anon_vma)
357  		return 0;
358  
359  	/* Then add our own anon_vma. */
360  	anon_vma = anon_vma_alloc();
361  	if (!anon_vma)
362  		goto out_error;
363  	anon_vma->num_active_vmas++;
364  	avc = anon_vma_chain_alloc(GFP_KERNEL);
365  	if (!avc)
366  		goto out_error_free_anon_vma;
367  
368  	/*
369  	 * The root anon_vma's rwsem is the lock actually used when we
370  	 * lock any of the anon_vmas in this anon_vma tree.
371  	 */
372  	anon_vma->root = pvma->anon_vma->root;
373  	anon_vma->parent = pvma->anon_vma;
374  	/*
375  	 * With refcounts, an anon_vma can stay around longer than the
376  	 * process it belongs to. The root anon_vma needs to be pinned until
377  	 * this anon_vma is freed, because the lock lives in the root.
378  	 */
379  	get_anon_vma(anon_vma->root);
380  	/* Mark this anon_vma as the one where our new (COWed) pages go. */
381  	vma->anon_vma = anon_vma;
382  	anon_vma_lock_write(anon_vma);
383  	anon_vma_chain_link(vma, avc, anon_vma);
384  	anon_vma->parent->num_children++;
385  	anon_vma_unlock_write(anon_vma);
386  
387  	return 0;
388  
389   out_error_free_anon_vma:
390  	put_anon_vma(anon_vma);
391   out_error:
392  	unlink_anon_vmas(vma);
393  	return -ENOMEM;
394  }
395  
396  void unlink_anon_vmas(struct vm_area_struct *vma)
397  {
398  	struct anon_vma_chain *avc, *next;
399  	struct anon_vma *root = NULL;
400  
401  	/*
402  	 * Unlink each anon_vma chained to the VMA.  This list is ordered
403  	 * from newest to oldest, ensuring the root anon_vma gets freed last.
404  	 */
405  	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
406  		struct anon_vma *anon_vma = avc->anon_vma;
407  
408  		root = lock_anon_vma_root(root, anon_vma);
409  		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
410  
411  		/*
412  		 * Leave empty anon_vmas on the list - we'll need
413  		 * to free them outside the lock.
414  		 */
415  		if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
416  			anon_vma->parent->num_children--;
417  			continue;
418  		}
419  
420  		list_del(&avc->same_vma);
421  		anon_vma_chain_free(avc);
422  	}
423  	if (vma->anon_vma) {
424  		vma->anon_vma->num_active_vmas--;
425  
426  		/*
427  		 * vma would still be needed after unlink, and anon_vma will be prepared
428  		 * when handle fault.
429  		 */
430  		vma->anon_vma = NULL;
431  	}
432  	unlock_anon_vma_root(root);
433  
434  	/*
435  	 * Iterate the list once more, it now only contains empty and unlinked
436  	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
437  	 * needing to write-acquire the anon_vma->root->rwsem.
438  	 */
439  	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
440  		struct anon_vma *anon_vma = avc->anon_vma;
441  
442  		VM_WARN_ON(anon_vma->num_children);
443  		VM_WARN_ON(anon_vma->num_active_vmas);
444  		put_anon_vma(anon_vma);
445  
446  		list_del(&avc->same_vma);
447  		anon_vma_chain_free(avc);
448  	}
449  }
450  
451  static void anon_vma_ctor(void *data)
452  {
453  	struct anon_vma *anon_vma = data;
454  
455  	init_rwsem(&anon_vma->rwsem);
456  	atomic_set(&anon_vma->refcount, 0);
457  	anon_vma->rb_root = RB_ROOT_CACHED;
458  }
459  
460  void __init anon_vma_init(void)
461  {
462  	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
463  			0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
464  			anon_vma_ctor);
465  	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
466  			SLAB_PANIC|SLAB_ACCOUNT);
467  }
468  
469  /*
470   * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
471   *
472   * Since there is no serialization what so ever against page_remove_rmap()
473   * the best this function can do is return a refcount increased anon_vma
474   * that might have been relevant to this page.
475   *
476   * The page might have been remapped to a different anon_vma or the anon_vma
477   * returned may already be freed (and even reused).
478   *
479   * In case it was remapped to a different anon_vma, the new anon_vma will be a
480   * child of the old anon_vma, and the anon_vma lifetime rules will therefore
481   * ensure that any anon_vma obtained from the page will still be valid for as
482   * long as we observe page_mapped() [ hence all those page_mapped() tests ].
483   *
484   * All users of this function must be very careful when walking the anon_vma
485   * chain and verify that the page in question is indeed mapped in it
486   * [ something equivalent to page_mapped_in_vma() ].
487   *
488   * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
489   * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
490   * if there is a mapcount, we can dereference the anon_vma after observing
491   * those.
492   */
493  struct anon_vma *folio_get_anon_vma(struct folio *folio)
494  {
495  	struct anon_vma *anon_vma = NULL;
496  	unsigned long anon_mapping;
497  
498  	rcu_read_lock();
499  	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
500  	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
501  		goto out;
502  	if (!folio_mapped(folio))
503  		goto out;
504  
505  	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
506  	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
507  		anon_vma = NULL;
508  		goto out;
509  	}
510  
511  	/*
512  	 * If this folio is still mapped, then its anon_vma cannot have been
513  	 * freed.  But if it has been unmapped, we have no security against the
514  	 * anon_vma structure being freed and reused (for another anon_vma:
515  	 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
516  	 * above cannot corrupt).
517  	 */
518  	if (!folio_mapped(folio)) {
519  		rcu_read_unlock();
520  		put_anon_vma(anon_vma);
521  		return NULL;
522  	}
523  out:
524  	rcu_read_unlock();
525  
526  	return anon_vma;
527  }
528  
529  /*
530   * Similar to folio_get_anon_vma() except it locks the anon_vma.
531   *
532   * Its a little more complex as it tries to keep the fast path to a single
533   * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
534   * reference like with folio_get_anon_vma() and then block on the mutex
535   * on !rwc->try_lock case.
536   */
537  struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
538  					  struct rmap_walk_control *rwc)
539  {
540  	struct anon_vma *anon_vma = NULL;
541  	struct anon_vma *root_anon_vma;
542  	unsigned long anon_mapping;
543  
544  	rcu_read_lock();
545  	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
546  	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
547  		goto out;
548  	if (!folio_mapped(folio))
549  		goto out;
550  
551  	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
552  	root_anon_vma = READ_ONCE(anon_vma->root);
553  	if (down_read_trylock(&root_anon_vma->rwsem)) {
554  		/*
555  		 * If the folio is still mapped, then this anon_vma is still
556  		 * its anon_vma, and holding the mutex ensures that it will
557  		 * not go away, see anon_vma_free().
558  		 */
559  		if (!folio_mapped(folio)) {
560  			up_read(&root_anon_vma->rwsem);
561  			anon_vma = NULL;
562  		}
563  		goto out;
564  	}
565  
566  	if (rwc && rwc->try_lock) {
567  		anon_vma = NULL;
568  		rwc->contended = true;
569  		goto out;
570  	}
571  
572  	/* trylock failed, we got to sleep */
573  	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
574  		anon_vma = NULL;
575  		goto out;
576  	}
577  
578  	if (!folio_mapped(folio)) {
579  		rcu_read_unlock();
580  		put_anon_vma(anon_vma);
581  		return NULL;
582  	}
583  
584  	/* we pinned the anon_vma, its safe to sleep */
585  	rcu_read_unlock();
586  	anon_vma_lock_read(anon_vma);
587  
588  	if (atomic_dec_and_test(&anon_vma->refcount)) {
589  		/*
590  		 * Oops, we held the last refcount, release the lock
591  		 * and bail -- can't simply use put_anon_vma() because
592  		 * we'll deadlock on the anon_vma_lock_write() recursion.
593  		 */
594  		anon_vma_unlock_read(anon_vma);
595  		__put_anon_vma(anon_vma);
596  		anon_vma = NULL;
597  	}
598  
599  	return anon_vma;
600  
601  out:
602  	rcu_read_unlock();
603  	return anon_vma;
604  }
605  
606  #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
607  /*
608   * Flush TLB entries for recently unmapped pages from remote CPUs. It is
609   * important if a PTE was dirty when it was unmapped that it's flushed
610   * before any IO is initiated on the page to prevent lost writes. Similarly,
611   * it must be flushed before freeing to prevent data leakage.
612   */
613  void try_to_unmap_flush(void)
614  {
615  	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
616  
617  	if (!tlb_ubc->flush_required)
618  		return;
619  
620  	arch_tlbbatch_flush(&tlb_ubc->arch);
621  	tlb_ubc->flush_required = false;
622  	tlb_ubc->writable = false;
623  }
624  
625  /* Flush iff there are potentially writable TLB entries that can race with IO */
626  void try_to_unmap_flush_dirty(void)
627  {
628  	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
629  
630  	if (tlb_ubc->writable)
631  		try_to_unmap_flush();
632  }
633  
634  /*
635   * Bits 0-14 of mm->tlb_flush_batched record pending generations.
636   * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations.
637   */
638  #define TLB_FLUSH_BATCH_FLUSHED_SHIFT	16
639  #define TLB_FLUSH_BATCH_PENDING_MASK			\
640  	((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1)
641  #define TLB_FLUSH_BATCH_PENDING_LARGE			\
642  	(TLB_FLUSH_BATCH_PENDING_MASK / 2)
643  
644  static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
645  {
646  	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
647  	int batch, nbatch;
648  
649  	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
650  	tlb_ubc->flush_required = true;
651  
652  	/*
653  	 * Ensure compiler does not re-order the setting of tlb_flush_batched
654  	 * before the PTE is cleared.
655  	 */
656  	barrier();
657  	batch = atomic_read(&mm->tlb_flush_batched);
658  retry:
659  	if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) {
660  		/*
661  		 * Prevent `pending' from catching up with `flushed' because of
662  		 * overflow.  Reset `pending' and `flushed' to be 1 and 0 if
663  		 * `pending' becomes large.
664  		 */
665  		nbatch = atomic_cmpxchg(&mm->tlb_flush_batched, batch, 1);
666  		if (nbatch != batch) {
667  			batch = nbatch;
668  			goto retry;
669  		}
670  	} else {
671  		atomic_inc(&mm->tlb_flush_batched);
672  	}
673  
674  	/*
675  	 * If the PTE was dirty then it's best to assume it's writable. The
676  	 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
677  	 * before the page is queued for IO.
678  	 */
679  	if (writable)
680  		tlb_ubc->writable = true;
681  }
682  
683  /*
684   * Returns true if the TLB flush should be deferred to the end of a batch of
685   * unmap operations to reduce IPIs.
686   */
687  static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
688  {
689  	bool should_defer = false;
690  
691  	if (!(flags & TTU_BATCH_FLUSH))
692  		return false;
693  
694  	/* If remote CPUs need to be flushed then defer batch the flush */
695  	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
696  		should_defer = true;
697  	put_cpu();
698  
699  	return should_defer;
700  }
701  
702  /*
703   * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
704   * releasing the PTL if TLB flushes are batched. It's possible for a parallel
705   * operation such as mprotect or munmap to race between reclaim unmapping
706   * the page and flushing the page. If this race occurs, it potentially allows
707   * access to data via a stale TLB entry. Tracking all mm's that have TLB
708   * batching in flight would be expensive during reclaim so instead track
709   * whether TLB batching occurred in the past and if so then do a flush here
710   * if required. This will cost one additional flush per reclaim cycle paid
711   * by the first operation at risk such as mprotect and mumap.
712   *
713   * This must be called under the PTL so that an access to tlb_flush_batched
714   * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
715   * via the PTL.
716   */
717  void flush_tlb_batched_pending(struct mm_struct *mm)
718  {
719  	int batch = atomic_read(&mm->tlb_flush_batched);
720  	int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK;
721  	int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT;
722  
723  	if (pending != flushed) {
724  		flush_tlb_mm(mm);
725  		/*
726  		 * If the new TLB flushing is pending during flushing, leave
727  		 * mm->tlb_flush_batched as is, to avoid losing flushing.
728  		 */
729  		atomic_cmpxchg(&mm->tlb_flush_batched, batch,
730  			       pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT));
731  	}
732  }
733  #else
734  static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
735  {
736  }
737  
738  static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
739  {
740  	return false;
741  }
742  #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
743  
744  /*
745   * At what user virtual address is page expected in vma?
746   * Caller should check the page is actually part of the vma.
747   */
748  unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
749  {
750  	struct folio *folio = page_folio(page);
751  	if (folio_test_anon(folio)) {
752  		struct anon_vma *page__anon_vma = folio_anon_vma(folio);
753  		/*
754  		 * Note: swapoff's unuse_vma() is more efficient with this
755  		 * check, and needs it to match anon_vma when KSM is active.
756  		 */
757  		if (!vma->anon_vma || !page__anon_vma ||
758  		    vma->anon_vma->root != page__anon_vma->root)
759  			return -EFAULT;
760  	} else if (!vma->vm_file) {
761  		return -EFAULT;
762  	} else if (vma->vm_file->f_mapping != folio->mapping) {
763  		return -EFAULT;
764  	}
765  
766  	return vma_address(page, vma);
767  }
768  
769  /*
770   * Returns the actual pmd_t* where we expect 'address' to be mapped from, or
771   * NULL if it doesn't exist.  No guarantees / checks on what the pmd_t*
772   * represents.
773   */
774  pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
775  {
776  	pgd_t *pgd;
777  	p4d_t *p4d;
778  	pud_t *pud;
779  	pmd_t *pmd = NULL;
780  
781  	pgd = pgd_offset(mm, address);
782  	if (!pgd_present(*pgd))
783  		goto out;
784  
785  	p4d = p4d_offset(pgd, address);
786  	if (!p4d_present(*p4d))
787  		goto out;
788  
789  	pud = pud_offset(p4d, address);
790  	if (!pud_present(*pud))
791  		goto out;
792  
793  	pmd = pmd_offset(pud, address);
794  out:
795  	return pmd;
796  }
797  
798  struct folio_referenced_arg {
799  	int mapcount;
800  	int referenced;
801  	unsigned long vm_flags;
802  	struct mem_cgroup *memcg;
803  };
804  /*
805   * arg: folio_referenced_arg will be passed
806   */
807  static bool folio_referenced_one(struct folio *folio,
808  		struct vm_area_struct *vma, unsigned long address, void *arg)
809  {
810  	struct folio_referenced_arg *pra = arg;
811  	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
812  	int referenced = 0;
813  
814  	while (page_vma_mapped_walk(&pvmw)) {
815  		address = pvmw.address;
816  
817  		if ((vma->vm_flags & VM_LOCKED) &&
818  		    (!folio_test_large(folio) || !pvmw.pte)) {
819  			/* Restore the mlock which got missed */
820  			mlock_vma_folio(folio, vma, !pvmw.pte);
821  			page_vma_mapped_walk_done(&pvmw);
822  			pra->vm_flags |= VM_LOCKED;
823  			return false; /* To break the loop */
824  		}
825  
826  		if (pvmw.pte) {
827  			if (lru_gen_enabled() && pte_young(*pvmw.pte)) {
828  				lru_gen_look_around(&pvmw);
829  				referenced++;
830  			}
831  
832  			if (ptep_clear_flush_young_notify(vma, address,
833  						pvmw.pte))
834  				referenced++;
835  		} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
836  			if (pmdp_clear_flush_young_notify(vma, address,
837  						pvmw.pmd))
838  				referenced++;
839  		} else {
840  			/* unexpected pmd-mapped folio? */
841  			WARN_ON_ONCE(1);
842  		}
843  
844  		pra->mapcount--;
845  	}
846  
847  	if (referenced)
848  		folio_clear_idle(folio);
849  	if (folio_test_clear_young(folio))
850  		referenced++;
851  
852  	if (referenced) {
853  		pra->referenced++;
854  		pra->vm_flags |= vma->vm_flags & ~VM_LOCKED;
855  	}
856  
857  	if (!pra->mapcount)
858  		return false; /* To break the loop */
859  
860  	return true;
861  }
862  
863  static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg)
864  {
865  	struct folio_referenced_arg *pra = arg;
866  	struct mem_cgroup *memcg = pra->memcg;
867  
868  	/*
869  	 * Ignore references from this mapping if it has no recency. If the
870  	 * folio has been used in another mapping, we will catch it; if this
871  	 * other mapping is already gone, the unmap path will have set the
872  	 * referenced flag or activated the folio in zap_pte_range().
873  	 */
874  	if (!vma_has_recency(vma))
875  		return true;
876  
877  	/*
878  	 * If we are reclaiming on behalf of a cgroup, skip counting on behalf
879  	 * of references from different cgroups.
880  	 */
881  	if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
882  		return true;
883  
884  	return false;
885  }
886  
887  /**
888   * folio_referenced() - Test if the folio was referenced.
889   * @folio: The folio to test.
890   * @is_locked: Caller holds lock on the folio.
891   * @memcg: target memory cgroup
892   * @vm_flags: A combination of all the vma->vm_flags which referenced the folio.
893   *
894   * Quick test_and_clear_referenced for all mappings of a folio,
895   *
896   * Return: The number of mappings which referenced the folio. Return -1 if
897   * the function bailed out due to rmap lock contention.
898   */
899  int folio_referenced(struct folio *folio, int is_locked,
900  		     struct mem_cgroup *memcg, unsigned long *vm_flags)
901  {
902  	int we_locked = 0;
903  	struct folio_referenced_arg pra = {
904  		.mapcount = folio_mapcount(folio),
905  		.memcg = memcg,
906  	};
907  	struct rmap_walk_control rwc = {
908  		.rmap_one = folio_referenced_one,
909  		.arg = (void *)&pra,
910  		.anon_lock = folio_lock_anon_vma_read,
911  		.try_lock = true,
912  		.invalid_vma = invalid_folio_referenced_vma,
913  	};
914  
915  	*vm_flags = 0;
916  	if (!pra.mapcount)
917  		return 0;
918  
919  	if (!folio_raw_mapping(folio))
920  		return 0;
921  
922  	if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) {
923  		we_locked = folio_trylock(folio);
924  		if (!we_locked)
925  			return 1;
926  	}
927  
928  	rmap_walk(folio, &rwc);
929  	*vm_flags = pra.vm_flags;
930  
931  	if (we_locked)
932  		folio_unlock(folio);
933  
934  	return rwc.contended ? -1 : pra.referenced;
935  }
936  
937  static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw)
938  {
939  	int cleaned = 0;
940  	struct vm_area_struct *vma = pvmw->vma;
941  	struct mmu_notifier_range range;
942  	unsigned long address = pvmw->address;
943  
944  	/*
945  	 * We have to assume the worse case ie pmd for invalidation. Note that
946  	 * the folio can not be freed from this function.
947  	 */
948  	mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 0,
949  				vma->vm_mm, address, vma_address_end(pvmw));
950  	mmu_notifier_invalidate_range_start(&range);
951  
952  	while (page_vma_mapped_walk(pvmw)) {
953  		int ret = 0;
954  
955  		address = pvmw->address;
956  		if (pvmw->pte) {
957  			pte_t entry;
958  			pte_t *pte = pvmw->pte;
959  
960  			if (!pte_dirty(*pte) && !pte_write(*pte))
961  				continue;
962  
963  			flush_cache_page(vma, address, pte_pfn(*pte));
964  			entry = ptep_clear_flush(vma, address, pte);
965  			entry = pte_wrprotect(entry);
966  			entry = pte_mkclean(entry);
967  			set_pte_at(vma->vm_mm, address, pte, entry);
968  			ret = 1;
969  		} else {
970  #ifdef CONFIG_TRANSPARENT_HUGEPAGE
971  			pmd_t *pmd = pvmw->pmd;
972  			pmd_t entry;
973  
974  			if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
975  				continue;
976  
977  			flush_cache_range(vma, address,
978  					  address + HPAGE_PMD_SIZE);
979  			entry = pmdp_invalidate(vma, address, pmd);
980  			entry = pmd_wrprotect(entry);
981  			entry = pmd_mkclean(entry);
982  			set_pmd_at(vma->vm_mm, address, pmd, entry);
983  			ret = 1;
984  #else
985  			/* unexpected pmd-mapped folio? */
986  			WARN_ON_ONCE(1);
987  #endif
988  		}
989  
990  		/*
991  		 * No need to call mmu_notifier_invalidate_range() as we are
992  		 * downgrading page table protection not changing it to point
993  		 * to a new page.
994  		 *
995  		 * See Documentation/mm/mmu_notifier.rst
996  		 */
997  		if (ret)
998  			cleaned++;
999  	}
1000  
1001  	mmu_notifier_invalidate_range_end(&range);
1002  
1003  	return cleaned;
1004  }
1005  
1006  static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma,
1007  			     unsigned long address, void *arg)
1008  {
1009  	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC);
1010  	int *cleaned = arg;
1011  
1012  	*cleaned += page_vma_mkclean_one(&pvmw);
1013  
1014  	return true;
1015  }
1016  
1017  static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1018  {
1019  	if (vma->vm_flags & VM_SHARED)
1020  		return false;
1021  
1022  	return true;
1023  }
1024  
1025  int folio_mkclean(struct folio *folio)
1026  {
1027  	int cleaned = 0;
1028  	struct address_space *mapping;
1029  	struct rmap_walk_control rwc = {
1030  		.arg = (void *)&cleaned,
1031  		.rmap_one = page_mkclean_one,
1032  		.invalid_vma = invalid_mkclean_vma,
1033  	};
1034  
1035  	BUG_ON(!folio_test_locked(folio));
1036  
1037  	if (!folio_mapped(folio))
1038  		return 0;
1039  
1040  	mapping = folio_mapping(folio);
1041  	if (!mapping)
1042  		return 0;
1043  
1044  	rmap_walk(folio, &rwc);
1045  
1046  	return cleaned;
1047  }
1048  EXPORT_SYMBOL_GPL(folio_mkclean);
1049  
1050  /**
1051   * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of
1052   *                     [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff)
1053   *                     within the @vma of shared mappings. And since clean PTEs
1054   *                     should also be readonly, write protects them too.
1055   * @pfn: start pfn.
1056   * @nr_pages: number of physically contiguous pages srarting with @pfn.
1057   * @pgoff: page offset that the @pfn mapped with.
1058   * @vma: vma that @pfn mapped within.
1059   *
1060   * Returns the number of cleaned PTEs (including PMDs).
1061   */
1062  int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff,
1063  		      struct vm_area_struct *vma)
1064  {
1065  	struct page_vma_mapped_walk pvmw = {
1066  		.pfn		= pfn,
1067  		.nr_pages	= nr_pages,
1068  		.pgoff		= pgoff,
1069  		.vma		= vma,
1070  		.flags		= PVMW_SYNC,
1071  	};
1072  
1073  	if (invalid_mkclean_vma(vma, NULL))
1074  		return 0;
1075  
1076  	pvmw.address = vma_pgoff_address(pgoff, nr_pages, vma);
1077  	VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma);
1078  
1079  	return page_vma_mkclean_one(&pvmw);
1080  }
1081  
1082  int folio_total_mapcount(struct folio *folio)
1083  {
1084  	int mapcount = folio_entire_mapcount(folio);
1085  	int nr_pages;
1086  	int i;
1087  
1088  	/* In the common case, avoid the loop when no pages mapped by PTE */
1089  	if (folio_nr_pages_mapped(folio) == 0)
1090  		return mapcount;
1091  	/*
1092  	 * Add all the PTE mappings of those pages mapped by PTE.
1093  	 * Limit the loop to folio_nr_pages_mapped()?
1094  	 * Perhaps: given all the raciness, that may be a good or a bad idea.
1095  	 */
1096  	nr_pages = folio_nr_pages(folio);
1097  	for (i = 0; i < nr_pages; i++)
1098  		mapcount += atomic_read(&folio_page(folio, i)->_mapcount);
1099  
1100  	/* But each of those _mapcounts was based on -1 */
1101  	mapcount += nr_pages;
1102  	return mapcount;
1103  }
1104  
1105  /**
1106   * page_move_anon_rmap - move a page to our anon_vma
1107   * @page:	the page to move to our anon_vma
1108   * @vma:	the vma the page belongs to
1109   *
1110   * When a page belongs exclusively to one process after a COW event,
1111   * that page can be moved into the anon_vma that belongs to just that
1112   * process, so the rmap code will not search the parent or sibling
1113   * processes.
1114   */
1115  void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1116  {
1117  	void *anon_vma = vma->anon_vma;
1118  	struct folio *folio = page_folio(page);
1119  
1120  	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1121  	VM_BUG_ON_VMA(!anon_vma, vma);
1122  
1123  	anon_vma += PAGE_MAPPING_ANON;
1124  	/*
1125  	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1126  	 * simultaneously, so a concurrent reader (eg folio_referenced()'s
1127  	 * folio_test_anon()) will not see one without the other.
1128  	 */
1129  	WRITE_ONCE(folio->mapping, anon_vma);
1130  	SetPageAnonExclusive(page);
1131  }
1132  
1133  /**
1134   * __page_set_anon_rmap - set up new anonymous rmap
1135   * @folio:	Folio which contains page.
1136   * @page:	Page to add to rmap.
1137   * @vma:	VM area to add page to.
1138   * @address:	User virtual address of the mapping
1139   * @exclusive:	the page is exclusively owned by the current process
1140   */
1141  static void __page_set_anon_rmap(struct folio *folio, struct page *page,
1142  	struct vm_area_struct *vma, unsigned long address, int exclusive)
1143  {
1144  	struct anon_vma *anon_vma = vma->anon_vma;
1145  
1146  	BUG_ON(!anon_vma);
1147  
1148  	if (folio_test_anon(folio))
1149  		goto out;
1150  
1151  	/*
1152  	 * If the page isn't exclusively mapped into this vma,
1153  	 * we must use the _oldest_ possible anon_vma for the
1154  	 * page mapping!
1155  	 */
1156  	if (!exclusive)
1157  		anon_vma = anon_vma->root;
1158  
1159  	/*
1160  	 * page_idle does a lockless/optimistic rmap scan on folio->mapping.
1161  	 * Make sure the compiler doesn't split the stores of anon_vma and
1162  	 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1163  	 * could mistake the mapping for a struct address_space and crash.
1164  	 */
1165  	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1166  	WRITE_ONCE(folio->mapping, (struct address_space *) anon_vma);
1167  	folio->index = linear_page_index(vma, address);
1168  out:
1169  	if (exclusive)
1170  		SetPageAnonExclusive(page);
1171  }
1172  
1173  /**
1174   * __page_check_anon_rmap - sanity check anonymous rmap addition
1175   * @page:	the page to add the mapping to
1176   * @vma:	the vm area in which the mapping is added
1177   * @address:	the user virtual address mapped
1178   */
1179  static void __page_check_anon_rmap(struct page *page,
1180  	struct vm_area_struct *vma, unsigned long address)
1181  {
1182  	struct folio *folio = page_folio(page);
1183  	/*
1184  	 * The page's anon-rmap details (mapping and index) are guaranteed to
1185  	 * be set up correctly at this point.
1186  	 *
1187  	 * We have exclusion against page_add_anon_rmap because the caller
1188  	 * always holds the page locked.
1189  	 *
1190  	 * We have exclusion against page_add_new_anon_rmap because those pages
1191  	 * are initially only visible via the pagetables, and the pte is locked
1192  	 * over the call to page_add_new_anon_rmap.
1193  	 */
1194  	VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root,
1195  			folio);
1196  	VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1197  		       page);
1198  }
1199  
1200  /**
1201   * page_add_anon_rmap - add pte mapping to an anonymous page
1202   * @page:	the page to add the mapping to
1203   * @vma:	the vm area in which the mapping is added
1204   * @address:	the user virtual address mapped
1205   * @flags:	the rmap flags
1206   *
1207   * The caller needs to hold the pte lock, and the page must be locked in
1208   * the anon_vma case: to serialize mapping,index checking after setting,
1209   * and to ensure that PageAnon is not being upgraded racily to PageKsm
1210   * (but PageKsm is never downgraded to PageAnon).
1211   */
1212  void page_add_anon_rmap(struct page *page, struct vm_area_struct *vma,
1213  		unsigned long address, rmap_t flags)
1214  {
1215  	struct folio *folio = page_folio(page);
1216  	atomic_t *mapped = &folio->_nr_pages_mapped;
1217  	int nr = 0, nr_pmdmapped = 0;
1218  	bool compound = flags & RMAP_COMPOUND;
1219  	bool first = true;
1220  
1221  	/* Is page being mapped by PTE? Is this its first map to be added? */
1222  	if (likely(!compound)) {
1223  		first = atomic_inc_and_test(&page->_mapcount);
1224  		nr = first;
1225  		if (first && folio_test_large(folio)) {
1226  			nr = atomic_inc_return_relaxed(mapped);
1227  			nr = (nr < COMPOUND_MAPPED);
1228  		}
1229  	} else if (folio_test_pmd_mappable(folio)) {
1230  		/* That test is redundant: it's for safety or to optimize out */
1231  
1232  		first = atomic_inc_and_test(&folio->_entire_mapcount);
1233  		if (first) {
1234  			nr = atomic_add_return_relaxed(COMPOUND_MAPPED, mapped);
1235  			if (likely(nr < COMPOUND_MAPPED + COMPOUND_MAPPED)) {
1236  				nr_pmdmapped = folio_nr_pages(folio);
1237  				nr = nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1238  				/* Raced ahead of a remove and another add? */
1239  				if (unlikely(nr < 0))
1240  					nr = 0;
1241  			} else {
1242  				/* Raced ahead of a remove of COMPOUND_MAPPED */
1243  				nr = 0;
1244  			}
1245  		}
1246  	}
1247  
1248  	VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
1249  	VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
1250  
1251  	if (nr_pmdmapped)
1252  		__lruvec_stat_mod_folio(folio, NR_ANON_THPS, nr_pmdmapped);
1253  	if (nr)
1254  		__lruvec_stat_mod_folio(folio, NR_ANON_MAPPED, nr);
1255  
1256  	if (likely(!folio_test_ksm(folio))) {
1257  		/* address might be in next vma when migration races vma_merge */
1258  		if (first)
1259  			__page_set_anon_rmap(folio, page, vma, address,
1260  					     !!(flags & RMAP_EXCLUSIVE));
1261  		else
1262  			__page_check_anon_rmap(page, vma, address);
1263  	}
1264  
1265  	mlock_vma_folio(folio, vma, compound);
1266  }
1267  
1268  /**
1269   * folio_add_new_anon_rmap - Add mapping to a new anonymous folio.
1270   * @folio:	The folio to add the mapping to.
1271   * @vma:	the vm area in which the mapping is added
1272   * @address:	the user virtual address mapped
1273   *
1274   * Like page_add_anon_rmap() but must only be called on *new* folios.
1275   * This means the inc-and-test can be bypassed.
1276   * The folio does not have to be locked.
1277   *
1278   * If the folio is large, it is accounted as a THP.  As the folio
1279   * is new, it's assumed to be mapped exclusively by a single process.
1280   */
1281  void folio_add_new_anon_rmap(struct folio *folio, struct vm_area_struct *vma,
1282  		unsigned long address)
1283  {
1284  	int nr;
1285  
1286  	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1287  	__folio_set_swapbacked(folio);
1288  
1289  	if (likely(!folio_test_pmd_mappable(folio))) {
1290  		/* increment count (starts at -1) */
1291  		atomic_set(&folio->_mapcount, 0);
1292  		nr = 1;
1293  	} else {
1294  		/* increment count (starts at -1) */
1295  		atomic_set(&folio->_entire_mapcount, 0);
1296  		atomic_set(&folio->_nr_pages_mapped, COMPOUND_MAPPED);
1297  		nr = folio_nr_pages(folio);
1298  		__lruvec_stat_mod_folio(folio, NR_ANON_THPS, nr);
1299  	}
1300  
1301  	__lruvec_stat_mod_folio(folio, NR_ANON_MAPPED, nr);
1302  	__page_set_anon_rmap(folio, &folio->page, vma, address, 1);
1303  }
1304  
1305  /**
1306   * page_add_file_rmap - add pte mapping to a file page
1307   * @page:	the page to add the mapping to
1308   * @vma:	the vm area in which the mapping is added
1309   * @compound:	charge the page as compound or small page
1310   *
1311   * The caller needs to hold the pte lock.
1312   */
1313  void page_add_file_rmap(struct page *page, struct vm_area_struct *vma,
1314  		bool compound)
1315  {
1316  	struct folio *folio = page_folio(page);
1317  	atomic_t *mapped = &folio->_nr_pages_mapped;
1318  	int nr = 0, nr_pmdmapped = 0;
1319  	bool first;
1320  
1321  	VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1322  
1323  	/* Is page being mapped by PTE? Is this its first map to be added? */
1324  	if (likely(!compound)) {
1325  		first = atomic_inc_and_test(&page->_mapcount);
1326  		nr = first;
1327  		if (first && folio_test_large(folio)) {
1328  			nr = atomic_inc_return_relaxed(mapped);
1329  			nr = (nr < COMPOUND_MAPPED);
1330  		}
1331  	} else if (folio_test_pmd_mappable(folio)) {
1332  		/* That test is redundant: it's for safety or to optimize out */
1333  
1334  		first = atomic_inc_and_test(&folio->_entire_mapcount);
1335  		if (first) {
1336  			nr = atomic_add_return_relaxed(COMPOUND_MAPPED, mapped);
1337  			if (likely(nr < COMPOUND_MAPPED + COMPOUND_MAPPED)) {
1338  				nr_pmdmapped = folio_nr_pages(folio);
1339  				nr = nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1340  				/* Raced ahead of a remove and another add? */
1341  				if (unlikely(nr < 0))
1342  					nr = 0;
1343  			} else {
1344  				/* Raced ahead of a remove of COMPOUND_MAPPED */
1345  				nr = 0;
1346  			}
1347  		}
1348  	}
1349  
1350  	if (nr_pmdmapped)
1351  		__lruvec_stat_mod_folio(folio, folio_test_swapbacked(folio) ?
1352  			NR_SHMEM_PMDMAPPED : NR_FILE_PMDMAPPED, nr_pmdmapped);
1353  	if (nr)
1354  		__lruvec_stat_mod_folio(folio, NR_FILE_MAPPED, nr);
1355  
1356  	mlock_vma_folio(folio, vma, compound);
1357  }
1358  
1359  /**
1360   * page_remove_rmap - take down pte mapping from a page
1361   * @page:	page to remove mapping from
1362   * @vma:	the vm area from which the mapping is removed
1363   * @compound:	uncharge the page as compound or small page
1364   *
1365   * The caller needs to hold the pte lock.
1366   */
1367  void page_remove_rmap(struct page *page, struct vm_area_struct *vma,
1368  		bool compound)
1369  {
1370  	struct folio *folio = page_folio(page);
1371  	atomic_t *mapped = &folio->_nr_pages_mapped;
1372  	int nr = 0, nr_pmdmapped = 0;
1373  	bool last;
1374  	enum node_stat_item idx;
1375  
1376  	VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1377  
1378  	/* Hugetlb pages are not counted in NR_*MAPPED */
1379  	if (unlikely(folio_test_hugetlb(folio))) {
1380  		/* hugetlb pages are always mapped with pmds */
1381  		atomic_dec(&folio->_entire_mapcount);
1382  		return;
1383  	}
1384  
1385  	/* Is page being unmapped by PTE? Is this its last map to be removed? */
1386  	if (likely(!compound)) {
1387  		last = atomic_add_negative(-1, &page->_mapcount);
1388  		nr = last;
1389  		if (last && folio_test_large(folio)) {
1390  			nr = atomic_dec_return_relaxed(mapped);
1391  			nr = (nr < COMPOUND_MAPPED);
1392  		}
1393  	} else if (folio_test_pmd_mappable(folio)) {
1394  		/* That test is redundant: it's for safety or to optimize out */
1395  
1396  		last = atomic_add_negative(-1, &folio->_entire_mapcount);
1397  		if (last) {
1398  			nr = atomic_sub_return_relaxed(COMPOUND_MAPPED, mapped);
1399  			if (likely(nr < COMPOUND_MAPPED)) {
1400  				nr_pmdmapped = folio_nr_pages(folio);
1401  				nr = nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1402  				/* Raced ahead of another remove and an add? */
1403  				if (unlikely(nr < 0))
1404  					nr = 0;
1405  			} else {
1406  				/* An add of COMPOUND_MAPPED raced ahead */
1407  				nr = 0;
1408  			}
1409  		}
1410  	}
1411  
1412  	if (nr_pmdmapped) {
1413  		if (folio_test_anon(folio))
1414  			idx = NR_ANON_THPS;
1415  		else if (folio_test_swapbacked(folio))
1416  			idx = NR_SHMEM_PMDMAPPED;
1417  		else
1418  			idx = NR_FILE_PMDMAPPED;
1419  		__lruvec_stat_mod_folio(folio, idx, -nr_pmdmapped);
1420  	}
1421  	if (nr) {
1422  		idx = folio_test_anon(folio) ? NR_ANON_MAPPED : NR_FILE_MAPPED;
1423  		__lruvec_stat_mod_folio(folio, idx, -nr);
1424  
1425  		/*
1426  		 * Queue anon THP for deferred split if at least one
1427  		 * page of the folio is unmapped and at least one page
1428  		 * is still mapped.
1429  		 */
1430  		if (folio_test_pmd_mappable(folio) && folio_test_anon(folio))
1431  			if (!compound || nr < nr_pmdmapped)
1432  				deferred_split_folio(folio);
1433  	}
1434  
1435  	/*
1436  	 * It would be tidy to reset folio_test_anon mapping when fully
1437  	 * unmapped, but that might overwrite a racing page_add_anon_rmap
1438  	 * which increments mapcount after us but sets mapping before us:
1439  	 * so leave the reset to free_pages_prepare, and remember that
1440  	 * it's only reliable while mapped.
1441  	 */
1442  
1443  	munlock_vma_folio(folio, vma, compound);
1444  }
1445  
1446  /*
1447   * @arg: enum ttu_flags will be passed to this argument
1448   */
1449  static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
1450  		     unsigned long address, void *arg)
1451  {
1452  	struct mm_struct *mm = vma->vm_mm;
1453  	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1454  	pte_t pteval;
1455  	struct page *subpage;
1456  	bool anon_exclusive, ret = true;
1457  	struct mmu_notifier_range range;
1458  	enum ttu_flags flags = (enum ttu_flags)(long)arg;
1459  
1460  	/*
1461  	 * When racing against e.g. zap_pte_range() on another cpu,
1462  	 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1463  	 * try_to_unmap() may return before page_mapped() has become false,
1464  	 * if page table locking is skipped: use TTU_SYNC to wait for that.
1465  	 */
1466  	if (flags & TTU_SYNC)
1467  		pvmw.flags = PVMW_SYNC;
1468  
1469  	if (flags & TTU_SPLIT_HUGE_PMD)
1470  		split_huge_pmd_address(vma, address, false, folio);
1471  
1472  	/*
1473  	 * For THP, we have to assume the worse case ie pmd for invalidation.
1474  	 * For hugetlb, it could be much worse if we need to do pud
1475  	 * invalidation in the case of pmd sharing.
1476  	 *
1477  	 * Note that the folio can not be freed in this function as call of
1478  	 * try_to_unmap() must hold a reference on the folio.
1479  	 */
1480  	range.end = vma_address_end(&pvmw);
1481  	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1482  				address, range.end);
1483  	if (folio_test_hugetlb(folio)) {
1484  		/*
1485  		 * If sharing is possible, start and end will be adjusted
1486  		 * accordingly.
1487  		 */
1488  		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1489  						     &range.end);
1490  	}
1491  	mmu_notifier_invalidate_range_start(&range);
1492  
1493  	while (page_vma_mapped_walk(&pvmw)) {
1494  		/* Unexpected PMD-mapped THP? */
1495  		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1496  
1497  		/*
1498  		 * If the folio is in an mlock()d vma, we must not swap it out.
1499  		 */
1500  		if (!(flags & TTU_IGNORE_MLOCK) &&
1501  		    (vma->vm_flags & VM_LOCKED)) {
1502  			/* Restore the mlock which got missed */
1503  			mlock_vma_folio(folio, vma, false);
1504  			page_vma_mapped_walk_done(&pvmw);
1505  			ret = false;
1506  			break;
1507  		}
1508  
1509  		subpage = folio_page(folio,
1510  					pte_pfn(*pvmw.pte) - folio_pfn(folio));
1511  		address = pvmw.address;
1512  		anon_exclusive = folio_test_anon(folio) &&
1513  				 PageAnonExclusive(subpage);
1514  
1515  		if (folio_test_hugetlb(folio)) {
1516  			bool anon = folio_test_anon(folio);
1517  
1518  			/*
1519  			 * The try_to_unmap() is only passed a hugetlb page
1520  			 * in the case where the hugetlb page is poisoned.
1521  			 */
1522  			VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage);
1523  			/*
1524  			 * huge_pmd_unshare may unmap an entire PMD page.
1525  			 * There is no way of knowing exactly which PMDs may
1526  			 * be cached for this mm, so we must flush them all.
1527  			 * start/end were already adjusted above to cover this
1528  			 * range.
1529  			 */
1530  			flush_cache_range(vma, range.start, range.end);
1531  
1532  			/*
1533  			 * To call huge_pmd_unshare, i_mmap_rwsem must be
1534  			 * held in write mode.  Caller needs to explicitly
1535  			 * do this outside rmap routines.
1536  			 *
1537  			 * We also must hold hugetlb vma_lock in write mode.
1538  			 * Lock order dictates acquiring vma_lock BEFORE
1539  			 * i_mmap_rwsem.  We can only try lock here and fail
1540  			 * if unsuccessful.
1541  			 */
1542  			if (!anon) {
1543  				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1544  				if (!hugetlb_vma_trylock_write(vma)) {
1545  					page_vma_mapped_walk_done(&pvmw);
1546  					ret = false;
1547  					break;
1548  				}
1549  				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1550  					hugetlb_vma_unlock_write(vma);
1551  					flush_tlb_range(vma,
1552  						range.start, range.end);
1553  					mmu_notifier_invalidate_range(mm,
1554  						range.start, range.end);
1555  					/*
1556  					 * The ref count of the PMD page was
1557  					 * dropped which is part of the way map
1558  					 * counting is done for shared PMDs.
1559  					 * Return 'true' here.  When there is
1560  					 * no other sharing, huge_pmd_unshare
1561  					 * returns false and we will unmap the
1562  					 * actual page and drop map count
1563  					 * to zero.
1564  					 */
1565  					page_vma_mapped_walk_done(&pvmw);
1566  					break;
1567  				}
1568  				hugetlb_vma_unlock_write(vma);
1569  			}
1570  			pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1571  		} else {
1572  			flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1573  			/* Nuke the page table entry. */
1574  			if (should_defer_flush(mm, flags)) {
1575  				/*
1576  				 * We clear the PTE but do not flush so potentially
1577  				 * a remote CPU could still be writing to the folio.
1578  				 * If the entry was previously clean then the
1579  				 * architecture must guarantee that a clear->dirty
1580  				 * transition on a cached TLB entry is written through
1581  				 * and traps if the PTE is unmapped.
1582  				 */
1583  				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1584  
1585  				set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1586  			} else {
1587  				pteval = ptep_clear_flush(vma, address, pvmw.pte);
1588  			}
1589  		}
1590  
1591  		/*
1592  		 * Now the pte is cleared. If this pte was uffd-wp armed,
1593  		 * we may want to replace a none pte with a marker pte if
1594  		 * it's file-backed, so we don't lose the tracking info.
1595  		 */
1596  		pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval);
1597  
1598  		/* Set the dirty flag on the folio now the pte is gone. */
1599  		if (pte_dirty(pteval))
1600  			folio_mark_dirty(folio);
1601  
1602  		/* Update high watermark before we lower rss */
1603  		update_hiwater_rss(mm);
1604  
1605  		if (PageHWPoison(subpage) && (flags & TTU_HWPOISON)) {
1606  			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1607  			if (folio_test_hugetlb(folio)) {
1608  				hugetlb_count_sub(folio_nr_pages(folio), mm);
1609  				set_huge_pte_at(mm, address, pvmw.pte, pteval);
1610  			} else {
1611  				dec_mm_counter(mm, mm_counter(&folio->page));
1612  				set_pte_at(mm, address, pvmw.pte, pteval);
1613  			}
1614  
1615  		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1616  			/*
1617  			 * The guest indicated that the page content is of no
1618  			 * interest anymore. Simply discard the pte, vmscan
1619  			 * will take care of the rest.
1620  			 * A future reference will then fault in a new zero
1621  			 * page. When userfaultfd is active, we must not drop
1622  			 * this page though, as its main user (postcopy
1623  			 * migration) will not expect userfaults on already
1624  			 * copied pages.
1625  			 */
1626  			dec_mm_counter(mm, mm_counter(&folio->page));
1627  			/* We have to invalidate as we cleared the pte */
1628  			mmu_notifier_invalidate_range(mm, address,
1629  						      address + PAGE_SIZE);
1630  		} else if (folio_test_anon(folio)) {
1631  			swp_entry_t entry = { .val = page_private(subpage) };
1632  			pte_t swp_pte;
1633  			/*
1634  			 * Store the swap location in the pte.
1635  			 * See handle_pte_fault() ...
1636  			 */
1637  			if (unlikely(folio_test_swapbacked(folio) !=
1638  					folio_test_swapcache(folio))) {
1639  				WARN_ON_ONCE(1);
1640  				ret = false;
1641  				/* We have to invalidate as we cleared the pte */
1642  				mmu_notifier_invalidate_range(mm, address,
1643  							address + PAGE_SIZE);
1644  				page_vma_mapped_walk_done(&pvmw);
1645  				break;
1646  			}
1647  
1648  			/* MADV_FREE page check */
1649  			if (!folio_test_swapbacked(folio)) {
1650  				int ref_count, map_count;
1651  
1652  				/*
1653  				 * Synchronize with gup_pte_range():
1654  				 * - clear PTE; barrier; read refcount
1655  				 * - inc refcount; barrier; read PTE
1656  				 */
1657  				smp_mb();
1658  
1659  				ref_count = folio_ref_count(folio);
1660  				map_count = folio_mapcount(folio);
1661  
1662  				/*
1663  				 * Order reads for page refcount and dirty flag
1664  				 * (see comments in __remove_mapping()).
1665  				 */
1666  				smp_rmb();
1667  
1668  				/*
1669  				 * The only page refs must be one from isolation
1670  				 * plus the rmap(s) (dropped by discard:).
1671  				 */
1672  				if (ref_count == 1 + map_count &&
1673  				    !folio_test_dirty(folio)) {
1674  					/* Invalidate as we cleared the pte */
1675  					mmu_notifier_invalidate_range(mm,
1676  						address, address + PAGE_SIZE);
1677  					dec_mm_counter(mm, MM_ANONPAGES);
1678  					goto discard;
1679  				}
1680  
1681  				/*
1682  				 * If the folio was redirtied, it cannot be
1683  				 * discarded. Remap the page to page table.
1684  				 */
1685  				set_pte_at(mm, address, pvmw.pte, pteval);
1686  				folio_set_swapbacked(folio);
1687  				ret = false;
1688  				page_vma_mapped_walk_done(&pvmw);
1689  				break;
1690  			}
1691  
1692  			if (swap_duplicate(entry) < 0) {
1693  				set_pte_at(mm, address, pvmw.pte, pteval);
1694  				ret = false;
1695  				page_vma_mapped_walk_done(&pvmw);
1696  				break;
1697  			}
1698  			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1699  				swap_free(entry);
1700  				set_pte_at(mm, address, pvmw.pte, pteval);
1701  				ret = false;
1702  				page_vma_mapped_walk_done(&pvmw);
1703  				break;
1704  			}
1705  
1706  			/* See page_try_share_anon_rmap(): clear PTE first. */
1707  			if (anon_exclusive &&
1708  			    page_try_share_anon_rmap(subpage)) {
1709  				swap_free(entry);
1710  				set_pte_at(mm, address, pvmw.pte, pteval);
1711  				ret = false;
1712  				page_vma_mapped_walk_done(&pvmw);
1713  				break;
1714  			}
1715  			if (list_empty(&mm->mmlist)) {
1716  				spin_lock(&mmlist_lock);
1717  				if (list_empty(&mm->mmlist))
1718  					list_add(&mm->mmlist, &init_mm.mmlist);
1719  				spin_unlock(&mmlist_lock);
1720  			}
1721  			dec_mm_counter(mm, MM_ANONPAGES);
1722  			inc_mm_counter(mm, MM_SWAPENTS);
1723  			swp_pte = swp_entry_to_pte(entry);
1724  			if (anon_exclusive)
1725  				swp_pte = pte_swp_mkexclusive(swp_pte);
1726  			if (pte_soft_dirty(pteval))
1727  				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1728  			if (pte_uffd_wp(pteval))
1729  				swp_pte = pte_swp_mkuffd_wp(swp_pte);
1730  			set_pte_at(mm, address, pvmw.pte, swp_pte);
1731  			/* Invalidate as we cleared the pte */
1732  			mmu_notifier_invalidate_range(mm, address,
1733  						      address + PAGE_SIZE);
1734  		} else {
1735  			/*
1736  			 * This is a locked file-backed folio,
1737  			 * so it cannot be removed from the page
1738  			 * cache and replaced by a new folio before
1739  			 * mmu_notifier_invalidate_range_end, so no
1740  			 * concurrent thread might update its page table
1741  			 * to point at a new folio while a device is
1742  			 * still using this folio.
1743  			 *
1744  			 * See Documentation/mm/mmu_notifier.rst
1745  			 */
1746  			dec_mm_counter(mm, mm_counter_file(&folio->page));
1747  		}
1748  discard:
1749  		/*
1750  		 * No need to call mmu_notifier_invalidate_range() it has be
1751  		 * done above for all cases requiring it to happen under page
1752  		 * table lock before mmu_notifier_invalidate_range_end()
1753  		 *
1754  		 * See Documentation/mm/mmu_notifier.rst
1755  		 */
1756  		page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
1757  		if (vma->vm_flags & VM_LOCKED)
1758  			mlock_drain_local();
1759  		folio_put(folio);
1760  	}
1761  
1762  	mmu_notifier_invalidate_range_end(&range);
1763  
1764  	return ret;
1765  }
1766  
1767  static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1768  {
1769  	return vma_is_temporary_stack(vma);
1770  }
1771  
1772  static int folio_not_mapped(struct folio *folio)
1773  {
1774  	return !folio_mapped(folio);
1775  }
1776  
1777  /**
1778   * try_to_unmap - Try to remove all page table mappings to a folio.
1779   * @folio: The folio to unmap.
1780   * @flags: action and flags
1781   *
1782   * Tries to remove all the page table entries which are mapping this
1783   * folio.  It is the caller's responsibility to check if the folio is
1784   * still mapped if needed (use TTU_SYNC to prevent accounting races).
1785   *
1786   * Context: Caller must hold the folio lock.
1787   */
1788  void try_to_unmap(struct folio *folio, enum ttu_flags flags)
1789  {
1790  	struct rmap_walk_control rwc = {
1791  		.rmap_one = try_to_unmap_one,
1792  		.arg = (void *)flags,
1793  		.done = folio_not_mapped,
1794  		.anon_lock = folio_lock_anon_vma_read,
1795  	};
1796  
1797  	if (flags & TTU_RMAP_LOCKED)
1798  		rmap_walk_locked(folio, &rwc);
1799  	else
1800  		rmap_walk(folio, &rwc);
1801  }
1802  
1803  /*
1804   * @arg: enum ttu_flags will be passed to this argument.
1805   *
1806   * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
1807   * containing migration entries.
1808   */
1809  static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
1810  		     unsigned long address, void *arg)
1811  {
1812  	struct mm_struct *mm = vma->vm_mm;
1813  	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1814  	pte_t pteval;
1815  	struct page *subpage;
1816  	bool anon_exclusive, ret = true;
1817  	struct mmu_notifier_range range;
1818  	enum ttu_flags flags = (enum ttu_flags)(long)arg;
1819  
1820  	/*
1821  	 * When racing against e.g. zap_pte_range() on another cpu,
1822  	 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1823  	 * try_to_migrate() may return before page_mapped() has become false,
1824  	 * if page table locking is skipped: use TTU_SYNC to wait for that.
1825  	 */
1826  	if (flags & TTU_SYNC)
1827  		pvmw.flags = PVMW_SYNC;
1828  
1829  	/*
1830  	 * unmap_page() in mm/huge_memory.c is the only user of migration with
1831  	 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
1832  	 */
1833  	if (flags & TTU_SPLIT_HUGE_PMD)
1834  		split_huge_pmd_address(vma, address, true, folio);
1835  
1836  	/*
1837  	 * For THP, we have to assume the worse case ie pmd for invalidation.
1838  	 * For hugetlb, it could be much worse if we need to do pud
1839  	 * invalidation in the case of pmd sharing.
1840  	 *
1841  	 * Note that the page can not be free in this function as call of
1842  	 * try_to_unmap() must hold a reference on the page.
1843  	 */
1844  	range.end = vma_address_end(&pvmw);
1845  	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1846  				address, range.end);
1847  	if (folio_test_hugetlb(folio)) {
1848  		/*
1849  		 * If sharing is possible, start and end will be adjusted
1850  		 * accordingly.
1851  		 */
1852  		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1853  						     &range.end);
1854  	}
1855  	mmu_notifier_invalidate_range_start(&range);
1856  
1857  	while (page_vma_mapped_walk(&pvmw)) {
1858  #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1859  		/* PMD-mapped THP migration entry */
1860  		if (!pvmw.pte) {
1861  			subpage = folio_page(folio,
1862  				pmd_pfn(*pvmw.pmd) - folio_pfn(folio));
1863  			VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
1864  					!folio_test_pmd_mappable(folio), folio);
1865  
1866  			if (set_pmd_migration_entry(&pvmw, subpage)) {
1867  				ret = false;
1868  				page_vma_mapped_walk_done(&pvmw);
1869  				break;
1870  			}
1871  			continue;
1872  		}
1873  #endif
1874  
1875  		/* Unexpected PMD-mapped THP? */
1876  		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1877  
1878  		if (folio_is_zone_device(folio)) {
1879  			/*
1880  			 * Our PTE is a non-present device exclusive entry and
1881  			 * calculating the subpage as for the common case would
1882  			 * result in an invalid pointer.
1883  			 *
1884  			 * Since only PAGE_SIZE pages can currently be
1885  			 * migrated, just set it to page. This will need to be
1886  			 * changed when hugepage migrations to device private
1887  			 * memory are supported.
1888  			 */
1889  			VM_BUG_ON_FOLIO(folio_nr_pages(folio) > 1, folio);
1890  			subpage = &folio->page;
1891  		} else {
1892  			subpage = folio_page(folio,
1893  					pte_pfn(*pvmw.pte) - folio_pfn(folio));
1894  		}
1895  		address = pvmw.address;
1896  		anon_exclusive = folio_test_anon(folio) &&
1897  				 PageAnonExclusive(subpage);
1898  
1899  		if (folio_test_hugetlb(folio)) {
1900  			bool anon = folio_test_anon(folio);
1901  
1902  			/*
1903  			 * huge_pmd_unshare may unmap an entire PMD page.
1904  			 * There is no way of knowing exactly which PMDs may
1905  			 * be cached for this mm, so we must flush them all.
1906  			 * start/end were already adjusted above to cover this
1907  			 * range.
1908  			 */
1909  			flush_cache_range(vma, range.start, range.end);
1910  
1911  			/*
1912  			 * To call huge_pmd_unshare, i_mmap_rwsem must be
1913  			 * held in write mode.  Caller needs to explicitly
1914  			 * do this outside rmap routines.
1915  			 *
1916  			 * We also must hold hugetlb vma_lock in write mode.
1917  			 * Lock order dictates acquiring vma_lock BEFORE
1918  			 * i_mmap_rwsem.  We can only try lock here and
1919  			 * fail if unsuccessful.
1920  			 */
1921  			if (!anon) {
1922  				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1923  				if (!hugetlb_vma_trylock_write(vma)) {
1924  					page_vma_mapped_walk_done(&pvmw);
1925  					ret = false;
1926  					break;
1927  				}
1928  				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1929  					hugetlb_vma_unlock_write(vma);
1930  					flush_tlb_range(vma,
1931  						range.start, range.end);
1932  					mmu_notifier_invalidate_range(mm,
1933  						range.start, range.end);
1934  
1935  					/*
1936  					 * The ref count of the PMD page was
1937  					 * dropped which is part of the way map
1938  					 * counting is done for shared PMDs.
1939  					 * Return 'true' here.  When there is
1940  					 * no other sharing, huge_pmd_unshare
1941  					 * returns false and we will unmap the
1942  					 * actual page and drop map count
1943  					 * to zero.
1944  					 */
1945  					page_vma_mapped_walk_done(&pvmw);
1946  					break;
1947  				}
1948  				hugetlb_vma_unlock_write(vma);
1949  			}
1950  			/* Nuke the hugetlb page table entry */
1951  			pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1952  		} else {
1953  			flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1954  			/* Nuke the page table entry. */
1955  			if (should_defer_flush(mm, flags)) {
1956  				/*
1957  				 * We clear the PTE but do not flush so potentially
1958  				 * a remote CPU could still be writing to the folio.
1959  				 * If the entry was previously clean then the
1960  				 * architecture must guarantee that a clear->dirty
1961  				 * transition on a cached TLB entry is written through
1962  				 * and traps if the PTE is unmapped.
1963  				 */
1964  				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1965  
1966  				set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1967  			} else {
1968  				pteval = ptep_clear_flush(vma, address, pvmw.pte);
1969  			}
1970  		}
1971  
1972  		/* Set the dirty flag on the folio now the pte is gone. */
1973  		if (pte_dirty(pteval))
1974  			folio_mark_dirty(folio);
1975  
1976  		/* Update high watermark before we lower rss */
1977  		update_hiwater_rss(mm);
1978  
1979  		if (folio_is_device_private(folio)) {
1980  			unsigned long pfn = folio_pfn(folio);
1981  			swp_entry_t entry;
1982  			pte_t swp_pte;
1983  
1984  			if (anon_exclusive)
1985  				BUG_ON(page_try_share_anon_rmap(subpage));
1986  
1987  			/*
1988  			 * Store the pfn of the page in a special migration
1989  			 * pte. do_swap_page() will wait until the migration
1990  			 * pte is removed and then restart fault handling.
1991  			 */
1992  			entry = pte_to_swp_entry(pteval);
1993  			if (is_writable_device_private_entry(entry))
1994  				entry = make_writable_migration_entry(pfn);
1995  			else if (anon_exclusive)
1996  				entry = make_readable_exclusive_migration_entry(pfn);
1997  			else
1998  				entry = make_readable_migration_entry(pfn);
1999  			swp_pte = swp_entry_to_pte(entry);
2000  
2001  			/*
2002  			 * pteval maps a zone device page and is therefore
2003  			 * a swap pte.
2004  			 */
2005  			if (pte_swp_soft_dirty(pteval))
2006  				swp_pte = pte_swp_mksoft_dirty(swp_pte);
2007  			if (pte_swp_uffd_wp(pteval))
2008  				swp_pte = pte_swp_mkuffd_wp(swp_pte);
2009  			set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
2010  			trace_set_migration_pte(pvmw.address, pte_val(swp_pte),
2011  						compound_order(&folio->page));
2012  			/*
2013  			 * No need to invalidate here it will synchronize on
2014  			 * against the special swap migration pte.
2015  			 */
2016  		} else if (PageHWPoison(subpage)) {
2017  			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
2018  			if (folio_test_hugetlb(folio)) {
2019  				hugetlb_count_sub(folio_nr_pages(folio), mm);
2020  				set_huge_pte_at(mm, address, pvmw.pte, pteval);
2021  			} else {
2022  				dec_mm_counter(mm, mm_counter(&folio->page));
2023  				set_pte_at(mm, address, pvmw.pte, pteval);
2024  			}
2025  
2026  		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
2027  			/*
2028  			 * The guest indicated that the page content is of no
2029  			 * interest anymore. Simply discard the pte, vmscan
2030  			 * will take care of the rest.
2031  			 * A future reference will then fault in a new zero
2032  			 * page. When userfaultfd is active, we must not drop
2033  			 * this page though, as its main user (postcopy
2034  			 * migration) will not expect userfaults on already
2035  			 * copied pages.
2036  			 */
2037  			dec_mm_counter(mm, mm_counter(&folio->page));
2038  			/* We have to invalidate as we cleared the pte */
2039  			mmu_notifier_invalidate_range(mm, address,
2040  						      address + PAGE_SIZE);
2041  		} else {
2042  			swp_entry_t entry;
2043  			pte_t swp_pte;
2044  
2045  			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
2046  				if (folio_test_hugetlb(folio))
2047  					set_huge_pte_at(mm, address, pvmw.pte, pteval);
2048  				else
2049  					set_pte_at(mm, address, pvmw.pte, pteval);
2050  				ret = false;
2051  				page_vma_mapped_walk_done(&pvmw);
2052  				break;
2053  			}
2054  			VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) &&
2055  				       !anon_exclusive, subpage);
2056  
2057  			/* See page_try_share_anon_rmap(): clear PTE first. */
2058  			if (anon_exclusive &&
2059  			    page_try_share_anon_rmap(subpage)) {
2060  				if (folio_test_hugetlb(folio))
2061  					set_huge_pte_at(mm, address, pvmw.pte, pteval);
2062  				else
2063  					set_pte_at(mm, address, pvmw.pte, pteval);
2064  				ret = false;
2065  				page_vma_mapped_walk_done(&pvmw);
2066  				break;
2067  			}
2068  
2069  			/*
2070  			 * Store the pfn of the page in a special migration
2071  			 * pte. do_swap_page() will wait until the migration
2072  			 * pte is removed and then restart fault handling.
2073  			 */
2074  			if (pte_write(pteval))
2075  				entry = make_writable_migration_entry(
2076  							page_to_pfn(subpage));
2077  			else if (anon_exclusive)
2078  				entry = make_readable_exclusive_migration_entry(
2079  							page_to_pfn(subpage));
2080  			else
2081  				entry = make_readable_migration_entry(
2082  							page_to_pfn(subpage));
2083  			if (pte_young(pteval))
2084  				entry = make_migration_entry_young(entry);
2085  			if (pte_dirty(pteval))
2086  				entry = make_migration_entry_dirty(entry);
2087  			swp_pte = swp_entry_to_pte(entry);
2088  			if (pte_soft_dirty(pteval))
2089  				swp_pte = pte_swp_mksoft_dirty(swp_pte);
2090  			if (pte_uffd_wp(pteval))
2091  				swp_pte = pte_swp_mkuffd_wp(swp_pte);
2092  			if (folio_test_hugetlb(folio))
2093  				set_huge_pte_at(mm, address, pvmw.pte, swp_pte);
2094  			else
2095  				set_pte_at(mm, address, pvmw.pte, swp_pte);
2096  			trace_set_migration_pte(address, pte_val(swp_pte),
2097  						compound_order(&folio->page));
2098  			/*
2099  			 * No need to invalidate here it will synchronize on
2100  			 * against the special swap migration pte.
2101  			 */
2102  		}
2103  
2104  		/*
2105  		 * No need to call mmu_notifier_invalidate_range() it has be
2106  		 * done above for all cases requiring it to happen under page
2107  		 * table lock before mmu_notifier_invalidate_range_end()
2108  		 *
2109  		 * See Documentation/mm/mmu_notifier.rst
2110  		 */
2111  		page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
2112  		if (vma->vm_flags & VM_LOCKED)
2113  			mlock_drain_local();
2114  		folio_put(folio);
2115  	}
2116  
2117  	mmu_notifier_invalidate_range_end(&range);
2118  
2119  	return ret;
2120  }
2121  
2122  /**
2123   * try_to_migrate - try to replace all page table mappings with swap entries
2124   * @folio: the folio to replace page table entries for
2125   * @flags: action and flags
2126   *
2127   * Tries to remove all the page table entries which are mapping this folio and
2128   * replace them with special swap entries. Caller must hold the folio lock.
2129   */
2130  void try_to_migrate(struct folio *folio, enum ttu_flags flags)
2131  {
2132  	struct rmap_walk_control rwc = {
2133  		.rmap_one = try_to_migrate_one,
2134  		.arg = (void *)flags,
2135  		.done = folio_not_mapped,
2136  		.anon_lock = folio_lock_anon_vma_read,
2137  	};
2138  
2139  	/*
2140  	 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
2141  	 * TTU_SPLIT_HUGE_PMD, TTU_SYNC, and TTU_BATCH_FLUSH flags.
2142  	 */
2143  	if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2144  					TTU_SYNC | TTU_BATCH_FLUSH)))
2145  		return;
2146  
2147  	if (folio_is_zone_device(folio) &&
2148  	    (!folio_is_device_private(folio) && !folio_is_device_coherent(folio)))
2149  		return;
2150  
2151  	/*
2152  	 * During exec, a temporary VMA is setup and later moved.
2153  	 * The VMA is moved under the anon_vma lock but not the
2154  	 * page tables leading to a race where migration cannot
2155  	 * find the migration ptes. Rather than increasing the
2156  	 * locking requirements of exec(), migration skips
2157  	 * temporary VMAs until after exec() completes.
2158  	 */
2159  	if (!folio_test_ksm(folio) && folio_test_anon(folio))
2160  		rwc.invalid_vma = invalid_migration_vma;
2161  
2162  	if (flags & TTU_RMAP_LOCKED)
2163  		rmap_walk_locked(folio, &rwc);
2164  	else
2165  		rmap_walk(folio, &rwc);
2166  }
2167  
2168  #ifdef CONFIG_DEVICE_PRIVATE
2169  struct make_exclusive_args {
2170  	struct mm_struct *mm;
2171  	unsigned long address;
2172  	void *owner;
2173  	bool valid;
2174  };
2175  
2176  static bool page_make_device_exclusive_one(struct folio *folio,
2177  		struct vm_area_struct *vma, unsigned long address, void *priv)
2178  {
2179  	struct mm_struct *mm = vma->vm_mm;
2180  	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
2181  	struct make_exclusive_args *args = priv;
2182  	pte_t pteval;
2183  	struct page *subpage;
2184  	bool ret = true;
2185  	struct mmu_notifier_range range;
2186  	swp_entry_t entry;
2187  	pte_t swp_pte;
2188  
2189  	mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
2190  				      vma->vm_mm, address, min(vma->vm_end,
2191  				      address + folio_size(folio)),
2192  				      args->owner);
2193  	mmu_notifier_invalidate_range_start(&range);
2194  
2195  	while (page_vma_mapped_walk(&pvmw)) {
2196  		/* Unexpected PMD-mapped THP? */
2197  		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2198  
2199  		if (!pte_present(*pvmw.pte)) {
2200  			ret = false;
2201  			page_vma_mapped_walk_done(&pvmw);
2202  			break;
2203  		}
2204  
2205  		subpage = folio_page(folio,
2206  				pte_pfn(*pvmw.pte) - folio_pfn(folio));
2207  		address = pvmw.address;
2208  
2209  		/* Nuke the page table entry. */
2210  		flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
2211  		pteval = ptep_clear_flush(vma, address, pvmw.pte);
2212  
2213  		/* Set the dirty flag on the folio now the pte is gone. */
2214  		if (pte_dirty(pteval))
2215  			folio_mark_dirty(folio);
2216  
2217  		/*
2218  		 * Check that our target page is still mapped at the expected
2219  		 * address.
2220  		 */
2221  		if (args->mm == mm && args->address == address &&
2222  		    pte_write(pteval))
2223  			args->valid = true;
2224  
2225  		/*
2226  		 * Store the pfn of the page in a special migration
2227  		 * pte. do_swap_page() will wait until the migration
2228  		 * pte is removed and then restart fault handling.
2229  		 */
2230  		if (pte_write(pteval))
2231  			entry = make_writable_device_exclusive_entry(
2232  							page_to_pfn(subpage));
2233  		else
2234  			entry = make_readable_device_exclusive_entry(
2235  							page_to_pfn(subpage));
2236  		swp_pte = swp_entry_to_pte(entry);
2237  		if (pte_soft_dirty(pteval))
2238  			swp_pte = pte_swp_mksoft_dirty(swp_pte);
2239  		if (pte_uffd_wp(pteval))
2240  			swp_pte = pte_swp_mkuffd_wp(swp_pte);
2241  
2242  		set_pte_at(mm, address, pvmw.pte, swp_pte);
2243  
2244  		/*
2245  		 * There is a reference on the page for the swap entry which has
2246  		 * been removed, so shouldn't take another.
2247  		 */
2248  		page_remove_rmap(subpage, vma, false);
2249  	}
2250  
2251  	mmu_notifier_invalidate_range_end(&range);
2252  
2253  	return ret;
2254  }
2255  
2256  /**
2257   * folio_make_device_exclusive - Mark the folio exclusively owned by a device.
2258   * @folio: The folio to replace page table entries for.
2259   * @mm: The mm_struct where the folio is expected to be mapped.
2260   * @address: Address where the folio is expected to be mapped.
2261   * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2262   *
2263   * Tries to remove all the page table entries which are mapping this
2264   * folio and replace them with special device exclusive swap entries to
2265   * grant a device exclusive access to the folio.
2266   *
2267   * Context: Caller must hold the folio lock.
2268   * Return: false if the page is still mapped, or if it could not be unmapped
2269   * from the expected address. Otherwise returns true (success).
2270   */
2271  static bool folio_make_device_exclusive(struct folio *folio,
2272  		struct mm_struct *mm, unsigned long address, void *owner)
2273  {
2274  	struct make_exclusive_args args = {
2275  		.mm = mm,
2276  		.address = address,
2277  		.owner = owner,
2278  		.valid = false,
2279  	};
2280  	struct rmap_walk_control rwc = {
2281  		.rmap_one = page_make_device_exclusive_one,
2282  		.done = folio_not_mapped,
2283  		.anon_lock = folio_lock_anon_vma_read,
2284  		.arg = &args,
2285  	};
2286  
2287  	/*
2288  	 * Restrict to anonymous folios for now to avoid potential writeback
2289  	 * issues.
2290  	 */
2291  	if (!folio_test_anon(folio))
2292  		return false;
2293  
2294  	rmap_walk(folio, &rwc);
2295  
2296  	return args.valid && !folio_mapcount(folio);
2297  }
2298  
2299  /**
2300   * make_device_exclusive_range() - Mark a range for exclusive use by a device
2301   * @mm: mm_struct of associated target process
2302   * @start: start of the region to mark for exclusive device access
2303   * @end: end address of region
2304   * @pages: returns the pages which were successfully marked for exclusive access
2305   * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2306   *
2307   * Returns: number of pages found in the range by GUP. A page is marked for
2308   * exclusive access only if the page pointer is non-NULL.
2309   *
2310   * This function finds ptes mapping page(s) to the given address range, locks
2311   * them and replaces mappings with special swap entries preventing userspace CPU
2312   * access. On fault these entries are replaced with the original mapping after
2313   * calling MMU notifiers.
2314   *
2315   * A driver using this to program access from a device must use a mmu notifier
2316   * critical section to hold a device specific lock during programming. Once
2317   * programming is complete it should drop the page lock and reference after
2318   * which point CPU access to the page will revoke the exclusive access.
2319   */
2320  int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2321  				unsigned long end, struct page **pages,
2322  				void *owner)
2323  {
2324  	long npages = (end - start) >> PAGE_SHIFT;
2325  	long i;
2326  
2327  	npages = get_user_pages_remote(mm, start, npages,
2328  				       FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2329  				       pages, NULL, NULL);
2330  	if (npages < 0)
2331  		return npages;
2332  
2333  	for (i = 0; i < npages; i++, start += PAGE_SIZE) {
2334  		struct folio *folio = page_folio(pages[i]);
2335  		if (PageTail(pages[i]) || !folio_trylock(folio)) {
2336  			folio_put(folio);
2337  			pages[i] = NULL;
2338  			continue;
2339  		}
2340  
2341  		if (!folio_make_device_exclusive(folio, mm, start, owner)) {
2342  			folio_unlock(folio);
2343  			folio_put(folio);
2344  			pages[i] = NULL;
2345  		}
2346  	}
2347  
2348  	return npages;
2349  }
2350  EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2351  #endif
2352  
2353  void __put_anon_vma(struct anon_vma *anon_vma)
2354  {
2355  	struct anon_vma *root = anon_vma->root;
2356  
2357  	anon_vma_free(anon_vma);
2358  	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2359  		anon_vma_free(root);
2360  }
2361  
2362  static struct anon_vma *rmap_walk_anon_lock(struct folio *folio,
2363  					    struct rmap_walk_control *rwc)
2364  {
2365  	struct anon_vma *anon_vma;
2366  
2367  	if (rwc->anon_lock)
2368  		return rwc->anon_lock(folio, rwc);
2369  
2370  	/*
2371  	 * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read()
2372  	 * because that depends on page_mapped(); but not all its usages
2373  	 * are holding mmap_lock. Users without mmap_lock are required to
2374  	 * take a reference count to prevent the anon_vma disappearing
2375  	 */
2376  	anon_vma = folio_anon_vma(folio);
2377  	if (!anon_vma)
2378  		return NULL;
2379  
2380  	if (anon_vma_trylock_read(anon_vma))
2381  		goto out;
2382  
2383  	if (rwc->try_lock) {
2384  		anon_vma = NULL;
2385  		rwc->contended = true;
2386  		goto out;
2387  	}
2388  
2389  	anon_vma_lock_read(anon_vma);
2390  out:
2391  	return anon_vma;
2392  }
2393  
2394  /*
2395   * rmap_walk_anon - do something to anonymous page using the object-based
2396   * rmap method
2397   * @page: the page to be handled
2398   * @rwc: control variable according to each walk type
2399   *
2400   * Find all the mappings of a page using the mapping pointer and the vma chains
2401   * contained in the anon_vma struct it points to.
2402   */
2403  static void rmap_walk_anon(struct folio *folio,
2404  		struct rmap_walk_control *rwc, bool locked)
2405  {
2406  	struct anon_vma *anon_vma;
2407  	pgoff_t pgoff_start, pgoff_end;
2408  	struct anon_vma_chain *avc;
2409  
2410  	if (locked) {
2411  		anon_vma = folio_anon_vma(folio);
2412  		/* anon_vma disappear under us? */
2413  		VM_BUG_ON_FOLIO(!anon_vma, folio);
2414  	} else {
2415  		anon_vma = rmap_walk_anon_lock(folio, rwc);
2416  	}
2417  	if (!anon_vma)
2418  		return;
2419  
2420  	pgoff_start = folio_pgoff(folio);
2421  	pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2422  	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2423  			pgoff_start, pgoff_end) {
2424  		struct vm_area_struct *vma = avc->vma;
2425  		unsigned long address = vma_address(&folio->page, vma);
2426  
2427  		VM_BUG_ON_VMA(address == -EFAULT, vma);
2428  		cond_resched();
2429  
2430  		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2431  			continue;
2432  
2433  		if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2434  			break;
2435  		if (rwc->done && rwc->done(folio))
2436  			break;
2437  	}
2438  
2439  	if (!locked)
2440  		anon_vma_unlock_read(anon_vma);
2441  }
2442  
2443  /*
2444   * rmap_walk_file - do something to file page using the object-based rmap method
2445   * @page: the page to be handled
2446   * @rwc: control variable according to each walk type
2447   *
2448   * Find all the mappings of a page using the mapping pointer and the vma chains
2449   * contained in the address_space struct it points to.
2450   */
2451  static void rmap_walk_file(struct folio *folio,
2452  		struct rmap_walk_control *rwc, bool locked)
2453  {
2454  	struct address_space *mapping = folio_mapping(folio);
2455  	pgoff_t pgoff_start, pgoff_end;
2456  	struct vm_area_struct *vma;
2457  
2458  	/*
2459  	 * The page lock not only makes sure that page->mapping cannot
2460  	 * suddenly be NULLified by truncation, it makes sure that the
2461  	 * structure at mapping cannot be freed and reused yet,
2462  	 * so we can safely take mapping->i_mmap_rwsem.
2463  	 */
2464  	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2465  
2466  	if (!mapping)
2467  		return;
2468  
2469  	pgoff_start = folio_pgoff(folio);
2470  	pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2471  	if (!locked) {
2472  		if (i_mmap_trylock_read(mapping))
2473  			goto lookup;
2474  
2475  		if (rwc->try_lock) {
2476  			rwc->contended = true;
2477  			return;
2478  		}
2479  
2480  		i_mmap_lock_read(mapping);
2481  	}
2482  lookup:
2483  	vma_interval_tree_foreach(vma, &mapping->i_mmap,
2484  			pgoff_start, pgoff_end) {
2485  		unsigned long address = vma_address(&folio->page, vma);
2486  
2487  		VM_BUG_ON_VMA(address == -EFAULT, vma);
2488  		cond_resched();
2489  
2490  		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2491  			continue;
2492  
2493  		if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2494  			goto done;
2495  		if (rwc->done && rwc->done(folio))
2496  			goto done;
2497  	}
2498  
2499  done:
2500  	if (!locked)
2501  		i_mmap_unlock_read(mapping);
2502  }
2503  
2504  void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc)
2505  {
2506  	if (unlikely(folio_test_ksm(folio)))
2507  		rmap_walk_ksm(folio, rwc);
2508  	else if (folio_test_anon(folio))
2509  		rmap_walk_anon(folio, rwc, false);
2510  	else
2511  		rmap_walk_file(folio, rwc, false);
2512  }
2513  
2514  /* Like rmap_walk, but caller holds relevant rmap lock */
2515  void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc)
2516  {
2517  	/* no ksm support for now */
2518  	VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio);
2519  	if (folio_test_anon(folio))
2520  		rmap_walk_anon(folio, rwc, true);
2521  	else
2522  		rmap_walk_file(folio, rwc, true);
2523  }
2524  
2525  #ifdef CONFIG_HUGETLB_PAGE
2526  /*
2527   * The following two functions are for anonymous (private mapped) hugepages.
2528   * Unlike common anonymous pages, anonymous hugepages have no accounting code
2529   * and no lru code, because we handle hugepages differently from common pages.
2530   *
2531   * RMAP_COMPOUND is ignored.
2532   */
2533  void hugepage_add_anon_rmap(struct page *page, struct vm_area_struct *vma,
2534  			    unsigned long address, rmap_t flags)
2535  {
2536  	struct folio *folio = page_folio(page);
2537  	struct anon_vma *anon_vma = vma->anon_vma;
2538  	int first;
2539  
2540  	BUG_ON(!folio_test_locked(folio));
2541  	BUG_ON(!anon_vma);
2542  	/* address might be in next vma when migration races vma_merge */
2543  	first = atomic_inc_and_test(&folio->_entire_mapcount);
2544  	VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
2545  	VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
2546  	if (first)
2547  		__page_set_anon_rmap(folio, page, vma, address,
2548  				     !!(flags & RMAP_EXCLUSIVE));
2549  }
2550  
2551  void hugepage_add_new_anon_rmap(struct folio *folio,
2552  			struct vm_area_struct *vma, unsigned long address)
2553  {
2554  	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2555  	/* increment count (starts at -1) */
2556  	atomic_set(&folio->_entire_mapcount, 0);
2557  	folio_clear_hugetlb_restore_reserve(folio);
2558  	__page_set_anon_rmap(folio, &folio->page, vma, address, 1);
2559  }
2560  #endif /* CONFIG_HUGETLB_PAGE */
2561