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 
anon_vma_alloc(void)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 
anon_vma_free(struct anon_vma * anon_vma)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 
anon_vma_chain_alloc(gfp_t gfp)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 
anon_vma_chain_free(struct anon_vma_chain * anon_vma_chain)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 
anon_vma_chain_link(struct vm_area_struct * vma,struct anon_vma_chain * avc,struct anon_vma * anon_vma)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  */
__anon_vma_prepare(struct vm_area_struct * vma)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  */
lock_anon_vma_root(struct anon_vma * root,struct anon_vma * anon_vma)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 
unlock_anon_vma_root(struct anon_vma * root)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_adjust(), __split_vma(), copy_vma() and
266  * anon_vma_fork(). The first three want an exact copy of src, while the last
267  * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
268  * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
269  * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
270  *
271  * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
272  * and reuse existing anon_vma which has no vmas and only one child anon_vma.
273  * This prevents degradation of anon_vma hierarchy to endless linear chain in
274  * case of constantly forking task. On the other hand, an anon_vma with more
275  * than one child isn't reused even if there was no alive vma, thus rmap
276  * walker has a good chance of avoiding scanning the whole hierarchy when it
277  * searches where page is mapped.
278  */
anon_vma_clone(struct vm_area_struct * dst,struct vm_area_struct * src)279 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
280 {
281 	struct anon_vma_chain *avc, *pavc;
282 	struct anon_vma *root = NULL;
283 
284 	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
285 		struct anon_vma *anon_vma;
286 
287 		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
288 		if (unlikely(!avc)) {
289 			unlock_anon_vma_root(root);
290 			root = NULL;
291 			avc = anon_vma_chain_alloc(GFP_KERNEL);
292 			if (!avc)
293 				goto enomem_failure;
294 		}
295 		anon_vma = pavc->anon_vma;
296 		root = lock_anon_vma_root(root, anon_vma);
297 		anon_vma_chain_link(dst, avc, anon_vma);
298 
299 		/*
300 		 * Reuse existing anon_vma if it has no vma and only one
301 		 * anon_vma child.
302 		 *
303 		 * Root anon_vma is never reused:
304 		 * it has self-parent reference and at least one child.
305 		 */
306 		if (!dst->anon_vma && src->anon_vma &&
307 		    anon_vma->num_children < 2 &&
308 		    anon_vma->num_active_vmas == 0)
309 			dst->anon_vma = anon_vma;
310 	}
311 	if (dst->anon_vma)
312 		dst->anon_vma->num_active_vmas++;
313 	unlock_anon_vma_root(root);
314 	return 0;
315 
316  enomem_failure:
317 	/*
318 	 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
319 	 * decremented in unlink_anon_vmas().
320 	 * We can safely do this because callers of anon_vma_clone() don't care
321 	 * about dst->anon_vma if anon_vma_clone() failed.
322 	 */
323 	dst->anon_vma = NULL;
324 	unlink_anon_vmas(dst);
325 	return -ENOMEM;
326 }
327 
328 /*
329  * Attach vma to its own anon_vma, as well as to the anon_vmas that
330  * the corresponding VMA in the parent process is attached to.
331  * Returns 0 on success, non-zero on failure.
332  */
anon_vma_fork(struct vm_area_struct * vma,struct vm_area_struct * pvma)333 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
334 {
335 	struct anon_vma_chain *avc;
336 	struct anon_vma *anon_vma;
337 	int error;
338 
339 	/* Don't bother if the parent process has no anon_vma here. */
340 	if (!pvma->anon_vma)
341 		return 0;
342 
343 	/* Drop inherited anon_vma, we'll reuse existing or allocate new. */
344 	vma->anon_vma = NULL;
345 
346 	/*
347 	 * First, attach the new VMA to the parent VMA's anon_vmas,
348 	 * so rmap can find non-COWed pages in child processes.
349 	 */
350 	error = anon_vma_clone(vma, pvma);
351 	if (error)
352 		return error;
353 
354 	/* An existing anon_vma has been reused, all done then. */
355 	if (vma->anon_vma)
356 		return 0;
357 
358 	/* Then add our own anon_vma. */
359 	anon_vma = anon_vma_alloc();
360 	if (!anon_vma)
361 		goto out_error;
362 	anon_vma->num_active_vmas++;
363 	avc = anon_vma_chain_alloc(GFP_KERNEL);
364 	if (!avc)
365 		goto out_error_free_anon_vma;
366 
367 	/*
368 	 * The root anon_vma's rwsem is the lock actually used when we
369 	 * lock any of the anon_vmas in this anon_vma tree.
370 	 */
371 	anon_vma->root = pvma->anon_vma->root;
372 	anon_vma->parent = pvma->anon_vma;
373 	/*
374 	 * With refcounts, an anon_vma can stay around longer than the
375 	 * process it belongs to. The root anon_vma needs to be pinned until
376 	 * this anon_vma is freed, because the lock lives in the root.
377 	 */
378 	get_anon_vma(anon_vma->root);
379 	/* Mark this anon_vma as the one where our new (COWed) pages go. */
380 	vma->anon_vma = anon_vma;
381 	anon_vma_lock_write(anon_vma);
382 	anon_vma_chain_link(vma, avc, anon_vma);
383 	anon_vma->parent->num_children++;
384 	anon_vma_unlock_write(anon_vma);
385 
386 	return 0;
387 
388  out_error_free_anon_vma:
389 	put_anon_vma(anon_vma);
390  out_error:
391 	unlink_anon_vmas(vma);
392 	return -ENOMEM;
393 }
394 
unlink_anon_vmas(struct vm_area_struct * vma)395 void unlink_anon_vmas(struct vm_area_struct *vma)
396 {
397 	struct anon_vma_chain *avc, *next;
398 	struct anon_vma *root = NULL;
399 
400 	/*
401 	 * Unlink each anon_vma chained to the VMA.  This list is ordered
402 	 * from newest to oldest, ensuring the root anon_vma gets freed last.
403 	 */
404 	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
405 		struct anon_vma *anon_vma = avc->anon_vma;
406 
407 		root = lock_anon_vma_root(root, anon_vma);
408 		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
409 
410 		/*
411 		 * Leave empty anon_vmas on the list - we'll need
412 		 * to free them outside the lock.
413 		 */
414 		if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
415 			anon_vma->parent->num_children--;
416 			continue;
417 		}
418 
419 		list_del(&avc->same_vma);
420 		anon_vma_chain_free(avc);
421 	}
422 	if (vma->anon_vma) {
423 		vma->anon_vma->num_active_vmas--;
424 
425 		/*
426 		 * vma would still be needed after unlink, and anon_vma will be prepared
427 		 * when handle fault.
428 		 */
429 		vma->anon_vma = NULL;
430 	}
431 	unlock_anon_vma_root(root);
432 
433 	/*
434 	 * Iterate the list once more, it now only contains empty and unlinked
435 	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
436 	 * needing to write-acquire the anon_vma->root->rwsem.
437 	 */
438 	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
439 		struct anon_vma *anon_vma = avc->anon_vma;
440 
441 		VM_WARN_ON(anon_vma->num_children);
442 		VM_WARN_ON(anon_vma->num_active_vmas);
443 		put_anon_vma(anon_vma);
444 
445 		list_del(&avc->same_vma);
446 		anon_vma_chain_free(avc);
447 	}
448 }
449 
anon_vma_ctor(void * data)450 static void anon_vma_ctor(void *data)
451 {
452 	struct anon_vma *anon_vma = data;
453 
454 	init_rwsem(&anon_vma->rwsem);
455 	atomic_set(&anon_vma->refcount, 0);
456 	anon_vma->rb_root = RB_ROOT_CACHED;
457 }
458 
anon_vma_init(void)459 void __init anon_vma_init(void)
460 {
461 	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
462 			0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
463 			anon_vma_ctor);
464 	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
465 			SLAB_PANIC|SLAB_ACCOUNT);
466 }
467 
468 /*
469  * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
470  *
471  * Since there is no serialization what so ever against page_remove_rmap()
472  * the best this function can do is return a refcount increased anon_vma
473  * that might have been relevant to this page.
474  *
475  * The page might have been remapped to a different anon_vma or the anon_vma
476  * returned may already be freed (and even reused).
477  *
478  * In case it was remapped to a different anon_vma, the new anon_vma will be a
479  * child of the old anon_vma, and the anon_vma lifetime rules will therefore
480  * ensure that any anon_vma obtained from the page will still be valid for as
481  * long as we observe page_mapped() [ hence all those page_mapped() tests ].
482  *
483  * All users of this function must be very careful when walking the anon_vma
484  * chain and verify that the page in question is indeed mapped in it
485  * [ something equivalent to page_mapped_in_vma() ].
486  *
487  * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
488  * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
489  * if there is a mapcount, we can dereference the anon_vma after observing
490  * those.
491  */
folio_get_anon_vma(struct folio * folio)492 struct anon_vma *folio_get_anon_vma(struct folio *folio)
493 {
494 	struct anon_vma *anon_vma = NULL;
495 	unsigned long anon_mapping;
496 
497 	rcu_read_lock();
498 	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
499 	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
500 		goto out;
501 	if (!folio_mapped(folio))
502 		goto out;
503 
504 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
505 	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
506 		anon_vma = NULL;
507 		goto out;
508 	}
509 
510 	/*
511 	 * If this folio is still mapped, then its anon_vma cannot have been
512 	 * freed.  But if it has been unmapped, we have no security against the
513 	 * anon_vma structure being freed and reused (for another anon_vma:
514 	 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
515 	 * above cannot corrupt).
516 	 */
517 	if (!folio_mapped(folio)) {
518 		rcu_read_unlock();
519 		put_anon_vma(anon_vma);
520 		return NULL;
521 	}
522 out:
523 	rcu_read_unlock();
524 
525 	return anon_vma;
526 }
527 
528 /*
529  * Similar to folio_get_anon_vma() except it locks the anon_vma.
530  *
531  * Its a little more complex as it tries to keep the fast path to a single
532  * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
533  * reference like with folio_get_anon_vma() and then block on the mutex
534  * on !rwc->try_lock case.
535  */
folio_lock_anon_vma_read(struct folio * folio,struct rmap_walk_control * rwc)536 struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
537 					  struct rmap_walk_control *rwc)
538 {
539 	struct anon_vma *anon_vma = NULL;
540 	struct anon_vma *root_anon_vma;
541 	unsigned long anon_mapping;
542 
543 	rcu_read_lock();
544 	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
545 	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
546 		goto out;
547 	if (!folio_mapped(folio))
548 		goto out;
549 
550 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
551 	root_anon_vma = READ_ONCE(anon_vma->root);
552 	if (down_read_trylock(&root_anon_vma->rwsem)) {
553 		/*
554 		 * If the folio is still mapped, then this anon_vma is still
555 		 * its anon_vma, and holding the mutex ensures that it will
556 		 * not go away, see anon_vma_free().
557 		 */
558 		if (!folio_mapped(folio)) {
559 			up_read(&root_anon_vma->rwsem);
560 			anon_vma = NULL;
561 		}
562 		goto out;
563 	}
564 
565 	if (rwc && rwc->try_lock) {
566 		anon_vma = NULL;
567 		rwc->contended = true;
568 		goto out;
569 	}
570 
571 	/* trylock failed, we got to sleep */
572 	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
573 		anon_vma = NULL;
574 		goto out;
575 	}
576 
577 	if (!folio_mapped(folio)) {
578 		rcu_read_unlock();
579 		put_anon_vma(anon_vma);
580 		return NULL;
581 	}
582 
583 	/* we pinned the anon_vma, its safe to sleep */
584 	rcu_read_unlock();
585 	anon_vma_lock_read(anon_vma);
586 
587 	if (atomic_dec_and_test(&anon_vma->refcount)) {
588 		/*
589 		 * Oops, we held the last refcount, release the lock
590 		 * and bail -- can't simply use put_anon_vma() because
591 		 * we'll deadlock on the anon_vma_lock_write() recursion.
592 		 */
593 		anon_vma_unlock_read(anon_vma);
594 		__put_anon_vma(anon_vma);
595 		anon_vma = NULL;
596 	}
597 
598 	return anon_vma;
599 
600 out:
601 	rcu_read_unlock();
602 	return anon_vma;
603 }
604 
605 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
606 /*
607  * Flush TLB entries for recently unmapped pages from remote CPUs. It is
608  * important if a PTE was dirty when it was unmapped that it's flushed
609  * before any IO is initiated on the page to prevent lost writes. Similarly,
610  * it must be flushed before freeing to prevent data leakage.
611  */
try_to_unmap_flush(void)612 void try_to_unmap_flush(void)
613 {
614 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
615 
616 	if (!tlb_ubc->flush_required)
617 		return;
618 
619 	arch_tlbbatch_flush(&tlb_ubc->arch);
620 	tlb_ubc->flush_required = false;
621 	tlb_ubc->writable = false;
622 }
623 
624 /* Flush iff there are potentially writable TLB entries that can race with IO */
try_to_unmap_flush_dirty(void)625 void try_to_unmap_flush_dirty(void)
626 {
627 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
628 
629 	if (tlb_ubc->writable)
630 		try_to_unmap_flush();
631 }
632 
633 /*
634  * Bits 0-14 of mm->tlb_flush_batched record pending generations.
635  * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations.
636  */
637 #define TLB_FLUSH_BATCH_FLUSHED_SHIFT	16
638 #define TLB_FLUSH_BATCH_PENDING_MASK			\
639 	((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1)
640 #define TLB_FLUSH_BATCH_PENDING_LARGE			\
641 	(TLB_FLUSH_BATCH_PENDING_MASK / 2)
642 
set_tlb_ubc_flush_pending(struct mm_struct * mm,bool writable)643 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
644 {
645 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
646 	int batch, nbatch;
647 
648 	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
649 	tlb_ubc->flush_required = true;
650 
651 	/*
652 	 * Ensure compiler does not re-order the setting of tlb_flush_batched
653 	 * before the PTE is cleared.
654 	 */
655 	barrier();
656 	batch = atomic_read(&mm->tlb_flush_batched);
657 retry:
658 	if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) {
659 		/*
660 		 * Prevent `pending' from catching up with `flushed' because of
661 		 * overflow.  Reset `pending' and `flushed' to be 1 and 0 if
662 		 * `pending' becomes large.
663 		 */
664 		nbatch = atomic_cmpxchg(&mm->tlb_flush_batched, batch, 1);
665 		if (nbatch != batch) {
666 			batch = nbatch;
667 			goto retry;
668 		}
669 	} else {
670 		atomic_inc(&mm->tlb_flush_batched);
671 	}
672 
673 	/*
674 	 * If the PTE was dirty then it's best to assume it's writable. The
675 	 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
676 	 * before the page is queued for IO.
677 	 */
678 	if (writable)
679 		tlb_ubc->writable = true;
680 }
681 
682 /*
683  * Returns true if the TLB flush should be deferred to the end of a batch of
684  * unmap operations to reduce IPIs.
685  */
should_defer_flush(struct mm_struct * mm,enum ttu_flags flags)686 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
687 {
688 	bool should_defer = false;
689 
690 	if (!(flags & TTU_BATCH_FLUSH))
691 		return false;
692 
693 	/* If remote CPUs need to be flushed then defer batch the flush */
694 	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
695 		should_defer = true;
696 	put_cpu();
697 
698 	return should_defer;
699 }
700 
701 /*
702  * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
703  * releasing the PTL if TLB flushes are batched. It's possible for a parallel
704  * operation such as mprotect or munmap to race between reclaim unmapping
705  * the page and flushing the page. If this race occurs, it potentially allows
706  * access to data via a stale TLB entry. Tracking all mm's that have TLB
707  * batching in flight would be expensive during reclaim so instead track
708  * whether TLB batching occurred in the past and if so then do a flush here
709  * if required. This will cost one additional flush per reclaim cycle paid
710  * by the first operation at risk such as mprotect and mumap.
711  *
712  * This must be called under the PTL so that an access to tlb_flush_batched
713  * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
714  * via the PTL.
715  */
flush_tlb_batched_pending(struct mm_struct * mm)716 void flush_tlb_batched_pending(struct mm_struct *mm)
717 {
718 	int batch = atomic_read(&mm->tlb_flush_batched);
719 	int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK;
720 	int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT;
721 
722 	if (pending != flushed) {
723 		flush_tlb_mm(mm);
724 		/*
725 		 * If the new TLB flushing is pending during flushing, leave
726 		 * mm->tlb_flush_batched as is, to avoid losing flushing.
727 		 */
728 		atomic_cmpxchg(&mm->tlb_flush_batched, batch,
729 			       pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT));
730 	}
731 }
732 #else
set_tlb_ubc_flush_pending(struct mm_struct * mm,bool writable)733 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
734 {
735 }
736 
should_defer_flush(struct mm_struct * mm,enum ttu_flags flags)737 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
738 {
739 	return false;
740 }
741 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
742 
743 /*
744  * At what user virtual address is page expected in vma?
745  * Caller should check the page is actually part of the vma.
746  */
page_address_in_vma(struct page * page,struct vm_area_struct * vma)747 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
748 {
749 	struct folio *folio = page_folio(page);
750 	if (folio_test_anon(folio)) {
751 		struct anon_vma *page__anon_vma = folio_anon_vma(folio);
752 		/*
753 		 * Note: swapoff's unuse_vma() is more efficient with this
754 		 * check, and needs it to match anon_vma when KSM is active.
755 		 */
756 		if (!vma->anon_vma || !page__anon_vma ||
757 		    vma->anon_vma->root != page__anon_vma->root)
758 			return -EFAULT;
759 	} else if (!vma->vm_file) {
760 		return -EFAULT;
761 	} else if (vma->vm_file->f_mapping != folio->mapping) {
762 		return -EFAULT;
763 	}
764 
765 	return vma_address(page, vma);
766 }
767 
768 /*
769  * Returns the actual pmd_t* where we expect 'address' to be mapped from, or
770  * NULL if it doesn't exist.  No guarantees / checks on what the pmd_t*
771  * represents.
772  */
mm_find_pmd(struct mm_struct * mm,unsigned long address)773 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
774 {
775 	pgd_t *pgd;
776 	p4d_t *p4d;
777 	pud_t *pud;
778 	pmd_t *pmd = NULL;
779 
780 	pgd = pgd_offset(mm, address);
781 	if (!pgd_present(*pgd))
782 		goto out;
783 
784 	p4d = p4d_offset(pgd, address);
785 	if (!p4d_present(*p4d))
786 		goto out;
787 
788 	pud = pud_offset(p4d, address);
789 	if (!pud_present(*pud))
790 		goto out;
791 
792 	pmd = pmd_offset(pud, address);
793 out:
794 	return pmd;
795 }
796 
797 struct folio_referenced_arg {
798 	int mapcount;
799 	int referenced;
800 	unsigned long vm_flags;
801 	struct mem_cgroup *memcg;
802 };
803 /*
804  * arg: folio_referenced_arg will be passed
805  */
folio_referenced_one(struct folio * folio,struct vm_area_struct * vma,unsigned long address,void * arg)806 static bool folio_referenced_one(struct folio *folio,
807 		struct vm_area_struct *vma, unsigned long address, void *arg)
808 {
809 	struct folio_referenced_arg *pra = arg;
810 	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
811 	int referenced = 0;
812 
813 	while (page_vma_mapped_walk(&pvmw)) {
814 		address = pvmw.address;
815 
816 		if ((vma->vm_flags & VM_LOCKED) &&
817 		    (!folio_test_large(folio) || !pvmw.pte)) {
818 			/* Restore the mlock which got missed */
819 			mlock_vma_folio(folio, vma, !pvmw.pte);
820 			page_vma_mapped_walk_done(&pvmw);
821 			pra->vm_flags |= VM_LOCKED;
822 			return false; /* To break the loop */
823 		}
824 
825 		if (pvmw.pte) {
826 			if (lru_gen_enabled() && pte_young(*pvmw.pte) &&
827 			    !(vma->vm_flags & (VM_SEQ_READ | VM_RAND_READ))) {
828 				lru_gen_look_around(&pvmw);
829 				referenced++;
830 			}
831 
832 			if (ptep_clear_flush_young_notify(vma, address,
833 						pvmw.pte)) {
834 				/*
835 				 * Don't treat a reference through
836 				 * a sequentially read mapping as such.
837 				 * If the folio has been used in another mapping,
838 				 * we will catch it; if this other mapping is
839 				 * already gone, the unmap path will have set
840 				 * the referenced flag or activated the folio.
841 				 */
842 				if (likely(!(vma->vm_flags & VM_SEQ_READ)))
843 					referenced++;
844 			}
845 		} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
846 			if (pmdp_clear_flush_young_notify(vma, address,
847 						pvmw.pmd))
848 				referenced++;
849 		} else {
850 			/* unexpected pmd-mapped folio? */
851 			WARN_ON_ONCE(1);
852 		}
853 
854 		pra->mapcount--;
855 	}
856 
857 	if (referenced)
858 		folio_clear_idle(folio);
859 	if (folio_test_clear_young(folio))
860 		referenced++;
861 
862 	if (referenced) {
863 		pra->referenced++;
864 		pra->vm_flags |= vma->vm_flags & ~VM_LOCKED;
865 	}
866 
867 	if (!pra->mapcount)
868 		return false; /* To break the loop */
869 
870 	return true;
871 }
872 
invalid_folio_referenced_vma(struct vm_area_struct * vma,void * arg)873 static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg)
874 {
875 	struct folio_referenced_arg *pra = arg;
876 	struct mem_cgroup *memcg = pra->memcg;
877 
878 	if (!mm_match_cgroup(vma->vm_mm, memcg))
879 		return true;
880 
881 	return false;
882 }
883 
884 /**
885  * folio_referenced() - Test if the folio was referenced.
886  * @folio: The folio to test.
887  * @is_locked: Caller holds lock on the folio.
888  * @memcg: target memory cgroup
889  * @vm_flags: A combination of all the vma->vm_flags which referenced the folio.
890  *
891  * Quick test_and_clear_referenced for all mappings of a folio,
892  *
893  * Return: The number of mappings which referenced the folio. Return -1 if
894  * the function bailed out due to rmap lock contention.
895  */
folio_referenced(struct folio * folio,int is_locked,struct mem_cgroup * memcg,unsigned long * vm_flags)896 int folio_referenced(struct folio *folio, int is_locked,
897 		     struct mem_cgroup *memcg, unsigned long *vm_flags)
898 {
899 	int we_locked = 0;
900 	struct folio_referenced_arg pra = {
901 		.mapcount = folio_mapcount(folio),
902 		.memcg = memcg,
903 	};
904 	struct rmap_walk_control rwc = {
905 		.rmap_one = folio_referenced_one,
906 		.arg = (void *)&pra,
907 		.anon_lock = folio_lock_anon_vma_read,
908 		.try_lock = true,
909 	};
910 
911 	*vm_flags = 0;
912 	if (!pra.mapcount)
913 		return 0;
914 
915 	if (!folio_raw_mapping(folio))
916 		return 0;
917 
918 	if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) {
919 		we_locked = folio_trylock(folio);
920 		if (!we_locked)
921 			return 1;
922 	}
923 
924 	/*
925 	 * If we are reclaiming on behalf of a cgroup, skip
926 	 * counting on behalf of references from different
927 	 * cgroups
928 	 */
929 	if (memcg) {
930 		rwc.invalid_vma = invalid_folio_referenced_vma;
931 	}
932 
933 	rmap_walk(folio, &rwc);
934 	*vm_flags = pra.vm_flags;
935 
936 	if (we_locked)
937 		folio_unlock(folio);
938 
939 	return rwc.contended ? -1 : pra.referenced;
940 }
941 
page_vma_mkclean_one(struct page_vma_mapped_walk * pvmw)942 static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw)
943 {
944 	int cleaned = 0;
945 	struct vm_area_struct *vma = pvmw->vma;
946 	struct mmu_notifier_range range;
947 	unsigned long address = pvmw->address;
948 
949 	/*
950 	 * We have to assume the worse case ie pmd for invalidation. Note that
951 	 * the folio can not be freed from this function.
952 	 */
953 	mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
954 				0, vma, vma->vm_mm, address,
955 				vma_address_end(pvmw));
956 	mmu_notifier_invalidate_range_start(&range);
957 
958 	while (page_vma_mapped_walk(pvmw)) {
959 		int ret = 0;
960 
961 		address = pvmw->address;
962 		if (pvmw->pte) {
963 			pte_t entry;
964 			pte_t *pte = pvmw->pte;
965 
966 			if (!pte_dirty(*pte) && !pte_write(*pte))
967 				continue;
968 
969 			flush_cache_page(vma, address, pte_pfn(*pte));
970 			entry = ptep_clear_flush(vma, address, pte);
971 			entry = pte_wrprotect(entry);
972 			entry = pte_mkclean(entry);
973 			set_pte_at(vma->vm_mm, address, pte, entry);
974 			ret = 1;
975 		} else {
976 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
977 			pmd_t *pmd = pvmw->pmd;
978 			pmd_t entry;
979 
980 			if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
981 				continue;
982 
983 			flush_cache_range(vma, address,
984 					  address + HPAGE_PMD_SIZE);
985 			entry = pmdp_invalidate(vma, address, pmd);
986 			entry = pmd_wrprotect(entry);
987 			entry = pmd_mkclean(entry);
988 			set_pmd_at(vma->vm_mm, address, pmd, entry);
989 			ret = 1;
990 #else
991 			/* unexpected pmd-mapped folio? */
992 			WARN_ON_ONCE(1);
993 #endif
994 		}
995 
996 		/*
997 		 * No need to call mmu_notifier_invalidate_range() as we are
998 		 * downgrading page table protection not changing it to point
999 		 * to a new page.
1000 		 *
1001 		 * See Documentation/mm/mmu_notifier.rst
1002 		 */
1003 		if (ret)
1004 			cleaned++;
1005 	}
1006 
1007 	mmu_notifier_invalidate_range_end(&range);
1008 
1009 	return cleaned;
1010 }
1011 
page_mkclean_one(struct folio * folio,struct vm_area_struct * vma,unsigned long address,void * arg)1012 static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma,
1013 			     unsigned long address, void *arg)
1014 {
1015 	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC);
1016 	int *cleaned = arg;
1017 
1018 	*cleaned += page_vma_mkclean_one(&pvmw);
1019 
1020 	return true;
1021 }
1022 
invalid_mkclean_vma(struct vm_area_struct * vma,void * arg)1023 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1024 {
1025 	if (vma->vm_flags & VM_SHARED)
1026 		return false;
1027 
1028 	return true;
1029 }
1030 
folio_mkclean(struct folio * folio)1031 int folio_mkclean(struct folio *folio)
1032 {
1033 	int cleaned = 0;
1034 	struct address_space *mapping;
1035 	struct rmap_walk_control rwc = {
1036 		.arg = (void *)&cleaned,
1037 		.rmap_one = page_mkclean_one,
1038 		.invalid_vma = invalid_mkclean_vma,
1039 	};
1040 
1041 	BUG_ON(!folio_test_locked(folio));
1042 
1043 	if (!folio_mapped(folio))
1044 		return 0;
1045 
1046 	mapping = folio_mapping(folio);
1047 	if (!mapping)
1048 		return 0;
1049 
1050 	rmap_walk(folio, &rwc);
1051 
1052 	return cleaned;
1053 }
1054 EXPORT_SYMBOL_GPL(folio_mkclean);
1055 
1056 /**
1057  * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of
1058  *                     [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff)
1059  *                     within the @vma of shared mappings. And since clean PTEs
1060  *                     should also be readonly, write protects them too.
1061  * @pfn: start pfn.
1062  * @nr_pages: number of physically contiguous pages srarting with @pfn.
1063  * @pgoff: page offset that the @pfn mapped with.
1064  * @vma: vma that @pfn mapped within.
1065  *
1066  * Returns the number of cleaned PTEs (including PMDs).
1067  */
pfn_mkclean_range(unsigned long pfn,unsigned long nr_pages,pgoff_t pgoff,struct vm_area_struct * vma)1068 int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff,
1069 		      struct vm_area_struct *vma)
1070 {
1071 	struct page_vma_mapped_walk pvmw = {
1072 		.pfn		= pfn,
1073 		.nr_pages	= nr_pages,
1074 		.pgoff		= pgoff,
1075 		.vma		= vma,
1076 		.flags		= PVMW_SYNC,
1077 	};
1078 
1079 	if (invalid_mkclean_vma(vma, NULL))
1080 		return 0;
1081 
1082 	pvmw.address = vma_pgoff_address(pgoff, nr_pages, vma);
1083 	VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma);
1084 
1085 	return page_vma_mkclean_one(&pvmw);
1086 }
1087 
1088 /**
1089  * page_move_anon_rmap - move a page to our anon_vma
1090  * @page:	the page to move to our anon_vma
1091  * @vma:	the vma the page belongs to
1092  *
1093  * When a page belongs exclusively to one process after a COW event,
1094  * that page can be moved into the anon_vma that belongs to just that
1095  * process, so the rmap code will not search the parent or sibling
1096  * processes.
1097  */
page_move_anon_rmap(struct page * page,struct vm_area_struct * vma)1098 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1099 {
1100 	void *anon_vma = vma->anon_vma;
1101 	struct folio *folio = page_folio(page);
1102 
1103 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1104 	VM_BUG_ON_VMA(!anon_vma, vma);
1105 
1106 	anon_vma += PAGE_MAPPING_ANON;
1107 	/*
1108 	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1109 	 * simultaneously, so a concurrent reader (eg folio_referenced()'s
1110 	 * folio_test_anon()) will not see one without the other.
1111 	 */
1112 	WRITE_ONCE(folio->mapping, anon_vma);
1113 	SetPageAnonExclusive(page);
1114 }
1115 
1116 /**
1117  * __page_set_anon_rmap - set up new anonymous rmap
1118  * @page:	Page or Hugepage to add to rmap
1119  * @vma:	VM area to add page to.
1120  * @address:	User virtual address of the mapping
1121  * @exclusive:	the page is exclusively owned by the current process
1122  */
__page_set_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address,int exclusive)1123 static void __page_set_anon_rmap(struct page *page,
1124 	struct vm_area_struct *vma, unsigned long address, int exclusive)
1125 {
1126 	struct anon_vma *anon_vma = vma->anon_vma;
1127 
1128 	BUG_ON(!anon_vma);
1129 
1130 	if (PageAnon(page))
1131 		goto out;
1132 
1133 	/*
1134 	 * If the page isn't exclusively mapped into this vma,
1135 	 * we must use the _oldest_ possible anon_vma for the
1136 	 * page mapping!
1137 	 */
1138 	if (!exclusive)
1139 		anon_vma = anon_vma->root;
1140 
1141 	/*
1142 	 * page_idle does a lockless/optimistic rmap scan on page->mapping.
1143 	 * Make sure the compiler doesn't split the stores of anon_vma and
1144 	 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1145 	 * could mistake the mapping for a struct address_space and crash.
1146 	 */
1147 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1148 	WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1149 	page->index = linear_page_index(vma, address);
1150 out:
1151 	if (exclusive)
1152 		SetPageAnonExclusive(page);
1153 }
1154 
1155 /**
1156  * __page_check_anon_rmap - sanity check anonymous rmap addition
1157  * @page:	the page to add the mapping to
1158  * @vma:	the vm area in which the mapping is added
1159  * @address:	the user virtual address mapped
1160  */
__page_check_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address)1161 static void __page_check_anon_rmap(struct page *page,
1162 	struct vm_area_struct *vma, unsigned long address)
1163 {
1164 	struct folio *folio = page_folio(page);
1165 	/*
1166 	 * The page's anon-rmap details (mapping and index) are guaranteed to
1167 	 * be set up correctly at this point.
1168 	 *
1169 	 * We have exclusion against page_add_anon_rmap because the caller
1170 	 * always holds the page locked.
1171 	 *
1172 	 * We have exclusion against page_add_new_anon_rmap because those pages
1173 	 * are initially only visible via the pagetables, and the pte is locked
1174 	 * over the call to page_add_new_anon_rmap.
1175 	 */
1176 	VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root,
1177 			folio);
1178 	VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1179 		       page);
1180 }
1181 
1182 /**
1183  * page_add_anon_rmap - add pte mapping to an anonymous page
1184  * @page:	the page to add the mapping to
1185  * @vma:	the vm area in which the mapping is added
1186  * @address:	the user virtual address mapped
1187  * @flags:	the rmap flags
1188  *
1189  * The caller needs to hold the pte lock, and the page must be locked in
1190  * the anon_vma case: to serialize mapping,index checking after setting,
1191  * and to ensure that PageAnon is not being upgraded racily to PageKsm
1192  * (but PageKsm is never downgraded to PageAnon).
1193  */
page_add_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address,rmap_t flags)1194 void page_add_anon_rmap(struct page *page,
1195 	struct vm_area_struct *vma, unsigned long address, rmap_t flags)
1196 {
1197 	bool compound = flags & RMAP_COMPOUND;
1198 	bool first;
1199 
1200 	if (unlikely(PageKsm(page)))
1201 		lock_page_memcg(page);
1202 	else
1203 		VM_BUG_ON_PAGE(!PageLocked(page), page);
1204 
1205 	if (compound) {
1206 		atomic_t *mapcount;
1207 		VM_BUG_ON_PAGE(!PageLocked(page), page);
1208 		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1209 		mapcount = compound_mapcount_ptr(page);
1210 		first = atomic_inc_and_test(mapcount);
1211 	} else {
1212 		first = atomic_inc_and_test(&page->_mapcount);
1213 	}
1214 	VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
1215 	VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
1216 
1217 	if (first) {
1218 		int nr = compound ? thp_nr_pages(page) : 1;
1219 		/*
1220 		 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1221 		 * these counters are not modified in interrupt context, and
1222 		 * pte lock(a spinlock) is held, which implies preemption
1223 		 * disabled.
1224 		 */
1225 		if (compound)
1226 			__mod_lruvec_page_state(page, NR_ANON_THPS, nr);
1227 		__mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1228 	}
1229 
1230 	if (unlikely(PageKsm(page)))
1231 		unlock_page_memcg(page);
1232 
1233 	/* address might be in next vma when migration races vma_adjust */
1234 	else if (first)
1235 		__page_set_anon_rmap(page, vma, address,
1236 				     !!(flags & RMAP_EXCLUSIVE));
1237 	else
1238 		__page_check_anon_rmap(page, vma, address);
1239 
1240 	mlock_vma_page(page, vma, compound);
1241 }
1242 
1243 /**
1244  * page_add_new_anon_rmap - add mapping to a new anonymous page
1245  * @page:	the page to add the mapping to
1246  * @vma:	the vm area in which the mapping is added
1247  * @address:	the user virtual address mapped
1248  *
1249  * If it's a compound page, it is accounted as a compound page. As the page
1250  * is new, it's assume to get mapped exclusively by a single process.
1251  *
1252  * Same as page_add_anon_rmap but must only be called on *new* pages.
1253  * This means the inc-and-test can be bypassed.
1254  * Page does not have to be locked.
1255  */
page_add_new_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address)1256 void page_add_new_anon_rmap(struct page *page,
1257 	struct vm_area_struct *vma, unsigned long address)
1258 {
1259 	const bool compound = PageCompound(page);
1260 	int nr = compound ? thp_nr_pages(page) : 1;
1261 
1262 	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1263 	__SetPageSwapBacked(page);
1264 	if (compound) {
1265 		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1266 		/* increment count (starts at -1) */
1267 		atomic_set(compound_mapcount_ptr(page), 0);
1268 		atomic_set(compound_pincount_ptr(page), 0);
1269 
1270 		__mod_lruvec_page_state(page, NR_ANON_THPS, nr);
1271 	} else {
1272 		/* increment count (starts at -1) */
1273 		atomic_set(&page->_mapcount, 0);
1274 	}
1275 	__mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1276 	__page_set_anon_rmap(page, vma, address, 1);
1277 }
1278 
1279 /**
1280  * page_add_file_rmap - add pte mapping to a file page
1281  * @page:	the page to add the mapping to
1282  * @vma:	the vm area in which the mapping is added
1283  * @compound:	charge the page as compound or small page
1284  *
1285  * The caller needs to hold the pte lock.
1286  */
page_add_file_rmap(struct page * page,struct vm_area_struct * vma,bool compound)1287 void page_add_file_rmap(struct page *page,
1288 	struct vm_area_struct *vma, bool compound)
1289 {
1290 	int i, nr = 0;
1291 
1292 	VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1293 	lock_page_memcg(page);
1294 	if (compound && PageTransHuge(page)) {
1295 		int nr_pages = thp_nr_pages(page);
1296 
1297 		for (i = 0; i < nr_pages; i++) {
1298 			if (atomic_inc_and_test(&page[i]._mapcount))
1299 				nr++;
1300 		}
1301 		if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1302 			goto out;
1303 
1304 		/*
1305 		 * It is racy to ClearPageDoubleMap in page_remove_file_rmap();
1306 		 * but page lock is held by all page_add_file_rmap() compound
1307 		 * callers, and SetPageDoubleMap below warns if !PageLocked:
1308 		 * so here is a place that DoubleMap can be safely cleared.
1309 		 */
1310 		VM_WARN_ON_ONCE(!PageLocked(page));
1311 		if (nr == nr_pages && PageDoubleMap(page))
1312 			ClearPageDoubleMap(page);
1313 
1314 		if (PageSwapBacked(page))
1315 			__mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED,
1316 						nr_pages);
1317 		else
1318 			__mod_lruvec_page_state(page, NR_FILE_PMDMAPPED,
1319 						nr_pages);
1320 	} else {
1321 		if (PageTransCompound(page) && page_mapping(page)) {
1322 			VM_WARN_ON_ONCE(!PageLocked(page));
1323 			SetPageDoubleMap(compound_head(page));
1324 		}
1325 		if (atomic_inc_and_test(&page->_mapcount))
1326 			nr++;
1327 	}
1328 out:
1329 	if (nr)
1330 		__mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1331 	unlock_page_memcg(page);
1332 
1333 	mlock_vma_page(page, vma, compound);
1334 }
1335 
page_remove_file_rmap(struct page * page,bool compound)1336 static void page_remove_file_rmap(struct page *page, bool compound)
1337 {
1338 	int i, nr = 0;
1339 
1340 	VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1341 
1342 	/* Hugepages are not counted in NR_FILE_MAPPED for now. */
1343 	if (unlikely(PageHuge(page))) {
1344 		/* hugetlb pages are always mapped with pmds */
1345 		atomic_dec(compound_mapcount_ptr(page));
1346 		return;
1347 	}
1348 
1349 	/* page still mapped by someone else? */
1350 	if (compound && PageTransHuge(page)) {
1351 		int nr_pages = thp_nr_pages(page);
1352 
1353 		for (i = 0; i < nr_pages; i++) {
1354 			if (atomic_add_negative(-1, &page[i]._mapcount))
1355 				nr++;
1356 		}
1357 		if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1358 			goto out;
1359 		if (PageSwapBacked(page))
1360 			__mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED,
1361 						-nr_pages);
1362 		else
1363 			__mod_lruvec_page_state(page, NR_FILE_PMDMAPPED,
1364 						-nr_pages);
1365 	} else {
1366 		if (atomic_add_negative(-1, &page->_mapcount))
1367 			nr++;
1368 	}
1369 out:
1370 	if (nr)
1371 		__mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1372 }
1373 
page_remove_anon_compound_rmap(struct page * page)1374 static void page_remove_anon_compound_rmap(struct page *page)
1375 {
1376 	int i, nr;
1377 
1378 	if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1379 		return;
1380 
1381 	/* Hugepages are not counted in NR_ANON_PAGES for now. */
1382 	if (unlikely(PageHuge(page)))
1383 		return;
1384 
1385 	if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1386 		return;
1387 
1388 	__mod_lruvec_page_state(page, NR_ANON_THPS, -thp_nr_pages(page));
1389 
1390 	if (TestClearPageDoubleMap(page)) {
1391 		/*
1392 		 * Subpages can be mapped with PTEs too. Check how many of
1393 		 * them are still mapped.
1394 		 */
1395 		for (i = 0, nr = 0; i < thp_nr_pages(page); i++) {
1396 			if (atomic_add_negative(-1, &page[i]._mapcount))
1397 				nr++;
1398 		}
1399 
1400 		/*
1401 		 * Queue the page for deferred split if at least one small
1402 		 * page of the compound page is unmapped, but at least one
1403 		 * small page is still mapped.
1404 		 */
1405 		if (nr && nr < thp_nr_pages(page))
1406 			deferred_split_huge_page(page);
1407 	} else {
1408 		nr = thp_nr_pages(page);
1409 	}
1410 
1411 	if (nr)
1412 		__mod_lruvec_page_state(page, NR_ANON_MAPPED, -nr);
1413 }
1414 
1415 /**
1416  * page_remove_rmap - take down pte mapping from a page
1417  * @page:	page to remove mapping from
1418  * @vma:	the vm area from which the mapping is removed
1419  * @compound:	uncharge the page as compound or small page
1420  *
1421  * The caller needs to hold the pte lock.
1422  */
page_remove_rmap(struct page * page,struct vm_area_struct * vma,bool compound)1423 void page_remove_rmap(struct page *page,
1424 	struct vm_area_struct *vma, bool compound)
1425 {
1426 	lock_page_memcg(page);
1427 
1428 	if (!PageAnon(page)) {
1429 		page_remove_file_rmap(page, compound);
1430 		goto out;
1431 	}
1432 
1433 	if (compound) {
1434 		page_remove_anon_compound_rmap(page);
1435 		goto out;
1436 	}
1437 
1438 	/* page still mapped by someone else? */
1439 	if (!atomic_add_negative(-1, &page->_mapcount))
1440 		goto out;
1441 
1442 	/*
1443 	 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1444 	 * these counters are not modified in interrupt context, and
1445 	 * pte lock(a spinlock) is held, which implies preemption disabled.
1446 	 */
1447 	__dec_lruvec_page_state(page, NR_ANON_MAPPED);
1448 
1449 	if (PageTransCompound(page))
1450 		deferred_split_huge_page(compound_head(page));
1451 
1452 	/*
1453 	 * It would be tidy to reset the PageAnon mapping here,
1454 	 * but that might overwrite a racing page_add_anon_rmap
1455 	 * which increments mapcount after us but sets mapping
1456 	 * before us: so leave the reset to free_unref_page,
1457 	 * and remember that it's only reliable while mapped.
1458 	 * Leaving it set also helps swapoff to reinstate ptes
1459 	 * faster for those pages still in swapcache.
1460 	 */
1461 out:
1462 	unlock_page_memcg(page);
1463 
1464 	munlock_vma_page(page, vma, compound);
1465 }
1466 
1467 /*
1468  * @arg: enum ttu_flags will be passed to this argument
1469  */
try_to_unmap_one(struct folio * folio,struct vm_area_struct * vma,unsigned long address,void * arg)1470 static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
1471 		     unsigned long address, void *arg)
1472 {
1473 	struct mm_struct *mm = vma->vm_mm;
1474 	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1475 	pte_t pteval;
1476 	struct page *subpage;
1477 	bool anon_exclusive, ret = true;
1478 	struct mmu_notifier_range range;
1479 	enum ttu_flags flags = (enum ttu_flags)(long)arg;
1480 
1481 	/*
1482 	 * When racing against e.g. zap_pte_range() on another cpu,
1483 	 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1484 	 * try_to_unmap() may return before page_mapped() has become false,
1485 	 * if page table locking is skipped: use TTU_SYNC to wait for that.
1486 	 */
1487 	if (flags & TTU_SYNC)
1488 		pvmw.flags = PVMW_SYNC;
1489 
1490 	if (flags & TTU_SPLIT_HUGE_PMD)
1491 		split_huge_pmd_address(vma, address, false, folio);
1492 
1493 	/*
1494 	 * For THP, we have to assume the worse case ie pmd for invalidation.
1495 	 * For hugetlb, it could be much worse if we need to do pud
1496 	 * invalidation in the case of pmd sharing.
1497 	 *
1498 	 * Note that the folio can not be freed in this function as call of
1499 	 * try_to_unmap() must hold a reference on the folio.
1500 	 */
1501 	range.end = vma_address_end(&pvmw);
1502 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1503 				address, range.end);
1504 	if (folio_test_hugetlb(folio)) {
1505 		/*
1506 		 * If sharing is possible, start and end will be adjusted
1507 		 * accordingly.
1508 		 */
1509 		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1510 						     &range.end);
1511 	}
1512 	mmu_notifier_invalidate_range_start(&range);
1513 
1514 	while (page_vma_mapped_walk(&pvmw)) {
1515 		/* Unexpected PMD-mapped THP? */
1516 		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1517 
1518 		/*
1519 		 * If the folio is in an mlock()d vma, we must not swap it out.
1520 		 */
1521 		if (!(flags & TTU_IGNORE_MLOCK) &&
1522 		    (vma->vm_flags & VM_LOCKED)) {
1523 			/* Restore the mlock which got missed */
1524 			mlock_vma_folio(folio, vma, false);
1525 			page_vma_mapped_walk_done(&pvmw);
1526 			ret = false;
1527 			break;
1528 		}
1529 
1530 		subpage = folio_page(folio,
1531 					pte_pfn(*pvmw.pte) - folio_pfn(folio));
1532 		address = pvmw.address;
1533 		anon_exclusive = folio_test_anon(folio) &&
1534 				 PageAnonExclusive(subpage);
1535 
1536 		if (folio_test_hugetlb(folio)) {
1537 			bool anon = folio_test_anon(folio);
1538 
1539 			/*
1540 			 * The try_to_unmap() is only passed a hugetlb page
1541 			 * in the case where the hugetlb page is poisoned.
1542 			 */
1543 			VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage);
1544 			/*
1545 			 * huge_pmd_unshare may unmap an entire PMD page.
1546 			 * There is no way of knowing exactly which PMDs may
1547 			 * be cached for this mm, so we must flush them all.
1548 			 * start/end were already adjusted above to cover this
1549 			 * range.
1550 			 */
1551 			flush_cache_range(vma, range.start, range.end);
1552 
1553 			/*
1554 			 * To call huge_pmd_unshare, i_mmap_rwsem must be
1555 			 * held in write mode.  Caller needs to explicitly
1556 			 * do this outside rmap routines.
1557 			 *
1558 			 * We also must hold hugetlb vma_lock in write mode.
1559 			 * Lock order dictates acquiring vma_lock BEFORE
1560 			 * i_mmap_rwsem.  We can only try lock here and fail
1561 			 * if unsuccessful.
1562 			 */
1563 			if (!anon) {
1564 				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1565 				if (!hugetlb_vma_trylock_write(vma)) {
1566 					page_vma_mapped_walk_done(&pvmw);
1567 					ret = false;
1568 					break;
1569 				}
1570 				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1571 					hugetlb_vma_unlock_write(vma);
1572 					flush_tlb_range(vma,
1573 						range.start, range.end);
1574 					mmu_notifier_invalidate_range(mm,
1575 						range.start, range.end);
1576 					/*
1577 					 * The ref count of the PMD page was
1578 					 * dropped which is part of the way map
1579 					 * counting is done for shared PMDs.
1580 					 * Return 'true' here.  When there is
1581 					 * no other sharing, huge_pmd_unshare
1582 					 * returns false and we will unmap the
1583 					 * actual page and drop map count
1584 					 * to zero.
1585 					 */
1586 					page_vma_mapped_walk_done(&pvmw);
1587 					break;
1588 				}
1589 				hugetlb_vma_unlock_write(vma);
1590 			}
1591 			pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1592 		} else {
1593 			flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1594 			/* Nuke the page table entry. */
1595 			if (should_defer_flush(mm, flags)) {
1596 				/*
1597 				 * We clear the PTE but do not flush so potentially
1598 				 * a remote CPU could still be writing to the folio.
1599 				 * If the entry was previously clean then the
1600 				 * architecture must guarantee that a clear->dirty
1601 				 * transition on a cached TLB entry is written through
1602 				 * and traps if the PTE is unmapped.
1603 				 */
1604 				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1605 
1606 				set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1607 			} else {
1608 				pteval = ptep_clear_flush(vma, address, pvmw.pte);
1609 			}
1610 		}
1611 
1612 		/*
1613 		 * Now the pte is cleared. If this pte was uffd-wp armed,
1614 		 * we may want to replace a none pte with a marker pte if
1615 		 * it's file-backed, so we don't lose the tracking info.
1616 		 */
1617 		pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval);
1618 
1619 		/* Set the dirty flag on the folio now the pte is gone. */
1620 		if (pte_dirty(pteval))
1621 			folio_mark_dirty(folio);
1622 
1623 		/* Update high watermark before we lower rss */
1624 		update_hiwater_rss(mm);
1625 
1626 		if (PageHWPoison(subpage) && !(flags & TTU_IGNORE_HWPOISON)) {
1627 			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1628 			if (folio_test_hugetlb(folio)) {
1629 				hugetlb_count_sub(folio_nr_pages(folio), mm);
1630 				set_huge_pte_at(mm, address, pvmw.pte, pteval);
1631 			} else {
1632 				dec_mm_counter(mm, mm_counter(&folio->page));
1633 				set_pte_at(mm, address, pvmw.pte, pteval);
1634 			}
1635 
1636 		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1637 			/*
1638 			 * The guest indicated that the page content is of no
1639 			 * interest anymore. Simply discard the pte, vmscan
1640 			 * will take care of the rest.
1641 			 * A future reference will then fault in a new zero
1642 			 * page. When userfaultfd is active, we must not drop
1643 			 * this page though, as its main user (postcopy
1644 			 * migration) will not expect userfaults on already
1645 			 * copied pages.
1646 			 */
1647 			dec_mm_counter(mm, mm_counter(&folio->page));
1648 			/* We have to invalidate as we cleared the pte */
1649 			mmu_notifier_invalidate_range(mm, address,
1650 						      address + PAGE_SIZE);
1651 		} else if (folio_test_anon(folio)) {
1652 			swp_entry_t entry = { .val = page_private(subpage) };
1653 			pte_t swp_pte;
1654 			/*
1655 			 * Store the swap location in the pte.
1656 			 * See handle_pte_fault() ...
1657 			 */
1658 			if (unlikely(folio_test_swapbacked(folio) !=
1659 					folio_test_swapcache(folio))) {
1660 				WARN_ON_ONCE(1);
1661 				ret = false;
1662 				/* We have to invalidate as we cleared the pte */
1663 				mmu_notifier_invalidate_range(mm, address,
1664 							address + PAGE_SIZE);
1665 				page_vma_mapped_walk_done(&pvmw);
1666 				break;
1667 			}
1668 
1669 			/* MADV_FREE page check */
1670 			if (!folio_test_swapbacked(folio)) {
1671 				int ref_count, map_count;
1672 
1673 				/*
1674 				 * Synchronize with gup_pte_range():
1675 				 * - clear PTE; barrier; read refcount
1676 				 * - inc refcount; barrier; read PTE
1677 				 */
1678 				smp_mb();
1679 
1680 				ref_count = folio_ref_count(folio);
1681 				map_count = folio_mapcount(folio);
1682 
1683 				/*
1684 				 * Order reads for page refcount and dirty flag
1685 				 * (see comments in __remove_mapping()).
1686 				 */
1687 				smp_rmb();
1688 
1689 				/*
1690 				 * The only page refs must be one from isolation
1691 				 * plus the rmap(s) (dropped by discard:).
1692 				 */
1693 				if (ref_count == 1 + map_count &&
1694 				    !folio_test_dirty(folio)) {
1695 					/* Invalidate as we cleared the pte */
1696 					mmu_notifier_invalidate_range(mm,
1697 						address, address + PAGE_SIZE);
1698 					dec_mm_counter(mm, MM_ANONPAGES);
1699 					goto discard;
1700 				}
1701 
1702 				/*
1703 				 * If the folio was redirtied, it cannot be
1704 				 * discarded. Remap the page to page table.
1705 				 */
1706 				set_pte_at(mm, address, pvmw.pte, pteval);
1707 				folio_set_swapbacked(folio);
1708 				ret = false;
1709 				page_vma_mapped_walk_done(&pvmw);
1710 				break;
1711 			}
1712 
1713 			if (swap_duplicate(entry) < 0) {
1714 				set_pte_at(mm, address, pvmw.pte, pteval);
1715 				ret = false;
1716 				page_vma_mapped_walk_done(&pvmw);
1717 				break;
1718 			}
1719 			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1720 				swap_free(entry);
1721 				set_pte_at(mm, address, pvmw.pte, pteval);
1722 				ret = false;
1723 				page_vma_mapped_walk_done(&pvmw);
1724 				break;
1725 			}
1726 
1727 			/* See page_try_share_anon_rmap(): clear PTE first. */
1728 			if (anon_exclusive &&
1729 			    page_try_share_anon_rmap(subpage)) {
1730 				swap_free(entry);
1731 				set_pte_at(mm, address, pvmw.pte, pteval);
1732 				ret = false;
1733 				page_vma_mapped_walk_done(&pvmw);
1734 				break;
1735 			}
1736 			/*
1737 			 * Note: We *don't* remember if the page was mapped
1738 			 * exclusively in the swap pte if the architecture
1739 			 * doesn't support __HAVE_ARCH_PTE_SWP_EXCLUSIVE. In
1740 			 * that case, swapin code has to re-determine that
1741 			 * manually and might detect the page as possibly
1742 			 * shared, for example, if there are other references on
1743 			 * the page or if the page is under writeback. We made
1744 			 * sure that there are no GUP pins on the page that
1745 			 * would rely on it, so for GUP pins this is fine.
1746 			 */
1747 			if (list_empty(&mm->mmlist)) {
1748 				spin_lock(&mmlist_lock);
1749 				if (list_empty(&mm->mmlist))
1750 					list_add(&mm->mmlist, &init_mm.mmlist);
1751 				spin_unlock(&mmlist_lock);
1752 			}
1753 			dec_mm_counter(mm, MM_ANONPAGES);
1754 			inc_mm_counter(mm, MM_SWAPENTS);
1755 			swp_pte = swp_entry_to_pte(entry);
1756 			if (anon_exclusive)
1757 				swp_pte = pte_swp_mkexclusive(swp_pte);
1758 			if (pte_soft_dirty(pteval))
1759 				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1760 			if (pte_uffd_wp(pteval))
1761 				swp_pte = pte_swp_mkuffd_wp(swp_pte);
1762 			set_pte_at(mm, address, pvmw.pte, swp_pte);
1763 			/* Invalidate as we cleared the pte */
1764 			mmu_notifier_invalidate_range(mm, address,
1765 						      address + PAGE_SIZE);
1766 		} else {
1767 			/*
1768 			 * This is a locked file-backed folio,
1769 			 * so it cannot be removed from the page
1770 			 * cache and replaced by a new folio before
1771 			 * mmu_notifier_invalidate_range_end, so no
1772 			 * concurrent thread might update its page table
1773 			 * to point at a new folio while a device is
1774 			 * still using this folio.
1775 			 *
1776 			 * See Documentation/mm/mmu_notifier.rst
1777 			 */
1778 			dec_mm_counter(mm, mm_counter_file(&folio->page));
1779 		}
1780 discard:
1781 		/*
1782 		 * No need to call mmu_notifier_invalidate_range() it has be
1783 		 * done above for all cases requiring it to happen under page
1784 		 * table lock before mmu_notifier_invalidate_range_end()
1785 		 *
1786 		 * See Documentation/mm/mmu_notifier.rst
1787 		 */
1788 		page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
1789 		if (vma->vm_flags & VM_LOCKED)
1790 			mlock_page_drain_local();
1791 		folio_put(folio);
1792 	}
1793 
1794 	mmu_notifier_invalidate_range_end(&range);
1795 
1796 	return ret;
1797 }
1798 
invalid_migration_vma(struct vm_area_struct * vma,void * arg)1799 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1800 {
1801 	return vma_is_temporary_stack(vma);
1802 }
1803 
page_not_mapped(struct folio * folio)1804 static int page_not_mapped(struct folio *folio)
1805 {
1806 	return !folio_mapped(folio);
1807 }
1808 
1809 /**
1810  * try_to_unmap - Try to remove all page table mappings to a folio.
1811  * @folio: The folio to unmap.
1812  * @flags: action and flags
1813  *
1814  * Tries to remove all the page table entries which are mapping this
1815  * folio.  It is the caller's responsibility to check if the folio is
1816  * still mapped if needed (use TTU_SYNC to prevent accounting races).
1817  *
1818  * Context: Caller must hold the folio lock.
1819  */
try_to_unmap(struct folio * folio,enum ttu_flags flags)1820 void try_to_unmap(struct folio *folio, enum ttu_flags flags)
1821 {
1822 	struct rmap_walk_control rwc = {
1823 		.rmap_one = try_to_unmap_one,
1824 		.arg = (void *)flags,
1825 		.done = page_not_mapped,
1826 		.anon_lock = folio_lock_anon_vma_read,
1827 	};
1828 
1829 	if (flags & TTU_RMAP_LOCKED)
1830 		rmap_walk_locked(folio, &rwc);
1831 	else
1832 		rmap_walk(folio, &rwc);
1833 }
1834 
1835 /*
1836  * @arg: enum ttu_flags will be passed to this argument.
1837  *
1838  * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
1839  * containing migration entries.
1840  */
try_to_migrate_one(struct folio * folio,struct vm_area_struct * vma,unsigned long address,void * arg)1841 static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
1842 		     unsigned long address, void *arg)
1843 {
1844 	struct mm_struct *mm = vma->vm_mm;
1845 	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1846 	pte_t pteval;
1847 	struct page *subpage;
1848 	bool anon_exclusive, ret = true;
1849 	struct mmu_notifier_range range;
1850 	enum ttu_flags flags = (enum ttu_flags)(long)arg;
1851 
1852 	/*
1853 	 * When racing against e.g. zap_pte_range() on another cpu,
1854 	 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1855 	 * try_to_migrate() may return before page_mapped() has become false,
1856 	 * if page table locking is skipped: use TTU_SYNC to wait for that.
1857 	 */
1858 	if (flags & TTU_SYNC)
1859 		pvmw.flags = PVMW_SYNC;
1860 
1861 	/*
1862 	 * unmap_page() in mm/huge_memory.c is the only user of migration with
1863 	 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
1864 	 */
1865 	if (flags & TTU_SPLIT_HUGE_PMD)
1866 		split_huge_pmd_address(vma, address, true, folio);
1867 
1868 	/*
1869 	 * For THP, we have to assume the worse case ie pmd for invalidation.
1870 	 * For hugetlb, it could be much worse if we need to do pud
1871 	 * invalidation in the case of pmd sharing.
1872 	 *
1873 	 * Note that the page can not be free in this function as call of
1874 	 * try_to_unmap() must hold a reference on the page.
1875 	 */
1876 	range.end = vma_address_end(&pvmw);
1877 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1878 				address, range.end);
1879 	if (folio_test_hugetlb(folio)) {
1880 		/*
1881 		 * If sharing is possible, start and end will be adjusted
1882 		 * accordingly.
1883 		 */
1884 		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1885 						     &range.end);
1886 	}
1887 	mmu_notifier_invalidate_range_start(&range);
1888 
1889 	while (page_vma_mapped_walk(&pvmw)) {
1890 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1891 		/* PMD-mapped THP migration entry */
1892 		if (!pvmw.pte) {
1893 			subpage = folio_page(folio,
1894 				pmd_pfn(*pvmw.pmd) - folio_pfn(folio));
1895 			VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
1896 					!folio_test_pmd_mappable(folio), folio);
1897 
1898 			if (set_pmd_migration_entry(&pvmw, subpage)) {
1899 				ret = false;
1900 				page_vma_mapped_walk_done(&pvmw);
1901 				break;
1902 			}
1903 			continue;
1904 		}
1905 #endif
1906 
1907 		/* Unexpected PMD-mapped THP? */
1908 		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1909 
1910 		if (folio_is_zone_device(folio)) {
1911 			/*
1912 			 * Our PTE is a non-present device exclusive entry and
1913 			 * calculating the subpage as for the common case would
1914 			 * result in an invalid pointer.
1915 			 *
1916 			 * Since only PAGE_SIZE pages can currently be
1917 			 * migrated, just set it to page. This will need to be
1918 			 * changed when hugepage migrations to device private
1919 			 * memory are supported.
1920 			 */
1921 			VM_BUG_ON_FOLIO(folio_nr_pages(folio) > 1, folio);
1922 			subpage = &folio->page;
1923 		} else {
1924 			subpage = folio_page(folio,
1925 					pte_pfn(*pvmw.pte) - folio_pfn(folio));
1926 		}
1927 		address = pvmw.address;
1928 		anon_exclusive = folio_test_anon(folio) &&
1929 				 PageAnonExclusive(subpage);
1930 
1931 		if (folio_test_hugetlb(folio)) {
1932 			bool anon = folio_test_anon(folio);
1933 
1934 			/*
1935 			 * huge_pmd_unshare may unmap an entire PMD page.
1936 			 * There is no way of knowing exactly which PMDs may
1937 			 * be cached for this mm, so we must flush them all.
1938 			 * start/end were already adjusted above to cover this
1939 			 * range.
1940 			 */
1941 			flush_cache_range(vma, range.start, range.end);
1942 
1943 			/*
1944 			 * To call huge_pmd_unshare, i_mmap_rwsem must be
1945 			 * held in write mode.  Caller needs to explicitly
1946 			 * do this outside rmap routines.
1947 			 *
1948 			 * We also must hold hugetlb vma_lock in write mode.
1949 			 * Lock order dictates acquiring vma_lock BEFORE
1950 			 * i_mmap_rwsem.  We can only try lock here and
1951 			 * fail if unsuccessful.
1952 			 */
1953 			if (!anon) {
1954 				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1955 				if (!hugetlb_vma_trylock_write(vma)) {
1956 					page_vma_mapped_walk_done(&pvmw);
1957 					ret = false;
1958 					break;
1959 				}
1960 				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1961 					hugetlb_vma_unlock_write(vma);
1962 					flush_tlb_range(vma,
1963 						range.start, range.end);
1964 					mmu_notifier_invalidate_range(mm,
1965 						range.start, range.end);
1966 
1967 					/*
1968 					 * The ref count of the PMD page was
1969 					 * dropped which is part of the way map
1970 					 * counting is done for shared PMDs.
1971 					 * Return 'true' here.  When there is
1972 					 * no other sharing, huge_pmd_unshare
1973 					 * returns false and we will unmap the
1974 					 * actual page and drop map count
1975 					 * to zero.
1976 					 */
1977 					page_vma_mapped_walk_done(&pvmw);
1978 					break;
1979 				}
1980 				hugetlb_vma_unlock_write(vma);
1981 			}
1982 			/* Nuke the hugetlb page table entry */
1983 			pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1984 		} else {
1985 			flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1986 			/* Nuke the page table entry. */
1987 			pteval = ptep_clear_flush(vma, address, pvmw.pte);
1988 		}
1989 
1990 		/* Set the dirty flag on the folio now the pte is gone. */
1991 		if (pte_dirty(pteval))
1992 			folio_mark_dirty(folio);
1993 
1994 		/* Update high watermark before we lower rss */
1995 		update_hiwater_rss(mm);
1996 
1997 		if (folio_is_device_private(folio)) {
1998 			unsigned long pfn = folio_pfn(folio);
1999 			swp_entry_t entry;
2000 			pte_t swp_pte;
2001 
2002 			if (anon_exclusive)
2003 				BUG_ON(page_try_share_anon_rmap(subpage));
2004 
2005 			/*
2006 			 * Store the pfn of the page in a special migration
2007 			 * pte. do_swap_page() will wait until the migration
2008 			 * pte is removed and then restart fault handling.
2009 			 */
2010 			entry = pte_to_swp_entry(pteval);
2011 			if (is_writable_device_private_entry(entry))
2012 				entry = make_writable_migration_entry(pfn);
2013 			else if (anon_exclusive)
2014 				entry = make_readable_exclusive_migration_entry(pfn);
2015 			else
2016 				entry = make_readable_migration_entry(pfn);
2017 			swp_pte = swp_entry_to_pte(entry);
2018 
2019 			/*
2020 			 * pteval maps a zone device page and is therefore
2021 			 * a swap pte.
2022 			 */
2023 			if (pte_swp_soft_dirty(pteval))
2024 				swp_pte = pte_swp_mksoft_dirty(swp_pte);
2025 			if (pte_swp_uffd_wp(pteval))
2026 				swp_pte = pte_swp_mkuffd_wp(swp_pte);
2027 			set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
2028 			trace_set_migration_pte(pvmw.address, pte_val(swp_pte),
2029 						compound_order(&folio->page));
2030 			/*
2031 			 * No need to invalidate here it will synchronize on
2032 			 * against the special swap migration pte.
2033 			 */
2034 		} else if (PageHWPoison(subpage)) {
2035 			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
2036 			if (folio_test_hugetlb(folio)) {
2037 				hugetlb_count_sub(folio_nr_pages(folio), mm);
2038 				set_huge_pte_at(mm, address, pvmw.pte, pteval);
2039 			} else {
2040 				dec_mm_counter(mm, mm_counter(&folio->page));
2041 				set_pte_at(mm, address, pvmw.pte, pteval);
2042 			}
2043 
2044 		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
2045 			/*
2046 			 * The guest indicated that the page content is of no
2047 			 * interest anymore. Simply discard the pte, vmscan
2048 			 * will take care of the rest.
2049 			 * A future reference will then fault in a new zero
2050 			 * page. When userfaultfd is active, we must not drop
2051 			 * this page though, as its main user (postcopy
2052 			 * migration) will not expect userfaults on already
2053 			 * copied pages.
2054 			 */
2055 			dec_mm_counter(mm, mm_counter(&folio->page));
2056 			/* We have to invalidate as we cleared the pte */
2057 			mmu_notifier_invalidate_range(mm, address,
2058 						      address + PAGE_SIZE);
2059 		} else {
2060 			swp_entry_t entry;
2061 			pte_t swp_pte;
2062 
2063 			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
2064 				if (folio_test_hugetlb(folio))
2065 					set_huge_pte_at(mm, address, pvmw.pte, pteval);
2066 				else
2067 					set_pte_at(mm, address, pvmw.pte, pteval);
2068 				ret = false;
2069 				page_vma_mapped_walk_done(&pvmw);
2070 				break;
2071 			}
2072 			VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) &&
2073 				       !anon_exclusive, subpage);
2074 
2075 			/* See page_try_share_anon_rmap(): clear PTE first. */
2076 			if (anon_exclusive &&
2077 			    page_try_share_anon_rmap(subpage)) {
2078 				if (folio_test_hugetlb(folio))
2079 					set_huge_pte_at(mm, address, pvmw.pte, pteval);
2080 				else
2081 					set_pte_at(mm, address, pvmw.pte, pteval);
2082 				ret = false;
2083 				page_vma_mapped_walk_done(&pvmw);
2084 				break;
2085 			}
2086 
2087 			/*
2088 			 * Store the pfn of the page in a special migration
2089 			 * pte. do_swap_page() will wait until the migration
2090 			 * pte is removed and then restart fault handling.
2091 			 */
2092 			if (pte_write(pteval))
2093 				entry = make_writable_migration_entry(
2094 							page_to_pfn(subpage));
2095 			else if (anon_exclusive)
2096 				entry = make_readable_exclusive_migration_entry(
2097 							page_to_pfn(subpage));
2098 			else
2099 				entry = make_readable_migration_entry(
2100 							page_to_pfn(subpage));
2101 			if (pte_young(pteval))
2102 				entry = make_migration_entry_young(entry);
2103 			if (pte_dirty(pteval))
2104 				entry = make_migration_entry_dirty(entry);
2105 			swp_pte = swp_entry_to_pte(entry);
2106 			if (pte_soft_dirty(pteval))
2107 				swp_pte = pte_swp_mksoft_dirty(swp_pte);
2108 			if (pte_uffd_wp(pteval))
2109 				swp_pte = pte_swp_mkuffd_wp(swp_pte);
2110 			if (folio_test_hugetlb(folio))
2111 				set_huge_pte_at(mm, address, pvmw.pte, swp_pte);
2112 			else
2113 				set_pte_at(mm, address, pvmw.pte, swp_pte);
2114 			trace_set_migration_pte(address, pte_val(swp_pte),
2115 						compound_order(&folio->page));
2116 			/*
2117 			 * No need to invalidate here it will synchronize on
2118 			 * against the special swap migration pte.
2119 			 */
2120 		}
2121 
2122 		/*
2123 		 * No need to call mmu_notifier_invalidate_range() it has be
2124 		 * done above for all cases requiring it to happen under page
2125 		 * table lock before mmu_notifier_invalidate_range_end()
2126 		 *
2127 		 * See Documentation/mm/mmu_notifier.rst
2128 		 */
2129 		page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
2130 		if (vma->vm_flags & VM_LOCKED)
2131 			mlock_page_drain_local();
2132 		folio_put(folio);
2133 	}
2134 
2135 	mmu_notifier_invalidate_range_end(&range);
2136 
2137 	return ret;
2138 }
2139 
2140 /**
2141  * try_to_migrate - try to replace all page table mappings with swap entries
2142  * @folio: the folio to replace page table entries for
2143  * @flags: action and flags
2144  *
2145  * Tries to remove all the page table entries which are mapping this folio and
2146  * replace them with special swap entries. Caller must hold the folio lock.
2147  */
try_to_migrate(struct folio * folio,enum ttu_flags flags)2148 void try_to_migrate(struct folio *folio, enum ttu_flags flags)
2149 {
2150 	struct rmap_walk_control rwc = {
2151 		.rmap_one = try_to_migrate_one,
2152 		.arg = (void *)flags,
2153 		.done = page_not_mapped,
2154 		.anon_lock = folio_lock_anon_vma_read,
2155 	};
2156 
2157 	/*
2158 	 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
2159 	 * TTU_SPLIT_HUGE_PMD and TTU_SYNC flags.
2160 	 */
2161 	if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2162 					TTU_SYNC)))
2163 		return;
2164 
2165 	if (folio_is_zone_device(folio) &&
2166 	    (!folio_is_device_private(folio) && !folio_is_device_coherent(folio)))
2167 		return;
2168 
2169 	/*
2170 	 * During exec, a temporary VMA is setup and later moved.
2171 	 * The VMA is moved under the anon_vma lock but not the
2172 	 * page tables leading to a race where migration cannot
2173 	 * find the migration ptes. Rather than increasing the
2174 	 * locking requirements of exec(), migration skips
2175 	 * temporary VMAs until after exec() completes.
2176 	 */
2177 	if (!folio_test_ksm(folio) && folio_test_anon(folio))
2178 		rwc.invalid_vma = invalid_migration_vma;
2179 
2180 	if (flags & TTU_RMAP_LOCKED)
2181 		rmap_walk_locked(folio, &rwc);
2182 	else
2183 		rmap_walk(folio, &rwc);
2184 }
2185 
2186 #ifdef CONFIG_DEVICE_PRIVATE
2187 struct make_exclusive_args {
2188 	struct mm_struct *mm;
2189 	unsigned long address;
2190 	void *owner;
2191 	bool valid;
2192 };
2193 
page_make_device_exclusive_one(struct folio * folio,struct vm_area_struct * vma,unsigned long address,void * priv)2194 static bool page_make_device_exclusive_one(struct folio *folio,
2195 		struct vm_area_struct *vma, unsigned long address, void *priv)
2196 {
2197 	struct mm_struct *mm = vma->vm_mm;
2198 	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
2199 	struct make_exclusive_args *args = priv;
2200 	pte_t pteval;
2201 	struct page *subpage;
2202 	bool ret = true;
2203 	struct mmu_notifier_range range;
2204 	swp_entry_t entry;
2205 	pte_t swp_pte;
2206 
2207 	mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
2208 				      vma->vm_mm, address, min(vma->vm_end,
2209 				      address + folio_size(folio)),
2210 				      args->owner);
2211 	mmu_notifier_invalidate_range_start(&range);
2212 
2213 	while (page_vma_mapped_walk(&pvmw)) {
2214 		/* Unexpected PMD-mapped THP? */
2215 		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2216 
2217 		if (!pte_present(*pvmw.pte)) {
2218 			ret = false;
2219 			page_vma_mapped_walk_done(&pvmw);
2220 			break;
2221 		}
2222 
2223 		subpage = folio_page(folio,
2224 				pte_pfn(*pvmw.pte) - folio_pfn(folio));
2225 		address = pvmw.address;
2226 
2227 		/* Nuke the page table entry. */
2228 		flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
2229 		pteval = ptep_clear_flush(vma, address, pvmw.pte);
2230 
2231 		/* Set the dirty flag on the folio now the pte is gone. */
2232 		if (pte_dirty(pteval))
2233 			folio_mark_dirty(folio);
2234 
2235 		/*
2236 		 * Check that our target page is still mapped at the expected
2237 		 * address.
2238 		 */
2239 		if (args->mm == mm && args->address == address &&
2240 		    pte_write(pteval))
2241 			args->valid = true;
2242 
2243 		/*
2244 		 * Store the pfn of the page in a special migration
2245 		 * pte. do_swap_page() will wait until the migration
2246 		 * pte is removed and then restart fault handling.
2247 		 */
2248 		if (pte_write(pteval))
2249 			entry = make_writable_device_exclusive_entry(
2250 							page_to_pfn(subpage));
2251 		else
2252 			entry = make_readable_device_exclusive_entry(
2253 							page_to_pfn(subpage));
2254 		swp_pte = swp_entry_to_pte(entry);
2255 		if (pte_soft_dirty(pteval))
2256 			swp_pte = pte_swp_mksoft_dirty(swp_pte);
2257 		if (pte_uffd_wp(pteval))
2258 			swp_pte = pte_swp_mkuffd_wp(swp_pte);
2259 
2260 		set_pte_at(mm, address, pvmw.pte, swp_pte);
2261 
2262 		/*
2263 		 * There is a reference on the page for the swap entry which has
2264 		 * been removed, so shouldn't take another.
2265 		 */
2266 		page_remove_rmap(subpage, vma, false);
2267 	}
2268 
2269 	mmu_notifier_invalidate_range_end(&range);
2270 
2271 	return ret;
2272 }
2273 
2274 /**
2275  * folio_make_device_exclusive - Mark the folio exclusively owned by a device.
2276  * @folio: The folio to replace page table entries for.
2277  * @mm: The mm_struct where the folio is expected to be mapped.
2278  * @address: Address where the folio is expected to be mapped.
2279  * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2280  *
2281  * Tries to remove all the page table entries which are mapping this
2282  * folio and replace them with special device exclusive swap entries to
2283  * grant a device exclusive access to the folio.
2284  *
2285  * Context: Caller must hold the folio lock.
2286  * Return: false if the page is still mapped, or if it could not be unmapped
2287  * from the expected address. Otherwise returns true (success).
2288  */
folio_make_device_exclusive(struct folio * folio,struct mm_struct * mm,unsigned long address,void * owner)2289 static bool folio_make_device_exclusive(struct folio *folio,
2290 		struct mm_struct *mm, unsigned long address, void *owner)
2291 {
2292 	struct make_exclusive_args args = {
2293 		.mm = mm,
2294 		.address = address,
2295 		.owner = owner,
2296 		.valid = false,
2297 	};
2298 	struct rmap_walk_control rwc = {
2299 		.rmap_one = page_make_device_exclusive_one,
2300 		.done = page_not_mapped,
2301 		.anon_lock = folio_lock_anon_vma_read,
2302 		.arg = &args,
2303 	};
2304 
2305 	/*
2306 	 * Restrict to anonymous folios for now to avoid potential writeback
2307 	 * issues.
2308 	 */
2309 	if (!folio_test_anon(folio))
2310 		return false;
2311 
2312 	rmap_walk(folio, &rwc);
2313 
2314 	return args.valid && !folio_mapcount(folio);
2315 }
2316 
2317 /**
2318  * make_device_exclusive_range() - Mark a range for exclusive use by a device
2319  * @mm: mm_struct of associated target process
2320  * @start: start of the region to mark for exclusive device access
2321  * @end: end address of region
2322  * @pages: returns the pages which were successfully marked for exclusive access
2323  * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2324  *
2325  * Returns: number of pages found in the range by GUP. A page is marked for
2326  * exclusive access only if the page pointer is non-NULL.
2327  *
2328  * This function finds ptes mapping page(s) to the given address range, locks
2329  * them and replaces mappings with special swap entries preventing userspace CPU
2330  * access. On fault these entries are replaced with the original mapping after
2331  * calling MMU notifiers.
2332  *
2333  * A driver using this to program access from a device must use a mmu notifier
2334  * critical section to hold a device specific lock during programming. Once
2335  * programming is complete it should drop the page lock and reference after
2336  * which point CPU access to the page will revoke the exclusive access.
2337  */
make_device_exclusive_range(struct mm_struct * mm,unsigned long start,unsigned long end,struct page ** pages,void * owner)2338 int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2339 				unsigned long end, struct page **pages,
2340 				void *owner)
2341 {
2342 	long npages = (end - start) >> PAGE_SHIFT;
2343 	long i;
2344 
2345 	npages = get_user_pages_remote(mm, start, npages,
2346 				       FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2347 				       pages, NULL, NULL);
2348 	if (npages < 0)
2349 		return npages;
2350 
2351 	for (i = 0; i < npages; i++, start += PAGE_SIZE) {
2352 		struct folio *folio = page_folio(pages[i]);
2353 		if (PageTail(pages[i]) || !folio_trylock(folio)) {
2354 			folio_put(folio);
2355 			pages[i] = NULL;
2356 			continue;
2357 		}
2358 
2359 		if (!folio_make_device_exclusive(folio, mm, start, owner)) {
2360 			folio_unlock(folio);
2361 			folio_put(folio);
2362 			pages[i] = NULL;
2363 		}
2364 	}
2365 
2366 	return npages;
2367 }
2368 EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2369 #endif
2370 
__put_anon_vma(struct anon_vma * anon_vma)2371 void __put_anon_vma(struct anon_vma *anon_vma)
2372 {
2373 	struct anon_vma *root = anon_vma->root;
2374 
2375 	anon_vma_free(anon_vma);
2376 	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2377 		anon_vma_free(root);
2378 }
2379 
rmap_walk_anon_lock(struct folio * folio,struct rmap_walk_control * rwc)2380 static struct anon_vma *rmap_walk_anon_lock(struct folio *folio,
2381 					    struct rmap_walk_control *rwc)
2382 {
2383 	struct anon_vma *anon_vma;
2384 
2385 	if (rwc->anon_lock)
2386 		return rwc->anon_lock(folio, rwc);
2387 
2388 	/*
2389 	 * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read()
2390 	 * because that depends on page_mapped(); but not all its usages
2391 	 * are holding mmap_lock. Users without mmap_lock are required to
2392 	 * take a reference count to prevent the anon_vma disappearing
2393 	 */
2394 	anon_vma = folio_anon_vma(folio);
2395 	if (!anon_vma)
2396 		return NULL;
2397 
2398 	if (anon_vma_trylock_read(anon_vma))
2399 		goto out;
2400 
2401 	if (rwc->try_lock) {
2402 		anon_vma = NULL;
2403 		rwc->contended = true;
2404 		goto out;
2405 	}
2406 
2407 	anon_vma_lock_read(anon_vma);
2408 out:
2409 	return anon_vma;
2410 }
2411 
2412 /*
2413  * rmap_walk_anon - do something to anonymous page using the object-based
2414  * rmap method
2415  * @page: the page to be handled
2416  * @rwc: control variable according to each walk type
2417  *
2418  * Find all the mappings of a page using the mapping pointer and the vma chains
2419  * contained in the anon_vma struct it points to.
2420  */
rmap_walk_anon(struct folio * folio,struct rmap_walk_control * rwc,bool locked)2421 static void rmap_walk_anon(struct folio *folio,
2422 		struct rmap_walk_control *rwc, bool locked)
2423 {
2424 	struct anon_vma *anon_vma;
2425 	pgoff_t pgoff_start, pgoff_end;
2426 	struct anon_vma_chain *avc;
2427 
2428 	if (locked) {
2429 		anon_vma = folio_anon_vma(folio);
2430 		/* anon_vma disappear under us? */
2431 		VM_BUG_ON_FOLIO(!anon_vma, folio);
2432 	} else {
2433 		anon_vma = rmap_walk_anon_lock(folio, rwc);
2434 	}
2435 	if (!anon_vma)
2436 		return;
2437 
2438 	pgoff_start = folio_pgoff(folio);
2439 	pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2440 	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2441 			pgoff_start, pgoff_end) {
2442 		struct vm_area_struct *vma = avc->vma;
2443 		unsigned long address = vma_address(&folio->page, vma);
2444 
2445 		VM_BUG_ON_VMA(address == -EFAULT, vma);
2446 		cond_resched();
2447 
2448 		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2449 			continue;
2450 
2451 		if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2452 			break;
2453 		if (rwc->done && rwc->done(folio))
2454 			break;
2455 	}
2456 
2457 	if (!locked)
2458 		anon_vma_unlock_read(anon_vma);
2459 }
2460 
2461 /*
2462  * rmap_walk_file - do something to file page using the object-based rmap method
2463  * @page: the page to be handled
2464  * @rwc: control variable according to each walk type
2465  *
2466  * Find all the mappings of a page using the mapping pointer and the vma chains
2467  * contained in the address_space struct it points to.
2468  */
rmap_walk_file(struct folio * folio,struct rmap_walk_control * rwc,bool locked)2469 static void rmap_walk_file(struct folio *folio,
2470 		struct rmap_walk_control *rwc, bool locked)
2471 {
2472 	struct address_space *mapping = folio_mapping(folio);
2473 	pgoff_t pgoff_start, pgoff_end;
2474 	struct vm_area_struct *vma;
2475 
2476 	/*
2477 	 * The page lock not only makes sure that page->mapping cannot
2478 	 * suddenly be NULLified by truncation, it makes sure that the
2479 	 * structure at mapping cannot be freed and reused yet,
2480 	 * so we can safely take mapping->i_mmap_rwsem.
2481 	 */
2482 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2483 
2484 	if (!mapping)
2485 		return;
2486 
2487 	pgoff_start = folio_pgoff(folio);
2488 	pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2489 	if (!locked) {
2490 		if (i_mmap_trylock_read(mapping))
2491 			goto lookup;
2492 
2493 		if (rwc->try_lock) {
2494 			rwc->contended = true;
2495 			return;
2496 		}
2497 
2498 		i_mmap_lock_read(mapping);
2499 	}
2500 lookup:
2501 	vma_interval_tree_foreach(vma, &mapping->i_mmap,
2502 			pgoff_start, pgoff_end) {
2503 		unsigned long address = vma_address(&folio->page, vma);
2504 
2505 		VM_BUG_ON_VMA(address == -EFAULT, vma);
2506 		cond_resched();
2507 
2508 		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2509 			continue;
2510 
2511 		if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2512 			goto done;
2513 		if (rwc->done && rwc->done(folio))
2514 			goto done;
2515 	}
2516 
2517 done:
2518 	if (!locked)
2519 		i_mmap_unlock_read(mapping);
2520 }
2521 
rmap_walk(struct folio * folio,struct rmap_walk_control * rwc)2522 void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc)
2523 {
2524 	if (unlikely(folio_test_ksm(folio)))
2525 		rmap_walk_ksm(folio, rwc);
2526 	else if (folio_test_anon(folio))
2527 		rmap_walk_anon(folio, rwc, false);
2528 	else
2529 		rmap_walk_file(folio, rwc, false);
2530 }
2531 
2532 /* Like rmap_walk, but caller holds relevant rmap lock */
rmap_walk_locked(struct folio * folio,struct rmap_walk_control * rwc)2533 void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc)
2534 {
2535 	/* no ksm support for now */
2536 	VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio);
2537 	if (folio_test_anon(folio))
2538 		rmap_walk_anon(folio, rwc, true);
2539 	else
2540 		rmap_walk_file(folio, rwc, true);
2541 }
2542 
2543 #ifdef CONFIG_HUGETLB_PAGE
2544 /*
2545  * The following two functions are for anonymous (private mapped) hugepages.
2546  * Unlike common anonymous pages, anonymous hugepages have no accounting code
2547  * and no lru code, because we handle hugepages differently from common pages.
2548  *
2549  * RMAP_COMPOUND is ignored.
2550  */
hugepage_add_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address,rmap_t flags)2551 void hugepage_add_anon_rmap(struct page *page, struct vm_area_struct *vma,
2552 			    unsigned long address, rmap_t flags)
2553 {
2554 	struct anon_vma *anon_vma = vma->anon_vma;
2555 	int first;
2556 
2557 	BUG_ON(!PageLocked(page));
2558 	BUG_ON(!anon_vma);
2559 	/* address might be in next vma when migration races vma_adjust */
2560 	first = atomic_inc_and_test(compound_mapcount_ptr(page));
2561 	VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
2562 	VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
2563 	if (first)
2564 		__page_set_anon_rmap(page, vma, address,
2565 				     !!(flags & RMAP_EXCLUSIVE));
2566 }
2567 
hugepage_add_new_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address)2568 void hugepage_add_new_anon_rmap(struct page *page,
2569 			struct vm_area_struct *vma, unsigned long address)
2570 {
2571 	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2572 	atomic_set(compound_mapcount_ptr(page), 0);
2573 	atomic_set(compound_pincount_ptr(page), 0);
2574 
2575 	__page_set_anon_rmap(page, vma, address, 1);
2576 }
2577 #endif /* CONFIG_HUGETLB_PAGE */
2578