Branch data Line data Source code
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_mutex (while writing or truncating, not reading or faulting)
24 : : * mm->mmap_sem
25 : : * page->flags PG_locked (lock_page)
26 : : * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
27 : : * mapping->i_mmap_rwsem
28 : : * anon_vma->rwsem
29 : : * mm->page_table_lock or pte_lock
30 : : * pgdat->lru_lock (in mark_page_accessed, isolate_lru_page)
31 : : * swap_lock (in swap_duplicate, swap_info_get)
32 : : * mmlist_lock (in mmput, drain_mmlist and others)
33 : : * mapping->private_lock (in __set_page_dirty_buffers)
34 : : * mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
35 : : * i_pages lock (widely used)
36 : : * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
37 : : * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
38 : : * sb_lock (within inode_lock in fs/fs-writeback.c)
39 : : * i_pages lock (widely used, in set_page_dirty,
40 : : * in arch-dependent flush_dcache_mmap_lock,
41 : : * within bdi.wb->list_lock in __sync_single_inode)
42 : : *
43 : : * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon)
44 : : * ->tasklist_lock
45 : : * pte map lock
46 : : */
47 : :
48 : : #include <linux/mm.h>
49 : : #include <linux/sched/mm.h>
50 : : #include <linux/sched/task.h>
51 : : #include <linux/pagemap.h>
52 : : #include <linux/swap.h>
53 : : #include <linux/swapops.h>
54 : : #include <linux/slab.h>
55 : : #include <linux/init.h>
56 : : #include <linux/ksm.h>
57 : : #include <linux/rmap.h>
58 : : #include <linux/rcupdate.h>
59 : : #include <linux/export.h>
60 : : #include <linux/memcontrol.h>
61 : : #include <linux/mmu_notifier.h>
62 : : #include <linux/migrate.h>
63 : : #include <linux/hugetlb.h>
64 : : #include <linux/huge_mm.h>
65 : : #include <linux/backing-dev.h>
66 : : #include <linux/page_idle.h>
67 : : #include <linux/memremap.h>
68 : : #include <linux/userfaultfd_k.h>
69 : :
70 : : #include <asm/tlbflush.h>
71 : :
72 : : #include <trace/events/tlb.h>
73 : :
74 : : #include "internal.h"
75 : :
76 : : static struct kmem_cache *anon_vma_cachep;
77 : : static struct kmem_cache *anon_vma_chain_cachep;
78 : :
79 : 658666 : static inline struct anon_vma *anon_vma_alloc(void)
80 : : {
81 : 658666 : struct anon_vma *anon_vma;
82 : :
83 : 658666 : anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
84 [ + - ]: 658666 : if (anon_vma) {
85 : 658666 : atomic_set(&anon_vma->refcount, 1);
86 : 658666 : anon_vma->degree = 1; /* Reference for first vma */
87 : 658666 : anon_vma->parent = anon_vma;
88 : : /*
89 : : * Initialise the anon_vma root to point to itself. If called
90 : : * from fork, the root will be reset to the parents anon_vma.
91 : : */
92 : 658666 : anon_vma->root = anon_vma;
93 : : }
94 : :
95 : 658666 : return anon_vma;
96 : : }
97 : :
98 : 642035 : static inline void anon_vma_free(struct anon_vma *anon_vma)
99 : : {
100 : 642035 : VM_BUG_ON(atomic_read(&anon_vma->refcount));
101 : :
102 : : /*
103 : : * Synchronize against page_lock_anon_vma_read() such that
104 : : * we can safely hold the lock without the anon_vma getting
105 : : * freed.
106 : : *
107 : : * Relies on the full mb implied by the atomic_dec_and_test() from
108 : : * put_anon_vma() against the acquire barrier implied by
109 : : * down_read_trylock() from page_lock_anon_vma_read(). This orders:
110 : : *
111 : : * page_lock_anon_vma_read() VS put_anon_vma()
112 : : * down_read_trylock() atomic_dec_and_test()
113 : : * LOCK MB
114 : : * atomic_read() rwsem_is_locked()
115 : : *
116 : : * LOCK should suffice since the actual taking of the lock must
117 : : * happen _before_ what follows.
118 : : */
119 : 642035 : might_sleep();
120 [ - + ]: 642035 : if (rwsem_is_locked(&anon_vma->root->rwsem)) {
121 : 0 : anon_vma_lock_write(anon_vma);
122 : 0 : anon_vma_unlock_write(anon_vma);
123 : : }
124 : :
125 : 642035 : kmem_cache_free(anon_vma_cachep, anon_vma);
126 : 642035 : }
127 : :
128 : 1369953 : static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
129 : : {
130 : 1369953 : return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
131 : : }
132 : :
133 : 1335522 : static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
134 : : {
135 : 1335522 : kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
136 : 693487 : }
137 : :
138 : 1369953 : static void anon_vma_chain_link(struct vm_area_struct *vma,
139 : : struct anon_vma_chain *avc,
140 : : struct anon_vma *anon_vma)
141 : : {
142 : 1369953 : avc->vma = vma;
143 : 1369953 : avc->anon_vma = anon_vma;
144 : 1369953 : list_add(&avc->same_vma, &vma->anon_vma_chain);
145 : 1369953 : anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
146 : : }
147 : :
148 : : /**
149 : : * __anon_vma_prepare - attach an anon_vma to a memory region
150 : : * @vma: the memory region in question
151 : : *
152 : : * This makes sure the memory mapping described by 'vma' has
153 : : * an 'anon_vma' attached to it, so that we can associate the
154 : : * anonymous pages mapped into it with that anon_vma.
155 : : *
156 : : * The common case will be that we already have one, which
157 : : * is handled inline by anon_vma_prepare(). But if
158 : : * not we either need to find an adjacent mapping that we
159 : : * can re-use the anon_vma from (very common when the only
160 : : * reason for splitting a vma has been mprotect()), or we
161 : : * allocate a new one.
162 : : *
163 : : * Anon-vma allocations are very subtle, because we may have
164 : : * optimistically looked up an anon_vma in page_lock_anon_vma_read()
165 : : * and that may actually touch the spinlock even in the newly
166 : : * allocated vma (it depends on RCU to make sure that the
167 : : * anon_vma isn't actually destroyed).
168 : : *
169 : : * As a result, we need to do proper anon_vma locking even
170 : : * for the new allocation. At the same time, we do not want
171 : : * to do any locking for the common case of already having
172 : : * an anon_vma.
173 : : *
174 : : * This must be called with the mmap_sem held for reading.
175 : : */
176 : 268795 : int __anon_vma_prepare(struct vm_area_struct *vma)
177 : : {
178 : 268795 : struct mm_struct *mm = vma->vm_mm;
179 : 268795 : struct anon_vma *anon_vma, *allocated;
180 : 268795 : struct anon_vma_chain *avc;
181 : :
182 : 268795 : might_sleep();
183 : :
184 : 268795 : avc = anon_vma_chain_alloc(GFP_KERNEL);
185 [ - + ]: 268795 : if (!avc)
186 : 0 : goto out_enomem;
187 : :
188 : 268795 : anon_vma = find_mergeable_anon_vma(vma);
189 : 268795 : allocated = NULL;
190 [ + - ]: 268795 : if (!anon_vma) {
191 : 268795 : anon_vma = anon_vma_alloc();
192 [ - + ]: 268795 : if (unlikely(!anon_vma))
193 : 0 : goto out_enomem_free_avc;
194 : : allocated = anon_vma;
195 : : }
196 : :
197 : 268795 : anon_vma_lock_write(anon_vma);
198 : : /* page_table_lock to protect against threads */
199 : 268795 : spin_lock(&mm->page_table_lock);
200 [ + - ]: 268795 : if (likely(!vma->anon_vma)) {
201 : 268795 : vma->anon_vma = anon_vma;
202 : 268795 : anon_vma_chain_link(vma, avc, anon_vma);
203 : : /* vma reference or self-parent link for new root */
204 : 268795 : anon_vma->degree++;
205 : 268795 : allocated = NULL;
206 : 268795 : avc = NULL;
207 : : }
208 : 268795 : spin_unlock(&mm->page_table_lock);
209 : 268795 : anon_vma_unlock_write(anon_vma);
210 : :
211 [ - + ]: 268795 : if (unlikely(allocated))
212 : 0 : put_anon_vma(allocated);
213 [ - + ]: 268795 : if (unlikely(avc))
214 : 0 : anon_vma_chain_free(avc);
215 : :
216 : : return 0;
217 : :
218 : : out_enomem_free_avc:
219 : 0 : anon_vma_chain_free(avc);
220 : : out_enomem:
221 : : return -ENOMEM;
222 : : }
223 : :
224 : : /*
225 : : * This is a useful helper function for locking the anon_vma root as
226 : : * we traverse the vma->anon_vma_chain, looping over anon_vma's that
227 : : * have the same vma.
228 : : *
229 : : * Such anon_vma's should have the same root, so you'd expect to see
230 : : * just a single mutex_lock for the whole traversal.
231 : : */
232 : : static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
233 : : {
234 : : struct anon_vma *new_root = anon_vma->root;
235 : : if (new_root != root) {
236 : : if (WARN_ON_ONCE(root))
237 : : up_write(&root->rwsem);
238 : : root = new_root;
239 : : down_write(&root->rwsem);
240 : : }
241 : : return root;
242 : : }
243 : :
244 : 3759686 : static inline void unlock_anon_vma_root(struct anon_vma *root)
245 : : {
246 : 3759686 : if (root)
247 : 1617738 : up_write(&root->rwsem);
248 : : }
249 : :
250 : : /*
251 : : * Attach the anon_vmas from src to dst.
252 : : * Returns 0 on success, -ENOMEM on failure.
253 : : *
254 : : * anon_vma_clone() is called by __vma_split(), __split_vma(), copy_vma() and
255 : : * anon_vma_fork(). The first three want an exact copy of src, while the last
256 : : * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
257 : : * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
258 : : * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
259 : : *
260 : : * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
261 : : * and reuse existing anon_vma which has no vmas and only one child anon_vma.
262 : : * This prevents degradation of anon_vma hierarchy to endless linear chain in
263 : : * case of constantly forking task. On the other hand, an anon_vma with more
264 : : * than one child isn't reused even if there was no alive vma, thus rmap
265 : : * walker has a good chance of avoiding scanning the whole hierarchy when it
266 : : * searches where page is mapped.
267 : : */
268 : 1079067 : int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
269 : : {
270 : 1079067 : struct anon_vma_chain *avc, *pavc;
271 : 1079067 : struct anon_vma *root = NULL;
272 : 1079067 : struct vm_area_struct *prev = dst->vm_prev, *pprev = src->vm_prev;
273 : :
274 : : /*
275 : : * If parent share anon_vma with its vm_prev, keep this sharing in in
276 : : * child.
277 : : *
278 : : * 1. Parent has vm_prev, which implies we have vm_prev.
279 : : * 2. Parent and its vm_prev have the same anon_vma.
280 : : */
281 [ + + + + : 1079067 : if (!dst->anon_vma && src->anon_vma &&
+ - ]
282 [ + + ]: 559971 : pprev && pprev->anon_vma == src->anon_vma)
283 : 170100 : dst->anon_vma = prev->anon_vma;
284 : :
285 : :
286 [ + + ]: 1790354 : list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
287 : 711287 : struct anon_vma *anon_vma;
288 : :
289 : 711287 : avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
290 [ - + ]: 711287 : if (unlikely(!avc)) {
291 [ # # ]: 0 : unlock_anon_vma_root(root);
292 : 0 : root = NULL;
293 : 0 : avc = anon_vma_chain_alloc(GFP_KERNEL);
294 [ # # ]: 0 : if (!avc)
295 : 0 : goto enomem_failure;
296 : : }
297 : 711287 : anon_vma = pavc->anon_vma;
298 : 711287 : root = lock_anon_vma_root(root, anon_vma);
299 : 711287 : anon_vma_chain_link(dst, avc, anon_vma);
300 : :
301 : : /*
302 : : * Reuse existing anon_vma if its degree lower than two,
303 : : * that means it has no vma and only one anon_vma child.
304 : : *
305 : : * Do not chose parent anon_vma, otherwise first child
306 : : * will always reuse it. Root anon_vma is never reused:
307 : : * it has self-parent reference and at least one child.
308 : : */
309 [ + + + - : 711287 : if (!dst->anon_vma && src->anon_vma &&
+ + ]
310 [ - + ]: 23240 : anon_vma != src->anon_vma && anon_vma->degree < 2)
311 : 0 : dst->anon_vma = anon_vma;
312 : : }
313 [ + + ]: 1079067 : if (dst->anon_vma)
314 : 298176 : dst->anon_vma->degree++;
315 [ + + ]: 1079067 : unlock_anon_vma_root(root);
316 : : return 0;
317 : :
318 : : enomem_failure:
319 : : /*
320 : : * dst->anon_vma is dropped here otherwise its degree can be incorrectly
321 : : * decremented in unlink_anon_vmas().
322 : : * We can safely do this because callers of anon_vma_clone() don't care
323 : : * about dst->anon_vma if anon_vma_clone() failed.
324 : : */
325 : 0 : dst->anon_vma = NULL;
326 : 0 : unlink_anon_vmas(dst);
327 : 0 : return -ENOMEM;
328 : : }
329 : :
330 : : /*
331 : : * Attach vma to its own anon_vma, as well as to the anon_vmas that
332 : : * the corresponding VMA in the parent process is attached to.
333 : : * Returns 0 on success, non-zero on failure.
334 : : */
335 : 1375835 : int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
336 : : {
337 : 1375835 : struct anon_vma_chain *avc;
338 : 1375835 : struct anon_vma *anon_vma;
339 : 1375835 : int error;
340 : :
341 : : /* Don't bother if the parent process has no anon_vma here. */
342 [ + + ]: 1375835 : if (!pvma->anon_vma)
343 : : return 0;
344 : :
345 : : /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
346 : 559971 : vma->anon_vma = NULL;
347 : :
348 : : /*
349 : : * First, attach the new VMA to the parent VMA's anon_vmas,
350 : : * so rmap can find non-COWed pages in child processes.
351 : : */
352 : 559971 : error = anon_vma_clone(vma, pvma);
353 [ + - ]: 559971 : if (error)
354 : : return error;
355 : :
356 : : /* An existing anon_vma has been reused, all done then. */
357 [ + + ]: 559971 : if (vma->anon_vma)
358 : : return 0;
359 : :
360 : : /* Then add our own anon_vma. */
361 : 389871 : anon_vma = anon_vma_alloc();
362 [ - + ]: 389871 : if (!anon_vma)
363 : 0 : goto out_error;
364 : 389871 : avc = anon_vma_chain_alloc(GFP_KERNEL);
365 [ - + ]: 389871 : if (!avc)
366 : 0 : goto out_error_free_anon_vma;
367 : :
368 : : /*
369 : : * The root anon_vma's spinlock is the lock actually used when we
370 : : * lock any of the anon_vmas in this anon_vma tree.
371 : : */
372 : 389871 : anon_vma->root = pvma->anon_vma->root;
373 : 389871 : 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 : 389871 : get_anon_vma(anon_vma->root);
380 : : /* Mark this anon_vma as the one where our new (COWed) pages go. */
381 : 389871 : vma->anon_vma = anon_vma;
382 : 389871 : anon_vma_lock_write(anon_vma);
383 : 389871 : anon_vma_chain_link(vma, avc, anon_vma);
384 : 389871 : anon_vma->parent->degree++;
385 : 389871 : anon_vma_unlock_write(anon_vma);
386 : :
387 : 389871 : return 0;
388 : :
389 : : out_error_free_anon_vma:
390 : 0 : put_anon_vma(anon_vma);
391 : 0 : out_error:
392 : 0 : unlink_anon_vmas(vma);
393 : 0 : return -ENOMEM;
394 : : }
395 : :
396 : 2680619 : void unlink_anon_vmas(struct vm_area_struct *vma)
397 : : {
398 : 2680619 : struct anon_vma_chain *avc, *next;
399 : 2680619 : 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 [ + + ]: 4016141 : list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
406 : 1335522 : struct anon_vma *anon_vma = avc->anon_vma;
407 : :
408 : 1335522 : root = lock_anon_vma_root(root, anon_vma);
409 : 1335522 : 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 [ + + ]: 1335522 : if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
416 : 642035 : anon_vma->parent->degree--;
417 : 642035 : continue;
418 : : }
419 : :
420 : 693487 : list_del(&avc->same_vma);
421 : 693487 : anon_vma_chain_free(avc);
422 : : }
423 [ + + ]: 2680619 : if (vma->anon_vma)
424 : 929691 : vma->anon_vma->degree--;
425 [ + + ]: 2680619 : unlock_anon_vma_root(root);
426 : :
427 : : /*
428 : : * Iterate the list once more, it now only contains empty and unlinked
429 : : * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
430 : : * needing to write-acquire the anon_vma->root->rwsem.
431 : : */
432 [ + + ]: 3322654 : list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
433 : 642035 : struct anon_vma *anon_vma = avc->anon_vma;
434 : :
435 : 642035 : VM_WARN_ON(anon_vma->degree);
436 : 642035 : put_anon_vma(anon_vma);
437 : :
438 : 642035 : list_del(&avc->same_vma);
439 : 642035 : anon_vma_chain_free(avc);
440 : : }
441 : 2680619 : }
442 : :
443 : 24588 : static void anon_vma_ctor(void *data)
444 : : {
445 : 24588 : struct anon_vma *anon_vma = data;
446 : :
447 : 24588 : init_rwsem(&anon_vma->rwsem);
448 : 24588 : atomic_set(&anon_vma->refcount, 0);
449 : 24588 : anon_vma->rb_root = RB_ROOT_CACHED;
450 : 24588 : }
451 : :
452 : 28 : void __init anon_vma_init(void)
453 : : {
454 : 28 : anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
455 : : 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
456 : : anon_vma_ctor);
457 : 28 : anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
458 : : SLAB_PANIC|SLAB_ACCOUNT);
459 : 28 : }
460 : :
461 : : /*
462 : : * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
463 : : *
464 : : * Since there is no serialization what so ever against page_remove_rmap()
465 : : * the best this function can do is return a locked anon_vma that might
466 : : * have been relevant to this page.
467 : : *
468 : : * The page might have been remapped to a different anon_vma or the anon_vma
469 : : * returned may already be freed (and even reused).
470 : : *
471 : : * In case it was remapped to a different anon_vma, the new anon_vma will be a
472 : : * child of the old anon_vma, and the anon_vma lifetime rules will therefore
473 : : * ensure that any anon_vma obtained from the page will still be valid for as
474 : : * long as we observe page_mapped() [ hence all those page_mapped() tests ].
475 : : *
476 : : * All users of this function must be very careful when walking the anon_vma
477 : : * chain and verify that the page in question is indeed mapped in it
478 : : * [ something equivalent to page_mapped_in_vma() ].
479 : : *
480 : : * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
481 : : * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
482 : : * if there is a mapcount, we can dereference the anon_vma after observing
483 : : * those.
484 : : */
485 : 0 : struct anon_vma *page_get_anon_vma(struct page *page)
486 : : {
487 : 0 : struct anon_vma *anon_vma = NULL;
488 : 0 : unsigned long anon_mapping;
489 : :
490 : 0 : rcu_read_lock();
491 [ # # ]: 0 : anon_mapping = (unsigned long)READ_ONCE(page->mapping);
492 [ # # ]: 0 : if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
493 : 0 : goto out;
494 [ # # ]: 0 : if (!page_mapped(page))
495 : 0 : goto out;
496 : :
497 : 0 : anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
498 [ # # ]: 0 : if (!atomic_inc_not_zero(&anon_vma->refcount)) {
499 : 0 : anon_vma = NULL;
500 : 0 : goto out;
501 : : }
502 : :
503 : : /*
504 : : * If this page is still mapped, then its anon_vma cannot have been
505 : : * freed. But if it has been unmapped, we have no security against the
506 : : * anon_vma structure being freed and reused (for another anon_vma:
507 : : * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
508 : : * above cannot corrupt).
509 : : */
510 [ # # ]: 0 : if (!page_mapped(page)) {
511 : 0 : rcu_read_unlock();
512 : 0 : put_anon_vma(anon_vma);
513 : 0 : return NULL;
514 : : }
515 : 0 : out:
516 : 0 : rcu_read_unlock();
517 : :
518 : 0 : return anon_vma;
519 : : }
520 : :
521 : : /*
522 : : * Similar to page_get_anon_vma() except it locks the anon_vma.
523 : : *
524 : : * Its a little more complex as it tries to keep the fast path to a single
525 : : * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
526 : : * reference like with page_get_anon_vma() and then block on the mutex.
527 : : */
528 : 0 : struct anon_vma *page_lock_anon_vma_read(struct page *page)
529 : : {
530 : 0 : struct anon_vma *anon_vma = NULL;
531 : 0 : struct anon_vma *root_anon_vma;
532 : 0 : unsigned long anon_mapping;
533 : :
534 : 0 : rcu_read_lock();
535 [ # # ]: 0 : anon_mapping = (unsigned long)READ_ONCE(page->mapping);
536 [ # # ]: 0 : if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
537 : 0 : goto out;
538 [ # # ]: 0 : if (!page_mapped(page))
539 : 0 : goto out;
540 : :
541 : 0 : anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
542 : 0 : root_anon_vma = READ_ONCE(anon_vma->root);
543 [ # # ]: 0 : if (down_read_trylock(&root_anon_vma->rwsem)) {
544 : : /*
545 : : * If the page is still mapped, then this anon_vma is still
546 : : * its anon_vma, and holding the mutex ensures that it will
547 : : * not go away, see anon_vma_free().
548 : : */
549 [ # # ]: 0 : if (!page_mapped(page)) {
550 : 0 : up_read(&root_anon_vma->rwsem);
551 : 0 : anon_vma = NULL;
552 : : }
553 : 0 : goto out;
554 : : }
555 : :
556 : : /* trylock failed, we got to sleep */
557 [ # # ]: 0 : if (!atomic_inc_not_zero(&anon_vma->refcount)) {
558 : 0 : anon_vma = NULL;
559 : 0 : goto out;
560 : : }
561 : :
562 [ # # ]: 0 : if (!page_mapped(page)) {
563 : 0 : rcu_read_unlock();
564 : 0 : put_anon_vma(anon_vma);
565 : 0 : return NULL;
566 : : }
567 : :
568 : : /* we pinned the anon_vma, its safe to sleep */
569 : 0 : rcu_read_unlock();
570 : 0 : anon_vma_lock_read(anon_vma);
571 : :
572 [ # # ]: 0 : if (atomic_dec_and_test(&anon_vma->refcount)) {
573 : : /*
574 : : * Oops, we held the last refcount, release the lock
575 : : * and bail -- can't simply use put_anon_vma() because
576 : : * we'll deadlock on the anon_vma_lock_write() recursion.
577 : : */
578 : 0 : anon_vma_unlock_read(anon_vma);
579 : 0 : __put_anon_vma(anon_vma);
580 : 0 : anon_vma = NULL;
581 : : }
582 : :
583 : : return anon_vma;
584 : :
585 : 0 : out:
586 : 0 : rcu_read_unlock();
587 : 0 : return anon_vma;
588 : : }
589 : :
590 : 0 : void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
591 : : {
592 : 0 : anon_vma_unlock_read(anon_vma);
593 : 0 : }
594 : :
595 : : #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
596 : : /*
597 : : * Flush TLB entries for recently unmapped pages from remote CPUs. It is
598 : : * important if a PTE was dirty when it was unmapped that it's flushed
599 : : * before any IO is initiated on the page to prevent lost writes. Similarly,
600 : : * it must be flushed before freeing to prevent data leakage.
601 : : */
602 : 0 : void try_to_unmap_flush(void)
603 : : {
604 [ # # ]: 0 : struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
605 : :
606 [ # # # # ]: 0 : if (!tlb_ubc->flush_required)
607 : : return;
608 : :
609 : 0 : arch_tlbbatch_flush(&tlb_ubc->arch);
610 : 0 : tlb_ubc->flush_required = false;
611 : 0 : tlb_ubc->writable = false;
612 : : }
613 : :
614 : : /* Flush iff there are potentially writable TLB entries that can race with IO */
615 : 0 : void try_to_unmap_flush_dirty(void)
616 : : {
617 [ # # ]: 0 : struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
618 : :
619 [ # # ]: 0 : if (tlb_ubc->writable)
620 [ # # ]: 0 : try_to_unmap_flush();
621 : 0 : }
622 : :
623 : 0 : static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
624 : : {
625 : 0 : struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
626 : :
627 : 0 : arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
628 : 0 : tlb_ubc->flush_required = true;
629 : :
630 : : /*
631 : : * Ensure compiler does not re-order the setting of tlb_flush_batched
632 : : * before the PTE is cleared.
633 : : */
634 : 0 : barrier();
635 : 0 : mm->tlb_flush_batched = true;
636 : :
637 : : /*
638 : : * If the PTE was dirty then it's best to assume it's writable. The
639 : : * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
640 : : * before the page is queued for IO.
641 : : */
642 [ # # ]: 0 : if (writable)
643 : 0 : tlb_ubc->writable = true;
644 : 0 : }
645 : :
646 : : /*
647 : : * Returns true if the TLB flush should be deferred to the end of a batch of
648 : : * unmap operations to reduce IPIs.
649 : : */
650 : 0 : static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
651 : : {
652 : 0 : bool should_defer = false;
653 : :
654 [ # # ]: 0 : if (!(flags & TTU_BATCH_FLUSH))
655 : : return false;
656 : :
657 : : /* If remote CPUs need to be flushed then defer batch the flush */
658 [ # # ]: 0 : if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
659 : 0 : should_defer = true;
660 : 0 : put_cpu();
661 : :
662 : 0 : return should_defer;
663 : : }
664 : :
665 : : /*
666 : : * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
667 : : * releasing the PTL if TLB flushes are batched. It's possible for a parallel
668 : : * operation such as mprotect or munmap to race between reclaim unmapping
669 : : * the page and flushing the page. If this race occurs, it potentially allows
670 : : * access to data via a stale TLB entry. Tracking all mm's that have TLB
671 : : * batching in flight would be expensive during reclaim so instead track
672 : : * whether TLB batching occurred in the past and if so then do a flush here
673 : : * if required. This will cost one additional flush per reclaim cycle paid
674 : : * by the first operation at risk such as mprotect and mumap.
675 : : *
676 : : * This must be called under the PTL so that an access to tlb_flush_batched
677 : : * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
678 : : * via the PTL.
679 : : */
680 : 2940441 : void flush_tlb_batched_pending(struct mm_struct *mm)
681 : : {
682 [ - + ]: 2940441 : if (mm->tlb_flush_batched) {
683 : 0 : flush_tlb_mm(mm);
684 : :
685 : : /*
686 : : * Do not allow the compiler to re-order the clearing of
687 : : * tlb_flush_batched before the tlb is flushed.
688 : : */
689 : 0 : barrier();
690 : 0 : mm->tlb_flush_batched = false;
691 : : }
692 : 2940441 : }
693 : : #else
694 : : static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
695 : : {
696 : : }
697 : :
698 : : static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
699 : : {
700 : : return false;
701 : : }
702 : : #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
703 : :
704 : : /*
705 : : * At what user virtual address is page expected in vma?
706 : : * Caller should check the page is actually part of the vma.
707 : : */
708 : 0 : unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
709 : : {
710 : 0 : unsigned long address;
711 [ # # # # ]: 0 : if (PageAnon(page)) {
712 : 0 : struct anon_vma *page__anon_vma = page_anon_vma(page);
713 : : /*
714 : : * Note: swapoff's unuse_vma() is more efficient with this
715 : : * check, and needs it to match anon_vma when KSM is active.
716 : : */
717 [ # # # # ]: 0 : if (!vma->anon_vma || !page__anon_vma ||
718 [ # # ]: 0 : vma->anon_vma->root != page__anon_vma->root)
719 : : return -EFAULT;
720 [ # # ]: 0 : } else if (page->mapping) {
721 [ # # # # ]: 0 : if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
722 : : return -EFAULT;
723 : : } else
724 : : return -EFAULT;
725 : 0 : address = __vma_address(page, vma);
726 [ # # # # ]: 0 : if (unlikely(address < vma->vm_start || address >= vma->vm_end))
727 : 0 : return -EFAULT;
728 : : return address;
729 : : }
730 : :
731 : 0 : pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
732 : : {
733 : 0 : pgd_t *pgd;
734 : 0 : p4d_t *p4d;
735 : 0 : pud_t *pud;
736 : 0 : pmd_t *pmd = NULL;
737 : 0 : pmd_t pmde;
738 : :
739 : 0 : pgd = pgd_offset(mm, address);
740 [ # # ]: 0 : if (!pgd_present(*pgd))
741 : 0 : goto out;
742 : :
743 : 0 : p4d = p4d_offset(pgd, address);
744 [ # # ]: 0 : if (!p4d_present(*p4d))
745 : 0 : goto out;
746 : :
747 [ # # ]: 0 : pud = pud_offset(p4d, address);
748 [ # # # # ]: 0 : if (!pud_present(*pud))
749 : 0 : goto out;
750 : :
751 : 0 : pmd = pmd_offset(pud, address);
752 : : /*
753 : : * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
754 : : * without holding anon_vma lock for write. So when looking for a
755 : : * genuine pmde (in which to find pte), test present and !THP together.
756 : : */
757 : 0 : pmde = *pmd;
758 : 0 : barrier();
759 [ # # # # ]: 0 : if (!pmd_present(pmde) || pmd_trans_huge(pmde))
760 : : pmd = NULL;
761 : 0 : out:
762 : 0 : return pmd;
763 : : }
764 : :
765 : : struct page_referenced_arg {
766 : : int mapcount;
767 : : int referenced;
768 : : unsigned long vm_flags;
769 : : struct mem_cgroup *memcg;
770 : : };
771 : : /*
772 : : * arg: page_referenced_arg will be passed
773 : : */
774 : 0 : static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
775 : : unsigned long address, void *arg)
776 : : {
777 : 0 : struct page_referenced_arg *pra = arg;
778 : 0 : struct page_vma_mapped_walk pvmw = {
779 : : .page = page,
780 : : .vma = vma,
781 : : .address = address,
782 : : };
783 : 0 : int referenced = 0;
784 : :
785 [ # # ]: 0 : while (page_vma_mapped_walk(&pvmw)) {
786 : 0 : address = pvmw.address;
787 : :
788 [ # # ]: 0 : if (vma->vm_flags & VM_LOCKED) {
789 [ # # ]: 0 : page_vma_mapped_walk_done(&pvmw);
790 : 0 : pra->vm_flags |= VM_LOCKED;
791 : 0 : return false; /* To break the loop */
792 : : }
793 : :
794 [ # # ]: 0 : if (pvmw.pte) {
795 [ # # # # ]: 0 : if (ptep_clear_flush_young_notify(vma, address,
796 : : pvmw.pte)) {
797 : : /*
798 : : * Don't treat a reference through
799 : : * a sequentially read mapping as such.
800 : : * If the page has been used in another mapping,
801 : : * we will catch it; if this other mapping is
802 : : * already gone, the unmap path will have set
803 : : * PG_referenced or activated the page.
804 : : */
805 [ # # ]: 0 : if (likely(!(vma->vm_flags & VM_SEQ_READ)))
806 : 0 : referenced++;
807 : : }
808 : 0 : } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
809 : : if (pmdp_clear_flush_young_notify(vma, address,
810 : : pvmw.pmd))
811 : : referenced++;
812 : : } else {
813 : : /* unexpected pmd-mapped page? */
814 : 0 : WARN_ON_ONCE(1);
815 : : }
816 : :
817 : 0 : pra->mapcount--;
818 : : }
819 : :
820 : 0 : if (referenced)
821 : : clear_page_idle(page);
822 [ # # ]: 0 : if (test_and_clear_page_young(page))
823 : : referenced++;
824 : :
825 [ # # ]: 0 : if (referenced) {
826 : 0 : pra->referenced++;
827 : 0 : pra->vm_flags |= vma->vm_flags;
828 : : }
829 : :
830 [ # # ]: 0 : if (!pra->mapcount)
831 : 0 : return false; /* To break the loop */
832 : :
833 : : return true;
834 : : }
835 : :
836 : 0 : static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
837 : : {
838 : 0 : struct page_referenced_arg *pra = arg;
839 : 0 : struct mem_cgroup *memcg = pra->memcg;
840 : :
841 : 0 : if (!mm_match_cgroup(vma->vm_mm, memcg))
842 : : return true;
843 : :
844 : 0 : return false;
845 : : }
846 : :
847 : : /**
848 : : * page_referenced - test if the page was referenced
849 : : * @page: the page to test
850 : : * @is_locked: caller holds lock on the page
851 : : * @memcg: target memory cgroup
852 : : * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
853 : : *
854 : : * Quick test_and_clear_referenced for all mappings to a page,
855 : : * returns the number of ptes which referenced the page.
856 : : */
857 : 0 : int page_referenced(struct page *page,
858 : : int is_locked,
859 : : struct mem_cgroup *memcg,
860 : : unsigned long *vm_flags)
861 : : {
862 : 0 : int we_locked = 0;
863 : 0 : struct page_referenced_arg pra = {
864 : : .mapcount = total_mapcount(page),
865 : : .memcg = memcg,
866 : : };
867 : 0 : struct rmap_walk_control rwc = {
868 : : .rmap_one = page_referenced_one,
869 : : .arg = (void *)&pra,
870 : : .anon_lock = page_lock_anon_vma_read,
871 : : };
872 : :
873 : 0 : *vm_flags = 0;
874 [ # # ]: 0 : if (!pra.mapcount)
875 : : return 0;
876 : :
877 [ # # ]: 0 : if (!page_rmapping(page))
878 : : return 0;
879 : :
880 [ # # # # ]: 0 : if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
881 [ # # ]: 0 : we_locked = trylock_page(page);
882 [ # # ]: 0 : if (!we_locked)
883 : : return 1;
884 : : }
885 : :
886 : : /*
887 : : * If we are reclaiming on behalf of a cgroup, skip
888 : : * counting on behalf of references from different
889 : : * cgroups
890 : : */
891 [ # # ]: 0 : if (memcg) {
892 : 0 : rwc.invalid_vma = invalid_page_referenced_vma;
893 : : }
894 : :
895 : 0 : rmap_walk(page, &rwc);
896 : 0 : *vm_flags = pra.vm_flags;
897 : :
898 [ # # ]: 0 : if (we_locked)
899 : 0 : unlock_page(page);
900 : :
901 : 0 : return pra.referenced;
902 : : }
903 : :
904 : 0 : static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
905 : : unsigned long address, void *arg)
906 : : {
907 : 0 : struct page_vma_mapped_walk pvmw = {
908 : : .page = page,
909 : : .vma = vma,
910 : : .address = address,
911 : : .flags = PVMW_SYNC,
912 : : };
913 : 0 : struct mmu_notifier_range range;
914 : 0 : int *cleaned = arg;
915 : :
916 : : /*
917 : : * We have to assume the worse case ie pmd for invalidation. Note that
918 : : * the page can not be free from this function.
919 : : */
920 : 0 : mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
921 : : 0, vma, vma->vm_mm, address,
922 : 0 : min(vma->vm_end, address + page_size(page)));
923 : 0 : mmu_notifier_invalidate_range_start(&range);
924 : :
925 [ # # ]: 0 : while (page_vma_mapped_walk(&pvmw)) {
926 : 0 : int ret = 0;
927 : :
928 : 0 : address = pvmw.address;
929 [ # # ]: 0 : if (pvmw.pte) {
930 : 0 : pte_t entry;
931 : 0 : pte_t *pte = pvmw.pte;
932 : :
933 [ # # # # ]: 0 : if (!pte_dirty(*pte) && !pte_write(*pte))
934 : 0 : continue;
935 : :
936 : 0 : flush_cache_page(vma, address, pte_pfn(*pte));
937 : 0 : entry = ptep_clear_flush(vma, address, pte);
938 : 0 : entry = pte_wrprotect(entry);
939 : 0 : entry = pte_mkclean(entry);
940 : 0 : set_pte_at(vma->vm_mm, address, pte, entry);
941 : 0 : ret = 1;
942 : : } else {
943 : : #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
944 : : pmd_t *pmd = pvmw.pmd;
945 : : pmd_t entry;
946 : :
947 : : if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
948 : : continue;
949 : :
950 : : flush_cache_page(vma, address, page_to_pfn(page));
951 : : entry = pmdp_invalidate(vma, address, pmd);
952 : : entry = pmd_wrprotect(entry);
953 : : entry = pmd_mkclean(entry);
954 : : set_pmd_at(vma->vm_mm, address, pmd, entry);
955 : : ret = 1;
956 : : #else
957 : : /* unexpected pmd-mapped page? */
958 : 0 : WARN_ON_ONCE(1);
959 : : #endif
960 : : }
961 : :
962 : : /*
963 : : * No need to call mmu_notifier_invalidate_range() as we are
964 : : * downgrading page table protection not changing it to point
965 : : * to a new page.
966 : : *
967 : : * See Documentation/vm/mmu_notifier.rst
968 : : */
969 : 0 : if (ret)
970 : 0 : (*cleaned)++;
971 : : }
972 : :
973 : 0 : mmu_notifier_invalidate_range_end(&range);
974 : :
975 : 0 : return true;
976 : : }
977 : :
978 : 0 : static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
979 : : {
980 [ # # ]: 0 : if (vma->vm_flags & VM_SHARED)
981 : 0 : return false;
982 : :
983 : : return true;
984 : : }
985 : :
986 : 224 : int page_mkclean(struct page *page)
987 : : {
988 : 224 : int cleaned = 0;
989 : 224 : struct address_space *mapping;
990 : 224 : struct rmap_walk_control rwc = {
991 : : .arg = (void *)&cleaned,
992 : : .rmap_one = page_mkclean_one,
993 : : .invalid_vma = invalid_mkclean_vma,
994 : : };
995 : :
996 [ - + - + ]: 448 : BUG_ON(!PageLocked(page));
997 : :
998 [ - + ]: 224 : if (!page_mapped(page))
999 : : return 0;
1000 : :
1001 : 0 : mapping = page_mapping(page);
1002 [ # # ]: 0 : if (!mapping)
1003 : : return 0;
1004 : :
1005 : 0 : rmap_walk(page, &rwc);
1006 : :
1007 : 0 : return cleaned;
1008 : : }
1009 : : EXPORT_SYMBOL_GPL(page_mkclean);
1010 : :
1011 : : /**
1012 : : * page_move_anon_rmap - move a page to our anon_vma
1013 : : * @page: the page to move to our anon_vma
1014 : : * @vma: the vma the page belongs to
1015 : : *
1016 : : * When a page belongs exclusively to one process after a COW event,
1017 : : * that page can be moved into the anon_vma that belongs to just that
1018 : : * process, so the rmap code will not search the parent or sibling
1019 : : * processes.
1020 : : */
1021 : 354653 : void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1022 : : {
1023 : 354653 : struct anon_vma *anon_vma = vma->anon_vma;
1024 : :
1025 [ - + ]: 354653 : page = compound_head(page);
1026 : :
1027 : 354653 : VM_BUG_ON_PAGE(!PageLocked(page), page);
1028 : 354653 : VM_BUG_ON_VMA(!anon_vma, vma);
1029 : :
1030 : 354653 : anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1031 : : /*
1032 : : * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1033 : : * simultaneously, so a concurrent reader (eg page_referenced()'s
1034 : : * PageAnon()) will not see one without the other.
1035 : : */
1036 : 354653 : WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1037 : 354653 : }
1038 : :
1039 : : /**
1040 : : * __page_set_anon_rmap - set up new anonymous rmap
1041 : : * @page: Page or Hugepage to add to rmap
1042 : : * @vma: VM area to add page to.
1043 : : * @address: User virtual address of the mapping
1044 : : * @exclusive: the page is exclusively owned by the current process
1045 : : */
1046 : 1066236 : static void __page_set_anon_rmap(struct page *page,
1047 : : struct vm_area_struct *vma, unsigned long address, int exclusive)
1048 : : {
1049 : 1066236 : struct anon_vma *anon_vma = vma->anon_vma;
1050 : :
1051 [ - + ]: 1066236 : BUG_ON(!anon_vma);
1052 : :
1053 [ - + + - ]: 1066236 : if (PageAnon(page))
1054 : : return;
1055 : :
1056 : : /*
1057 : : * If the page isn't exclusively mapped into this vma,
1058 : : * we must use the _oldest_ possible anon_vma for the
1059 : : * page mapping!
1060 : : */
1061 [ - + ]: 1066236 : if (!exclusive)
1062 : 0 : anon_vma = anon_vma->root;
1063 : :
1064 : 1066236 : anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1065 : 1066236 : page->mapping = (struct address_space *) anon_vma;
1066 : 1066236 : page->index = linear_page_index(vma, address);
1067 : : }
1068 : :
1069 : : /**
1070 : : * __page_check_anon_rmap - sanity check anonymous rmap addition
1071 : : * @page: the page to add the mapping to
1072 : : * @vma: the vm area in which the mapping is added
1073 : : * @address: the user virtual address mapped
1074 : : */
1075 : : static void __page_check_anon_rmap(struct page *page,
1076 : : struct vm_area_struct *vma, unsigned long address)
1077 : : {
1078 : : /*
1079 : : * The page's anon-rmap details (mapping and index) are guaranteed to
1080 : : * be set up correctly at this point.
1081 : : *
1082 : : * We have exclusion against page_add_anon_rmap because the caller
1083 : : * always holds the page locked, except if called from page_dup_rmap,
1084 : : * in which case the page is already known to be setup.
1085 : : *
1086 : : * We have exclusion against page_add_new_anon_rmap because those pages
1087 : : * are initially only visible via the pagetables, and the pte is locked
1088 : : * over the call to page_add_new_anon_rmap.
1089 : : */
1090 : : VM_BUG_ON_PAGE(page_anon_vma(page)->root != vma->anon_vma->root, page);
1091 : : VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1092 : : page);
1093 : : }
1094 : :
1095 : : /**
1096 : : * page_add_anon_rmap - add pte mapping to an anonymous page
1097 : : * @page: the page to add the mapping to
1098 : : * @vma: the vm area in which the mapping is added
1099 : : * @address: the user virtual address mapped
1100 : : * @compound: charge the page as compound or small page
1101 : : *
1102 : : * The caller needs to hold the pte lock, and the page must be locked in
1103 : : * the anon_vma case: to serialize mapping,index checking after setting,
1104 : : * and to ensure that PageAnon is not being upgraded racily to PageKsm
1105 : : * (but PageKsm is never downgraded to PageAnon).
1106 : : */
1107 : 0 : void page_add_anon_rmap(struct page *page,
1108 : : struct vm_area_struct *vma, unsigned long address, bool compound)
1109 : : {
1110 [ # # ]: 0 : do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1111 : 0 : }
1112 : :
1113 : : /*
1114 : : * Special version of the above for do_swap_page, which often runs
1115 : : * into pages that are exclusively owned by the current process.
1116 : : * Everybody else should continue to use page_add_anon_rmap above.
1117 : : */
1118 : 0 : void do_page_add_anon_rmap(struct page *page,
1119 : : struct vm_area_struct *vma, unsigned long address, int flags)
1120 : : {
1121 : 0 : bool compound = flags & RMAP_COMPOUND;
1122 : 0 : bool first;
1123 : :
1124 [ # # ]: 0 : if (compound) {
1125 : 0 : atomic_t *mapcount;
1126 : 0 : VM_BUG_ON_PAGE(!PageLocked(page), page);
1127 : 0 : VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1128 : 0 : mapcount = compound_mapcount_ptr(page);
1129 : 0 : first = atomic_inc_and_test(mapcount);
1130 : : } else {
1131 : 0 : first = atomic_inc_and_test(&page->_mapcount);
1132 : : }
1133 : :
1134 [ # # ]: 0 : if (first) {
1135 : 0 : int nr = compound ? hpage_nr_pages(page) : 1;
1136 : : /*
1137 : : * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1138 : : * these counters are not modified in interrupt context, and
1139 : : * pte lock(a spinlock) is held, which implies preemption
1140 : : * disabled.
1141 : : */
1142 [ # # ]: 0 : if (compound)
1143 : 0 : __inc_node_page_state(page, NR_ANON_THPS);
1144 : 0 : __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1145 : : }
1146 [ # # ]: 0 : if (unlikely(PageKsm(page)))
1147 : : return;
1148 : :
1149 : 0 : VM_BUG_ON_PAGE(!PageLocked(page), page);
1150 : :
1151 : : /* address might be in next vma when migration races vma_adjust */
1152 [ # # ]: 0 : if (first)
1153 : 0 : __page_set_anon_rmap(page, vma, address,
1154 : : flags & RMAP_EXCLUSIVE);
1155 : : else
1156 : : __page_check_anon_rmap(page, vma, address);
1157 : : }
1158 : :
1159 : : /**
1160 : : * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1161 : : * @page: the page to add the mapping to
1162 : : * @vma: the vm area in which the mapping is added
1163 : : * @address: the user virtual address mapped
1164 : : * @compound: charge the page as compound or small page
1165 : : *
1166 : : * Same as page_add_anon_rmap but must only be called on *new* pages.
1167 : : * This means the inc-and-test can be bypassed.
1168 : : * Page does not have to be locked.
1169 : : */
1170 : 1066236 : void page_add_new_anon_rmap(struct page *page,
1171 : : struct vm_area_struct *vma, unsigned long address, bool compound)
1172 : : {
1173 : 1066236 : int nr = compound ? hpage_nr_pages(page) : 1;
1174 : :
1175 : 1066236 : VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1176 [ - + ]: 1066236 : __SetPageSwapBacked(page);
1177 [ - + ]: 1066236 : if (compound) {
1178 : 0 : VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1179 : : /* increment count (starts at -1) */
1180 : 0 : atomic_set(compound_mapcount_ptr(page), 0);
1181 : 0 : __inc_node_page_state(page, NR_ANON_THPS);
1182 : : } else {
1183 : : /* Anon THP always mapped first with PMD */
1184 : 1066236 : VM_BUG_ON_PAGE(PageTransCompound(page), page);
1185 : : /* increment count (starts at -1) */
1186 : 1066236 : atomic_set(&page->_mapcount, 0);
1187 : : }
1188 : 1066236 : __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1189 : 1066236 : __page_set_anon_rmap(page, vma, address, 1);
1190 : 1066236 : }
1191 : :
1192 : : /**
1193 : : * page_add_file_rmap - add pte mapping to a file page
1194 : : * @page: the page to add the mapping to
1195 : : * @compound: charge the page as compound or small page
1196 : : *
1197 : : * The caller needs to hold the pte lock.
1198 : : */
1199 : 15664555 : void page_add_file_rmap(struct page *page, bool compound)
1200 : : {
1201 : 15664555 : int i, nr = 1;
1202 : :
1203 : 15664555 : VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1204 : 15664555 : lock_page_memcg(page);
1205 : 15664555 : if (compound && PageTransHuge(page)) {
1206 : : for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1207 : : if (atomic_inc_and_test(&page[i]._mapcount))
1208 : : nr++;
1209 : : }
1210 : : if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1211 : : goto out;
1212 : : if (PageSwapBacked(page))
1213 : : __inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1214 : : else
1215 : : __inc_node_page_state(page, NR_FILE_PMDMAPPED);
1216 : : } else {
1217 : 15664555 : if (PageTransCompound(page) && page_mapping(page)) {
1218 : : VM_WARN_ON_ONCE(!PageLocked(page));
1219 : :
1220 : : SetPageDoubleMap(compound_head(page));
1221 : : if (PageMlocked(page))
1222 : : clear_page_mlock(compound_head(page));
1223 : : }
1224 [ + + ]: 15664555 : if (!atomic_inc_and_test(&page->_mapcount))
1225 : 14805482 : goto out;
1226 : : }
1227 : 859073 : __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1228 : 15664555 : out:
1229 : 15664555 : unlock_page_memcg(page);
1230 : 15664555 : }
1231 : :
1232 : 15153327 : static void page_remove_file_rmap(struct page *page, bool compound)
1233 : : {
1234 : 15153327 : int i, nr = 1;
1235 : :
1236 : 15153327 : VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1237 : 15153327 : lock_page_memcg(page);
1238 : :
1239 : : /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1240 [ - + ]: 15153327 : if (unlikely(PageHuge(page))) {
1241 : : /* hugetlb pages are always mapped with pmds */
1242 : 0 : atomic_dec(compound_mapcount_ptr(page));
1243 : 0 : goto out;
1244 : : }
1245 : :
1246 : : /* page still mapped by someone else? */
1247 : 15153327 : if (compound && PageTransHuge(page)) {
1248 : : for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1249 : : if (atomic_add_negative(-1, &page[i]._mapcount))
1250 : : nr++;
1251 : : }
1252 : : if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1253 : : goto out;
1254 : : if (PageSwapBacked(page))
1255 : : __dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1256 : : else
1257 : : __dec_node_page_state(page, NR_FILE_PMDMAPPED);
1258 : : } else {
1259 [ + + ]: 15153327 : if (!atomic_add_negative(-1, &page->_mapcount))
1260 : 14419082 : goto out;
1261 : : }
1262 : :
1263 : : /*
1264 : : * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1265 : : * these counters are not modified in interrupt context, and
1266 : : * pte lock(a spinlock) is held, which implies preemption disabled.
1267 : : */
1268 : 734245 : __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1269 : :
1270 [ - + + - ]: 1468490 : if (unlikely(PageMlocked(page)))
1271 : 0 : clear_page_mlock(page);
1272 : 734245 : out:
1273 : 15153327 : unlock_page_memcg(page);
1274 : 15153327 : }
1275 : :
1276 : 0 : static void page_remove_anon_compound_rmap(struct page *page)
1277 : : {
1278 : 0 : int i, nr;
1279 : :
1280 [ # # ]: 0 : if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1281 : : return;
1282 : :
1283 : : /* Hugepages are not counted in NR_ANON_PAGES for now. */
1284 : 0 : if (unlikely(PageHuge(page)))
1285 : : return;
1286 : :
1287 : 0 : if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1288 : 0 : return;
1289 : :
1290 : : __dec_node_page_state(page, NR_ANON_THPS);
1291 : :
1292 : : if (TestClearPageDoubleMap(page)) {
1293 : : /*
1294 : : * Subpages can be mapped with PTEs too. Check how many of
1295 : : * them are still mapped.
1296 : : */
1297 : : for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1298 : : if (atomic_add_negative(-1, &page[i]._mapcount))
1299 : : nr++;
1300 : : }
1301 : :
1302 : : /*
1303 : : * Queue the page for deferred split if at least one small
1304 : : * page of the compound page is unmapped, but at least one
1305 : : * small page is still mapped.
1306 : : */
1307 : : if (nr && nr < HPAGE_PMD_NR)
1308 : : deferred_split_huge_page(page);
1309 : : } else {
1310 : : nr = HPAGE_PMD_NR;
1311 : : }
1312 : :
1313 : : if (unlikely(PageMlocked(page)))
1314 : : clear_page_mlock(page);
1315 : :
1316 : : if (nr)
1317 : : __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
1318 : : }
1319 : :
1320 : : /**
1321 : : * page_remove_rmap - take down pte mapping from a page
1322 : : * @page: page to remove mapping from
1323 : : * @compound: uncharge the page as compound or small page
1324 : : *
1325 : : * The caller needs to hold the pte lock.
1326 : : */
1327 : 17744708 : void page_remove_rmap(struct page *page, bool compound)
1328 : : {
1329 [ - + + + ]: 17744708 : if (!PageAnon(page))
1330 : 15153327 : return page_remove_file_rmap(page, compound);
1331 : :
1332 [ - + ]: 2591381 : if (compound)
1333 : 0 : return page_remove_anon_compound_rmap(page);
1334 : :
1335 : : /* page still mapped by someone else? */
1336 [ + + ]: 2591381 : if (!atomic_add_negative(-1, &page->_mapcount))
1337 : : return;
1338 : :
1339 : : /*
1340 : : * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1341 : : * these counters are not modified in interrupt context, and
1342 : : * pte lock(a spinlock) is held, which implies preemption disabled.
1343 : : */
1344 : 995174 : __dec_node_page_state(page, NR_ANON_MAPPED);
1345 : :
1346 [ - + - + ]: 1990348 : if (unlikely(PageMlocked(page)))
1347 : 0 : clear_page_mlock(page);
1348 : :
1349 : : if (PageTransCompound(page))
1350 : : deferred_split_huge_page(compound_head(page));
1351 : :
1352 : : /*
1353 : : * It would be tidy to reset the PageAnon mapping here,
1354 : : * but that might overwrite a racing page_add_anon_rmap
1355 : : * which increments mapcount after us but sets mapping
1356 : : * before us: so leave the reset to free_unref_page,
1357 : : * and remember that it's only reliable while mapped.
1358 : : * Leaving it set also helps swapoff to reinstate ptes
1359 : : * faster for those pages still in swapcache.
1360 : : */
1361 : : }
1362 : :
1363 : : /*
1364 : : * @arg: enum ttu_flags will be passed to this argument
1365 : : */
1366 : 0 : static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1367 : : unsigned long address, void *arg)
1368 : : {
1369 : 0 : struct mm_struct *mm = vma->vm_mm;
1370 : 0 : struct page_vma_mapped_walk pvmw = {
1371 : : .page = page,
1372 : : .vma = vma,
1373 : : .address = address,
1374 : : };
1375 : 0 : pte_t pteval;
1376 : 0 : struct page *subpage;
1377 : 0 : bool ret = true;
1378 : 0 : struct mmu_notifier_range range;
1379 : 0 : enum ttu_flags flags = (enum ttu_flags)arg;
1380 : :
1381 : : /* munlock has nothing to gain from examining un-locked vmas */
1382 [ # # # # ]: 0 : if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1383 : : return true;
1384 : :
1385 : 0 : if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1386 : : is_zone_device_page(page) && !is_device_private_page(page))
1387 : : return true;
1388 : :
1389 : 0 : if (flags & TTU_SPLIT_HUGE_PMD) {
1390 : : split_huge_pmd_address(vma, address,
1391 : : flags & TTU_SPLIT_FREEZE, page);
1392 : : }
1393 : :
1394 : : /*
1395 : : * For THP, we have to assume the worse case ie pmd for invalidation.
1396 : : * For hugetlb, it could be much worse if we need to do pud
1397 : : * invalidation in the case of pmd sharing.
1398 : : *
1399 : : * Note that the page can not be free in this function as call of
1400 : : * try_to_unmap() must hold a reference on the page.
1401 : : */
1402 : 0 : mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1403 : : address,
1404 : 0 : min(vma->vm_end, address + page_size(page)));
1405 [ # # ]: 0 : if (PageHuge(page)) {
1406 : : /*
1407 : : * If sharing is possible, start and end will be adjusted
1408 : : * accordingly.
1409 : : */
1410 : 0 : adjust_range_if_pmd_sharing_possible(vma, &range.start,
1411 : : &range.end);
1412 : : }
1413 : 0 : mmu_notifier_invalidate_range_start(&range);
1414 : :
1415 [ # # ]: 0 : while (page_vma_mapped_walk(&pvmw)) {
1416 : : #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1417 : : /* PMD-mapped THP migration entry */
1418 : : if (!pvmw.pte && (flags & TTU_MIGRATION)) {
1419 : : VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1420 : :
1421 : : set_pmd_migration_entry(&pvmw, page);
1422 : : continue;
1423 : : }
1424 : : #endif
1425 : :
1426 : : /*
1427 : : * If the page is mlock()d, we cannot swap it out.
1428 : : * If it's recently referenced (perhaps page_referenced
1429 : : * skipped over this mm) then we should reactivate it.
1430 : : */
1431 [ # # ]: 0 : if (!(flags & TTU_IGNORE_MLOCK)) {
1432 [ # # ]: 0 : if (vma->vm_flags & VM_LOCKED) {
1433 : : /* PTE-mapped THP are never mlocked */
1434 : 0 : if (!PageTransCompound(page)) {
1435 : : /*
1436 : : * Holding pte lock, we do *not* need
1437 : : * mmap_sem here
1438 : : */
1439 : 0 : mlock_vma_page(page);
1440 : : }
1441 : 0 : ret = false;
1442 [ # # ]: 0 : page_vma_mapped_walk_done(&pvmw);
1443 : : break;
1444 : : }
1445 [ # # ]: 0 : if (flags & TTU_MUNLOCK)
1446 : 0 : continue;
1447 : : }
1448 : :
1449 : : /* Unexpected PMD-mapped THP? */
1450 : 0 : VM_BUG_ON_PAGE(!pvmw.pte, page);
1451 : :
1452 [ # # ]: 0 : subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1453 : 0 : address = pvmw.address;
1454 : :
1455 [ # # ]: 0 : if (PageHuge(page)) {
1456 [ # # ]: 0 : if (huge_pmd_unshare(mm, &address, pvmw.pte)) {
1457 : : /*
1458 : : * huge_pmd_unshare unmapped an entire PMD
1459 : : * page. There is no way of knowing exactly
1460 : : * which PMDs may be cached for this mm, so
1461 : : * we must flush them all. start/end were
1462 : : * already adjusted above to cover this range.
1463 : : */
1464 [ # # ]: 0 : flush_cache_range(vma, range.start, range.end);
1465 [ # # ]: 0 : flush_tlb_range(vma, range.start, range.end);
1466 [ # # ]: 0 : mmu_notifier_invalidate_range(mm, range.start,
1467 : : range.end);
1468 : :
1469 : : /*
1470 : : * The ref count of the PMD page was dropped
1471 : : * which is part of the way map counting
1472 : : * is done for shared PMDs. Return 'true'
1473 : : * here. When there is no other sharing,
1474 : : * huge_pmd_unshare returns false and we will
1475 : : * unmap the actual page and drop map count
1476 : : * to zero.
1477 : : */
1478 [ # # ]: 0 : page_vma_mapped_walk_done(&pvmw);
1479 : : break;
1480 : : }
1481 : : }
1482 : :
1483 : 0 : if (IS_ENABLED(CONFIG_MIGRATION) &&
1484 : : (flags & TTU_MIGRATION) &&
1485 : : is_zone_device_page(page)) {
1486 : : swp_entry_t entry;
1487 : : pte_t swp_pte;
1488 : :
1489 : : pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
1490 : :
1491 : : /*
1492 : : * Store the pfn of the page in a special migration
1493 : : * pte. do_swap_page() will wait until the migration
1494 : : * pte is removed and then restart fault handling.
1495 : : */
1496 : : entry = make_migration_entry(page, 0);
1497 : : swp_pte = swp_entry_to_pte(entry);
1498 : : if (pte_soft_dirty(pteval))
1499 : : swp_pte = pte_swp_mksoft_dirty(swp_pte);
1500 : : set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1501 : : /*
1502 : : * No need to invalidate here it will synchronize on
1503 : : * against the special swap migration pte.
1504 : : *
1505 : : * The assignment to subpage above was computed from a
1506 : : * swap PTE which results in an invalid pointer.
1507 : : * Since only PAGE_SIZE pages can currently be
1508 : : * migrated, just set it to page. This will need to be
1509 : : * changed when hugepage migrations to device private
1510 : : * memory are supported.
1511 : : */
1512 : : subpage = page;
1513 : : goto discard;
1514 : : }
1515 : :
1516 [ # # ]: 0 : if (!(flags & TTU_IGNORE_ACCESS)) {
1517 [ # # # # ]: 0 : if (ptep_clear_flush_young_notify(vma, address,
1518 : : pvmw.pte)) {
1519 : 0 : ret = false;
1520 [ # # ]: 0 : page_vma_mapped_walk_done(&pvmw);
1521 : : break;
1522 : : }
1523 : : }
1524 : :
1525 : : /* Nuke the page table entry. */
1526 : 0 : flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1527 [ # # ]: 0 : if (should_defer_flush(mm, flags)) {
1528 : : /*
1529 : : * We clear the PTE but do not flush so potentially
1530 : : * a remote CPU could still be writing to the page.
1531 : : * If the entry was previously clean then the
1532 : : * architecture must guarantee that a clear->dirty
1533 : : * transition on a cached TLB entry is written through
1534 : : * and traps if the PTE is unmapped.
1535 : : */
1536 : 0 : pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1537 : :
1538 : 0 : set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1539 : : } else {
1540 : 0 : pteval = ptep_clear_flush(vma, address, pvmw.pte);
1541 : : }
1542 : :
1543 : : /* Move the dirty bit to the page. Now the pte is gone. */
1544 [ # # ]: 0 : if (pte_dirty(pteval))
1545 : 0 : set_page_dirty(page);
1546 : :
1547 : : /* Update high watermark before we lower rss */
1548 : 0 : update_hiwater_rss(mm);
1549 : :
1550 [ # # ]: 0 : if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1551 : : pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1552 : : if (PageHuge(page)) {
1553 : : hugetlb_count_sub(compound_nr(page), mm);
1554 : : set_huge_swap_pte_at(mm, address,
1555 : : pvmw.pte, pteval,
1556 : : vma_mmu_pagesize(vma));
1557 : : } else {
1558 : : dec_mm_counter(mm, mm_counter(page));
1559 : : set_pte_at(mm, address, pvmw.pte, pteval);
1560 : : }
1561 : :
1562 [ # # ]: 0 : } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1563 : : /*
1564 : : * The guest indicated that the page content is of no
1565 : : * interest anymore. Simply discard the pte, vmscan
1566 : : * will take care of the rest.
1567 : : * A future reference will then fault in a new zero
1568 : : * page. When userfaultfd is active, we must not drop
1569 : : * this page though, as its main user (postcopy
1570 : : * migration) will not expect userfaults on already
1571 : : * copied pages.
1572 : : */
1573 : : dec_mm_counter(mm, mm_counter(page));
1574 : : /* We have to invalidate as we cleared the pte */
1575 : : mmu_notifier_invalidate_range(mm, address,
1576 : : address + PAGE_SIZE);
1577 : 0 : } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1578 [ # # ]: 0 : (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
1579 : 0 : swp_entry_t entry;
1580 : 0 : pte_t swp_pte;
1581 : :
1582 : 0 : if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1583 : : set_pte_at(mm, address, pvmw.pte, pteval);
1584 : : ret = false;
1585 : : page_vma_mapped_walk_done(&pvmw);
1586 : : break;
1587 : : }
1588 : :
1589 : : /*
1590 : : * Store the pfn of the page in a special migration
1591 : : * pte. do_swap_page() will wait until the migration
1592 : : * pte is removed and then restart fault handling.
1593 : : */
1594 : 0 : entry = make_migration_entry(subpage,
1595 : : pte_write(pteval));
1596 : 0 : swp_pte = swp_entry_to_pte(entry);
1597 : 0 : if (pte_soft_dirty(pteval))
1598 : : swp_pte = pte_swp_mksoft_dirty(swp_pte);
1599 : 0 : set_pte_at(mm, address, pvmw.pte, swp_pte);
1600 : : /*
1601 : : * No need to invalidate here it will synchronize on
1602 : : * against the special swap migration pte.
1603 : : */
1604 [ # # # # ]: 0 : } else if (PageAnon(page)) {
1605 : 0 : swp_entry_t entry = { .val = page_private(subpage) };
1606 : 0 : pte_t swp_pte;
1607 : : /*
1608 : : * Store the swap location in the pte.
1609 : : * See handle_pte_fault() ...
1610 : : */
1611 [ # # # # ]: 0 : if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1612 : 0 : WARN_ON_ONCE(1);
1613 : 0 : ret = false;
1614 : : /* We have to invalidate as we cleared the pte */
1615 [ # # ]: 0 : mmu_notifier_invalidate_range(mm, address,
1616 : : address + PAGE_SIZE);
1617 [ # # ]: 0 : page_vma_mapped_walk_done(&pvmw);
1618 : : break;
1619 : : }
1620 : :
1621 : : /* MADV_FREE page check */
1622 [ # # # # ]: 0 : if (!PageSwapBacked(page)) {
1623 [ # # # # ]: 0 : if (!PageDirty(page)) {
1624 : : /* Invalidate as we cleared the pte */
1625 [ # # ]: 0 : mmu_notifier_invalidate_range(mm,
1626 : : address, address + PAGE_SIZE);
1627 : 0 : dec_mm_counter(mm, MM_ANONPAGES);
1628 : 0 : goto discard;
1629 : : }
1630 : :
1631 : : /*
1632 : : * If the page was redirtied, it cannot be
1633 : : * discarded. Remap the page to page table.
1634 : : */
1635 [ # # ]: 0 : set_pte_at(mm, address, pvmw.pte, pteval);
1636 [ # # ]: 0 : SetPageSwapBacked(page);
1637 : 0 : ret = false;
1638 [ # # ]: 0 : page_vma_mapped_walk_done(&pvmw);
1639 : : break;
1640 : : }
1641 : :
1642 [ # # ]: 0 : if (swap_duplicate(entry) < 0) {
1643 [ # # ]: 0 : set_pte_at(mm, address, pvmw.pte, pteval);
1644 : 0 : ret = false;
1645 [ # # ]: 0 : page_vma_mapped_walk_done(&pvmw);
1646 : : break;
1647 : : }
1648 [ # # ]: 0 : if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1649 : : set_pte_at(mm, address, pvmw.pte, pteval);
1650 : : ret = false;
1651 : : page_vma_mapped_walk_done(&pvmw);
1652 : : break;
1653 : : }
1654 [ # # ]: 0 : if (list_empty(&mm->mmlist)) {
1655 : 0 : spin_lock(&mmlist_lock);
1656 [ # # ]: 0 : if (list_empty(&mm->mmlist))
1657 : 0 : list_add(&mm->mmlist, &init_mm.mmlist);
1658 : 0 : spin_unlock(&mmlist_lock);
1659 : : }
1660 : 0 : dec_mm_counter(mm, MM_ANONPAGES);
1661 : 0 : inc_mm_counter(mm, MM_SWAPENTS);
1662 [ # # ]: 0 : swp_pte = swp_entry_to_pte(entry);
1663 [ # # ]: 0 : if (pte_soft_dirty(pteval))
1664 : : swp_pte = pte_swp_mksoft_dirty(swp_pte);
1665 [ # # ]: 0 : set_pte_at(mm, address, pvmw.pte, swp_pte);
1666 : : /* Invalidate as we cleared the pte */
1667 [ # # ]: 0 : mmu_notifier_invalidate_range(mm, address,
1668 : : address + PAGE_SIZE);
1669 : : } else {
1670 : : /*
1671 : : * This is a locked file-backed page, thus it cannot
1672 : : * be removed from the page cache and replaced by a new
1673 : : * page before mmu_notifier_invalidate_range_end, so no
1674 : : * concurrent thread might update its page table to
1675 : : * point at new page while a device still is using this
1676 : : * page.
1677 : : *
1678 : : * See Documentation/vm/mmu_notifier.rst
1679 : : */
1680 : 0 : dec_mm_counter(mm, mm_counter_file(page));
1681 : : }
1682 : 0 : discard:
1683 : : /*
1684 : : * No need to call mmu_notifier_invalidate_range() it has be
1685 : : * done above for all cases requiring it to happen under page
1686 : : * table lock before mmu_notifier_invalidate_range_end()
1687 : : *
1688 : : * See Documentation/vm/mmu_notifier.rst
1689 : : */
1690 : 0 : page_remove_rmap(subpage, PageHuge(page));
1691 : 0 : put_page(page);
1692 : : }
1693 : :
1694 : 0 : mmu_notifier_invalidate_range_end(&range);
1695 : :
1696 : 0 : return ret;
1697 : : }
1698 : :
1699 : 0 : bool is_vma_temporary_stack(struct vm_area_struct *vma)
1700 : : {
1701 : 0 : int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1702 : :
1703 [ # # ]: 0 : if (!maybe_stack)
1704 : : return false;
1705 : :
1706 [ # # # # ]: 0 : if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1707 : : VM_STACK_INCOMPLETE_SETUP)
1708 : 0 : return true;
1709 : :
1710 : : return false;
1711 : : }
1712 : :
1713 : 0 : static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1714 : : {
1715 [ # # ]: 0 : return is_vma_temporary_stack(vma);
1716 : : }
1717 : :
1718 : 0 : static int page_mapcount_is_zero(struct page *page)
1719 : : {
1720 : 0 : return !total_mapcount(page);
1721 : : }
1722 : :
1723 : : /**
1724 : : * try_to_unmap - try to remove all page table mappings to a page
1725 : : * @page: the page to get unmapped
1726 : : * @flags: action and flags
1727 : : *
1728 : : * Tries to remove all the page table entries which are mapping this
1729 : : * page, used in the pageout path. Caller must hold the page lock.
1730 : : *
1731 : : * If unmap is successful, return true. Otherwise, false.
1732 : : */
1733 : 0 : bool try_to_unmap(struct page *page, enum ttu_flags flags)
1734 : : {
1735 : 0 : struct rmap_walk_control rwc = {
1736 : : .rmap_one = try_to_unmap_one,
1737 : 0 : .arg = (void *)flags,
1738 : : .done = page_mapcount_is_zero,
1739 : : .anon_lock = page_lock_anon_vma_read,
1740 : : };
1741 : :
1742 : : /*
1743 : : * During exec, a temporary VMA is setup and later moved.
1744 : : * The VMA is moved under the anon_vma lock but not the
1745 : : * page tables leading to a race where migration cannot
1746 : : * find the migration ptes. Rather than increasing the
1747 : : * locking requirements of exec(), migration skips
1748 : : * temporary VMAs until after exec() completes.
1749 : : */
1750 [ # # ]: 0 : if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
1751 [ # # ]: 0 : && !PageKsm(page) && PageAnon(page))
1752 : 0 : rwc.invalid_vma = invalid_migration_vma;
1753 : :
1754 [ # # ]: 0 : if (flags & TTU_RMAP_LOCKED)
1755 : 0 : rmap_walk_locked(page, &rwc);
1756 : : else
1757 : 0 : rmap_walk(page, &rwc);
1758 : :
1759 : 0 : return !page_mapcount(page) ? true : false;
1760 : : }
1761 : :
1762 : 0 : static int page_not_mapped(struct page *page)
1763 : : {
1764 : 0 : return !page_mapped(page);
1765 : : };
1766 : :
1767 : : /**
1768 : : * try_to_munlock - try to munlock a page
1769 : : * @page: the page to be munlocked
1770 : : *
1771 : : * Called from munlock code. Checks all of the VMAs mapping the page
1772 : : * to make sure nobody else has this page mlocked. The page will be
1773 : : * returned with PG_mlocked cleared if no other vmas have it mlocked.
1774 : : */
1775 : :
1776 : 0 : void try_to_munlock(struct page *page)
1777 : : {
1778 : 0 : struct rmap_walk_control rwc = {
1779 : : .rmap_one = try_to_unmap_one,
1780 : : .arg = (void *)TTU_MUNLOCK,
1781 : : .done = page_not_mapped,
1782 : : .anon_lock = page_lock_anon_vma_read,
1783 : :
1784 : : };
1785 : :
1786 : 0 : VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1787 : 0 : VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
1788 : :
1789 : 0 : rmap_walk(page, &rwc);
1790 : 0 : }
1791 : :
1792 : 642035 : void __put_anon_vma(struct anon_vma *anon_vma)
1793 : : {
1794 : 642035 : struct anon_vma *root = anon_vma->root;
1795 : :
1796 : 642035 : anon_vma_free(anon_vma);
1797 [ + + - + ]: 642035 : if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1798 : 0 : anon_vma_free(root);
1799 : 642035 : }
1800 : :
1801 : : static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1802 : : struct rmap_walk_control *rwc)
1803 : : {
1804 : : struct anon_vma *anon_vma;
1805 : :
1806 : : if (rwc->anon_lock)
1807 : : return rwc->anon_lock(page);
1808 : :
1809 : : /*
1810 : : * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1811 : : * because that depends on page_mapped(); but not all its usages
1812 : : * are holding mmap_sem. Users without mmap_sem are required to
1813 : : * take a reference count to prevent the anon_vma disappearing
1814 : : */
1815 : : anon_vma = page_anon_vma(page);
1816 : : if (!anon_vma)
1817 : : return NULL;
1818 : :
1819 : : anon_vma_lock_read(anon_vma);
1820 : : return anon_vma;
1821 : : }
1822 : :
1823 : : /*
1824 : : * rmap_walk_anon - do something to anonymous page using the object-based
1825 : : * rmap method
1826 : : * @page: the page to be handled
1827 : : * @rwc: control variable according to each walk type
1828 : : *
1829 : : * Find all the mappings of a page using the mapping pointer and the vma chains
1830 : : * contained in the anon_vma struct it points to.
1831 : : *
1832 : : * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1833 : : * where the page was found will be held for write. So, we won't recheck
1834 : : * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1835 : : * LOCKED.
1836 : : */
1837 : 0 : static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1838 : : bool locked)
1839 : : {
1840 : 0 : struct anon_vma *anon_vma;
1841 : 0 : pgoff_t pgoff_start, pgoff_end;
1842 : 0 : struct anon_vma_chain *avc;
1843 : :
1844 [ # # ]: 0 : if (locked) {
1845 : 0 : anon_vma = page_anon_vma(page);
1846 : : /* anon_vma disappear under us? */
1847 : 0 : VM_BUG_ON_PAGE(!anon_vma, page);
1848 : : } else {
1849 : 0 : anon_vma = rmap_walk_anon_lock(page, rwc);
1850 : : }
1851 [ # # ]: 0 : if (!anon_vma)
1852 : : return;
1853 : :
1854 : 0 : pgoff_start = page_to_pgoff(page);
1855 : 0 : pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1856 [ # # ]: 0 : anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1857 : : pgoff_start, pgoff_end) {
1858 : 0 : struct vm_area_struct *vma = avc->vma;
1859 : 0 : unsigned long address = vma_address(page, vma);
1860 : :
1861 : 0 : cond_resched();
1862 : :
1863 [ # # # # ]: 0 : if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1864 : 0 : continue;
1865 : :
1866 [ # # ]: 0 : if (!rwc->rmap_one(page, vma, address, rwc->arg))
1867 : : break;
1868 [ # # # # ]: 0 : if (rwc->done && rwc->done(page))
1869 : : break;
1870 : : }
1871 : :
1872 [ # # ]: 0 : if (!locked)
1873 : 0 : anon_vma_unlock_read(anon_vma);
1874 : : }
1875 : :
1876 : : /*
1877 : : * rmap_walk_file - do something to file page using the object-based rmap method
1878 : : * @page: the page to be handled
1879 : : * @rwc: control variable according to each walk type
1880 : : *
1881 : : * Find all the mappings of a page using the mapping pointer and the vma chains
1882 : : * contained in the address_space struct it points to.
1883 : : *
1884 : : * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1885 : : * where the page was found will be held for write. So, we won't recheck
1886 : : * vm_flags for that VMA. That should be OK, because that vma shouldn't be
1887 : : * LOCKED.
1888 : : */
1889 : 0 : static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1890 : : bool locked)
1891 : : {
1892 : 0 : struct address_space *mapping = page_mapping(page);
1893 : 0 : pgoff_t pgoff_start, pgoff_end;
1894 : 0 : struct vm_area_struct *vma;
1895 : :
1896 : : /*
1897 : : * The page lock not only makes sure that page->mapping cannot
1898 : : * suddenly be NULLified by truncation, it makes sure that the
1899 : : * structure at mapping cannot be freed and reused yet,
1900 : : * so we can safely take mapping->i_mmap_rwsem.
1901 : : */
1902 : 0 : VM_BUG_ON_PAGE(!PageLocked(page), page);
1903 : :
1904 [ # # ]: 0 : if (!mapping)
1905 : : return;
1906 : :
1907 : 0 : pgoff_start = page_to_pgoff(page);
1908 : 0 : pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1909 [ # # ]: 0 : if (!locked)
1910 : 0 : i_mmap_lock_read(mapping);
1911 [ # # ]: 0 : vma_interval_tree_foreach(vma, &mapping->i_mmap,
1912 : : pgoff_start, pgoff_end) {
1913 : 0 : unsigned long address = vma_address(page, vma);
1914 : :
1915 : 0 : cond_resched();
1916 : :
1917 [ # # # # ]: 0 : if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1918 : 0 : continue;
1919 : :
1920 [ # # ]: 0 : if (!rwc->rmap_one(page, vma, address, rwc->arg))
1921 : 0 : goto done;
1922 [ # # # # ]: 0 : if (rwc->done && rwc->done(page))
1923 : 0 : goto done;
1924 : : }
1925 : :
1926 : 0 : done:
1927 [ # # ]: 0 : if (!locked)
1928 : 0 : i_mmap_unlock_read(mapping);
1929 : : }
1930 : :
1931 : 0 : void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1932 : : {
1933 [ # # ]: 0 : if (unlikely(PageKsm(page)))
1934 : : rmap_walk_ksm(page, rwc);
1935 [ # # # # ]: 0 : else if (PageAnon(page))
1936 : 0 : rmap_walk_anon(page, rwc, false);
1937 : : else
1938 : 0 : rmap_walk_file(page, rwc, false);
1939 : 0 : }
1940 : :
1941 : : /* Like rmap_walk, but caller holds relevant rmap lock */
1942 : 0 : void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1943 : : {
1944 : : /* no ksm support for now */
1945 : 0 : VM_BUG_ON_PAGE(PageKsm(page), page);
1946 [ # # # # ]: 0 : if (PageAnon(page))
1947 : 0 : rmap_walk_anon(page, rwc, true);
1948 : : else
1949 : 0 : rmap_walk_file(page, rwc, true);
1950 : 0 : }
1951 : :
1952 : : #ifdef CONFIG_HUGETLB_PAGE
1953 : : /*
1954 : : * The following two functions are for anonymous (private mapped) hugepages.
1955 : : * Unlike common anonymous pages, anonymous hugepages have no accounting code
1956 : : * and no lru code, because we handle hugepages differently from common pages.
1957 : : */
1958 : 0 : void hugepage_add_anon_rmap(struct page *page,
1959 : : struct vm_area_struct *vma, unsigned long address)
1960 : : {
1961 : 0 : struct anon_vma *anon_vma = vma->anon_vma;
1962 : 0 : int first;
1963 : :
1964 [ # # # # ]: 0 : BUG_ON(!PageLocked(page));
1965 [ # # ]: 0 : BUG_ON(!anon_vma);
1966 : : /* address might be in next vma when migration races vma_adjust */
1967 : 0 : first = atomic_inc_and_test(compound_mapcount_ptr(page));
1968 [ # # ]: 0 : if (first)
1969 : 0 : __page_set_anon_rmap(page, vma, address, 0);
1970 : 0 : }
1971 : :
1972 : 0 : void hugepage_add_new_anon_rmap(struct page *page,
1973 : : struct vm_area_struct *vma, unsigned long address)
1974 : : {
1975 [ # # # # ]: 0 : BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1976 : 0 : atomic_set(compound_mapcount_ptr(page), 0);
1977 : 0 : __page_set_anon_rmap(page, vma, address, 1);
1978 : 0 : }
1979 : : #endif /* CONFIG_HUGETLB_PAGE */
|
