]>
Commit | Line | Data |
---|---|---|
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 | * folio_lock | |
27 | * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share, see hugetlbfs below) | |
28 | * vma_start_write | |
29 | * mapping->i_mmap_rwsem | |
30 | * anon_vma->rwsem | |
31 | * mm->page_table_lock or pte_lock | |
32 | * swap_lock (in swap_duplicate, swap_info_get) | |
33 | * mmlist_lock (in mmput, drain_mmlist and others) | |
34 | * mapping->private_lock (in block_dirty_folio) | |
35 | * 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 | * folio_lock | |
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 | #include <linux/oom.h> | |
78 | ||
79 | #include <asm/tlbflush.h> | |
80 | ||
81 | #define CREATE_TRACE_POINTS | |
82 | #include <trace/events/tlb.h> | |
83 | #include <trace/events/migrate.h> | |
84 | ||
85 | #include "internal.h" | |
86 | ||
87 | static struct kmem_cache *anon_vma_cachep; | |
88 | static struct kmem_cache *anon_vma_chain_cachep; | |
89 | ||
90 | static inline struct anon_vma *anon_vma_alloc(void) | |
91 | { | |
92 | struct anon_vma *anon_vma; | |
93 | ||
94 | anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); | |
95 | if (anon_vma) { | |
96 | atomic_set(&anon_vma->refcount, 1); | |
97 | anon_vma->num_children = 0; | |
98 | anon_vma->num_active_vmas = 0; | |
99 | anon_vma->parent = anon_vma; | |
100 | /* | |
101 | * Initialise the anon_vma root to point to itself. If called | |
102 | * from fork, the root will be reset to the parents anon_vma. | |
103 | */ | |
104 | anon_vma->root = anon_vma; | |
105 | } | |
106 | ||
107 | return anon_vma; | |
108 | } | |
109 | ||
110 | static inline void anon_vma_free(struct anon_vma *anon_vma) | |
111 | { | |
112 | VM_BUG_ON(atomic_read(&anon_vma->refcount)); | |
113 | ||
114 | /* | |
115 | * Synchronize against folio_lock_anon_vma_read() such that | |
116 | * we can safely hold the lock without the anon_vma getting | |
117 | * freed. | |
118 | * | |
119 | * Relies on the full mb implied by the atomic_dec_and_test() from | |
120 | * put_anon_vma() against the acquire barrier implied by | |
121 | * down_read_trylock() from folio_lock_anon_vma_read(). This orders: | |
122 | * | |
123 | * folio_lock_anon_vma_read() VS put_anon_vma() | |
124 | * down_read_trylock() atomic_dec_and_test() | |
125 | * LOCK MB | |
126 | * atomic_read() rwsem_is_locked() | |
127 | * | |
128 | * LOCK should suffice since the actual taking of the lock must | |
129 | * happen _before_ what follows. | |
130 | */ | |
131 | might_sleep(); | |
132 | if (rwsem_is_locked(&anon_vma->root->rwsem)) { | |
133 | anon_vma_lock_write(anon_vma); | |
134 | anon_vma_unlock_write(anon_vma); | |
135 | } | |
136 | ||
137 | kmem_cache_free(anon_vma_cachep, anon_vma); | |
138 | } | |
139 | ||
140 | static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp) | |
141 | { | |
142 | return kmem_cache_alloc(anon_vma_chain_cachep, gfp); | |
143 | } | |
144 | ||
145 | static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) | |
146 | { | |
147 | kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); | |
148 | } | |
149 | ||
150 | static void anon_vma_chain_link(struct vm_area_struct *vma, | |
151 | struct anon_vma_chain *avc, | |
152 | struct anon_vma *anon_vma) | |
153 | { | |
154 | avc->vma = vma; | |
155 | avc->anon_vma = anon_vma; | |
156 | list_add(&avc->same_vma, &vma->anon_vma_chain); | |
157 | anon_vma_interval_tree_insert(avc, &anon_vma->rb_root); | |
158 | } | |
159 | ||
160 | /** | |
161 | * __anon_vma_prepare - attach an anon_vma to a memory region | |
162 | * @vma: the memory region in question | |
163 | * | |
164 | * This makes sure the memory mapping described by 'vma' has | |
165 | * an 'anon_vma' attached to it, so that we can associate the | |
166 | * anonymous pages mapped into it with that anon_vma. | |
167 | * | |
168 | * The common case will be that we already have one, which | |
169 | * is handled inline by anon_vma_prepare(). But if | |
170 | * not we either need to find an adjacent mapping that we | |
171 | * can re-use the anon_vma from (very common when the only | |
172 | * reason for splitting a vma has been mprotect()), or we | |
173 | * allocate a new one. | |
174 | * | |
175 | * Anon-vma allocations are very subtle, because we may have | |
176 | * optimistically looked up an anon_vma in folio_lock_anon_vma_read() | |
177 | * and that may actually touch the rwsem even in the newly | |
178 | * allocated vma (it depends on RCU to make sure that the | |
179 | * anon_vma isn't actually destroyed). | |
180 | * | |
181 | * As a result, we need to do proper anon_vma locking even | |
182 | * for the new allocation. At the same time, we do not want | |
183 | * to do any locking for the common case of already having | |
184 | * an anon_vma. | |
185 | */ | |
186 | int __anon_vma_prepare(struct vm_area_struct *vma) | |
187 | { | |
188 | struct mm_struct *mm = vma->vm_mm; | |
189 | struct anon_vma *anon_vma, *allocated; | |
190 | struct anon_vma_chain *avc; | |
191 | ||
192 | mmap_assert_locked(mm); | |
193 | might_sleep(); | |
194 | ||
195 | avc = anon_vma_chain_alloc(GFP_KERNEL); | |
196 | if (!avc) | |
197 | goto out_enomem; | |
198 | ||
199 | anon_vma = find_mergeable_anon_vma(vma); | |
200 | allocated = NULL; | |
201 | if (!anon_vma) { | |
202 | anon_vma = anon_vma_alloc(); | |
203 | if (unlikely(!anon_vma)) | |
204 | goto out_enomem_free_avc; | |
205 | anon_vma->num_children++; /* self-parent link for new root */ | |
206 | allocated = anon_vma; | |
207 | } | |
208 | ||
209 | anon_vma_lock_write(anon_vma); | |
210 | /* page_table_lock to protect against threads */ | |
211 | spin_lock(&mm->page_table_lock); | |
212 | if (likely(!vma->anon_vma)) { | |
213 | vma->anon_vma = anon_vma; | |
214 | anon_vma_chain_link(vma, avc, anon_vma); | |
215 | anon_vma->num_active_vmas++; | |
216 | allocated = NULL; | |
217 | avc = NULL; | |
218 | } | |
219 | spin_unlock(&mm->page_table_lock); | |
220 | anon_vma_unlock_write(anon_vma); | |
221 | ||
222 | if (unlikely(allocated)) | |
223 | put_anon_vma(allocated); | |
224 | if (unlikely(avc)) | |
225 | anon_vma_chain_free(avc); | |
226 | ||
227 | return 0; | |
228 | ||
229 | out_enomem_free_avc: | |
230 | anon_vma_chain_free(avc); | |
231 | out_enomem: | |
232 | return -ENOMEM; | |
233 | } | |
234 | ||
235 | /* | |
236 | * This is a useful helper function for locking the anon_vma root as | |
237 | * we traverse the vma->anon_vma_chain, looping over anon_vma's that | |
238 | * have the same vma. | |
239 | * | |
240 | * Such anon_vma's should have the same root, so you'd expect to see | |
241 | * just a single mutex_lock for the whole traversal. | |
242 | */ | |
243 | static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma) | |
244 | { | |
245 | struct anon_vma *new_root = anon_vma->root; | |
246 | if (new_root != root) { | |
247 | if (WARN_ON_ONCE(root)) | |
248 | up_write(&root->rwsem); | |
249 | root = new_root; | |
250 | down_write(&root->rwsem); | |
251 | } | |
252 | return root; | |
253 | } | |
254 | ||
255 | static inline void unlock_anon_vma_root(struct anon_vma *root) | |
256 | { | |
257 | if (root) | |
258 | up_write(&root->rwsem); | |
259 | } | |
260 | ||
261 | /* | |
262 | * Attach the anon_vmas from src to dst. | |
263 | * Returns 0 on success, -ENOMEM on failure. | |
264 | * | |
265 | * anon_vma_clone() is called by vma_expand(), vma_merge(), __split_vma(), | |
266 | * copy_vma() and anon_vma_fork(). The first four want an exact copy of src, | |
267 | * while the last one, anon_vma_fork(), may try to reuse an existing anon_vma to | |
268 | * prevent endless growth of anon_vma. Since dst->anon_vma is set to NULL before | |
269 | * call, we can identify this case by checking (!dst->anon_vma && | |
270 | * src->anon_vma). | |
271 | * | |
272 | * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find | |
273 | * and reuse existing anon_vma which has no vmas and only one child anon_vma. | |
274 | * This prevents degradation of anon_vma hierarchy to endless linear chain in | |
275 | * case of constantly forking task. On the other hand, an anon_vma with more | |
276 | * than one child isn't reused even if there was no alive vma, thus rmap | |
277 | * walker has a good chance of avoiding scanning the whole hierarchy when it | |
278 | * searches where page is mapped. | |
279 | */ | |
280 | int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) | |
281 | { | |
282 | struct anon_vma_chain *avc, *pavc; | |
283 | struct anon_vma *root = NULL; | |
284 | ||
285 | list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { | |
286 | struct anon_vma *anon_vma; | |
287 | ||
288 | avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN); | |
289 | if (unlikely(!avc)) { | |
290 | unlock_anon_vma_root(root); | |
291 | root = NULL; | |
292 | avc = anon_vma_chain_alloc(GFP_KERNEL); | |
293 | if (!avc) | |
294 | goto enomem_failure; | |
295 | } | |
296 | anon_vma = pavc->anon_vma; | |
297 | root = lock_anon_vma_root(root, anon_vma); | |
298 | anon_vma_chain_link(dst, avc, anon_vma); | |
299 | ||
300 | /* | |
301 | * Reuse existing anon_vma if it has no vma and only one | |
302 | * anon_vma child. | |
303 | * | |
304 | * Root anon_vma is never reused: | |
305 | * it has self-parent reference and at least one child. | |
306 | */ | |
307 | if (!dst->anon_vma && src->anon_vma && | |
308 | anon_vma->num_children < 2 && | |
309 | anon_vma->num_active_vmas == 0) | |
310 | dst->anon_vma = anon_vma; | |
311 | } | |
312 | if (dst->anon_vma) | |
313 | dst->anon_vma->num_active_vmas++; | |
314 | unlock_anon_vma_root(root); | |
315 | return 0; | |
316 | ||
317 | enomem_failure: | |
318 | /* | |
319 | * dst->anon_vma is dropped here otherwise its num_active_vmas can | |
320 | * be incorrectly decremented in unlink_anon_vmas(). | |
321 | * We can safely do this because callers of anon_vma_clone() don't care | |
322 | * about dst->anon_vma if anon_vma_clone() failed. | |
323 | */ | |
324 | dst->anon_vma = NULL; | |
325 | unlink_anon_vmas(dst); | |
326 | return -ENOMEM; | |
327 | } | |
328 | ||
329 | /* | |
330 | * Attach vma to its own anon_vma, as well as to the anon_vmas that | |
331 | * the corresponding VMA in the parent process is attached to. | |
332 | * Returns 0 on success, non-zero on failure. | |
333 | */ | |
334 | int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) | |
335 | { | |
336 | struct anon_vma_chain *avc; | |
337 | struct anon_vma *anon_vma; | |
338 | int error; | |
339 | ||
340 | /* Don't bother if the parent process has no anon_vma here. */ | |
341 | if (!pvma->anon_vma) | |
342 | return 0; | |
343 | ||
344 | /* Drop inherited anon_vma, we'll reuse existing or allocate new. */ | |
345 | vma->anon_vma = NULL; | |
346 | ||
347 | /* | |
348 | * First, attach the new VMA to the parent VMA's anon_vmas, | |
349 | * so rmap can find non-COWed pages in child processes. | |
350 | */ | |
351 | error = anon_vma_clone(vma, pvma); | |
352 | if (error) | |
353 | return error; | |
354 | ||
355 | /* An existing anon_vma has been reused, all done then. */ | |
356 | if (vma->anon_vma) | |
357 | return 0; | |
358 | ||
359 | /* Then add our own anon_vma. */ | |
360 | anon_vma = anon_vma_alloc(); | |
361 | if (!anon_vma) | |
362 | goto out_error; | |
363 | anon_vma->num_active_vmas++; | |
364 | avc = anon_vma_chain_alloc(GFP_KERNEL); | |
365 | if (!avc) | |
366 | goto out_error_free_anon_vma; | |
367 | ||
368 | /* | |
369 | * The root anon_vma's rwsem is the lock actually used when we | |
370 | * lock any of the anon_vmas in this anon_vma tree. | |
371 | */ | |
372 | anon_vma->root = pvma->anon_vma->root; | |
373 | anon_vma->parent = pvma->anon_vma; | |
374 | /* | |
375 | * With refcounts, an anon_vma can stay around longer than the | |
376 | * process it belongs to. The root anon_vma needs to be pinned until | |
377 | * this anon_vma is freed, because the lock lives in the root. | |
378 | */ | |
379 | get_anon_vma(anon_vma->root); | |
380 | /* Mark this anon_vma as the one where our new (COWed) pages go. */ | |
381 | vma->anon_vma = anon_vma; | |
382 | anon_vma_lock_write(anon_vma); | |
383 | anon_vma_chain_link(vma, avc, anon_vma); | |
384 | anon_vma->parent->num_children++; | |
385 | anon_vma_unlock_write(anon_vma); | |
386 | ||
387 | return 0; | |
388 | ||
389 | out_error_free_anon_vma: | |
390 | put_anon_vma(anon_vma); | |
391 | out_error: | |
392 | unlink_anon_vmas(vma); | |
393 | return -ENOMEM; | |
394 | } | |
395 | ||
396 | void unlink_anon_vmas(struct vm_area_struct *vma) | |
397 | { | |
398 | struct anon_vma_chain *avc, *next; | |
399 | struct anon_vma *root = NULL; | |
400 | ||
401 | /* | |
402 | * Unlink each anon_vma chained to the VMA. This list is ordered | |
403 | * from newest to oldest, ensuring the root anon_vma gets freed last. | |
404 | */ | |
405 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { | |
406 | struct anon_vma *anon_vma = avc->anon_vma; | |
407 | ||
408 | root = lock_anon_vma_root(root, anon_vma); | |
409 | anon_vma_interval_tree_remove(avc, &anon_vma->rb_root); | |
410 | ||
411 | /* | |
412 | * Leave empty anon_vmas on the list - we'll need | |
413 | * to free them outside the lock. | |
414 | */ | |
415 | if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) { | |
416 | anon_vma->parent->num_children--; | |
417 | continue; | |
418 | } | |
419 | ||
420 | list_del(&avc->same_vma); | |
421 | anon_vma_chain_free(avc); | |
422 | } | |
423 | if (vma->anon_vma) { | |
424 | vma->anon_vma->num_active_vmas--; | |
425 | ||
426 | /* | |
427 | * vma would still be needed after unlink, and anon_vma will be prepared | |
428 | * when handle fault. | |
429 | */ | |
430 | vma->anon_vma = NULL; | |
431 | } | |
432 | unlock_anon_vma_root(root); | |
433 | ||
434 | /* | |
435 | * Iterate the list once more, it now only contains empty and unlinked | |
436 | * anon_vmas, destroy them. Could not do before due to __put_anon_vma() | |
437 | * needing to write-acquire the anon_vma->root->rwsem. | |
438 | */ | |
439 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { | |
440 | struct anon_vma *anon_vma = avc->anon_vma; | |
441 | ||
442 | VM_WARN_ON(anon_vma->num_children); | |
443 | VM_WARN_ON(anon_vma->num_active_vmas); | |
444 | put_anon_vma(anon_vma); | |
445 | ||
446 | list_del(&avc->same_vma); | |
447 | anon_vma_chain_free(avc); | |
448 | } | |
449 | } | |
450 | ||
451 | static void anon_vma_ctor(void *data) | |
452 | { | |
453 | struct anon_vma *anon_vma = data; | |
454 | ||
455 | init_rwsem(&anon_vma->rwsem); | |
456 | atomic_set(&anon_vma->refcount, 0); | |
457 | anon_vma->rb_root = RB_ROOT_CACHED; | |
458 | } | |
459 | ||
460 | void __init anon_vma_init(void) | |
461 | { | |
462 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), | |
463 | 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT, | |
464 | anon_vma_ctor); | |
465 | anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, | |
466 | SLAB_PANIC|SLAB_ACCOUNT); | |
467 | } | |
468 | ||
469 | /* | |
470 | * Getting a lock on a stable anon_vma from a page off the LRU is tricky! | |
471 | * | |
472 | * Since there is no serialization what so ever against folio_remove_rmap_*() | |
473 | * the best this function can do is return a refcount increased anon_vma | |
474 | * that might have been relevant to this page. | |
475 | * | |
476 | * The page might have been remapped to a different anon_vma or the anon_vma | |
477 | * returned may already be freed (and even reused). | |
478 | * | |
479 | * In case it was remapped to a different anon_vma, the new anon_vma will be a | |
480 | * child of the old anon_vma, and the anon_vma lifetime rules will therefore | |
481 | * ensure that any anon_vma obtained from the page will still be valid for as | |
482 | * long as we observe page_mapped() [ hence all those page_mapped() tests ]. | |
483 | * | |
484 | * All users of this function must be very careful when walking the anon_vma | |
485 | * chain and verify that the page in question is indeed mapped in it | |
486 | * [ something equivalent to page_mapped_in_vma() ]. | |
487 | * | |
488 | * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from | |
489 | * folio_remove_rmap_*() that the anon_vma pointer from page->mapping is valid | |
490 | * if there is a mapcount, we can dereference the anon_vma after observing | |
491 | * those. | |
492 | * | |
493 | * NOTE: the caller should normally hold folio lock when calling this. If | |
494 | * not, the caller needs to double check the anon_vma didn't change after | |
495 | * taking the anon_vma lock for either read or write (UFFDIO_MOVE can modify it | |
496 | * concurrently without folio lock protection). See folio_lock_anon_vma_read() | |
497 | * which has already covered that, and comment above remap_pages(). | |
498 | */ | |
499 | struct anon_vma *folio_get_anon_vma(const struct folio *folio) | |
500 | { | |
501 | struct anon_vma *anon_vma = NULL; | |
502 | unsigned long anon_mapping; | |
503 | ||
504 | rcu_read_lock(); | |
505 | anon_mapping = (unsigned long)READ_ONCE(folio->mapping); | |
506 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) | |
507 | goto out; | |
508 | if (!folio_mapped(folio)) | |
509 | goto out; | |
510 | ||
511 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
512 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { | |
513 | anon_vma = NULL; | |
514 | goto out; | |
515 | } | |
516 | ||
517 | /* | |
518 | * If this folio is still mapped, then its anon_vma cannot have been | |
519 | * freed. But if it has been unmapped, we have no security against the | |
520 | * anon_vma structure being freed and reused (for another anon_vma: | |
521 | * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero() | |
522 | * above cannot corrupt). | |
523 | */ | |
524 | if (!folio_mapped(folio)) { | |
525 | rcu_read_unlock(); | |
526 | put_anon_vma(anon_vma); | |
527 | return NULL; | |
528 | } | |
529 | out: | |
530 | rcu_read_unlock(); | |
531 | ||
532 | return anon_vma; | |
533 | } | |
534 | ||
535 | /* | |
536 | * Similar to folio_get_anon_vma() except it locks the anon_vma. | |
537 | * | |
538 | * Its a little more complex as it tries to keep the fast path to a single | |
539 | * atomic op -- the trylock. If we fail the trylock, we fall back to getting a | |
540 | * reference like with folio_get_anon_vma() and then block on the mutex | |
541 | * on !rwc->try_lock case. | |
542 | */ | |
543 | struct anon_vma *folio_lock_anon_vma_read(const struct folio *folio, | |
544 | struct rmap_walk_control *rwc) | |
545 | { | |
546 | struct anon_vma *anon_vma = NULL; | |
547 | struct anon_vma *root_anon_vma; | |
548 | unsigned long anon_mapping; | |
549 | ||
550 | retry: | |
551 | rcu_read_lock(); | |
552 | anon_mapping = (unsigned long)READ_ONCE(folio->mapping); | |
553 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) | |
554 | goto out; | |
555 | if (!folio_mapped(folio)) | |
556 | goto out; | |
557 | ||
558 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); | |
559 | root_anon_vma = READ_ONCE(anon_vma->root); | |
560 | if (down_read_trylock(&root_anon_vma->rwsem)) { | |
561 | /* | |
562 | * folio_move_anon_rmap() might have changed the anon_vma as we | |
563 | * might not hold the folio lock here. | |
564 | */ | |
565 | if (unlikely((unsigned long)READ_ONCE(folio->mapping) != | |
566 | anon_mapping)) { | |
567 | up_read(&root_anon_vma->rwsem); | |
568 | rcu_read_unlock(); | |
569 | goto retry; | |
570 | } | |
571 | ||
572 | /* | |
573 | * If the folio is still mapped, then this anon_vma is still | |
574 | * its anon_vma, and holding the mutex ensures that it will | |
575 | * not go away, see anon_vma_free(). | |
576 | */ | |
577 | if (!folio_mapped(folio)) { | |
578 | up_read(&root_anon_vma->rwsem); | |
579 | anon_vma = NULL; | |
580 | } | |
581 | goto out; | |
582 | } | |
583 | ||
584 | if (rwc && rwc->try_lock) { | |
585 | anon_vma = NULL; | |
586 | rwc->contended = true; | |
587 | goto out; | |
588 | } | |
589 | ||
590 | /* trylock failed, we got to sleep */ | |
591 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { | |
592 | anon_vma = NULL; | |
593 | goto out; | |
594 | } | |
595 | ||
596 | if (!folio_mapped(folio)) { | |
597 | rcu_read_unlock(); | |
598 | put_anon_vma(anon_vma); | |
599 | return NULL; | |
600 | } | |
601 | ||
602 | /* we pinned the anon_vma, its safe to sleep */ | |
603 | rcu_read_unlock(); | |
604 | anon_vma_lock_read(anon_vma); | |
605 | ||
606 | /* | |
607 | * folio_move_anon_rmap() might have changed the anon_vma as we might | |
608 | * not hold the folio lock here. | |
609 | */ | |
610 | if (unlikely((unsigned long)READ_ONCE(folio->mapping) != | |
611 | anon_mapping)) { | |
612 | anon_vma_unlock_read(anon_vma); | |
613 | put_anon_vma(anon_vma); | |
614 | anon_vma = NULL; | |
615 | goto retry; | |
616 | } | |
617 | ||
618 | if (atomic_dec_and_test(&anon_vma->refcount)) { | |
619 | /* | |
620 | * Oops, we held the last refcount, release the lock | |
621 | * and bail -- can't simply use put_anon_vma() because | |
622 | * we'll deadlock on the anon_vma_lock_write() recursion. | |
623 | */ | |
624 | anon_vma_unlock_read(anon_vma); | |
625 | __put_anon_vma(anon_vma); | |
626 | anon_vma = NULL; | |
627 | } | |
628 | ||
629 | return anon_vma; | |
630 | ||
631 | out: | |
632 | rcu_read_unlock(); | |
633 | return anon_vma; | |
634 | } | |
635 | ||
636 | #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH | |
637 | /* | |
638 | * Flush TLB entries for recently unmapped pages from remote CPUs. It is | |
639 | * important if a PTE was dirty when it was unmapped that it's flushed | |
640 | * before any IO is initiated on the page to prevent lost writes. Similarly, | |
641 | * it must be flushed before freeing to prevent data leakage. | |
642 | */ | |
643 | void try_to_unmap_flush(void) | |
644 | { | |
645 | struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; | |
646 | ||
647 | if (!tlb_ubc->flush_required) | |
648 | return; | |
649 | ||
650 | arch_tlbbatch_flush(&tlb_ubc->arch); | |
651 | tlb_ubc->flush_required = false; | |
652 | tlb_ubc->writable = false; | |
653 | } | |
654 | ||
655 | /* Flush iff there are potentially writable TLB entries that can race with IO */ | |
656 | void try_to_unmap_flush_dirty(void) | |
657 | { | |
658 | struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; | |
659 | ||
660 | if (tlb_ubc->writable) | |
661 | try_to_unmap_flush(); | |
662 | } | |
663 | ||
664 | /* | |
665 | * Bits 0-14 of mm->tlb_flush_batched record pending generations. | |
666 | * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations. | |
667 | */ | |
668 | #define TLB_FLUSH_BATCH_FLUSHED_SHIFT 16 | |
669 | #define TLB_FLUSH_BATCH_PENDING_MASK \ | |
670 | ((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1) | |
671 | #define TLB_FLUSH_BATCH_PENDING_LARGE \ | |
672 | (TLB_FLUSH_BATCH_PENDING_MASK / 2) | |
673 | ||
674 | static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval, | |
675 | unsigned long start, unsigned long end) | |
676 | { | |
677 | struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; | |
678 | int batch; | |
679 | bool writable = pte_dirty(pteval); | |
680 | ||
681 | if (!pte_accessible(mm, pteval)) | |
682 | return; | |
683 | ||
684 | arch_tlbbatch_add_pending(&tlb_ubc->arch, mm, start, end); | |
685 | tlb_ubc->flush_required = true; | |
686 | ||
687 | /* | |
688 | * Ensure compiler does not re-order the setting of tlb_flush_batched | |
689 | * before the PTE is cleared. | |
690 | */ | |
691 | barrier(); | |
692 | batch = atomic_read(&mm->tlb_flush_batched); | |
693 | retry: | |
694 | if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) { | |
695 | /* | |
696 | * Prevent `pending' from catching up with `flushed' because of | |
697 | * overflow. Reset `pending' and `flushed' to be 1 and 0 if | |
698 | * `pending' becomes large. | |
699 | */ | |
700 | if (!atomic_try_cmpxchg(&mm->tlb_flush_batched, &batch, 1)) | |
701 | goto retry; | |
702 | } else { | |
703 | atomic_inc(&mm->tlb_flush_batched); | |
704 | } | |
705 | ||
706 | /* | |
707 | * If the PTE was dirty then it's best to assume it's writable. The | |
708 | * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush() | |
709 | * before the page is queued for IO. | |
710 | */ | |
711 | if (writable) | |
712 | tlb_ubc->writable = true; | |
713 | } | |
714 | ||
715 | /* | |
716 | * Returns true if the TLB flush should be deferred to the end of a batch of | |
717 | * unmap operations to reduce IPIs. | |
718 | */ | |
719 | static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) | |
720 | { | |
721 | if (!(flags & TTU_BATCH_FLUSH)) | |
722 | return false; | |
723 | ||
724 | return arch_tlbbatch_should_defer(mm); | |
725 | } | |
726 | ||
727 | /* | |
728 | * Reclaim unmaps pages under the PTL but do not flush the TLB prior to | |
729 | * releasing the PTL if TLB flushes are batched. It's possible for a parallel | |
730 | * operation such as mprotect or munmap to race between reclaim unmapping | |
731 | * the page and flushing the page. If this race occurs, it potentially allows | |
732 | * access to data via a stale TLB entry. Tracking all mm's that have TLB | |
733 | * batching in flight would be expensive during reclaim so instead track | |
734 | * whether TLB batching occurred in the past and if so then do a flush here | |
735 | * if required. This will cost one additional flush per reclaim cycle paid | |
736 | * by the first operation at risk such as mprotect and mumap. | |
737 | * | |
738 | * This must be called under the PTL so that an access to tlb_flush_batched | |
739 | * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise | |
740 | * via the PTL. | |
741 | */ | |
742 | void flush_tlb_batched_pending(struct mm_struct *mm) | |
743 | { | |
744 | int batch = atomic_read(&mm->tlb_flush_batched); | |
745 | int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK; | |
746 | int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT; | |
747 | ||
748 | if (pending != flushed) { | |
749 | arch_flush_tlb_batched_pending(mm); | |
750 | /* | |
751 | * If the new TLB flushing is pending during flushing, leave | |
752 | * mm->tlb_flush_batched as is, to avoid losing flushing. | |
753 | */ | |
754 | atomic_cmpxchg(&mm->tlb_flush_batched, batch, | |
755 | pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT)); | |
756 | } | |
757 | } | |
758 | #else | |
759 | static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval, | |
760 | unsigned long start, unsigned long end) | |
761 | { | |
762 | } | |
763 | ||
764 | static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) | |
765 | { | |
766 | return false; | |
767 | } | |
768 | #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */ | |
769 | ||
770 | /** | |
771 | * page_address_in_vma - The virtual address of a page in this VMA. | |
772 | * @folio: The folio containing the page. | |
773 | * @page: The page within the folio. | |
774 | * @vma: The VMA we need to know the address in. | |
775 | * | |
776 | * Calculates the user virtual address of this page in the specified VMA. | |
777 | * It is the caller's responsibility to check the page is actually | |
778 | * within the VMA. There may not currently be a PTE pointing at this | |
779 | * page, but if a page fault occurs at this address, this is the page | |
780 | * which will be accessed. | |
781 | * | |
782 | * Context: Caller should hold a reference to the folio. Caller should | |
783 | * hold a lock (eg the i_mmap_lock or the mmap_lock) which keeps the | |
784 | * VMA from being altered. | |
785 | * | |
786 | * Return: The virtual address corresponding to this page in the VMA. | |
787 | */ | |
788 | unsigned long page_address_in_vma(const struct folio *folio, | |
789 | const struct page *page, const struct vm_area_struct *vma) | |
790 | { | |
791 | if (folio_test_anon(folio)) { | |
792 | struct anon_vma *anon_vma = folio_anon_vma(folio); | |
793 | /* | |
794 | * Note: swapoff's unuse_vma() is more efficient with this | |
795 | * check, and needs it to match anon_vma when KSM is active. | |
796 | */ | |
797 | if (!vma->anon_vma || !anon_vma || | |
798 | vma->anon_vma->root != anon_vma->root) | |
799 | return -EFAULT; | |
800 | } else if (!vma->vm_file) { | |
801 | return -EFAULT; | |
802 | } else if (vma->vm_file->f_mapping != folio->mapping) { | |
803 | return -EFAULT; | |
804 | } | |
805 | ||
806 | /* KSM folios don't reach here because of the !anon_vma check */ | |
807 | return vma_address(vma, page_pgoff(folio, page), 1); | |
808 | } | |
809 | ||
810 | /* | |
811 | * Returns the actual pmd_t* where we expect 'address' to be mapped from, or | |
812 | * NULL if it doesn't exist. No guarantees / checks on what the pmd_t* | |
813 | * represents. | |
814 | */ | |
815 | pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address) | |
816 | { | |
817 | pgd_t *pgd; | |
818 | p4d_t *p4d; | |
819 | pud_t *pud; | |
820 | pmd_t *pmd = NULL; | |
821 | ||
822 | pgd = pgd_offset(mm, address); | |
823 | if (!pgd_present(*pgd)) | |
824 | goto out; | |
825 | ||
826 | p4d = p4d_offset(pgd, address); | |
827 | if (!p4d_present(*p4d)) | |
828 | goto out; | |
829 | ||
830 | pud = pud_offset(p4d, address); | |
831 | if (!pud_present(*pud)) | |
832 | goto out; | |
833 | ||
834 | pmd = pmd_offset(pud, address); | |
835 | out: | |
836 | return pmd; | |
837 | } | |
838 | ||
839 | struct folio_referenced_arg { | |
840 | int mapcount; | |
841 | int referenced; | |
842 | unsigned long vm_flags; | |
843 | struct mem_cgroup *memcg; | |
844 | }; | |
845 | ||
846 | /* | |
847 | * arg: folio_referenced_arg will be passed | |
848 | */ | |
849 | static bool folio_referenced_one(struct folio *folio, | |
850 | struct vm_area_struct *vma, unsigned long address, void *arg) | |
851 | { | |
852 | struct folio_referenced_arg *pra = arg; | |
853 | DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0); | |
854 | int referenced = 0; | |
855 | unsigned long start = address, ptes = 0; | |
856 | ||
857 | while (page_vma_mapped_walk(&pvmw)) { | |
858 | address = pvmw.address; | |
859 | ||
860 | if (vma->vm_flags & VM_LOCKED) { | |
861 | if (!folio_test_large(folio) || !pvmw.pte) { | |
862 | /* Restore the mlock which got missed */ | |
863 | mlock_vma_folio(folio, vma); | |
864 | page_vma_mapped_walk_done(&pvmw); | |
865 | pra->vm_flags |= VM_LOCKED; | |
866 | return false; /* To break the loop */ | |
867 | } | |
868 | /* | |
869 | * For large folio fully mapped to VMA, will | |
870 | * be handled after the pvmw loop. | |
871 | * | |
872 | * For large folio cross VMA boundaries, it's | |
873 | * expected to be picked by page reclaim. But | |
874 | * should skip reference of pages which are in | |
875 | * the range of VM_LOCKED vma. As page reclaim | |
876 | * should just count the reference of pages out | |
877 | * the range of VM_LOCKED vma. | |
878 | */ | |
879 | ptes++; | |
880 | pra->mapcount--; | |
881 | continue; | |
882 | } | |
883 | ||
884 | /* | |
885 | * Skip the non-shared swapbacked folio mapped solely by | |
886 | * the exiting or OOM-reaped process. This avoids redundant | |
887 | * swap-out followed by an immediate unmap. | |
888 | */ | |
889 | if ((!atomic_read(&vma->vm_mm->mm_users) || | |
890 | check_stable_address_space(vma->vm_mm)) && | |
891 | folio_test_anon(folio) && folio_test_swapbacked(folio) && | |
892 | !folio_maybe_mapped_shared(folio)) { | |
893 | pra->referenced = -1; | |
894 | page_vma_mapped_walk_done(&pvmw); | |
895 | return false; | |
896 | } | |
897 | ||
898 | if (lru_gen_enabled() && pvmw.pte) { | |
899 | if (lru_gen_look_around(&pvmw)) | |
900 | referenced++; | |
901 | } else if (pvmw.pte) { | |
902 | if (ptep_clear_flush_young_notify(vma, address, | |
903 | pvmw.pte)) | |
904 | referenced++; | |
905 | } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) { | |
906 | if (pmdp_clear_flush_young_notify(vma, address, | |
907 | pvmw.pmd)) | |
908 | referenced++; | |
909 | } else { | |
910 | /* unexpected pmd-mapped folio? */ | |
911 | WARN_ON_ONCE(1); | |
912 | } | |
913 | ||
914 | pra->mapcount--; | |
915 | } | |
916 | ||
917 | if ((vma->vm_flags & VM_LOCKED) && | |
918 | folio_test_large(folio) && | |
919 | folio_within_vma(folio, vma)) { | |
920 | unsigned long s_align, e_align; | |
921 | ||
922 | s_align = ALIGN_DOWN(start, PMD_SIZE); | |
923 | e_align = ALIGN_DOWN(start + folio_size(folio) - 1, PMD_SIZE); | |
924 | ||
925 | /* folio doesn't cross page table boundary and fully mapped */ | |
926 | if ((s_align == e_align) && (ptes == folio_nr_pages(folio))) { | |
927 | /* Restore the mlock which got missed */ | |
928 | mlock_vma_folio(folio, vma); | |
929 | pra->vm_flags |= VM_LOCKED; | |
930 | return false; /* To break the loop */ | |
931 | } | |
932 | } | |
933 | ||
934 | if (referenced) | |
935 | folio_clear_idle(folio); | |
936 | if (folio_test_clear_young(folio)) | |
937 | referenced++; | |
938 | ||
939 | if (referenced) { | |
940 | pra->referenced++; | |
941 | pra->vm_flags |= vma->vm_flags & ~VM_LOCKED; | |
942 | } | |
943 | ||
944 | if (!pra->mapcount) | |
945 | return false; /* To break the loop */ | |
946 | ||
947 | return true; | |
948 | } | |
949 | ||
950 | static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg) | |
951 | { | |
952 | struct folio_referenced_arg *pra = arg; | |
953 | struct mem_cgroup *memcg = pra->memcg; | |
954 | ||
955 | /* | |
956 | * Ignore references from this mapping if it has no recency. If the | |
957 | * folio has been used in another mapping, we will catch it; if this | |
958 | * other mapping is already gone, the unmap path will have set the | |
959 | * referenced flag or activated the folio in zap_pte_range(). | |
960 | */ | |
961 | if (!vma_has_recency(vma)) | |
962 | return true; | |
963 | ||
964 | /* | |
965 | * If we are reclaiming on behalf of a cgroup, skip counting on behalf | |
966 | * of references from different cgroups. | |
967 | */ | |
968 | if (memcg && !mm_match_cgroup(vma->vm_mm, memcg)) | |
969 | return true; | |
970 | ||
971 | return false; | |
972 | } | |
973 | ||
974 | /** | |
975 | * folio_referenced() - Test if the folio was referenced. | |
976 | * @folio: The folio to test. | |
977 | * @is_locked: Caller holds lock on the folio. | |
978 | * @memcg: target memory cgroup | |
979 | * @vm_flags: A combination of all the vma->vm_flags which referenced the folio. | |
980 | * | |
981 | * Quick test_and_clear_referenced for all mappings of a folio, | |
982 | * | |
983 | * Return: The number of mappings which referenced the folio. Return -1 if | |
984 | * the function bailed out due to rmap lock contention. | |
985 | */ | |
986 | int folio_referenced(struct folio *folio, int is_locked, | |
987 | struct mem_cgroup *memcg, unsigned long *vm_flags) | |
988 | { | |
989 | bool we_locked = false; | |
990 | struct folio_referenced_arg pra = { | |
991 | .mapcount = folio_mapcount(folio), | |
992 | .memcg = memcg, | |
993 | }; | |
994 | struct rmap_walk_control rwc = { | |
995 | .rmap_one = folio_referenced_one, | |
996 | .arg = (void *)&pra, | |
997 | .anon_lock = folio_lock_anon_vma_read, | |
998 | .try_lock = true, | |
999 | .invalid_vma = invalid_folio_referenced_vma, | |
1000 | }; | |
1001 | ||
1002 | *vm_flags = 0; | |
1003 | if (!pra.mapcount) | |
1004 | return 0; | |
1005 | ||
1006 | if (!folio_raw_mapping(folio)) | |
1007 | return 0; | |
1008 | ||
1009 | if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) { | |
1010 | we_locked = folio_trylock(folio); | |
1011 | if (!we_locked) | |
1012 | return 1; | |
1013 | } | |
1014 | ||
1015 | rmap_walk(folio, &rwc); | |
1016 | *vm_flags = pra.vm_flags; | |
1017 | ||
1018 | if (we_locked) | |
1019 | folio_unlock(folio); | |
1020 | ||
1021 | return rwc.contended ? -1 : pra.referenced; | |
1022 | } | |
1023 | ||
1024 | static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw) | |
1025 | { | |
1026 | int cleaned = 0; | |
1027 | struct vm_area_struct *vma = pvmw->vma; | |
1028 | struct mmu_notifier_range range; | |
1029 | unsigned long address = pvmw->address; | |
1030 | ||
1031 | /* | |
1032 | * We have to assume the worse case ie pmd for invalidation. Note that | |
1033 | * the folio can not be freed from this function. | |
1034 | */ | |
1035 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 0, | |
1036 | vma->vm_mm, address, vma_address_end(pvmw)); | |
1037 | mmu_notifier_invalidate_range_start(&range); | |
1038 | ||
1039 | while (page_vma_mapped_walk(pvmw)) { | |
1040 | int ret = 0; | |
1041 | ||
1042 | address = pvmw->address; | |
1043 | if (pvmw->pte) { | |
1044 | pte_t *pte = pvmw->pte; | |
1045 | pte_t entry = ptep_get(pte); | |
1046 | ||
1047 | /* | |
1048 | * PFN swap PTEs, such as device-exclusive ones, that | |
1049 | * actually map pages are clean and not writable from a | |
1050 | * CPU perspective. The MMU notifier takes care of any | |
1051 | * device aspects. | |
1052 | */ | |
1053 | if (!pte_present(entry)) | |
1054 | continue; | |
1055 | if (!pte_dirty(entry) && !pte_write(entry)) | |
1056 | continue; | |
1057 | ||
1058 | flush_cache_page(vma, address, pte_pfn(entry)); | |
1059 | entry = ptep_clear_flush(vma, address, pte); | |
1060 | entry = pte_wrprotect(entry); | |
1061 | entry = pte_mkclean(entry); | |
1062 | set_pte_at(vma->vm_mm, address, pte, entry); | |
1063 | ret = 1; | |
1064 | } else { | |
1065 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
1066 | pmd_t *pmd = pvmw->pmd; | |
1067 | pmd_t entry; | |
1068 | ||
1069 | if (!pmd_dirty(*pmd) && !pmd_write(*pmd)) | |
1070 | continue; | |
1071 | ||
1072 | flush_cache_range(vma, address, | |
1073 | address + HPAGE_PMD_SIZE); | |
1074 | entry = pmdp_invalidate(vma, address, pmd); | |
1075 | entry = pmd_wrprotect(entry); | |
1076 | entry = pmd_mkclean(entry); | |
1077 | set_pmd_at(vma->vm_mm, address, pmd, entry); | |
1078 | ret = 1; | |
1079 | #else | |
1080 | /* unexpected pmd-mapped folio? */ | |
1081 | WARN_ON_ONCE(1); | |
1082 | #endif | |
1083 | } | |
1084 | ||
1085 | if (ret) | |
1086 | cleaned++; | |
1087 | } | |
1088 | ||
1089 | mmu_notifier_invalidate_range_end(&range); | |
1090 | ||
1091 | return cleaned; | |
1092 | } | |
1093 | ||
1094 | static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma, | |
1095 | unsigned long address, void *arg) | |
1096 | { | |
1097 | DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC); | |
1098 | int *cleaned = arg; | |
1099 | ||
1100 | *cleaned += page_vma_mkclean_one(&pvmw); | |
1101 | ||
1102 | return true; | |
1103 | } | |
1104 | ||
1105 | static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg) | |
1106 | { | |
1107 | if (vma->vm_flags & VM_SHARED) | |
1108 | return false; | |
1109 | ||
1110 | return true; | |
1111 | } | |
1112 | ||
1113 | int folio_mkclean(struct folio *folio) | |
1114 | { | |
1115 | int cleaned = 0; | |
1116 | struct address_space *mapping; | |
1117 | struct rmap_walk_control rwc = { | |
1118 | .arg = (void *)&cleaned, | |
1119 | .rmap_one = page_mkclean_one, | |
1120 | .invalid_vma = invalid_mkclean_vma, | |
1121 | }; | |
1122 | ||
1123 | BUG_ON(!folio_test_locked(folio)); | |
1124 | ||
1125 | if (!folio_mapped(folio)) | |
1126 | return 0; | |
1127 | ||
1128 | mapping = folio_mapping(folio); | |
1129 | if (!mapping) | |
1130 | return 0; | |
1131 | ||
1132 | rmap_walk(folio, &rwc); | |
1133 | ||
1134 | return cleaned; | |
1135 | } | |
1136 | EXPORT_SYMBOL_GPL(folio_mkclean); | |
1137 | ||
1138 | struct wrprotect_file_state { | |
1139 | int cleaned; | |
1140 | pgoff_t pgoff; | |
1141 | unsigned long pfn; | |
1142 | unsigned long nr_pages; | |
1143 | }; | |
1144 | ||
1145 | static bool mapping_wrprotect_range_one(struct folio *folio, | |
1146 | struct vm_area_struct *vma, unsigned long address, void *arg) | |
1147 | { | |
1148 | struct wrprotect_file_state *state = (struct wrprotect_file_state *)arg; | |
1149 | struct page_vma_mapped_walk pvmw = { | |
1150 | .pfn = state->pfn, | |
1151 | .nr_pages = state->nr_pages, | |
1152 | .pgoff = state->pgoff, | |
1153 | .vma = vma, | |
1154 | .address = address, | |
1155 | .flags = PVMW_SYNC, | |
1156 | }; | |
1157 | ||
1158 | state->cleaned += page_vma_mkclean_one(&pvmw); | |
1159 | ||
1160 | return true; | |
1161 | } | |
1162 | ||
1163 | static void __rmap_walk_file(struct folio *folio, struct address_space *mapping, | |
1164 | pgoff_t pgoff_start, unsigned long nr_pages, | |
1165 | struct rmap_walk_control *rwc, bool locked); | |
1166 | ||
1167 | /** | |
1168 | * mapping_wrprotect_range() - Write-protect all mappings in a specified range. | |
1169 | * | |
1170 | * @mapping: The mapping whose reverse mapping should be traversed. | |
1171 | * @pgoff: The page offset at which @pfn is mapped within @mapping. | |
1172 | * @pfn: The PFN of the page mapped in @mapping at @pgoff. | |
1173 | * @nr_pages: The number of physically contiguous base pages spanned. | |
1174 | * | |
1175 | * Traverses the reverse mapping, finding all VMAs which contain a shared | |
1176 | * mapping of the pages in the specified range in @mapping, and write-protects | |
1177 | * them (that is, updates the page tables to mark the mappings read-only such | |
1178 | * that a write protection fault arises when the mappings are written to). | |
1179 | * | |
1180 | * The @pfn value need not refer to a folio, but rather can reference a kernel | |
1181 | * allocation which is mapped into userland. We therefore do not require that | |
1182 | * the page maps to a folio with a valid mapping or index field, rather the | |
1183 | * caller specifies these in @mapping and @pgoff. | |
1184 | * | |
1185 | * Return: the number of write-protected PTEs, or an error. | |
1186 | */ | |
1187 | int mapping_wrprotect_range(struct address_space *mapping, pgoff_t pgoff, | |
1188 | unsigned long pfn, unsigned long nr_pages) | |
1189 | { | |
1190 | struct wrprotect_file_state state = { | |
1191 | .cleaned = 0, | |
1192 | .pgoff = pgoff, | |
1193 | .pfn = pfn, | |
1194 | .nr_pages = nr_pages, | |
1195 | }; | |
1196 | struct rmap_walk_control rwc = { | |
1197 | .arg = (void *)&state, | |
1198 | .rmap_one = mapping_wrprotect_range_one, | |
1199 | .invalid_vma = invalid_mkclean_vma, | |
1200 | }; | |
1201 | ||
1202 | if (!mapping) | |
1203 | return 0; | |
1204 | ||
1205 | __rmap_walk_file(/* folio = */NULL, mapping, pgoff, nr_pages, &rwc, | |
1206 | /* locked = */false); | |
1207 | ||
1208 | return state.cleaned; | |
1209 | } | |
1210 | EXPORT_SYMBOL_GPL(mapping_wrprotect_range); | |
1211 | ||
1212 | /** | |
1213 | * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of | |
1214 | * [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff) | |
1215 | * within the @vma of shared mappings. And since clean PTEs | |
1216 | * should also be readonly, write protects them too. | |
1217 | * @pfn: start pfn. | |
1218 | * @nr_pages: number of physically contiguous pages srarting with @pfn. | |
1219 | * @pgoff: page offset that the @pfn mapped with. | |
1220 | * @vma: vma that @pfn mapped within. | |
1221 | * | |
1222 | * Returns the number of cleaned PTEs (including PMDs). | |
1223 | */ | |
1224 | int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff, | |
1225 | struct vm_area_struct *vma) | |
1226 | { | |
1227 | struct page_vma_mapped_walk pvmw = { | |
1228 | .pfn = pfn, | |
1229 | .nr_pages = nr_pages, | |
1230 | .pgoff = pgoff, | |
1231 | .vma = vma, | |
1232 | .flags = PVMW_SYNC, | |
1233 | }; | |
1234 | ||
1235 | if (invalid_mkclean_vma(vma, NULL)) | |
1236 | return 0; | |
1237 | ||
1238 | pvmw.address = vma_address(vma, pgoff, nr_pages); | |
1239 | VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma); | |
1240 | ||
1241 | return page_vma_mkclean_one(&pvmw); | |
1242 | } | |
1243 | ||
1244 | static __always_inline unsigned int __folio_add_rmap(struct folio *folio, | |
1245 | struct page *page, int nr_pages, struct vm_area_struct *vma, | |
1246 | enum rmap_level level, int *nr_pmdmapped) | |
1247 | { | |
1248 | atomic_t *mapped = &folio->_nr_pages_mapped; | |
1249 | const int orig_nr_pages = nr_pages; | |
1250 | int first = 0, nr = 0; | |
1251 | ||
1252 | __folio_rmap_sanity_checks(folio, page, nr_pages, level); | |
1253 | ||
1254 | switch (level) { | |
1255 | case RMAP_LEVEL_PTE: | |
1256 | if (!folio_test_large(folio)) { | |
1257 | nr = atomic_inc_and_test(&folio->_mapcount); | |
1258 | break; | |
1259 | } | |
1260 | ||
1261 | if (IS_ENABLED(CONFIG_NO_PAGE_MAPCOUNT)) { | |
1262 | nr = folio_add_return_large_mapcount(folio, orig_nr_pages, vma); | |
1263 | if (nr == orig_nr_pages) | |
1264 | /* Was completely unmapped. */ | |
1265 | nr = folio_large_nr_pages(folio); | |
1266 | else | |
1267 | nr = 0; | |
1268 | break; | |
1269 | } | |
1270 | ||
1271 | do { | |
1272 | first += atomic_inc_and_test(&page->_mapcount); | |
1273 | } while (page++, --nr_pages > 0); | |
1274 | ||
1275 | if (first && | |
1276 | atomic_add_return_relaxed(first, mapped) < ENTIRELY_MAPPED) | |
1277 | nr = first; | |
1278 | ||
1279 | folio_add_large_mapcount(folio, orig_nr_pages, vma); | |
1280 | break; | |
1281 | case RMAP_LEVEL_PMD: | |
1282 | case RMAP_LEVEL_PUD: | |
1283 | first = atomic_inc_and_test(&folio->_entire_mapcount); | |
1284 | if (IS_ENABLED(CONFIG_NO_PAGE_MAPCOUNT)) { | |
1285 | if (level == RMAP_LEVEL_PMD && first) | |
1286 | *nr_pmdmapped = folio_large_nr_pages(folio); | |
1287 | nr = folio_inc_return_large_mapcount(folio, vma); | |
1288 | if (nr == 1) | |
1289 | /* Was completely unmapped. */ | |
1290 | nr = folio_large_nr_pages(folio); | |
1291 | else | |
1292 | nr = 0; | |
1293 | break; | |
1294 | } | |
1295 | ||
1296 | if (first) { | |
1297 | nr = atomic_add_return_relaxed(ENTIRELY_MAPPED, mapped); | |
1298 | if (likely(nr < ENTIRELY_MAPPED + ENTIRELY_MAPPED)) { | |
1299 | nr_pages = folio_large_nr_pages(folio); | |
1300 | /* | |
1301 | * We only track PMD mappings of PMD-sized | |
1302 | * folios separately. | |
1303 | */ | |
1304 | if (level == RMAP_LEVEL_PMD) | |
1305 | *nr_pmdmapped = nr_pages; | |
1306 | nr = nr_pages - (nr & FOLIO_PAGES_MAPPED); | |
1307 | /* Raced ahead of a remove and another add? */ | |
1308 | if (unlikely(nr < 0)) | |
1309 | nr = 0; | |
1310 | } else { | |
1311 | /* Raced ahead of a remove of ENTIRELY_MAPPED */ | |
1312 | nr = 0; | |
1313 | } | |
1314 | } | |
1315 | folio_inc_large_mapcount(folio, vma); | |
1316 | break; | |
1317 | } | |
1318 | return nr; | |
1319 | } | |
1320 | ||
1321 | /** | |
1322 | * folio_move_anon_rmap - move a folio to our anon_vma | |
1323 | * @folio: The folio to move to our anon_vma | |
1324 | * @vma: The vma the folio belongs to | |
1325 | * | |
1326 | * When a folio belongs exclusively to one process after a COW event, | |
1327 | * that folio can be moved into the anon_vma that belongs to just that | |
1328 | * process, so the rmap code will not search the parent or sibling processes. | |
1329 | */ | |
1330 | void folio_move_anon_rmap(struct folio *folio, struct vm_area_struct *vma) | |
1331 | { | |
1332 | void *anon_vma = vma->anon_vma; | |
1333 | ||
1334 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); | |
1335 | VM_BUG_ON_VMA(!anon_vma, vma); | |
1336 | ||
1337 | anon_vma += PAGE_MAPPING_ANON; | |
1338 | /* | |
1339 | * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written | |
1340 | * simultaneously, so a concurrent reader (eg folio_referenced()'s | |
1341 | * folio_test_anon()) will not see one without the other. | |
1342 | */ | |
1343 | WRITE_ONCE(folio->mapping, anon_vma); | |
1344 | } | |
1345 | ||
1346 | /** | |
1347 | * __folio_set_anon - set up a new anonymous rmap for a folio | |
1348 | * @folio: The folio to set up the new anonymous rmap for. | |
1349 | * @vma: VM area to add the folio to. | |
1350 | * @address: User virtual address of the mapping | |
1351 | * @exclusive: Whether the folio is exclusive to the process. | |
1352 | */ | |
1353 | static void __folio_set_anon(struct folio *folio, struct vm_area_struct *vma, | |
1354 | unsigned long address, bool exclusive) | |
1355 | { | |
1356 | struct anon_vma *anon_vma = vma->anon_vma; | |
1357 | ||
1358 | BUG_ON(!anon_vma); | |
1359 | ||
1360 | /* | |
1361 | * If the folio isn't exclusive to this vma, we must use the _oldest_ | |
1362 | * possible anon_vma for the folio mapping! | |
1363 | */ | |
1364 | if (!exclusive) | |
1365 | anon_vma = anon_vma->root; | |
1366 | ||
1367 | /* | |
1368 | * page_idle does a lockless/optimistic rmap scan on folio->mapping. | |
1369 | * Make sure the compiler doesn't split the stores of anon_vma and | |
1370 | * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code | |
1371 | * could mistake the mapping for a struct address_space and crash. | |
1372 | */ | |
1373 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; | |
1374 | WRITE_ONCE(folio->mapping, (struct address_space *) anon_vma); | |
1375 | folio->index = linear_page_index(vma, address); | |
1376 | } | |
1377 | ||
1378 | /** | |
1379 | * __page_check_anon_rmap - sanity check anonymous rmap addition | |
1380 | * @folio: The folio containing @page. | |
1381 | * @page: the page to check the mapping of | |
1382 | * @vma: the vm area in which the mapping is added | |
1383 | * @address: the user virtual address mapped | |
1384 | */ | |
1385 | static void __page_check_anon_rmap(const struct folio *folio, | |
1386 | const struct page *page, struct vm_area_struct *vma, | |
1387 | unsigned long address) | |
1388 | { | |
1389 | /* | |
1390 | * The page's anon-rmap details (mapping and index) are guaranteed to | |
1391 | * be set up correctly at this point. | |
1392 | * | |
1393 | * We have exclusion against folio_add_anon_rmap_*() because the caller | |
1394 | * always holds the page locked. | |
1395 | * | |
1396 | * We have exclusion against folio_add_new_anon_rmap because those pages | |
1397 | * are initially only visible via the pagetables, and the pte is locked | |
1398 | * over the call to folio_add_new_anon_rmap. | |
1399 | */ | |
1400 | VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root, | |
1401 | folio); | |
1402 | VM_BUG_ON_PAGE(page_pgoff(folio, page) != linear_page_index(vma, address), | |
1403 | page); | |
1404 | } | |
1405 | ||
1406 | static void __folio_mod_stat(struct folio *folio, int nr, int nr_pmdmapped) | |
1407 | { | |
1408 | int idx; | |
1409 | ||
1410 | if (nr) { | |
1411 | idx = folio_test_anon(folio) ? NR_ANON_MAPPED : NR_FILE_MAPPED; | |
1412 | __lruvec_stat_mod_folio(folio, idx, nr); | |
1413 | } | |
1414 | if (nr_pmdmapped) { | |
1415 | if (folio_test_anon(folio)) { | |
1416 | idx = NR_ANON_THPS; | |
1417 | __lruvec_stat_mod_folio(folio, idx, nr_pmdmapped); | |
1418 | } else { | |
1419 | /* NR_*_PMDMAPPED are not maintained per-memcg */ | |
1420 | idx = folio_test_swapbacked(folio) ? | |
1421 | NR_SHMEM_PMDMAPPED : NR_FILE_PMDMAPPED; | |
1422 | __mod_node_page_state(folio_pgdat(folio), idx, | |
1423 | nr_pmdmapped); | |
1424 | } | |
1425 | } | |
1426 | } | |
1427 | ||
1428 | static __always_inline void __folio_add_anon_rmap(struct folio *folio, | |
1429 | struct page *page, int nr_pages, struct vm_area_struct *vma, | |
1430 | unsigned long address, rmap_t flags, enum rmap_level level) | |
1431 | { | |
1432 | int i, nr, nr_pmdmapped = 0; | |
1433 | ||
1434 | VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio); | |
1435 | ||
1436 | nr = __folio_add_rmap(folio, page, nr_pages, vma, level, &nr_pmdmapped); | |
1437 | ||
1438 | if (likely(!folio_test_ksm(folio))) | |
1439 | __page_check_anon_rmap(folio, page, vma, address); | |
1440 | ||
1441 | __folio_mod_stat(folio, nr, nr_pmdmapped); | |
1442 | ||
1443 | if (flags & RMAP_EXCLUSIVE) { | |
1444 | switch (level) { | |
1445 | case RMAP_LEVEL_PTE: | |
1446 | for (i = 0; i < nr_pages; i++) | |
1447 | SetPageAnonExclusive(page + i); | |
1448 | break; | |
1449 | case RMAP_LEVEL_PMD: | |
1450 | SetPageAnonExclusive(page); | |
1451 | break; | |
1452 | case RMAP_LEVEL_PUD: | |
1453 | /* | |
1454 | * Keep the compiler happy, we don't support anonymous | |
1455 | * PUD mappings. | |
1456 | */ | |
1457 | WARN_ON_ONCE(1); | |
1458 | break; | |
1459 | } | |
1460 | } | |
1461 | ||
1462 | VM_WARN_ON_FOLIO(!folio_test_large(folio) && PageAnonExclusive(page) && | |
1463 | atomic_read(&folio->_mapcount) > 0, folio); | |
1464 | for (i = 0; i < nr_pages; i++) { | |
1465 | struct page *cur_page = page + i; | |
1466 | ||
1467 | VM_WARN_ON_FOLIO(folio_test_large(folio) && | |
1468 | folio_entire_mapcount(folio) > 1 && | |
1469 | PageAnonExclusive(cur_page), folio); | |
1470 | if (IS_ENABLED(CONFIG_NO_PAGE_MAPCOUNT)) | |
1471 | continue; | |
1472 | ||
1473 | /* | |
1474 | * While PTE-mapping a THP we have a PMD and a PTE | |
1475 | * mapping. | |
1476 | */ | |
1477 | VM_WARN_ON_FOLIO(atomic_read(&cur_page->_mapcount) > 0 && | |
1478 | PageAnonExclusive(cur_page), folio); | |
1479 | } | |
1480 | ||
1481 | /* | |
1482 | * For large folio, only mlock it if it's fully mapped to VMA. It's | |
1483 | * not easy to check whether the large folio is fully mapped to VMA | |
1484 | * here. Only mlock normal 4K folio and leave page reclaim to handle | |
1485 | * large folio. | |
1486 | */ | |
1487 | if (!folio_test_large(folio)) | |
1488 | mlock_vma_folio(folio, vma); | |
1489 | } | |
1490 | ||
1491 | /** | |
1492 | * folio_add_anon_rmap_ptes - add PTE mappings to a page range of an anon folio | |
1493 | * @folio: The folio to add the mappings to | |
1494 | * @page: The first page to add | |
1495 | * @nr_pages: The number of pages which will be mapped | |
1496 | * @vma: The vm area in which the mappings are added | |
1497 | * @address: The user virtual address of the first page to map | |
1498 | * @flags: The rmap flags | |
1499 | * | |
1500 | * The page range of folio is defined by [first_page, first_page + nr_pages) | |
1501 | * | |
1502 | * The caller needs to hold the page table lock, and the page must be locked in | |
1503 | * the anon_vma case: to serialize mapping,index checking after setting, | |
1504 | * and to ensure that an anon folio is not being upgraded racily to a KSM folio | |
1505 | * (but KSM folios are never downgraded). | |
1506 | */ | |
1507 | void folio_add_anon_rmap_ptes(struct folio *folio, struct page *page, | |
1508 | int nr_pages, struct vm_area_struct *vma, unsigned long address, | |
1509 | rmap_t flags) | |
1510 | { | |
1511 | __folio_add_anon_rmap(folio, page, nr_pages, vma, address, flags, | |
1512 | RMAP_LEVEL_PTE); | |
1513 | } | |
1514 | ||
1515 | /** | |
1516 | * folio_add_anon_rmap_pmd - add a PMD mapping to a page range of an anon folio | |
1517 | * @folio: The folio to add the mapping to | |
1518 | * @page: The first page to add | |
1519 | * @vma: The vm area in which the mapping is added | |
1520 | * @address: The user virtual address of the first page to map | |
1521 | * @flags: The rmap flags | |
1522 | * | |
1523 | * The page range of folio is defined by [first_page, first_page + HPAGE_PMD_NR) | |
1524 | * | |
1525 | * The caller needs to hold the page table lock, and the page must be locked in | |
1526 | * the anon_vma case: to serialize mapping,index checking after setting. | |
1527 | */ | |
1528 | void folio_add_anon_rmap_pmd(struct folio *folio, struct page *page, | |
1529 | struct vm_area_struct *vma, unsigned long address, rmap_t flags) | |
1530 | { | |
1531 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
1532 | __folio_add_anon_rmap(folio, page, HPAGE_PMD_NR, vma, address, flags, | |
1533 | RMAP_LEVEL_PMD); | |
1534 | #else | |
1535 | WARN_ON_ONCE(true); | |
1536 | #endif | |
1537 | } | |
1538 | ||
1539 | /** | |
1540 | * folio_add_new_anon_rmap - Add mapping to a new anonymous folio. | |
1541 | * @folio: The folio to add the mapping to. | |
1542 | * @vma: the vm area in which the mapping is added | |
1543 | * @address: the user virtual address mapped | |
1544 | * @flags: The rmap flags | |
1545 | * | |
1546 | * Like folio_add_anon_rmap_*() but must only be called on *new* folios. | |
1547 | * This means the inc-and-test can be bypassed. | |
1548 | * The folio doesn't necessarily need to be locked while it's exclusive | |
1549 | * unless two threads map it concurrently. However, the folio must be | |
1550 | * locked if it's shared. | |
1551 | * | |
1552 | * If the folio is pmd-mappable, it is accounted as a THP. | |
1553 | */ | |
1554 | void folio_add_new_anon_rmap(struct folio *folio, struct vm_area_struct *vma, | |
1555 | unsigned long address, rmap_t flags) | |
1556 | { | |
1557 | const bool exclusive = flags & RMAP_EXCLUSIVE; | |
1558 | int nr = 1, nr_pmdmapped = 0; | |
1559 | ||
1560 | VM_WARN_ON_FOLIO(folio_test_hugetlb(folio), folio); | |
1561 | VM_WARN_ON_FOLIO(!exclusive && !folio_test_locked(folio), folio); | |
1562 | ||
1563 | /* | |
1564 | * VM_DROPPABLE mappings don't swap; instead they're just dropped when | |
1565 | * under memory pressure. | |
1566 | */ | |
1567 | if (!folio_test_swapbacked(folio) && !(vma->vm_flags & VM_DROPPABLE)) | |
1568 | __folio_set_swapbacked(folio); | |
1569 | __folio_set_anon(folio, vma, address, exclusive); | |
1570 | ||
1571 | if (likely(!folio_test_large(folio))) { | |
1572 | /* increment count (starts at -1) */ | |
1573 | atomic_set(&folio->_mapcount, 0); | |
1574 | if (exclusive) | |
1575 | SetPageAnonExclusive(&folio->page); | |
1576 | } else if (!folio_test_pmd_mappable(folio)) { | |
1577 | int i; | |
1578 | ||
1579 | nr = folio_large_nr_pages(folio); | |
1580 | for (i = 0; i < nr; i++) { | |
1581 | struct page *page = folio_page(folio, i); | |
1582 | ||
1583 | if (IS_ENABLED(CONFIG_PAGE_MAPCOUNT)) | |
1584 | /* increment count (starts at -1) */ | |
1585 | atomic_set(&page->_mapcount, 0); | |
1586 | if (exclusive) | |
1587 | SetPageAnonExclusive(page); | |
1588 | } | |
1589 | ||
1590 | folio_set_large_mapcount(folio, nr, vma); | |
1591 | if (IS_ENABLED(CONFIG_PAGE_MAPCOUNT)) | |
1592 | atomic_set(&folio->_nr_pages_mapped, nr); | |
1593 | } else { | |
1594 | nr = folio_large_nr_pages(folio); | |
1595 | /* increment count (starts at -1) */ | |
1596 | atomic_set(&folio->_entire_mapcount, 0); | |
1597 | folio_set_large_mapcount(folio, 1, vma); | |
1598 | if (IS_ENABLED(CONFIG_PAGE_MAPCOUNT)) | |
1599 | atomic_set(&folio->_nr_pages_mapped, ENTIRELY_MAPPED); | |
1600 | if (exclusive) | |
1601 | SetPageAnonExclusive(&folio->page); | |
1602 | nr_pmdmapped = nr; | |
1603 | } | |
1604 | ||
1605 | VM_WARN_ON_ONCE(address < vma->vm_start || | |
1606 | address + (nr << PAGE_SHIFT) > vma->vm_end); | |
1607 | ||
1608 | __folio_mod_stat(folio, nr, nr_pmdmapped); | |
1609 | mod_mthp_stat(folio_order(folio), MTHP_STAT_NR_ANON, 1); | |
1610 | } | |
1611 | ||
1612 | static __always_inline void __folio_add_file_rmap(struct folio *folio, | |
1613 | struct page *page, int nr_pages, struct vm_area_struct *vma, | |
1614 | enum rmap_level level) | |
1615 | { | |
1616 | int nr, nr_pmdmapped = 0; | |
1617 | ||
1618 | VM_WARN_ON_FOLIO(folio_test_anon(folio), folio); | |
1619 | ||
1620 | nr = __folio_add_rmap(folio, page, nr_pages, vma, level, &nr_pmdmapped); | |
1621 | __folio_mod_stat(folio, nr, nr_pmdmapped); | |
1622 | ||
1623 | /* See comments in folio_add_anon_rmap_*() */ | |
1624 | if (!folio_test_large(folio)) | |
1625 | mlock_vma_folio(folio, vma); | |
1626 | } | |
1627 | ||
1628 | /** | |
1629 | * folio_add_file_rmap_ptes - add PTE mappings to a page range of a folio | |
1630 | * @folio: The folio to add the mappings to | |
1631 | * @page: The first page to add | |
1632 | * @nr_pages: The number of pages that will be mapped using PTEs | |
1633 | * @vma: The vm area in which the mappings are added | |
1634 | * | |
1635 | * The page range of the folio is defined by [page, page + nr_pages) | |
1636 | * | |
1637 | * The caller needs to hold the page table lock. | |
1638 | */ | |
1639 | void folio_add_file_rmap_ptes(struct folio *folio, struct page *page, | |
1640 | int nr_pages, struct vm_area_struct *vma) | |
1641 | { | |
1642 | __folio_add_file_rmap(folio, page, nr_pages, vma, RMAP_LEVEL_PTE); | |
1643 | } | |
1644 | ||
1645 | /** | |
1646 | * folio_add_file_rmap_pmd - add a PMD mapping to a page range of a folio | |
1647 | * @folio: The folio to add the mapping to | |
1648 | * @page: The first page to add | |
1649 | * @vma: The vm area in which the mapping is added | |
1650 | * | |
1651 | * The page range of the folio is defined by [page, page + HPAGE_PMD_NR) | |
1652 | * | |
1653 | * The caller needs to hold the page table lock. | |
1654 | */ | |
1655 | void folio_add_file_rmap_pmd(struct folio *folio, struct page *page, | |
1656 | struct vm_area_struct *vma) | |
1657 | { | |
1658 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
1659 | __folio_add_file_rmap(folio, page, HPAGE_PMD_NR, vma, RMAP_LEVEL_PMD); | |
1660 | #else | |
1661 | WARN_ON_ONCE(true); | |
1662 | #endif | |
1663 | } | |
1664 | ||
1665 | /** | |
1666 | * folio_add_file_rmap_pud - add a PUD mapping to a page range of a folio | |
1667 | * @folio: The folio to add the mapping to | |
1668 | * @page: The first page to add | |
1669 | * @vma: The vm area in which the mapping is added | |
1670 | * | |
1671 | * The page range of the folio is defined by [page, page + HPAGE_PUD_NR) | |
1672 | * | |
1673 | * The caller needs to hold the page table lock. | |
1674 | */ | |
1675 | void folio_add_file_rmap_pud(struct folio *folio, struct page *page, | |
1676 | struct vm_area_struct *vma) | |
1677 | { | |
1678 | #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \ | |
1679 | defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) | |
1680 | __folio_add_file_rmap(folio, page, HPAGE_PUD_NR, vma, RMAP_LEVEL_PUD); | |
1681 | #else | |
1682 | WARN_ON_ONCE(true); | |
1683 | #endif | |
1684 | } | |
1685 | ||
1686 | static __always_inline void __folio_remove_rmap(struct folio *folio, | |
1687 | struct page *page, int nr_pages, struct vm_area_struct *vma, | |
1688 | enum rmap_level level) | |
1689 | { | |
1690 | atomic_t *mapped = &folio->_nr_pages_mapped; | |
1691 | int last = 0, nr = 0, nr_pmdmapped = 0; | |
1692 | bool partially_mapped = false; | |
1693 | ||
1694 | __folio_rmap_sanity_checks(folio, page, nr_pages, level); | |
1695 | ||
1696 | switch (level) { | |
1697 | case RMAP_LEVEL_PTE: | |
1698 | if (!folio_test_large(folio)) { | |
1699 | nr = atomic_add_negative(-1, &folio->_mapcount); | |
1700 | break; | |
1701 | } | |
1702 | ||
1703 | if (IS_ENABLED(CONFIG_NO_PAGE_MAPCOUNT)) { | |
1704 | nr = folio_sub_return_large_mapcount(folio, nr_pages, vma); | |
1705 | if (!nr) { | |
1706 | /* Now completely unmapped. */ | |
1707 | nr = folio_nr_pages(folio); | |
1708 | } else { | |
1709 | partially_mapped = nr < folio_large_nr_pages(folio) && | |
1710 | !folio_entire_mapcount(folio); | |
1711 | nr = 0; | |
1712 | } | |
1713 | break; | |
1714 | } | |
1715 | ||
1716 | folio_sub_large_mapcount(folio, nr_pages, vma); | |
1717 | do { | |
1718 | last += atomic_add_negative(-1, &page->_mapcount); | |
1719 | } while (page++, --nr_pages > 0); | |
1720 | ||
1721 | if (last && | |
1722 | atomic_sub_return_relaxed(last, mapped) < ENTIRELY_MAPPED) | |
1723 | nr = last; | |
1724 | ||
1725 | partially_mapped = nr && atomic_read(mapped); | |
1726 | break; | |
1727 | case RMAP_LEVEL_PMD: | |
1728 | case RMAP_LEVEL_PUD: | |
1729 | if (IS_ENABLED(CONFIG_NO_PAGE_MAPCOUNT)) { | |
1730 | last = atomic_add_negative(-1, &folio->_entire_mapcount); | |
1731 | if (level == RMAP_LEVEL_PMD && last) | |
1732 | nr_pmdmapped = folio_large_nr_pages(folio); | |
1733 | nr = folio_dec_return_large_mapcount(folio, vma); | |
1734 | if (!nr) { | |
1735 | /* Now completely unmapped. */ | |
1736 | nr = folio_large_nr_pages(folio); | |
1737 | } else { | |
1738 | partially_mapped = last && | |
1739 | nr < folio_large_nr_pages(folio); | |
1740 | nr = 0; | |
1741 | } | |
1742 | break; | |
1743 | } | |
1744 | ||
1745 | folio_dec_large_mapcount(folio, vma); | |
1746 | last = atomic_add_negative(-1, &folio->_entire_mapcount); | |
1747 | if (last) { | |
1748 | nr = atomic_sub_return_relaxed(ENTIRELY_MAPPED, mapped); | |
1749 | if (likely(nr < ENTIRELY_MAPPED)) { | |
1750 | nr_pages = folio_large_nr_pages(folio); | |
1751 | if (level == RMAP_LEVEL_PMD) | |
1752 | nr_pmdmapped = nr_pages; | |
1753 | nr = nr_pages - (nr & FOLIO_PAGES_MAPPED); | |
1754 | /* Raced ahead of another remove and an add? */ | |
1755 | if (unlikely(nr < 0)) | |
1756 | nr = 0; | |
1757 | } else { | |
1758 | /* An add of ENTIRELY_MAPPED raced ahead */ | |
1759 | nr = 0; | |
1760 | } | |
1761 | } | |
1762 | ||
1763 | partially_mapped = nr && nr < nr_pmdmapped; | |
1764 | break; | |
1765 | } | |
1766 | ||
1767 | /* | |
1768 | * Queue anon large folio for deferred split if at least one page of | |
1769 | * the folio is unmapped and at least one page is still mapped. | |
1770 | * | |
1771 | * Check partially_mapped first to ensure it is a large folio. | |
1772 | */ | |
1773 | if (partially_mapped && folio_test_anon(folio) && | |
1774 | !folio_test_partially_mapped(folio)) | |
1775 | deferred_split_folio(folio, true); | |
1776 | ||
1777 | __folio_mod_stat(folio, -nr, -nr_pmdmapped); | |
1778 | ||
1779 | /* | |
1780 | * It would be tidy to reset folio_test_anon mapping when fully | |
1781 | * unmapped, but that might overwrite a racing folio_add_anon_rmap_*() | |
1782 | * which increments mapcount after us but sets mapping before us: | |
1783 | * so leave the reset to free_pages_prepare, and remember that | |
1784 | * it's only reliable while mapped. | |
1785 | */ | |
1786 | ||
1787 | munlock_vma_folio(folio, vma); | |
1788 | } | |
1789 | ||
1790 | /** | |
1791 | * folio_remove_rmap_ptes - remove PTE mappings from a page range of a folio | |
1792 | * @folio: The folio to remove the mappings from | |
1793 | * @page: The first page to remove | |
1794 | * @nr_pages: The number of pages that will be removed from the mapping | |
1795 | * @vma: The vm area from which the mappings are removed | |
1796 | * | |
1797 | * The page range of the folio is defined by [page, page + nr_pages) | |
1798 | * | |
1799 | * The caller needs to hold the page table lock. | |
1800 | */ | |
1801 | void folio_remove_rmap_ptes(struct folio *folio, struct page *page, | |
1802 | int nr_pages, struct vm_area_struct *vma) | |
1803 | { | |
1804 | __folio_remove_rmap(folio, page, nr_pages, vma, RMAP_LEVEL_PTE); | |
1805 | } | |
1806 | ||
1807 | /** | |
1808 | * folio_remove_rmap_pmd - remove a PMD mapping from a page range of a folio | |
1809 | * @folio: The folio to remove the mapping from | |
1810 | * @page: The first page to remove | |
1811 | * @vma: The vm area from which the mapping is removed | |
1812 | * | |
1813 | * The page range of the folio is defined by [page, page + HPAGE_PMD_NR) | |
1814 | * | |
1815 | * The caller needs to hold the page table lock. | |
1816 | */ | |
1817 | void folio_remove_rmap_pmd(struct folio *folio, struct page *page, | |
1818 | struct vm_area_struct *vma) | |
1819 | { | |
1820 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
1821 | __folio_remove_rmap(folio, page, HPAGE_PMD_NR, vma, RMAP_LEVEL_PMD); | |
1822 | #else | |
1823 | WARN_ON_ONCE(true); | |
1824 | #endif | |
1825 | } | |
1826 | ||
1827 | /** | |
1828 | * folio_remove_rmap_pud - remove a PUD mapping from a page range of a folio | |
1829 | * @folio: The folio to remove the mapping from | |
1830 | * @page: The first page to remove | |
1831 | * @vma: The vm area from which the mapping is removed | |
1832 | * | |
1833 | * The page range of the folio is defined by [page, page + HPAGE_PUD_NR) | |
1834 | * | |
1835 | * The caller needs to hold the page table lock. | |
1836 | */ | |
1837 | void folio_remove_rmap_pud(struct folio *folio, struct page *page, | |
1838 | struct vm_area_struct *vma) | |
1839 | { | |
1840 | #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \ | |
1841 | defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) | |
1842 | __folio_remove_rmap(folio, page, HPAGE_PUD_NR, vma, RMAP_LEVEL_PUD); | |
1843 | #else | |
1844 | WARN_ON_ONCE(true); | |
1845 | #endif | |
1846 | } | |
1847 | ||
1848 | /* We support batch unmapping of PTEs for lazyfree large folios */ | |
1849 | static inline bool can_batch_unmap_folio_ptes(unsigned long addr, | |
1850 | struct folio *folio, pte_t *ptep) | |
1851 | { | |
1852 | const fpb_t fpb_flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY; | |
1853 | int max_nr = folio_nr_pages(folio); | |
1854 | pte_t pte = ptep_get(ptep); | |
1855 | ||
1856 | if (!folio_test_anon(folio) || folio_test_swapbacked(folio)) | |
1857 | return false; | |
1858 | if (pte_unused(pte)) | |
1859 | return false; | |
1860 | if (pte_pfn(pte) != folio_pfn(folio)) | |
1861 | return false; | |
1862 | ||
1863 | return folio_pte_batch(folio, addr, ptep, pte, max_nr, fpb_flags, NULL, | |
1864 | NULL, NULL) == max_nr; | |
1865 | } | |
1866 | ||
1867 | /* | |
1868 | * @arg: enum ttu_flags will be passed to this argument | |
1869 | */ | |
1870 | static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, | |
1871 | unsigned long address, void *arg) | |
1872 | { | |
1873 | struct mm_struct *mm = vma->vm_mm; | |
1874 | DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0); | |
1875 | bool anon_exclusive, ret = true; | |
1876 | pte_t pteval; | |
1877 | struct page *subpage; | |
1878 | struct mmu_notifier_range range; | |
1879 | enum ttu_flags flags = (enum ttu_flags)(long)arg; | |
1880 | unsigned long nr_pages = 1, end_addr; | |
1881 | unsigned long pfn; | |
1882 | unsigned long hsz = 0; | |
1883 | ||
1884 | /* | |
1885 | * When racing against e.g. zap_pte_range() on another cpu, | |
1886 | * in between its ptep_get_and_clear_full() and folio_remove_rmap_*(), | |
1887 | * try_to_unmap() may return before page_mapped() has become false, | |
1888 | * if page table locking is skipped: use TTU_SYNC to wait for that. | |
1889 | */ | |
1890 | if (flags & TTU_SYNC) | |
1891 | pvmw.flags = PVMW_SYNC; | |
1892 | ||
1893 | /* | |
1894 | * For THP, we have to assume the worse case ie pmd for invalidation. | |
1895 | * For hugetlb, it could be much worse if we need to do pud | |
1896 | * invalidation in the case of pmd sharing. | |
1897 | * | |
1898 | * Note that the folio can not be freed in this function as call of | |
1899 | * try_to_unmap() must hold a reference on the folio. | |
1900 | */ | |
1901 | range.end = vma_address_end(&pvmw); | |
1902 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, | |
1903 | address, range.end); | |
1904 | if (folio_test_hugetlb(folio)) { | |
1905 | /* | |
1906 | * If sharing is possible, start and end will be adjusted | |
1907 | * accordingly. | |
1908 | */ | |
1909 | adjust_range_if_pmd_sharing_possible(vma, &range.start, | |
1910 | &range.end); | |
1911 | ||
1912 | /* We need the huge page size for set_huge_pte_at() */ | |
1913 | hsz = huge_page_size(hstate_vma(vma)); | |
1914 | } | |
1915 | mmu_notifier_invalidate_range_start(&range); | |
1916 | ||
1917 | while (page_vma_mapped_walk(&pvmw)) { | |
1918 | /* | |
1919 | * If the folio is in an mlock()d vma, we must not swap it out. | |
1920 | */ | |
1921 | if (!(flags & TTU_IGNORE_MLOCK) && | |
1922 | (vma->vm_flags & VM_LOCKED)) { | |
1923 | /* Restore the mlock which got missed */ | |
1924 | if (!folio_test_large(folio)) | |
1925 | mlock_vma_folio(folio, vma); | |
1926 | goto walk_abort; | |
1927 | } | |
1928 | ||
1929 | if (!pvmw.pte) { | |
1930 | if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) { | |
1931 | if (unmap_huge_pmd_locked(vma, pvmw.address, pvmw.pmd, folio)) | |
1932 | goto walk_done; | |
1933 | /* | |
1934 | * unmap_huge_pmd_locked has either already marked | |
1935 | * the folio as swap-backed or decided to retain it | |
1936 | * due to GUP or speculative references. | |
1937 | */ | |
1938 | goto walk_abort; | |
1939 | } | |
1940 | ||
1941 | if (flags & TTU_SPLIT_HUGE_PMD) { | |
1942 | /* | |
1943 | * We temporarily have to drop the PTL and | |
1944 | * restart so we can process the PTE-mapped THP. | |
1945 | */ | |
1946 | split_huge_pmd_locked(vma, pvmw.address, | |
1947 | pvmw.pmd, false); | |
1948 | flags &= ~TTU_SPLIT_HUGE_PMD; | |
1949 | page_vma_mapped_walk_restart(&pvmw); | |
1950 | continue; | |
1951 | } | |
1952 | } | |
1953 | ||
1954 | /* Unexpected PMD-mapped THP? */ | |
1955 | VM_BUG_ON_FOLIO(!pvmw.pte, folio); | |
1956 | ||
1957 | /* | |
1958 | * Handle PFN swap PTEs, such as device-exclusive ones, that | |
1959 | * actually map pages. | |
1960 | */ | |
1961 | pteval = ptep_get(pvmw.pte); | |
1962 | if (likely(pte_present(pteval))) { | |
1963 | pfn = pte_pfn(pteval); | |
1964 | } else { | |
1965 | pfn = swp_offset_pfn(pte_to_swp_entry(pteval)); | |
1966 | VM_WARN_ON_FOLIO(folio_test_hugetlb(folio), folio); | |
1967 | } | |
1968 | ||
1969 | subpage = folio_page(folio, pfn - folio_pfn(folio)); | |
1970 | address = pvmw.address; | |
1971 | anon_exclusive = folio_test_anon(folio) && | |
1972 | PageAnonExclusive(subpage); | |
1973 | ||
1974 | if (folio_test_hugetlb(folio)) { | |
1975 | bool anon = folio_test_anon(folio); | |
1976 | ||
1977 | /* | |
1978 | * The try_to_unmap() is only passed a hugetlb page | |
1979 | * in the case where the hugetlb page is poisoned. | |
1980 | */ | |
1981 | VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage); | |
1982 | /* | |
1983 | * huge_pmd_unshare may unmap an entire PMD page. | |
1984 | * There is no way of knowing exactly which PMDs may | |
1985 | * be cached for this mm, so we must flush them all. | |
1986 | * start/end were already adjusted above to cover this | |
1987 | * range. | |
1988 | */ | |
1989 | flush_cache_range(vma, range.start, range.end); | |
1990 | ||
1991 | /* | |
1992 | * To call huge_pmd_unshare, i_mmap_rwsem must be | |
1993 | * held in write mode. Caller needs to explicitly | |
1994 | * do this outside rmap routines. | |
1995 | * | |
1996 | * We also must hold hugetlb vma_lock in write mode. | |
1997 | * Lock order dictates acquiring vma_lock BEFORE | |
1998 | * i_mmap_rwsem. We can only try lock here and fail | |
1999 | * if unsuccessful. | |
2000 | */ | |
2001 | if (!anon) { | |
2002 | VM_BUG_ON(!(flags & TTU_RMAP_LOCKED)); | |
2003 | if (!hugetlb_vma_trylock_write(vma)) | |
2004 | goto walk_abort; | |
2005 | if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) { | |
2006 | hugetlb_vma_unlock_write(vma); | |
2007 | flush_tlb_range(vma, | |
2008 | range.start, range.end); | |
2009 | /* | |
2010 | * The ref count of the PMD page was | |
2011 | * dropped which is part of the way map | |
2012 | * counting is done for shared PMDs. | |
2013 | * Return 'true' here. When there is | |
2014 | * no other sharing, huge_pmd_unshare | |
2015 | * returns false and we will unmap the | |
2016 | * actual page and drop map count | |
2017 | * to zero. | |
2018 | */ | |
2019 | goto walk_done; | |
2020 | } | |
2021 | hugetlb_vma_unlock_write(vma); | |
2022 | } | |
2023 | pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); | |
2024 | if (pte_dirty(pteval)) | |
2025 | folio_mark_dirty(folio); | |
2026 | } else if (likely(pte_present(pteval))) { | |
2027 | if (folio_test_large(folio) && !(flags & TTU_HWPOISON) && | |
2028 | can_batch_unmap_folio_ptes(address, folio, pvmw.pte)) | |
2029 | nr_pages = folio_nr_pages(folio); | |
2030 | end_addr = address + nr_pages * PAGE_SIZE; | |
2031 | flush_cache_range(vma, address, end_addr); | |
2032 | ||
2033 | /* Nuke the page table entry. */ | |
2034 | pteval = get_and_clear_full_ptes(mm, address, pvmw.pte, nr_pages, 0); | |
2035 | /* | |
2036 | * We clear the PTE but do not flush so potentially | |
2037 | * a remote CPU could still be writing to the folio. | |
2038 | * If the entry was previously clean then the | |
2039 | * architecture must guarantee that a clear->dirty | |
2040 | * transition on a cached TLB entry is written through | |
2041 | * and traps if the PTE is unmapped. | |
2042 | */ | |
2043 | if (should_defer_flush(mm, flags)) | |
2044 | set_tlb_ubc_flush_pending(mm, pteval, address, end_addr); | |
2045 | else | |
2046 | flush_tlb_range(vma, address, end_addr); | |
2047 | if (pte_dirty(pteval)) | |
2048 | folio_mark_dirty(folio); | |
2049 | } else { | |
2050 | pte_clear(mm, address, pvmw.pte); | |
2051 | } | |
2052 | ||
2053 | /* | |
2054 | * Now the pte is cleared. If this pte was uffd-wp armed, | |
2055 | * we may want to replace a none pte with a marker pte if | |
2056 | * it's file-backed, so we don't lose the tracking info. | |
2057 | */ | |
2058 | pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval); | |
2059 | ||
2060 | /* Update high watermark before we lower rss */ | |
2061 | update_hiwater_rss(mm); | |
2062 | ||
2063 | if (PageHWPoison(subpage) && (flags & TTU_HWPOISON)) { | |
2064 | pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); | |
2065 | if (folio_test_hugetlb(folio)) { | |
2066 | hugetlb_count_sub(folio_nr_pages(folio), mm); | |
2067 | set_huge_pte_at(mm, address, pvmw.pte, pteval, | |
2068 | hsz); | |
2069 | } else { | |
2070 | dec_mm_counter(mm, mm_counter(folio)); | |
2071 | set_pte_at(mm, address, pvmw.pte, pteval); | |
2072 | } | |
2073 | } else if (likely(pte_present(pteval)) && pte_unused(pteval) && | |
2074 | !userfaultfd_armed(vma)) { | |
2075 | /* | |
2076 | * The guest indicated that the page content is of no | |
2077 | * interest anymore. Simply discard the pte, vmscan | |
2078 | * will take care of the rest. | |
2079 | * A future reference will then fault in a new zero | |
2080 | * page. When userfaultfd is active, we must not drop | |
2081 | * this page though, as its main user (postcopy | |
2082 | * migration) will not expect userfaults on already | |
2083 | * copied pages. | |
2084 | */ | |
2085 | dec_mm_counter(mm, mm_counter(folio)); | |
2086 | } else if (folio_test_anon(folio)) { | |
2087 | swp_entry_t entry = page_swap_entry(subpage); | |
2088 | pte_t swp_pte; | |
2089 | /* | |
2090 | * Store the swap location in the pte. | |
2091 | * See handle_pte_fault() ... | |
2092 | */ | |
2093 | if (unlikely(folio_test_swapbacked(folio) != | |
2094 | folio_test_swapcache(folio))) { | |
2095 | WARN_ON_ONCE(1); | |
2096 | goto walk_abort; | |
2097 | } | |
2098 | ||
2099 | /* MADV_FREE page check */ | |
2100 | if (!folio_test_swapbacked(folio)) { | |
2101 | int ref_count, map_count; | |
2102 | ||
2103 | /* | |
2104 | * Synchronize with gup_pte_range(): | |
2105 | * - clear PTE; barrier; read refcount | |
2106 | * - inc refcount; barrier; read PTE | |
2107 | */ | |
2108 | smp_mb(); | |
2109 | ||
2110 | ref_count = folio_ref_count(folio); | |
2111 | map_count = folio_mapcount(folio); | |
2112 | ||
2113 | /* | |
2114 | * Order reads for page refcount and dirty flag | |
2115 | * (see comments in __remove_mapping()). | |
2116 | */ | |
2117 | smp_rmb(); | |
2118 | ||
2119 | if (folio_test_dirty(folio) && !(vma->vm_flags & VM_DROPPABLE)) { | |
2120 | /* | |
2121 | * redirtied either using the page table or a previously | |
2122 | * obtained GUP reference. | |
2123 | */ | |
2124 | set_ptes(mm, address, pvmw.pte, pteval, nr_pages); | |
2125 | folio_set_swapbacked(folio); | |
2126 | goto walk_abort; | |
2127 | } else if (ref_count != 1 + map_count) { | |
2128 | /* | |
2129 | * Additional reference. Could be a GUP reference or any | |
2130 | * speculative reference. GUP users must mark the folio | |
2131 | * dirty if there was a modification. This folio cannot be | |
2132 | * reclaimed right now either way, so act just like nothing | |
2133 | * happened. | |
2134 | * We'll come back here later and detect if the folio was | |
2135 | * dirtied when the additional reference is gone. | |
2136 | */ | |
2137 | set_ptes(mm, address, pvmw.pte, pteval, nr_pages); | |
2138 | goto walk_abort; | |
2139 | } | |
2140 | add_mm_counter(mm, MM_ANONPAGES, -nr_pages); | |
2141 | goto discard; | |
2142 | } | |
2143 | ||
2144 | if (swap_duplicate(entry) < 0) { | |
2145 | set_pte_at(mm, address, pvmw.pte, pteval); | |
2146 | goto walk_abort; | |
2147 | } | |
2148 | ||
2149 | /* | |
2150 | * arch_unmap_one() is expected to be a NOP on | |
2151 | * architectures where we could have PFN swap PTEs, | |
2152 | * so we'll not check/care. | |
2153 | */ | |
2154 | if (arch_unmap_one(mm, vma, address, pteval) < 0) { | |
2155 | swap_free(entry); | |
2156 | set_pte_at(mm, address, pvmw.pte, pteval); | |
2157 | goto walk_abort; | |
2158 | } | |
2159 | ||
2160 | /* See folio_try_share_anon_rmap(): clear PTE first. */ | |
2161 | if (anon_exclusive && | |
2162 | folio_try_share_anon_rmap_pte(folio, subpage)) { | |
2163 | swap_free(entry); | |
2164 | set_pte_at(mm, address, pvmw.pte, pteval); | |
2165 | goto walk_abort; | |
2166 | } | |
2167 | if (list_empty(&mm->mmlist)) { | |
2168 | spin_lock(&mmlist_lock); | |
2169 | if (list_empty(&mm->mmlist)) | |
2170 | list_add(&mm->mmlist, &init_mm.mmlist); | |
2171 | spin_unlock(&mmlist_lock); | |
2172 | } | |
2173 | dec_mm_counter(mm, MM_ANONPAGES); | |
2174 | inc_mm_counter(mm, MM_SWAPENTS); | |
2175 | swp_pte = swp_entry_to_pte(entry); | |
2176 | if (anon_exclusive) | |
2177 | swp_pte = pte_swp_mkexclusive(swp_pte); | |
2178 | if (likely(pte_present(pteval))) { | |
2179 | if (pte_soft_dirty(pteval)) | |
2180 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2181 | if (pte_uffd_wp(pteval)) | |
2182 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
2183 | } else { | |
2184 | if (pte_swp_soft_dirty(pteval)) | |
2185 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2186 | if (pte_swp_uffd_wp(pteval)) | |
2187 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
2188 | } | |
2189 | set_pte_at(mm, address, pvmw.pte, swp_pte); | |
2190 | } else { | |
2191 | /* | |
2192 | * This is a locked file-backed folio, | |
2193 | * so it cannot be removed from the page | |
2194 | * cache and replaced by a new folio before | |
2195 | * mmu_notifier_invalidate_range_end, so no | |
2196 | * concurrent thread might update its page table | |
2197 | * to point at a new folio while a device is | |
2198 | * still using this folio. | |
2199 | * | |
2200 | * See Documentation/mm/mmu_notifier.rst | |
2201 | */ | |
2202 | dec_mm_counter(mm, mm_counter_file(folio)); | |
2203 | } | |
2204 | discard: | |
2205 | if (unlikely(folio_test_hugetlb(folio))) { | |
2206 | hugetlb_remove_rmap(folio); | |
2207 | } else { | |
2208 | folio_remove_rmap_ptes(folio, subpage, nr_pages, vma); | |
2209 | folio_ref_sub(folio, nr_pages - 1); | |
2210 | } | |
2211 | if (vma->vm_flags & VM_LOCKED) | |
2212 | mlock_drain_local(); | |
2213 | folio_put(folio); | |
2214 | /* We have already batched the entire folio */ | |
2215 | if (nr_pages > 1) | |
2216 | goto walk_done; | |
2217 | continue; | |
2218 | walk_abort: | |
2219 | ret = false; | |
2220 | walk_done: | |
2221 | page_vma_mapped_walk_done(&pvmw); | |
2222 | break; | |
2223 | } | |
2224 | ||
2225 | mmu_notifier_invalidate_range_end(&range); | |
2226 | ||
2227 | return ret; | |
2228 | } | |
2229 | ||
2230 | static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg) | |
2231 | { | |
2232 | return vma_is_temporary_stack(vma); | |
2233 | } | |
2234 | ||
2235 | static int folio_not_mapped(struct folio *folio) | |
2236 | { | |
2237 | return !folio_mapped(folio); | |
2238 | } | |
2239 | ||
2240 | /** | |
2241 | * try_to_unmap - Try to remove all page table mappings to a folio. | |
2242 | * @folio: The folio to unmap. | |
2243 | * @flags: action and flags | |
2244 | * | |
2245 | * Tries to remove all the page table entries which are mapping this | |
2246 | * folio. It is the caller's responsibility to check if the folio is | |
2247 | * still mapped if needed (use TTU_SYNC to prevent accounting races). | |
2248 | * | |
2249 | * Context: Caller must hold the folio lock. | |
2250 | */ | |
2251 | void try_to_unmap(struct folio *folio, enum ttu_flags flags) | |
2252 | { | |
2253 | struct rmap_walk_control rwc = { | |
2254 | .rmap_one = try_to_unmap_one, | |
2255 | .arg = (void *)flags, | |
2256 | .done = folio_not_mapped, | |
2257 | .anon_lock = folio_lock_anon_vma_read, | |
2258 | }; | |
2259 | ||
2260 | if (flags & TTU_RMAP_LOCKED) | |
2261 | rmap_walk_locked(folio, &rwc); | |
2262 | else | |
2263 | rmap_walk(folio, &rwc); | |
2264 | } | |
2265 | ||
2266 | /* | |
2267 | * @arg: enum ttu_flags will be passed to this argument. | |
2268 | * | |
2269 | * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs | |
2270 | * containing migration entries. | |
2271 | */ | |
2272 | static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma, | |
2273 | unsigned long address, void *arg) | |
2274 | { | |
2275 | struct mm_struct *mm = vma->vm_mm; | |
2276 | DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0); | |
2277 | bool anon_exclusive, writable, ret = true; | |
2278 | pte_t pteval; | |
2279 | struct page *subpage; | |
2280 | struct mmu_notifier_range range; | |
2281 | enum ttu_flags flags = (enum ttu_flags)(long)arg; | |
2282 | unsigned long pfn; | |
2283 | unsigned long hsz = 0; | |
2284 | ||
2285 | /* | |
2286 | * When racing against e.g. zap_pte_range() on another cpu, | |
2287 | * in between its ptep_get_and_clear_full() and folio_remove_rmap_*(), | |
2288 | * try_to_migrate() may return before page_mapped() has become false, | |
2289 | * if page table locking is skipped: use TTU_SYNC to wait for that. | |
2290 | */ | |
2291 | if (flags & TTU_SYNC) | |
2292 | pvmw.flags = PVMW_SYNC; | |
2293 | ||
2294 | /* | |
2295 | * For THP, we have to assume the worse case ie pmd for invalidation. | |
2296 | * For hugetlb, it could be much worse if we need to do pud | |
2297 | * invalidation in the case of pmd sharing. | |
2298 | * | |
2299 | * Note that the page can not be free in this function as call of | |
2300 | * try_to_unmap() must hold a reference on the page. | |
2301 | */ | |
2302 | range.end = vma_address_end(&pvmw); | |
2303 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, | |
2304 | address, range.end); | |
2305 | if (folio_test_hugetlb(folio)) { | |
2306 | /* | |
2307 | * If sharing is possible, start and end will be adjusted | |
2308 | * accordingly. | |
2309 | */ | |
2310 | adjust_range_if_pmd_sharing_possible(vma, &range.start, | |
2311 | &range.end); | |
2312 | ||
2313 | /* We need the huge page size for set_huge_pte_at() */ | |
2314 | hsz = huge_page_size(hstate_vma(vma)); | |
2315 | } | |
2316 | mmu_notifier_invalidate_range_start(&range); | |
2317 | ||
2318 | while (page_vma_mapped_walk(&pvmw)) { | |
2319 | /* PMD-mapped THP migration entry */ | |
2320 | if (!pvmw.pte) { | |
2321 | if (flags & TTU_SPLIT_HUGE_PMD) { | |
2322 | split_huge_pmd_locked(vma, pvmw.address, | |
2323 | pvmw.pmd, true); | |
2324 | ret = false; | |
2325 | page_vma_mapped_walk_done(&pvmw); | |
2326 | break; | |
2327 | } | |
2328 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION | |
2329 | subpage = folio_page(folio, | |
2330 | pmd_pfn(*pvmw.pmd) - folio_pfn(folio)); | |
2331 | VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) || | |
2332 | !folio_test_pmd_mappable(folio), folio); | |
2333 | ||
2334 | if (set_pmd_migration_entry(&pvmw, subpage)) { | |
2335 | ret = false; | |
2336 | page_vma_mapped_walk_done(&pvmw); | |
2337 | break; | |
2338 | } | |
2339 | continue; | |
2340 | #endif | |
2341 | } | |
2342 | ||
2343 | /* Unexpected PMD-mapped THP? */ | |
2344 | VM_BUG_ON_FOLIO(!pvmw.pte, folio); | |
2345 | ||
2346 | /* | |
2347 | * Handle PFN swap PTEs, such as device-exclusive ones, that | |
2348 | * actually map pages. | |
2349 | */ | |
2350 | pteval = ptep_get(pvmw.pte); | |
2351 | if (likely(pte_present(pteval))) { | |
2352 | pfn = pte_pfn(pteval); | |
2353 | } else { | |
2354 | pfn = swp_offset_pfn(pte_to_swp_entry(pteval)); | |
2355 | VM_WARN_ON_FOLIO(folio_test_hugetlb(folio), folio); | |
2356 | } | |
2357 | ||
2358 | subpage = folio_page(folio, pfn - folio_pfn(folio)); | |
2359 | address = pvmw.address; | |
2360 | anon_exclusive = folio_test_anon(folio) && | |
2361 | PageAnonExclusive(subpage); | |
2362 | ||
2363 | if (folio_test_hugetlb(folio)) { | |
2364 | bool anon = folio_test_anon(folio); | |
2365 | ||
2366 | /* | |
2367 | * huge_pmd_unshare may unmap an entire PMD page. | |
2368 | * There is no way of knowing exactly which PMDs may | |
2369 | * be cached for this mm, so we must flush them all. | |
2370 | * start/end were already adjusted above to cover this | |
2371 | * range. | |
2372 | */ | |
2373 | flush_cache_range(vma, range.start, range.end); | |
2374 | ||
2375 | /* | |
2376 | * To call huge_pmd_unshare, i_mmap_rwsem must be | |
2377 | * held in write mode. Caller needs to explicitly | |
2378 | * do this outside rmap routines. | |
2379 | * | |
2380 | * We also must hold hugetlb vma_lock in write mode. | |
2381 | * Lock order dictates acquiring vma_lock BEFORE | |
2382 | * i_mmap_rwsem. We can only try lock here and | |
2383 | * fail if unsuccessful. | |
2384 | */ | |
2385 | if (!anon) { | |
2386 | VM_BUG_ON(!(flags & TTU_RMAP_LOCKED)); | |
2387 | if (!hugetlb_vma_trylock_write(vma)) { | |
2388 | page_vma_mapped_walk_done(&pvmw); | |
2389 | ret = false; | |
2390 | break; | |
2391 | } | |
2392 | if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) { | |
2393 | hugetlb_vma_unlock_write(vma); | |
2394 | flush_tlb_range(vma, | |
2395 | range.start, range.end); | |
2396 | ||
2397 | /* | |
2398 | * The ref count of the PMD page was | |
2399 | * dropped which is part of the way map | |
2400 | * counting is done for shared PMDs. | |
2401 | * Return 'true' here. When there is | |
2402 | * no other sharing, huge_pmd_unshare | |
2403 | * returns false and we will unmap the | |
2404 | * actual page and drop map count | |
2405 | * to zero. | |
2406 | */ | |
2407 | page_vma_mapped_walk_done(&pvmw); | |
2408 | break; | |
2409 | } | |
2410 | hugetlb_vma_unlock_write(vma); | |
2411 | } | |
2412 | /* Nuke the hugetlb page table entry */ | |
2413 | pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); | |
2414 | if (pte_dirty(pteval)) | |
2415 | folio_mark_dirty(folio); | |
2416 | writable = pte_write(pteval); | |
2417 | } else if (likely(pte_present(pteval))) { | |
2418 | flush_cache_page(vma, address, pfn); | |
2419 | /* Nuke the page table entry. */ | |
2420 | if (should_defer_flush(mm, flags)) { | |
2421 | /* | |
2422 | * We clear the PTE but do not flush so potentially | |
2423 | * a remote CPU could still be writing to the folio. | |
2424 | * If the entry was previously clean then the | |
2425 | * architecture must guarantee that a clear->dirty | |
2426 | * transition on a cached TLB entry is written through | |
2427 | * and traps if the PTE is unmapped. | |
2428 | */ | |
2429 | pteval = ptep_get_and_clear(mm, address, pvmw.pte); | |
2430 | ||
2431 | set_tlb_ubc_flush_pending(mm, pteval, address, address + PAGE_SIZE); | |
2432 | } else { | |
2433 | pteval = ptep_clear_flush(vma, address, pvmw.pte); | |
2434 | } | |
2435 | if (pte_dirty(pteval)) | |
2436 | folio_mark_dirty(folio); | |
2437 | writable = pte_write(pteval); | |
2438 | } else { | |
2439 | pte_clear(mm, address, pvmw.pte); | |
2440 | writable = is_writable_device_private_entry(pte_to_swp_entry(pteval)); | |
2441 | } | |
2442 | ||
2443 | VM_WARN_ON_FOLIO(writable && folio_test_anon(folio) && | |
2444 | !anon_exclusive, folio); | |
2445 | ||
2446 | /* Update high watermark before we lower rss */ | |
2447 | update_hiwater_rss(mm); | |
2448 | ||
2449 | if (PageHWPoison(subpage)) { | |
2450 | VM_WARN_ON_FOLIO(folio_is_device_private(folio), folio); | |
2451 | ||
2452 | pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); | |
2453 | if (folio_test_hugetlb(folio)) { | |
2454 | hugetlb_count_sub(folio_nr_pages(folio), mm); | |
2455 | set_huge_pte_at(mm, address, pvmw.pte, pteval, | |
2456 | hsz); | |
2457 | } else { | |
2458 | dec_mm_counter(mm, mm_counter(folio)); | |
2459 | set_pte_at(mm, address, pvmw.pte, pteval); | |
2460 | } | |
2461 | } else if (likely(pte_present(pteval)) && pte_unused(pteval) && | |
2462 | !userfaultfd_armed(vma)) { | |
2463 | /* | |
2464 | * The guest indicated that the page content is of no | |
2465 | * interest anymore. Simply discard the pte, vmscan | |
2466 | * will take care of the rest. | |
2467 | * A future reference will then fault in a new zero | |
2468 | * page. When userfaultfd is active, we must not drop | |
2469 | * this page though, as its main user (postcopy | |
2470 | * migration) will not expect userfaults on already | |
2471 | * copied pages. | |
2472 | */ | |
2473 | dec_mm_counter(mm, mm_counter(folio)); | |
2474 | } else { | |
2475 | swp_entry_t entry; | |
2476 | pte_t swp_pte; | |
2477 | ||
2478 | /* | |
2479 | * arch_unmap_one() is expected to be a NOP on | |
2480 | * architectures where we could have PFN swap PTEs, | |
2481 | * so we'll not check/care. | |
2482 | */ | |
2483 | if (arch_unmap_one(mm, vma, address, pteval) < 0) { | |
2484 | if (folio_test_hugetlb(folio)) | |
2485 | set_huge_pte_at(mm, address, pvmw.pte, | |
2486 | pteval, hsz); | |
2487 | else | |
2488 | set_pte_at(mm, address, pvmw.pte, pteval); | |
2489 | ret = false; | |
2490 | page_vma_mapped_walk_done(&pvmw); | |
2491 | break; | |
2492 | } | |
2493 | ||
2494 | /* See folio_try_share_anon_rmap_pte(): clear PTE first. */ | |
2495 | if (folio_test_hugetlb(folio)) { | |
2496 | if (anon_exclusive && | |
2497 | hugetlb_try_share_anon_rmap(folio)) { | |
2498 | set_huge_pte_at(mm, address, pvmw.pte, | |
2499 | pteval, hsz); | |
2500 | ret = false; | |
2501 | page_vma_mapped_walk_done(&pvmw); | |
2502 | break; | |
2503 | } | |
2504 | } else if (anon_exclusive && | |
2505 | folio_try_share_anon_rmap_pte(folio, subpage)) { | |
2506 | set_pte_at(mm, address, pvmw.pte, pteval); | |
2507 | ret = false; | |
2508 | page_vma_mapped_walk_done(&pvmw); | |
2509 | break; | |
2510 | } | |
2511 | ||
2512 | /* | |
2513 | * Store the pfn of the page in a special migration | |
2514 | * pte. do_swap_page() will wait until the migration | |
2515 | * pte is removed and then restart fault handling. | |
2516 | */ | |
2517 | if (writable) | |
2518 | entry = make_writable_migration_entry( | |
2519 | page_to_pfn(subpage)); | |
2520 | else if (anon_exclusive) | |
2521 | entry = make_readable_exclusive_migration_entry( | |
2522 | page_to_pfn(subpage)); | |
2523 | else | |
2524 | entry = make_readable_migration_entry( | |
2525 | page_to_pfn(subpage)); | |
2526 | if (likely(pte_present(pteval))) { | |
2527 | if (pte_young(pteval)) | |
2528 | entry = make_migration_entry_young(entry); | |
2529 | if (pte_dirty(pteval)) | |
2530 | entry = make_migration_entry_dirty(entry); | |
2531 | swp_pte = swp_entry_to_pte(entry); | |
2532 | if (pte_soft_dirty(pteval)) | |
2533 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2534 | if (pte_uffd_wp(pteval)) | |
2535 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
2536 | } else { | |
2537 | swp_pte = swp_entry_to_pte(entry); | |
2538 | if (pte_swp_soft_dirty(pteval)) | |
2539 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2540 | if (pte_swp_uffd_wp(pteval)) | |
2541 | swp_pte = pte_swp_mkuffd_wp(swp_pte); | |
2542 | } | |
2543 | if (folio_test_hugetlb(folio)) | |
2544 | set_huge_pte_at(mm, address, pvmw.pte, swp_pte, | |
2545 | hsz); | |
2546 | else | |
2547 | set_pte_at(mm, address, pvmw.pte, swp_pte); | |
2548 | trace_set_migration_pte(address, pte_val(swp_pte), | |
2549 | folio_order(folio)); | |
2550 | /* | |
2551 | * No need to invalidate here it will synchronize on | |
2552 | * against the special swap migration pte. | |
2553 | */ | |
2554 | } | |
2555 | ||
2556 | if (unlikely(folio_test_hugetlb(folio))) | |
2557 | hugetlb_remove_rmap(folio); | |
2558 | else | |
2559 | folio_remove_rmap_pte(folio, subpage, vma); | |
2560 | if (vma->vm_flags & VM_LOCKED) | |
2561 | mlock_drain_local(); | |
2562 | folio_put(folio); | |
2563 | } | |
2564 | ||
2565 | mmu_notifier_invalidate_range_end(&range); | |
2566 | ||
2567 | return ret; | |
2568 | } | |
2569 | ||
2570 | /** | |
2571 | * try_to_migrate - try to replace all page table mappings with swap entries | |
2572 | * @folio: the folio to replace page table entries for | |
2573 | * @flags: action and flags | |
2574 | * | |
2575 | * Tries to remove all the page table entries which are mapping this folio and | |
2576 | * replace them with special swap entries. Caller must hold the folio lock. | |
2577 | */ | |
2578 | void try_to_migrate(struct folio *folio, enum ttu_flags flags) | |
2579 | { | |
2580 | struct rmap_walk_control rwc = { | |
2581 | .rmap_one = try_to_migrate_one, | |
2582 | .arg = (void *)flags, | |
2583 | .done = folio_not_mapped, | |
2584 | .anon_lock = folio_lock_anon_vma_read, | |
2585 | }; | |
2586 | ||
2587 | /* | |
2588 | * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and | |
2589 | * TTU_SPLIT_HUGE_PMD, TTU_SYNC, and TTU_BATCH_FLUSH flags. | |
2590 | */ | |
2591 | if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD | | |
2592 | TTU_SYNC | TTU_BATCH_FLUSH))) | |
2593 | return; | |
2594 | ||
2595 | if (folio_is_zone_device(folio) && | |
2596 | (!folio_is_device_private(folio) && !folio_is_device_coherent(folio))) | |
2597 | return; | |
2598 | ||
2599 | /* | |
2600 | * During exec, a temporary VMA is setup and later moved. | |
2601 | * The VMA is moved under the anon_vma lock but not the | |
2602 | * page tables leading to a race where migration cannot | |
2603 | * find the migration ptes. Rather than increasing the | |
2604 | * locking requirements of exec(), migration skips | |
2605 | * temporary VMAs until after exec() completes. | |
2606 | */ | |
2607 | if (!folio_test_ksm(folio) && folio_test_anon(folio)) | |
2608 | rwc.invalid_vma = invalid_migration_vma; | |
2609 | ||
2610 | if (flags & TTU_RMAP_LOCKED) | |
2611 | rmap_walk_locked(folio, &rwc); | |
2612 | else | |
2613 | rmap_walk(folio, &rwc); | |
2614 | } | |
2615 | ||
2616 | #ifdef CONFIG_DEVICE_PRIVATE | |
2617 | /** | |
2618 | * make_device_exclusive() - Mark a page for exclusive use by a device | |
2619 | * @mm: mm_struct of associated target process | |
2620 | * @addr: the virtual address to mark for exclusive device access | |
2621 | * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering | |
2622 | * @foliop: folio pointer will be stored here on success. | |
2623 | * | |
2624 | * This function looks up the page mapped at the given address, grabs a | |
2625 | * folio reference, locks the folio and replaces the PTE with special | |
2626 | * device-exclusive PFN swap entry, preventing access through the process | |
2627 | * page tables. The function will return with the folio locked and referenced. | |
2628 | * | |
2629 | * On fault, the device-exclusive entries are replaced with the original PTE | |
2630 | * under folio lock, after calling MMU notifiers. | |
2631 | * | |
2632 | * Only anonymous non-hugetlb folios are supported and the VMA must have | |
2633 | * write permissions such that we can fault in the anonymous page writable | |
2634 | * in order to mark it exclusive. The caller must hold the mmap_lock in read | |
2635 | * mode. | |
2636 | * | |
2637 | * A driver using this to program access from a device must use a mmu notifier | |
2638 | * critical section to hold a device specific lock during programming. Once | |
2639 | * programming is complete it should drop the folio lock and reference after | |
2640 | * which point CPU access to the page will revoke the exclusive access. | |
2641 | * | |
2642 | * Notes: | |
2643 | * #. This function always operates on individual PTEs mapping individual | |
2644 | * pages. PMD-sized THPs are first remapped to be mapped by PTEs before | |
2645 | * the conversion happens on a single PTE corresponding to @addr. | |
2646 | * #. While concurrent access through the process page tables is prevented, | |
2647 | * concurrent access through other page references (e.g., earlier GUP | |
2648 | * invocation) is not handled and not supported. | |
2649 | * #. device-exclusive entries are considered "clean" and "old" by core-mm. | |
2650 | * Device drivers must update the folio state when informed by MMU | |
2651 | * notifiers. | |
2652 | * | |
2653 | * Returns: pointer to mapped page on success, otherwise a negative error. | |
2654 | */ | |
2655 | struct page *make_device_exclusive(struct mm_struct *mm, unsigned long addr, | |
2656 | void *owner, struct folio **foliop) | |
2657 | { | |
2658 | struct mmu_notifier_range range; | |
2659 | struct folio *folio, *fw_folio; | |
2660 | struct vm_area_struct *vma; | |
2661 | struct folio_walk fw; | |
2662 | struct page *page; | |
2663 | swp_entry_t entry; | |
2664 | pte_t swp_pte; | |
2665 | int ret; | |
2666 | ||
2667 | mmap_assert_locked(mm); | |
2668 | addr = PAGE_ALIGN_DOWN(addr); | |
2669 | ||
2670 | /* | |
2671 | * Fault in the page writable and try to lock it; note that if the | |
2672 | * address would already be marked for exclusive use by a device, | |
2673 | * the GUP call would undo that first by triggering a fault. | |
2674 | * | |
2675 | * If any other device would already map this page exclusively, the | |
2676 | * fault will trigger a conversion to an ordinary | |
2677 | * (non-device-exclusive) PTE and issue a MMU_NOTIFY_EXCLUSIVE. | |
2678 | */ | |
2679 | retry: | |
2680 | page = get_user_page_vma_remote(mm, addr, | |
2681 | FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD, | |
2682 | &vma); | |
2683 | if (IS_ERR(page)) | |
2684 | return page; | |
2685 | folio = page_folio(page); | |
2686 | ||
2687 | if (!folio_test_anon(folio) || folio_test_hugetlb(folio)) { | |
2688 | folio_put(folio); | |
2689 | return ERR_PTR(-EOPNOTSUPP); | |
2690 | } | |
2691 | ||
2692 | ret = folio_lock_killable(folio); | |
2693 | if (ret) { | |
2694 | folio_put(folio); | |
2695 | return ERR_PTR(ret); | |
2696 | } | |
2697 | ||
2698 | /* | |
2699 | * Inform secondary MMUs that we are going to convert this PTE to | |
2700 | * device-exclusive, such that they unmap it now. Note that the | |
2701 | * caller must filter this event out to prevent livelocks. | |
2702 | */ | |
2703 | mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, | |
2704 | mm, addr, addr + PAGE_SIZE, owner); | |
2705 | mmu_notifier_invalidate_range_start(&range); | |
2706 | ||
2707 | /* | |
2708 | * Let's do a second walk and make sure we still find the same page | |
2709 | * mapped writable. Note that any page of an anonymous folio can | |
2710 | * only be mapped writable using exactly one PTE ("exclusive"), so | |
2711 | * there cannot be other mappings. | |
2712 | */ | |
2713 | fw_folio = folio_walk_start(&fw, vma, addr, 0); | |
2714 | if (fw_folio != folio || fw.page != page || | |
2715 | fw.level != FW_LEVEL_PTE || !pte_write(fw.pte)) { | |
2716 | if (fw_folio) | |
2717 | folio_walk_end(&fw, vma); | |
2718 | mmu_notifier_invalidate_range_end(&range); | |
2719 | folio_unlock(folio); | |
2720 | folio_put(folio); | |
2721 | goto retry; | |
2722 | } | |
2723 | ||
2724 | /* Nuke the page table entry so we get the uptodate dirty bit. */ | |
2725 | flush_cache_page(vma, addr, page_to_pfn(page)); | |
2726 | fw.pte = ptep_clear_flush(vma, addr, fw.ptep); | |
2727 | ||
2728 | /* Set the dirty flag on the folio now the PTE is gone. */ | |
2729 | if (pte_dirty(fw.pte)) | |
2730 | folio_mark_dirty(folio); | |
2731 | ||
2732 | /* | |
2733 | * Store the pfn of the page in a special device-exclusive PFN swap PTE. | |
2734 | * do_swap_page() will trigger the conversion back while holding the | |
2735 | * folio lock. | |
2736 | */ | |
2737 | entry = make_device_exclusive_entry(page_to_pfn(page)); | |
2738 | swp_pte = swp_entry_to_pte(entry); | |
2739 | if (pte_soft_dirty(fw.pte)) | |
2740 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2741 | /* The pte is writable, uffd-wp does not apply. */ | |
2742 | set_pte_at(mm, addr, fw.ptep, swp_pte); | |
2743 | ||
2744 | folio_walk_end(&fw, vma); | |
2745 | mmu_notifier_invalidate_range_end(&range); | |
2746 | *foliop = folio; | |
2747 | return page; | |
2748 | } | |
2749 | EXPORT_SYMBOL_GPL(make_device_exclusive); | |
2750 | #endif | |
2751 | ||
2752 | void __put_anon_vma(struct anon_vma *anon_vma) | |
2753 | { | |
2754 | struct anon_vma *root = anon_vma->root; | |
2755 | ||
2756 | anon_vma_free(anon_vma); | |
2757 | if (root != anon_vma && atomic_dec_and_test(&root->refcount)) | |
2758 | anon_vma_free(root); | |
2759 | } | |
2760 | ||
2761 | static struct anon_vma *rmap_walk_anon_lock(const struct folio *folio, | |
2762 | struct rmap_walk_control *rwc) | |
2763 | { | |
2764 | struct anon_vma *anon_vma; | |
2765 | ||
2766 | if (rwc->anon_lock) | |
2767 | return rwc->anon_lock(folio, rwc); | |
2768 | ||
2769 | /* | |
2770 | * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read() | |
2771 | * because that depends on page_mapped(); but not all its usages | |
2772 | * are holding mmap_lock. Users without mmap_lock are required to | |
2773 | * take a reference count to prevent the anon_vma disappearing | |
2774 | */ | |
2775 | anon_vma = folio_anon_vma(folio); | |
2776 | if (!anon_vma) | |
2777 | return NULL; | |
2778 | ||
2779 | if (anon_vma_trylock_read(anon_vma)) | |
2780 | goto out; | |
2781 | ||
2782 | if (rwc->try_lock) { | |
2783 | anon_vma = NULL; | |
2784 | rwc->contended = true; | |
2785 | goto out; | |
2786 | } | |
2787 | ||
2788 | anon_vma_lock_read(anon_vma); | |
2789 | out: | |
2790 | return anon_vma; | |
2791 | } | |
2792 | ||
2793 | /* | |
2794 | * rmap_walk_anon - do something to anonymous page using the object-based | |
2795 | * rmap method | |
2796 | * @folio: the folio to be handled | |
2797 | * @rwc: control variable according to each walk type | |
2798 | * @locked: caller holds relevant rmap lock | |
2799 | * | |
2800 | * Find all the mappings of a folio using the mapping pointer and the vma | |
2801 | * chains contained in the anon_vma struct it points to. | |
2802 | */ | |
2803 | static void rmap_walk_anon(struct folio *folio, | |
2804 | struct rmap_walk_control *rwc, bool locked) | |
2805 | { | |
2806 | struct anon_vma *anon_vma; | |
2807 | pgoff_t pgoff_start, pgoff_end; | |
2808 | struct anon_vma_chain *avc; | |
2809 | ||
2810 | if (locked) { | |
2811 | anon_vma = folio_anon_vma(folio); | |
2812 | /* anon_vma disappear under us? */ | |
2813 | VM_BUG_ON_FOLIO(!anon_vma, folio); | |
2814 | } else { | |
2815 | anon_vma = rmap_walk_anon_lock(folio, rwc); | |
2816 | } | |
2817 | if (!anon_vma) | |
2818 | return; | |
2819 | ||
2820 | pgoff_start = folio_pgoff(folio); | |
2821 | pgoff_end = pgoff_start + folio_nr_pages(folio) - 1; | |
2822 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, | |
2823 | pgoff_start, pgoff_end) { | |
2824 | struct vm_area_struct *vma = avc->vma; | |
2825 | unsigned long address = vma_address(vma, pgoff_start, | |
2826 | folio_nr_pages(folio)); | |
2827 | ||
2828 | VM_BUG_ON_VMA(address == -EFAULT, vma); | |
2829 | cond_resched(); | |
2830 | ||
2831 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) | |
2832 | continue; | |
2833 | ||
2834 | if (!rwc->rmap_one(folio, vma, address, rwc->arg)) | |
2835 | break; | |
2836 | if (rwc->done && rwc->done(folio)) | |
2837 | break; | |
2838 | } | |
2839 | ||
2840 | if (!locked) | |
2841 | anon_vma_unlock_read(anon_vma); | |
2842 | } | |
2843 | ||
2844 | /** | |
2845 | * __rmap_walk_file() - Traverse the reverse mapping for a file-backed mapping | |
2846 | * of a page mapped within a specified page cache object at a specified offset. | |
2847 | * | |
2848 | * @folio: Either the folio whose mappings to traverse, or if NULL, | |
2849 | * the callbacks specified in @rwc will be configured such | |
2850 | * as to be able to look up mappings correctly. | |
2851 | * @mapping: The page cache object whose mapping VMAs we intend to | |
2852 | * traverse. If @folio is non-NULL, this should be equal to | |
2853 | * folio_mapping(folio). | |
2854 | * @pgoff_start: The offset within @mapping of the page which we are | |
2855 | * looking up. If @folio is non-NULL, this should be equal | |
2856 | * to folio_pgoff(folio). | |
2857 | * @nr_pages: The number of pages mapped by the mapping. If @folio is | |
2858 | * non-NULL, this should be equal to folio_nr_pages(folio). | |
2859 | * @rwc: The reverse mapping walk control object describing how | |
2860 | * the traversal should proceed. | |
2861 | * @locked: Is the @mapping already locked? If not, we acquire the | |
2862 | * lock. | |
2863 | */ | |
2864 | static void __rmap_walk_file(struct folio *folio, struct address_space *mapping, | |
2865 | pgoff_t pgoff_start, unsigned long nr_pages, | |
2866 | struct rmap_walk_control *rwc, bool locked) | |
2867 | { | |
2868 | pgoff_t pgoff_end = pgoff_start + nr_pages - 1; | |
2869 | struct vm_area_struct *vma; | |
2870 | ||
2871 | VM_WARN_ON_FOLIO(folio && mapping != folio_mapping(folio), folio); | |
2872 | VM_WARN_ON_FOLIO(folio && pgoff_start != folio_pgoff(folio), folio); | |
2873 | VM_WARN_ON_FOLIO(folio && nr_pages != folio_nr_pages(folio), folio); | |
2874 | ||
2875 | if (!locked) { | |
2876 | if (i_mmap_trylock_read(mapping)) | |
2877 | goto lookup; | |
2878 | ||
2879 | if (rwc->try_lock) { | |
2880 | rwc->contended = true; | |
2881 | return; | |
2882 | } | |
2883 | ||
2884 | i_mmap_lock_read(mapping); | |
2885 | } | |
2886 | lookup: | |
2887 | vma_interval_tree_foreach(vma, &mapping->i_mmap, | |
2888 | pgoff_start, pgoff_end) { | |
2889 | unsigned long address = vma_address(vma, pgoff_start, nr_pages); | |
2890 | ||
2891 | VM_BUG_ON_VMA(address == -EFAULT, vma); | |
2892 | cond_resched(); | |
2893 | ||
2894 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) | |
2895 | continue; | |
2896 | ||
2897 | if (!rwc->rmap_one(folio, vma, address, rwc->arg)) | |
2898 | goto done; | |
2899 | if (rwc->done && rwc->done(folio)) | |
2900 | goto done; | |
2901 | } | |
2902 | done: | |
2903 | if (!locked) | |
2904 | i_mmap_unlock_read(mapping); | |
2905 | } | |
2906 | ||
2907 | /* | |
2908 | * rmap_walk_file - do something to file page using the object-based rmap method | |
2909 | * @folio: the folio to be handled | |
2910 | * @rwc: control variable according to each walk type | |
2911 | * @locked: caller holds relevant rmap lock | |
2912 | * | |
2913 | * Find all the mappings of a folio using the mapping pointer and the vma chains | |
2914 | * contained in the address_space struct it points to. | |
2915 | */ | |
2916 | static void rmap_walk_file(struct folio *folio, | |
2917 | struct rmap_walk_control *rwc, bool locked) | |
2918 | { | |
2919 | /* | |
2920 | * The folio lock not only makes sure that folio->mapping cannot | |
2921 | * suddenly be NULLified by truncation, it makes sure that the structure | |
2922 | * at mapping cannot be freed and reused yet, so we can safely take | |
2923 | * mapping->i_mmap_rwsem. | |
2924 | */ | |
2925 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); | |
2926 | ||
2927 | if (!folio->mapping) | |
2928 | return; | |
2929 | ||
2930 | __rmap_walk_file(folio, folio->mapping, folio->index, | |
2931 | folio_nr_pages(folio), rwc, locked); | |
2932 | } | |
2933 | ||
2934 | void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc) | |
2935 | { | |
2936 | if (unlikely(folio_test_ksm(folio))) | |
2937 | rmap_walk_ksm(folio, rwc); | |
2938 | else if (folio_test_anon(folio)) | |
2939 | rmap_walk_anon(folio, rwc, false); | |
2940 | else | |
2941 | rmap_walk_file(folio, rwc, false); | |
2942 | } | |
2943 | ||
2944 | /* Like rmap_walk, but caller holds relevant rmap lock */ | |
2945 | void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc) | |
2946 | { | |
2947 | /* no ksm support for now */ | |
2948 | VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio); | |
2949 | if (folio_test_anon(folio)) | |
2950 | rmap_walk_anon(folio, rwc, true); | |
2951 | else | |
2952 | rmap_walk_file(folio, rwc, true); | |
2953 | } | |
2954 | ||
2955 | #ifdef CONFIG_HUGETLB_PAGE | |
2956 | /* | |
2957 | * The following two functions are for anonymous (private mapped) hugepages. | |
2958 | * Unlike common anonymous pages, anonymous hugepages have no accounting code | |
2959 | * and no lru code, because we handle hugepages differently from common pages. | |
2960 | */ | |
2961 | void hugetlb_add_anon_rmap(struct folio *folio, struct vm_area_struct *vma, | |
2962 | unsigned long address, rmap_t flags) | |
2963 | { | |
2964 | VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio); | |
2965 | VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio); | |
2966 | ||
2967 | atomic_inc(&folio->_entire_mapcount); | |
2968 | atomic_inc(&folio->_large_mapcount); | |
2969 | if (flags & RMAP_EXCLUSIVE) | |
2970 | SetPageAnonExclusive(&folio->page); | |
2971 | VM_WARN_ON_FOLIO(folio_entire_mapcount(folio) > 1 && | |
2972 | PageAnonExclusive(&folio->page), folio); | |
2973 | } | |
2974 | ||
2975 | void hugetlb_add_new_anon_rmap(struct folio *folio, | |
2976 | struct vm_area_struct *vma, unsigned long address) | |
2977 | { | |
2978 | VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio); | |
2979 | ||
2980 | BUG_ON(address < vma->vm_start || address >= vma->vm_end); | |
2981 | /* increment count (starts at -1) */ | |
2982 | atomic_set(&folio->_entire_mapcount, 0); | |
2983 | atomic_set(&folio->_large_mapcount, 0); | |
2984 | folio_clear_hugetlb_restore_reserve(folio); | |
2985 | __folio_set_anon(folio, vma, address, true); | |
2986 | SetPageAnonExclusive(&folio->page); | |
2987 | } | |
2988 | #endif /* CONFIG_HUGETLB_PAGE */ |