1/* SPDX-License-Identifier: GPL-2.0 */
2#ifndef _LINUX_RMAP_H
3#define _LINUX_RMAP_H
4/*
5 * Declarations for Reverse Mapping functions in mm/rmap.c
6 */
7
8#include <linux/list.h>
9#include <linux/slab.h>
10#include <linux/mm.h>
11#include <linux/rwsem.h>
12#include <linux/memcontrol.h>
13#include <linux/highmem.h>
14#include <linux/pagemap.h>
15#include <linux/memremap.h>
16#include <linux/bit_spinlock.h>
17
18/*
19 * The anon_vma heads a list of private "related" vmas, to scan if
20 * an anonymous page pointing to this anon_vma needs to be unmapped:
21 * the vmas on the list will be related by forking, or by splitting.
22 *
23 * Since vmas come and go as they are split and merged (particularly
24 * in mprotect), the mapping field of an anonymous page cannot point
25 * directly to a vma: instead it points to an anon_vma, on whose list
26 * the related vmas can be easily linked or unlinked.
27 *
28 * After unlinking the last vma on the list, we must garbage collect
29 * the anon_vma object itself: we're guaranteed no page can be
30 * pointing to this anon_vma once its vma list is empty.
31 */
32struct anon_vma {
33 struct anon_vma *root; /* Root of this anon_vma tree */
34 struct rw_semaphore rwsem; /* W: modification, R: walking the list */
35 /*
36 * The refcount is taken on an anon_vma when there is no
37 * guarantee that the vma of page tables will exist for
38 * the duration of the operation. A caller that takes
39 * the reference is responsible for clearing up the
40 * anon_vma if they are the last user on release
41 */
42 atomic_t refcount;
43
44 /*
45 * Count of child anon_vmas. Equals to the count of all anon_vmas that
46 * have ->parent pointing to this one, including itself.
47 *
48 * This counter is used for making decision about reusing anon_vma
49 * instead of forking new one. See comments in function anon_vma_clone.
50 */
51 unsigned long num_children;
52 /* Count of VMAs whose ->anon_vma pointer points to this object. */
53 unsigned long num_active_vmas;
54
55 struct anon_vma *parent; /* Parent of this anon_vma */
56
57 /*
58 * NOTE: the LSB of the rb_root.rb_node is set by
59 * mm_take_all_locks() _after_ taking the above lock. So the
60 * rb_root must only be read/written after taking the above lock
61 * to be sure to see a valid next pointer. The LSB bit itself
62 * is serialized by a system wide lock only visible to
63 * mm_take_all_locks() (mm_all_locks_mutex).
64 */
65
66 /* Interval tree of private "related" vmas */
67 struct rb_root_cached rb_root;
68};
69
70/*
71 * The copy-on-write semantics of fork mean that an anon_vma
72 * can become associated with multiple processes. Furthermore,
73 * each child process will have its own anon_vma, where new
74 * pages for that process are instantiated.
75 *
76 * This structure allows us to find the anon_vmas associated
77 * with a VMA, or the VMAs associated with an anon_vma.
78 * The "same_vma" list contains the anon_vma_chains linking
79 * all the anon_vmas associated with this VMA.
80 * The "rb" field indexes on an interval tree the anon_vma_chains
81 * which link all the VMAs associated with this anon_vma.
82 */
83struct anon_vma_chain {
84 struct vm_area_struct *vma;
85 struct anon_vma *anon_vma;
86 struct list_head same_vma; /* locked by mmap_lock & page_table_lock */
87 struct rb_node rb; /* locked by anon_vma->rwsem */
88 unsigned long rb_subtree_last;
89#ifdef CONFIG_DEBUG_VM_RB
90 unsigned long cached_vma_start, cached_vma_last;
91#endif
92};
93
94enum ttu_flags {
95 TTU_SPLIT_HUGE_PMD = 0x4, /* split huge PMD if any */
96 TTU_IGNORE_MLOCK = 0x8, /* ignore mlock */
97 TTU_SYNC = 0x10, /* avoid racy checks with PVMW_SYNC */
98 TTU_HWPOISON = 0x20, /* do convert pte to hwpoison entry */
99 TTU_BATCH_FLUSH = 0x40, /* Batch TLB flushes where possible
100 * and caller guarantees they will
101 * do a final flush if necessary */
102 TTU_RMAP_LOCKED = 0x80, /* do not grab rmap lock:
103 * caller holds it */
104};
105
106#ifdef CONFIG_MMU
107static inline void get_anon_vma(struct anon_vma *anon_vma)
108{
109 atomic_inc(v: &anon_vma->refcount);
110}
111
112void __put_anon_vma(struct anon_vma *anon_vma);
113
114static inline void put_anon_vma(struct anon_vma *anon_vma)
115{
116 if (atomic_dec_and_test(v: &anon_vma->refcount))
117 __put_anon_vma(anon_vma);
118}
119
120static inline void anon_vma_lock_write(struct anon_vma *anon_vma)
121{
122 down_write(sem: &anon_vma->root->rwsem);
123}
124
125static inline int anon_vma_trylock_write(struct anon_vma *anon_vma)
126{
127 return down_write_trylock(sem: &anon_vma->root->rwsem);
128}
129
130static inline void anon_vma_unlock_write(struct anon_vma *anon_vma)
131{
132 up_write(sem: &anon_vma->root->rwsem);
133}
134
135static inline void anon_vma_lock_read(struct anon_vma *anon_vma)
136{
137 down_read(sem: &anon_vma->root->rwsem);
138}
139
140static inline int anon_vma_trylock_read(struct anon_vma *anon_vma)
141{
142 return down_read_trylock(sem: &anon_vma->root->rwsem);
143}
144
145static inline void anon_vma_unlock_read(struct anon_vma *anon_vma)
146{
147 up_read(sem: &anon_vma->root->rwsem);
148}
149
150
151/*
152 * anon_vma helper functions.
153 */
154void anon_vma_init(void); /* create anon_vma_cachep */
155int __anon_vma_prepare(struct vm_area_struct *);
156void unlink_anon_vmas(struct vm_area_struct *);
157int anon_vma_clone(struct vm_area_struct *, struct vm_area_struct *);
158int anon_vma_fork(struct vm_area_struct *, struct vm_area_struct *);
159
160static inline int anon_vma_prepare(struct vm_area_struct *vma)
161{
162 if (likely(vma->anon_vma))
163 return 0;
164
165 return __anon_vma_prepare(vma);
166}
167
168static inline void anon_vma_merge(struct vm_area_struct *vma,
169 struct vm_area_struct *next)
170{
171 VM_BUG_ON_VMA(vma->anon_vma != next->anon_vma, vma);
172 unlink_anon_vmas(next);
173}
174
175struct anon_vma *folio_get_anon_vma(const struct folio *folio);
176
177#ifdef CONFIG_MM_ID
178static __always_inline void folio_lock_large_mapcount(struct folio *folio)
179{
180 bit_spin_lock(FOLIO_MM_IDS_LOCK_BITNUM, addr: &folio->_mm_ids);
181}
182
183static __always_inline void folio_unlock_large_mapcount(struct folio *folio)
184{
185 __bit_spin_unlock(FOLIO_MM_IDS_LOCK_BITNUM, addr: &folio->_mm_ids);
186}
187
188static inline unsigned int folio_mm_id(const struct folio *folio, int idx)
189{
190 VM_WARN_ON_ONCE(idx != 0 && idx != 1);
191 return folio->_mm_id[idx] & MM_ID_MASK;
192}
193
194static inline void folio_set_mm_id(struct folio *folio, int idx, mm_id_t id)
195{
196 VM_WARN_ON_ONCE(idx != 0 && idx != 1);
197 folio->_mm_id[idx] &= ~MM_ID_MASK;
198 folio->_mm_id[idx] |= id;
199}
200
201static inline void __folio_large_mapcount_sanity_checks(const struct folio *folio,
202 int diff, mm_id_t mm_id)
203{
204 VM_WARN_ON_ONCE(!folio_test_large(folio) || folio_test_hugetlb(folio));
205 VM_WARN_ON_ONCE(diff <= 0);
206 VM_WARN_ON_ONCE(mm_id < MM_ID_MIN || mm_id > MM_ID_MAX);
207
208 /*
209 * Make sure we can detect at least one complete PTE mapping of the
210 * folio in a single MM as "exclusively mapped". This is primarily
211 * a check on 32bit, where we currently reduce the size of the per-MM
212 * mapcount to a short.
213 */
214 VM_WARN_ON_ONCE(diff > folio_large_nr_pages(folio));
215 VM_WARN_ON_ONCE(folio_large_nr_pages(folio) - 1 > MM_ID_MAPCOUNT_MAX);
216
217 VM_WARN_ON_ONCE(folio_mm_id(folio, 0) == MM_ID_DUMMY &&
218 folio->_mm_id_mapcount[0] != -1);
219 VM_WARN_ON_ONCE(folio_mm_id(folio, 0) != MM_ID_DUMMY &&
220 folio->_mm_id_mapcount[0] < 0);
221 VM_WARN_ON_ONCE(folio_mm_id(folio, 1) == MM_ID_DUMMY &&
222 folio->_mm_id_mapcount[1] != -1);
223 VM_WARN_ON_ONCE(folio_mm_id(folio, 1) != MM_ID_DUMMY &&
224 folio->_mm_id_mapcount[1] < 0);
225 VM_WARN_ON_ONCE(!folio_mapped(folio) &&
226 test_bit(FOLIO_MM_IDS_SHARED_BITNUM, &folio->_mm_ids));
227}
228
229static __always_inline void folio_set_large_mapcount(struct folio *folio,
230 int mapcount, struct vm_area_struct *vma)
231{
232 __folio_large_mapcount_sanity_checks(folio, diff: mapcount, mm_id: vma->vm_mm->mm_id);
233
234 VM_WARN_ON_ONCE(folio_mm_id(folio, 0) != MM_ID_DUMMY);
235 VM_WARN_ON_ONCE(folio_mm_id(folio, 1) != MM_ID_DUMMY);
236
237 /* Note: mapcounts start at -1. */
238 atomic_set(v: &folio->_large_mapcount, i: mapcount - 1);
239 folio->_mm_id_mapcount[0] = mapcount - 1;
240 folio_set_mm_id(folio, idx: 0, id: vma->vm_mm->mm_id);
241}
242
243static __always_inline int folio_add_return_large_mapcount(struct folio *folio,
244 int diff, struct vm_area_struct *vma)
245{
246 const mm_id_t mm_id = vma->vm_mm->mm_id;
247 int new_mapcount_val;
248
249 folio_lock_large_mapcount(folio);
250 __folio_large_mapcount_sanity_checks(folio, diff, mm_id);
251
252 new_mapcount_val = atomic_read(v: &folio->_large_mapcount) + diff;
253 atomic_set(v: &folio->_large_mapcount, i: new_mapcount_val);
254
255 /*
256 * If a folio is mapped more than once into an MM on 32bit, we
257 * can in theory overflow the per-MM mapcount (although only for
258 * fairly large folios), turning it negative. In that case, just
259 * free up the slot and mark the folio "mapped shared", otherwise
260 * we might be in trouble when unmapping pages later.
261 */
262 if (folio_mm_id(folio, idx: 0) == mm_id) {
263 folio->_mm_id_mapcount[0] += diff;
264 if (!IS_ENABLED(CONFIG_64BIT) && unlikely(folio->_mm_id_mapcount[0] < 0)) {
265 folio->_mm_id_mapcount[0] = -1;
266 folio_set_mm_id(folio, idx: 0, MM_ID_DUMMY);
267 folio->_mm_ids |= FOLIO_MM_IDS_SHARED_BIT;
268 }
269 } else if (folio_mm_id(folio, idx: 1) == mm_id) {
270 folio->_mm_id_mapcount[1] += diff;
271 if (!IS_ENABLED(CONFIG_64BIT) && unlikely(folio->_mm_id_mapcount[1] < 0)) {
272 folio->_mm_id_mapcount[1] = -1;
273 folio_set_mm_id(folio, idx: 1, MM_ID_DUMMY);
274 folio->_mm_ids |= FOLIO_MM_IDS_SHARED_BIT;
275 }
276 } else if (folio_mm_id(folio, idx: 0) == MM_ID_DUMMY) {
277 folio_set_mm_id(folio, idx: 0, id: mm_id);
278 folio->_mm_id_mapcount[0] = diff - 1;
279 /* We might have other mappings already. */
280 if (new_mapcount_val != diff - 1)
281 folio->_mm_ids |= FOLIO_MM_IDS_SHARED_BIT;
282 } else if (folio_mm_id(folio, idx: 1) == MM_ID_DUMMY) {
283 folio_set_mm_id(folio, idx: 1, id: mm_id);
284 folio->_mm_id_mapcount[1] = diff - 1;
285 /* Slot 0 certainly has mappings as well. */
286 folio->_mm_ids |= FOLIO_MM_IDS_SHARED_BIT;
287 }
288 folio_unlock_large_mapcount(folio);
289 return new_mapcount_val + 1;
290}
291#define folio_add_large_mapcount folio_add_return_large_mapcount
292
293static __always_inline int folio_sub_return_large_mapcount(struct folio *folio,
294 int diff, struct vm_area_struct *vma)
295{
296 const mm_id_t mm_id = vma->vm_mm->mm_id;
297 int new_mapcount_val;
298
299 folio_lock_large_mapcount(folio);
300 __folio_large_mapcount_sanity_checks(folio, diff, mm_id);
301
302 new_mapcount_val = atomic_read(v: &folio->_large_mapcount) - diff;
303 atomic_set(v: &folio->_large_mapcount, i: new_mapcount_val);
304
305 /*
306 * There are valid corner cases where we might underflow a per-MM
307 * mapcount (some mappings added when no slot was free, some mappings
308 * added once a slot was free), so we always set it to -1 once we go
309 * negative.
310 */
311 if (folio_mm_id(folio, idx: 0) == mm_id) {
312 folio->_mm_id_mapcount[0] -= diff;
313 if (folio->_mm_id_mapcount[0] >= 0)
314 goto out;
315 folio->_mm_id_mapcount[0] = -1;
316 folio_set_mm_id(folio, idx: 0, MM_ID_DUMMY);
317 } else if (folio_mm_id(folio, idx: 1) == mm_id) {
318 folio->_mm_id_mapcount[1] -= diff;
319 if (folio->_mm_id_mapcount[1] >= 0)
320 goto out;
321 folio->_mm_id_mapcount[1] = -1;
322 folio_set_mm_id(folio, idx: 1, MM_ID_DUMMY);
323 }
324
325 /*
326 * If one MM slot owns all mappings, the folio is mapped exclusively.
327 * Note that if the folio is now unmapped (new_mapcount_val == -1), both
328 * slots must be free (mapcount == -1), and we'll also mark it as
329 * exclusive.
330 */
331 if (folio->_mm_id_mapcount[0] == new_mapcount_val ||
332 folio->_mm_id_mapcount[1] == new_mapcount_val)
333 folio->_mm_ids &= ~FOLIO_MM_IDS_SHARED_BIT;
334out:
335 folio_unlock_large_mapcount(folio);
336 return new_mapcount_val + 1;
337}
338#define folio_sub_large_mapcount folio_sub_return_large_mapcount
339#else /* !CONFIG_MM_ID */
340/*
341 * See __folio_rmap_sanity_checks(), we might map large folios even without
342 * CONFIG_TRANSPARENT_HUGEPAGE. We'll keep that working for now.
343 */
344static inline void folio_set_large_mapcount(struct folio *folio, int mapcount,
345 struct vm_area_struct *vma)
346{
347 /* Note: mapcounts start at -1. */
348 atomic_set(&folio->_large_mapcount, mapcount - 1);
349}
350
351static inline void folio_add_large_mapcount(struct folio *folio,
352 int diff, struct vm_area_struct *vma)
353{
354 atomic_add(diff, &folio->_large_mapcount);
355}
356
357static inline int folio_add_return_large_mapcount(struct folio *folio,
358 int diff, struct vm_area_struct *vma)
359{
360 BUILD_BUG();
361}
362
363static inline void folio_sub_large_mapcount(struct folio *folio,
364 int diff, struct vm_area_struct *vma)
365{
366 atomic_sub(diff, &folio->_large_mapcount);
367}
368
369static inline int folio_sub_return_large_mapcount(struct folio *folio,
370 int diff, struct vm_area_struct *vma)
371{
372 BUILD_BUG();
373}
374#endif /* CONFIG_MM_ID */
375
376#define folio_inc_large_mapcount(folio, vma) \
377 folio_add_large_mapcount(folio, 1, vma)
378#define folio_inc_return_large_mapcount(folio, vma) \
379 folio_add_return_large_mapcount(folio, 1, vma)
380#define folio_dec_large_mapcount(folio, vma) \
381 folio_sub_large_mapcount(folio, 1, vma)
382#define folio_dec_return_large_mapcount(folio, vma) \
383 folio_sub_return_large_mapcount(folio, 1, vma)
384
385/* RMAP flags, currently only relevant for some anon rmap operations. */
386typedef int __bitwise rmap_t;
387
388/*
389 * No special request: A mapped anonymous (sub)page is possibly shared between
390 * processes.
391 */
392#define RMAP_NONE ((__force rmap_t)0)
393
394/* The anonymous (sub)page is exclusive to a single process. */
395#define RMAP_EXCLUSIVE ((__force rmap_t)BIT(0))
396
397/*
398 * Internally, we're using an enum to specify the granularity. We make the
399 * compiler emit specialized code for each granularity.
400 */
401enum rmap_level {
402 RMAP_LEVEL_PTE = 0,
403 RMAP_LEVEL_PMD,
404 RMAP_LEVEL_PUD,
405};
406
407static inline void __folio_rmap_sanity_checks(const struct folio *folio,
408 const struct page *page, int nr_pages, enum rmap_level level)
409{
410 /* hugetlb folios are handled separately. */
411 VM_WARN_ON_FOLIO(folio_test_hugetlb(folio), folio);
412
413 /* When (un)mapping zeropages, we should never touch ref+mapcount. */
414 VM_WARN_ON_FOLIO(is_zero_folio(folio), folio);
415
416 /*
417 * TODO: we get driver-allocated folios that have nothing to do with
418 * the rmap using vm_insert_page(); therefore, we cannot assume that
419 * folio_test_large_rmappable() holds for large folios. We should
420 * handle any desired mapcount+stats accounting for these folios in
421 * VM_MIXEDMAP VMAs separately, and then sanity-check here that
422 * we really only get rmappable folios.
423 */
424
425 VM_WARN_ON_ONCE(nr_pages <= 0);
426 VM_WARN_ON_FOLIO(page_folio(page) != folio, folio);
427 VM_WARN_ON_FOLIO(page_folio(page + nr_pages - 1) != folio, folio);
428
429 switch (level) {
430 case RMAP_LEVEL_PTE:
431 break;
432 case RMAP_LEVEL_PMD:
433 /*
434 * We don't support folios larger than a single PMD yet. So
435 * when RMAP_LEVEL_PMD is set, we assume that we are creating
436 * a single "entire" mapping of the folio.
437 */
438 VM_WARN_ON_FOLIO(folio_nr_pages(folio) != HPAGE_PMD_NR, folio);
439 VM_WARN_ON_FOLIO(nr_pages != HPAGE_PMD_NR, folio);
440 break;
441 case RMAP_LEVEL_PUD:
442 /*
443 * Assume that we are creating a single "entire" mapping of the
444 * folio.
445 */
446 VM_WARN_ON_FOLIO(folio_nr_pages(folio) != HPAGE_PUD_NR, folio);
447 VM_WARN_ON_FOLIO(nr_pages != HPAGE_PUD_NR, folio);
448 break;
449 default:
450 VM_WARN_ON_ONCE(true);
451 }
452}
453
454/*
455 * rmap interfaces called when adding or removing pte of page
456 */
457void folio_move_anon_rmap(struct folio *, struct vm_area_struct *);
458void folio_add_anon_rmap_ptes(struct folio *, struct page *, int nr_pages,
459 struct vm_area_struct *, unsigned long address, rmap_t flags);
460#define folio_add_anon_rmap_pte(folio, page, vma, address, flags) \
461 folio_add_anon_rmap_ptes(folio, page, 1, vma, address, flags)
462void folio_add_anon_rmap_pmd(struct folio *, struct page *,
463 struct vm_area_struct *, unsigned long address, rmap_t flags);
464void folio_add_new_anon_rmap(struct folio *, struct vm_area_struct *,
465 unsigned long address, rmap_t flags);
466void folio_add_file_rmap_ptes(struct folio *, struct page *, int nr_pages,
467 struct vm_area_struct *);
468#define folio_add_file_rmap_pte(folio, page, vma) \
469 folio_add_file_rmap_ptes(folio, page, 1, vma)
470void folio_add_file_rmap_pmd(struct folio *, struct page *,
471 struct vm_area_struct *);
472void folio_add_file_rmap_pud(struct folio *, struct page *,
473 struct vm_area_struct *);
474void folio_remove_rmap_ptes(struct folio *, struct page *, int nr_pages,
475 struct vm_area_struct *);
476#define folio_remove_rmap_pte(folio, page, vma) \
477 folio_remove_rmap_ptes(folio, page, 1, vma)
478void folio_remove_rmap_pmd(struct folio *, struct page *,
479 struct vm_area_struct *);
480void folio_remove_rmap_pud(struct folio *, struct page *,
481 struct vm_area_struct *);
482
483void hugetlb_add_anon_rmap(struct folio *, struct vm_area_struct *,
484 unsigned long address, rmap_t flags);
485void hugetlb_add_new_anon_rmap(struct folio *, struct vm_area_struct *,
486 unsigned long address);
487
488/* See folio_try_dup_anon_rmap_*() */
489static inline int hugetlb_try_dup_anon_rmap(struct folio *folio,
490 struct vm_area_struct *vma)
491{
492 VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio);
493 VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
494
495 if (PageAnonExclusive(page: &folio->page)) {
496 if (unlikely(folio_needs_cow_for_dma(vma, folio)))
497 return -EBUSY;
498 ClearPageAnonExclusive(page: &folio->page);
499 }
500 atomic_inc(v: &folio->_entire_mapcount);
501 atomic_inc(v: &folio->_large_mapcount);
502 return 0;
503}
504
505/* See folio_try_share_anon_rmap_*() */
506static inline int hugetlb_try_share_anon_rmap(struct folio *folio)
507{
508 VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio);
509 VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
510 VM_WARN_ON_FOLIO(!PageAnonExclusive(&folio->page), folio);
511
512 /* Paired with the memory barrier in try_grab_folio(). */
513 if (IS_ENABLED(CONFIG_HAVE_GUP_FAST))
514 smp_mb();
515
516 if (unlikely(folio_maybe_dma_pinned(folio)))
517 return -EBUSY;
518 ClearPageAnonExclusive(page: &folio->page);
519
520 /*
521 * This is conceptually a smp_wmb() paired with the smp_rmb() in
522 * gup_must_unshare().
523 */
524 if (IS_ENABLED(CONFIG_HAVE_GUP_FAST))
525 smp_mb__after_atomic();
526 return 0;
527}
528
529static inline void hugetlb_add_file_rmap(struct folio *folio)
530{
531 VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio);
532 VM_WARN_ON_FOLIO(folio_test_anon(folio), folio);
533
534 atomic_inc(v: &folio->_entire_mapcount);
535 atomic_inc(v: &folio->_large_mapcount);
536}
537
538static inline void hugetlb_remove_rmap(struct folio *folio)
539{
540 VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio);
541
542 atomic_dec(v: &folio->_entire_mapcount);
543 atomic_dec(v: &folio->_large_mapcount);
544}
545
546static __always_inline void __folio_dup_file_rmap(struct folio *folio,
547 struct page *page, int nr_pages, struct vm_area_struct *dst_vma,
548 enum rmap_level level)
549{
550 const int orig_nr_pages = nr_pages;
551
552 __folio_rmap_sanity_checks(folio, page, nr_pages, level);
553
554 switch (level) {
555 case RMAP_LEVEL_PTE:
556 if (!folio_test_large(folio)) {
557 atomic_inc(v: &folio->_mapcount);
558 break;
559 }
560
561 if (IS_ENABLED(CONFIG_PAGE_MAPCOUNT)) {
562 do {
563 atomic_inc(v: &page->_mapcount);
564 } while (page++, --nr_pages > 0);
565 }
566 folio_add_large_mapcount(folio, diff: orig_nr_pages, vma: dst_vma);
567 break;
568 case RMAP_LEVEL_PMD:
569 case RMAP_LEVEL_PUD:
570 atomic_inc(v: &folio->_entire_mapcount);
571 folio_inc_large_mapcount(folio, dst_vma);
572 break;
573 }
574}
575
576/**
577 * folio_dup_file_rmap_ptes - duplicate PTE mappings of a page range of a folio
578 * @folio: The folio to duplicate the mappings of
579 * @page: The first page to duplicate the mappings of
580 * @nr_pages: The number of pages of which the mapping will be duplicated
581 * @dst_vma: The destination vm area
582 *
583 * The page range of the folio is defined by [page, page + nr_pages)
584 *
585 * The caller needs to hold the page table lock.
586 */
587static inline void folio_dup_file_rmap_ptes(struct folio *folio,
588 struct page *page, int nr_pages, struct vm_area_struct *dst_vma)
589{
590 __folio_dup_file_rmap(folio, page, nr_pages, dst_vma, level: RMAP_LEVEL_PTE);
591}
592
593static __always_inline void folio_dup_file_rmap_pte(struct folio *folio,
594 struct page *page, struct vm_area_struct *dst_vma)
595{
596 __folio_dup_file_rmap(folio, page, nr_pages: 1, dst_vma, level: RMAP_LEVEL_PTE);
597}
598
599/**
600 * folio_dup_file_rmap_pmd - duplicate a PMD mapping of a page range of a folio
601 * @folio: The folio to duplicate the mapping of
602 * @page: The first page to duplicate the mapping of
603 * @dst_vma: The destination vm area
604 *
605 * The page range of the folio is defined by [page, page + HPAGE_PMD_NR)
606 *
607 * The caller needs to hold the page table lock.
608 */
609static inline void folio_dup_file_rmap_pmd(struct folio *folio,
610 struct page *page, struct vm_area_struct *dst_vma)
611{
612#ifdef CONFIG_TRANSPARENT_HUGEPAGE
613 __folio_dup_file_rmap(folio, page, HPAGE_PMD_NR, dst_vma, level: RMAP_LEVEL_PTE);
614#else
615 WARN_ON_ONCE(true);
616#endif
617}
618
619static __always_inline int __folio_try_dup_anon_rmap(struct folio *folio,
620 struct page *page, int nr_pages, struct vm_area_struct *dst_vma,
621 struct vm_area_struct *src_vma, enum rmap_level level)
622{
623 const int orig_nr_pages = nr_pages;
624 bool maybe_pinned;
625 int i;
626
627 VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
628 __folio_rmap_sanity_checks(folio, page, nr_pages, level);
629
630 /*
631 * If this folio may have been pinned by the parent process,
632 * don't allow to duplicate the mappings but instead require to e.g.,
633 * copy the subpage immediately for the child so that we'll always
634 * guarantee the pinned folio won't be randomly replaced in the
635 * future on write faults.
636 */
637 maybe_pinned = likely(!folio_is_device_private(folio)) &&
638 unlikely(folio_needs_cow_for_dma(src_vma, folio));
639
640 /*
641 * No need to check+clear for already shared PTEs/PMDs of the
642 * folio. But if any page is PageAnonExclusive, we must fallback to
643 * copying if the folio maybe pinned.
644 */
645 switch (level) {
646 case RMAP_LEVEL_PTE:
647 if (unlikely(maybe_pinned)) {
648 for (i = 0; i < nr_pages; i++)
649 if (PageAnonExclusive(page: page + i))
650 return -EBUSY;
651 }
652
653 if (!folio_test_large(folio)) {
654 if (PageAnonExclusive(page))
655 ClearPageAnonExclusive(page);
656 atomic_inc(v: &folio->_mapcount);
657 break;
658 }
659
660 do {
661 if (PageAnonExclusive(page))
662 ClearPageAnonExclusive(page);
663 if (IS_ENABLED(CONFIG_PAGE_MAPCOUNT))
664 atomic_inc(v: &page->_mapcount);
665 } while (page++, --nr_pages > 0);
666 folio_add_large_mapcount(folio, diff: orig_nr_pages, vma: dst_vma);
667 break;
668 case RMAP_LEVEL_PMD:
669 case RMAP_LEVEL_PUD:
670 if (PageAnonExclusive(page)) {
671 if (unlikely(maybe_pinned))
672 return -EBUSY;
673 ClearPageAnonExclusive(page);
674 }
675 atomic_inc(v: &folio->_entire_mapcount);
676 folio_inc_large_mapcount(folio, dst_vma);
677 break;
678 }
679 return 0;
680}
681
682/**
683 * folio_try_dup_anon_rmap_ptes - try duplicating PTE mappings of a page range
684 * of a folio
685 * @folio: The folio to duplicate the mappings of
686 * @page: The first page to duplicate the mappings of
687 * @nr_pages: The number of pages of which the mapping will be duplicated
688 * @dst_vma: The destination vm area
689 * @src_vma: The vm area from which the mappings are duplicated
690 *
691 * The page range of the folio is defined by [page, page + nr_pages)
692 *
693 * The caller needs to hold the page table lock and the
694 * vma->vma_mm->write_protect_seq.
695 *
696 * Duplicating the mappings can only fail if the folio may be pinned; device
697 * private folios cannot get pinned and consequently this function cannot fail
698 * for them.
699 *
700 * If duplicating the mappings succeeded, the duplicated PTEs have to be R/O in
701 * the parent and the child. They must *not* be writable after this call
702 * succeeded.
703 *
704 * Returns 0 if duplicating the mappings succeeded. Returns -EBUSY otherwise.
705 */
706static inline int folio_try_dup_anon_rmap_ptes(struct folio *folio,
707 struct page *page, int nr_pages, struct vm_area_struct *dst_vma,
708 struct vm_area_struct *src_vma)
709{
710 return __folio_try_dup_anon_rmap(folio, page, nr_pages, dst_vma,
711 src_vma, level: RMAP_LEVEL_PTE);
712}
713
714static __always_inline int folio_try_dup_anon_rmap_pte(struct folio *folio,
715 struct page *page, struct vm_area_struct *dst_vma,
716 struct vm_area_struct *src_vma)
717{
718 return __folio_try_dup_anon_rmap(folio, page, nr_pages: 1, dst_vma, src_vma,
719 level: RMAP_LEVEL_PTE);
720}
721
722/**
723 * folio_try_dup_anon_rmap_pmd - try duplicating a PMD mapping of a page range
724 * of a folio
725 * @folio: The folio to duplicate the mapping of
726 * @page: The first page to duplicate the mapping of
727 * @dst_vma: The destination vm area
728 * @src_vma: The vm area from which the mapping is duplicated
729 *
730 * The page range of the folio is defined by [page, page + HPAGE_PMD_NR)
731 *
732 * The caller needs to hold the page table lock and the
733 * vma->vma_mm->write_protect_seq.
734 *
735 * Duplicating the mapping can only fail if the folio may be pinned; device
736 * private folios cannot get pinned and consequently this function cannot fail
737 * for them.
738 *
739 * If duplicating the mapping succeeds, the duplicated PMD has to be R/O in
740 * the parent and the child. They must *not* be writable after this call
741 * succeeded.
742 *
743 * Returns 0 if duplicating the mapping succeeded. Returns -EBUSY otherwise.
744 */
745static inline int folio_try_dup_anon_rmap_pmd(struct folio *folio,
746 struct page *page, struct vm_area_struct *dst_vma,
747 struct vm_area_struct *src_vma)
748{
749#ifdef CONFIG_TRANSPARENT_HUGEPAGE
750 return __folio_try_dup_anon_rmap(folio, page, HPAGE_PMD_NR, dst_vma,
751 src_vma, level: RMAP_LEVEL_PMD);
752#else
753 WARN_ON_ONCE(true);
754 return -EBUSY;
755#endif
756}
757
758static __always_inline int __folio_try_share_anon_rmap(struct folio *folio,
759 struct page *page, int nr_pages, enum rmap_level level)
760{
761 VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
762 VM_WARN_ON_FOLIO(!PageAnonExclusive(page), folio);
763 __folio_rmap_sanity_checks(folio, page, nr_pages, level);
764
765 /* device private folios cannot get pinned via GUP. */
766 if (unlikely(folio_is_device_private(folio))) {
767 ClearPageAnonExclusive(page);
768 return 0;
769 }
770
771 /*
772 * We have to make sure that when we clear PageAnonExclusive, that
773 * the page is not pinned and that concurrent GUP-fast won't succeed in
774 * concurrently pinning the page.
775 *
776 * Conceptually, PageAnonExclusive clearing consists of:
777 * (A1) Clear PTE
778 * (A2) Check if the page is pinned; back off if so.
779 * (A3) Clear PageAnonExclusive
780 * (A4) Restore PTE (optional, but certainly not writable)
781 *
782 * When clearing PageAnonExclusive, we cannot possibly map the page
783 * writable again, because anon pages that may be shared must never
784 * be writable. So in any case, if the PTE was writable it cannot
785 * be writable anymore afterwards and there would be a PTE change. Only
786 * if the PTE wasn't writable, there might not be a PTE change.
787 *
788 * Conceptually, GUP-fast pinning of an anon page consists of:
789 * (B1) Read the PTE
790 * (B2) FOLL_WRITE: check if the PTE is not writable; back off if so.
791 * (B3) Pin the mapped page
792 * (B4) Check if the PTE changed by re-reading it; back off if so.
793 * (B5) If the original PTE is not writable, check if
794 * PageAnonExclusive is not set; back off if so.
795 *
796 * If the PTE was writable, we only have to make sure that GUP-fast
797 * observes a PTE change and properly backs off.
798 *
799 * If the PTE was not writable, we have to make sure that GUP-fast either
800 * detects a (temporary) PTE change or that PageAnonExclusive is cleared
801 * and properly backs off.
802 *
803 * Consequently, when clearing PageAnonExclusive(), we have to make
804 * sure that (A1), (A2)/(A3) and (A4) happen in the right memory
805 * order. In GUP-fast pinning code, we have to make sure that (B3),(B4)
806 * and (B5) happen in the right memory order.
807 *
808 * We assume that there might not be a memory barrier after
809 * clearing/invalidating the PTE (A1) and before restoring the PTE (A4),
810 * so we use explicit ones here.
811 */
812
813 /* Paired with the memory barrier in try_grab_folio(). */
814 if (IS_ENABLED(CONFIG_HAVE_GUP_FAST))
815 smp_mb();
816
817 if (unlikely(folio_maybe_dma_pinned(folio)))
818 return -EBUSY;
819 ClearPageAnonExclusive(page);
820
821 /*
822 * This is conceptually a smp_wmb() paired with the smp_rmb() in
823 * gup_must_unshare().
824 */
825 if (IS_ENABLED(CONFIG_HAVE_GUP_FAST))
826 smp_mb__after_atomic();
827 return 0;
828}
829
830/**
831 * folio_try_share_anon_rmap_pte - try marking an exclusive anonymous page
832 * mapped by a PTE possibly shared to prepare
833 * for KSM or temporary unmapping
834 * @folio: The folio to share a mapping of
835 * @page: The mapped exclusive page
836 *
837 * The caller needs to hold the page table lock and has to have the page table
838 * entries cleared/invalidated.
839 *
840 * This is similar to folio_try_dup_anon_rmap_pte(), however, not used during
841 * fork() to duplicate mappings, but instead to prepare for KSM or temporarily
842 * unmapping parts of a folio (swap, migration) via folio_remove_rmap_pte().
843 *
844 * Marking the mapped page shared can only fail if the folio maybe pinned;
845 * device private folios cannot get pinned and consequently this function cannot
846 * fail.
847 *
848 * Returns 0 if marking the mapped page possibly shared succeeded. Returns
849 * -EBUSY otherwise.
850 */
851static inline int folio_try_share_anon_rmap_pte(struct folio *folio,
852 struct page *page)
853{
854 return __folio_try_share_anon_rmap(folio, page, nr_pages: 1, level: RMAP_LEVEL_PTE);
855}
856
857/**
858 * folio_try_share_anon_rmap_pmd - try marking an exclusive anonymous page
859 * range mapped by a PMD possibly shared to
860 * prepare for temporary unmapping
861 * @folio: The folio to share the mapping of
862 * @page: The first page to share the mapping of
863 *
864 * The page range of the folio is defined by [page, page + HPAGE_PMD_NR)
865 *
866 * The caller needs to hold the page table lock and has to have the page table
867 * entries cleared/invalidated.
868 *
869 * This is similar to folio_try_dup_anon_rmap_pmd(), however, not used during
870 * fork() to duplicate a mapping, but instead to prepare for temporarily
871 * unmapping parts of a folio (swap, migration) via folio_remove_rmap_pmd().
872 *
873 * Marking the mapped pages shared can only fail if the folio maybe pinned;
874 * device private folios cannot get pinned and consequently this function cannot
875 * fail.
876 *
877 * Returns 0 if marking the mapped pages possibly shared succeeded. Returns
878 * -EBUSY otherwise.
879 */
880static inline int folio_try_share_anon_rmap_pmd(struct folio *folio,
881 struct page *page)
882{
883#ifdef CONFIG_TRANSPARENT_HUGEPAGE
884 return __folio_try_share_anon_rmap(folio, page, HPAGE_PMD_NR,
885 level: RMAP_LEVEL_PMD);
886#else
887 WARN_ON_ONCE(true);
888 return -EBUSY;
889#endif
890}
891
892/*
893 * Called from mm/vmscan.c to handle paging out
894 */
895int folio_referenced(struct folio *, int is_locked,
896 struct mem_cgroup *memcg, unsigned long *vm_flags);
897
898void try_to_migrate(struct folio *folio, enum ttu_flags flags);
899void try_to_unmap(struct folio *, enum ttu_flags flags);
900
901struct page *make_device_exclusive(struct mm_struct *mm, unsigned long addr,
902 void *owner, struct folio **foliop);
903
904/* Avoid racy checks */
905#define PVMW_SYNC (1 << 0)
906/* Look for migration entries rather than present PTEs */
907#define PVMW_MIGRATION (1 << 1)
908
909struct page_vma_mapped_walk {
910 unsigned long pfn;
911 unsigned long nr_pages;
912 pgoff_t pgoff;
913 struct vm_area_struct *vma;
914 unsigned long address;
915 pmd_t *pmd;
916 pte_t *pte;
917 spinlock_t *ptl;
918 unsigned int flags;
919};
920
921#define DEFINE_FOLIO_VMA_WALK(name, _folio, _vma, _address, _flags) \
922 struct page_vma_mapped_walk name = { \
923 .pfn = folio_pfn(_folio), \
924 .nr_pages = folio_nr_pages(_folio), \
925 .pgoff = folio_pgoff(_folio), \
926 .vma = _vma, \
927 .address = _address, \
928 .flags = _flags, \
929 }
930
931static inline void page_vma_mapped_walk_done(struct page_vma_mapped_walk *pvmw)
932{
933 /* HugeTLB pte is set to the relevant page table entry without pte_mapped. */
934 if (pvmw->pte && !is_vm_hugetlb_page(vma: pvmw->vma))
935 pte_unmap(pte: pvmw->pte);
936 if (pvmw->ptl)
937 spin_unlock(lock: pvmw->ptl);
938}
939
940/**
941 * page_vma_mapped_walk_restart - Restart the page table walk.
942 * @pvmw: Pointer to struct page_vma_mapped_walk.
943 *
944 * It restarts the page table walk when changes occur in the page
945 * table, such as splitting a PMD. Ensures that the PTL held during
946 * the previous walk is released and resets the state to allow for
947 * a new walk starting at the current address stored in pvmw->address.
948 */
949static inline void
950page_vma_mapped_walk_restart(struct page_vma_mapped_walk *pvmw)
951{
952 WARN_ON_ONCE(!pvmw->pmd && !pvmw->pte);
953
954 if (likely(pvmw->ptl))
955 spin_unlock(lock: pvmw->ptl);
956 else
957 WARN_ON_ONCE(1);
958
959 pvmw->ptl = NULL;
960 pvmw->pmd = NULL;
961 pvmw->pte = NULL;
962}
963
964bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw);
965unsigned long page_address_in_vma(const struct folio *folio,
966 const struct page *, const struct vm_area_struct *);
967
968/*
969 * Cleans the PTEs of shared mappings.
970 * (and since clean PTEs should also be readonly, write protects them too)
971 *
972 * returns the number of cleaned PTEs.
973 */
974int folio_mkclean(struct folio *);
975
976int mapping_wrprotect_range(struct address_space *mapping, pgoff_t pgoff,
977 unsigned long pfn, unsigned long nr_pages);
978
979int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff,
980 struct vm_area_struct *vma);
981
982enum rmp_flags {
983 RMP_LOCKED = 1 << 0,
984 RMP_USE_SHARED_ZEROPAGE = 1 << 1,
985};
986
987void remove_migration_ptes(struct folio *src, struct folio *dst, int flags);
988
989/*
990 * rmap_walk_control: To control rmap traversing for specific needs
991 *
992 * arg: passed to rmap_one() and invalid_vma()
993 * try_lock: bail out if the rmap lock is contended
994 * contended: indicate the rmap traversal bailed out due to lock contention
995 * rmap_one: executed on each vma where page is mapped
996 * done: for checking traversing termination condition
997 * anon_lock: for getting anon_lock by optimized way rather than default
998 * invalid_vma: for skipping uninterested vma
999 */
1000struct rmap_walk_control {
1001 void *arg;
1002 bool try_lock;
1003 bool contended;
1004 /*
1005 * Return false if page table scanning in rmap_walk should be stopped.
1006 * Otherwise, return true.
1007 */
1008 bool (*rmap_one)(struct folio *folio, struct vm_area_struct *vma,
1009 unsigned long addr, void *arg);
1010 int (*done)(struct folio *folio);
1011 struct anon_vma *(*anon_lock)(const struct folio *folio,
1012 struct rmap_walk_control *rwc);
1013 bool (*invalid_vma)(struct vm_area_struct *vma, void *arg);
1014};
1015
1016void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc);
1017void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc);
1018struct anon_vma *folio_lock_anon_vma_read(const struct folio *folio,
1019 struct rmap_walk_control *rwc);
1020
1021#else /* !CONFIG_MMU */
1022
1023#define anon_vma_init() do {} while (0)
1024#define anon_vma_prepare(vma) (0)
1025
1026static inline int folio_referenced(struct folio *folio, int is_locked,
1027 struct mem_cgroup *memcg,
1028 unsigned long *vm_flags)
1029{
1030 *vm_flags = 0;
1031 return 0;
1032}
1033
1034static inline void try_to_unmap(struct folio *folio, enum ttu_flags flags)
1035{
1036}
1037
1038static inline int folio_mkclean(struct folio *folio)
1039{
1040 return 0;
1041}
1042#endif /* CONFIG_MMU */
1043
1044#endif /* _LINUX_RMAP_H */
1045

source code of linux/include/linux/rmap.h

Morty Proxy This is a proxified and sanitized view of the page, visit original site.