~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/mm/rmap.c

Version: ~ [ linux-5.13-rc5 ] ~ [ linux-5.12.9 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.42 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.124 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.193 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.235 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.271 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.271 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * mm/rmap.c - physical to virtual reverse mappings
  3  *
  4  * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
  5  * Released under the General Public License (GPL).
  6  *
  7  * Simple, low overhead reverse mapping scheme.
  8  * Please try to keep this thing as modular as possible.
  9  *
 10  * Provides methods for unmapping each kind of mapped page:
 11  * the anon methods track anonymous pages, and
 12  * the file methods track pages belonging to an inode.
 13  *
 14  * Original design by Rik van Riel <riel@conectiva.com.br> 2001
 15  * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
 16  * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
 17  * Contributions by Hugh Dickins 2003, 2004
 18  */
 19 
 20 /*
 21  * Lock ordering in mm:
 22  *
 23  * inode->i_mutex       (while writing or truncating, not reading or faulting)
 24  *   mm->mmap_sem
 25  *     page->flags PG_locked (lock_page)
 26  *       mapping->i_mmap_rwsem
 27  *         anon_vma->rwsem
 28  *           mm->page_table_lock or pte_lock
 29  *             zone->lru_lock (in mark_page_accessed, isolate_lru_page)
 30  *             swap_lock (in swap_duplicate, swap_info_get)
 31  *               mmlist_lock (in mmput, drain_mmlist and others)
 32  *               mapping->private_lock (in __set_page_dirty_buffers)
 33  *                 mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
 34  *                   mapping->tree_lock (widely used)
 35  *               inode->i_lock (in set_page_dirty's __mark_inode_dirty)
 36  *               bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
 37  *                 sb_lock (within inode_lock in fs/fs-writeback.c)
 38  *                 mapping->tree_lock (widely used, in set_page_dirty,
 39  *                           in arch-dependent flush_dcache_mmap_lock,
 40  *                           within bdi.wb->list_lock in __sync_single_inode)
 41  *
 42  * anon_vma->rwsem,mapping->i_mutex      (memory_failure, collect_procs_anon)
 43  *   ->tasklist_lock
 44  *     pte map lock
 45  */
 46 
 47 #include <linux/mm.h>
 48 #include <linux/pagemap.h>
 49 #include <linux/swap.h>
 50 #include <linux/swapops.h>
 51 #include <linux/slab.h>
 52 #include <linux/init.h>
 53 #include <linux/ksm.h>
 54 #include <linux/rmap.h>
 55 #include <linux/rcupdate.h>
 56 #include <linux/export.h>
 57 #include <linux/memcontrol.h>
 58 #include <linux/mmu_notifier.h>
 59 #include <linux/migrate.h>
 60 #include <linux/hugetlb.h>
 61 #include <linux/backing-dev.h>
 62 #include <linux/page_idle.h>
 63 
 64 #include <asm/tlbflush.h>
 65 
 66 #include <trace/events/tlb.h>
 67 
 68 #include "internal.h"
 69 
 70 static struct kmem_cache *anon_vma_cachep;
 71 static struct kmem_cache *anon_vma_chain_cachep;
 72 
 73 static inline struct anon_vma *anon_vma_alloc(void)
 74 {
 75         struct anon_vma *anon_vma;
 76 
 77         anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
 78         if (anon_vma) {
 79                 atomic_set(&anon_vma->refcount, 1);
 80                 anon_vma->degree = 1;   /* Reference for first vma */
 81                 anon_vma->parent = anon_vma;
 82                 /*
 83                  * Initialise the anon_vma root to point to itself. If called
 84                  * from fork, the root will be reset to the parents anon_vma.
 85                  */
 86                 anon_vma->root = anon_vma;
 87         }
 88 
 89         return anon_vma;
 90 }
 91 
 92 static inline void anon_vma_free(struct anon_vma *anon_vma)
 93 {
 94         VM_BUG_ON(atomic_read(&anon_vma->refcount));
 95 
 96         /*
 97          * Synchronize against page_lock_anon_vma_read() such that
 98          * we can safely hold the lock without the anon_vma getting
 99          * freed.
100          *
101          * Relies on the full mb implied by the atomic_dec_and_test() from
102          * put_anon_vma() against the acquire barrier implied by
103          * down_read_trylock() from page_lock_anon_vma_read(). This orders:
104          *
105          * page_lock_anon_vma_read()    VS      put_anon_vma()
106          *   down_read_trylock()                  atomic_dec_and_test()
107          *   LOCK                                 MB
108          *   atomic_read()                        rwsem_is_locked()
109          *
110          * LOCK should suffice since the actual taking of the lock must
111          * happen _before_ what follows.
112          */
113         might_sleep();
114         if (rwsem_is_locked(&anon_vma->root->rwsem)) {
115                 anon_vma_lock_write(anon_vma);
116                 anon_vma_unlock_write(anon_vma);
117         }
118 
119         kmem_cache_free(anon_vma_cachep, anon_vma);
120 }
121 
122 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
123 {
124         return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
125 }
126 
127 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
128 {
129         kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
130 }
131 
132 static void anon_vma_chain_link(struct vm_area_struct *vma,
133                                 struct anon_vma_chain *avc,
134                                 struct anon_vma *anon_vma)
135 {
136         avc->vma = vma;
137         avc->anon_vma = anon_vma;
138         list_add(&avc->same_vma, &vma->anon_vma_chain);
139         anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
140 }
141 
142 /**
143  * anon_vma_prepare - attach an anon_vma to a memory region
144  * @vma: the memory region in question
145  *
146  * This makes sure the memory mapping described by 'vma' has
147  * an 'anon_vma' attached to it, so that we can associate the
148  * anonymous pages mapped into it with that anon_vma.
149  *
150  * The common case will be that we already have one, but if
151  * not we either need to find an adjacent mapping that we
152  * can re-use the anon_vma from (very common when the only
153  * reason for splitting a vma has been mprotect()), or we
154  * allocate a new one.
155  *
156  * Anon-vma allocations are very subtle, because we may have
157  * optimistically looked up an anon_vma in page_lock_anon_vma_read()
158  * and that may actually touch the spinlock even in the newly
159  * allocated vma (it depends on RCU to make sure that the
160  * anon_vma isn't actually destroyed).
161  *
162  * As a result, we need to do proper anon_vma locking even
163  * for the new allocation. At the same time, we do not want
164  * to do any locking for the common case of already having
165  * an anon_vma.
166  *
167  * This must be called with the mmap_sem held for reading.
168  */
169 int anon_vma_prepare(struct vm_area_struct *vma)
170 {
171         struct anon_vma *anon_vma = vma->anon_vma;
172         struct anon_vma_chain *avc;
173 
174         might_sleep();
175         if (unlikely(!anon_vma)) {
176                 struct mm_struct *mm = vma->vm_mm;
177                 struct anon_vma *allocated;
178 
179                 avc = anon_vma_chain_alloc(GFP_KERNEL);
180                 if (!avc)
181                         goto out_enomem;
182 
183                 anon_vma = find_mergeable_anon_vma(vma);
184                 allocated = NULL;
185                 if (!anon_vma) {
186                         anon_vma = anon_vma_alloc();
187                         if (unlikely(!anon_vma))
188                                 goto out_enomem_free_avc;
189                         allocated = anon_vma;
190                 }
191 
192                 anon_vma_lock_write(anon_vma);
193                 /* page_table_lock to protect against threads */
194                 spin_lock(&mm->page_table_lock);
195                 if (likely(!vma->anon_vma)) {
196                         vma->anon_vma = anon_vma;
197                         anon_vma_chain_link(vma, avc, anon_vma);
198                         /* vma reference or self-parent link for new root */
199                         anon_vma->degree++;
200                         allocated = NULL;
201                         avc = NULL;
202                 }
203                 spin_unlock(&mm->page_table_lock);
204                 anon_vma_unlock_write(anon_vma);
205 
206                 if (unlikely(allocated))
207                         put_anon_vma(allocated);
208                 if (unlikely(avc))
209                         anon_vma_chain_free(avc);
210         }
211         return 0;
212 
213  out_enomem_free_avc:
214         anon_vma_chain_free(avc);
215  out_enomem:
216         return -ENOMEM;
217 }
218 
219 /*
220  * This is a useful helper function for locking the anon_vma root as
221  * we traverse the vma->anon_vma_chain, looping over anon_vma's that
222  * have the same vma.
223  *
224  * Such anon_vma's should have the same root, so you'd expect to see
225  * just a single mutex_lock for the whole traversal.
226  */
227 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
228 {
229         struct anon_vma *new_root = anon_vma->root;
230         if (new_root != root) {
231                 if (WARN_ON_ONCE(root))
232                         up_write(&root->rwsem);
233                 root = new_root;
234                 down_write(&root->rwsem);
235         }
236         return root;
237 }
238 
239 static inline void unlock_anon_vma_root(struct anon_vma *root)
240 {
241         if (root)
242                 up_write(&root->rwsem);
243 }
244 
245 /*
246  * Attach the anon_vmas from src to dst.
247  * Returns 0 on success, -ENOMEM on failure.
248  *
249  * If dst->anon_vma is NULL this function tries to find and reuse existing
250  * anon_vma which has no vmas and only one child anon_vma. This prevents
251  * degradation of anon_vma hierarchy to endless linear chain in case of
252  * constantly forking task. On the other hand, an anon_vma with more than one
253  * child isn't reused even if there was no alive vma, thus rmap walker has a
254  * good chance of avoiding scanning the whole hierarchy when it searches where
255  * page is mapped.
256  */
257 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
258 {
259         struct anon_vma_chain *avc, *pavc;
260         struct anon_vma *root = NULL;
261 
262         list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
263                 struct anon_vma *anon_vma;
264 
265                 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
266                 if (unlikely(!avc)) {
267                         unlock_anon_vma_root(root);
268                         root = NULL;
269                         avc = anon_vma_chain_alloc(GFP_KERNEL);
270                         if (!avc)
271                                 goto enomem_failure;
272                 }
273                 anon_vma = pavc->anon_vma;
274                 root = lock_anon_vma_root(root, anon_vma);
275                 anon_vma_chain_link(dst, avc, anon_vma);
276 
277                 /*
278                  * Reuse existing anon_vma if its degree lower than two,
279                  * that means it has no vma and only one anon_vma child.
280                  *
281                  * Do not chose parent anon_vma, otherwise first child
282                  * will always reuse it. Root anon_vma is never reused:
283                  * it has self-parent reference and at least one child.
284                  */
285                 if (!dst->anon_vma && anon_vma != src->anon_vma &&
286                                 anon_vma->degree < 2)
287                         dst->anon_vma = anon_vma;
288         }
289         if (dst->anon_vma)
290                 dst->anon_vma->degree++;
291         unlock_anon_vma_root(root);
292         return 0;
293 
294  enomem_failure:
295         /*
296          * dst->anon_vma is dropped here otherwise its degree can be incorrectly
297          * decremented in unlink_anon_vmas().
298          * We can safely do this because callers of anon_vma_clone() don't care
299          * about dst->anon_vma if anon_vma_clone() failed.
300          */
301         dst->anon_vma = NULL;
302         unlink_anon_vmas(dst);
303         return -ENOMEM;
304 }
305 
306 /*
307  * Attach vma to its own anon_vma, as well as to the anon_vmas that
308  * the corresponding VMA in the parent process is attached to.
309  * Returns 0 on success, non-zero on failure.
310  */
311 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
312 {
313         struct anon_vma_chain *avc;
314         struct anon_vma *anon_vma;
315         int error;
316 
317         /* Don't bother if the parent process has no anon_vma here. */
318         if (!pvma->anon_vma)
319                 return 0;
320 
321         /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
322         vma->anon_vma = NULL;
323 
324         /*
325          * First, attach the new VMA to the parent VMA's anon_vmas,
326          * so rmap can find non-COWed pages in child processes.
327          */
328         error = anon_vma_clone(vma, pvma);
329         if (error)
330                 return error;
331 
332         /* An existing anon_vma has been reused, all done then. */
333         if (vma->anon_vma)
334                 return 0;
335 
336         /* Then add our own anon_vma. */
337         anon_vma = anon_vma_alloc();
338         if (!anon_vma)
339                 goto out_error;
340         avc = anon_vma_chain_alloc(GFP_KERNEL);
341         if (!avc)
342                 goto out_error_free_anon_vma;
343 
344         /*
345          * The root anon_vma's spinlock is the lock actually used when we
346          * lock any of the anon_vmas in this anon_vma tree.
347          */
348         anon_vma->root = pvma->anon_vma->root;
349         anon_vma->parent = pvma->anon_vma;
350         /*
351          * With refcounts, an anon_vma can stay around longer than the
352          * process it belongs to. The root anon_vma needs to be pinned until
353          * this anon_vma is freed, because the lock lives in the root.
354          */
355         get_anon_vma(anon_vma->root);
356         /* Mark this anon_vma as the one where our new (COWed) pages go. */
357         vma->anon_vma = anon_vma;
358         anon_vma_lock_write(anon_vma);
359         anon_vma_chain_link(vma, avc, anon_vma);
360         anon_vma->parent->degree++;
361         anon_vma_unlock_write(anon_vma);
362 
363         return 0;
364 
365  out_error_free_anon_vma:
366         put_anon_vma(anon_vma);
367  out_error:
368         unlink_anon_vmas(vma);
369         return -ENOMEM;
370 }
371 
372 void unlink_anon_vmas(struct vm_area_struct *vma)
373 {
374         struct anon_vma_chain *avc, *next;
375         struct anon_vma *root = NULL;
376 
377         /*
378          * Unlink each anon_vma chained to the VMA.  This list is ordered
379          * from newest to oldest, ensuring the root anon_vma gets freed last.
380          */
381         list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
382                 struct anon_vma *anon_vma = avc->anon_vma;
383 
384                 root = lock_anon_vma_root(root, anon_vma);
385                 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
386 
387                 /*
388                  * Leave empty anon_vmas on the list - we'll need
389                  * to free them outside the lock.
390                  */
391                 if (RB_EMPTY_ROOT(&anon_vma->rb_root)) {
392                         anon_vma->parent->degree--;
393                         continue;
394                 }
395 
396                 list_del(&avc->same_vma);
397                 anon_vma_chain_free(avc);
398         }
399         if (vma->anon_vma)
400                 vma->anon_vma->degree--;
401         unlock_anon_vma_root(root);
402 
403         /*
404          * Iterate the list once more, it now only contains empty and unlinked
405          * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
406          * needing to write-acquire the anon_vma->root->rwsem.
407          */
408         list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
409                 struct anon_vma *anon_vma = avc->anon_vma;
410 
411                 BUG_ON(anon_vma->degree);
412                 put_anon_vma(anon_vma);
413 
414                 list_del(&avc->same_vma);
415                 anon_vma_chain_free(avc);
416         }
417 }
418 
419 static void anon_vma_ctor(void *data)
420 {
421         struct anon_vma *anon_vma = data;
422 
423         init_rwsem(&anon_vma->rwsem);
424         atomic_set(&anon_vma->refcount, 0);
425         anon_vma->rb_root = RB_ROOT;
426 }
427 
428 void __init anon_vma_init(void)
429 {
430         anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
431                         0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
432         anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
433 }
434 
435 /*
436  * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
437  *
438  * Since there is no serialization what so ever against page_remove_rmap()
439  * the best this function can do is return a locked anon_vma that might
440  * have been relevant to this page.
441  *
442  * The page might have been remapped to a different anon_vma or the anon_vma
443  * returned may already be freed (and even reused).
444  *
445  * In case it was remapped to a different anon_vma, the new anon_vma will be a
446  * child of the old anon_vma, and the anon_vma lifetime rules will therefore
447  * ensure that any anon_vma obtained from the page will still be valid for as
448  * long as we observe page_mapped() [ hence all those page_mapped() tests ].
449  *
450  * All users of this function must be very careful when walking the anon_vma
451  * chain and verify that the page in question is indeed mapped in it
452  * [ something equivalent to page_mapped_in_vma() ].
453  *
454  * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
455  * that the anon_vma pointer from page->mapping is valid if there is a
456  * mapcount, we can dereference the anon_vma after observing those.
457  */
458 struct anon_vma *page_get_anon_vma(struct page *page)
459 {
460         struct anon_vma *anon_vma = NULL;
461         unsigned long anon_mapping;
462 
463         rcu_read_lock();
464         anon_mapping = (unsigned long)READ_ONCE(page->mapping);
465         if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
466                 goto out;
467         if (!page_mapped(page))
468                 goto out;
469 
470         anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
471         if (!atomic_inc_not_zero(&anon_vma->refcount)) {
472                 anon_vma = NULL;
473                 goto out;
474         }
475 
476         /*
477          * If this page is still mapped, then its anon_vma cannot have been
478          * freed.  But if it has been unmapped, we have no security against the
479          * anon_vma structure being freed and reused (for another anon_vma:
480          * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
481          * above cannot corrupt).
482          */
483         if (!page_mapped(page)) {
484                 rcu_read_unlock();
485                 put_anon_vma(anon_vma);
486                 return NULL;
487         }
488 out:
489         rcu_read_unlock();
490 
491         return anon_vma;
492 }
493 
494 /*
495  * Similar to page_get_anon_vma() except it locks the anon_vma.
496  *
497  * Its a little more complex as it tries to keep the fast path to a single
498  * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
499  * reference like with page_get_anon_vma() and then block on the mutex.
500  */
501 struct anon_vma *page_lock_anon_vma_read(struct page *page)
502 {
503         struct anon_vma *anon_vma = NULL;
504         struct anon_vma *root_anon_vma;
505         unsigned long anon_mapping;
506 
507         rcu_read_lock();
508         anon_mapping = (unsigned long)READ_ONCE(page->mapping);
509         if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
510                 goto out;
511         if (!page_mapped(page))
512                 goto out;
513 
514         anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
515         root_anon_vma = READ_ONCE(anon_vma->root);
516         if (down_read_trylock(&root_anon_vma->rwsem)) {
517                 /*
518                  * If the page is still mapped, then this anon_vma is still
519                  * its anon_vma, and holding the mutex ensures that it will
520                  * not go away, see anon_vma_free().
521                  */
522                 if (!page_mapped(page)) {
523                         up_read(&root_anon_vma->rwsem);
524                         anon_vma = NULL;
525                 }
526                 goto out;
527         }
528 
529         /* trylock failed, we got to sleep */
530         if (!atomic_inc_not_zero(&anon_vma->refcount)) {
531                 anon_vma = NULL;
532                 goto out;
533         }
534 
535         if (!page_mapped(page)) {
536                 rcu_read_unlock();
537                 put_anon_vma(anon_vma);
538                 return NULL;
539         }
540 
541         /* we pinned the anon_vma, its safe to sleep */
542         rcu_read_unlock();
543         anon_vma_lock_read(anon_vma);
544 
545         if (atomic_dec_and_test(&anon_vma->refcount)) {
546                 /*
547                  * Oops, we held the last refcount, release the lock
548                  * and bail -- can't simply use put_anon_vma() because
549                  * we'll deadlock on the anon_vma_lock_write() recursion.
550                  */
551                 anon_vma_unlock_read(anon_vma);
552                 __put_anon_vma(anon_vma);
553                 anon_vma = NULL;
554         }
555 
556         return anon_vma;
557 
558 out:
559         rcu_read_unlock();
560         return anon_vma;
561 }
562 
563 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
564 {
565         anon_vma_unlock_read(anon_vma);
566 }
567 
568 /*
569  * At what user virtual address is page expected in @vma?
570  */
571 static inline unsigned long
572 __vma_address(struct page *page, struct vm_area_struct *vma)
573 {
574         pgoff_t pgoff = page_to_pgoff(page);
575         return vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
576 }
577 
578 inline unsigned long
579 vma_address(struct page *page, struct vm_area_struct *vma)
580 {
581         unsigned long address = __vma_address(page, vma);
582 
583         /* page should be within @vma mapping range */
584         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
585 
586         return address;
587 }
588 
589 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
590 static void percpu_flush_tlb_batch_pages(void *data)
591 {
592         /*
593          * All TLB entries are flushed on the assumption that it is
594          * cheaper to flush all TLBs and let them be refilled than
595          * flushing individual PFNs. Note that we do not track mm's
596          * to flush as that might simply be multiple full TLB flushes
597          * for no gain.
598          */
599         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
600         flush_tlb_local();
601 }
602 
603 /*
604  * Flush TLB entries for recently unmapped pages from remote CPUs. It is
605  * important if a PTE was dirty when it was unmapped that it's flushed
606  * before any IO is initiated on the page to prevent lost writes. Similarly,
607  * it must be flushed before freeing to prevent data leakage.
608  */
609 void try_to_unmap_flush(void)
610 {
611         struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
612         int cpu;
613 
614         if (!tlb_ubc->flush_required)
615                 return;
616 
617         cpu = get_cpu();
618 
619         trace_tlb_flush(TLB_REMOTE_SHOOTDOWN, -1UL);
620 
621         if (cpumask_test_cpu(cpu, &tlb_ubc->cpumask))
622                 percpu_flush_tlb_batch_pages(&tlb_ubc->cpumask);
623 
624         if (cpumask_any_but(&tlb_ubc->cpumask, cpu) < nr_cpu_ids) {
625                 smp_call_function_many(&tlb_ubc->cpumask,
626                         percpu_flush_tlb_batch_pages, (void *)tlb_ubc, true);
627         }
628         cpumask_clear(&tlb_ubc->cpumask);
629         tlb_ubc->flush_required = false;
630         tlb_ubc->writable = false;
631         put_cpu();
632 }
633 
634 /* Flush iff there are potentially writable TLB entries that can race with IO */
635 void try_to_unmap_flush_dirty(void)
636 {
637         struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
638 
639         if (tlb_ubc->writable)
640                 try_to_unmap_flush();
641 }
642 
643 static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
644                 struct page *page, bool writable)
645 {
646         struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
647 
648         cpumask_or(&tlb_ubc->cpumask, &tlb_ubc->cpumask, mm_cpumask(mm));
649         tlb_ubc->flush_required = true;
650 
651         /*
652          * If the PTE was dirty then it's best to assume it's writable. The
653          * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
654          * before the page is queued for IO.
655          */
656         if (writable)
657                 tlb_ubc->writable = true;
658 }
659 
660 /*
661  * Returns true if the TLB flush should be deferred to the end of a batch of
662  * unmap operations to reduce IPIs.
663  */
664 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
665 {
666         bool should_defer = false;
667 
668         if (!(flags & TTU_BATCH_FLUSH))
669                 return false;
670 
671         /* If remote CPUs need to be flushed then defer batch the flush */
672         if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
673                 should_defer = true;
674         put_cpu();
675 
676         return should_defer;
677 }
678 #else
679 static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
680                 struct page *page, bool writable)
681 {
682 }
683 
684 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
685 {
686         return false;
687 }
688 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
689 
690 /*
691  * At what user virtual address is page expected in vma?
692  * Caller should check the page is actually part of the vma.
693  */
694 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
695 {
696         unsigned long address;
697         if (PageAnon(page)) {
698                 struct anon_vma *page__anon_vma = page_anon_vma(page);
699                 /*
700                  * Note: swapoff's unuse_vma() is more efficient with this
701                  * check, and needs it to match anon_vma when KSM is active.
702                  */
703                 if (!vma->anon_vma || !page__anon_vma ||
704                     vma->anon_vma->root != page__anon_vma->root)
705                         return -EFAULT;
706         } else if (page->mapping) {
707                 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
708                         return -EFAULT;
709         } else
710                 return -EFAULT;
711         address = __vma_address(page, vma);
712         if (unlikely(address < vma->vm_start || address >= vma->vm_end))
713                 return -EFAULT;
714         return address;
715 }
716 
717 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
718 {
719         pgd_t *pgd;
720         pud_t *pud;
721         pmd_t *pmd = NULL;
722         pmd_t pmde;
723 
724         pgd = pgd_offset(mm, address);
725         if (!pgd_present(*pgd))
726                 goto out;
727 
728         pud = pud_offset(pgd, address);
729         if (!pud_present(*pud))
730                 goto out;
731 
732         pmd = pmd_offset(pud, address);
733         /*
734          * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
735          * without holding anon_vma lock for write.  So when looking for a
736          * genuine pmde (in which to find pte), test present and !THP together.
737          */
738         pmde = *pmd;
739         barrier();
740         if (!pmd_present(pmde) || pmd_trans_huge(pmde))
741                 pmd = NULL;
742 out:
743         return pmd;
744 }
745 
746 /*
747  * Check that @page is mapped at @address into @mm.
748  *
749  * If @sync is false, page_check_address may perform a racy check to avoid
750  * the page table lock when the pte is not present (helpful when reclaiming
751  * highly shared pages).
752  *
753  * On success returns with pte mapped and locked.
754  */
755 pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
756                           unsigned long address, spinlock_t **ptlp, int sync)
757 {
758         pmd_t *pmd;
759         pte_t *pte;
760         spinlock_t *ptl;
761 
762         if (unlikely(PageHuge(page))) {
763                 /* when pud is not present, pte will be NULL */
764                 pte = huge_pte_offset(mm, address);
765                 if (!pte)
766                         return NULL;
767 
768                 ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
769                 goto check;
770         }
771 
772         pmd = mm_find_pmd(mm, address);
773         if (!pmd)
774                 return NULL;
775 
776         pte = pte_offset_map(pmd, address);
777         /* Make a quick check before getting the lock */
778         if (!sync && !pte_present(*pte)) {
779                 pte_unmap(pte);
780                 return NULL;
781         }
782 
783         ptl = pte_lockptr(mm, pmd);
784 check:
785         spin_lock(ptl);
786         if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
787                 *ptlp = ptl;
788                 return pte;
789         }
790         pte_unmap_unlock(pte, ptl);
791         return NULL;
792 }
793 
794 /**
795  * page_mapped_in_vma - check whether a page is really mapped in a VMA
796  * @page: the page to test
797  * @vma: the VMA to test
798  *
799  * Returns 1 if the page is mapped into the page tables of the VMA, 0
800  * if the page is not mapped into the page tables of this VMA.  Only
801  * valid for normal file or anonymous VMAs.
802  */
803 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
804 {
805         unsigned long address;
806         pte_t *pte;
807         spinlock_t *ptl;
808 
809         address = __vma_address(page, vma);
810         if (unlikely(address < vma->vm_start || address >= vma->vm_end))
811                 return 0;
812         pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
813         if (!pte)                       /* the page is not in this mm */
814                 return 0;
815         pte_unmap_unlock(pte, ptl);
816 
817         return 1;
818 }
819 
820 struct page_referenced_arg {
821         int mapcount;
822         int referenced;
823         unsigned long vm_flags;
824         struct mem_cgroup *memcg;
825 };
826 /*
827  * arg: page_referenced_arg will be passed
828  */
829 static int page_referenced_one(struct page *page, struct vm_area_struct *vma,
830                         unsigned long address, void *arg)
831 {
832         struct mm_struct *mm = vma->vm_mm;
833         spinlock_t *ptl;
834         int referenced = 0;
835         struct page_referenced_arg *pra = arg;
836 
837         if (unlikely(PageTransHuge(page))) {
838                 pmd_t *pmd;
839 
840                 /*
841                  * rmap might return false positives; we must filter
842                  * these out using page_check_address_pmd().
843                  */
844                 pmd = page_check_address_pmd(page, mm, address,
845                                              PAGE_CHECK_ADDRESS_PMD_FLAG, &ptl);
846                 if (!pmd)
847                         return SWAP_AGAIN;
848 
849                 if (vma->vm_flags & VM_LOCKED) {
850                         spin_unlock(ptl);
851                         pra->vm_flags |= VM_LOCKED;
852                         return SWAP_FAIL; /* To break the loop */
853                 }
854 
855                 /* go ahead even if the pmd is pmd_trans_splitting() */
856                 if (pmdp_clear_flush_young_notify(vma, address, pmd))
857                         referenced++;
858                 spin_unlock(ptl);
859         } else {
860                 pte_t *pte;
861 
862                 /*
863                  * rmap might return false positives; we must filter
864                  * these out using page_check_address().
865                  */
866                 pte = page_check_address(page, mm, address, &ptl, 0);
867                 if (!pte)
868                         return SWAP_AGAIN;
869 
870                 if (vma->vm_flags & VM_LOCKED) {
871                         pte_unmap_unlock(pte, ptl);
872                         pra->vm_flags |= VM_LOCKED;
873                         return SWAP_FAIL; /* To break the loop */
874                 }
875 
876                 if (ptep_clear_flush_young_notify(vma, address, pte)) {
877                         /*
878                          * Don't treat a reference through a sequentially read
879                          * mapping as such.  If the page has been used in
880                          * another mapping, we will catch it; if this other
881                          * mapping is already gone, the unmap path will have
882                          * set PG_referenced or activated the page.
883                          */
884                         if (likely(!(vma->vm_flags & VM_SEQ_READ)))
885                                 referenced++;
886                 }
887                 pte_unmap_unlock(pte, ptl);
888         }
889 
890         if (referenced)
891                 clear_page_idle(page);
892         if (test_and_clear_page_young(page))
893                 referenced++;
894 
895         if (referenced) {
896                 pra->referenced++;
897                 pra->vm_flags |= vma->vm_flags;
898         }
899 
900         pra->mapcount--;
901         if (!pra->mapcount)
902                 return SWAP_SUCCESS; /* To break the loop */
903 
904         return SWAP_AGAIN;
905 }
906 
907 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
908 {
909         struct page_referenced_arg *pra = arg;
910         struct mem_cgroup *memcg = pra->memcg;
911 
912         if (!mm_match_cgroup(vma->vm_mm, memcg))
913                 return true;
914 
915         return false;
916 }
917 
918 /**
919  * page_referenced - test if the page was referenced
920  * @page: the page to test
921  * @is_locked: caller holds lock on the page
922  * @memcg: target memory cgroup
923  * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
924  *
925  * Quick test_and_clear_referenced for all mappings to a page,
926  * returns the number of ptes which referenced the page.
927  */
928 int page_referenced(struct page *page,
929                     int is_locked,
930                     struct mem_cgroup *memcg,
931                     unsigned long *vm_flags)
932 {
933         int ret;
934         int we_locked = 0;
935         struct page_referenced_arg pra = {
936                 .mapcount = page_mapcount(page),
937                 .memcg = memcg,
938         };
939         struct rmap_walk_control rwc = {
940                 .rmap_one = page_referenced_one,
941                 .arg = (void *)&pra,
942                 .anon_lock = page_lock_anon_vma_read,
943         };
944 
945         *vm_flags = 0;
946         if (!page_mapped(page))
947                 return 0;
948 
949         if (!page_rmapping(page))
950                 return 0;
951 
952         if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
953                 we_locked = trylock_page(page);
954                 if (!we_locked)
955                         return 1;
956         }
957 
958         /*
959          * If we are reclaiming on behalf of a cgroup, skip
960          * counting on behalf of references from different
961          * cgroups
962          */
963         if (memcg) {
964                 rwc.invalid_vma = invalid_page_referenced_vma;
965         }
966 
967         ret = rmap_walk(page, &rwc);
968         *vm_flags = pra.vm_flags;
969 
970         if (we_locked)
971                 unlock_page(page);
972 
973         return pra.referenced;
974 }
975 
976 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
977                             unsigned long address, void *arg)
978 {
979         struct mm_struct *mm = vma->vm_mm;
980         pte_t *pte;
981         spinlock_t *ptl;
982         int ret = 0;
983         int *cleaned = arg;
984 
985         pte = page_check_address(page, mm, address, &ptl, 1);
986         if (!pte)
987                 goto out;
988 
989         if (pte_dirty(*pte) || pte_write(*pte)) {
990                 pte_t entry;
991 
992                 flush_cache_page(vma, address, pte_pfn(*pte));
993                 entry = ptep_clear_flush(vma, address, pte);
994                 entry = pte_wrprotect(entry);
995                 entry = pte_mkclean(entry);
996                 set_pte_at(mm, address, pte, entry);
997                 ret = 1;
998         }
999 
1000         pte_unmap_unlock(pte, ptl);
1001 
1002         if (ret) {
1003                 mmu_notifier_invalidate_page(mm, address);
1004                 (*cleaned)++;
1005         }
1006 out:
1007         return SWAP_AGAIN;
1008 }
1009 
1010 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1011 {
1012         if (vma->vm_flags & VM_SHARED)
1013                 return false;
1014 
1015         return true;
1016 }
1017 
1018 int page_mkclean(struct page *page)
1019 {
1020         int cleaned = 0;
1021         struct address_space *mapping;
1022         struct rmap_walk_control rwc = {
1023                 .arg = (void *)&cleaned,
1024                 .rmap_one = page_mkclean_one,
1025                 .invalid_vma = invalid_mkclean_vma,
1026         };
1027 
1028         BUG_ON(!PageLocked(page));
1029 
1030         if (!page_mapped(page))
1031                 return 0;
1032 
1033         mapping = page_mapping(page);
1034         if (!mapping)
1035                 return 0;
1036 
1037         rmap_walk(page, &rwc);
1038 
1039         return cleaned;
1040 }
1041 EXPORT_SYMBOL_GPL(page_mkclean);
1042 
1043 /**
1044  * page_move_anon_rmap - move a page to our anon_vma
1045  * @page:       the page to move to our anon_vma
1046  * @vma:        the vma the page belongs to
1047  * @address:    the user virtual address mapped
1048  *
1049  * When a page belongs exclusively to one process after a COW event,
1050  * that page can be moved into the anon_vma that belongs to just that
1051  * process, so the rmap code will not search the parent or sibling
1052  * processes.
1053  */
1054 void page_move_anon_rmap(struct page *page,
1055         struct vm_area_struct *vma, unsigned long address)
1056 {
1057         struct anon_vma *anon_vma = vma->anon_vma;
1058 
1059         VM_BUG_ON_PAGE(!PageLocked(page), page);
1060         VM_BUG_ON_VMA(!anon_vma, vma);
1061         VM_BUG_ON_PAGE(page->index != linear_page_index(vma, address), page);
1062 
1063         anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1064         /*
1065          * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1066          * simultaneously, so a concurrent reader (eg page_referenced()'s
1067          * PageAnon()) will not see one without the other.
1068          */
1069         WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1070 }
1071 
1072 /**
1073  * __page_set_anon_rmap - set up new anonymous rmap
1074  * @page:       Page to add to rmap     
1075  * @vma:        VM area to add page to.
1076  * @address:    User virtual address of the mapping     
1077  * @exclusive:  the page is exclusively owned by the current process
1078  */
1079 static void __page_set_anon_rmap(struct page *page,
1080         struct vm_area_struct *vma, unsigned long address, int exclusive)
1081 {
1082         struct anon_vma *anon_vma = vma->anon_vma;
1083 
1084         BUG_ON(!anon_vma);
1085 
1086         if (PageAnon(page))
1087                 return;
1088 
1089         /*
1090          * If the page isn't exclusively mapped into this vma,
1091          * we must use the _oldest_ possible anon_vma for the
1092          * page mapping!
1093          */
1094         if (!exclusive)
1095                 anon_vma = anon_vma->root;
1096 
1097         anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1098         page->mapping = (struct address_space *) anon_vma;
1099         page->index = linear_page_index(vma, address);
1100 }
1101 
1102 /**
1103  * __page_check_anon_rmap - sanity check anonymous rmap addition
1104  * @page:       the page to add the mapping to
1105  * @vma:        the vm area in which the mapping is added
1106  * @address:    the user virtual address mapped
1107  */
1108 static void __page_check_anon_rmap(struct page *page,
1109         struct vm_area_struct *vma, unsigned long address)
1110 {
1111 #ifdef CONFIG_DEBUG_VM
1112         /*
1113          * The page's anon-rmap details (mapping and index) are guaranteed to
1114          * be set up correctly at this point.
1115          *
1116          * We have exclusion against page_add_anon_rmap because the caller
1117          * always holds the page locked, except if called from page_dup_rmap,
1118          * in which case the page is already known to be setup.
1119          *
1120          * We have exclusion against page_add_new_anon_rmap because those pages
1121          * are initially only visible via the pagetables, and the pte is locked
1122          * over the call to page_add_new_anon_rmap.
1123          */
1124         BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1125         BUG_ON(page->index != linear_page_index(vma, address));
1126 #endif
1127 }
1128 
1129 /**
1130  * page_add_anon_rmap - add pte mapping to an anonymous page
1131  * @page:       the page to add the mapping to
1132  * @vma:        the vm area in which the mapping is added
1133  * @address:    the user virtual address mapped
1134  *
1135  * The caller needs to hold the pte lock, and the page must be locked in
1136  * the anon_vma case: to serialize mapping,index checking after setting,
1137  * and to ensure that PageAnon is not being upgraded racily to PageKsm
1138  * (but PageKsm is never downgraded to PageAnon).
1139  */
1140 void page_add_anon_rmap(struct page *page,
1141         struct vm_area_struct *vma, unsigned long address)
1142 {
1143         do_page_add_anon_rmap(page, vma, address, 0);
1144 }
1145 
1146 /*
1147  * Special version of the above for do_swap_page, which often runs
1148  * into pages that are exclusively owned by the current process.
1149  * Everybody else should continue to use page_add_anon_rmap above.
1150  */
1151 void do_page_add_anon_rmap(struct page *page,
1152         struct vm_area_struct *vma, unsigned long address, int exclusive)
1153 {
1154         int first = atomic_inc_and_test(&page->_mapcount);
1155         if (first) {
1156                 /*
1157                  * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1158                  * these counters are not modified in interrupt context, and
1159                  * pte lock(a spinlock) is held, which implies preemption
1160                  * disabled.
1161                  */
1162                 if (PageTransHuge(page))
1163                         __inc_zone_page_state(page,
1164                                               NR_ANON_TRANSPARENT_HUGEPAGES);
1165                 __mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1166                                 hpage_nr_pages(page));
1167         }
1168         if (unlikely(PageKsm(page)))
1169                 return;
1170 
1171         VM_BUG_ON_PAGE(!PageLocked(page), page);
1172         /* address might be in next vma when migration races vma_adjust */
1173         if (first)
1174                 __page_set_anon_rmap(page, vma, address, exclusive);
1175         else
1176                 __page_check_anon_rmap(page, vma, address);
1177 }
1178 
1179 /**
1180  * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1181  * @page:       the page to add the mapping to
1182  * @vma:        the vm area in which the mapping is added
1183  * @address:    the user virtual address mapped
1184  *
1185  * Same as page_add_anon_rmap but must only be called on *new* pages.
1186  * This means the inc-and-test can be bypassed.
1187  * Page does not have to be locked.
1188  */
1189 void page_add_new_anon_rmap(struct page *page,
1190         struct vm_area_struct *vma, unsigned long address)
1191 {
1192         VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1193         SetPageSwapBacked(page);
1194         atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1195         if (PageTransHuge(page))
1196                 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1197         __mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1198                         hpage_nr_pages(page));
1199         __page_set_anon_rmap(page, vma, address, 1);
1200 }
1201 
1202 /**
1203  * page_add_file_rmap - add pte mapping to a file page
1204  * @page: the page to add the mapping to
1205  *
1206  * The caller needs to hold the pte lock.
1207  */
1208 void page_add_file_rmap(struct page *page)
1209 {
1210         struct mem_cgroup *memcg;
1211 
1212         memcg = mem_cgroup_begin_page_stat(page);
1213         if (atomic_inc_and_test(&page->_mapcount)) {
1214                 __inc_zone_page_state(page, NR_FILE_MAPPED);
1215                 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
1216         }
1217         mem_cgroup_end_page_stat(memcg);
1218 }
1219 
1220 static void page_remove_file_rmap(struct page *page)
1221 {
1222         struct mem_cgroup *memcg;
1223 
1224         memcg = mem_cgroup_begin_page_stat(page);
1225 
1226         /* page still mapped by someone else? */
1227         if (!atomic_add_negative(-1, &page->_mapcount))
1228                 goto out;
1229 
1230         /* Hugepages are not counted in NR_FILE_MAPPED for now. */
1231         if (unlikely(PageHuge(page)))
1232                 goto out;
1233 
1234         /*
1235          * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1236          * these counters are not modified in interrupt context, and
1237          * pte lock(a spinlock) is held, which implies preemption disabled.
1238          */
1239         __dec_zone_page_state(page, NR_FILE_MAPPED);
1240         mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
1241 
1242         if (unlikely(PageMlocked(page)))
1243                 clear_page_mlock(page);
1244 out:
1245         mem_cgroup_end_page_stat(memcg);
1246 }
1247 
1248 /**
1249  * page_remove_rmap - take down pte mapping from a page
1250  * @page: page to remove mapping from
1251  *
1252  * The caller needs to hold the pte lock.
1253  */
1254 void page_remove_rmap(struct page *page)
1255 {
1256         if (!PageAnon(page)) {
1257                 page_remove_file_rmap(page);
1258                 return;
1259         }
1260 
1261         /* page still mapped by someone else? */
1262         if (!atomic_add_negative(-1, &page->_mapcount))
1263                 return;
1264 
1265         /* Hugepages are not counted in NR_ANON_PAGES for now. */
1266         if (unlikely(PageHuge(page)))
1267                 return;
1268 
1269         /*
1270          * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1271          * these counters are not modified in interrupt context, and
1272          * pte lock(a spinlock) is held, which implies preemption disabled.
1273          */
1274         if (PageTransHuge(page))
1275                 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1276 
1277         __mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1278                               -hpage_nr_pages(page));
1279 
1280         if (unlikely(PageMlocked(page)))
1281                 clear_page_mlock(page);
1282 
1283         /*
1284          * It would be tidy to reset the PageAnon mapping here,
1285          * but that might overwrite a racing page_add_anon_rmap
1286          * which increments mapcount after us but sets mapping
1287          * before us: so leave the reset to free_hot_cold_page,
1288          * and remember that it's only reliable while mapped.
1289          * Leaving it set also helps swapoff to reinstate ptes
1290          * faster for those pages still in swapcache.
1291          */
1292 }
1293 
1294 /*
1295  * @arg: enum ttu_flags will be passed to this argument
1296  */
1297 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1298                      unsigned long address, void *arg)
1299 {
1300         struct mm_struct *mm = vma->vm_mm;
1301         pte_t *pte;
1302         pte_t pteval;
1303         spinlock_t *ptl;
1304         int ret = SWAP_AGAIN;
1305         enum ttu_flags flags = (enum ttu_flags)arg;
1306 
1307         pte = page_check_address(page, mm, address, &ptl, 0);
1308         if (!pte)
1309                 goto out;
1310 
1311         /*
1312          * If the page is mlock()d, we cannot swap it out.
1313          * If it's recently referenced (perhaps page_referenced
1314          * skipped over this mm) then we should reactivate it.
1315          */
1316         if (!(flags & TTU_IGNORE_MLOCK)) {
1317                 if (vma->vm_flags & VM_LOCKED)
1318                         goto out_mlock;
1319 
1320                 if (flags & TTU_MUNLOCK)
1321                         goto out_unmap;
1322         }
1323         if (!(flags & TTU_IGNORE_ACCESS)) {
1324                 if (ptep_clear_flush_young_notify(vma, address, pte)) {
1325                         ret = SWAP_FAIL;
1326                         goto out_unmap;
1327                 }
1328         }
1329 
1330         /* Nuke the page table entry. */
1331         flush_cache_page(vma, address, page_to_pfn(page));
1332         if (should_defer_flush(mm, flags)) {
1333                 /*
1334                  * We clear the PTE but do not flush so potentially a remote
1335                  * CPU could still be writing to the page. If the entry was
1336                  * previously clean then the architecture must guarantee that
1337                  * a clear->dirty transition on a cached TLB entry is written
1338                  * through and traps if the PTE is unmapped.
1339                  */
1340                 pteval = ptep_get_and_clear(mm, address, pte);
1341 
1342                 set_tlb_ubc_flush_pending(mm, page, pte_dirty(pteval));
1343         } else {
1344                 pteval = ptep_clear_flush(vma, address, pte);
1345         }
1346 
1347         /* Move the dirty bit to the physical page now the pte is gone. */
1348         if (pte_dirty(pteval))
1349                 set_page_dirty(page);
1350 
1351         /* Update high watermark before we lower rss */
1352         update_hiwater_rss(mm);
1353 
1354         if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1355                 if (!PageHuge(page)) {
1356                         if (PageAnon(page))
1357                                 dec_mm_counter(mm, MM_ANONPAGES);
1358                         else
1359                                 dec_mm_counter(mm, MM_FILEPAGES);
1360                 }
1361                 set_pte_at(mm, address, pte,
1362                            swp_entry_to_pte(make_hwpoison_entry(page)));
1363         } else if (pte_unused(pteval)) {
1364                 /*
1365                  * The guest indicated that the page content is of no
1366                  * interest anymore. Simply discard the pte, vmscan
1367                  * will take care of the rest.
1368                  */
1369                 if (PageAnon(page))
1370                         dec_mm_counter(mm, MM_ANONPAGES);
1371                 else
1372                         dec_mm_counter(mm, MM_FILEPAGES);
1373         } else if (PageAnon(page)) {
1374                 swp_entry_t entry = { .val = page_private(page) };
1375                 pte_t swp_pte;
1376 
1377                 if (PageSwapCache(page)) {
1378                         /*
1379                          * Store the swap location in the pte.
1380                          * See handle_pte_fault() ...
1381                          */
1382                         if (swap_duplicate(entry) < 0) {
1383                                 set_pte_at(mm, address, pte, pteval);
1384                                 ret = SWAP_FAIL;
1385                                 goto out_unmap;
1386                         }
1387                         if (list_empty(&mm->mmlist)) {
1388                                 spin_lock(&mmlist_lock);
1389                                 if (list_empty(&mm->mmlist))
1390                                         list_add(&mm->mmlist, &init_mm.mmlist);
1391                                 spin_unlock(&mmlist_lock);
1392                         }
1393                         dec_mm_counter(mm, MM_ANONPAGES);
1394                         inc_mm_counter(mm, MM_SWAPENTS);
1395                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
1396                         /*
1397                          * Store the pfn of the page in a special migration
1398                          * pte. do_swap_page() will wait until the migration
1399                          * pte is removed and then restart fault handling.
1400                          */
1401                         BUG_ON(!(flags & TTU_MIGRATION));
1402                         entry = make_migration_entry(page, pte_write(pteval));
1403                 }
1404                 swp_pte = swp_entry_to_pte(entry);
1405                 if (pte_soft_dirty(pteval))
1406                         swp_pte = pte_swp_mksoft_dirty(swp_pte);
1407                 set_pte_at(mm, address, pte, swp_pte);
1408         } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1409                    (flags & TTU_MIGRATION)) {
1410                 /* Establish migration entry for a file page */
1411                 swp_entry_t entry;
1412                 entry = make_migration_entry(page, pte_write(pteval));
1413                 set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1414         } else
1415                 dec_mm_counter(mm, MM_FILEPAGES);
1416 
1417         page_remove_rmap(page);
1418         page_cache_release(page);
1419 
1420 out_unmap:
1421         pte_unmap_unlock(pte, ptl);
1422         if (ret != SWAP_FAIL && !(flags & TTU_MUNLOCK))
1423                 mmu_notifier_invalidate_page(mm, address);
1424 out:
1425         return ret;
1426 
1427 out_mlock:
1428         pte_unmap_unlock(pte, ptl);
1429 
1430 
1431         /*
1432          * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1433          * unstable result and race. Plus, We can't wait here because
1434          * we now hold anon_vma->rwsem or mapping->i_mmap_rwsem.
1435          * if trylock failed, the page remain in evictable lru and later
1436          * vmscan could retry to move the page to unevictable lru if the
1437          * page is actually mlocked.
1438          */
1439         if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1440                 if (vma->vm_flags & VM_LOCKED) {
1441                         mlock_vma_page(page);
1442                         ret = SWAP_MLOCK;
1443                 }
1444                 up_read(&vma->vm_mm->mmap_sem);
1445         }
1446         return ret;
1447 }
1448 
1449 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1450 {
1451         int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1452 
1453         if (!maybe_stack)
1454                 return false;
1455 
1456         if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1457                                                 VM_STACK_INCOMPLETE_SETUP)
1458                 return true;
1459 
1460         return false;
1461 }
1462 
1463 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1464 {
1465         return is_vma_temporary_stack(vma);
1466 }
1467 
1468 static int page_not_mapped(struct page *page)
1469 {
1470         return !page_mapped(page);
1471 };
1472 
1473 /**
1474  * try_to_unmap - try to remove all page table mappings to a page
1475  * @page: the page to get unmapped
1476  * @flags: action and flags
1477  *
1478  * Tries to remove all the page table entries which are mapping this
1479  * page, used in the pageout path.  Caller must hold the page lock.
1480  * Return values are:
1481  *
1482  * SWAP_SUCCESS - we succeeded in removing all mappings
1483  * SWAP_AGAIN   - we missed a mapping, try again later
1484  * SWAP_FAIL    - the page is unswappable
1485  * SWAP_MLOCK   - page is mlocked.
1486  */
1487 int try_to_unmap(struct page *page, enum ttu_flags flags)
1488 {
1489         int ret;
1490         struct rmap_walk_control rwc = {
1491                 .rmap_one = try_to_unmap_one,
1492                 .arg = (void *)flags,
1493                 .done = page_not_mapped,
1494                 .anon_lock = page_lock_anon_vma_read,
1495         };
1496 
1497         VM_BUG_ON_PAGE(!PageHuge(page) && PageTransHuge(page), page);
1498 
1499         /*
1500          * During exec, a temporary VMA is setup and later moved.
1501          * The VMA is moved under the anon_vma lock but not the
1502          * page tables leading to a race where migration cannot
1503          * find the migration ptes. Rather than increasing the
1504          * locking requirements of exec(), migration skips
1505          * temporary VMAs until after exec() completes.
1506          */
1507         if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
1508                 rwc.invalid_vma = invalid_migration_vma;
1509 
1510         ret = rmap_walk(page, &rwc);
1511 
1512         if (ret != SWAP_MLOCK && !page_mapped(page))
1513                 ret = SWAP_SUCCESS;
1514         return ret;
1515 }
1516 
1517 /**
1518  * try_to_munlock - try to munlock a page
1519  * @page: the page to be munlocked
1520  *
1521  * Called from munlock code.  Checks all of the VMAs mapping the page
1522  * to make sure nobody else has this page mlocked. The page will be
1523  * returned with PG_mlocked cleared if no other vmas have it mlocked.
1524  *
1525  * Return values are:
1526  *
1527  * SWAP_AGAIN   - no vma is holding page mlocked, or,
1528  * SWAP_AGAIN   - page mapped in mlocked vma -- couldn't acquire mmap sem
1529  * SWAP_FAIL    - page cannot be located at present
1530  * SWAP_MLOCK   - page is now mlocked.
1531  */
1532 int try_to_munlock(struct page *page)
1533 {
1534         int ret;
1535         struct rmap_walk_control rwc = {
1536                 .rmap_one = try_to_unmap_one,
1537                 .arg = (void *)TTU_MUNLOCK,
1538                 .done = page_not_mapped,
1539                 .anon_lock = page_lock_anon_vma_read,
1540 
1541         };
1542 
1543         VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1544 
1545         ret = rmap_walk(page, &rwc);
1546         return ret;
1547 }
1548 
1549 void __put_anon_vma(struct anon_vma *anon_vma)
1550 {
1551         struct anon_vma *root = anon_vma->root;
1552 
1553         anon_vma_free(anon_vma);
1554         if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1555                 anon_vma_free(root);
1556 }
1557 
1558 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1559                                         struct rmap_walk_control *rwc)
1560 {
1561         struct anon_vma *anon_vma;
1562 
1563         if (rwc->anon_lock)
1564                 return rwc->anon_lock(page);
1565 
1566         /*
1567          * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1568          * because that depends on page_mapped(); but not all its usages
1569          * are holding mmap_sem. Users without mmap_sem are required to
1570          * take a reference count to prevent the anon_vma disappearing
1571          */
1572         anon_vma = page_anon_vma(page);
1573         if (!anon_vma)
1574                 return NULL;
1575 
1576         anon_vma_lock_read(anon_vma);
1577         return anon_vma;
1578 }
1579 
1580 /*
1581  * rmap_walk_anon - do something to anonymous page using the object-based
1582  * rmap method
1583  * @page: the page to be handled
1584  * @rwc: control variable according to each walk type
1585  *
1586  * Find all the mappings of a page using the mapping pointer and the vma chains
1587  * contained in the anon_vma struct it points to.
1588  *
1589  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1590  * where the page was found will be held for write.  So, we won't recheck
1591  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1592  * LOCKED.
1593  */
1594 static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc)
1595 {
1596         struct anon_vma *anon_vma;
1597         pgoff_t pgoff;
1598         struct anon_vma_chain *avc;
1599         int ret = SWAP_AGAIN;
1600 
1601         anon_vma = rmap_walk_anon_lock(page, rwc);
1602         if (!anon_vma)
1603                 return ret;
1604 
1605         pgoff = page_to_pgoff(page);
1606         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1607                 struct vm_area_struct *vma = avc->vma;
1608                 unsigned long address = vma_address(page, vma);
1609 
1610                 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1611                         continue;
1612 
1613                 ret = rwc->rmap_one(page, vma, address, rwc->arg);
1614                 if (ret != SWAP_AGAIN)
1615                         break;
1616                 if (rwc->done && rwc->done(page))
1617                         break;
1618         }
1619         anon_vma_unlock_read(anon_vma);
1620         return ret;
1621 }
1622 
1623 /*
1624  * rmap_walk_file - do something to file page using the object-based rmap method
1625  * @page: the page to be handled
1626  * @rwc: control variable according to each walk type
1627  *
1628  * Find all the mappings of a page using the mapping pointer and the vma chains
1629  * contained in the address_space struct it points to.
1630  *
1631  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1632  * where the page was found will be held for write.  So, we won't recheck
1633  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1634  * LOCKED.
1635  */
1636 static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc)
1637 {
1638         struct address_space *mapping = page->mapping;
1639         pgoff_t pgoff;
1640         struct vm_area_struct *vma;
1641         int ret = SWAP_AGAIN;
1642 
1643         /*
1644          * The page lock not only makes sure that page->mapping cannot
1645          * suddenly be NULLified by truncation, it makes sure that the
1646          * structure at mapping cannot be freed and reused yet,
1647          * so we can safely take mapping->i_mmap_rwsem.
1648          */
1649         VM_BUG_ON_PAGE(!PageLocked(page), page);
1650 
1651         if (!mapping)
1652                 return ret;
1653 
1654         pgoff = page_to_pgoff(page);
1655         i_mmap_lock_read(mapping);
1656         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1657                 unsigned long address = vma_address(page, vma);
1658 
1659                 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1660                         continue;
1661 
1662                 ret = rwc->rmap_one(page, vma, address, rwc->arg);
1663                 if (ret != SWAP_AGAIN)
1664                         goto done;
1665                 if (rwc->done && rwc->done(page))
1666                         goto done;
1667         }
1668 
1669 done:
1670         i_mmap_unlock_read(mapping);
1671         return ret;
1672 }
1673 
1674 int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1675 {
1676         if (unlikely(PageKsm(page)))
1677                 return rmap_walk_ksm(page, rwc);
1678         else if (PageAnon(page))
1679                 return rmap_walk_anon(page, rwc);
1680         else
1681                 return rmap_walk_file(page, rwc);
1682 }
1683 
1684 #ifdef CONFIG_HUGETLB_PAGE
1685 /*
1686  * The following three functions are for anonymous (private mapped) hugepages.
1687  * Unlike common anonymous pages, anonymous hugepages have no accounting code
1688  * and no lru code, because we handle hugepages differently from common pages.
1689  */
1690 static void __hugepage_set_anon_rmap(struct page *page,
1691         struct vm_area_struct *vma, unsigned long address, int exclusive)
1692 {
1693         struct anon_vma *anon_vma = vma->anon_vma;
1694 
1695         BUG_ON(!anon_vma);
1696 
1697         if (PageAnon(page))
1698                 return;
1699         if (!exclusive)
1700                 anon_vma = anon_vma->root;
1701 
1702         anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1703         page->mapping = (struct address_space *) anon_vma;
1704         page->index = linear_page_index(vma, address);
1705 }
1706 
1707 void hugepage_add_anon_rmap(struct page *page,
1708                             struct vm_area_struct *vma, unsigned long address)
1709 {
1710         struct anon_vma *anon_vma = vma->anon_vma;
1711         int first;
1712 
1713         BUG_ON(!PageLocked(page));
1714         BUG_ON(!anon_vma);
1715         /* address might be in next vma when migration races vma_adjust */
1716         first = atomic_inc_and_test(&page->_mapcount);
1717         if (first)
1718                 __hugepage_set_anon_rmap(page, vma, address, 0);
1719 }
1720 
1721 void hugepage_add_new_anon_rmap(struct page *page,
1722                         struct vm_area_struct *vma, unsigned long address)
1723 {
1724         BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1725         atomic_set(&page->_mapcount, 0);
1726         __hugepage_set_anon_rmap(page, vma, address, 1);
1727 }
1728 #endif /* CONFIG_HUGETLB_PAGE */
1729 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp