Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
ksm.c
Go to the documentation of this file.
1 /*
2  * Memory merging support.
3  *
4  * This code enables dynamic sharing of identical pages found in different
5  * memory areas, even if they are not shared by fork()
6  *
7  * Copyright (C) 2008-2009 Red Hat, Inc.
8  * Authors:
9  * Izik Eidus
10  * Andrea Arcangeli
11  * Chris Wright
12  * Hugh Dickins
13  *
14  * This work is licensed under the terms of the GNU GPL, version 2.
15  */
16 
17 #include <linux/errno.h>
18 #include <linux/mm.h>
19 #include <linux/fs.h>
20 #include <linux/mman.h>
21 #include <linux/sched.h>
22 #include <linux/rwsem.h>
23 #include <linux/pagemap.h>
24 #include <linux/rmap.h>
25 #include <linux/spinlock.h>
26 #include <linux/jhash.h>
27 #include <linux/delay.h>
28 #include <linux/kthread.h>
29 #include <linux/wait.h>
30 #include <linux/slab.h>
31 #include <linux/rbtree.h>
32 #include <linux/memory.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/swap.h>
35 #include <linux/ksm.h>
36 #include <linux/hash.h>
37 #include <linux/freezer.h>
38 #include <linux/oom.h>
39 
40 #include <asm/tlbflush.h>
41 #include "internal.h"
42 
43 /*
44  * A few notes about the KSM scanning process,
45  * to make it easier to understand the data structures below:
46  *
47  * In order to reduce excessive scanning, KSM sorts the memory pages by their
48  * contents into a data structure that holds pointers to the pages' locations.
49  *
50  * Since the contents of the pages may change at any moment, KSM cannot just
51  * insert the pages into a normal sorted tree and expect it to find anything.
52  * Therefore KSM uses two data structures - the stable and the unstable tree.
53  *
54  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
55  * by their contents. Because each such page is write-protected, searching on
56  * this tree is fully assured to be working (except when pages are unmapped),
57  * and therefore this tree is called the stable tree.
58  *
59  * In addition to the stable tree, KSM uses a second data structure called the
60  * unstable tree: this tree holds pointers to pages which have been found to
61  * be "unchanged for a period of time". The unstable tree sorts these pages
62  * by their contents, but since they are not write-protected, KSM cannot rely
63  * upon the unstable tree to work correctly - the unstable tree is liable to
64  * be corrupted as its contents are modified, and so it is called unstable.
65  *
66  * KSM solves this problem by several techniques:
67  *
68  * 1) The unstable tree is flushed every time KSM completes scanning all
69  * memory areas, and then the tree is rebuilt again from the beginning.
70  * 2) KSM will only insert into the unstable tree, pages whose hash value
71  * has not changed since the previous scan of all memory areas.
72  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
73  * colors of the nodes and not on their contents, assuring that even when
74  * the tree gets "corrupted" it won't get out of balance, so scanning time
75  * remains the same (also, searching and inserting nodes in an rbtree uses
76  * the same algorithm, so we have no overhead when we flush and rebuild).
77  * 4) KSM never flushes the stable tree, which means that even if it were to
78  * take 10 attempts to find a page in the unstable tree, once it is found,
79  * it is secured in the stable tree. (When we scan a new page, we first
80  * compare it against the stable tree, and then against the unstable tree.)
81  */
82 
90 struct mm_slot {
91  struct hlist_node link;
94  struct mm_struct *mm;
95 };
96 
106 struct ksm_scan {
107  struct mm_slot *mm_slot;
108  unsigned long address;
110  unsigned long seqnr;
111 };
112 
119 struct stable_node {
120  struct rb_node node;
122  unsigned long kpfn;
123 };
124 
136 struct rmap_item {
138  struct anon_vma *anon_vma; /* when stable */
139  struct mm_struct *mm;
140  unsigned long address; /* + low bits used for flags below */
141  unsigned int oldchecksum; /* when unstable */
142  union {
143  struct rb_node node; /* when node of unstable tree */
144  struct { /* when listed from stable tree */
145  struct stable_node *head;
147  };
148  };
149 };
150 
151 #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
152 #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
153 #define STABLE_FLAG 0x200 /* is listed from the stable tree */
154 
155 /* The stable and unstable tree heads */
156 static struct rb_root root_stable_tree = RB_ROOT;
157 static struct rb_root root_unstable_tree = RB_ROOT;
158 
159 #define MM_SLOTS_HASH_SHIFT 10
160 #define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT)
161 static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS];
162 
163 static struct mm_slot ksm_mm_head = {
164  .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
165 };
166 static struct ksm_scan ksm_scan = {
167  .mm_slot = &ksm_mm_head,
168 };
169 
170 static struct kmem_cache *rmap_item_cache;
171 static struct kmem_cache *stable_node_cache;
172 static struct kmem_cache *mm_slot_cache;
173 
174 /* The number of nodes in the stable tree */
175 static unsigned long ksm_pages_shared;
176 
177 /* The number of page slots additionally sharing those nodes */
178 static unsigned long ksm_pages_sharing;
179 
180 /* The number of nodes in the unstable tree */
181 static unsigned long ksm_pages_unshared;
182 
183 /* The number of rmap_items in use: to calculate pages_volatile */
184 static unsigned long ksm_rmap_items;
185 
186 /* Number of pages ksmd should scan in one batch */
187 static unsigned int ksm_thread_pages_to_scan = 100;
188 
189 /* Milliseconds ksmd should sleep between batches */
190 static unsigned int ksm_thread_sleep_millisecs = 20;
191 
192 #define KSM_RUN_STOP 0
193 #define KSM_RUN_MERGE 1
194 #define KSM_RUN_UNMERGE 2
195 static unsigned int ksm_run = KSM_RUN_STOP;
196 
197 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
198 static DEFINE_MUTEX(ksm_thread_mutex);
199 static DEFINE_SPINLOCK(ksm_mmlist_lock);
200 
201 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
202  sizeof(struct __struct), __alignof__(struct __struct),\
203  (__flags), NULL)
204 
205 static int __init ksm_slab_init(void)
206 {
207  rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
208  if (!rmap_item_cache)
209  goto out;
210 
211  stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
212  if (!stable_node_cache)
213  goto out_free1;
214 
215  mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
216  if (!mm_slot_cache)
217  goto out_free2;
218 
219  return 0;
220 
221 out_free2:
222  kmem_cache_destroy(stable_node_cache);
223 out_free1:
224  kmem_cache_destroy(rmap_item_cache);
225 out:
226  return -ENOMEM;
227 }
228 
229 static void __init ksm_slab_free(void)
230 {
231  kmem_cache_destroy(mm_slot_cache);
232  kmem_cache_destroy(stable_node_cache);
233  kmem_cache_destroy(rmap_item_cache);
234  mm_slot_cache = NULL;
235 }
236 
237 static inline struct rmap_item *alloc_rmap_item(void)
238 {
239  struct rmap_item *rmap_item;
240 
241  rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
242  if (rmap_item)
243  ksm_rmap_items++;
244  return rmap_item;
245 }
246 
247 static inline void free_rmap_item(struct rmap_item *rmap_item)
248 {
249  ksm_rmap_items--;
250  rmap_item->mm = NULL; /* debug safety */
251  kmem_cache_free(rmap_item_cache, rmap_item);
252 }
253 
254 static inline struct stable_node *alloc_stable_node(void)
255 {
256  return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
257 }
258 
259 static inline void free_stable_node(struct stable_node *stable_node)
260 {
261  kmem_cache_free(stable_node_cache, stable_node);
262 }
263 
264 static inline struct mm_slot *alloc_mm_slot(void)
265 {
266  if (!mm_slot_cache) /* initialization failed */
267  return NULL;
268  return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
269 }
270 
271 static inline void free_mm_slot(struct mm_slot *mm_slot)
272 {
273  kmem_cache_free(mm_slot_cache, mm_slot);
274 }
275 
276 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
277 {
278  struct mm_slot *mm_slot;
279  struct hlist_head *bucket;
280  struct hlist_node *node;
281 
282  bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
283  hlist_for_each_entry(mm_slot, node, bucket, link) {
284  if (mm == mm_slot->mm)
285  return mm_slot;
286  }
287  return NULL;
288 }
289 
290 static void insert_to_mm_slots_hash(struct mm_struct *mm,
291  struct mm_slot *mm_slot)
292 {
293  struct hlist_head *bucket;
294 
295  bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)];
296  mm_slot->mm = mm;
297  hlist_add_head(&mm_slot->link, bucket);
298 }
299 
300 static inline int in_stable_tree(struct rmap_item *rmap_item)
301 {
302  return rmap_item->address & STABLE_FLAG;
303 }
304 
305 /*
306  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
307  * page tables after it has passed through ksm_exit() - which, if necessary,
308  * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
309  * a special flag: they can just back out as soon as mm_users goes to zero.
310  * ksm_test_exit() is used throughout to make this test for exit: in some
311  * places for correctness, in some places just to avoid unnecessary work.
312  */
313 static inline bool ksm_test_exit(struct mm_struct *mm)
314 {
315  return atomic_read(&mm->mm_users) == 0;
316 }
317 
318 /*
319  * We use break_ksm to break COW on a ksm page: it's a stripped down
320  *
321  * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
322  * put_page(page);
323  *
324  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
325  * in case the application has unmapped and remapped mm,addr meanwhile.
326  * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
327  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
328  */
329 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
330 {
331  struct page *page;
332  int ret = 0;
333 
334  do {
335  cond_resched();
336  page = follow_page(vma, addr, FOLL_GET);
337  if (IS_ERR_OR_NULL(page))
338  break;
339  if (PageKsm(page))
340  ret = handle_mm_fault(vma->vm_mm, vma, addr,
341  FAULT_FLAG_WRITE);
342  else
343  ret = VM_FAULT_WRITE;
344  put_page(page);
345  } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
346  /*
347  * We must loop because handle_mm_fault() may back out if there's
348  * any difficulty e.g. if pte accessed bit gets updated concurrently.
349  *
350  * VM_FAULT_WRITE is what we have been hoping for: it indicates that
351  * COW has been broken, even if the vma does not permit VM_WRITE;
352  * but note that a concurrent fault might break PageKsm for us.
353  *
354  * VM_FAULT_SIGBUS could occur if we race with truncation of the
355  * backing file, which also invalidates anonymous pages: that's
356  * okay, that truncation will have unmapped the PageKsm for us.
357  *
358  * VM_FAULT_OOM: at the time of writing (late July 2009), setting
359  * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
360  * current task has TIF_MEMDIE set, and will be OOM killed on return
361  * to user; and ksmd, having no mm, would never be chosen for that.
362  *
363  * But if the mm is in a limited mem_cgroup, then the fault may fail
364  * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
365  * even ksmd can fail in this way - though it's usually breaking ksm
366  * just to undo a merge it made a moment before, so unlikely to oom.
367  *
368  * That's a pity: we might therefore have more kernel pages allocated
369  * than we're counting as nodes in the stable tree; but ksm_do_scan
370  * will retry to break_cow on each pass, so should recover the page
371  * in due course. The important thing is to not let VM_MERGEABLE
372  * be cleared while any such pages might remain in the area.
373  */
374  return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
375 }
376 
377 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
378  unsigned long addr)
379 {
380  struct vm_area_struct *vma;
381  if (ksm_test_exit(mm))
382  return NULL;
383  vma = find_vma(mm, addr);
384  if (!vma || vma->vm_start > addr)
385  return NULL;
386  if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
387  return NULL;
388  return vma;
389 }
390 
391 static void break_cow(struct rmap_item *rmap_item)
392 {
393  struct mm_struct *mm = rmap_item->mm;
394  unsigned long addr = rmap_item->address;
395  struct vm_area_struct *vma;
396 
397  /*
398  * It is not an accident that whenever we want to break COW
399  * to undo, we also need to drop a reference to the anon_vma.
400  */
401  put_anon_vma(rmap_item->anon_vma);
402 
403  down_read(&mm->mmap_sem);
404  vma = find_mergeable_vma(mm, addr);
405  if (vma)
406  break_ksm(vma, addr);
407  up_read(&mm->mmap_sem);
408 }
409 
410 static struct page *page_trans_compound_anon(struct page *page)
411 {
412  if (PageTransCompound(page)) {
413  struct page *head = compound_trans_head(page);
414  /*
415  * head may actually be splitted and freed from under
416  * us but it's ok here.
417  */
418  if (PageAnon(head))
419  return head;
420  }
421  return NULL;
422 }
423 
424 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
425 {
426  struct mm_struct *mm = rmap_item->mm;
427  unsigned long addr = rmap_item->address;
428  struct vm_area_struct *vma;
429  struct page *page;
430 
431  down_read(&mm->mmap_sem);
432  vma = find_mergeable_vma(mm, addr);
433  if (!vma)
434  goto out;
435 
436  page = follow_page(vma, addr, FOLL_GET);
437  if (IS_ERR_OR_NULL(page))
438  goto out;
439  if (PageAnon(page) || page_trans_compound_anon(page)) {
440  flush_anon_page(vma, page, addr);
441  flush_dcache_page(page);
442  } else {
443  put_page(page);
444 out: page = NULL;
445  }
446  up_read(&mm->mmap_sem);
447  return page;
448 }
449 
450 static void remove_node_from_stable_tree(struct stable_node *stable_node)
451 {
452  struct rmap_item *rmap_item;
453  struct hlist_node *hlist;
454 
455  hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
456  if (rmap_item->hlist.next)
457  ksm_pages_sharing--;
458  else
459  ksm_pages_shared--;
460  put_anon_vma(rmap_item->anon_vma);
461  rmap_item->address &= PAGE_MASK;
462  cond_resched();
463  }
464 
465  rb_erase(&stable_node->node, &root_stable_tree);
466  free_stable_node(stable_node);
467 }
468 
469 /*
470  * get_ksm_page: checks if the page indicated by the stable node
471  * is still its ksm page, despite having held no reference to it.
472  * In which case we can trust the content of the page, and it
473  * returns the gotten page; but if the page has now been zapped,
474  * remove the stale node from the stable tree and return NULL.
475  *
476  * You would expect the stable_node to hold a reference to the ksm page.
477  * But if it increments the page's count, swapping out has to wait for
478  * ksmd to come around again before it can free the page, which may take
479  * seconds or even minutes: much too unresponsive. So instead we use a
480  * "keyhole reference": access to the ksm page from the stable node peeps
481  * out through its keyhole to see if that page still holds the right key,
482  * pointing back to this stable node. This relies on freeing a PageAnon
483  * page to reset its page->mapping to NULL, and relies on no other use of
484  * a page to put something that might look like our key in page->mapping.
485  *
486  * include/linux/pagemap.h page_cache_get_speculative() is a good reference,
487  * but this is different - made simpler by ksm_thread_mutex being held, but
488  * interesting for assuming that no other use of the struct page could ever
489  * put our expected_mapping into page->mapping (or a field of the union which
490  * coincides with page->mapping). The RCU calls are not for KSM at all, but
491  * to keep the page_count protocol described with page_cache_get_speculative.
492  *
493  * Note: it is possible that get_ksm_page() will return NULL one moment,
494  * then page the next, if the page is in between page_freeze_refs() and
495  * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page
496  * is on its way to being freed; but it is an anomaly to bear in mind.
497  */
498 static struct page *get_ksm_page(struct stable_node *stable_node)
499 {
500  struct page *page;
501  void *expected_mapping;
502 
503  page = pfn_to_page(stable_node->kpfn);
504  expected_mapping = (void *)stable_node +
505  (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
506  rcu_read_lock();
507  if (page->mapping != expected_mapping)
508  goto stale;
509  if (!get_page_unless_zero(page))
510  goto stale;
511  if (page->mapping != expected_mapping) {
512  put_page(page);
513  goto stale;
514  }
515  rcu_read_unlock();
516  return page;
517 stale:
518  rcu_read_unlock();
519  remove_node_from_stable_tree(stable_node);
520  return NULL;
521 }
522 
523 /*
524  * Removing rmap_item from stable or unstable tree.
525  * This function will clean the information from the stable/unstable tree.
526  */
527 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
528 {
529  if (rmap_item->address & STABLE_FLAG) {
530  struct stable_node *stable_node;
531  struct page *page;
532 
533  stable_node = rmap_item->head;
534  page = get_ksm_page(stable_node);
535  if (!page)
536  goto out;
537 
538  lock_page(page);
539  hlist_del(&rmap_item->hlist);
540  unlock_page(page);
541  put_page(page);
542 
543  if (stable_node->hlist.first)
544  ksm_pages_sharing--;
545  else
546  ksm_pages_shared--;
547 
548  put_anon_vma(rmap_item->anon_vma);
549  rmap_item->address &= PAGE_MASK;
550 
551  } else if (rmap_item->address & UNSTABLE_FLAG) {
552  unsigned char age;
553  /*
554  * Usually ksmd can and must skip the rb_erase, because
555  * root_unstable_tree was already reset to RB_ROOT.
556  * But be careful when an mm is exiting: do the rb_erase
557  * if this rmap_item was inserted by this scan, rather
558  * than left over from before.
559  */
560  age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
561  BUG_ON(age > 1);
562  if (!age)
563  rb_erase(&rmap_item->node, &root_unstable_tree);
564 
565  ksm_pages_unshared--;
566  rmap_item->address &= PAGE_MASK;
567  }
568 out:
569  cond_resched(); /* we're called from many long loops */
570 }
571 
572 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
573  struct rmap_item **rmap_list)
574 {
575  while (*rmap_list) {
576  struct rmap_item *rmap_item = *rmap_list;
577  *rmap_list = rmap_item->rmap_list;
578  remove_rmap_item_from_tree(rmap_item);
579  free_rmap_item(rmap_item);
580  }
581 }
582 
583 /*
584  * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
585  * than check every pte of a given vma, the locking doesn't quite work for
586  * that - an rmap_item is assigned to the stable tree after inserting ksm
587  * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
588  * rmap_items from parent to child at fork time (so as not to waste time
589  * if exit comes before the next scan reaches it).
590  *
591  * Similarly, although we'd like to remove rmap_items (so updating counts
592  * and freeing memory) when unmerging an area, it's easier to leave that
593  * to the next pass of ksmd - consider, for example, how ksmd might be
594  * in cmp_and_merge_page on one of the rmap_items we would be removing.
595  */
596 static int unmerge_ksm_pages(struct vm_area_struct *vma,
597  unsigned long start, unsigned long end)
598 {
599  unsigned long addr;
600  int err = 0;
601 
602  for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
603  if (ksm_test_exit(vma->vm_mm))
604  break;
605  if (signal_pending(current))
606  err = -ERESTARTSYS;
607  else
608  err = break_ksm(vma, addr);
609  }
610  return err;
611 }
612 
613 #ifdef CONFIG_SYSFS
614 /*
615  * Only called through the sysfs control interface:
616  */
617 static int unmerge_and_remove_all_rmap_items(void)
618 {
619  struct mm_slot *mm_slot;
620  struct mm_struct *mm;
621  struct vm_area_struct *vma;
622  int err = 0;
623 
624  spin_lock(&ksm_mmlist_lock);
625  ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
626  struct mm_slot, mm_list);
627  spin_unlock(&ksm_mmlist_lock);
628 
629  for (mm_slot = ksm_scan.mm_slot;
630  mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
631  mm = mm_slot->mm;
632  down_read(&mm->mmap_sem);
633  for (vma = mm->mmap; vma; vma = vma->vm_next) {
634  if (ksm_test_exit(mm))
635  break;
636  if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
637  continue;
638  err = unmerge_ksm_pages(vma,
639  vma->vm_start, vma->vm_end);
640  if (err)
641  goto error;
642  }
643 
644  remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
645 
646  spin_lock(&ksm_mmlist_lock);
647  ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
648  struct mm_slot, mm_list);
649  if (ksm_test_exit(mm)) {
650  hlist_del(&mm_slot->link);
651  list_del(&mm_slot->mm_list);
652  spin_unlock(&ksm_mmlist_lock);
653 
654  free_mm_slot(mm_slot);
656  up_read(&mm->mmap_sem);
657  mmdrop(mm);
658  } else {
659  spin_unlock(&ksm_mmlist_lock);
660  up_read(&mm->mmap_sem);
661  }
662  }
663 
664  ksm_scan.seqnr = 0;
665  return 0;
666 
667 error:
668  up_read(&mm->mmap_sem);
669  spin_lock(&ksm_mmlist_lock);
670  ksm_scan.mm_slot = &ksm_mm_head;
671  spin_unlock(&ksm_mmlist_lock);
672  return err;
673 }
674 #endif /* CONFIG_SYSFS */
675 
676 static u32 calc_checksum(struct page *page)
677 {
678  u32 checksum;
679  void *addr = kmap_atomic(page);
680  checksum = jhash2(addr, PAGE_SIZE / 4, 17);
681  kunmap_atomic(addr);
682  return checksum;
683 }
684 
685 static int memcmp_pages(struct page *page1, struct page *page2)
686 {
687  char *addr1, *addr2;
688  int ret;
689 
690  addr1 = kmap_atomic(page1);
691  addr2 = kmap_atomic(page2);
692  ret = memcmp(addr1, addr2, PAGE_SIZE);
693  kunmap_atomic(addr2);
694  kunmap_atomic(addr1);
695  return ret;
696 }
697 
698 static inline int pages_identical(struct page *page1, struct page *page2)
699 {
700  return !memcmp_pages(page1, page2);
701 }
702 
703 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
704  pte_t *orig_pte)
705 {
706  struct mm_struct *mm = vma->vm_mm;
707  unsigned long addr;
708  pte_t *ptep;
709  spinlock_t *ptl;
710  int swapped;
711  int err = -EFAULT;
712  unsigned long mmun_start; /* For mmu_notifiers */
713  unsigned long mmun_end; /* For mmu_notifiers */
714 
715  addr = page_address_in_vma(page, vma);
716  if (addr == -EFAULT)
717  goto out;
718 
719  BUG_ON(PageTransCompound(page));
720 
721  mmun_start = addr;
722  mmun_end = addr + PAGE_SIZE;
723  mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
724 
725  ptep = page_check_address(page, mm, addr, &ptl, 0);
726  if (!ptep)
727  goto out_mn;
728 
729  if (pte_write(*ptep) || pte_dirty(*ptep)) {
730  pte_t entry;
731 
732  swapped = PageSwapCache(page);
733  flush_cache_page(vma, addr, page_to_pfn(page));
734  /*
735  * Ok this is tricky, when get_user_pages_fast() run it doesn't
736  * take any lock, therefore the check that we are going to make
737  * with the pagecount against the mapcount is racey and
738  * O_DIRECT can happen right after the check.
739  * So we clear the pte and flush the tlb before the check
740  * this assure us that no O_DIRECT can happen after the check
741  * or in the middle of the check.
742  */
743  entry = ptep_clear_flush(vma, addr, ptep);
744  /*
745  * Check that no O_DIRECT or similar I/O is in progress on the
746  * page
747  */
748  if (page_mapcount(page) + 1 + swapped != page_count(page)) {
749  set_pte_at(mm, addr, ptep, entry);
750  goto out_unlock;
751  }
752  if (pte_dirty(entry))
753  set_page_dirty(page);
754  entry = pte_mkclean(pte_wrprotect(entry));
755  set_pte_at_notify(mm, addr, ptep, entry);
756  }
757  *orig_pte = *ptep;
758  err = 0;
759 
760 out_unlock:
761  pte_unmap_unlock(ptep, ptl);
762 out_mn:
763  mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
764 out:
765  return err;
766 }
767 
777 static int replace_page(struct vm_area_struct *vma, struct page *page,
778  struct page *kpage, pte_t orig_pte)
779 {
780  struct mm_struct *mm = vma->vm_mm;
781  pgd_t *pgd;
782  pud_t *pud;
783  pmd_t *pmd;
784  pte_t *ptep;
785  spinlock_t *ptl;
786  unsigned long addr;
787  int err = -EFAULT;
788  unsigned long mmun_start; /* For mmu_notifiers */
789  unsigned long mmun_end; /* For mmu_notifiers */
790 
791  addr = page_address_in_vma(page, vma);
792  if (addr == -EFAULT)
793  goto out;
794 
795  pgd = pgd_offset(mm, addr);
796  if (!pgd_present(*pgd))
797  goto out;
798 
799  pud = pud_offset(pgd, addr);
800  if (!pud_present(*pud))
801  goto out;
802 
803  pmd = pmd_offset(pud, addr);
804  BUG_ON(pmd_trans_huge(*pmd));
805  if (!pmd_present(*pmd))
806  goto out;
807 
808  mmun_start = addr;
809  mmun_end = addr + PAGE_SIZE;
810  mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
811 
812  ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
813  if (!pte_same(*ptep, orig_pte)) {
814  pte_unmap_unlock(ptep, ptl);
815  goto out_mn;
816  }
817 
818  get_page(kpage);
819  page_add_anon_rmap(kpage, vma, addr);
820 
821  flush_cache_page(vma, addr, pte_pfn(*ptep));
822  ptep_clear_flush(vma, addr, ptep);
823  set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
824 
825  page_remove_rmap(page);
826  if (!page_mapped(page))
827  try_to_free_swap(page);
828  put_page(page);
829 
830  pte_unmap_unlock(ptep, ptl);
831  err = 0;
832 out_mn:
833  mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
834 out:
835  return err;
836 }
837 
838 static int page_trans_compound_anon_split(struct page *page)
839 {
840  int ret = 0;
841  struct page *transhuge_head = page_trans_compound_anon(page);
842  if (transhuge_head) {
843  /* Get the reference on the head to split it. */
844  if (get_page_unless_zero(transhuge_head)) {
845  /*
846  * Recheck we got the reference while the head
847  * was still anonymous.
848  */
849  if (PageAnon(transhuge_head))
850  ret = split_huge_page(transhuge_head);
851  else
852  /*
853  * Retry later if split_huge_page run
854  * from under us.
855  */
856  ret = 1;
857  put_page(transhuge_head);
858  } else
859  /* Retry later if split_huge_page run from under us. */
860  ret = 1;
861  }
862  return ret;
863 }
864 
865 /*
866  * try_to_merge_one_page - take two pages and merge them into one
867  * @vma: the vma that holds the pte pointing to page
868  * @page: the PageAnon page that we want to replace with kpage
869  * @kpage: the PageKsm page that we want to map instead of page,
870  * or NULL the first time when we want to use page as kpage.
871  *
872  * This function returns 0 if the pages were merged, -EFAULT otherwise.
873  */
874 static int try_to_merge_one_page(struct vm_area_struct *vma,
875  struct page *page, struct page *kpage)
876 {
877  pte_t orig_pte = __pte(0);
878  int err = -EFAULT;
879 
880  if (page == kpage) /* ksm page forked */
881  return 0;
882 
883  if (!(vma->vm_flags & VM_MERGEABLE))
884  goto out;
885  if (PageTransCompound(page) && page_trans_compound_anon_split(page))
886  goto out;
887  BUG_ON(PageTransCompound(page));
888  if (!PageAnon(page))
889  goto out;
890 
891  /*
892  * We need the page lock to read a stable PageSwapCache in
893  * write_protect_page(). We use trylock_page() instead of
894  * lock_page() because we don't want to wait here - we
895  * prefer to continue scanning and merging different pages,
896  * then come back to this page when it is unlocked.
897  */
898  if (!trylock_page(page))
899  goto out;
900  /*
901  * If this anonymous page is mapped only here, its pte may need
902  * to be write-protected. If it's mapped elsewhere, all of its
903  * ptes are necessarily already write-protected. But in either
904  * case, we need to lock and check page_count is not raised.
905  */
906  if (write_protect_page(vma, page, &orig_pte) == 0) {
907  if (!kpage) {
908  /*
909  * While we hold page lock, upgrade page from
910  * PageAnon+anon_vma to PageKsm+NULL stable_node:
911  * stable_tree_insert() will update stable_node.
912  */
913  set_page_stable_node(page, NULL);
914  mark_page_accessed(page);
915  err = 0;
916  } else if (pages_identical(page, kpage))
917  err = replace_page(vma, page, kpage, orig_pte);
918  }
919 
920  if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
921  munlock_vma_page(page);
922  if (!PageMlocked(kpage)) {
923  unlock_page(page);
924  lock_page(kpage);
925  mlock_vma_page(kpage);
926  page = kpage; /* for final unlock */
927  }
928  }
929 
930  unlock_page(page);
931 out:
932  return err;
933 }
934 
935 /*
936  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
937  * but no new kernel page is allocated: kpage must already be a ksm page.
938  *
939  * This function returns 0 if the pages were merged, -EFAULT otherwise.
940  */
941 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
942  struct page *page, struct page *kpage)
943 {
944  struct mm_struct *mm = rmap_item->mm;
945  struct vm_area_struct *vma;
946  int err = -EFAULT;
947 
948  down_read(&mm->mmap_sem);
949  if (ksm_test_exit(mm))
950  goto out;
951  vma = find_vma(mm, rmap_item->address);
952  if (!vma || vma->vm_start > rmap_item->address)
953  goto out;
954 
955  err = try_to_merge_one_page(vma, page, kpage);
956  if (err)
957  goto out;
958 
959  /* Must get reference to anon_vma while still holding mmap_sem */
960  rmap_item->anon_vma = vma->anon_vma;
961  get_anon_vma(vma->anon_vma);
962 out:
963  up_read(&mm->mmap_sem);
964  return err;
965 }
966 
967 /*
968  * try_to_merge_two_pages - take two identical pages and prepare them
969  * to be merged into one page.
970  *
971  * This function returns the kpage if we successfully merged two identical
972  * pages into one ksm page, NULL otherwise.
973  *
974  * Note that this function upgrades page to ksm page: if one of the pages
975  * is already a ksm page, try_to_merge_with_ksm_page should be used.
976  */
977 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
978  struct page *page,
979  struct rmap_item *tree_rmap_item,
980  struct page *tree_page)
981 {
982  int err;
983 
984  err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
985  if (!err) {
986  err = try_to_merge_with_ksm_page(tree_rmap_item,
987  tree_page, page);
988  /*
989  * If that fails, we have a ksm page with only one pte
990  * pointing to it: so break it.
991  */
992  if (err)
993  break_cow(rmap_item);
994  }
995  return err ? NULL : page;
996 }
997 
998 /*
999  * stable_tree_search - search for page inside the stable tree
1000  *
1001  * This function checks if there is a page inside the stable tree
1002  * with identical content to the page that we are scanning right now.
1003  *
1004  * This function returns the stable tree node of identical content if found,
1005  * NULL otherwise.
1006  */
1007 static struct page *stable_tree_search(struct page *page)
1008 {
1009  struct rb_node *node = root_stable_tree.rb_node;
1010  struct stable_node *stable_node;
1011 
1012  stable_node = page_stable_node(page);
1013  if (stable_node) { /* ksm page forked */
1014  get_page(page);
1015  return page;
1016  }
1017 
1018  while (node) {
1019  struct page *tree_page;
1020  int ret;
1021 
1022  cond_resched();
1023  stable_node = rb_entry(node, struct stable_node, node);
1024  tree_page = get_ksm_page(stable_node);
1025  if (!tree_page)
1026  return NULL;
1027 
1028  ret = memcmp_pages(page, tree_page);
1029 
1030  if (ret < 0) {
1031  put_page(tree_page);
1032  node = node->rb_left;
1033  } else if (ret > 0) {
1034  put_page(tree_page);
1035  node = node->rb_right;
1036  } else
1037  return tree_page;
1038  }
1039 
1040  return NULL;
1041 }
1042 
1043 /*
1044  * stable_tree_insert - insert rmap_item pointing to new ksm page
1045  * into the stable tree.
1046  *
1047  * This function returns the stable tree node just allocated on success,
1048  * NULL otherwise.
1049  */
1050 static struct stable_node *stable_tree_insert(struct page *kpage)
1051 {
1052  struct rb_node **new = &root_stable_tree.rb_node;
1053  struct rb_node *parent = NULL;
1054  struct stable_node *stable_node;
1055 
1056  while (*new) {
1057  struct page *tree_page;
1058  int ret;
1059 
1060  cond_resched();
1061  stable_node = rb_entry(*new, struct stable_node, node);
1062  tree_page = get_ksm_page(stable_node);
1063  if (!tree_page)
1064  return NULL;
1065 
1066  ret = memcmp_pages(kpage, tree_page);
1067  put_page(tree_page);
1068 
1069  parent = *new;
1070  if (ret < 0)
1071  new = &parent->rb_left;
1072  else if (ret > 0)
1073  new = &parent->rb_right;
1074  else {
1075  /*
1076  * It is not a bug that stable_tree_search() didn't
1077  * find this node: because at that time our page was
1078  * not yet write-protected, so may have changed since.
1079  */
1080  return NULL;
1081  }
1082  }
1083 
1084  stable_node = alloc_stable_node();
1085  if (!stable_node)
1086  return NULL;
1087 
1088  rb_link_node(&stable_node->node, parent, new);
1089  rb_insert_color(&stable_node->node, &root_stable_tree);
1090 
1091  INIT_HLIST_HEAD(&stable_node->hlist);
1092 
1093  stable_node->kpfn = page_to_pfn(kpage);
1094  set_page_stable_node(kpage, stable_node);
1095 
1096  return stable_node;
1097 }
1098 
1099 /*
1100  * unstable_tree_search_insert - search for identical page,
1101  * else insert rmap_item into the unstable tree.
1102  *
1103  * This function searches for a page in the unstable tree identical to the
1104  * page currently being scanned; and if no identical page is found in the
1105  * tree, we insert rmap_item as a new object into the unstable tree.
1106  *
1107  * This function returns pointer to rmap_item found to be identical
1108  * to the currently scanned page, NULL otherwise.
1109  *
1110  * This function does both searching and inserting, because they share
1111  * the same walking algorithm in an rbtree.
1112  */
1113 static
1114 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1115  struct page *page,
1116  struct page **tree_pagep)
1117 
1118 {
1119  struct rb_node **new = &root_unstable_tree.rb_node;
1120  struct rb_node *parent = NULL;
1121 
1122  while (*new) {
1123  struct rmap_item *tree_rmap_item;
1124  struct page *tree_page;
1125  int ret;
1126 
1127  cond_resched();
1128  tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1129  tree_page = get_mergeable_page(tree_rmap_item);
1130  if (IS_ERR_OR_NULL(tree_page))
1131  return NULL;
1132 
1133  /*
1134  * Don't substitute a ksm page for a forked page.
1135  */
1136  if (page == tree_page) {
1137  put_page(tree_page);
1138  return NULL;
1139  }
1140 
1141  ret = memcmp_pages(page, tree_page);
1142 
1143  parent = *new;
1144  if (ret < 0) {
1145  put_page(tree_page);
1146  new = &parent->rb_left;
1147  } else if (ret > 0) {
1148  put_page(tree_page);
1149  new = &parent->rb_right;
1150  } else {
1151  *tree_pagep = tree_page;
1152  return tree_rmap_item;
1153  }
1154  }
1155 
1156  rmap_item->address |= UNSTABLE_FLAG;
1157  rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1158  rb_link_node(&rmap_item->node, parent, new);
1159  rb_insert_color(&rmap_item->node, &root_unstable_tree);
1160 
1161  ksm_pages_unshared++;
1162  return NULL;
1163 }
1164 
1165 /*
1166  * stable_tree_append - add another rmap_item to the linked list of
1167  * rmap_items hanging off a given node of the stable tree, all sharing
1168  * the same ksm page.
1169  */
1170 static void stable_tree_append(struct rmap_item *rmap_item,
1171  struct stable_node *stable_node)
1172 {
1173  rmap_item->head = stable_node;
1174  rmap_item->address |= STABLE_FLAG;
1175  hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1176 
1177  if (rmap_item->hlist.next)
1178  ksm_pages_sharing++;
1179  else
1180  ksm_pages_shared++;
1181 }
1182 
1183 /*
1184  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1185  * if not, compare checksum to previous and if it's the same, see if page can
1186  * be inserted into the unstable tree, or merged with a page already there and
1187  * both transferred to the stable tree.
1188  *
1189  * @page: the page that we are searching identical page to.
1190  * @rmap_item: the reverse mapping into the virtual address of this page
1191  */
1192 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1193 {
1194  struct rmap_item *tree_rmap_item;
1195  struct page *tree_page = NULL;
1196  struct stable_node *stable_node;
1197  struct page *kpage;
1198  unsigned int checksum;
1199  int err;
1200 
1201  remove_rmap_item_from_tree(rmap_item);
1202 
1203  /* We first start with searching the page inside the stable tree */
1204  kpage = stable_tree_search(page);
1205  if (kpage) {
1206  err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1207  if (!err) {
1208  /*
1209  * The page was successfully merged:
1210  * add its rmap_item to the stable tree.
1211  */
1212  lock_page(kpage);
1213  stable_tree_append(rmap_item, page_stable_node(kpage));
1214  unlock_page(kpage);
1215  }
1216  put_page(kpage);
1217  return;
1218  }
1219 
1220  /*
1221  * If the hash value of the page has changed from the last time
1222  * we calculated it, this page is changing frequently: therefore we
1223  * don't want to insert it in the unstable tree, and we don't want
1224  * to waste our time searching for something identical to it there.
1225  */
1226  checksum = calc_checksum(page);
1227  if (rmap_item->oldchecksum != checksum) {
1228  rmap_item->oldchecksum = checksum;
1229  return;
1230  }
1231 
1232  tree_rmap_item =
1233  unstable_tree_search_insert(rmap_item, page, &tree_page);
1234  if (tree_rmap_item) {
1235  kpage = try_to_merge_two_pages(rmap_item, page,
1236  tree_rmap_item, tree_page);
1237  put_page(tree_page);
1238  /*
1239  * As soon as we merge this page, we want to remove the
1240  * rmap_item of the page we have merged with from the unstable
1241  * tree, and insert it instead as new node in the stable tree.
1242  */
1243  if (kpage) {
1244  remove_rmap_item_from_tree(tree_rmap_item);
1245 
1246  lock_page(kpage);
1247  stable_node = stable_tree_insert(kpage);
1248  if (stable_node) {
1249  stable_tree_append(tree_rmap_item, stable_node);
1250  stable_tree_append(rmap_item, stable_node);
1251  }
1252  unlock_page(kpage);
1253 
1254  /*
1255  * If we fail to insert the page into the stable tree,
1256  * we will have 2 virtual addresses that are pointing
1257  * to a ksm page left outside the stable tree,
1258  * in which case we need to break_cow on both.
1259  */
1260  if (!stable_node) {
1261  break_cow(tree_rmap_item);
1262  break_cow(rmap_item);
1263  }
1264  }
1265  }
1266 }
1267 
1268 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1269  struct rmap_item **rmap_list,
1270  unsigned long addr)
1271 {
1272  struct rmap_item *rmap_item;
1273 
1274  while (*rmap_list) {
1275  rmap_item = *rmap_list;
1276  if ((rmap_item->address & PAGE_MASK) == addr)
1277  return rmap_item;
1278  if (rmap_item->address > addr)
1279  break;
1280  *rmap_list = rmap_item->rmap_list;
1281  remove_rmap_item_from_tree(rmap_item);
1282  free_rmap_item(rmap_item);
1283  }
1284 
1285  rmap_item = alloc_rmap_item();
1286  if (rmap_item) {
1287  /* It has already been zeroed */
1288  rmap_item->mm = mm_slot->mm;
1289  rmap_item->address = addr;
1290  rmap_item->rmap_list = *rmap_list;
1291  *rmap_list = rmap_item;
1292  }
1293  return rmap_item;
1294 }
1295 
1296 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1297 {
1298  struct mm_struct *mm;
1299  struct mm_slot *slot;
1300  struct vm_area_struct *vma;
1301  struct rmap_item *rmap_item;
1302 
1303  if (list_empty(&ksm_mm_head.mm_list))
1304  return NULL;
1305 
1306  slot = ksm_scan.mm_slot;
1307  if (slot == &ksm_mm_head) {
1308  /*
1309  * A number of pages can hang around indefinitely on per-cpu
1310  * pagevecs, raised page count preventing write_protect_page
1311  * from merging them. Though it doesn't really matter much,
1312  * it is puzzling to see some stuck in pages_volatile until
1313  * other activity jostles them out, and they also prevented
1314  * LTP's KSM test from succeeding deterministically; so drain
1315  * them here (here rather than on entry to ksm_do_scan(),
1316  * so we don't IPI too often when pages_to_scan is set low).
1317  */
1319 
1320  root_unstable_tree = RB_ROOT;
1321 
1322  spin_lock(&ksm_mmlist_lock);
1323  slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1324  ksm_scan.mm_slot = slot;
1325  spin_unlock(&ksm_mmlist_lock);
1326  /*
1327  * Although we tested list_empty() above, a racing __ksm_exit
1328  * of the last mm on the list may have removed it since then.
1329  */
1330  if (slot == &ksm_mm_head)
1331  return NULL;
1332 next_mm:
1333  ksm_scan.address = 0;
1334  ksm_scan.rmap_list = &slot->rmap_list;
1335  }
1336 
1337  mm = slot->mm;
1338  down_read(&mm->mmap_sem);
1339  if (ksm_test_exit(mm))
1340  vma = NULL;
1341  else
1342  vma = find_vma(mm, ksm_scan.address);
1343 
1344  for (; vma; vma = vma->vm_next) {
1345  if (!(vma->vm_flags & VM_MERGEABLE))
1346  continue;
1347  if (ksm_scan.address < vma->vm_start)
1348  ksm_scan.address = vma->vm_start;
1349  if (!vma->anon_vma)
1350  ksm_scan.address = vma->vm_end;
1351 
1352  while (ksm_scan.address < vma->vm_end) {
1353  if (ksm_test_exit(mm))
1354  break;
1355  *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1356  if (IS_ERR_OR_NULL(*page)) {
1357  ksm_scan.address += PAGE_SIZE;
1358  cond_resched();
1359  continue;
1360  }
1361  if (PageAnon(*page) ||
1362  page_trans_compound_anon(*page)) {
1363  flush_anon_page(vma, *page, ksm_scan.address);
1364  flush_dcache_page(*page);
1365  rmap_item = get_next_rmap_item(slot,
1366  ksm_scan.rmap_list, ksm_scan.address);
1367  if (rmap_item) {
1368  ksm_scan.rmap_list =
1369  &rmap_item->rmap_list;
1370  ksm_scan.address += PAGE_SIZE;
1371  } else
1372  put_page(*page);
1373  up_read(&mm->mmap_sem);
1374  return rmap_item;
1375  }
1376  put_page(*page);
1377  ksm_scan.address += PAGE_SIZE;
1378  cond_resched();
1379  }
1380  }
1381 
1382  if (ksm_test_exit(mm)) {
1383  ksm_scan.address = 0;
1384  ksm_scan.rmap_list = &slot->rmap_list;
1385  }
1386  /*
1387  * Nuke all the rmap_items that are above this current rmap:
1388  * because there were no VM_MERGEABLE vmas with such addresses.
1389  */
1390  remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1391 
1392  spin_lock(&ksm_mmlist_lock);
1393  ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1394  struct mm_slot, mm_list);
1395  if (ksm_scan.address == 0) {
1396  /*
1397  * We've completed a full scan of all vmas, holding mmap_sem
1398  * throughout, and found no VM_MERGEABLE: so do the same as
1399  * __ksm_exit does to remove this mm from all our lists now.
1400  * This applies either when cleaning up after __ksm_exit
1401  * (but beware: we can reach here even before __ksm_exit),
1402  * or when all VM_MERGEABLE areas have been unmapped (and
1403  * mmap_sem then protects against race with MADV_MERGEABLE).
1404  */
1405  hlist_del(&slot->link);
1406  list_del(&slot->mm_list);
1407  spin_unlock(&ksm_mmlist_lock);
1408 
1409  free_mm_slot(slot);
1411  up_read(&mm->mmap_sem);
1412  mmdrop(mm);
1413  } else {
1414  spin_unlock(&ksm_mmlist_lock);
1415  up_read(&mm->mmap_sem);
1416  }
1417 
1418  /* Repeat until we've completed scanning the whole list */
1419  slot = ksm_scan.mm_slot;
1420  if (slot != &ksm_mm_head)
1421  goto next_mm;
1422 
1423  ksm_scan.seqnr++;
1424  return NULL;
1425 }
1426 
1431 static void ksm_do_scan(unsigned int scan_npages)
1432 {
1433  struct rmap_item *rmap_item;
1434  struct page *uninitialized_var(page);
1435 
1436  while (scan_npages-- && likely(!freezing(current))) {
1437  cond_resched();
1438  rmap_item = scan_get_next_rmap_item(&page);
1439  if (!rmap_item)
1440  return;
1441  if (!PageKsm(page) || !in_stable_tree(rmap_item))
1442  cmp_and_merge_page(page, rmap_item);
1443  put_page(page);
1444  }
1445 }
1446 
1447 static int ksmd_should_run(void)
1448 {
1449  return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1450 }
1451 
1452 static int ksm_scan_thread(void *nothing)
1453 {
1454  set_freezable();
1455  set_user_nice(current, 5);
1456 
1457  while (!kthread_should_stop()) {
1458  mutex_lock(&ksm_thread_mutex);
1459  if (ksmd_should_run())
1460  ksm_do_scan(ksm_thread_pages_to_scan);
1461  mutex_unlock(&ksm_thread_mutex);
1462 
1463  try_to_freeze();
1464 
1465  if (ksmd_should_run()) {
1467  msecs_to_jiffies(ksm_thread_sleep_millisecs));
1468  } else {
1469  wait_event_freezable(ksm_thread_wait,
1470  ksmd_should_run() || kthread_should_stop());
1471  }
1472  }
1473  return 0;
1474 }
1475 
1476 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1477  unsigned long end, int advice, unsigned long *vm_flags)
1478 {
1479  struct mm_struct *mm = vma->vm_mm;
1480  int err;
1481 
1482  switch (advice) {
1483  case MADV_MERGEABLE:
1484  /*
1485  * Be somewhat over-protective for now!
1486  */
1487  if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1488  VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1489  VM_HUGETLB | VM_NONLINEAR | VM_MIXEDMAP))
1490  return 0; /* just ignore the advice */
1491 
1492 #ifdef VM_SAO
1493  if (*vm_flags & VM_SAO)
1494  return 0;
1495 #endif
1496 
1497  if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1498  err = __ksm_enter(mm);
1499  if (err)
1500  return err;
1501  }
1502 
1503  *vm_flags |= VM_MERGEABLE;
1504  break;
1505 
1506  case MADV_UNMERGEABLE:
1507  if (!(*vm_flags & VM_MERGEABLE))
1508  return 0; /* just ignore the advice */
1509 
1510  if (vma->anon_vma) {
1511  err = unmerge_ksm_pages(vma, start, end);
1512  if (err)
1513  return err;
1514  }
1515 
1516  *vm_flags &= ~VM_MERGEABLE;
1517  break;
1518  }
1519 
1520  return 0;
1521 }
1522 
1523 int __ksm_enter(struct mm_struct *mm)
1524 {
1525  struct mm_slot *mm_slot;
1526  int needs_wakeup;
1527 
1528  mm_slot = alloc_mm_slot();
1529  if (!mm_slot)
1530  return -ENOMEM;
1531 
1532  /* Check ksm_run too? Would need tighter locking */
1533  needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1534 
1535  spin_lock(&ksm_mmlist_lock);
1536  insert_to_mm_slots_hash(mm, mm_slot);
1537  /*
1538  * Insert just behind the scanning cursor, to let the area settle
1539  * down a little; when fork is followed by immediate exec, we don't
1540  * want ksmd to waste time setting up and tearing down an rmap_list.
1541  */
1542  list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1543  spin_unlock(&ksm_mmlist_lock);
1544 
1546  atomic_inc(&mm->mm_count);
1547 
1548  if (needs_wakeup)
1549  wake_up_interruptible(&ksm_thread_wait);
1550 
1551  return 0;
1552 }
1553 
1554 void __ksm_exit(struct mm_struct *mm)
1555 {
1556  struct mm_slot *mm_slot;
1557  int easy_to_free = 0;
1558 
1559  /*
1560  * This process is exiting: if it's straightforward (as is the
1561  * case when ksmd was never running), free mm_slot immediately.
1562  * But if it's at the cursor or has rmap_items linked to it, use
1563  * mmap_sem to synchronize with any break_cows before pagetables
1564  * are freed, and leave the mm_slot on the list for ksmd to free.
1565  * Beware: ksm may already have noticed it exiting and freed the slot.
1566  */
1567 
1568  spin_lock(&ksm_mmlist_lock);
1569  mm_slot = get_mm_slot(mm);
1570  if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1571  if (!mm_slot->rmap_list) {
1572  hlist_del(&mm_slot->link);
1573  list_del(&mm_slot->mm_list);
1574  easy_to_free = 1;
1575  } else {
1576  list_move(&mm_slot->mm_list,
1577  &ksm_scan.mm_slot->mm_list);
1578  }
1579  }
1580  spin_unlock(&ksm_mmlist_lock);
1581 
1582  if (easy_to_free) {
1583  free_mm_slot(mm_slot);
1585  mmdrop(mm);
1586  } else if (mm_slot) {
1587  down_write(&mm->mmap_sem);
1588  up_write(&mm->mmap_sem);
1589  }
1590 }
1591 
1592 struct page *ksm_does_need_to_copy(struct page *page,
1593  struct vm_area_struct *vma, unsigned long address)
1594 {
1595  struct page *new_page;
1596 
1597  new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1598  if (new_page) {
1599  copy_user_highpage(new_page, page, address, vma);
1600 
1601  SetPageDirty(new_page);
1602  __SetPageUptodate(new_page);
1603  SetPageSwapBacked(new_page);
1604  __set_page_locked(new_page);
1605 
1606  if (!mlocked_vma_newpage(vma, new_page))
1608  else
1609  add_page_to_unevictable_list(new_page);
1610  }
1611 
1612  return new_page;
1613 }
1614 
1615 int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1616  unsigned long *vm_flags)
1617 {
1618  struct stable_node *stable_node;
1619  struct rmap_item *rmap_item;
1620  struct hlist_node *hlist;
1621  unsigned int mapcount = page_mapcount(page);
1622  int referenced = 0;
1623  int search_new_forks = 0;
1624 
1625  VM_BUG_ON(!PageKsm(page));
1626  VM_BUG_ON(!PageLocked(page));
1627 
1628  stable_node = page_stable_node(page);
1629  if (!stable_node)
1630  return 0;
1631 again:
1632  hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1633  struct anon_vma *anon_vma = rmap_item->anon_vma;
1634  struct anon_vma_chain *vmac;
1635  struct vm_area_struct *vma;
1636 
1637  anon_vma_lock(anon_vma);
1638  anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1639  0, ULONG_MAX) {
1640  vma = vmac->vma;
1641  if (rmap_item->address < vma->vm_start ||
1642  rmap_item->address >= vma->vm_end)
1643  continue;
1644  /*
1645  * Initially we examine only the vma which covers this
1646  * rmap_item; but later, if there is still work to do,
1647  * we examine covering vmas in other mms: in case they
1648  * were forked from the original since ksmd passed.
1649  */
1650  if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1651  continue;
1652 
1653  if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1654  continue;
1655 
1656  referenced += page_referenced_one(page, vma,
1657  rmap_item->address, &mapcount, vm_flags);
1658  if (!search_new_forks || !mapcount)
1659  break;
1660  }
1661  anon_vma_unlock(anon_vma);
1662  if (!mapcount)
1663  goto out;
1664  }
1665  if (!search_new_forks++)
1666  goto again;
1667 out:
1668  return referenced;
1669 }
1670 
1671 int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1672 {
1673  struct stable_node *stable_node;
1674  struct hlist_node *hlist;
1675  struct rmap_item *rmap_item;
1676  int ret = SWAP_AGAIN;
1677  int search_new_forks = 0;
1678 
1679  VM_BUG_ON(!PageKsm(page));
1680  VM_BUG_ON(!PageLocked(page));
1681 
1682  stable_node = page_stable_node(page);
1683  if (!stable_node)
1684  return SWAP_FAIL;
1685 again:
1686  hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1687  struct anon_vma *anon_vma = rmap_item->anon_vma;
1688  struct anon_vma_chain *vmac;
1689  struct vm_area_struct *vma;
1690 
1691  anon_vma_lock(anon_vma);
1692  anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1693  0, ULONG_MAX) {
1694  vma = vmac->vma;
1695  if (rmap_item->address < vma->vm_start ||
1696  rmap_item->address >= vma->vm_end)
1697  continue;
1698  /*
1699  * Initially we examine only the vma which covers this
1700  * rmap_item; but later, if there is still work to do,
1701  * we examine covering vmas in other mms: in case they
1702  * were forked from the original since ksmd passed.
1703  */
1704  if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1705  continue;
1706 
1707  ret = try_to_unmap_one(page, vma,
1708  rmap_item->address, flags);
1709  if (ret != SWAP_AGAIN || !page_mapped(page)) {
1710  anon_vma_unlock(anon_vma);
1711  goto out;
1712  }
1713  }
1714  anon_vma_unlock(anon_vma);
1715  }
1716  if (!search_new_forks++)
1717  goto again;
1718 out:
1719  return ret;
1720 }
1721 
1722 #ifdef CONFIG_MIGRATION
1723 int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1724  struct vm_area_struct *, unsigned long, void *), void *arg)
1725 {
1726  struct stable_node *stable_node;
1727  struct hlist_node *hlist;
1728  struct rmap_item *rmap_item;
1729  int ret = SWAP_AGAIN;
1730  int search_new_forks = 0;
1731 
1732  VM_BUG_ON(!PageKsm(page));
1733  VM_BUG_ON(!PageLocked(page));
1734 
1735  stable_node = page_stable_node(page);
1736  if (!stable_node)
1737  return ret;
1738 again:
1739  hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1740  struct anon_vma *anon_vma = rmap_item->anon_vma;
1741  struct anon_vma_chain *vmac;
1742  struct vm_area_struct *vma;
1743 
1744  anon_vma_lock(anon_vma);
1745  anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1746  0, ULONG_MAX) {
1747  vma = vmac->vma;
1748  if (rmap_item->address < vma->vm_start ||
1749  rmap_item->address >= vma->vm_end)
1750  continue;
1751  /*
1752  * Initially we examine only the vma which covers this
1753  * rmap_item; but later, if there is still work to do,
1754  * we examine covering vmas in other mms: in case they
1755  * were forked from the original since ksmd passed.
1756  */
1757  if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1758  continue;
1759 
1760  ret = rmap_one(page, vma, rmap_item->address, arg);
1761  if (ret != SWAP_AGAIN) {
1762  anon_vma_unlock(anon_vma);
1763  goto out;
1764  }
1765  }
1766  anon_vma_unlock(anon_vma);
1767  }
1768  if (!search_new_forks++)
1769  goto again;
1770 out:
1771  return ret;
1772 }
1773 
1774 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1775 {
1776  struct stable_node *stable_node;
1777 
1778  VM_BUG_ON(!PageLocked(oldpage));
1779  VM_BUG_ON(!PageLocked(newpage));
1780  VM_BUG_ON(newpage->mapping != oldpage->mapping);
1781 
1782  stable_node = page_stable_node(newpage);
1783  if (stable_node) {
1784  VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
1785  stable_node->kpfn = page_to_pfn(newpage);
1786  }
1787 }
1788 #endif /* CONFIG_MIGRATION */
1789 
1790 #ifdef CONFIG_MEMORY_HOTREMOVE
1791 static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn,
1792  unsigned long end_pfn)
1793 {
1794  struct rb_node *node;
1795 
1796  for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) {
1797  struct stable_node *stable_node;
1798 
1799  stable_node = rb_entry(node, struct stable_node, node);
1800  if (stable_node->kpfn >= start_pfn &&
1801  stable_node->kpfn < end_pfn)
1802  return stable_node;
1803  }
1804  return NULL;
1805 }
1806 
1807 static int ksm_memory_callback(struct notifier_block *self,
1808  unsigned long action, void *arg)
1809 {
1810  struct memory_notify *mn = arg;
1811  struct stable_node *stable_node;
1812 
1813  switch (action) {
1814  case MEM_GOING_OFFLINE:
1815  /*
1816  * Keep it very simple for now: just lock out ksmd and
1817  * MADV_UNMERGEABLE while any memory is going offline.
1818  * mutex_lock_nested() is necessary because lockdep was alarmed
1819  * that here we take ksm_thread_mutex inside notifier chain
1820  * mutex, and later take notifier chain mutex inside
1821  * ksm_thread_mutex to unlock it. But that's safe because both
1822  * are inside mem_hotplug_mutex.
1823  */
1824  mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING);
1825  break;
1826 
1827  case MEM_OFFLINE:
1828  /*
1829  * Most of the work is done by page migration; but there might
1830  * be a few stable_nodes left over, still pointing to struct
1831  * pages which have been offlined: prune those from the tree.
1832  */
1833  while ((stable_node = ksm_check_stable_tree(mn->start_pfn,
1834  mn->start_pfn + mn->nr_pages)) != NULL)
1835  remove_node_from_stable_tree(stable_node);
1836  /* fallthrough */
1837 
1838  case MEM_CANCEL_OFFLINE:
1839  mutex_unlock(&ksm_thread_mutex);
1840  break;
1841  }
1842  return NOTIFY_OK;
1843 }
1844 #endif /* CONFIG_MEMORY_HOTREMOVE */
1845 
1846 #ifdef CONFIG_SYSFS
1847 /*
1848  * This all compiles without CONFIG_SYSFS, but is a waste of space.
1849  */
1850 
1851 #define KSM_ATTR_RO(_name) \
1852  static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1853 #define KSM_ATTR(_name) \
1854  static struct kobj_attribute _name##_attr = \
1855  __ATTR(_name, 0644, _name##_show, _name##_store)
1856 
1857 static ssize_t sleep_millisecs_show(struct kobject *kobj,
1858  struct kobj_attribute *attr, char *buf)
1859 {
1860  return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1861 }
1862 
1863 static ssize_t sleep_millisecs_store(struct kobject *kobj,
1864  struct kobj_attribute *attr,
1865  const char *buf, size_t count)
1866 {
1867  unsigned long msecs;
1868  int err;
1869 
1870  err = strict_strtoul(buf, 10, &msecs);
1871  if (err || msecs > UINT_MAX)
1872  return -EINVAL;
1873 
1874  ksm_thread_sleep_millisecs = msecs;
1875 
1876  return count;
1877 }
1878 KSM_ATTR(sleep_millisecs);
1879 
1880 static ssize_t pages_to_scan_show(struct kobject *kobj,
1881  struct kobj_attribute *attr, char *buf)
1882 {
1883  return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1884 }
1885 
1886 static ssize_t pages_to_scan_store(struct kobject *kobj,
1887  struct kobj_attribute *attr,
1888  const char *buf, size_t count)
1889 {
1890  int err;
1891  unsigned long nr_pages;
1892 
1893  err = strict_strtoul(buf, 10, &nr_pages);
1894  if (err || nr_pages > UINT_MAX)
1895  return -EINVAL;
1896 
1897  ksm_thread_pages_to_scan = nr_pages;
1898 
1899  return count;
1900 }
1901 KSM_ATTR(pages_to_scan);
1902 
1903 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1904  char *buf)
1905 {
1906  return sprintf(buf, "%u\n", ksm_run);
1907 }
1908 
1909 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1910  const char *buf, size_t count)
1911 {
1912  int err;
1913  unsigned long flags;
1914 
1915  err = strict_strtoul(buf, 10, &flags);
1916  if (err || flags > UINT_MAX)
1917  return -EINVAL;
1918  if (flags > KSM_RUN_UNMERGE)
1919  return -EINVAL;
1920 
1921  /*
1922  * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1923  * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
1924  * breaking COW to free the pages_shared (but leaves mm_slots
1925  * on the list for when ksmd may be set running again).
1926  */
1927 
1928  mutex_lock(&ksm_thread_mutex);
1929  if (ksm_run != flags) {
1930  ksm_run = flags;
1931  if (flags & KSM_RUN_UNMERGE) {
1932  int oom_score_adj;
1933 
1934  oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1935  err = unmerge_and_remove_all_rmap_items();
1937  oom_score_adj);
1938  if (err) {
1939  ksm_run = KSM_RUN_STOP;
1940  count = err;
1941  }
1942  }
1943  }
1944  mutex_unlock(&ksm_thread_mutex);
1945 
1946  if (flags & KSM_RUN_MERGE)
1947  wake_up_interruptible(&ksm_thread_wait);
1948 
1949  return count;
1950 }
1951 KSM_ATTR(run);
1952 
1953 static ssize_t pages_shared_show(struct kobject *kobj,
1954  struct kobj_attribute *attr, char *buf)
1955 {
1956  return sprintf(buf, "%lu\n", ksm_pages_shared);
1957 }
1958 KSM_ATTR_RO(pages_shared);
1959 
1960 static ssize_t pages_sharing_show(struct kobject *kobj,
1961  struct kobj_attribute *attr, char *buf)
1962 {
1963  return sprintf(buf, "%lu\n", ksm_pages_sharing);
1964 }
1965 KSM_ATTR_RO(pages_sharing);
1966 
1967 static ssize_t pages_unshared_show(struct kobject *kobj,
1968  struct kobj_attribute *attr, char *buf)
1969 {
1970  return sprintf(buf, "%lu\n", ksm_pages_unshared);
1971 }
1972 KSM_ATTR_RO(pages_unshared);
1973 
1974 static ssize_t pages_volatile_show(struct kobject *kobj,
1975  struct kobj_attribute *attr, char *buf)
1976 {
1977  long ksm_pages_volatile;
1978 
1979  ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1980  - ksm_pages_sharing - ksm_pages_unshared;
1981  /*
1982  * It was not worth any locking to calculate that statistic,
1983  * but it might therefore sometimes be negative: conceal that.
1984  */
1985  if (ksm_pages_volatile < 0)
1986  ksm_pages_volatile = 0;
1987  return sprintf(buf, "%ld\n", ksm_pages_volatile);
1988 }
1989 KSM_ATTR_RO(pages_volatile);
1990 
1991 static ssize_t full_scans_show(struct kobject *kobj,
1992  struct kobj_attribute *attr, char *buf)
1993 {
1994  return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1995 }
1996 KSM_ATTR_RO(full_scans);
1997 
1998 static struct attribute *ksm_attrs[] = {
1999  &sleep_millisecs_attr.attr,
2000  &pages_to_scan_attr.attr,
2001  &run_attr.attr,
2002  &pages_shared_attr.attr,
2003  &pages_sharing_attr.attr,
2004  &pages_unshared_attr.attr,
2005  &pages_volatile_attr.attr,
2006  &full_scans_attr.attr,
2007  NULL,
2008 };
2009 
2010 static struct attribute_group ksm_attr_group = {
2011  .attrs = ksm_attrs,
2012  .name = "ksm",
2013 };
2014 #endif /* CONFIG_SYSFS */
2015 
2016 static int __init ksm_init(void)
2017 {
2018  struct task_struct *ksm_thread;
2019  int err;
2020 
2021  err = ksm_slab_init();
2022  if (err)
2023  goto out;
2024 
2025  ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2026  if (IS_ERR(ksm_thread)) {
2027  printk(KERN_ERR "ksm: creating kthread failed\n");
2028  err = PTR_ERR(ksm_thread);
2029  goto out_free;
2030  }
2031 
2032 #ifdef CONFIG_SYSFS
2033  err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2034  if (err) {
2035  printk(KERN_ERR "ksm: register sysfs failed\n");
2036  kthread_stop(ksm_thread);
2037  goto out_free;
2038  }
2039 #else
2040  ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2041 
2042 #endif /* CONFIG_SYSFS */
2043 
2044 #ifdef CONFIG_MEMORY_HOTREMOVE
2045  /*
2046  * Choose a high priority since the callback takes ksm_thread_mutex:
2047  * later callbacks could only be taking locks which nest within that.
2048  */
2049  hotplug_memory_notifier(ksm_memory_callback, 100);
2050 #endif
2051  return 0;
2052 
2053 out_free:
2054  ksm_slab_free();
2055 out:
2056  return err;
2057 }
2058 module_init(ksm_init)