Merge tag 'devicetree-for-4.4' of git://git.kernel.org/pub/scm/linux/kernel/git/robh...
[firefly-linux-kernel-4.4.55.git] / mm / ksm.c
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/hashtable.h>
37 #include <linux/freezer.h>
38 #include <linux/oom.h>
39 #include <linux/numa.h>
40
41 #include <asm/tlbflush.h>
42 #include "internal.h"
43
44 #ifdef CONFIG_NUMA
45 #define NUMA(x)         (x)
46 #define DO_NUMA(x)      do { (x); } while (0)
47 #else
48 #define NUMA(x)         (0)
49 #define DO_NUMA(x)      do { } while (0)
50 #endif
51
52 /*
53  * A few notes about the KSM scanning process,
54  * to make it easier to understand the data structures below:
55  *
56  * In order to reduce excessive scanning, KSM sorts the memory pages by their
57  * contents into a data structure that holds pointers to the pages' locations.
58  *
59  * Since the contents of the pages may change at any moment, KSM cannot just
60  * insert the pages into a normal sorted tree and expect it to find anything.
61  * Therefore KSM uses two data structures - the stable and the unstable tree.
62  *
63  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
64  * by their contents.  Because each such page is write-protected, searching on
65  * this tree is fully assured to be working (except when pages are unmapped),
66  * and therefore this tree is called the stable tree.
67  *
68  * In addition to the stable tree, KSM uses a second data structure called the
69  * unstable tree: this tree holds pointers to pages which have been found to
70  * be "unchanged for a period of time".  The unstable tree sorts these pages
71  * by their contents, but since they are not write-protected, KSM cannot rely
72  * upon the unstable tree to work correctly - the unstable tree is liable to
73  * be corrupted as its contents are modified, and so it is called unstable.
74  *
75  * KSM solves this problem by several techniques:
76  *
77  * 1) The unstable tree is flushed every time KSM completes scanning all
78  *    memory areas, and then the tree is rebuilt again from the beginning.
79  * 2) KSM will only insert into the unstable tree, pages whose hash value
80  *    has not changed since the previous scan of all memory areas.
81  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
82  *    colors of the nodes and not on their contents, assuring that even when
83  *    the tree gets "corrupted" it won't get out of balance, so scanning time
84  *    remains the same (also, searching and inserting nodes in an rbtree uses
85  *    the same algorithm, so we have no overhead when we flush and rebuild).
86  * 4) KSM never flushes the stable tree, which means that even if it were to
87  *    take 10 attempts to find a page in the unstable tree, once it is found,
88  *    it is secured in the stable tree.  (When we scan a new page, we first
89  *    compare it against the stable tree, and then against the unstable tree.)
90  *
91  * If the merge_across_nodes tunable is unset, then KSM maintains multiple
92  * stable trees and multiple unstable trees: one of each for each NUMA node.
93  */
94
95 /**
96  * struct mm_slot - ksm information per mm that is being scanned
97  * @link: link to the mm_slots hash list
98  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
99  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
100  * @mm: the mm that this information is valid for
101  */
102 struct mm_slot {
103         struct hlist_node link;
104         struct list_head mm_list;
105         struct rmap_item *rmap_list;
106         struct mm_struct *mm;
107 };
108
109 /**
110  * struct ksm_scan - cursor for scanning
111  * @mm_slot: the current mm_slot we are scanning
112  * @address: the next address inside that to be scanned
113  * @rmap_list: link to the next rmap to be scanned in the rmap_list
114  * @seqnr: count of completed full scans (needed when removing unstable node)
115  *
116  * There is only the one ksm_scan instance of this cursor structure.
117  */
118 struct ksm_scan {
119         struct mm_slot *mm_slot;
120         unsigned long address;
121         struct rmap_item **rmap_list;
122         unsigned long seqnr;
123 };
124
125 /**
126  * struct stable_node - node of the stable rbtree
127  * @node: rb node of this ksm page in the stable tree
128  * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
129  * @list: linked into migrate_nodes, pending placement in the proper node tree
130  * @hlist: hlist head of rmap_items using this ksm page
131  * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
132  * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
133  */
134 struct stable_node {
135         union {
136                 struct rb_node node;    /* when node of stable tree */
137                 struct {                /* when listed for migration */
138                         struct list_head *head;
139                         struct list_head list;
140                 };
141         };
142         struct hlist_head hlist;
143         unsigned long kpfn;
144 #ifdef CONFIG_NUMA
145         int nid;
146 #endif
147 };
148
149 /**
150  * struct rmap_item - reverse mapping item for virtual addresses
151  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
152  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
153  * @nid: NUMA node id of unstable tree in which linked (may not match page)
154  * @mm: the memory structure this rmap_item is pointing into
155  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
156  * @oldchecksum: previous checksum of the page at that virtual address
157  * @node: rb node of this rmap_item in the unstable tree
158  * @head: pointer to stable_node heading this list in the stable tree
159  * @hlist: link into hlist of rmap_items hanging off that stable_node
160  */
161 struct rmap_item {
162         struct rmap_item *rmap_list;
163         union {
164                 struct anon_vma *anon_vma;      /* when stable */
165 #ifdef CONFIG_NUMA
166                 int nid;                /* when node of unstable tree */
167 #endif
168         };
169         struct mm_struct *mm;
170         unsigned long address;          /* + low bits used for flags below */
171         unsigned int oldchecksum;       /* when unstable */
172         union {
173                 struct rb_node node;    /* when node of unstable tree */
174                 struct {                /* when listed from stable tree */
175                         struct stable_node *head;
176                         struct hlist_node hlist;
177                 };
178         };
179 };
180
181 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
182 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
183 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
184
185 /* The stable and unstable tree heads */
186 static struct rb_root one_stable_tree[1] = { RB_ROOT };
187 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
188 static struct rb_root *root_stable_tree = one_stable_tree;
189 static struct rb_root *root_unstable_tree = one_unstable_tree;
190
191 /* Recently migrated nodes of stable tree, pending proper placement */
192 static LIST_HEAD(migrate_nodes);
193
194 #define MM_SLOTS_HASH_BITS 10
195 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
196
197 static struct mm_slot ksm_mm_head = {
198         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
199 };
200 static struct ksm_scan ksm_scan = {
201         .mm_slot = &ksm_mm_head,
202 };
203
204 static struct kmem_cache *rmap_item_cache;
205 static struct kmem_cache *stable_node_cache;
206 static struct kmem_cache *mm_slot_cache;
207
208 /* The number of nodes in the stable tree */
209 static unsigned long ksm_pages_shared;
210
211 /* The number of page slots additionally sharing those nodes */
212 static unsigned long ksm_pages_sharing;
213
214 /* The number of nodes in the unstable tree */
215 static unsigned long ksm_pages_unshared;
216
217 /* The number of rmap_items in use: to calculate pages_volatile */
218 static unsigned long ksm_rmap_items;
219
220 /* Number of pages ksmd should scan in one batch */
221 static unsigned int ksm_thread_pages_to_scan = 100;
222
223 /* Milliseconds ksmd should sleep between batches */
224 static unsigned int ksm_thread_sleep_millisecs = 20;
225
226 #ifdef CONFIG_NUMA
227 /* Zeroed when merging across nodes is not allowed */
228 static unsigned int ksm_merge_across_nodes = 1;
229 static int ksm_nr_node_ids = 1;
230 #else
231 #define ksm_merge_across_nodes  1U
232 #define ksm_nr_node_ids         1
233 #endif
234
235 #define KSM_RUN_STOP    0
236 #define KSM_RUN_MERGE   1
237 #define KSM_RUN_UNMERGE 2
238 #define KSM_RUN_OFFLINE 4
239 static unsigned long ksm_run = KSM_RUN_STOP;
240 static void wait_while_offlining(void);
241
242 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
243 static DEFINE_MUTEX(ksm_thread_mutex);
244 static DEFINE_SPINLOCK(ksm_mmlist_lock);
245
246 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
247                 sizeof(struct __struct), __alignof__(struct __struct),\
248                 (__flags), NULL)
249
250 static int __init ksm_slab_init(void)
251 {
252         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
253         if (!rmap_item_cache)
254                 goto out;
255
256         stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
257         if (!stable_node_cache)
258                 goto out_free1;
259
260         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
261         if (!mm_slot_cache)
262                 goto out_free2;
263
264         return 0;
265
266 out_free2:
267         kmem_cache_destroy(stable_node_cache);
268 out_free1:
269         kmem_cache_destroy(rmap_item_cache);
270 out:
271         return -ENOMEM;
272 }
273
274 static void __init ksm_slab_free(void)
275 {
276         kmem_cache_destroy(mm_slot_cache);
277         kmem_cache_destroy(stable_node_cache);
278         kmem_cache_destroy(rmap_item_cache);
279         mm_slot_cache = NULL;
280 }
281
282 static inline struct rmap_item *alloc_rmap_item(void)
283 {
284         struct rmap_item *rmap_item;
285
286         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
287         if (rmap_item)
288                 ksm_rmap_items++;
289         return rmap_item;
290 }
291
292 static inline void free_rmap_item(struct rmap_item *rmap_item)
293 {
294         ksm_rmap_items--;
295         rmap_item->mm = NULL;   /* debug safety */
296         kmem_cache_free(rmap_item_cache, rmap_item);
297 }
298
299 static inline struct stable_node *alloc_stable_node(void)
300 {
301         return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
302 }
303
304 static inline void free_stable_node(struct stable_node *stable_node)
305 {
306         kmem_cache_free(stable_node_cache, stable_node);
307 }
308
309 static inline struct mm_slot *alloc_mm_slot(void)
310 {
311         if (!mm_slot_cache)     /* initialization failed */
312                 return NULL;
313         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
314 }
315
316 static inline void free_mm_slot(struct mm_slot *mm_slot)
317 {
318         kmem_cache_free(mm_slot_cache, mm_slot);
319 }
320
321 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
322 {
323         struct mm_slot *slot;
324
325         hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
326                 if (slot->mm == mm)
327                         return slot;
328
329         return NULL;
330 }
331
332 static void insert_to_mm_slots_hash(struct mm_struct *mm,
333                                     struct mm_slot *mm_slot)
334 {
335         mm_slot->mm = mm;
336         hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
337 }
338
339 /*
340  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
341  * page tables after it has passed through ksm_exit() - which, if necessary,
342  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
343  * a special flag: they can just back out as soon as mm_users goes to zero.
344  * ksm_test_exit() is used throughout to make this test for exit: in some
345  * places for correctness, in some places just to avoid unnecessary work.
346  */
347 static inline bool ksm_test_exit(struct mm_struct *mm)
348 {
349         return atomic_read(&mm->mm_users) == 0;
350 }
351
352 /*
353  * We use break_ksm to break COW on a ksm page: it's a stripped down
354  *
355  *      if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
356  *              put_page(page);
357  *
358  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
359  * in case the application has unmapped and remapped mm,addr meanwhile.
360  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
361  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
362  */
363 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
364 {
365         struct page *page;
366         int ret = 0;
367
368         do {
369                 cond_resched();
370                 page = follow_page(vma, addr, FOLL_GET | FOLL_MIGRATION);
371                 if (IS_ERR_OR_NULL(page))
372                         break;
373                 if (PageKsm(page))
374                         ret = handle_mm_fault(vma->vm_mm, vma, addr,
375                                                         FAULT_FLAG_WRITE);
376                 else
377                         ret = VM_FAULT_WRITE;
378                 put_page(page);
379         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
380         /*
381          * We must loop because handle_mm_fault() may back out if there's
382          * any difficulty e.g. if pte accessed bit gets updated concurrently.
383          *
384          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
385          * COW has been broken, even if the vma does not permit VM_WRITE;
386          * but note that a concurrent fault might break PageKsm for us.
387          *
388          * VM_FAULT_SIGBUS could occur if we race with truncation of the
389          * backing file, which also invalidates anonymous pages: that's
390          * okay, that truncation will have unmapped the PageKsm for us.
391          *
392          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
393          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
394          * current task has TIF_MEMDIE set, and will be OOM killed on return
395          * to user; and ksmd, having no mm, would never be chosen for that.
396          *
397          * But if the mm is in a limited mem_cgroup, then the fault may fail
398          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
399          * even ksmd can fail in this way - though it's usually breaking ksm
400          * just to undo a merge it made a moment before, so unlikely to oom.
401          *
402          * That's a pity: we might therefore have more kernel pages allocated
403          * than we're counting as nodes in the stable tree; but ksm_do_scan
404          * will retry to break_cow on each pass, so should recover the page
405          * in due course.  The important thing is to not let VM_MERGEABLE
406          * be cleared while any such pages might remain in the area.
407          */
408         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
409 }
410
411 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
412                 unsigned long addr)
413 {
414         struct vm_area_struct *vma;
415         if (ksm_test_exit(mm))
416                 return NULL;
417         vma = find_vma(mm, addr);
418         if (!vma || vma->vm_start > addr)
419                 return NULL;
420         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
421                 return NULL;
422         return vma;
423 }
424
425 static void break_cow(struct rmap_item *rmap_item)
426 {
427         struct mm_struct *mm = rmap_item->mm;
428         unsigned long addr = rmap_item->address;
429         struct vm_area_struct *vma;
430
431         /*
432          * It is not an accident that whenever we want to break COW
433          * to undo, we also need to drop a reference to the anon_vma.
434          */
435         put_anon_vma(rmap_item->anon_vma);
436
437         down_read(&mm->mmap_sem);
438         vma = find_mergeable_vma(mm, addr);
439         if (vma)
440                 break_ksm(vma, addr);
441         up_read(&mm->mmap_sem);
442 }
443
444 static struct page *page_trans_compound_anon(struct page *page)
445 {
446         if (PageTransCompound(page)) {
447                 struct page *head = compound_head(page);
448                 /*
449                  * head may actually be splitted and freed from under
450                  * us but it's ok here.
451                  */
452                 if (PageAnon(head))
453                         return head;
454         }
455         return NULL;
456 }
457
458 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
459 {
460         struct mm_struct *mm = rmap_item->mm;
461         unsigned long addr = rmap_item->address;
462         struct vm_area_struct *vma;
463         struct page *page;
464
465         down_read(&mm->mmap_sem);
466         vma = find_mergeable_vma(mm, addr);
467         if (!vma)
468                 goto out;
469
470         page = follow_page(vma, addr, FOLL_GET);
471         if (IS_ERR_OR_NULL(page))
472                 goto out;
473         if (PageAnon(page) || page_trans_compound_anon(page)) {
474                 flush_anon_page(vma, page, addr);
475                 flush_dcache_page(page);
476         } else {
477                 put_page(page);
478 out:
479                 page = NULL;
480         }
481         up_read(&mm->mmap_sem);
482         return page;
483 }
484
485 /*
486  * This helper is used for getting right index into array of tree roots.
487  * When merge_across_nodes knob is set to 1, there are only two rb-trees for
488  * stable and unstable pages from all nodes with roots in index 0. Otherwise,
489  * every node has its own stable and unstable tree.
490  */
491 static inline int get_kpfn_nid(unsigned long kpfn)
492 {
493         return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
494 }
495
496 static void remove_node_from_stable_tree(struct stable_node *stable_node)
497 {
498         struct rmap_item *rmap_item;
499
500         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
501                 if (rmap_item->hlist.next)
502                         ksm_pages_sharing--;
503                 else
504                         ksm_pages_shared--;
505                 put_anon_vma(rmap_item->anon_vma);
506                 rmap_item->address &= PAGE_MASK;
507                 cond_resched();
508         }
509
510         if (stable_node->head == &migrate_nodes)
511                 list_del(&stable_node->list);
512         else
513                 rb_erase(&stable_node->node,
514                          root_stable_tree + NUMA(stable_node->nid));
515         free_stable_node(stable_node);
516 }
517
518 /*
519  * get_ksm_page: checks if the page indicated by the stable node
520  * is still its ksm page, despite having held no reference to it.
521  * In which case we can trust the content of the page, and it
522  * returns the gotten page; but if the page has now been zapped,
523  * remove the stale node from the stable tree and return NULL.
524  * But beware, the stable node's page might be being migrated.
525  *
526  * You would expect the stable_node to hold a reference to the ksm page.
527  * But if it increments the page's count, swapping out has to wait for
528  * ksmd to come around again before it can free the page, which may take
529  * seconds or even minutes: much too unresponsive.  So instead we use a
530  * "keyhole reference": access to the ksm page from the stable node peeps
531  * out through its keyhole to see if that page still holds the right key,
532  * pointing back to this stable node.  This relies on freeing a PageAnon
533  * page to reset its page->mapping to NULL, and relies on no other use of
534  * a page to put something that might look like our key in page->mapping.
535  * is on its way to being freed; but it is an anomaly to bear in mind.
536  */
537 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
538 {
539         struct page *page;
540         void *expected_mapping;
541         unsigned long kpfn;
542
543         expected_mapping = (void *)stable_node +
544                                 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
545 again:
546         kpfn = READ_ONCE(stable_node->kpfn);
547         page = pfn_to_page(kpfn);
548
549         /*
550          * page is computed from kpfn, so on most architectures reading
551          * page->mapping is naturally ordered after reading node->kpfn,
552          * but on Alpha we need to be more careful.
553          */
554         smp_read_barrier_depends();
555         if (READ_ONCE(page->mapping) != expected_mapping)
556                 goto stale;
557
558         /*
559          * We cannot do anything with the page while its refcount is 0.
560          * Usually 0 means free, or tail of a higher-order page: in which
561          * case this node is no longer referenced, and should be freed;
562          * however, it might mean that the page is under page_freeze_refs().
563          * The __remove_mapping() case is easy, again the node is now stale;
564          * but if page is swapcache in migrate_page_move_mapping(), it might
565          * still be our page, in which case it's essential to keep the node.
566          */
567         while (!get_page_unless_zero(page)) {
568                 /*
569                  * Another check for page->mapping != expected_mapping would
570                  * work here too.  We have chosen the !PageSwapCache test to
571                  * optimize the common case, when the page is or is about to
572                  * be freed: PageSwapCache is cleared (under spin_lock_irq)
573                  * in the freeze_refs section of __remove_mapping(); but Anon
574                  * page->mapping reset to NULL later, in free_pages_prepare().
575                  */
576                 if (!PageSwapCache(page))
577                         goto stale;
578                 cpu_relax();
579         }
580
581         if (READ_ONCE(page->mapping) != expected_mapping) {
582                 put_page(page);
583                 goto stale;
584         }
585
586         if (lock_it) {
587                 lock_page(page);
588                 if (READ_ONCE(page->mapping) != expected_mapping) {
589                         unlock_page(page);
590                         put_page(page);
591                         goto stale;
592                 }
593         }
594         return page;
595
596 stale:
597         /*
598          * We come here from above when page->mapping or !PageSwapCache
599          * suggests that the node is stale; but it might be under migration.
600          * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
601          * before checking whether node->kpfn has been changed.
602          */
603         smp_rmb();
604         if (READ_ONCE(stable_node->kpfn) != kpfn)
605                 goto again;
606         remove_node_from_stable_tree(stable_node);
607         return NULL;
608 }
609
610 /*
611  * Removing rmap_item from stable or unstable tree.
612  * This function will clean the information from the stable/unstable tree.
613  */
614 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
615 {
616         if (rmap_item->address & STABLE_FLAG) {
617                 struct stable_node *stable_node;
618                 struct page *page;
619
620                 stable_node = rmap_item->head;
621                 page = get_ksm_page(stable_node, true);
622                 if (!page)
623                         goto out;
624
625                 hlist_del(&rmap_item->hlist);
626                 unlock_page(page);
627                 put_page(page);
628
629                 if (!hlist_empty(&stable_node->hlist))
630                         ksm_pages_sharing--;
631                 else
632                         ksm_pages_shared--;
633
634                 put_anon_vma(rmap_item->anon_vma);
635                 rmap_item->address &= PAGE_MASK;
636
637         } else if (rmap_item->address & UNSTABLE_FLAG) {
638                 unsigned char age;
639                 /*
640                  * Usually ksmd can and must skip the rb_erase, because
641                  * root_unstable_tree was already reset to RB_ROOT.
642                  * But be careful when an mm is exiting: do the rb_erase
643                  * if this rmap_item was inserted by this scan, rather
644                  * than left over from before.
645                  */
646                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
647                 BUG_ON(age > 1);
648                 if (!age)
649                         rb_erase(&rmap_item->node,
650                                  root_unstable_tree + NUMA(rmap_item->nid));
651                 ksm_pages_unshared--;
652                 rmap_item->address &= PAGE_MASK;
653         }
654 out:
655         cond_resched();         /* we're called from many long loops */
656 }
657
658 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
659                                        struct rmap_item **rmap_list)
660 {
661         while (*rmap_list) {
662                 struct rmap_item *rmap_item = *rmap_list;
663                 *rmap_list = rmap_item->rmap_list;
664                 remove_rmap_item_from_tree(rmap_item);
665                 free_rmap_item(rmap_item);
666         }
667 }
668
669 /*
670  * Though it's very tempting to unmerge rmap_items from stable tree rather
671  * than check every pte of a given vma, the locking doesn't quite work for
672  * that - an rmap_item is assigned to the stable tree after inserting ksm
673  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
674  * rmap_items from parent to child at fork time (so as not to waste time
675  * if exit comes before the next scan reaches it).
676  *
677  * Similarly, although we'd like to remove rmap_items (so updating counts
678  * and freeing memory) when unmerging an area, it's easier to leave that
679  * to the next pass of ksmd - consider, for example, how ksmd might be
680  * in cmp_and_merge_page on one of the rmap_items we would be removing.
681  */
682 static int unmerge_ksm_pages(struct vm_area_struct *vma,
683                              unsigned long start, unsigned long end)
684 {
685         unsigned long addr;
686         int err = 0;
687
688         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
689                 if (ksm_test_exit(vma->vm_mm))
690                         break;
691                 if (signal_pending(current))
692                         err = -ERESTARTSYS;
693                 else
694                         err = break_ksm(vma, addr);
695         }
696         return err;
697 }
698
699 #ifdef CONFIG_SYSFS
700 /*
701  * Only called through the sysfs control interface:
702  */
703 static int remove_stable_node(struct stable_node *stable_node)
704 {
705         struct page *page;
706         int err;
707
708         page = get_ksm_page(stable_node, true);
709         if (!page) {
710                 /*
711                  * get_ksm_page did remove_node_from_stable_tree itself.
712                  */
713                 return 0;
714         }
715
716         if (WARN_ON_ONCE(page_mapped(page))) {
717                 /*
718                  * This should not happen: but if it does, just refuse to let
719                  * merge_across_nodes be switched - there is no need to panic.
720                  */
721                 err = -EBUSY;
722         } else {
723                 /*
724                  * The stable node did not yet appear stale to get_ksm_page(),
725                  * since that allows for an unmapped ksm page to be recognized
726                  * right up until it is freed; but the node is safe to remove.
727                  * This page might be in a pagevec waiting to be freed,
728                  * or it might be PageSwapCache (perhaps under writeback),
729                  * or it might have been removed from swapcache a moment ago.
730                  */
731                 set_page_stable_node(page, NULL);
732                 remove_node_from_stable_tree(stable_node);
733                 err = 0;
734         }
735
736         unlock_page(page);
737         put_page(page);
738         return err;
739 }
740
741 static int remove_all_stable_nodes(void)
742 {
743         struct stable_node *stable_node;
744         struct list_head *this, *next;
745         int nid;
746         int err = 0;
747
748         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
749                 while (root_stable_tree[nid].rb_node) {
750                         stable_node = rb_entry(root_stable_tree[nid].rb_node,
751                                                 struct stable_node, node);
752                         if (remove_stable_node(stable_node)) {
753                                 err = -EBUSY;
754                                 break;  /* proceed to next nid */
755                         }
756                         cond_resched();
757                 }
758         }
759         list_for_each_safe(this, next, &migrate_nodes) {
760                 stable_node = list_entry(this, struct stable_node, list);
761                 if (remove_stable_node(stable_node))
762                         err = -EBUSY;
763                 cond_resched();
764         }
765         return err;
766 }
767
768 static int unmerge_and_remove_all_rmap_items(void)
769 {
770         struct mm_slot *mm_slot;
771         struct mm_struct *mm;
772         struct vm_area_struct *vma;
773         int err = 0;
774
775         spin_lock(&ksm_mmlist_lock);
776         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
777                                                 struct mm_slot, mm_list);
778         spin_unlock(&ksm_mmlist_lock);
779
780         for (mm_slot = ksm_scan.mm_slot;
781                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
782                 mm = mm_slot->mm;
783                 down_read(&mm->mmap_sem);
784                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
785                         if (ksm_test_exit(mm))
786                                 break;
787                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
788                                 continue;
789                         err = unmerge_ksm_pages(vma,
790                                                 vma->vm_start, vma->vm_end);
791                         if (err)
792                                 goto error;
793                 }
794
795                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
796
797                 spin_lock(&ksm_mmlist_lock);
798                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
799                                                 struct mm_slot, mm_list);
800                 if (ksm_test_exit(mm)) {
801                         hash_del(&mm_slot->link);
802                         list_del(&mm_slot->mm_list);
803                         spin_unlock(&ksm_mmlist_lock);
804
805                         free_mm_slot(mm_slot);
806                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
807                         up_read(&mm->mmap_sem);
808                         mmdrop(mm);
809                 } else {
810                         spin_unlock(&ksm_mmlist_lock);
811                         up_read(&mm->mmap_sem);
812                 }
813         }
814
815         /* Clean up stable nodes, but don't worry if some are still busy */
816         remove_all_stable_nodes();
817         ksm_scan.seqnr = 0;
818         return 0;
819
820 error:
821         up_read(&mm->mmap_sem);
822         spin_lock(&ksm_mmlist_lock);
823         ksm_scan.mm_slot = &ksm_mm_head;
824         spin_unlock(&ksm_mmlist_lock);
825         return err;
826 }
827 #endif /* CONFIG_SYSFS */
828
829 static u32 calc_checksum(struct page *page)
830 {
831         u32 checksum;
832         void *addr = kmap_atomic(page);
833         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
834         kunmap_atomic(addr);
835         return checksum;
836 }
837
838 static int memcmp_pages(struct page *page1, struct page *page2)
839 {
840         char *addr1, *addr2;
841         int ret;
842
843         addr1 = kmap_atomic(page1);
844         addr2 = kmap_atomic(page2);
845         ret = memcmp(addr1, addr2, PAGE_SIZE);
846         kunmap_atomic(addr2);
847         kunmap_atomic(addr1);
848         return ret;
849 }
850
851 static inline int pages_identical(struct page *page1, struct page *page2)
852 {
853         return !memcmp_pages(page1, page2);
854 }
855
856 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
857                               pte_t *orig_pte)
858 {
859         struct mm_struct *mm = vma->vm_mm;
860         unsigned long addr;
861         pte_t *ptep;
862         spinlock_t *ptl;
863         int swapped;
864         int err = -EFAULT;
865         unsigned long mmun_start;       /* For mmu_notifiers */
866         unsigned long mmun_end;         /* For mmu_notifiers */
867
868         addr = page_address_in_vma(page, vma);
869         if (addr == -EFAULT)
870                 goto out;
871
872         BUG_ON(PageTransCompound(page));
873
874         mmun_start = addr;
875         mmun_end   = addr + PAGE_SIZE;
876         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
877
878         ptep = page_check_address(page, mm, addr, &ptl, 0);
879         if (!ptep)
880                 goto out_mn;
881
882         if (pte_write(*ptep) || pte_dirty(*ptep)) {
883                 pte_t entry;
884
885                 swapped = PageSwapCache(page);
886                 flush_cache_page(vma, addr, page_to_pfn(page));
887                 /*
888                  * Ok this is tricky, when get_user_pages_fast() run it doesn't
889                  * take any lock, therefore the check that we are going to make
890                  * with the pagecount against the mapcount is racey and
891                  * O_DIRECT can happen right after the check.
892                  * So we clear the pte and flush the tlb before the check
893                  * this assure us that no O_DIRECT can happen after the check
894                  * or in the middle of the check.
895                  */
896                 entry = ptep_clear_flush_notify(vma, addr, ptep);
897                 /*
898                  * Check that no O_DIRECT or similar I/O is in progress on the
899                  * page
900                  */
901                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
902                         set_pte_at(mm, addr, ptep, entry);
903                         goto out_unlock;
904                 }
905                 if (pte_dirty(entry))
906                         set_page_dirty(page);
907                 entry = pte_mkclean(pte_wrprotect(entry));
908                 set_pte_at_notify(mm, addr, ptep, entry);
909         }
910         *orig_pte = *ptep;
911         err = 0;
912
913 out_unlock:
914         pte_unmap_unlock(ptep, ptl);
915 out_mn:
916         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
917 out:
918         return err;
919 }
920
921 /**
922  * replace_page - replace page in vma by new ksm page
923  * @vma:      vma that holds the pte pointing to page
924  * @page:     the page we are replacing by kpage
925  * @kpage:    the ksm page we replace page by
926  * @orig_pte: the original value of the pte
927  *
928  * Returns 0 on success, -EFAULT on failure.
929  */
930 static int replace_page(struct vm_area_struct *vma, struct page *page,
931                         struct page *kpage, pte_t orig_pte)
932 {
933         struct mm_struct *mm = vma->vm_mm;
934         pmd_t *pmd;
935         pte_t *ptep;
936         spinlock_t *ptl;
937         unsigned long addr;
938         int err = -EFAULT;
939         unsigned long mmun_start;       /* For mmu_notifiers */
940         unsigned long mmun_end;         /* For mmu_notifiers */
941
942         addr = page_address_in_vma(page, vma);
943         if (addr == -EFAULT)
944                 goto out;
945
946         pmd = mm_find_pmd(mm, addr);
947         if (!pmd)
948                 goto out;
949
950         mmun_start = addr;
951         mmun_end   = addr + PAGE_SIZE;
952         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
953
954         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
955         if (!pte_same(*ptep, orig_pte)) {
956                 pte_unmap_unlock(ptep, ptl);
957                 goto out_mn;
958         }
959
960         get_page(kpage);
961         page_add_anon_rmap(kpage, vma, addr);
962
963         flush_cache_page(vma, addr, pte_pfn(*ptep));
964         ptep_clear_flush_notify(vma, addr, ptep);
965         set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
966
967         page_remove_rmap(page);
968         if (!page_mapped(page))
969                 try_to_free_swap(page);
970         put_page(page);
971
972         pte_unmap_unlock(ptep, ptl);
973         err = 0;
974 out_mn:
975         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
976 out:
977         return err;
978 }
979
980 static int page_trans_compound_anon_split(struct page *page)
981 {
982         int ret = 0;
983         struct page *transhuge_head = page_trans_compound_anon(page);
984         if (transhuge_head) {
985                 /* Get the reference on the head to split it. */
986                 if (get_page_unless_zero(transhuge_head)) {
987                         /*
988                          * Recheck we got the reference while the head
989                          * was still anonymous.
990                          */
991                         if (PageAnon(transhuge_head))
992                                 ret = split_huge_page(transhuge_head);
993                         else
994                                 /*
995                                  * Retry later if split_huge_page run
996                                  * from under us.
997                                  */
998                                 ret = 1;
999                         put_page(transhuge_head);
1000                 } else
1001                         /* Retry later if split_huge_page run from under us. */
1002                         ret = 1;
1003         }
1004         return ret;
1005 }
1006
1007 /*
1008  * try_to_merge_one_page - take two pages and merge them into one
1009  * @vma: the vma that holds the pte pointing to page
1010  * @page: the PageAnon page that we want to replace with kpage
1011  * @kpage: the PageKsm page that we want to map instead of page,
1012  *         or NULL the first time when we want to use page as kpage.
1013  *
1014  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1015  */
1016 static int try_to_merge_one_page(struct vm_area_struct *vma,
1017                                  struct page *page, struct page *kpage)
1018 {
1019         pte_t orig_pte = __pte(0);
1020         int err = -EFAULT;
1021
1022         if (page == kpage)                      /* ksm page forked */
1023                 return 0;
1024
1025         if (PageTransCompound(page) && page_trans_compound_anon_split(page))
1026                 goto out;
1027         BUG_ON(PageTransCompound(page));
1028         if (!PageAnon(page))
1029                 goto out;
1030
1031         /*
1032          * We need the page lock to read a stable PageSwapCache in
1033          * write_protect_page().  We use trylock_page() instead of
1034          * lock_page() because we don't want to wait here - we
1035          * prefer to continue scanning and merging different pages,
1036          * then come back to this page when it is unlocked.
1037          */
1038         if (!trylock_page(page))
1039                 goto out;
1040         /*
1041          * If this anonymous page is mapped only here, its pte may need
1042          * to be write-protected.  If it's mapped elsewhere, all of its
1043          * ptes are necessarily already write-protected.  But in either
1044          * case, we need to lock and check page_count is not raised.
1045          */
1046         if (write_protect_page(vma, page, &orig_pte) == 0) {
1047                 if (!kpage) {
1048                         /*
1049                          * While we hold page lock, upgrade page from
1050                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
1051                          * stable_tree_insert() will update stable_node.
1052                          */
1053                         set_page_stable_node(page, NULL);
1054                         mark_page_accessed(page);
1055                         err = 0;
1056                 } else if (pages_identical(page, kpage))
1057                         err = replace_page(vma, page, kpage, orig_pte);
1058         }
1059
1060         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1061                 munlock_vma_page(page);
1062                 if (!PageMlocked(kpage)) {
1063                         unlock_page(page);
1064                         lock_page(kpage);
1065                         mlock_vma_page(kpage);
1066                         page = kpage;           /* for final unlock */
1067                 }
1068         }
1069
1070         unlock_page(page);
1071 out:
1072         return err;
1073 }
1074
1075 /*
1076  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1077  * but no new kernel page is allocated: kpage must already be a ksm page.
1078  *
1079  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1080  */
1081 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1082                                       struct page *page, struct page *kpage)
1083 {
1084         struct mm_struct *mm = rmap_item->mm;
1085         struct vm_area_struct *vma;
1086         int err = -EFAULT;
1087
1088         down_read(&mm->mmap_sem);
1089         vma = find_mergeable_vma(mm, rmap_item->address);
1090         if (!vma)
1091                 goto out;
1092
1093         err = try_to_merge_one_page(vma, page, kpage);
1094         if (err)
1095                 goto out;
1096
1097         /* Unstable nid is in union with stable anon_vma: remove first */
1098         remove_rmap_item_from_tree(rmap_item);
1099
1100         /* Must get reference to anon_vma while still holding mmap_sem */
1101         rmap_item->anon_vma = vma->anon_vma;
1102         get_anon_vma(vma->anon_vma);
1103 out:
1104         up_read(&mm->mmap_sem);
1105         return err;
1106 }
1107
1108 /*
1109  * try_to_merge_two_pages - take two identical pages and prepare them
1110  * to be merged into one page.
1111  *
1112  * This function returns the kpage if we successfully merged two identical
1113  * pages into one ksm page, NULL otherwise.
1114  *
1115  * Note that this function upgrades page to ksm page: if one of the pages
1116  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1117  */
1118 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1119                                            struct page *page,
1120                                            struct rmap_item *tree_rmap_item,
1121                                            struct page *tree_page)
1122 {
1123         int err;
1124
1125         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1126         if (!err) {
1127                 err = try_to_merge_with_ksm_page(tree_rmap_item,
1128                                                         tree_page, page);
1129                 /*
1130                  * If that fails, we have a ksm page with only one pte
1131                  * pointing to it: so break it.
1132                  */
1133                 if (err)
1134                         break_cow(rmap_item);
1135         }
1136         return err ? NULL : page;
1137 }
1138
1139 /*
1140  * stable_tree_search - search for page inside the stable tree
1141  *
1142  * This function checks if there is a page inside the stable tree
1143  * with identical content to the page that we are scanning right now.
1144  *
1145  * This function returns the stable tree node of identical content if found,
1146  * NULL otherwise.
1147  */
1148 static struct page *stable_tree_search(struct page *page)
1149 {
1150         int nid;
1151         struct rb_root *root;
1152         struct rb_node **new;
1153         struct rb_node *parent;
1154         struct stable_node *stable_node;
1155         struct stable_node *page_node;
1156
1157         page_node = page_stable_node(page);
1158         if (page_node && page_node->head != &migrate_nodes) {
1159                 /* ksm page forked */
1160                 get_page(page);
1161                 return page;
1162         }
1163
1164         nid = get_kpfn_nid(page_to_pfn(page));
1165         root = root_stable_tree + nid;
1166 again:
1167         new = &root->rb_node;
1168         parent = NULL;
1169
1170         while (*new) {
1171                 struct page *tree_page;
1172                 int ret;
1173
1174                 cond_resched();
1175                 stable_node = rb_entry(*new, struct stable_node, node);
1176                 tree_page = get_ksm_page(stable_node, false);
1177                 if (!tree_page) {
1178                         /*
1179                          * If we walked over a stale stable_node,
1180                          * get_ksm_page() will call rb_erase() and it
1181                          * may rebalance the tree from under us. So
1182                          * restart the search from scratch. Returning
1183                          * NULL would be safe too, but we'd generate
1184                          * false negative insertions just because some
1185                          * stable_node was stale.
1186                          */
1187                         goto again;
1188                 }
1189
1190                 ret = memcmp_pages(page, tree_page);
1191                 put_page(tree_page);
1192
1193                 parent = *new;
1194                 if (ret < 0)
1195                         new = &parent->rb_left;
1196                 else if (ret > 0)
1197                         new = &parent->rb_right;
1198                 else {
1199                         /*
1200                          * Lock and unlock the stable_node's page (which
1201                          * might already have been migrated) so that page
1202                          * migration is sure to notice its raised count.
1203                          * It would be more elegant to return stable_node
1204                          * than kpage, but that involves more changes.
1205                          */
1206                         tree_page = get_ksm_page(stable_node, true);
1207                         if (tree_page) {
1208                                 unlock_page(tree_page);
1209                                 if (get_kpfn_nid(stable_node->kpfn) !=
1210                                                 NUMA(stable_node->nid)) {
1211                                         put_page(tree_page);
1212                                         goto replace;
1213                                 }
1214                                 return tree_page;
1215                         }
1216                         /*
1217                          * There is now a place for page_node, but the tree may
1218                          * have been rebalanced, so re-evaluate parent and new.
1219                          */
1220                         if (page_node)
1221                                 goto again;
1222                         return NULL;
1223                 }
1224         }
1225
1226         if (!page_node)
1227                 return NULL;
1228
1229         list_del(&page_node->list);
1230         DO_NUMA(page_node->nid = nid);
1231         rb_link_node(&page_node->node, parent, new);
1232         rb_insert_color(&page_node->node, root);
1233         get_page(page);
1234         return page;
1235
1236 replace:
1237         if (page_node) {
1238                 list_del(&page_node->list);
1239                 DO_NUMA(page_node->nid = nid);
1240                 rb_replace_node(&stable_node->node, &page_node->node, root);
1241                 get_page(page);
1242         } else {
1243                 rb_erase(&stable_node->node, root);
1244                 page = NULL;
1245         }
1246         stable_node->head = &migrate_nodes;
1247         list_add(&stable_node->list, stable_node->head);
1248         return page;
1249 }
1250
1251 /*
1252  * stable_tree_insert - insert stable tree node pointing to new ksm page
1253  * into the stable tree.
1254  *
1255  * This function returns the stable tree node just allocated on success,
1256  * NULL otherwise.
1257  */
1258 static struct stable_node *stable_tree_insert(struct page *kpage)
1259 {
1260         int nid;
1261         unsigned long kpfn;
1262         struct rb_root *root;
1263         struct rb_node **new;
1264         struct rb_node *parent;
1265         struct stable_node *stable_node;
1266
1267         kpfn = page_to_pfn(kpage);
1268         nid = get_kpfn_nid(kpfn);
1269         root = root_stable_tree + nid;
1270 again:
1271         parent = NULL;
1272         new = &root->rb_node;
1273
1274         while (*new) {
1275                 struct page *tree_page;
1276                 int ret;
1277
1278                 cond_resched();
1279                 stable_node = rb_entry(*new, struct stable_node, node);
1280                 tree_page = get_ksm_page(stable_node, false);
1281                 if (!tree_page) {
1282                         /*
1283                          * If we walked over a stale stable_node,
1284                          * get_ksm_page() will call rb_erase() and it
1285                          * may rebalance the tree from under us. So
1286                          * restart the search from scratch. Returning
1287                          * NULL would be safe too, but we'd generate
1288                          * false negative insertions just because some
1289                          * stable_node was stale.
1290                          */
1291                         goto again;
1292                 }
1293
1294                 ret = memcmp_pages(kpage, tree_page);
1295                 put_page(tree_page);
1296
1297                 parent = *new;
1298                 if (ret < 0)
1299                         new = &parent->rb_left;
1300                 else if (ret > 0)
1301                         new = &parent->rb_right;
1302                 else {
1303                         /*
1304                          * It is not a bug that stable_tree_search() didn't
1305                          * find this node: because at that time our page was
1306                          * not yet write-protected, so may have changed since.
1307                          */
1308                         return NULL;
1309                 }
1310         }
1311
1312         stable_node = alloc_stable_node();
1313         if (!stable_node)
1314                 return NULL;
1315
1316         INIT_HLIST_HEAD(&stable_node->hlist);
1317         stable_node->kpfn = kpfn;
1318         set_page_stable_node(kpage, stable_node);
1319         DO_NUMA(stable_node->nid = nid);
1320         rb_link_node(&stable_node->node, parent, new);
1321         rb_insert_color(&stable_node->node, root);
1322
1323         return stable_node;
1324 }
1325
1326 /*
1327  * unstable_tree_search_insert - search for identical page,
1328  * else insert rmap_item into the unstable tree.
1329  *
1330  * This function searches for a page in the unstable tree identical to the
1331  * page currently being scanned; and if no identical page is found in the
1332  * tree, we insert rmap_item as a new object into the unstable tree.
1333  *
1334  * This function returns pointer to rmap_item found to be identical
1335  * to the currently scanned page, NULL otherwise.
1336  *
1337  * This function does both searching and inserting, because they share
1338  * the same walking algorithm in an rbtree.
1339  */
1340 static
1341 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1342                                               struct page *page,
1343                                               struct page **tree_pagep)
1344 {
1345         struct rb_node **new;
1346         struct rb_root *root;
1347         struct rb_node *parent = NULL;
1348         int nid;
1349
1350         nid = get_kpfn_nid(page_to_pfn(page));
1351         root = root_unstable_tree + nid;
1352         new = &root->rb_node;
1353
1354         while (*new) {
1355                 struct rmap_item *tree_rmap_item;
1356                 struct page *tree_page;
1357                 int ret;
1358
1359                 cond_resched();
1360                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1361                 tree_page = get_mergeable_page(tree_rmap_item);
1362                 if (!tree_page)
1363                         return NULL;
1364
1365                 /*
1366                  * Don't substitute a ksm page for a forked page.
1367                  */
1368                 if (page == tree_page) {
1369                         put_page(tree_page);
1370                         return NULL;
1371                 }
1372
1373                 ret = memcmp_pages(page, tree_page);
1374
1375                 parent = *new;
1376                 if (ret < 0) {
1377                         put_page(tree_page);
1378                         new = &parent->rb_left;
1379                 } else if (ret > 0) {
1380                         put_page(tree_page);
1381                         new = &parent->rb_right;
1382                 } else if (!ksm_merge_across_nodes &&
1383                            page_to_nid(tree_page) != nid) {
1384                         /*
1385                          * If tree_page has been migrated to another NUMA node,
1386                          * it will be flushed out and put in the right unstable
1387                          * tree next time: only merge with it when across_nodes.
1388                          */
1389                         put_page(tree_page);
1390                         return NULL;
1391                 } else {
1392                         *tree_pagep = tree_page;
1393                         return tree_rmap_item;
1394                 }
1395         }
1396
1397         rmap_item->address |= UNSTABLE_FLAG;
1398         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1399         DO_NUMA(rmap_item->nid = nid);
1400         rb_link_node(&rmap_item->node, parent, new);
1401         rb_insert_color(&rmap_item->node, root);
1402
1403         ksm_pages_unshared++;
1404         return NULL;
1405 }
1406
1407 /*
1408  * stable_tree_append - add another rmap_item to the linked list of
1409  * rmap_items hanging off a given node of the stable tree, all sharing
1410  * the same ksm page.
1411  */
1412 static void stable_tree_append(struct rmap_item *rmap_item,
1413                                struct stable_node *stable_node)
1414 {
1415         rmap_item->head = stable_node;
1416         rmap_item->address |= STABLE_FLAG;
1417         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1418
1419         if (rmap_item->hlist.next)
1420                 ksm_pages_sharing++;
1421         else
1422                 ksm_pages_shared++;
1423 }
1424
1425 /*
1426  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1427  * if not, compare checksum to previous and if it's the same, see if page can
1428  * be inserted into the unstable tree, or merged with a page already there and
1429  * both transferred to the stable tree.
1430  *
1431  * @page: the page that we are searching identical page to.
1432  * @rmap_item: the reverse mapping into the virtual address of this page
1433  */
1434 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1435 {
1436         struct rmap_item *tree_rmap_item;
1437         struct page *tree_page = NULL;
1438         struct stable_node *stable_node;
1439         struct page *kpage;
1440         unsigned int checksum;
1441         int err;
1442
1443         stable_node = page_stable_node(page);
1444         if (stable_node) {
1445                 if (stable_node->head != &migrate_nodes &&
1446                     get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1447                         rb_erase(&stable_node->node,
1448                                  root_stable_tree + NUMA(stable_node->nid));
1449                         stable_node->head = &migrate_nodes;
1450                         list_add(&stable_node->list, stable_node->head);
1451                 }
1452                 if (stable_node->head != &migrate_nodes &&
1453                     rmap_item->head == stable_node)
1454                         return;
1455         }
1456
1457         /* We first start with searching the page inside the stable tree */
1458         kpage = stable_tree_search(page);
1459         if (kpage == page && rmap_item->head == stable_node) {
1460                 put_page(kpage);
1461                 return;
1462         }
1463
1464         remove_rmap_item_from_tree(rmap_item);
1465
1466         if (kpage) {
1467                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
1468                 if (!err) {
1469                         /*
1470                          * The page was successfully merged:
1471                          * add its rmap_item to the stable tree.
1472                          */
1473                         lock_page(kpage);
1474                         stable_tree_append(rmap_item, page_stable_node(kpage));
1475                         unlock_page(kpage);
1476                 }
1477                 put_page(kpage);
1478                 return;
1479         }
1480
1481         /*
1482          * If the hash value of the page has changed from the last time
1483          * we calculated it, this page is changing frequently: therefore we
1484          * don't want to insert it in the unstable tree, and we don't want
1485          * to waste our time searching for something identical to it there.
1486          */
1487         checksum = calc_checksum(page);
1488         if (rmap_item->oldchecksum != checksum) {
1489                 rmap_item->oldchecksum = checksum;
1490                 return;
1491         }
1492
1493         tree_rmap_item =
1494                 unstable_tree_search_insert(rmap_item, page, &tree_page);
1495         if (tree_rmap_item) {
1496                 kpage = try_to_merge_two_pages(rmap_item, page,
1497                                                 tree_rmap_item, tree_page);
1498                 put_page(tree_page);
1499                 if (kpage) {
1500                         /*
1501                          * The pages were successfully merged: insert new
1502                          * node in the stable tree and add both rmap_items.
1503                          */
1504                         lock_page(kpage);
1505                         stable_node = stable_tree_insert(kpage);
1506                         if (stable_node) {
1507                                 stable_tree_append(tree_rmap_item, stable_node);
1508                                 stable_tree_append(rmap_item, stable_node);
1509                         }
1510                         unlock_page(kpage);
1511
1512                         /*
1513                          * If we fail to insert the page into the stable tree,
1514                          * we will have 2 virtual addresses that are pointing
1515                          * to a ksm page left outside the stable tree,
1516                          * in which case we need to break_cow on both.
1517                          */
1518                         if (!stable_node) {
1519                                 break_cow(tree_rmap_item);
1520                                 break_cow(rmap_item);
1521                         }
1522                 }
1523         }
1524 }
1525
1526 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
1527                                             struct rmap_item **rmap_list,
1528                                             unsigned long addr)
1529 {
1530         struct rmap_item *rmap_item;
1531
1532         while (*rmap_list) {
1533                 rmap_item = *rmap_list;
1534                 if ((rmap_item->address & PAGE_MASK) == addr)
1535                         return rmap_item;
1536                 if (rmap_item->address > addr)
1537                         break;
1538                 *rmap_list = rmap_item->rmap_list;
1539                 remove_rmap_item_from_tree(rmap_item);
1540                 free_rmap_item(rmap_item);
1541         }
1542
1543         rmap_item = alloc_rmap_item();
1544         if (rmap_item) {
1545                 /* It has already been zeroed */
1546                 rmap_item->mm = mm_slot->mm;
1547                 rmap_item->address = addr;
1548                 rmap_item->rmap_list = *rmap_list;
1549                 *rmap_list = rmap_item;
1550         }
1551         return rmap_item;
1552 }
1553
1554 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1555 {
1556         struct mm_struct *mm;
1557         struct mm_slot *slot;
1558         struct vm_area_struct *vma;
1559         struct rmap_item *rmap_item;
1560         int nid;
1561
1562         if (list_empty(&ksm_mm_head.mm_list))
1563                 return NULL;
1564
1565         slot = ksm_scan.mm_slot;
1566         if (slot == &ksm_mm_head) {
1567                 /*
1568                  * A number of pages can hang around indefinitely on per-cpu
1569                  * pagevecs, raised page count preventing write_protect_page
1570                  * from merging them.  Though it doesn't really matter much,
1571                  * it is puzzling to see some stuck in pages_volatile until
1572                  * other activity jostles them out, and they also prevented
1573                  * LTP's KSM test from succeeding deterministically; so drain
1574                  * them here (here rather than on entry to ksm_do_scan(),
1575                  * so we don't IPI too often when pages_to_scan is set low).
1576                  */
1577                 lru_add_drain_all();
1578
1579                 /*
1580                  * Whereas stale stable_nodes on the stable_tree itself
1581                  * get pruned in the regular course of stable_tree_search(),
1582                  * those moved out to the migrate_nodes list can accumulate:
1583                  * so prune them once before each full scan.
1584                  */
1585                 if (!ksm_merge_across_nodes) {
1586                         struct stable_node *stable_node;
1587                         struct list_head *this, *next;
1588                         struct page *page;
1589
1590                         list_for_each_safe(this, next, &migrate_nodes) {
1591                                 stable_node = list_entry(this,
1592                                                 struct stable_node, list);
1593                                 page = get_ksm_page(stable_node, false);
1594                                 if (page)
1595                                         put_page(page);
1596                                 cond_resched();
1597                         }
1598                 }
1599
1600                 for (nid = 0; nid < ksm_nr_node_ids; nid++)
1601                         root_unstable_tree[nid] = RB_ROOT;
1602
1603                 spin_lock(&ksm_mmlist_lock);
1604                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1605                 ksm_scan.mm_slot = slot;
1606                 spin_unlock(&ksm_mmlist_lock);
1607                 /*
1608                  * Although we tested list_empty() above, a racing __ksm_exit
1609                  * of the last mm on the list may have removed it since then.
1610                  */
1611                 if (slot == &ksm_mm_head)
1612                         return NULL;
1613 next_mm:
1614                 ksm_scan.address = 0;
1615                 ksm_scan.rmap_list = &slot->rmap_list;
1616         }
1617
1618         mm = slot->mm;
1619         down_read(&mm->mmap_sem);
1620         if (ksm_test_exit(mm))
1621                 vma = NULL;
1622         else
1623                 vma = find_vma(mm, ksm_scan.address);
1624
1625         for (; vma; vma = vma->vm_next) {
1626                 if (!(vma->vm_flags & VM_MERGEABLE))
1627                         continue;
1628                 if (ksm_scan.address < vma->vm_start)
1629                         ksm_scan.address = vma->vm_start;
1630                 if (!vma->anon_vma)
1631                         ksm_scan.address = vma->vm_end;
1632
1633                 while (ksm_scan.address < vma->vm_end) {
1634                         if (ksm_test_exit(mm))
1635                                 break;
1636                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1637                         if (IS_ERR_OR_NULL(*page)) {
1638                                 ksm_scan.address += PAGE_SIZE;
1639                                 cond_resched();
1640                                 continue;
1641                         }
1642                         if (PageAnon(*page) ||
1643                             page_trans_compound_anon(*page)) {
1644                                 flush_anon_page(vma, *page, ksm_scan.address);
1645                                 flush_dcache_page(*page);
1646                                 rmap_item = get_next_rmap_item(slot,
1647                                         ksm_scan.rmap_list, ksm_scan.address);
1648                                 if (rmap_item) {
1649                                         ksm_scan.rmap_list =
1650                                                         &rmap_item->rmap_list;
1651                                         ksm_scan.address += PAGE_SIZE;
1652                                 } else
1653                                         put_page(*page);
1654                                 up_read(&mm->mmap_sem);
1655                                 return rmap_item;
1656                         }
1657                         put_page(*page);
1658                         ksm_scan.address += PAGE_SIZE;
1659                         cond_resched();
1660                 }
1661         }
1662
1663         if (ksm_test_exit(mm)) {
1664                 ksm_scan.address = 0;
1665                 ksm_scan.rmap_list = &slot->rmap_list;
1666         }
1667         /*
1668          * Nuke all the rmap_items that are above this current rmap:
1669          * because there were no VM_MERGEABLE vmas with such addresses.
1670          */
1671         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
1672
1673         spin_lock(&ksm_mmlist_lock);
1674         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1675                                                 struct mm_slot, mm_list);
1676         if (ksm_scan.address == 0) {
1677                 /*
1678                  * We've completed a full scan of all vmas, holding mmap_sem
1679                  * throughout, and found no VM_MERGEABLE: so do the same as
1680                  * __ksm_exit does to remove this mm from all our lists now.
1681                  * This applies either when cleaning up after __ksm_exit
1682                  * (but beware: we can reach here even before __ksm_exit),
1683                  * or when all VM_MERGEABLE areas have been unmapped (and
1684                  * mmap_sem then protects against race with MADV_MERGEABLE).
1685                  */
1686                 hash_del(&slot->link);
1687                 list_del(&slot->mm_list);
1688                 spin_unlock(&ksm_mmlist_lock);
1689
1690                 free_mm_slot(slot);
1691                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1692                 up_read(&mm->mmap_sem);
1693                 mmdrop(mm);
1694         } else {
1695                 spin_unlock(&ksm_mmlist_lock);
1696                 up_read(&mm->mmap_sem);
1697         }
1698
1699         /* Repeat until we've completed scanning the whole list */
1700         slot = ksm_scan.mm_slot;
1701         if (slot != &ksm_mm_head)
1702                 goto next_mm;
1703
1704         ksm_scan.seqnr++;
1705         return NULL;
1706 }
1707
1708 /**
1709  * ksm_do_scan  - the ksm scanner main worker function.
1710  * @scan_npages - number of pages we want to scan before we return.
1711  */
1712 static void ksm_do_scan(unsigned int scan_npages)
1713 {
1714         struct rmap_item *rmap_item;
1715         struct page *uninitialized_var(page);
1716
1717         while (scan_npages-- && likely(!freezing(current))) {
1718                 cond_resched();
1719                 rmap_item = scan_get_next_rmap_item(&page);
1720                 if (!rmap_item)
1721                         return;
1722                 cmp_and_merge_page(page, rmap_item);
1723                 put_page(page);
1724         }
1725 }
1726
1727 static int ksmd_should_run(void)
1728 {
1729         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1730 }
1731
1732 static int ksm_scan_thread(void *nothing)
1733 {
1734         set_freezable();
1735         set_user_nice(current, 5);
1736
1737         while (!kthread_should_stop()) {
1738                 mutex_lock(&ksm_thread_mutex);
1739                 wait_while_offlining();
1740                 if (ksmd_should_run())
1741                         ksm_do_scan(ksm_thread_pages_to_scan);
1742                 mutex_unlock(&ksm_thread_mutex);
1743
1744                 try_to_freeze();
1745
1746                 if (ksmd_should_run()) {
1747                         schedule_timeout_interruptible(
1748                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1749                 } else {
1750                         wait_event_freezable(ksm_thread_wait,
1751                                 ksmd_should_run() || kthread_should_stop());
1752                 }
1753         }
1754         return 0;
1755 }
1756
1757 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1758                 unsigned long end, int advice, unsigned long *vm_flags)
1759 {
1760         struct mm_struct *mm = vma->vm_mm;
1761         int err;
1762
1763         switch (advice) {
1764         case MADV_MERGEABLE:
1765                 /*
1766                  * Be somewhat over-protective for now!
1767                  */
1768                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
1769                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
1770                                  VM_HUGETLB | VM_MIXEDMAP))
1771                         return 0;               /* just ignore the advice */
1772
1773 #ifdef VM_SAO
1774                 if (*vm_flags & VM_SAO)
1775                         return 0;
1776 #endif
1777
1778                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1779                         err = __ksm_enter(mm);
1780                         if (err)
1781                                 return err;
1782                 }
1783
1784                 *vm_flags |= VM_MERGEABLE;
1785                 break;
1786
1787         case MADV_UNMERGEABLE:
1788                 if (!(*vm_flags & VM_MERGEABLE))
1789                         return 0;               /* just ignore the advice */
1790
1791                 if (vma->anon_vma) {
1792                         err = unmerge_ksm_pages(vma, start, end);
1793                         if (err)
1794                                 return err;
1795                 }
1796
1797                 *vm_flags &= ~VM_MERGEABLE;
1798                 break;
1799         }
1800
1801         return 0;
1802 }
1803
1804 int __ksm_enter(struct mm_struct *mm)
1805 {
1806         struct mm_slot *mm_slot;
1807         int needs_wakeup;
1808
1809         mm_slot = alloc_mm_slot();
1810         if (!mm_slot)
1811                 return -ENOMEM;
1812
1813         /* Check ksm_run too?  Would need tighter locking */
1814         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1815
1816         spin_lock(&ksm_mmlist_lock);
1817         insert_to_mm_slots_hash(mm, mm_slot);
1818         /*
1819          * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1820          * insert just behind the scanning cursor, to let the area settle
1821          * down a little; when fork is followed by immediate exec, we don't
1822          * want ksmd to waste time setting up and tearing down an rmap_list.
1823          *
1824          * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1825          * scanning cursor, otherwise KSM pages in newly forked mms will be
1826          * missed: then we might as well insert at the end of the list.
1827          */
1828         if (ksm_run & KSM_RUN_UNMERGE)
1829                 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1830         else
1831                 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1832         spin_unlock(&ksm_mmlist_lock);
1833
1834         set_bit(MMF_VM_MERGEABLE, &mm->flags);
1835         atomic_inc(&mm->mm_count);
1836
1837         if (needs_wakeup)
1838                 wake_up_interruptible(&ksm_thread_wait);
1839
1840         return 0;
1841 }
1842
1843 void __ksm_exit(struct mm_struct *mm)
1844 {
1845         struct mm_slot *mm_slot;
1846         int easy_to_free = 0;
1847
1848         /*
1849          * This process is exiting: if it's straightforward (as is the
1850          * case when ksmd was never running), free mm_slot immediately.
1851          * But if it's at the cursor or has rmap_items linked to it, use
1852          * mmap_sem to synchronize with any break_cows before pagetables
1853          * are freed, and leave the mm_slot on the list for ksmd to free.
1854          * Beware: ksm may already have noticed it exiting and freed the slot.
1855          */
1856
1857         spin_lock(&ksm_mmlist_lock);
1858         mm_slot = get_mm_slot(mm);
1859         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
1860                 if (!mm_slot->rmap_list) {
1861                         hash_del(&mm_slot->link);
1862                         list_del(&mm_slot->mm_list);
1863                         easy_to_free = 1;
1864                 } else {
1865                         list_move(&mm_slot->mm_list,
1866                                   &ksm_scan.mm_slot->mm_list);
1867                 }
1868         }
1869         spin_unlock(&ksm_mmlist_lock);
1870
1871         if (easy_to_free) {
1872                 free_mm_slot(mm_slot);
1873                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1874                 mmdrop(mm);
1875         } else if (mm_slot) {
1876                 down_write(&mm->mmap_sem);
1877                 up_write(&mm->mmap_sem);
1878         }
1879 }
1880
1881 struct page *ksm_might_need_to_copy(struct page *page,
1882                         struct vm_area_struct *vma, unsigned long address)
1883 {
1884         struct anon_vma *anon_vma = page_anon_vma(page);
1885         struct page *new_page;
1886
1887         if (PageKsm(page)) {
1888                 if (page_stable_node(page) &&
1889                     !(ksm_run & KSM_RUN_UNMERGE))
1890                         return page;    /* no need to copy it */
1891         } else if (!anon_vma) {
1892                 return page;            /* no need to copy it */
1893         } else if (anon_vma->root == vma->anon_vma->root &&
1894                  page->index == linear_page_index(vma, address)) {
1895                 return page;            /* still no need to copy it */
1896         }
1897         if (!PageUptodate(page))
1898                 return page;            /* let do_swap_page report the error */
1899
1900         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1901         if (new_page) {
1902                 copy_user_highpage(new_page, page, address, vma);
1903
1904                 SetPageDirty(new_page);
1905                 __SetPageUptodate(new_page);
1906                 __set_page_locked(new_page);
1907         }
1908
1909         return new_page;
1910 }
1911
1912 int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
1913 {
1914         struct stable_node *stable_node;
1915         struct rmap_item *rmap_item;
1916         int ret = SWAP_AGAIN;
1917         int search_new_forks = 0;
1918
1919         VM_BUG_ON_PAGE(!PageKsm(page), page);
1920
1921         /*
1922          * Rely on the page lock to protect against concurrent modifications
1923          * to that page's node of the stable tree.
1924          */
1925         VM_BUG_ON_PAGE(!PageLocked(page), page);
1926
1927         stable_node = page_stable_node(page);
1928         if (!stable_node)
1929                 return ret;
1930 again:
1931         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
1932                 struct anon_vma *anon_vma = rmap_item->anon_vma;
1933                 struct anon_vma_chain *vmac;
1934                 struct vm_area_struct *vma;
1935
1936                 cond_resched();
1937                 anon_vma_lock_read(anon_vma);
1938                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1939                                                0, ULONG_MAX) {
1940                         cond_resched();
1941                         vma = vmac->vma;
1942                         if (rmap_item->address < vma->vm_start ||
1943                             rmap_item->address >= vma->vm_end)
1944                                 continue;
1945                         /*
1946                          * Initially we examine only the vma which covers this
1947                          * rmap_item; but later, if there is still work to do,
1948                          * we examine covering vmas in other mms: in case they
1949                          * were forked from the original since ksmd passed.
1950                          */
1951                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1952                                 continue;
1953
1954                         if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1955                                 continue;
1956
1957                         ret = rwc->rmap_one(page, vma,
1958                                         rmap_item->address, rwc->arg);
1959                         if (ret != SWAP_AGAIN) {
1960                                 anon_vma_unlock_read(anon_vma);
1961                                 goto out;
1962                         }
1963                         if (rwc->done && rwc->done(page)) {
1964                                 anon_vma_unlock_read(anon_vma);
1965                                 goto out;
1966                         }
1967                 }
1968                 anon_vma_unlock_read(anon_vma);
1969         }
1970         if (!search_new_forks++)
1971                 goto again;
1972 out:
1973         return ret;
1974 }
1975
1976 #ifdef CONFIG_MIGRATION
1977 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1978 {
1979         struct stable_node *stable_node;
1980
1981         VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
1982         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1983         VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
1984
1985         stable_node = page_stable_node(newpage);
1986         if (stable_node) {
1987                 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
1988                 stable_node->kpfn = page_to_pfn(newpage);
1989                 /*
1990                  * newpage->mapping was set in advance; now we need smp_wmb()
1991                  * to make sure that the new stable_node->kpfn is visible
1992                  * to get_ksm_page() before it can see that oldpage->mapping
1993                  * has gone stale (or that PageSwapCache has been cleared).
1994                  */
1995                 smp_wmb();
1996                 set_page_stable_node(oldpage, NULL);
1997         }
1998 }
1999 #endif /* CONFIG_MIGRATION */
2000
2001 #ifdef CONFIG_MEMORY_HOTREMOVE
2002 static void wait_while_offlining(void)
2003 {
2004         while (ksm_run & KSM_RUN_OFFLINE) {
2005                 mutex_unlock(&ksm_thread_mutex);
2006                 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2007                             TASK_UNINTERRUPTIBLE);
2008                 mutex_lock(&ksm_thread_mutex);
2009         }
2010 }
2011
2012 static void ksm_check_stable_tree(unsigned long start_pfn,
2013                                   unsigned long end_pfn)
2014 {
2015         struct stable_node *stable_node;
2016         struct list_head *this, *next;
2017         struct rb_node *node;
2018         int nid;
2019
2020         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2021                 node = rb_first(root_stable_tree + nid);
2022                 while (node) {
2023                         stable_node = rb_entry(node, struct stable_node, node);
2024                         if (stable_node->kpfn >= start_pfn &&
2025                             stable_node->kpfn < end_pfn) {
2026                                 /*
2027                                  * Don't get_ksm_page, page has already gone:
2028                                  * which is why we keep kpfn instead of page*
2029                                  */
2030                                 remove_node_from_stable_tree(stable_node);
2031                                 node = rb_first(root_stable_tree + nid);
2032                         } else
2033                                 node = rb_next(node);
2034                         cond_resched();
2035                 }
2036         }
2037         list_for_each_safe(this, next, &migrate_nodes) {
2038                 stable_node = list_entry(this, struct stable_node, list);
2039                 if (stable_node->kpfn >= start_pfn &&
2040                     stable_node->kpfn < end_pfn)
2041                         remove_node_from_stable_tree(stable_node);
2042                 cond_resched();
2043         }
2044 }
2045
2046 static int ksm_memory_callback(struct notifier_block *self,
2047                                unsigned long action, void *arg)
2048 {
2049         struct memory_notify *mn = arg;
2050
2051         switch (action) {
2052         case MEM_GOING_OFFLINE:
2053                 /*
2054                  * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2055                  * and remove_all_stable_nodes() while memory is going offline:
2056                  * it is unsafe for them to touch the stable tree at this time.
2057                  * But unmerge_ksm_pages(), rmap lookups and other entry points
2058                  * which do not need the ksm_thread_mutex are all safe.
2059                  */
2060                 mutex_lock(&ksm_thread_mutex);
2061                 ksm_run |= KSM_RUN_OFFLINE;
2062                 mutex_unlock(&ksm_thread_mutex);
2063                 break;
2064
2065         case MEM_OFFLINE:
2066                 /*
2067                  * Most of the work is done by page migration; but there might
2068                  * be a few stable_nodes left over, still pointing to struct
2069                  * pages which have been offlined: prune those from the tree,
2070                  * otherwise get_ksm_page() might later try to access a
2071                  * non-existent struct page.
2072                  */
2073                 ksm_check_stable_tree(mn->start_pfn,
2074                                       mn->start_pfn + mn->nr_pages);
2075                 /* fallthrough */
2076
2077         case MEM_CANCEL_OFFLINE:
2078                 mutex_lock(&ksm_thread_mutex);
2079                 ksm_run &= ~KSM_RUN_OFFLINE;
2080                 mutex_unlock(&ksm_thread_mutex);
2081
2082                 smp_mb();       /* wake_up_bit advises this */
2083                 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2084                 break;
2085         }
2086         return NOTIFY_OK;
2087 }
2088 #else
2089 static void wait_while_offlining(void)
2090 {
2091 }
2092 #endif /* CONFIG_MEMORY_HOTREMOVE */
2093
2094 #ifdef CONFIG_SYSFS
2095 /*
2096  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2097  */
2098
2099 #define KSM_ATTR_RO(_name) \
2100         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2101 #define KSM_ATTR(_name) \
2102         static struct kobj_attribute _name##_attr = \
2103                 __ATTR(_name, 0644, _name##_show, _name##_store)
2104
2105 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2106                                     struct kobj_attribute *attr, char *buf)
2107 {
2108         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2109 }
2110
2111 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2112                                      struct kobj_attribute *attr,
2113                                      const char *buf, size_t count)
2114 {
2115         unsigned long msecs;
2116         int err;
2117
2118         err = kstrtoul(buf, 10, &msecs);
2119         if (err || msecs > UINT_MAX)
2120                 return -EINVAL;
2121
2122         ksm_thread_sleep_millisecs = msecs;
2123
2124         return count;
2125 }
2126 KSM_ATTR(sleep_millisecs);
2127
2128 static ssize_t pages_to_scan_show(struct kobject *kobj,
2129                                   struct kobj_attribute *attr, char *buf)
2130 {
2131         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2132 }
2133
2134 static ssize_t pages_to_scan_store(struct kobject *kobj,
2135                                    struct kobj_attribute *attr,
2136                                    const char *buf, size_t count)
2137 {
2138         int err;
2139         unsigned long nr_pages;
2140
2141         err = kstrtoul(buf, 10, &nr_pages);
2142         if (err || nr_pages > UINT_MAX)
2143                 return -EINVAL;
2144
2145         ksm_thread_pages_to_scan = nr_pages;
2146
2147         return count;
2148 }
2149 KSM_ATTR(pages_to_scan);
2150
2151 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2152                         char *buf)
2153 {
2154         return sprintf(buf, "%lu\n", ksm_run);
2155 }
2156
2157 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2158                          const char *buf, size_t count)
2159 {
2160         int err;
2161         unsigned long flags;
2162
2163         err = kstrtoul(buf, 10, &flags);
2164         if (err || flags > UINT_MAX)
2165                 return -EINVAL;
2166         if (flags > KSM_RUN_UNMERGE)
2167                 return -EINVAL;
2168
2169         /*
2170          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2171          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2172          * breaking COW to free the pages_shared (but leaves mm_slots
2173          * on the list for when ksmd may be set running again).
2174          */
2175
2176         mutex_lock(&ksm_thread_mutex);
2177         wait_while_offlining();
2178         if (ksm_run != flags) {
2179                 ksm_run = flags;
2180                 if (flags & KSM_RUN_UNMERGE) {
2181                         set_current_oom_origin();
2182                         err = unmerge_and_remove_all_rmap_items();
2183                         clear_current_oom_origin();
2184                         if (err) {
2185                                 ksm_run = KSM_RUN_STOP;
2186                                 count = err;
2187                         }
2188                 }
2189         }
2190         mutex_unlock(&ksm_thread_mutex);
2191
2192         if (flags & KSM_RUN_MERGE)
2193                 wake_up_interruptible(&ksm_thread_wait);
2194
2195         return count;
2196 }
2197 KSM_ATTR(run);
2198
2199 #ifdef CONFIG_NUMA
2200 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2201                                 struct kobj_attribute *attr, char *buf)
2202 {
2203         return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2204 }
2205
2206 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2207                                    struct kobj_attribute *attr,
2208                                    const char *buf, size_t count)
2209 {
2210         int err;
2211         unsigned long knob;
2212
2213         err = kstrtoul(buf, 10, &knob);
2214         if (err)
2215                 return err;
2216         if (knob > 1)
2217                 return -EINVAL;
2218
2219         mutex_lock(&ksm_thread_mutex);
2220         wait_while_offlining();
2221         if (ksm_merge_across_nodes != knob) {
2222                 if (ksm_pages_shared || remove_all_stable_nodes())
2223                         err = -EBUSY;
2224                 else if (root_stable_tree == one_stable_tree) {
2225                         struct rb_root *buf;
2226                         /*
2227                          * This is the first time that we switch away from the
2228                          * default of merging across nodes: must now allocate
2229                          * a buffer to hold as many roots as may be needed.
2230                          * Allocate stable and unstable together:
2231                          * MAXSMP NODES_SHIFT 10 will use 16kB.
2232                          */
2233                         buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2234                                       GFP_KERNEL);
2235                         /* Let us assume that RB_ROOT is NULL is zero */
2236                         if (!buf)
2237                                 err = -ENOMEM;
2238                         else {
2239                                 root_stable_tree = buf;
2240                                 root_unstable_tree = buf + nr_node_ids;
2241                                 /* Stable tree is empty but not the unstable */
2242                                 root_unstable_tree[0] = one_unstable_tree[0];
2243                         }
2244                 }
2245                 if (!err) {
2246                         ksm_merge_across_nodes = knob;
2247                         ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2248                 }
2249         }
2250         mutex_unlock(&ksm_thread_mutex);
2251
2252         return err ? err : count;
2253 }
2254 KSM_ATTR(merge_across_nodes);
2255 #endif
2256
2257 static ssize_t pages_shared_show(struct kobject *kobj,
2258                                  struct kobj_attribute *attr, char *buf)
2259 {
2260         return sprintf(buf, "%lu\n", ksm_pages_shared);
2261 }
2262 KSM_ATTR_RO(pages_shared);
2263
2264 static ssize_t pages_sharing_show(struct kobject *kobj,
2265                                   struct kobj_attribute *attr, char *buf)
2266 {
2267         return sprintf(buf, "%lu\n", ksm_pages_sharing);
2268 }
2269 KSM_ATTR_RO(pages_sharing);
2270
2271 static ssize_t pages_unshared_show(struct kobject *kobj,
2272                                    struct kobj_attribute *attr, char *buf)
2273 {
2274         return sprintf(buf, "%lu\n", ksm_pages_unshared);
2275 }
2276 KSM_ATTR_RO(pages_unshared);
2277
2278 static ssize_t pages_volatile_show(struct kobject *kobj,
2279                                    struct kobj_attribute *attr, char *buf)
2280 {
2281         long ksm_pages_volatile;
2282
2283         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2284                                 - ksm_pages_sharing - ksm_pages_unshared;
2285         /*
2286          * It was not worth any locking to calculate that statistic,
2287          * but it might therefore sometimes be negative: conceal that.
2288          */
2289         if (ksm_pages_volatile < 0)
2290                 ksm_pages_volatile = 0;
2291         return sprintf(buf, "%ld\n", ksm_pages_volatile);
2292 }
2293 KSM_ATTR_RO(pages_volatile);
2294
2295 static ssize_t full_scans_show(struct kobject *kobj,
2296                                struct kobj_attribute *attr, char *buf)
2297 {
2298         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2299 }
2300 KSM_ATTR_RO(full_scans);
2301
2302 static struct attribute *ksm_attrs[] = {
2303         &sleep_millisecs_attr.attr,
2304         &pages_to_scan_attr.attr,
2305         &run_attr.attr,
2306         &pages_shared_attr.attr,
2307         &pages_sharing_attr.attr,
2308         &pages_unshared_attr.attr,
2309         &pages_volatile_attr.attr,
2310         &full_scans_attr.attr,
2311 #ifdef CONFIG_NUMA
2312         &merge_across_nodes_attr.attr,
2313 #endif
2314         NULL,
2315 };
2316
2317 static struct attribute_group ksm_attr_group = {
2318         .attrs = ksm_attrs,
2319         .name = "ksm",
2320 };
2321 #endif /* CONFIG_SYSFS */
2322
2323 static int __init ksm_init(void)
2324 {
2325         struct task_struct *ksm_thread;
2326         int err;
2327
2328         err = ksm_slab_init();
2329         if (err)
2330                 goto out;
2331
2332         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2333         if (IS_ERR(ksm_thread)) {
2334                 pr_err("ksm: creating kthread failed\n");
2335                 err = PTR_ERR(ksm_thread);
2336                 goto out_free;
2337         }
2338
2339 #ifdef CONFIG_SYSFS
2340         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2341         if (err) {
2342                 pr_err("ksm: register sysfs failed\n");
2343                 kthread_stop(ksm_thread);
2344                 goto out_free;
2345         }
2346 #else
2347         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
2348
2349 #endif /* CONFIG_SYSFS */
2350
2351 #ifdef CONFIG_MEMORY_HOTREMOVE
2352         /* There is no significance to this priority 100 */
2353         hotplug_memory_notifier(ksm_memory_callback, 100);
2354 #endif
2355         return 0;
2356
2357 out_free:
2358         ksm_slab_free();
2359 out:
2360         return err;
2361 }
2362 subsys_initcall(ksm_init);