]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - mm/ksm.c
Merge branch 'core-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[karo-tx-linux.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/sched/mm.h>
23 #include <linux/sched/coredump.h>
24 #include <linux/rwsem.h>
25 #include <linux/pagemap.h>
26 #include <linux/rmap.h>
27 #include <linux/spinlock.h>
28 #include <linux/jhash.h>
29 #include <linux/delay.h>
30 #include <linux/kthread.h>
31 #include <linux/wait.h>
32 #include <linux/slab.h>
33 #include <linux/rbtree.h>
34 #include <linux/memory.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/swap.h>
37 #include <linux/ksm.h>
38 #include <linux/hashtable.h>
39 #include <linux/freezer.h>
40 #include <linux/oom.h>
41 #include <linux/numa.h>
42
43 #include <asm/tlbflush.h>
44 #include "internal.h"
45
46 #ifdef CONFIG_NUMA
47 #define NUMA(x)         (x)
48 #define DO_NUMA(x)      do { (x); } while (0)
49 #else
50 #define NUMA(x)         (0)
51 #define DO_NUMA(x)      do { } while (0)
52 #endif
53
54 /*
55  * A few notes about the KSM scanning process,
56  * to make it easier to understand the data structures below:
57  *
58  * In order to reduce excessive scanning, KSM sorts the memory pages by their
59  * contents into a data structure that holds pointers to the pages' locations.
60  *
61  * Since the contents of the pages may change at any moment, KSM cannot just
62  * insert the pages into a normal sorted tree and expect it to find anything.
63  * Therefore KSM uses two data structures - the stable and the unstable tree.
64  *
65  * The stable tree holds pointers to all the merged pages (ksm pages), sorted
66  * by their contents.  Because each such page is write-protected, searching on
67  * this tree is fully assured to be working (except when pages are unmapped),
68  * and therefore this tree is called the stable tree.
69  *
70  * In addition to the stable tree, KSM uses a second data structure called the
71  * unstable tree: this tree holds pointers to pages which have been found to
72  * be "unchanged for a period of time".  The unstable tree sorts these pages
73  * by their contents, but since they are not write-protected, KSM cannot rely
74  * upon the unstable tree to work correctly - the unstable tree is liable to
75  * be corrupted as its contents are modified, and so it is called unstable.
76  *
77  * KSM solves this problem by several techniques:
78  *
79  * 1) The unstable tree is flushed every time KSM completes scanning all
80  *    memory areas, and then the tree is rebuilt again from the beginning.
81  * 2) KSM will only insert into the unstable tree, pages whose hash value
82  *    has not changed since the previous scan of all memory areas.
83  * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
84  *    colors of the nodes and not on their contents, assuring that even when
85  *    the tree gets "corrupted" it won't get out of balance, so scanning time
86  *    remains the same (also, searching and inserting nodes in an rbtree uses
87  *    the same algorithm, so we have no overhead when we flush and rebuild).
88  * 4) KSM never flushes the stable tree, which means that even if it were to
89  *    take 10 attempts to find a page in the unstable tree, once it is found,
90  *    it is secured in the stable tree.  (When we scan a new page, we first
91  *    compare it against the stable tree, and then against the unstable tree.)
92  *
93  * If the merge_across_nodes tunable is unset, then KSM maintains multiple
94  * stable trees and multiple unstable trees: one of each for each NUMA node.
95  */
96
97 /**
98  * struct mm_slot - ksm information per mm that is being scanned
99  * @link: link to the mm_slots hash list
100  * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
101  * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
102  * @mm: the mm that this information is valid for
103  */
104 struct mm_slot {
105         struct hlist_node link;
106         struct list_head mm_list;
107         struct rmap_item *rmap_list;
108         struct mm_struct *mm;
109 };
110
111 /**
112  * struct ksm_scan - cursor for scanning
113  * @mm_slot: the current mm_slot we are scanning
114  * @address: the next address inside that to be scanned
115  * @rmap_list: link to the next rmap to be scanned in the rmap_list
116  * @seqnr: count of completed full scans (needed when removing unstable node)
117  *
118  * There is only the one ksm_scan instance of this cursor structure.
119  */
120 struct ksm_scan {
121         struct mm_slot *mm_slot;
122         unsigned long address;
123         struct rmap_item **rmap_list;
124         unsigned long seqnr;
125 };
126
127 /**
128  * struct stable_node - node of the stable rbtree
129  * @node: rb node of this ksm page in the stable tree
130  * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
131  * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
132  * @list: linked into migrate_nodes, pending placement in the proper node tree
133  * @hlist: hlist head of rmap_items using this ksm page
134  * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
135  * @chain_prune_time: time of the last full garbage collection
136  * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
137  * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
138  */
139 struct stable_node {
140         union {
141                 struct rb_node node;    /* when node of stable tree */
142                 struct {                /* when listed for migration */
143                         struct list_head *head;
144                         struct {
145                                 struct hlist_node hlist_dup;
146                                 struct list_head list;
147                         };
148                 };
149         };
150         struct hlist_head hlist;
151         union {
152                 unsigned long kpfn;
153                 unsigned long chain_prune_time;
154         };
155         /*
156          * STABLE_NODE_CHAIN can be any negative number in
157          * rmap_hlist_len negative range, but better not -1 to be able
158          * to reliably detect underflows.
159          */
160 #define STABLE_NODE_CHAIN -1024
161         int rmap_hlist_len;
162 #ifdef CONFIG_NUMA
163         int nid;
164 #endif
165 };
166
167 /**
168  * struct rmap_item - reverse mapping item for virtual addresses
169  * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
170  * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
171  * @nid: NUMA node id of unstable tree in which linked (may not match page)
172  * @mm: the memory structure this rmap_item is pointing into
173  * @address: the virtual address this rmap_item tracks (+ flags in low bits)
174  * @oldchecksum: previous checksum of the page at that virtual address
175  * @node: rb node of this rmap_item in the unstable tree
176  * @head: pointer to stable_node heading this list in the stable tree
177  * @hlist: link into hlist of rmap_items hanging off that stable_node
178  */
179 struct rmap_item {
180         struct rmap_item *rmap_list;
181         union {
182                 struct anon_vma *anon_vma;      /* when stable */
183 #ifdef CONFIG_NUMA
184                 int nid;                /* when node of unstable tree */
185 #endif
186         };
187         struct mm_struct *mm;
188         unsigned long address;          /* + low bits used for flags below */
189         unsigned int oldchecksum;       /* when unstable */
190         union {
191                 struct rb_node node;    /* when node of unstable tree */
192                 struct {                /* when listed from stable tree */
193                         struct stable_node *head;
194                         struct hlist_node hlist;
195                 };
196         };
197 };
198
199 #define SEQNR_MASK      0x0ff   /* low bits of unstable tree seqnr */
200 #define UNSTABLE_FLAG   0x100   /* is a node of the unstable tree */
201 #define STABLE_FLAG     0x200   /* is listed from the stable tree */
202
203 /* The stable and unstable tree heads */
204 static struct rb_root one_stable_tree[1] = { RB_ROOT };
205 static struct rb_root one_unstable_tree[1] = { RB_ROOT };
206 static struct rb_root *root_stable_tree = one_stable_tree;
207 static struct rb_root *root_unstable_tree = one_unstable_tree;
208
209 /* Recently migrated nodes of stable tree, pending proper placement */
210 static LIST_HEAD(migrate_nodes);
211 #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
212
213 #define MM_SLOTS_HASH_BITS 10
214 static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
215
216 static struct mm_slot ksm_mm_head = {
217         .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
218 };
219 static struct ksm_scan ksm_scan = {
220         .mm_slot = &ksm_mm_head,
221 };
222
223 static struct kmem_cache *rmap_item_cache;
224 static struct kmem_cache *stable_node_cache;
225 static struct kmem_cache *mm_slot_cache;
226
227 /* The number of nodes in the stable tree */
228 static unsigned long ksm_pages_shared;
229
230 /* The number of page slots additionally sharing those nodes */
231 static unsigned long ksm_pages_sharing;
232
233 /* The number of nodes in the unstable tree */
234 static unsigned long ksm_pages_unshared;
235
236 /* The number of rmap_items in use: to calculate pages_volatile */
237 static unsigned long ksm_rmap_items;
238
239 /* The number of stable_node chains */
240 static unsigned long ksm_stable_node_chains;
241
242 /* The number of stable_node dups linked to the stable_node chains */
243 static unsigned long ksm_stable_node_dups;
244
245 /* Delay in pruning stale stable_node_dups in the stable_node_chains */
246 static int ksm_stable_node_chains_prune_millisecs = 2000;
247
248 /* Maximum number of page slots sharing a stable node */
249 static int ksm_max_page_sharing = 256;
250
251 /* Number of pages ksmd should scan in one batch */
252 static unsigned int ksm_thread_pages_to_scan = 100;
253
254 /* Milliseconds ksmd should sleep between batches */
255 static unsigned int ksm_thread_sleep_millisecs = 20;
256
257 /* Checksum of an empty (zeroed) page */
258 static unsigned int zero_checksum __read_mostly;
259
260 /* Whether to merge empty (zeroed) pages with actual zero pages */
261 static bool ksm_use_zero_pages __read_mostly;
262
263 #ifdef CONFIG_NUMA
264 /* Zeroed when merging across nodes is not allowed */
265 static unsigned int ksm_merge_across_nodes = 1;
266 static int ksm_nr_node_ids = 1;
267 #else
268 #define ksm_merge_across_nodes  1U
269 #define ksm_nr_node_ids         1
270 #endif
271
272 #define KSM_RUN_STOP    0
273 #define KSM_RUN_MERGE   1
274 #define KSM_RUN_UNMERGE 2
275 #define KSM_RUN_OFFLINE 4
276 static unsigned long ksm_run = KSM_RUN_STOP;
277 static void wait_while_offlining(void);
278
279 static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
280 static DEFINE_MUTEX(ksm_thread_mutex);
281 static DEFINE_SPINLOCK(ksm_mmlist_lock);
282
283 #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
284                 sizeof(struct __struct), __alignof__(struct __struct),\
285                 (__flags), NULL)
286
287 static int __init ksm_slab_init(void)
288 {
289         rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
290         if (!rmap_item_cache)
291                 goto out;
292
293         stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
294         if (!stable_node_cache)
295                 goto out_free1;
296
297         mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
298         if (!mm_slot_cache)
299                 goto out_free2;
300
301         return 0;
302
303 out_free2:
304         kmem_cache_destroy(stable_node_cache);
305 out_free1:
306         kmem_cache_destroy(rmap_item_cache);
307 out:
308         return -ENOMEM;
309 }
310
311 static void __init ksm_slab_free(void)
312 {
313         kmem_cache_destroy(mm_slot_cache);
314         kmem_cache_destroy(stable_node_cache);
315         kmem_cache_destroy(rmap_item_cache);
316         mm_slot_cache = NULL;
317 }
318
319 static __always_inline bool is_stable_node_chain(struct stable_node *chain)
320 {
321         return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
322 }
323
324 static __always_inline bool is_stable_node_dup(struct stable_node *dup)
325 {
326         return dup->head == STABLE_NODE_DUP_HEAD;
327 }
328
329 static inline void stable_node_chain_add_dup(struct stable_node *dup,
330                                              struct stable_node *chain)
331 {
332         VM_BUG_ON(is_stable_node_dup(dup));
333         dup->head = STABLE_NODE_DUP_HEAD;
334         VM_BUG_ON(!is_stable_node_chain(chain));
335         hlist_add_head(&dup->hlist_dup, &chain->hlist);
336         ksm_stable_node_dups++;
337 }
338
339 static inline void __stable_node_dup_del(struct stable_node *dup)
340 {
341         VM_BUG_ON(!is_stable_node_dup(dup));
342         hlist_del(&dup->hlist_dup);
343         ksm_stable_node_dups--;
344 }
345
346 static inline void stable_node_dup_del(struct stable_node *dup)
347 {
348         VM_BUG_ON(is_stable_node_chain(dup));
349         if (is_stable_node_dup(dup))
350                 __stable_node_dup_del(dup);
351         else
352                 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
353 #ifdef CONFIG_DEBUG_VM
354         dup->head = NULL;
355 #endif
356 }
357
358 static inline struct rmap_item *alloc_rmap_item(void)
359 {
360         struct rmap_item *rmap_item;
361
362         rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
363                                                 __GFP_NORETRY | __GFP_NOWARN);
364         if (rmap_item)
365                 ksm_rmap_items++;
366         return rmap_item;
367 }
368
369 static inline void free_rmap_item(struct rmap_item *rmap_item)
370 {
371         ksm_rmap_items--;
372         rmap_item->mm = NULL;   /* debug safety */
373         kmem_cache_free(rmap_item_cache, rmap_item);
374 }
375
376 static inline struct stable_node *alloc_stable_node(void)
377 {
378         /*
379          * The allocation can take too long with GFP_KERNEL when memory is under
380          * pressure, which may lead to hung task warnings.  Adding __GFP_HIGH
381          * grants access to memory reserves, helping to avoid this problem.
382          */
383         return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
384 }
385
386 static inline void free_stable_node(struct stable_node *stable_node)
387 {
388         VM_BUG_ON(stable_node->rmap_hlist_len &&
389                   !is_stable_node_chain(stable_node));
390         kmem_cache_free(stable_node_cache, stable_node);
391 }
392
393 static inline struct mm_slot *alloc_mm_slot(void)
394 {
395         if (!mm_slot_cache)     /* initialization failed */
396                 return NULL;
397         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
398 }
399
400 static inline void free_mm_slot(struct mm_slot *mm_slot)
401 {
402         kmem_cache_free(mm_slot_cache, mm_slot);
403 }
404
405 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
406 {
407         struct mm_slot *slot;
408
409         hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
410                 if (slot->mm == mm)
411                         return slot;
412
413         return NULL;
414 }
415
416 static void insert_to_mm_slots_hash(struct mm_struct *mm,
417                                     struct mm_slot *mm_slot)
418 {
419         mm_slot->mm = mm;
420         hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
421 }
422
423 /*
424  * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
425  * page tables after it has passed through ksm_exit() - which, if necessary,
426  * takes mmap_sem briefly to serialize against them.  ksm_exit() does not set
427  * a special flag: they can just back out as soon as mm_users goes to zero.
428  * ksm_test_exit() is used throughout to make this test for exit: in some
429  * places for correctness, in some places just to avoid unnecessary work.
430  */
431 static inline bool ksm_test_exit(struct mm_struct *mm)
432 {
433         return atomic_read(&mm->mm_users) == 0;
434 }
435
436 /*
437  * We use break_ksm to break COW on a ksm page: it's a stripped down
438  *
439  *      if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
440  *              put_page(page);
441  *
442  * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
443  * in case the application has unmapped and remapped mm,addr meanwhile.
444  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
445  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
446  *
447  * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
448  * of the process that owns 'vma'.  We also do not want to enforce
449  * protection keys here anyway.
450  */
451 static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
452 {
453         struct page *page;
454         int ret = 0;
455
456         do {
457                 cond_resched();
458                 page = follow_page(vma, addr,
459                                 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
460                 if (IS_ERR_OR_NULL(page))
461                         break;
462                 if (PageKsm(page))
463                         ret = handle_mm_fault(vma, addr,
464                                         FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
465                 else
466                         ret = VM_FAULT_WRITE;
467                 put_page(page);
468         } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
469         /*
470          * We must loop because handle_mm_fault() may back out if there's
471          * any difficulty e.g. if pte accessed bit gets updated concurrently.
472          *
473          * VM_FAULT_WRITE is what we have been hoping for: it indicates that
474          * COW has been broken, even if the vma does not permit VM_WRITE;
475          * but note that a concurrent fault might break PageKsm for us.
476          *
477          * VM_FAULT_SIGBUS could occur if we race with truncation of the
478          * backing file, which also invalidates anonymous pages: that's
479          * okay, that truncation will have unmapped the PageKsm for us.
480          *
481          * VM_FAULT_OOM: at the time of writing (late July 2009), setting
482          * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
483          * current task has TIF_MEMDIE set, and will be OOM killed on return
484          * to user; and ksmd, having no mm, would never be chosen for that.
485          *
486          * But if the mm is in a limited mem_cgroup, then the fault may fail
487          * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
488          * even ksmd can fail in this way - though it's usually breaking ksm
489          * just to undo a merge it made a moment before, so unlikely to oom.
490          *
491          * That's a pity: we might therefore have more kernel pages allocated
492          * than we're counting as nodes in the stable tree; but ksm_do_scan
493          * will retry to break_cow on each pass, so should recover the page
494          * in due course.  The important thing is to not let VM_MERGEABLE
495          * be cleared while any such pages might remain in the area.
496          */
497         return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
498 }
499
500 static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
501                 unsigned long addr)
502 {
503         struct vm_area_struct *vma;
504         if (ksm_test_exit(mm))
505                 return NULL;
506         vma = find_vma(mm, addr);
507         if (!vma || vma->vm_start > addr)
508                 return NULL;
509         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
510                 return NULL;
511         return vma;
512 }
513
514 static void break_cow(struct rmap_item *rmap_item)
515 {
516         struct mm_struct *mm = rmap_item->mm;
517         unsigned long addr = rmap_item->address;
518         struct vm_area_struct *vma;
519
520         /*
521          * It is not an accident that whenever we want to break COW
522          * to undo, we also need to drop a reference to the anon_vma.
523          */
524         put_anon_vma(rmap_item->anon_vma);
525
526         down_read(&mm->mmap_sem);
527         vma = find_mergeable_vma(mm, addr);
528         if (vma)
529                 break_ksm(vma, addr);
530         up_read(&mm->mmap_sem);
531 }
532
533 static struct page *get_mergeable_page(struct rmap_item *rmap_item)
534 {
535         struct mm_struct *mm = rmap_item->mm;
536         unsigned long addr = rmap_item->address;
537         struct vm_area_struct *vma;
538         struct page *page;
539
540         down_read(&mm->mmap_sem);
541         vma = find_mergeable_vma(mm, addr);
542         if (!vma)
543                 goto out;
544
545         page = follow_page(vma, addr, FOLL_GET);
546         if (IS_ERR_OR_NULL(page))
547                 goto out;
548         if (PageAnon(page)) {
549                 flush_anon_page(vma, page, addr);
550                 flush_dcache_page(page);
551         } else {
552                 put_page(page);
553 out:
554                 page = NULL;
555         }
556         up_read(&mm->mmap_sem);
557         return page;
558 }
559
560 /*
561  * This helper is used for getting right index into array of tree roots.
562  * When merge_across_nodes knob is set to 1, there are only two rb-trees for
563  * stable and unstable pages from all nodes with roots in index 0. Otherwise,
564  * every node has its own stable and unstable tree.
565  */
566 static inline int get_kpfn_nid(unsigned long kpfn)
567 {
568         return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
569 }
570
571 static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
572                                                    struct rb_root *root)
573 {
574         struct stable_node *chain = alloc_stable_node();
575         VM_BUG_ON(is_stable_node_chain(dup));
576         if (likely(chain)) {
577                 INIT_HLIST_HEAD(&chain->hlist);
578                 chain->chain_prune_time = jiffies;
579                 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
580 #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
581                 chain->nid = -1; /* debug */
582 #endif
583                 ksm_stable_node_chains++;
584
585                 /*
586                  * Put the stable node chain in the first dimension of
587                  * the stable tree and at the same time remove the old
588                  * stable node.
589                  */
590                 rb_replace_node(&dup->node, &chain->node, root);
591
592                 /*
593                  * Move the old stable node to the second dimension
594                  * queued in the hlist_dup. The invariant is that all
595                  * dup stable_nodes in the chain->hlist point to pages
596                  * that are wrprotected and have the exact same
597                  * content.
598                  */
599                 stable_node_chain_add_dup(dup, chain);
600         }
601         return chain;
602 }
603
604 static inline void free_stable_node_chain(struct stable_node *chain,
605                                           struct rb_root *root)
606 {
607         rb_erase(&chain->node, root);
608         free_stable_node(chain);
609         ksm_stable_node_chains--;
610 }
611
612 static void remove_node_from_stable_tree(struct stable_node *stable_node)
613 {
614         struct rmap_item *rmap_item;
615
616         /* check it's not STABLE_NODE_CHAIN or negative */
617         BUG_ON(stable_node->rmap_hlist_len < 0);
618
619         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
620                 if (rmap_item->hlist.next)
621                         ksm_pages_sharing--;
622                 else
623                         ksm_pages_shared--;
624                 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
625                 stable_node->rmap_hlist_len--;
626                 put_anon_vma(rmap_item->anon_vma);
627                 rmap_item->address &= PAGE_MASK;
628                 cond_resched();
629         }
630
631         /*
632          * We need the second aligned pointer of the migrate_nodes
633          * list_head to stay clear from the rb_parent_color union
634          * (aligned and different than any node) and also different
635          * from &migrate_nodes. This will verify that future list.h changes
636          * don't break STABLE_NODE_DUP_HEAD.
637          */
638 #if GCC_VERSION >= 40903 /* only recent gcc can handle it */
639         BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
640         BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
641 #endif
642
643         if (stable_node->head == &migrate_nodes)
644                 list_del(&stable_node->list);
645         else
646                 stable_node_dup_del(stable_node);
647         free_stable_node(stable_node);
648 }
649
650 /*
651  * get_ksm_page: checks if the page indicated by the stable node
652  * is still its ksm page, despite having held no reference to it.
653  * In which case we can trust the content of the page, and it
654  * returns the gotten page; but if the page has now been zapped,
655  * remove the stale node from the stable tree and return NULL.
656  * But beware, the stable node's page might be being migrated.
657  *
658  * You would expect the stable_node to hold a reference to the ksm page.
659  * But if it increments the page's count, swapping out has to wait for
660  * ksmd to come around again before it can free the page, which may take
661  * seconds or even minutes: much too unresponsive.  So instead we use a
662  * "keyhole reference": access to the ksm page from the stable node peeps
663  * out through its keyhole to see if that page still holds the right key,
664  * pointing back to this stable node.  This relies on freeing a PageAnon
665  * page to reset its page->mapping to NULL, and relies on no other use of
666  * a page to put something that might look like our key in page->mapping.
667  * is on its way to being freed; but it is an anomaly to bear in mind.
668  */
669 static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
670 {
671         struct page *page;
672         void *expected_mapping;
673         unsigned long kpfn;
674
675         expected_mapping = (void *)((unsigned long)stable_node |
676                                         PAGE_MAPPING_KSM);
677 again:
678         kpfn = READ_ONCE(stable_node->kpfn);
679         page = pfn_to_page(kpfn);
680
681         /*
682          * page is computed from kpfn, so on most architectures reading
683          * page->mapping is naturally ordered after reading node->kpfn,
684          * but on Alpha we need to be more careful.
685          */
686         smp_read_barrier_depends();
687         if (READ_ONCE(page->mapping) != expected_mapping)
688                 goto stale;
689
690         /*
691          * We cannot do anything with the page while its refcount is 0.
692          * Usually 0 means free, or tail of a higher-order page: in which
693          * case this node is no longer referenced, and should be freed;
694          * however, it might mean that the page is under page_freeze_refs().
695          * The __remove_mapping() case is easy, again the node is now stale;
696          * but if page is swapcache in migrate_page_move_mapping(), it might
697          * still be our page, in which case it's essential to keep the node.
698          */
699         while (!get_page_unless_zero(page)) {
700                 /*
701                  * Another check for page->mapping != expected_mapping would
702                  * work here too.  We have chosen the !PageSwapCache test to
703                  * optimize the common case, when the page is or is about to
704                  * be freed: PageSwapCache is cleared (under spin_lock_irq)
705                  * in the freeze_refs section of __remove_mapping(); but Anon
706                  * page->mapping reset to NULL later, in free_pages_prepare().
707                  */
708                 if (!PageSwapCache(page))
709                         goto stale;
710                 cpu_relax();
711         }
712
713         if (READ_ONCE(page->mapping) != expected_mapping) {
714                 put_page(page);
715                 goto stale;
716         }
717
718         if (lock_it) {
719                 lock_page(page);
720                 if (READ_ONCE(page->mapping) != expected_mapping) {
721                         unlock_page(page);
722                         put_page(page);
723                         goto stale;
724                 }
725         }
726         return page;
727
728 stale:
729         /*
730          * We come here from above when page->mapping or !PageSwapCache
731          * suggests that the node is stale; but it might be under migration.
732          * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
733          * before checking whether node->kpfn has been changed.
734          */
735         smp_rmb();
736         if (READ_ONCE(stable_node->kpfn) != kpfn)
737                 goto again;
738         remove_node_from_stable_tree(stable_node);
739         return NULL;
740 }
741
742 /*
743  * Removing rmap_item from stable or unstable tree.
744  * This function will clean the information from the stable/unstable tree.
745  */
746 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
747 {
748         if (rmap_item->address & STABLE_FLAG) {
749                 struct stable_node *stable_node;
750                 struct page *page;
751
752                 stable_node = rmap_item->head;
753                 page = get_ksm_page(stable_node, true);
754                 if (!page)
755                         goto out;
756
757                 hlist_del(&rmap_item->hlist);
758                 unlock_page(page);
759                 put_page(page);
760
761                 if (!hlist_empty(&stable_node->hlist))
762                         ksm_pages_sharing--;
763                 else
764                         ksm_pages_shared--;
765                 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
766                 stable_node->rmap_hlist_len--;
767
768                 put_anon_vma(rmap_item->anon_vma);
769                 rmap_item->address &= PAGE_MASK;
770
771         } else if (rmap_item->address & UNSTABLE_FLAG) {
772                 unsigned char age;
773                 /*
774                  * Usually ksmd can and must skip the rb_erase, because
775                  * root_unstable_tree was already reset to RB_ROOT.
776                  * But be careful when an mm is exiting: do the rb_erase
777                  * if this rmap_item was inserted by this scan, rather
778                  * than left over from before.
779                  */
780                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
781                 BUG_ON(age > 1);
782                 if (!age)
783                         rb_erase(&rmap_item->node,
784                                  root_unstable_tree + NUMA(rmap_item->nid));
785                 ksm_pages_unshared--;
786                 rmap_item->address &= PAGE_MASK;
787         }
788 out:
789         cond_resched();         /* we're called from many long loops */
790 }
791
792 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
793                                        struct rmap_item **rmap_list)
794 {
795         while (*rmap_list) {
796                 struct rmap_item *rmap_item = *rmap_list;
797                 *rmap_list = rmap_item->rmap_list;
798                 remove_rmap_item_from_tree(rmap_item);
799                 free_rmap_item(rmap_item);
800         }
801 }
802
803 /*
804  * Though it's very tempting to unmerge rmap_items from stable tree rather
805  * than check every pte of a given vma, the locking doesn't quite work for
806  * that - an rmap_item is assigned to the stable tree after inserting ksm
807  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
808  * rmap_items from parent to child at fork time (so as not to waste time
809  * if exit comes before the next scan reaches it).
810  *
811  * Similarly, although we'd like to remove rmap_items (so updating counts
812  * and freeing memory) when unmerging an area, it's easier to leave that
813  * to the next pass of ksmd - consider, for example, how ksmd might be
814  * in cmp_and_merge_page on one of the rmap_items we would be removing.
815  */
816 static int unmerge_ksm_pages(struct vm_area_struct *vma,
817                              unsigned long start, unsigned long end)
818 {
819         unsigned long addr;
820         int err = 0;
821
822         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
823                 if (ksm_test_exit(vma->vm_mm))
824                         break;
825                 if (signal_pending(current))
826                         err = -ERESTARTSYS;
827                 else
828                         err = break_ksm(vma, addr);
829         }
830         return err;
831 }
832
833 #ifdef CONFIG_SYSFS
834 /*
835  * Only called through the sysfs control interface:
836  */
837 static int remove_stable_node(struct stable_node *stable_node)
838 {
839         struct page *page;
840         int err;
841
842         page = get_ksm_page(stable_node, true);
843         if (!page) {
844                 /*
845                  * get_ksm_page did remove_node_from_stable_tree itself.
846                  */
847                 return 0;
848         }
849
850         if (WARN_ON_ONCE(page_mapped(page))) {
851                 /*
852                  * This should not happen: but if it does, just refuse to let
853                  * merge_across_nodes be switched - there is no need to panic.
854                  */
855                 err = -EBUSY;
856         } else {
857                 /*
858                  * The stable node did not yet appear stale to get_ksm_page(),
859                  * since that allows for an unmapped ksm page to be recognized
860                  * right up until it is freed; but the node is safe to remove.
861                  * This page might be in a pagevec waiting to be freed,
862                  * or it might be PageSwapCache (perhaps under writeback),
863                  * or it might have been removed from swapcache a moment ago.
864                  */
865                 set_page_stable_node(page, NULL);
866                 remove_node_from_stable_tree(stable_node);
867                 err = 0;
868         }
869
870         unlock_page(page);
871         put_page(page);
872         return err;
873 }
874
875 static int remove_stable_node_chain(struct stable_node *stable_node,
876                                     struct rb_root *root)
877 {
878         struct stable_node *dup;
879         struct hlist_node *hlist_safe;
880
881         if (!is_stable_node_chain(stable_node)) {
882                 VM_BUG_ON(is_stable_node_dup(stable_node));
883                 if (remove_stable_node(stable_node))
884                         return true;
885                 else
886                         return false;
887         }
888
889         hlist_for_each_entry_safe(dup, hlist_safe,
890                                   &stable_node->hlist, hlist_dup) {
891                 VM_BUG_ON(!is_stable_node_dup(dup));
892                 if (remove_stable_node(dup))
893                         return true;
894         }
895         BUG_ON(!hlist_empty(&stable_node->hlist));
896         free_stable_node_chain(stable_node, root);
897         return false;
898 }
899
900 static int remove_all_stable_nodes(void)
901 {
902         struct stable_node *stable_node, *next;
903         int nid;
904         int err = 0;
905
906         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
907                 while (root_stable_tree[nid].rb_node) {
908                         stable_node = rb_entry(root_stable_tree[nid].rb_node,
909                                                 struct stable_node, node);
910                         if (remove_stable_node_chain(stable_node,
911                                                      root_stable_tree + nid)) {
912                                 err = -EBUSY;
913                                 break;  /* proceed to next nid */
914                         }
915                         cond_resched();
916                 }
917         }
918         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
919                 if (remove_stable_node(stable_node))
920                         err = -EBUSY;
921                 cond_resched();
922         }
923         return err;
924 }
925
926 static int unmerge_and_remove_all_rmap_items(void)
927 {
928         struct mm_slot *mm_slot;
929         struct mm_struct *mm;
930         struct vm_area_struct *vma;
931         int err = 0;
932
933         spin_lock(&ksm_mmlist_lock);
934         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
935                                                 struct mm_slot, mm_list);
936         spin_unlock(&ksm_mmlist_lock);
937
938         for (mm_slot = ksm_scan.mm_slot;
939                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
940                 mm = mm_slot->mm;
941                 down_read(&mm->mmap_sem);
942                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
943                         if (ksm_test_exit(mm))
944                                 break;
945                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
946                                 continue;
947                         err = unmerge_ksm_pages(vma,
948                                                 vma->vm_start, vma->vm_end);
949                         if (err)
950                                 goto error;
951                 }
952
953                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
954                 up_read(&mm->mmap_sem);
955
956                 spin_lock(&ksm_mmlist_lock);
957                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
958                                                 struct mm_slot, mm_list);
959                 if (ksm_test_exit(mm)) {
960                         hash_del(&mm_slot->link);
961                         list_del(&mm_slot->mm_list);
962                         spin_unlock(&ksm_mmlist_lock);
963
964                         free_mm_slot(mm_slot);
965                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
966                         mmdrop(mm);
967                 } else
968                         spin_unlock(&ksm_mmlist_lock);
969         }
970
971         /* Clean up stable nodes, but don't worry if some are still busy */
972         remove_all_stable_nodes();
973         ksm_scan.seqnr = 0;
974         return 0;
975
976 error:
977         up_read(&mm->mmap_sem);
978         spin_lock(&ksm_mmlist_lock);
979         ksm_scan.mm_slot = &ksm_mm_head;
980         spin_unlock(&ksm_mmlist_lock);
981         return err;
982 }
983 #endif /* CONFIG_SYSFS */
984
985 static u32 calc_checksum(struct page *page)
986 {
987         u32 checksum;
988         void *addr = kmap_atomic(page);
989         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
990         kunmap_atomic(addr);
991         return checksum;
992 }
993
994 static int memcmp_pages(struct page *page1, struct page *page2)
995 {
996         char *addr1, *addr2;
997         int ret;
998
999         addr1 = kmap_atomic(page1);
1000         addr2 = kmap_atomic(page2);
1001         ret = memcmp(addr1, addr2, PAGE_SIZE);
1002         kunmap_atomic(addr2);
1003         kunmap_atomic(addr1);
1004         return ret;
1005 }
1006
1007 static inline int pages_identical(struct page *page1, struct page *page2)
1008 {
1009         return !memcmp_pages(page1, page2);
1010 }
1011
1012 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1013                               pte_t *orig_pte)
1014 {
1015         struct mm_struct *mm = vma->vm_mm;
1016         struct page_vma_mapped_walk pvmw = {
1017                 .page = page,
1018                 .vma = vma,
1019         };
1020         int swapped;
1021         int err = -EFAULT;
1022         unsigned long mmun_start;       /* For mmu_notifiers */
1023         unsigned long mmun_end;         /* For mmu_notifiers */
1024
1025         pvmw.address = page_address_in_vma(page, vma);
1026         if (pvmw.address == -EFAULT)
1027                 goto out;
1028
1029         BUG_ON(PageTransCompound(page));
1030
1031         mmun_start = pvmw.address;
1032         mmun_end   = pvmw.address + PAGE_SIZE;
1033         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1034
1035         if (!page_vma_mapped_walk(&pvmw))
1036                 goto out_mn;
1037         if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1038                 goto out_unlock;
1039
1040         if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1041             (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte))) {
1042                 pte_t entry;
1043
1044                 swapped = PageSwapCache(page);
1045                 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1046                 /*
1047                  * Ok this is tricky, when get_user_pages_fast() run it doesn't
1048                  * take any lock, therefore the check that we are going to make
1049                  * with the pagecount against the mapcount is racey and
1050                  * O_DIRECT can happen right after the check.
1051                  * So we clear the pte and flush the tlb before the check
1052                  * this assure us that no O_DIRECT can happen after the check
1053                  * or in the middle of the check.
1054                  */
1055                 entry = ptep_clear_flush_notify(vma, pvmw.address, pvmw.pte);
1056                 /*
1057                  * Check that no O_DIRECT or similar I/O is in progress on the
1058                  * page
1059                  */
1060                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1061                         set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1062                         goto out_unlock;
1063                 }
1064                 if (pte_dirty(entry))
1065                         set_page_dirty(page);
1066
1067                 if (pte_protnone(entry))
1068                         entry = pte_mkclean(pte_clear_savedwrite(entry));
1069                 else
1070                         entry = pte_mkclean(pte_wrprotect(entry));
1071                 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1072         }
1073         *orig_pte = *pvmw.pte;
1074         err = 0;
1075
1076 out_unlock:
1077         page_vma_mapped_walk_done(&pvmw);
1078 out_mn:
1079         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1080 out:
1081         return err;
1082 }
1083
1084 /**
1085  * replace_page - replace page in vma by new ksm page
1086  * @vma:      vma that holds the pte pointing to page
1087  * @page:     the page we are replacing by kpage
1088  * @kpage:    the ksm page we replace page by
1089  * @orig_pte: the original value of the pte
1090  *
1091  * Returns 0 on success, -EFAULT on failure.
1092  */
1093 static int replace_page(struct vm_area_struct *vma, struct page *page,
1094                         struct page *kpage, pte_t orig_pte)
1095 {
1096         struct mm_struct *mm = vma->vm_mm;
1097         pmd_t *pmd;
1098         pte_t *ptep;
1099         pte_t newpte;
1100         spinlock_t *ptl;
1101         unsigned long addr;
1102         int err = -EFAULT;
1103         unsigned long mmun_start;       /* For mmu_notifiers */
1104         unsigned long mmun_end;         /* For mmu_notifiers */
1105
1106         addr = page_address_in_vma(page, vma);
1107         if (addr == -EFAULT)
1108                 goto out;
1109
1110         pmd = mm_find_pmd(mm, addr);
1111         if (!pmd)
1112                 goto out;
1113
1114         mmun_start = addr;
1115         mmun_end   = addr + PAGE_SIZE;
1116         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1117
1118         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1119         if (!pte_same(*ptep, orig_pte)) {
1120                 pte_unmap_unlock(ptep, ptl);
1121                 goto out_mn;
1122         }
1123
1124         /*
1125          * No need to check ksm_use_zero_pages here: we can only have a
1126          * zero_page here if ksm_use_zero_pages was enabled alreaady.
1127          */
1128         if (!is_zero_pfn(page_to_pfn(kpage))) {
1129                 get_page(kpage);
1130                 page_add_anon_rmap(kpage, vma, addr, false);
1131                 newpte = mk_pte(kpage, vma->vm_page_prot);
1132         } else {
1133                 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1134                                                vma->vm_page_prot));
1135         }
1136
1137         flush_cache_page(vma, addr, pte_pfn(*ptep));
1138         ptep_clear_flush_notify(vma, addr, ptep);
1139         set_pte_at_notify(mm, addr, ptep, newpte);
1140
1141         page_remove_rmap(page, false);
1142         if (!page_mapped(page))
1143                 try_to_free_swap(page);
1144         put_page(page);
1145
1146         pte_unmap_unlock(ptep, ptl);
1147         err = 0;
1148 out_mn:
1149         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1150 out:
1151         return err;
1152 }
1153
1154 /*
1155  * try_to_merge_one_page - take two pages and merge them into one
1156  * @vma: the vma that holds the pte pointing to page
1157  * @page: the PageAnon page that we want to replace with kpage
1158  * @kpage: the PageKsm page that we want to map instead of page,
1159  *         or NULL the first time when we want to use page as kpage.
1160  *
1161  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1162  */
1163 static int try_to_merge_one_page(struct vm_area_struct *vma,
1164                                  struct page *page, struct page *kpage)
1165 {
1166         pte_t orig_pte = __pte(0);
1167         int err = -EFAULT;
1168
1169         if (page == kpage)                      /* ksm page forked */
1170                 return 0;
1171
1172         if (!PageAnon(page))
1173                 goto out;
1174
1175         /*
1176          * We need the page lock to read a stable PageSwapCache in
1177          * write_protect_page().  We use trylock_page() instead of
1178          * lock_page() because we don't want to wait here - we
1179          * prefer to continue scanning and merging different pages,
1180          * then come back to this page when it is unlocked.
1181          */
1182         if (!trylock_page(page))
1183                 goto out;
1184
1185         if (PageTransCompound(page)) {
1186                 if (split_huge_page(page))
1187                         goto out_unlock;
1188         }
1189
1190         /*
1191          * If this anonymous page is mapped only here, its pte may need
1192          * to be write-protected.  If it's mapped elsewhere, all of its
1193          * ptes are necessarily already write-protected.  But in either
1194          * case, we need to lock and check page_count is not raised.
1195          */
1196         if (write_protect_page(vma, page, &orig_pte) == 0) {
1197                 if (!kpage) {
1198                         /*
1199                          * While we hold page lock, upgrade page from
1200                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
1201                          * stable_tree_insert() will update stable_node.
1202                          */
1203                         set_page_stable_node(page, NULL);
1204                         mark_page_accessed(page);
1205                         /*
1206                          * Page reclaim just frees a clean page with no dirty
1207                          * ptes: make sure that the ksm page would be swapped.
1208                          */
1209                         if (!PageDirty(page))
1210                                 SetPageDirty(page);
1211                         err = 0;
1212                 } else if (pages_identical(page, kpage))
1213                         err = replace_page(vma, page, kpage, orig_pte);
1214         }
1215
1216         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1217                 munlock_vma_page(page);
1218                 if (!PageMlocked(kpage)) {
1219                         unlock_page(page);
1220                         lock_page(kpage);
1221                         mlock_vma_page(kpage);
1222                         page = kpage;           /* for final unlock */
1223                 }
1224         }
1225
1226 out_unlock:
1227         unlock_page(page);
1228 out:
1229         return err;
1230 }
1231
1232 /*
1233  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1234  * but no new kernel page is allocated: kpage must already be a ksm page.
1235  *
1236  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1237  */
1238 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1239                                       struct page *page, struct page *kpage)
1240 {
1241         struct mm_struct *mm = rmap_item->mm;
1242         struct vm_area_struct *vma;
1243         int err = -EFAULT;
1244
1245         down_read(&mm->mmap_sem);
1246         vma = find_mergeable_vma(mm, rmap_item->address);
1247         if (!vma)
1248                 goto out;
1249
1250         err = try_to_merge_one_page(vma, page, kpage);
1251         if (err)
1252                 goto out;
1253
1254         /* Unstable nid is in union with stable anon_vma: remove first */
1255         remove_rmap_item_from_tree(rmap_item);
1256
1257         /* Must get reference to anon_vma while still holding mmap_sem */
1258         rmap_item->anon_vma = vma->anon_vma;
1259         get_anon_vma(vma->anon_vma);
1260 out:
1261         up_read(&mm->mmap_sem);
1262         return err;
1263 }
1264
1265 /*
1266  * try_to_merge_two_pages - take two identical pages and prepare them
1267  * to be merged into one page.
1268  *
1269  * This function returns the kpage if we successfully merged two identical
1270  * pages into one ksm page, NULL otherwise.
1271  *
1272  * Note that this function upgrades page to ksm page: if one of the pages
1273  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1274  */
1275 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1276                                            struct page *page,
1277                                            struct rmap_item *tree_rmap_item,
1278                                            struct page *tree_page)
1279 {
1280         int err;
1281
1282         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1283         if (!err) {
1284                 err = try_to_merge_with_ksm_page(tree_rmap_item,
1285                                                         tree_page, page);
1286                 /*
1287                  * If that fails, we have a ksm page with only one pte
1288                  * pointing to it: so break it.
1289                  */
1290                 if (err)
1291                         break_cow(rmap_item);
1292         }
1293         return err ? NULL : page;
1294 }
1295
1296 static __always_inline
1297 bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1298 {
1299         VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1300         /*
1301          * Check that at least one mapping still exists, otherwise
1302          * there's no much point to merge and share with this
1303          * stable_node, as the underlying tree_page of the other
1304          * sharer is going to be freed soon.
1305          */
1306         return stable_node->rmap_hlist_len &&
1307                 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1308 }
1309
1310 static __always_inline
1311 bool is_page_sharing_candidate(struct stable_node *stable_node)
1312 {
1313         return __is_page_sharing_candidate(stable_node, 0);
1314 }
1315
1316 struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1317                              struct stable_node **_stable_node,
1318                              struct rb_root *root,
1319                              bool prune_stale_stable_nodes)
1320 {
1321         struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1322         struct hlist_node *hlist_safe;
1323         struct page *_tree_page, *tree_page = NULL;
1324         int nr = 0;
1325         int found_rmap_hlist_len;
1326
1327         if (!prune_stale_stable_nodes ||
1328             time_before(jiffies, stable_node->chain_prune_time +
1329                         msecs_to_jiffies(
1330                                 ksm_stable_node_chains_prune_millisecs)))
1331                 prune_stale_stable_nodes = false;
1332         else
1333                 stable_node->chain_prune_time = jiffies;
1334
1335         hlist_for_each_entry_safe(dup, hlist_safe,
1336                                   &stable_node->hlist, hlist_dup) {
1337                 cond_resched();
1338                 /*
1339                  * We must walk all stable_node_dup to prune the stale
1340                  * stable nodes during lookup.
1341                  *
1342                  * get_ksm_page can drop the nodes from the
1343                  * stable_node->hlist if they point to freed pages
1344                  * (that's why we do a _safe walk). The "dup"
1345                  * stable_node parameter itself will be freed from
1346                  * under us if it returns NULL.
1347                  */
1348                 _tree_page = get_ksm_page(dup, false);
1349                 if (!_tree_page)
1350                         continue;
1351                 nr += 1;
1352                 if (is_page_sharing_candidate(dup)) {
1353                         if (!found ||
1354                             dup->rmap_hlist_len > found_rmap_hlist_len) {
1355                                 if (found)
1356                                         put_page(tree_page);
1357                                 found = dup;
1358                                 found_rmap_hlist_len = found->rmap_hlist_len;
1359                                 tree_page = _tree_page;
1360
1361                                 /* skip put_page for found dup */
1362                                 if (!prune_stale_stable_nodes)
1363                                         break;
1364                                 continue;
1365                         }
1366                 }
1367                 put_page(_tree_page);
1368         }
1369
1370         if (found) {
1371                 /*
1372                  * nr is counting all dups in the chain only if
1373                  * prune_stale_stable_nodes is true, otherwise we may
1374                  * break the loop at nr == 1 even if there are
1375                  * multiple entries.
1376                  */
1377                 if (prune_stale_stable_nodes && nr == 1) {
1378                         /*
1379                          * If there's not just one entry it would
1380                          * corrupt memory, better BUG_ON. In KSM
1381                          * context with no lock held it's not even
1382                          * fatal.
1383                          */
1384                         BUG_ON(stable_node->hlist.first->next);
1385
1386                         /*
1387                          * There's just one entry and it is below the
1388                          * deduplication limit so drop the chain.
1389                          */
1390                         rb_replace_node(&stable_node->node, &found->node,
1391                                         root);
1392                         free_stable_node(stable_node);
1393                         ksm_stable_node_chains--;
1394                         ksm_stable_node_dups--;
1395                         /*
1396                          * NOTE: the caller depends on the stable_node
1397                          * to be equal to stable_node_dup if the chain
1398                          * was collapsed.
1399                          */
1400                         *_stable_node = found;
1401                         /*
1402                          * Just for robustneess as stable_node is
1403                          * otherwise left as a stable pointer, the
1404                          * compiler shall optimize it away at build
1405                          * time.
1406                          */
1407                         stable_node = NULL;
1408                 } else if (stable_node->hlist.first != &found->hlist_dup &&
1409                            __is_page_sharing_candidate(found, 1)) {
1410                         /*
1411                          * If the found stable_node dup can accept one
1412                          * more future merge (in addition to the one
1413                          * that is underway) and is not at the head of
1414                          * the chain, put it there so next search will
1415                          * be quicker in the !prune_stale_stable_nodes
1416                          * case.
1417                          *
1418                          * NOTE: it would be inaccurate to use nr > 1
1419                          * instead of checking the hlist.first pointer
1420                          * directly, because in the
1421                          * prune_stale_stable_nodes case "nr" isn't
1422                          * the position of the found dup in the chain,
1423                          * but the total number of dups in the chain.
1424                          */
1425                         hlist_del(&found->hlist_dup);
1426                         hlist_add_head(&found->hlist_dup,
1427                                        &stable_node->hlist);
1428                 }
1429         }
1430
1431         *_stable_node_dup = found;
1432         return tree_page;
1433 }
1434
1435 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1436                                                struct rb_root *root)
1437 {
1438         if (!is_stable_node_chain(stable_node))
1439                 return stable_node;
1440         if (hlist_empty(&stable_node->hlist)) {
1441                 free_stable_node_chain(stable_node, root);
1442                 return NULL;
1443         }
1444         return hlist_entry(stable_node->hlist.first,
1445                            typeof(*stable_node), hlist_dup);
1446 }
1447
1448 /*
1449  * Like for get_ksm_page, this function can free the *_stable_node and
1450  * *_stable_node_dup if the returned tree_page is NULL.
1451  *
1452  * It can also free and overwrite *_stable_node with the found
1453  * stable_node_dup if the chain is collapsed (in which case
1454  * *_stable_node will be equal to *_stable_node_dup like if the chain
1455  * never existed). It's up to the caller to verify tree_page is not
1456  * NULL before dereferencing *_stable_node or *_stable_node_dup.
1457  *
1458  * *_stable_node_dup is really a second output parameter of this
1459  * function and will be overwritten in all cases, the caller doesn't
1460  * need to initialize it.
1461  */
1462 static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1463                                         struct stable_node **_stable_node,
1464                                         struct rb_root *root,
1465                                         bool prune_stale_stable_nodes)
1466 {
1467         struct stable_node *stable_node = *_stable_node;
1468         if (!is_stable_node_chain(stable_node)) {
1469                 if (is_page_sharing_candidate(stable_node)) {
1470                         *_stable_node_dup = stable_node;
1471                         return get_ksm_page(stable_node, false);
1472                 }
1473                 /*
1474                  * _stable_node_dup set to NULL means the stable_node
1475                  * reached the ksm_max_page_sharing limit.
1476                  */
1477                 *_stable_node_dup = NULL;
1478                 return NULL;
1479         }
1480         return stable_node_dup(_stable_node_dup, _stable_node, root,
1481                                prune_stale_stable_nodes);
1482 }
1483
1484 static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1485                                                 struct stable_node **s_n,
1486                                                 struct rb_root *root)
1487 {
1488         return __stable_node_chain(s_n_d, s_n, root, true);
1489 }
1490
1491 static __always_inline struct page *chain(struct stable_node **s_n_d,
1492                                           struct stable_node *s_n,
1493                                           struct rb_root *root)
1494 {
1495         struct stable_node *old_stable_node = s_n;
1496         struct page *tree_page;
1497
1498         tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1499         /* not pruning dups so s_n cannot have changed */
1500         VM_BUG_ON(s_n != old_stable_node);
1501         return tree_page;
1502 }
1503
1504 /*
1505  * stable_tree_search - search for page inside the stable tree
1506  *
1507  * This function checks if there is a page inside the stable tree
1508  * with identical content to the page that we are scanning right now.
1509  *
1510  * This function returns the stable tree node of identical content if found,
1511  * NULL otherwise.
1512  */
1513 static struct page *stable_tree_search(struct page *page)
1514 {
1515         int nid;
1516         struct rb_root *root;
1517         struct rb_node **new;
1518         struct rb_node *parent;
1519         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1520         struct stable_node *page_node;
1521
1522         page_node = page_stable_node(page);
1523         if (page_node && page_node->head != &migrate_nodes) {
1524                 /* ksm page forked */
1525                 get_page(page);
1526                 return page;
1527         }
1528
1529         nid = get_kpfn_nid(page_to_pfn(page));
1530         root = root_stable_tree + nid;
1531 again:
1532         new = &root->rb_node;
1533         parent = NULL;
1534
1535         while (*new) {
1536                 struct page *tree_page;
1537                 int ret;
1538
1539                 cond_resched();
1540                 stable_node = rb_entry(*new, struct stable_node, node);
1541                 stable_node_any = NULL;
1542                 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1543                 /*
1544                  * NOTE: stable_node may have been freed by
1545                  * chain_prune() if the returned stable_node_dup is
1546                  * not NULL. stable_node_dup may have been inserted in
1547                  * the rbtree instead as a regular stable_node (in
1548                  * order to collapse the stable_node chain if a single
1549                  * stable_node dup was found in it). In such case the
1550                  * stable_node is overwritten by the calleee to point
1551                  * to the stable_node_dup that was collapsed in the
1552                  * stable rbtree and stable_node will be equal to
1553                  * stable_node_dup like if the chain never existed.
1554                  */
1555                 if (!stable_node_dup) {
1556                         /*
1557                          * Either all stable_node dups were full in
1558                          * this stable_node chain, or this chain was
1559                          * empty and should be rb_erased.
1560                          */
1561                         stable_node_any = stable_node_dup_any(stable_node,
1562                                                               root);
1563                         if (!stable_node_any) {
1564                                 /* rb_erase just run */
1565                                 goto again;
1566                         }
1567                         /*
1568                          * Take any of the stable_node dups page of
1569                          * this stable_node chain to let the tree walk
1570                          * continue. All KSM pages belonging to the
1571                          * stable_node dups in a stable_node chain
1572                          * have the same content and they're
1573                          * wrprotected at all times. Any will work
1574                          * fine to continue the walk.
1575                          */
1576                         tree_page = get_ksm_page(stable_node_any, false);
1577                 }
1578                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1579                 if (!tree_page) {
1580                         /*
1581                          * If we walked over a stale stable_node,
1582                          * get_ksm_page() will call rb_erase() and it
1583                          * may rebalance the tree from under us. So
1584                          * restart the search from scratch. Returning
1585                          * NULL would be safe too, but we'd generate
1586                          * false negative insertions just because some
1587                          * stable_node was stale.
1588                          */
1589                         goto again;
1590                 }
1591
1592                 ret = memcmp_pages(page, tree_page);
1593                 put_page(tree_page);
1594
1595                 parent = *new;
1596                 if (ret < 0)
1597                         new = &parent->rb_left;
1598                 else if (ret > 0)
1599                         new = &parent->rb_right;
1600                 else {
1601                         if (page_node) {
1602                                 VM_BUG_ON(page_node->head != &migrate_nodes);
1603                                 /*
1604                                  * Test if the migrated page should be merged
1605                                  * into a stable node dup. If the mapcount is
1606                                  * 1 we can migrate it with another KSM page
1607                                  * without adding it to the chain.
1608                                  */
1609                                 if (page_mapcount(page) > 1)
1610                                         goto chain_append;
1611                         }
1612
1613                         if (!stable_node_dup) {
1614                                 /*
1615                                  * If the stable_node is a chain and
1616                                  * we got a payload match in memcmp
1617                                  * but we cannot merge the scanned
1618                                  * page in any of the existing
1619                                  * stable_node dups because they're
1620                                  * all full, we need to wait the
1621                                  * scanned page to find itself a match
1622                                  * in the unstable tree to create a
1623                                  * brand new KSM page to add later to
1624                                  * the dups of this stable_node.
1625                                  */
1626                                 return NULL;
1627                         }
1628
1629                         /*
1630                          * Lock and unlock the stable_node's page (which
1631                          * might already have been migrated) so that page
1632                          * migration is sure to notice its raised count.
1633                          * It would be more elegant to return stable_node
1634                          * than kpage, but that involves more changes.
1635                          */
1636                         tree_page = get_ksm_page(stable_node_dup, true);
1637                         if (unlikely(!tree_page))
1638                                 /*
1639                                  * The tree may have been rebalanced,
1640                                  * so re-evaluate parent and new.
1641                                  */
1642                                 goto again;
1643                         unlock_page(tree_page);
1644
1645                         if (get_kpfn_nid(stable_node_dup->kpfn) !=
1646                             NUMA(stable_node_dup->nid)) {
1647                                 put_page(tree_page);
1648                                 goto replace;
1649                         }
1650                         return tree_page;
1651                 }
1652         }
1653
1654         if (!page_node)
1655                 return NULL;
1656
1657         list_del(&page_node->list);
1658         DO_NUMA(page_node->nid = nid);
1659         rb_link_node(&page_node->node, parent, new);
1660         rb_insert_color(&page_node->node, root);
1661 out:
1662         if (is_page_sharing_candidate(page_node)) {
1663                 get_page(page);
1664                 return page;
1665         } else
1666                 return NULL;
1667
1668 replace:
1669         /*
1670          * If stable_node was a chain and chain_prune collapsed it,
1671          * stable_node has been updated to be the new regular
1672          * stable_node. A collapse of the chain is indistinguishable
1673          * from the case there was no chain in the stable
1674          * rbtree. Otherwise stable_node is the chain and
1675          * stable_node_dup is the dup to replace.
1676          */
1677         if (stable_node_dup == stable_node) {
1678                 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1679                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1680                 /* there is no chain */
1681                 if (page_node) {
1682                         VM_BUG_ON(page_node->head != &migrate_nodes);
1683                         list_del(&page_node->list);
1684                         DO_NUMA(page_node->nid = nid);
1685                         rb_replace_node(&stable_node_dup->node,
1686                                         &page_node->node,
1687                                         root);
1688                         if (is_page_sharing_candidate(page_node))
1689                                 get_page(page);
1690                         else
1691                                 page = NULL;
1692                 } else {
1693                         rb_erase(&stable_node_dup->node, root);
1694                         page = NULL;
1695                 }
1696         } else {
1697                 VM_BUG_ON(!is_stable_node_chain(stable_node));
1698                 __stable_node_dup_del(stable_node_dup);
1699                 if (page_node) {
1700                         VM_BUG_ON(page_node->head != &migrate_nodes);
1701                         list_del(&page_node->list);
1702                         DO_NUMA(page_node->nid = nid);
1703                         stable_node_chain_add_dup(page_node, stable_node);
1704                         if (is_page_sharing_candidate(page_node))
1705                                 get_page(page);
1706                         else
1707                                 page = NULL;
1708                 } else {
1709                         page = NULL;
1710                 }
1711         }
1712         stable_node_dup->head = &migrate_nodes;
1713         list_add(&stable_node_dup->list, stable_node_dup->head);
1714         return page;
1715
1716 chain_append:
1717         /* stable_node_dup could be null if it reached the limit */
1718         if (!stable_node_dup)
1719                 stable_node_dup = stable_node_any;
1720         /*
1721          * If stable_node was a chain and chain_prune collapsed it,
1722          * stable_node has been updated to be the new regular
1723          * stable_node. A collapse of the chain is indistinguishable
1724          * from the case there was no chain in the stable
1725          * rbtree. Otherwise stable_node is the chain and
1726          * stable_node_dup is the dup to replace.
1727          */
1728         if (stable_node_dup == stable_node) {
1729                 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1730                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1731                 /* chain is missing so create it */
1732                 stable_node = alloc_stable_node_chain(stable_node_dup,
1733                                                       root);
1734                 if (!stable_node)
1735                         return NULL;
1736         }
1737         /*
1738          * Add this stable_node dup that was
1739          * migrated to the stable_node chain
1740          * of the current nid for this page
1741          * content.
1742          */
1743         VM_BUG_ON(!is_stable_node_chain(stable_node));
1744         VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1745         VM_BUG_ON(page_node->head != &migrate_nodes);
1746         list_del(&page_node->list);
1747         DO_NUMA(page_node->nid = nid);
1748         stable_node_chain_add_dup(page_node, stable_node);
1749         goto out;
1750 }
1751
1752 /*
1753  * stable_tree_insert - insert stable tree node pointing to new ksm page
1754  * into the stable tree.
1755  *
1756  * This function returns the stable tree node just allocated on success,
1757  * NULL otherwise.
1758  */
1759 static struct stable_node *stable_tree_insert(struct page *kpage)
1760 {
1761         int nid;
1762         unsigned long kpfn;
1763         struct rb_root *root;
1764         struct rb_node **new;
1765         struct rb_node *parent;
1766         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1767         bool need_chain = false;
1768
1769         kpfn = page_to_pfn(kpage);
1770         nid = get_kpfn_nid(kpfn);
1771         root = root_stable_tree + nid;
1772 again:
1773         parent = NULL;
1774         new = &root->rb_node;
1775
1776         while (*new) {
1777                 struct page *tree_page;
1778                 int ret;
1779
1780                 cond_resched();
1781                 stable_node = rb_entry(*new, struct stable_node, node);
1782                 stable_node_any = NULL;
1783                 tree_page = chain(&stable_node_dup, stable_node, root);
1784                 if (!stable_node_dup) {
1785                         /*
1786                          * Either all stable_node dups were full in
1787                          * this stable_node chain, or this chain was
1788                          * empty and should be rb_erased.
1789                          */
1790                         stable_node_any = stable_node_dup_any(stable_node,
1791                                                               root);
1792                         if (!stable_node_any) {
1793                                 /* rb_erase just run */
1794                                 goto again;
1795                         }
1796                         /*
1797                          * Take any of the stable_node dups page of
1798                          * this stable_node chain to let the tree walk
1799                          * continue. All KSM pages belonging to the
1800                          * stable_node dups in a stable_node chain
1801                          * have the same content and they're
1802                          * wrprotected at all times. Any will work
1803                          * fine to continue the walk.
1804                          */
1805                         tree_page = get_ksm_page(stable_node_any, false);
1806                 }
1807                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1808                 if (!tree_page) {
1809                         /*
1810                          * If we walked over a stale stable_node,
1811                          * get_ksm_page() will call rb_erase() and it
1812                          * may rebalance the tree from under us. So
1813                          * restart the search from scratch. Returning
1814                          * NULL would be safe too, but we'd generate
1815                          * false negative insertions just because some
1816                          * stable_node was stale.
1817                          */
1818                         goto again;
1819                 }
1820
1821                 ret = memcmp_pages(kpage, tree_page);
1822                 put_page(tree_page);
1823
1824                 parent = *new;
1825                 if (ret < 0)
1826                         new = &parent->rb_left;
1827                 else if (ret > 0)
1828                         new = &parent->rb_right;
1829                 else {
1830                         need_chain = true;
1831                         break;
1832                 }
1833         }
1834
1835         stable_node_dup = alloc_stable_node();
1836         if (!stable_node_dup)
1837                 return NULL;
1838
1839         INIT_HLIST_HEAD(&stable_node_dup->hlist);
1840         stable_node_dup->kpfn = kpfn;
1841         set_page_stable_node(kpage, stable_node_dup);
1842         stable_node_dup->rmap_hlist_len = 0;
1843         DO_NUMA(stable_node_dup->nid = nid);
1844         if (!need_chain) {
1845                 rb_link_node(&stable_node_dup->node, parent, new);
1846                 rb_insert_color(&stable_node_dup->node, root);
1847         } else {
1848                 if (!is_stable_node_chain(stable_node)) {
1849                         struct stable_node *orig = stable_node;
1850                         /* chain is missing so create it */
1851                         stable_node = alloc_stable_node_chain(orig, root);
1852                         if (!stable_node) {
1853                                 free_stable_node(stable_node_dup);
1854                                 return NULL;
1855                         }
1856                 }
1857                 stable_node_chain_add_dup(stable_node_dup, stable_node);
1858         }
1859
1860         return stable_node_dup;
1861 }
1862
1863 /*
1864  * unstable_tree_search_insert - search for identical page,
1865  * else insert rmap_item into the unstable tree.
1866  *
1867  * This function searches for a page in the unstable tree identical to the
1868  * page currently being scanned; and if no identical page is found in the
1869  * tree, we insert rmap_item as a new object into the unstable tree.
1870  *
1871  * This function returns pointer to rmap_item found to be identical
1872  * to the currently scanned page, NULL otherwise.
1873  *
1874  * This function does both searching and inserting, because they share
1875  * the same walking algorithm in an rbtree.
1876  */
1877 static
1878 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1879                                               struct page *page,
1880                                               struct page **tree_pagep)
1881 {
1882         struct rb_node **new;
1883         struct rb_root *root;
1884         struct rb_node *parent = NULL;
1885         int nid;
1886
1887         nid = get_kpfn_nid(page_to_pfn(page));
1888         root = root_unstable_tree + nid;
1889         new = &root->rb_node;
1890
1891         while (*new) {
1892                 struct rmap_item *tree_rmap_item;
1893                 struct page *tree_page;
1894                 int ret;
1895
1896                 cond_resched();
1897                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1898                 tree_page = get_mergeable_page(tree_rmap_item);
1899                 if (!tree_page)
1900                         return NULL;
1901
1902                 /*
1903                  * Don't substitute a ksm page for a forked page.
1904                  */
1905                 if (page == tree_page) {
1906                         put_page(tree_page);
1907                         return NULL;
1908                 }
1909
1910                 ret = memcmp_pages(page, tree_page);
1911
1912                 parent = *new;
1913                 if (ret < 0) {
1914                         put_page(tree_page);
1915                         new = &parent->rb_left;
1916                 } else if (ret > 0) {
1917                         put_page(tree_page);
1918                         new = &parent->rb_right;
1919                 } else if (!ksm_merge_across_nodes &&
1920                            page_to_nid(tree_page) != nid) {
1921                         /*
1922                          * If tree_page has been migrated to another NUMA node,
1923                          * it will be flushed out and put in the right unstable
1924                          * tree next time: only merge with it when across_nodes.
1925                          */
1926                         put_page(tree_page);
1927                         return NULL;
1928                 } else {
1929                         *tree_pagep = tree_page;
1930                         return tree_rmap_item;
1931                 }
1932         }
1933
1934         rmap_item->address |= UNSTABLE_FLAG;
1935         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1936         DO_NUMA(rmap_item->nid = nid);
1937         rb_link_node(&rmap_item->node, parent, new);
1938         rb_insert_color(&rmap_item->node, root);
1939
1940         ksm_pages_unshared++;
1941         return NULL;
1942 }
1943
1944 /*
1945  * stable_tree_append - add another rmap_item to the linked list of
1946  * rmap_items hanging off a given node of the stable tree, all sharing
1947  * the same ksm page.
1948  */
1949 static void stable_tree_append(struct rmap_item *rmap_item,
1950                                struct stable_node *stable_node,
1951                                bool max_page_sharing_bypass)
1952 {
1953         /*
1954          * rmap won't find this mapping if we don't insert the
1955          * rmap_item in the right stable_node
1956          * duplicate. page_migration could break later if rmap breaks,
1957          * so we can as well crash here. We really need to check for
1958          * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
1959          * for other negative values as an undeflow if detected here
1960          * for the first time (and not when decreasing rmap_hlist_len)
1961          * would be sign of memory corruption in the stable_node.
1962          */
1963         BUG_ON(stable_node->rmap_hlist_len < 0);
1964
1965         stable_node->rmap_hlist_len++;
1966         if (!max_page_sharing_bypass)
1967                 /* possibly non fatal but unexpected overflow, only warn */
1968                 WARN_ON_ONCE(stable_node->rmap_hlist_len >
1969                              ksm_max_page_sharing);
1970
1971         rmap_item->head = stable_node;
1972         rmap_item->address |= STABLE_FLAG;
1973         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1974
1975         if (rmap_item->hlist.next)
1976                 ksm_pages_sharing++;
1977         else
1978                 ksm_pages_shared++;
1979 }
1980
1981 /*
1982  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1983  * if not, compare checksum to previous and if it's the same, see if page can
1984  * be inserted into the unstable tree, or merged with a page already there and
1985  * both transferred to the stable tree.
1986  *
1987  * @page: the page that we are searching identical page to.
1988  * @rmap_item: the reverse mapping into the virtual address of this page
1989  */
1990 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1991 {
1992         struct rmap_item *tree_rmap_item;
1993         struct page *tree_page = NULL;
1994         struct stable_node *stable_node;
1995         struct page *kpage;
1996         unsigned int checksum;
1997         int err;
1998         bool max_page_sharing_bypass = false;
1999
2000         stable_node = page_stable_node(page);
2001         if (stable_node) {
2002                 if (stable_node->head != &migrate_nodes &&
2003                     get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2004                     NUMA(stable_node->nid)) {
2005                         stable_node_dup_del(stable_node);
2006                         stable_node->head = &migrate_nodes;
2007                         list_add(&stable_node->list, stable_node->head);
2008                 }
2009                 if (stable_node->head != &migrate_nodes &&
2010                     rmap_item->head == stable_node)
2011                         return;
2012                 /*
2013                  * If it's a KSM fork, allow it to go over the sharing limit
2014                  * without warnings.
2015                  */
2016                 if (!is_page_sharing_candidate(stable_node))
2017                         max_page_sharing_bypass = true;
2018         }
2019
2020         /* We first start with searching the page inside the stable tree */
2021         kpage = stable_tree_search(page);
2022         if (kpage == page && rmap_item->head == stable_node) {
2023                 put_page(kpage);
2024                 return;
2025         }
2026
2027         remove_rmap_item_from_tree(rmap_item);
2028
2029         if (kpage) {
2030                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2031                 if (!err) {
2032                         /*
2033                          * The page was successfully merged:
2034                          * add its rmap_item to the stable tree.
2035                          */
2036                         lock_page(kpage);
2037                         stable_tree_append(rmap_item, page_stable_node(kpage),
2038                                            max_page_sharing_bypass);
2039                         unlock_page(kpage);
2040                 }
2041                 put_page(kpage);
2042                 return;
2043         }
2044
2045         /*
2046          * If the hash value of the page has changed from the last time
2047          * we calculated it, this page is changing frequently: therefore we
2048          * don't want to insert it in the unstable tree, and we don't want
2049          * to waste our time searching for something identical to it there.
2050          */
2051         checksum = calc_checksum(page);
2052         if (rmap_item->oldchecksum != checksum) {
2053                 rmap_item->oldchecksum = checksum;
2054                 return;
2055         }
2056
2057         /*
2058          * Same checksum as an empty page. We attempt to merge it with the
2059          * appropriate zero page if the user enabled this via sysfs.
2060          */
2061         if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2062                 struct vm_area_struct *vma;
2063
2064                 vma = find_mergeable_vma(rmap_item->mm, rmap_item->address);
2065                 err = try_to_merge_one_page(vma, page,
2066                                             ZERO_PAGE(rmap_item->address));
2067                 /*
2068                  * In case of failure, the page was not really empty, so we
2069                  * need to continue. Otherwise we're done.
2070                  */
2071                 if (!err)
2072                         return;
2073         }
2074         tree_rmap_item =
2075                 unstable_tree_search_insert(rmap_item, page, &tree_page);
2076         if (tree_rmap_item) {
2077                 kpage = try_to_merge_two_pages(rmap_item, page,
2078                                                 tree_rmap_item, tree_page);
2079                 put_page(tree_page);
2080                 if (kpage) {
2081                         /*
2082                          * The pages were successfully merged: insert new
2083                          * node in the stable tree and add both rmap_items.
2084                          */
2085                         lock_page(kpage);
2086                         stable_node = stable_tree_insert(kpage);
2087                         if (stable_node) {
2088                                 stable_tree_append(tree_rmap_item, stable_node,
2089                                                    false);
2090                                 stable_tree_append(rmap_item, stable_node,
2091                                                    false);
2092                         }
2093                         unlock_page(kpage);
2094
2095                         /*
2096                          * If we fail to insert the page into the stable tree,
2097                          * we will have 2 virtual addresses that are pointing
2098                          * to a ksm page left outside the stable tree,
2099                          * in which case we need to break_cow on both.
2100                          */
2101                         if (!stable_node) {
2102                                 break_cow(tree_rmap_item);
2103                                 break_cow(rmap_item);
2104                         }
2105                 }
2106         }
2107 }
2108
2109 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2110                                             struct rmap_item **rmap_list,
2111                                             unsigned long addr)
2112 {
2113         struct rmap_item *rmap_item;
2114
2115         while (*rmap_list) {
2116                 rmap_item = *rmap_list;
2117                 if ((rmap_item->address & PAGE_MASK) == addr)
2118                         return rmap_item;
2119                 if (rmap_item->address > addr)
2120                         break;
2121                 *rmap_list = rmap_item->rmap_list;
2122                 remove_rmap_item_from_tree(rmap_item);
2123                 free_rmap_item(rmap_item);
2124         }
2125
2126         rmap_item = alloc_rmap_item();
2127         if (rmap_item) {
2128                 /* It has already been zeroed */
2129                 rmap_item->mm = mm_slot->mm;
2130                 rmap_item->address = addr;
2131                 rmap_item->rmap_list = *rmap_list;
2132                 *rmap_list = rmap_item;
2133         }
2134         return rmap_item;
2135 }
2136
2137 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2138 {
2139         struct mm_struct *mm;
2140         struct mm_slot *slot;
2141         struct vm_area_struct *vma;
2142         struct rmap_item *rmap_item;
2143         int nid;
2144
2145         if (list_empty(&ksm_mm_head.mm_list))
2146                 return NULL;
2147
2148         slot = ksm_scan.mm_slot;
2149         if (slot == &ksm_mm_head) {
2150                 /*
2151                  * A number of pages can hang around indefinitely on per-cpu
2152                  * pagevecs, raised page count preventing write_protect_page
2153                  * from merging them.  Though it doesn't really matter much,
2154                  * it is puzzling to see some stuck in pages_volatile until
2155                  * other activity jostles them out, and they also prevented
2156                  * LTP's KSM test from succeeding deterministically; so drain
2157                  * them here (here rather than on entry to ksm_do_scan(),
2158                  * so we don't IPI too often when pages_to_scan is set low).
2159                  */
2160                 lru_add_drain_all();
2161
2162                 /*
2163                  * Whereas stale stable_nodes on the stable_tree itself
2164                  * get pruned in the regular course of stable_tree_search(),
2165                  * those moved out to the migrate_nodes list can accumulate:
2166                  * so prune them once before each full scan.
2167                  */
2168                 if (!ksm_merge_across_nodes) {
2169                         struct stable_node *stable_node, *next;
2170                         struct page *page;
2171
2172                         list_for_each_entry_safe(stable_node, next,
2173                                                  &migrate_nodes, list) {
2174                                 page = get_ksm_page(stable_node, false);
2175                                 if (page)
2176                                         put_page(page);
2177                                 cond_resched();
2178                         }
2179                 }
2180
2181                 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2182                         root_unstable_tree[nid] = RB_ROOT;
2183
2184                 spin_lock(&ksm_mmlist_lock);
2185                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2186                 ksm_scan.mm_slot = slot;
2187                 spin_unlock(&ksm_mmlist_lock);
2188                 /*
2189                  * Although we tested list_empty() above, a racing __ksm_exit
2190                  * of the last mm on the list may have removed it since then.
2191                  */
2192                 if (slot == &ksm_mm_head)
2193                         return NULL;
2194 next_mm:
2195                 ksm_scan.address = 0;
2196                 ksm_scan.rmap_list = &slot->rmap_list;
2197         }
2198
2199         mm = slot->mm;
2200         down_read(&mm->mmap_sem);
2201         if (ksm_test_exit(mm))
2202                 vma = NULL;
2203         else
2204                 vma = find_vma(mm, ksm_scan.address);
2205
2206         for (; vma; vma = vma->vm_next) {
2207                 if (!(vma->vm_flags & VM_MERGEABLE))
2208                         continue;
2209                 if (ksm_scan.address < vma->vm_start)
2210                         ksm_scan.address = vma->vm_start;
2211                 if (!vma->anon_vma)
2212                         ksm_scan.address = vma->vm_end;
2213
2214                 while (ksm_scan.address < vma->vm_end) {
2215                         if (ksm_test_exit(mm))
2216                                 break;
2217                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2218                         if (IS_ERR_OR_NULL(*page)) {
2219                                 ksm_scan.address += PAGE_SIZE;
2220                                 cond_resched();
2221                                 continue;
2222                         }
2223                         if (PageAnon(*page)) {
2224                                 flush_anon_page(vma, *page, ksm_scan.address);
2225                                 flush_dcache_page(*page);
2226                                 rmap_item = get_next_rmap_item(slot,
2227                                         ksm_scan.rmap_list, ksm_scan.address);
2228                                 if (rmap_item) {
2229                                         ksm_scan.rmap_list =
2230                                                         &rmap_item->rmap_list;
2231                                         ksm_scan.address += PAGE_SIZE;
2232                                 } else
2233                                         put_page(*page);
2234                                 up_read(&mm->mmap_sem);
2235                                 return rmap_item;
2236                         }
2237                         put_page(*page);
2238                         ksm_scan.address += PAGE_SIZE;
2239                         cond_resched();
2240                 }
2241         }
2242
2243         if (ksm_test_exit(mm)) {
2244                 ksm_scan.address = 0;
2245                 ksm_scan.rmap_list = &slot->rmap_list;
2246         }
2247         /*
2248          * Nuke all the rmap_items that are above this current rmap:
2249          * because there were no VM_MERGEABLE vmas with such addresses.
2250          */
2251         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2252
2253         spin_lock(&ksm_mmlist_lock);
2254         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2255                                                 struct mm_slot, mm_list);
2256         if (ksm_scan.address == 0) {
2257                 /*
2258                  * We've completed a full scan of all vmas, holding mmap_sem
2259                  * throughout, and found no VM_MERGEABLE: so do the same as
2260                  * __ksm_exit does to remove this mm from all our lists now.
2261                  * This applies either when cleaning up after __ksm_exit
2262                  * (but beware: we can reach here even before __ksm_exit),
2263                  * or when all VM_MERGEABLE areas have been unmapped (and
2264                  * mmap_sem then protects against race with MADV_MERGEABLE).
2265                  */
2266                 hash_del(&slot->link);
2267                 list_del(&slot->mm_list);
2268                 spin_unlock(&ksm_mmlist_lock);
2269
2270                 free_mm_slot(slot);
2271                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2272                 up_read(&mm->mmap_sem);
2273                 mmdrop(mm);
2274         } else {
2275                 up_read(&mm->mmap_sem);
2276                 /*
2277                  * up_read(&mm->mmap_sem) first because after
2278                  * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2279                  * already have been freed under us by __ksm_exit()
2280                  * because the "mm_slot" is still hashed and
2281                  * ksm_scan.mm_slot doesn't point to it anymore.
2282                  */
2283                 spin_unlock(&ksm_mmlist_lock);
2284         }
2285
2286         /* Repeat until we've completed scanning the whole list */
2287         slot = ksm_scan.mm_slot;
2288         if (slot != &ksm_mm_head)
2289                 goto next_mm;
2290
2291         ksm_scan.seqnr++;
2292         return NULL;
2293 }
2294
2295 /**
2296  * ksm_do_scan  - the ksm scanner main worker function.
2297  * @scan_npages - number of pages we want to scan before we return.
2298  */
2299 static void ksm_do_scan(unsigned int scan_npages)
2300 {
2301         struct rmap_item *rmap_item;
2302         struct page *uninitialized_var(page);
2303
2304         while (scan_npages-- && likely(!freezing(current))) {
2305                 cond_resched();
2306                 rmap_item = scan_get_next_rmap_item(&page);
2307                 if (!rmap_item)
2308                         return;
2309                 cmp_and_merge_page(page, rmap_item);
2310                 put_page(page);
2311         }
2312 }
2313
2314 static int ksmd_should_run(void)
2315 {
2316         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2317 }
2318
2319 static int ksm_scan_thread(void *nothing)
2320 {
2321         set_freezable();
2322         set_user_nice(current, 5);
2323
2324         while (!kthread_should_stop()) {
2325                 mutex_lock(&ksm_thread_mutex);
2326                 wait_while_offlining();
2327                 if (ksmd_should_run())
2328                         ksm_do_scan(ksm_thread_pages_to_scan);
2329                 mutex_unlock(&ksm_thread_mutex);
2330
2331                 try_to_freeze();
2332
2333                 if (ksmd_should_run()) {
2334                         schedule_timeout_interruptible(
2335                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
2336                 } else {
2337                         wait_event_freezable(ksm_thread_wait,
2338                                 ksmd_should_run() || kthread_should_stop());
2339                 }
2340         }
2341         return 0;
2342 }
2343
2344 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2345                 unsigned long end, int advice, unsigned long *vm_flags)
2346 {
2347         struct mm_struct *mm = vma->vm_mm;
2348         int err;
2349
2350         switch (advice) {
2351         case MADV_MERGEABLE:
2352                 /*
2353                  * Be somewhat over-protective for now!
2354                  */
2355                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
2356                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
2357                                  VM_HUGETLB | VM_MIXEDMAP))
2358                         return 0;               /* just ignore the advice */
2359
2360 #ifdef VM_SAO
2361                 if (*vm_flags & VM_SAO)
2362                         return 0;
2363 #endif
2364
2365                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2366                         err = __ksm_enter(mm);
2367                         if (err)
2368                                 return err;
2369                 }
2370
2371                 *vm_flags |= VM_MERGEABLE;
2372                 break;
2373
2374         case MADV_UNMERGEABLE:
2375                 if (!(*vm_flags & VM_MERGEABLE))
2376                         return 0;               /* just ignore the advice */
2377
2378                 if (vma->anon_vma) {
2379                         err = unmerge_ksm_pages(vma, start, end);
2380                         if (err)
2381                                 return err;
2382                 }
2383
2384                 *vm_flags &= ~VM_MERGEABLE;
2385                 break;
2386         }
2387
2388         return 0;
2389 }
2390
2391 int __ksm_enter(struct mm_struct *mm)
2392 {
2393         struct mm_slot *mm_slot;
2394         int needs_wakeup;
2395
2396         mm_slot = alloc_mm_slot();
2397         if (!mm_slot)
2398                 return -ENOMEM;
2399
2400         /* Check ksm_run too?  Would need tighter locking */
2401         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2402
2403         spin_lock(&ksm_mmlist_lock);
2404         insert_to_mm_slots_hash(mm, mm_slot);
2405         /*
2406          * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2407          * insert just behind the scanning cursor, to let the area settle
2408          * down a little; when fork is followed by immediate exec, we don't
2409          * want ksmd to waste time setting up and tearing down an rmap_list.
2410          *
2411          * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2412          * scanning cursor, otherwise KSM pages in newly forked mms will be
2413          * missed: then we might as well insert at the end of the list.
2414          */
2415         if (ksm_run & KSM_RUN_UNMERGE)
2416                 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2417         else
2418                 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2419         spin_unlock(&ksm_mmlist_lock);
2420
2421         set_bit(MMF_VM_MERGEABLE, &mm->flags);
2422         mmgrab(mm);
2423
2424         if (needs_wakeup)
2425                 wake_up_interruptible(&ksm_thread_wait);
2426
2427         return 0;
2428 }
2429
2430 void __ksm_exit(struct mm_struct *mm)
2431 {
2432         struct mm_slot *mm_slot;
2433         int easy_to_free = 0;
2434
2435         /*
2436          * This process is exiting: if it's straightforward (as is the
2437          * case when ksmd was never running), free mm_slot immediately.
2438          * But if it's at the cursor or has rmap_items linked to it, use
2439          * mmap_sem to synchronize with any break_cows before pagetables
2440          * are freed, and leave the mm_slot on the list for ksmd to free.
2441          * Beware: ksm may already have noticed it exiting and freed the slot.
2442          */
2443
2444         spin_lock(&ksm_mmlist_lock);
2445         mm_slot = get_mm_slot(mm);
2446         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2447                 if (!mm_slot->rmap_list) {
2448                         hash_del(&mm_slot->link);
2449                         list_del(&mm_slot->mm_list);
2450                         easy_to_free = 1;
2451                 } else {
2452                         list_move(&mm_slot->mm_list,
2453                                   &ksm_scan.mm_slot->mm_list);
2454                 }
2455         }
2456         spin_unlock(&ksm_mmlist_lock);
2457
2458         if (easy_to_free) {
2459                 free_mm_slot(mm_slot);
2460                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2461                 mmdrop(mm);
2462         } else if (mm_slot) {
2463                 down_write(&mm->mmap_sem);
2464                 up_write(&mm->mmap_sem);
2465         }
2466 }
2467
2468 struct page *ksm_might_need_to_copy(struct page *page,
2469                         struct vm_area_struct *vma, unsigned long address)
2470 {
2471         struct anon_vma *anon_vma = page_anon_vma(page);
2472         struct page *new_page;
2473
2474         if (PageKsm(page)) {
2475                 if (page_stable_node(page) &&
2476                     !(ksm_run & KSM_RUN_UNMERGE))
2477                         return page;    /* no need to copy it */
2478         } else if (!anon_vma) {
2479                 return page;            /* no need to copy it */
2480         } else if (anon_vma->root == vma->anon_vma->root &&
2481                  page->index == linear_page_index(vma, address)) {
2482                 return page;            /* still no need to copy it */
2483         }
2484         if (!PageUptodate(page))
2485                 return page;            /* let do_swap_page report the error */
2486
2487         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2488         if (new_page) {
2489                 copy_user_highpage(new_page, page, address, vma);
2490
2491                 SetPageDirty(new_page);
2492                 __SetPageUptodate(new_page);
2493                 __SetPageLocked(new_page);
2494         }
2495
2496         return new_page;
2497 }
2498
2499 void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2500 {
2501         struct stable_node *stable_node;
2502         struct rmap_item *rmap_item;
2503         int search_new_forks = 0;
2504
2505         VM_BUG_ON_PAGE(!PageKsm(page), page);
2506
2507         /*
2508          * Rely on the page lock to protect against concurrent modifications
2509          * to that page's node of the stable tree.
2510          */
2511         VM_BUG_ON_PAGE(!PageLocked(page), page);
2512
2513         stable_node = page_stable_node(page);
2514         if (!stable_node)
2515                 return;
2516 again:
2517         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2518                 struct anon_vma *anon_vma = rmap_item->anon_vma;
2519                 struct anon_vma_chain *vmac;
2520                 struct vm_area_struct *vma;
2521
2522                 cond_resched();
2523                 anon_vma_lock_read(anon_vma);
2524                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2525                                                0, ULONG_MAX) {
2526                         cond_resched();
2527                         vma = vmac->vma;
2528                         if (rmap_item->address < vma->vm_start ||
2529                             rmap_item->address >= vma->vm_end)
2530                                 continue;
2531                         /*
2532                          * Initially we examine only the vma which covers this
2533                          * rmap_item; but later, if there is still work to do,
2534                          * we examine covering vmas in other mms: in case they
2535                          * were forked from the original since ksmd passed.
2536                          */
2537                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2538                                 continue;
2539
2540                         if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2541                                 continue;
2542
2543                         if (!rwc->rmap_one(page, vma,
2544                                         rmap_item->address, rwc->arg)) {
2545                                 anon_vma_unlock_read(anon_vma);
2546                                 return;
2547                         }
2548                         if (rwc->done && rwc->done(page)) {
2549                                 anon_vma_unlock_read(anon_vma);
2550                                 return;
2551                         }
2552                 }
2553                 anon_vma_unlock_read(anon_vma);
2554         }
2555         if (!search_new_forks++)
2556                 goto again;
2557 }
2558
2559 #ifdef CONFIG_MIGRATION
2560 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2561 {
2562         struct stable_node *stable_node;
2563
2564         VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2565         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2566         VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2567
2568         stable_node = page_stable_node(newpage);
2569         if (stable_node) {
2570                 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2571                 stable_node->kpfn = page_to_pfn(newpage);
2572                 /*
2573                  * newpage->mapping was set in advance; now we need smp_wmb()
2574                  * to make sure that the new stable_node->kpfn is visible
2575                  * to get_ksm_page() before it can see that oldpage->mapping
2576                  * has gone stale (or that PageSwapCache has been cleared).
2577                  */
2578                 smp_wmb();
2579                 set_page_stable_node(oldpage, NULL);
2580         }
2581 }
2582 #endif /* CONFIG_MIGRATION */
2583
2584 #ifdef CONFIG_MEMORY_HOTREMOVE
2585 static void wait_while_offlining(void)
2586 {
2587         while (ksm_run & KSM_RUN_OFFLINE) {
2588                 mutex_unlock(&ksm_thread_mutex);
2589                 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2590                             TASK_UNINTERRUPTIBLE);
2591                 mutex_lock(&ksm_thread_mutex);
2592         }
2593 }
2594
2595 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2596                                          unsigned long start_pfn,
2597                                          unsigned long end_pfn)
2598 {
2599         if (stable_node->kpfn >= start_pfn &&
2600             stable_node->kpfn < end_pfn) {
2601                 /*
2602                  * Don't get_ksm_page, page has already gone:
2603                  * which is why we keep kpfn instead of page*
2604                  */
2605                 remove_node_from_stable_tree(stable_node);
2606                 return true;
2607         }
2608         return false;
2609 }
2610
2611 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2612                                            unsigned long start_pfn,
2613                                            unsigned long end_pfn,
2614                                            struct rb_root *root)
2615 {
2616         struct stable_node *dup;
2617         struct hlist_node *hlist_safe;
2618
2619         if (!is_stable_node_chain(stable_node)) {
2620                 VM_BUG_ON(is_stable_node_dup(stable_node));
2621                 return stable_node_dup_remove_range(stable_node, start_pfn,
2622                                                     end_pfn);
2623         }
2624
2625         hlist_for_each_entry_safe(dup, hlist_safe,
2626                                   &stable_node->hlist, hlist_dup) {
2627                 VM_BUG_ON(!is_stable_node_dup(dup));
2628                 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2629         }
2630         if (hlist_empty(&stable_node->hlist)) {
2631                 free_stable_node_chain(stable_node, root);
2632                 return true; /* notify caller that tree was rebalanced */
2633         } else
2634                 return false;
2635 }
2636
2637 static void ksm_check_stable_tree(unsigned long start_pfn,
2638                                   unsigned long end_pfn)
2639 {
2640         struct stable_node *stable_node, *next;
2641         struct rb_node *node;
2642         int nid;
2643
2644         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2645                 node = rb_first(root_stable_tree + nid);
2646                 while (node) {
2647                         stable_node = rb_entry(node, struct stable_node, node);
2648                         if (stable_node_chain_remove_range(stable_node,
2649                                                            start_pfn, end_pfn,
2650                                                            root_stable_tree +
2651                                                            nid))
2652                                 node = rb_first(root_stable_tree + nid);
2653                         else
2654                                 node = rb_next(node);
2655                         cond_resched();
2656                 }
2657         }
2658         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2659                 if (stable_node->kpfn >= start_pfn &&
2660                     stable_node->kpfn < end_pfn)
2661                         remove_node_from_stable_tree(stable_node);
2662                 cond_resched();
2663         }
2664 }
2665
2666 static int ksm_memory_callback(struct notifier_block *self,
2667                                unsigned long action, void *arg)
2668 {
2669         struct memory_notify *mn = arg;
2670
2671         switch (action) {
2672         case MEM_GOING_OFFLINE:
2673                 /*
2674                  * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2675                  * and remove_all_stable_nodes() while memory is going offline:
2676                  * it is unsafe for them to touch the stable tree at this time.
2677                  * But unmerge_ksm_pages(), rmap lookups and other entry points
2678                  * which do not need the ksm_thread_mutex are all safe.
2679                  */
2680                 mutex_lock(&ksm_thread_mutex);
2681                 ksm_run |= KSM_RUN_OFFLINE;
2682                 mutex_unlock(&ksm_thread_mutex);
2683                 break;
2684
2685         case MEM_OFFLINE:
2686                 /*
2687                  * Most of the work is done by page migration; but there might
2688                  * be a few stable_nodes left over, still pointing to struct
2689                  * pages which have been offlined: prune those from the tree,
2690                  * otherwise get_ksm_page() might later try to access a
2691                  * non-existent struct page.
2692                  */
2693                 ksm_check_stable_tree(mn->start_pfn,
2694                                       mn->start_pfn + mn->nr_pages);
2695                 /* fallthrough */
2696
2697         case MEM_CANCEL_OFFLINE:
2698                 mutex_lock(&ksm_thread_mutex);
2699                 ksm_run &= ~KSM_RUN_OFFLINE;
2700                 mutex_unlock(&ksm_thread_mutex);
2701
2702                 smp_mb();       /* wake_up_bit advises this */
2703                 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2704                 break;
2705         }
2706         return NOTIFY_OK;
2707 }
2708 #else
2709 static void wait_while_offlining(void)
2710 {
2711 }
2712 #endif /* CONFIG_MEMORY_HOTREMOVE */
2713
2714 #ifdef CONFIG_SYSFS
2715 /*
2716  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2717  */
2718
2719 #define KSM_ATTR_RO(_name) \
2720         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2721 #define KSM_ATTR(_name) \
2722         static struct kobj_attribute _name##_attr = \
2723                 __ATTR(_name, 0644, _name##_show, _name##_store)
2724
2725 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2726                                     struct kobj_attribute *attr, char *buf)
2727 {
2728         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2729 }
2730
2731 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2732                                      struct kobj_attribute *attr,
2733                                      const char *buf, size_t count)
2734 {
2735         unsigned long msecs;
2736         int err;
2737
2738         err = kstrtoul(buf, 10, &msecs);
2739         if (err || msecs > UINT_MAX)
2740                 return -EINVAL;
2741
2742         ksm_thread_sleep_millisecs = msecs;
2743
2744         return count;
2745 }
2746 KSM_ATTR(sleep_millisecs);
2747
2748 static ssize_t pages_to_scan_show(struct kobject *kobj,
2749                                   struct kobj_attribute *attr, char *buf)
2750 {
2751         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2752 }
2753
2754 static ssize_t pages_to_scan_store(struct kobject *kobj,
2755                                    struct kobj_attribute *attr,
2756                                    const char *buf, size_t count)
2757 {
2758         int err;
2759         unsigned long nr_pages;
2760
2761         err = kstrtoul(buf, 10, &nr_pages);
2762         if (err || nr_pages > UINT_MAX)
2763                 return -EINVAL;
2764
2765         ksm_thread_pages_to_scan = nr_pages;
2766
2767         return count;
2768 }
2769 KSM_ATTR(pages_to_scan);
2770
2771 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2772                         char *buf)
2773 {
2774         return sprintf(buf, "%lu\n", ksm_run);
2775 }
2776
2777 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2778                          const char *buf, size_t count)
2779 {
2780         int err;
2781         unsigned long flags;
2782
2783         err = kstrtoul(buf, 10, &flags);
2784         if (err || flags > UINT_MAX)
2785                 return -EINVAL;
2786         if (flags > KSM_RUN_UNMERGE)
2787                 return -EINVAL;
2788
2789         /*
2790          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2791          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2792          * breaking COW to free the pages_shared (but leaves mm_slots
2793          * on the list for when ksmd may be set running again).
2794          */
2795
2796         mutex_lock(&ksm_thread_mutex);
2797         wait_while_offlining();
2798         if (ksm_run != flags) {
2799                 ksm_run = flags;
2800                 if (flags & KSM_RUN_UNMERGE) {
2801                         set_current_oom_origin();
2802                         err = unmerge_and_remove_all_rmap_items();
2803                         clear_current_oom_origin();
2804                         if (err) {
2805                                 ksm_run = KSM_RUN_STOP;
2806                                 count = err;
2807                         }
2808                 }
2809         }
2810         mutex_unlock(&ksm_thread_mutex);
2811
2812         if (flags & KSM_RUN_MERGE)
2813                 wake_up_interruptible(&ksm_thread_wait);
2814
2815         return count;
2816 }
2817 KSM_ATTR(run);
2818
2819 #ifdef CONFIG_NUMA
2820 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2821                                 struct kobj_attribute *attr, char *buf)
2822 {
2823         return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2824 }
2825
2826 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2827                                    struct kobj_attribute *attr,
2828                                    const char *buf, size_t count)
2829 {
2830         int err;
2831         unsigned long knob;
2832
2833         err = kstrtoul(buf, 10, &knob);
2834         if (err)
2835                 return err;
2836         if (knob > 1)
2837                 return -EINVAL;
2838
2839         mutex_lock(&ksm_thread_mutex);
2840         wait_while_offlining();
2841         if (ksm_merge_across_nodes != knob) {
2842                 if (ksm_pages_shared || remove_all_stable_nodes())
2843                         err = -EBUSY;
2844                 else if (root_stable_tree == one_stable_tree) {
2845                         struct rb_root *buf;
2846                         /*
2847                          * This is the first time that we switch away from the
2848                          * default of merging across nodes: must now allocate
2849                          * a buffer to hold as many roots as may be needed.
2850                          * Allocate stable and unstable together:
2851                          * MAXSMP NODES_SHIFT 10 will use 16kB.
2852                          */
2853                         buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2854                                       GFP_KERNEL);
2855                         /* Let us assume that RB_ROOT is NULL is zero */
2856                         if (!buf)
2857                                 err = -ENOMEM;
2858                         else {
2859                                 root_stable_tree = buf;
2860                                 root_unstable_tree = buf + nr_node_ids;
2861                                 /* Stable tree is empty but not the unstable */
2862                                 root_unstable_tree[0] = one_unstable_tree[0];
2863                         }
2864                 }
2865                 if (!err) {
2866                         ksm_merge_across_nodes = knob;
2867                         ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2868                 }
2869         }
2870         mutex_unlock(&ksm_thread_mutex);
2871
2872         return err ? err : count;
2873 }
2874 KSM_ATTR(merge_across_nodes);
2875 #endif
2876
2877 static ssize_t use_zero_pages_show(struct kobject *kobj,
2878                                 struct kobj_attribute *attr, char *buf)
2879 {
2880         return sprintf(buf, "%u\n", ksm_use_zero_pages);
2881 }
2882 static ssize_t use_zero_pages_store(struct kobject *kobj,
2883                                    struct kobj_attribute *attr,
2884                                    const char *buf, size_t count)
2885 {
2886         int err;
2887         bool value;
2888
2889         err = kstrtobool(buf, &value);
2890         if (err)
2891                 return -EINVAL;
2892
2893         ksm_use_zero_pages = value;
2894
2895         return count;
2896 }
2897 KSM_ATTR(use_zero_pages);
2898
2899 static ssize_t max_page_sharing_show(struct kobject *kobj,
2900                                      struct kobj_attribute *attr, char *buf)
2901 {
2902         return sprintf(buf, "%u\n", ksm_max_page_sharing);
2903 }
2904
2905 static ssize_t max_page_sharing_store(struct kobject *kobj,
2906                                       struct kobj_attribute *attr,
2907                                       const char *buf, size_t count)
2908 {
2909         int err;
2910         int knob;
2911
2912         err = kstrtoint(buf, 10, &knob);
2913         if (err)
2914                 return err;
2915         /*
2916          * When a KSM page is created it is shared by 2 mappings. This
2917          * being a signed comparison, it implicitly verifies it's not
2918          * negative.
2919          */
2920         if (knob < 2)
2921                 return -EINVAL;
2922
2923         if (READ_ONCE(ksm_max_page_sharing) == knob)
2924                 return count;
2925
2926         mutex_lock(&ksm_thread_mutex);
2927         wait_while_offlining();
2928         if (ksm_max_page_sharing != knob) {
2929                 if (ksm_pages_shared || remove_all_stable_nodes())
2930                         err = -EBUSY;
2931                 else
2932                         ksm_max_page_sharing = knob;
2933         }
2934         mutex_unlock(&ksm_thread_mutex);
2935
2936         return err ? err : count;
2937 }
2938 KSM_ATTR(max_page_sharing);
2939
2940 static ssize_t pages_shared_show(struct kobject *kobj,
2941                                  struct kobj_attribute *attr, char *buf)
2942 {
2943         return sprintf(buf, "%lu\n", ksm_pages_shared);
2944 }
2945 KSM_ATTR_RO(pages_shared);
2946
2947 static ssize_t pages_sharing_show(struct kobject *kobj,
2948                                   struct kobj_attribute *attr, char *buf)
2949 {
2950         return sprintf(buf, "%lu\n", ksm_pages_sharing);
2951 }
2952 KSM_ATTR_RO(pages_sharing);
2953
2954 static ssize_t pages_unshared_show(struct kobject *kobj,
2955                                    struct kobj_attribute *attr, char *buf)
2956 {
2957         return sprintf(buf, "%lu\n", ksm_pages_unshared);
2958 }
2959 KSM_ATTR_RO(pages_unshared);
2960
2961 static ssize_t pages_volatile_show(struct kobject *kobj,
2962                                    struct kobj_attribute *attr, char *buf)
2963 {
2964         long ksm_pages_volatile;
2965
2966         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2967                                 - ksm_pages_sharing - ksm_pages_unshared;
2968         /*
2969          * It was not worth any locking to calculate that statistic,
2970          * but it might therefore sometimes be negative: conceal that.
2971          */
2972         if (ksm_pages_volatile < 0)
2973                 ksm_pages_volatile = 0;
2974         return sprintf(buf, "%ld\n", ksm_pages_volatile);
2975 }
2976 KSM_ATTR_RO(pages_volatile);
2977
2978 static ssize_t stable_node_dups_show(struct kobject *kobj,
2979                                      struct kobj_attribute *attr, char *buf)
2980 {
2981         return sprintf(buf, "%lu\n", ksm_stable_node_dups);
2982 }
2983 KSM_ATTR_RO(stable_node_dups);
2984
2985 static ssize_t stable_node_chains_show(struct kobject *kobj,
2986                                        struct kobj_attribute *attr, char *buf)
2987 {
2988         return sprintf(buf, "%lu\n", ksm_stable_node_chains);
2989 }
2990 KSM_ATTR_RO(stable_node_chains);
2991
2992 static ssize_t
2993 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
2994                                         struct kobj_attribute *attr,
2995                                         char *buf)
2996 {
2997         return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
2998 }
2999
3000 static ssize_t
3001 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3002                                          struct kobj_attribute *attr,
3003                                          const char *buf, size_t count)
3004 {
3005         unsigned long msecs;
3006         int err;
3007
3008         err = kstrtoul(buf, 10, &msecs);
3009         if (err || msecs > UINT_MAX)
3010                 return -EINVAL;
3011
3012         ksm_stable_node_chains_prune_millisecs = msecs;
3013
3014         return count;
3015 }
3016 KSM_ATTR(stable_node_chains_prune_millisecs);
3017
3018 static ssize_t full_scans_show(struct kobject *kobj,
3019                                struct kobj_attribute *attr, char *buf)
3020 {
3021         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3022 }
3023 KSM_ATTR_RO(full_scans);
3024
3025 static struct attribute *ksm_attrs[] = {
3026         &sleep_millisecs_attr.attr,
3027         &pages_to_scan_attr.attr,
3028         &run_attr.attr,
3029         &pages_shared_attr.attr,
3030         &pages_sharing_attr.attr,
3031         &pages_unshared_attr.attr,
3032         &pages_volatile_attr.attr,
3033         &full_scans_attr.attr,
3034 #ifdef CONFIG_NUMA
3035         &merge_across_nodes_attr.attr,
3036 #endif
3037         &max_page_sharing_attr.attr,
3038         &stable_node_chains_attr.attr,
3039         &stable_node_dups_attr.attr,
3040         &stable_node_chains_prune_millisecs_attr.attr,
3041         &use_zero_pages_attr.attr,
3042         NULL,
3043 };
3044
3045 static struct attribute_group ksm_attr_group = {
3046         .attrs = ksm_attrs,
3047         .name = "ksm",
3048 };
3049 #endif /* CONFIG_SYSFS */
3050
3051 static int __init ksm_init(void)
3052 {
3053         struct task_struct *ksm_thread;
3054         int err;
3055
3056         /* The correct value depends on page size and endianness */
3057         zero_checksum = calc_checksum(ZERO_PAGE(0));
3058         /* Default to false for backwards compatibility */
3059         ksm_use_zero_pages = false;
3060
3061         err = ksm_slab_init();
3062         if (err)
3063                 goto out;
3064
3065         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3066         if (IS_ERR(ksm_thread)) {
3067                 pr_err("ksm: creating kthread failed\n");
3068                 err = PTR_ERR(ksm_thread);
3069                 goto out_free;
3070         }
3071
3072 #ifdef CONFIG_SYSFS
3073         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3074         if (err) {
3075                 pr_err("ksm: register sysfs failed\n");
3076                 kthread_stop(ksm_thread);
3077                 goto out_free;
3078         }
3079 #else
3080         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
3081
3082 #endif /* CONFIG_SYSFS */
3083
3084 #ifdef CONFIG_MEMORY_HOTREMOVE
3085         /* There is no significance to this priority 100 */
3086         hotplug_memory_notifier(ksm_memory_callback, 100);
3087 #endif
3088         return 0;
3089
3090 out_free:
3091         ksm_slab_free();
3092 out:
3093         return err;
3094 }
3095 subsys_initcall(ksm_init);