mm/ksm.c: make stable_node_dup() static
[muen/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); /* Address dependency. */
679         page = pfn_to_page(kpfn);
680         if (READ_ONCE(page->mapping) != expected_mapping)
681                 goto stale;
682
683         /*
684          * We cannot do anything with the page while its refcount is 0.
685          * Usually 0 means free, or tail of a higher-order page: in which
686          * case this node is no longer referenced, and should be freed;
687          * however, it might mean that the page is under page_freeze_refs().
688          * The __remove_mapping() case is easy, again the node is now stale;
689          * but if page is swapcache in migrate_page_move_mapping(), it might
690          * still be our page, in which case it's essential to keep the node.
691          */
692         while (!get_page_unless_zero(page)) {
693                 /*
694                  * Another check for page->mapping != expected_mapping would
695                  * work here too.  We have chosen the !PageSwapCache test to
696                  * optimize the common case, when the page is or is about to
697                  * be freed: PageSwapCache is cleared (under spin_lock_irq)
698                  * in the freeze_refs section of __remove_mapping(); but Anon
699                  * page->mapping reset to NULL later, in free_pages_prepare().
700                  */
701                 if (!PageSwapCache(page))
702                         goto stale;
703                 cpu_relax();
704         }
705
706         if (READ_ONCE(page->mapping) != expected_mapping) {
707                 put_page(page);
708                 goto stale;
709         }
710
711         if (lock_it) {
712                 lock_page(page);
713                 if (READ_ONCE(page->mapping) != expected_mapping) {
714                         unlock_page(page);
715                         put_page(page);
716                         goto stale;
717                 }
718         }
719         return page;
720
721 stale:
722         /*
723          * We come here from above when page->mapping or !PageSwapCache
724          * suggests that the node is stale; but it might be under migration.
725          * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
726          * before checking whether node->kpfn has been changed.
727          */
728         smp_rmb();
729         if (READ_ONCE(stable_node->kpfn) != kpfn)
730                 goto again;
731         remove_node_from_stable_tree(stable_node);
732         return NULL;
733 }
734
735 /*
736  * Removing rmap_item from stable or unstable tree.
737  * This function will clean the information from the stable/unstable tree.
738  */
739 static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
740 {
741         if (rmap_item->address & STABLE_FLAG) {
742                 struct stable_node *stable_node;
743                 struct page *page;
744
745                 stable_node = rmap_item->head;
746                 page = get_ksm_page(stable_node, true);
747                 if (!page)
748                         goto out;
749
750                 hlist_del(&rmap_item->hlist);
751                 unlock_page(page);
752                 put_page(page);
753
754                 if (!hlist_empty(&stable_node->hlist))
755                         ksm_pages_sharing--;
756                 else
757                         ksm_pages_shared--;
758                 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
759                 stable_node->rmap_hlist_len--;
760
761                 put_anon_vma(rmap_item->anon_vma);
762                 rmap_item->address &= PAGE_MASK;
763
764         } else if (rmap_item->address & UNSTABLE_FLAG) {
765                 unsigned char age;
766                 /*
767                  * Usually ksmd can and must skip the rb_erase, because
768                  * root_unstable_tree was already reset to RB_ROOT.
769                  * But be careful when an mm is exiting: do the rb_erase
770                  * if this rmap_item was inserted by this scan, rather
771                  * than left over from before.
772                  */
773                 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
774                 BUG_ON(age > 1);
775                 if (!age)
776                         rb_erase(&rmap_item->node,
777                                  root_unstable_tree + NUMA(rmap_item->nid));
778                 ksm_pages_unshared--;
779                 rmap_item->address &= PAGE_MASK;
780         }
781 out:
782         cond_resched();         /* we're called from many long loops */
783 }
784
785 static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
786                                        struct rmap_item **rmap_list)
787 {
788         while (*rmap_list) {
789                 struct rmap_item *rmap_item = *rmap_list;
790                 *rmap_list = rmap_item->rmap_list;
791                 remove_rmap_item_from_tree(rmap_item);
792                 free_rmap_item(rmap_item);
793         }
794 }
795
796 /*
797  * Though it's very tempting to unmerge rmap_items from stable tree rather
798  * than check every pte of a given vma, the locking doesn't quite work for
799  * that - an rmap_item is assigned to the stable tree after inserting ksm
800  * page and upping mmap_sem.  Nor does it fit with the way we skip dup'ing
801  * rmap_items from parent to child at fork time (so as not to waste time
802  * if exit comes before the next scan reaches it).
803  *
804  * Similarly, although we'd like to remove rmap_items (so updating counts
805  * and freeing memory) when unmerging an area, it's easier to leave that
806  * to the next pass of ksmd - consider, for example, how ksmd might be
807  * in cmp_and_merge_page on one of the rmap_items we would be removing.
808  */
809 static int unmerge_ksm_pages(struct vm_area_struct *vma,
810                              unsigned long start, unsigned long end)
811 {
812         unsigned long addr;
813         int err = 0;
814
815         for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
816                 if (ksm_test_exit(vma->vm_mm))
817                         break;
818                 if (signal_pending(current))
819                         err = -ERESTARTSYS;
820                 else
821                         err = break_ksm(vma, addr);
822         }
823         return err;
824 }
825
826 #ifdef CONFIG_SYSFS
827 /*
828  * Only called through the sysfs control interface:
829  */
830 static int remove_stable_node(struct stable_node *stable_node)
831 {
832         struct page *page;
833         int err;
834
835         page = get_ksm_page(stable_node, true);
836         if (!page) {
837                 /*
838                  * get_ksm_page did remove_node_from_stable_tree itself.
839                  */
840                 return 0;
841         }
842
843         if (WARN_ON_ONCE(page_mapped(page))) {
844                 /*
845                  * This should not happen: but if it does, just refuse to let
846                  * merge_across_nodes be switched - there is no need to panic.
847                  */
848                 err = -EBUSY;
849         } else {
850                 /*
851                  * The stable node did not yet appear stale to get_ksm_page(),
852                  * since that allows for an unmapped ksm page to be recognized
853                  * right up until it is freed; but the node is safe to remove.
854                  * This page might be in a pagevec waiting to be freed,
855                  * or it might be PageSwapCache (perhaps under writeback),
856                  * or it might have been removed from swapcache a moment ago.
857                  */
858                 set_page_stable_node(page, NULL);
859                 remove_node_from_stable_tree(stable_node);
860                 err = 0;
861         }
862
863         unlock_page(page);
864         put_page(page);
865         return err;
866 }
867
868 static int remove_stable_node_chain(struct stable_node *stable_node,
869                                     struct rb_root *root)
870 {
871         struct stable_node *dup;
872         struct hlist_node *hlist_safe;
873
874         if (!is_stable_node_chain(stable_node)) {
875                 VM_BUG_ON(is_stable_node_dup(stable_node));
876                 if (remove_stable_node(stable_node))
877                         return true;
878                 else
879                         return false;
880         }
881
882         hlist_for_each_entry_safe(dup, hlist_safe,
883                                   &stable_node->hlist, hlist_dup) {
884                 VM_BUG_ON(!is_stable_node_dup(dup));
885                 if (remove_stable_node(dup))
886                         return true;
887         }
888         BUG_ON(!hlist_empty(&stable_node->hlist));
889         free_stable_node_chain(stable_node, root);
890         return false;
891 }
892
893 static int remove_all_stable_nodes(void)
894 {
895         struct stable_node *stable_node, *next;
896         int nid;
897         int err = 0;
898
899         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
900                 while (root_stable_tree[nid].rb_node) {
901                         stable_node = rb_entry(root_stable_tree[nid].rb_node,
902                                                 struct stable_node, node);
903                         if (remove_stable_node_chain(stable_node,
904                                                      root_stable_tree + nid)) {
905                                 err = -EBUSY;
906                                 break;  /* proceed to next nid */
907                         }
908                         cond_resched();
909                 }
910         }
911         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
912                 if (remove_stable_node(stable_node))
913                         err = -EBUSY;
914                 cond_resched();
915         }
916         return err;
917 }
918
919 static int unmerge_and_remove_all_rmap_items(void)
920 {
921         struct mm_slot *mm_slot;
922         struct mm_struct *mm;
923         struct vm_area_struct *vma;
924         int err = 0;
925
926         spin_lock(&ksm_mmlist_lock);
927         ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
928                                                 struct mm_slot, mm_list);
929         spin_unlock(&ksm_mmlist_lock);
930
931         for (mm_slot = ksm_scan.mm_slot;
932                         mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
933                 mm = mm_slot->mm;
934                 down_read(&mm->mmap_sem);
935                 for (vma = mm->mmap; vma; vma = vma->vm_next) {
936                         if (ksm_test_exit(mm))
937                                 break;
938                         if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
939                                 continue;
940                         err = unmerge_ksm_pages(vma,
941                                                 vma->vm_start, vma->vm_end);
942                         if (err)
943                                 goto error;
944                 }
945
946                 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
947                 up_read(&mm->mmap_sem);
948
949                 spin_lock(&ksm_mmlist_lock);
950                 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
951                                                 struct mm_slot, mm_list);
952                 if (ksm_test_exit(mm)) {
953                         hash_del(&mm_slot->link);
954                         list_del(&mm_slot->mm_list);
955                         spin_unlock(&ksm_mmlist_lock);
956
957                         free_mm_slot(mm_slot);
958                         clear_bit(MMF_VM_MERGEABLE, &mm->flags);
959                         mmdrop(mm);
960                 } else
961                         spin_unlock(&ksm_mmlist_lock);
962         }
963
964         /* Clean up stable nodes, but don't worry if some are still busy */
965         remove_all_stable_nodes();
966         ksm_scan.seqnr = 0;
967         return 0;
968
969 error:
970         up_read(&mm->mmap_sem);
971         spin_lock(&ksm_mmlist_lock);
972         ksm_scan.mm_slot = &ksm_mm_head;
973         spin_unlock(&ksm_mmlist_lock);
974         return err;
975 }
976 #endif /* CONFIG_SYSFS */
977
978 static u32 calc_checksum(struct page *page)
979 {
980         u32 checksum;
981         void *addr = kmap_atomic(page);
982         checksum = jhash2(addr, PAGE_SIZE / 4, 17);
983         kunmap_atomic(addr);
984         return checksum;
985 }
986
987 static int memcmp_pages(struct page *page1, struct page *page2)
988 {
989         char *addr1, *addr2;
990         int ret;
991
992         addr1 = kmap_atomic(page1);
993         addr2 = kmap_atomic(page2);
994         ret = memcmp(addr1, addr2, PAGE_SIZE);
995         kunmap_atomic(addr2);
996         kunmap_atomic(addr1);
997         return ret;
998 }
999
1000 static inline int pages_identical(struct page *page1, struct page *page2)
1001 {
1002         return !memcmp_pages(page1, page2);
1003 }
1004
1005 static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1006                               pte_t *orig_pte)
1007 {
1008         struct mm_struct *mm = vma->vm_mm;
1009         struct page_vma_mapped_walk pvmw = {
1010                 .page = page,
1011                 .vma = vma,
1012         };
1013         int swapped;
1014         int err = -EFAULT;
1015         unsigned long mmun_start;       /* For mmu_notifiers */
1016         unsigned long mmun_end;         /* For mmu_notifiers */
1017
1018         pvmw.address = page_address_in_vma(page, vma);
1019         if (pvmw.address == -EFAULT)
1020                 goto out;
1021
1022         BUG_ON(PageTransCompound(page));
1023
1024         mmun_start = pvmw.address;
1025         mmun_end   = pvmw.address + PAGE_SIZE;
1026         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1027
1028         if (!page_vma_mapped_walk(&pvmw))
1029                 goto out_mn;
1030         if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1031                 goto out_unlock;
1032
1033         if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1034             (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1035                                                 mm_tlb_flush_pending(mm)) {
1036                 pte_t entry;
1037
1038                 swapped = PageSwapCache(page);
1039                 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
1040                 /*
1041                  * Ok this is tricky, when get_user_pages_fast() run it doesn't
1042                  * take any lock, therefore the check that we are going to make
1043                  * with the pagecount against the mapcount is racey and
1044                  * O_DIRECT can happen right after the check.
1045                  * So we clear the pte and flush the tlb before the check
1046                  * this assure us that no O_DIRECT can happen after the check
1047                  * or in the middle of the check.
1048                  *
1049                  * No need to notify as we are downgrading page table to read
1050                  * only not changing it to point to a new page.
1051                  *
1052                  * See Documentation/vm/mmu_notifier.txt
1053                  */
1054                 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
1055                 /*
1056                  * Check that no O_DIRECT or similar I/O is in progress on the
1057                  * page
1058                  */
1059                 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
1060                         set_pte_at(mm, pvmw.address, pvmw.pte, entry);
1061                         goto out_unlock;
1062                 }
1063                 if (pte_dirty(entry))
1064                         set_page_dirty(page);
1065
1066                 if (pte_protnone(entry))
1067                         entry = pte_mkclean(pte_clear_savedwrite(entry));
1068                 else
1069                         entry = pte_mkclean(pte_wrprotect(entry));
1070                 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
1071         }
1072         *orig_pte = *pvmw.pte;
1073         err = 0;
1074
1075 out_unlock:
1076         page_vma_mapped_walk_done(&pvmw);
1077 out_mn:
1078         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1079 out:
1080         return err;
1081 }
1082
1083 /**
1084  * replace_page - replace page in vma by new ksm page
1085  * @vma:      vma that holds the pte pointing to page
1086  * @page:     the page we are replacing by kpage
1087  * @kpage:    the ksm page we replace page by
1088  * @orig_pte: the original value of the pte
1089  *
1090  * Returns 0 on success, -EFAULT on failure.
1091  */
1092 static int replace_page(struct vm_area_struct *vma, struct page *page,
1093                         struct page *kpage, pte_t orig_pte)
1094 {
1095         struct mm_struct *mm = vma->vm_mm;
1096         pmd_t *pmd;
1097         pte_t *ptep;
1098         pte_t newpte;
1099         spinlock_t *ptl;
1100         unsigned long addr;
1101         int err = -EFAULT;
1102         unsigned long mmun_start;       /* For mmu_notifiers */
1103         unsigned long mmun_end;         /* For mmu_notifiers */
1104
1105         addr = page_address_in_vma(page, vma);
1106         if (addr == -EFAULT)
1107                 goto out;
1108
1109         pmd = mm_find_pmd(mm, addr);
1110         if (!pmd)
1111                 goto out;
1112
1113         mmun_start = addr;
1114         mmun_end   = addr + PAGE_SIZE;
1115         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1116
1117         ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1118         if (!pte_same(*ptep, orig_pte)) {
1119                 pte_unmap_unlock(ptep, ptl);
1120                 goto out_mn;
1121         }
1122
1123         /*
1124          * No need to check ksm_use_zero_pages here: we can only have a
1125          * zero_page here if ksm_use_zero_pages was enabled alreaady.
1126          */
1127         if (!is_zero_pfn(page_to_pfn(kpage))) {
1128                 get_page(kpage);
1129                 page_add_anon_rmap(kpage, vma, addr, false);
1130                 newpte = mk_pte(kpage, vma->vm_page_prot);
1131         } else {
1132                 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1133                                                vma->vm_page_prot));
1134         }
1135
1136         flush_cache_page(vma, addr, pte_pfn(*ptep));
1137         /*
1138          * No need to notify as we are replacing a read only page with another
1139          * read only page with the same content.
1140          *
1141          * See Documentation/vm/mmu_notifier.txt
1142          */
1143         ptep_clear_flush(vma, addr, ptep);
1144         set_pte_at_notify(mm, addr, ptep, newpte);
1145
1146         page_remove_rmap(page, false);
1147         if (!page_mapped(page))
1148                 try_to_free_swap(page);
1149         put_page(page);
1150
1151         pte_unmap_unlock(ptep, ptl);
1152         err = 0;
1153 out_mn:
1154         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1155 out:
1156         return err;
1157 }
1158
1159 /*
1160  * try_to_merge_one_page - take two pages and merge them into one
1161  * @vma: the vma that holds the pte pointing to page
1162  * @page: the PageAnon page that we want to replace with kpage
1163  * @kpage: the PageKsm page that we want to map instead of page,
1164  *         or NULL the first time when we want to use page as kpage.
1165  *
1166  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1167  */
1168 static int try_to_merge_one_page(struct vm_area_struct *vma,
1169                                  struct page *page, struct page *kpage)
1170 {
1171         pte_t orig_pte = __pte(0);
1172         int err = -EFAULT;
1173
1174         if (page == kpage)                      /* ksm page forked */
1175                 return 0;
1176
1177         if (!PageAnon(page))
1178                 goto out;
1179
1180         /*
1181          * We need the page lock to read a stable PageSwapCache in
1182          * write_protect_page().  We use trylock_page() instead of
1183          * lock_page() because we don't want to wait here - we
1184          * prefer to continue scanning and merging different pages,
1185          * then come back to this page when it is unlocked.
1186          */
1187         if (!trylock_page(page))
1188                 goto out;
1189
1190         if (PageTransCompound(page)) {
1191                 if (split_huge_page(page))
1192                         goto out_unlock;
1193         }
1194
1195         /*
1196          * If this anonymous page is mapped only here, its pte may need
1197          * to be write-protected.  If it's mapped elsewhere, all of its
1198          * ptes are necessarily already write-protected.  But in either
1199          * case, we need to lock and check page_count is not raised.
1200          */
1201         if (write_protect_page(vma, page, &orig_pte) == 0) {
1202                 if (!kpage) {
1203                         /*
1204                          * While we hold page lock, upgrade page from
1205                          * PageAnon+anon_vma to PageKsm+NULL stable_node:
1206                          * stable_tree_insert() will update stable_node.
1207                          */
1208                         set_page_stable_node(page, NULL);
1209                         mark_page_accessed(page);
1210                         /*
1211                          * Page reclaim just frees a clean page with no dirty
1212                          * ptes: make sure that the ksm page would be swapped.
1213                          */
1214                         if (!PageDirty(page))
1215                                 SetPageDirty(page);
1216                         err = 0;
1217                 } else if (pages_identical(page, kpage))
1218                         err = replace_page(vma, page, kpage, orig_pte);
1219         }
1220
1221         if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
1222                 munlock_vma_page(page);
1223                 if (!PageMlocked(kpage)) {
1224                         unlock_page(page);
1225                         lock_page(kpage);
1226                         mlock_vma_page(kpage);
1227                         page = kpage;           /* for final unlock */
1228                 }
1229         }
1230
1231 out_unlock:
1232         unlock_page(page);
1233 out:
1234         return err;
1235 }
1236
1237 /*
1238  * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1239  * but no new kernel page is allocated: kpage must already be a ksm page.
1240  *
1241  * This function returns 0 if the pages were merged, -EFAULT otherwise.
1242  */
1243 static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1244                                       struct page *page, struct page *kpage)
1245 {
1246         struct mm_struct *mm = rmap_item->mm;
1247         struct vm_area_struct *vma;
1248         int err = -EFAULT;
1249
1250         down_read(&mm->mmap_sem);
1251         vma = find_mergeable_vma(mm, rmap_item->address);
1252         if (!vma)
1253                 goto out;
1254
1255         err = try_to_merge_one_page(vma, page, kpage);
1256         if (err)
1257                 goto out;
1258
1259         /* Unstable nid is in union with stable anon_vma: remove first */
1260         remove_rmap_item_from_tree(rmap_item);
1261
1262         /* Must get reference to anon_vma while still holding mmap_sem */
1263         rmap_item->anon_vma = vma->anon_vma;
1264         get_anon_vma(vma->anon_vma);
1265 out:
1266         up_read(&mm->mmap_sem);
1267         return err;
1268 }
1269
1270 /*
1271  * try_to_merge_two_pages - take two identical pages and prepare them
1272  * to be merged into one page.
1273  *
1274  * This function returns the kpage if we successfully merged two identical
1275  * pages into one ksm page, NULL otherwise.
1276  *
1277  * Note that this function upgrades page to ksm page: if one of the pages
1278  * is already a ksm page, try_to_merge_with_ksm_page should be used.
1279  */
1280 static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1281                                            struct page *page,
1282                                            struct rmap_item *tree_rmap_item,
1283                                            struct page *tree_page)
1284 {
1285         int err;
1286
1287         err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
1288         if (!err) {
1289                 err = try_to_merge_with_ksm_page(tree_rmap_item,
1290                                                         tree_page, page);
1291                 /*
1292                  * If that fails, we have a ksm page with only one pte
1293                  * pointing to it: so break it.
1294                  */
1295                 if (err)
1296                         break_cow(rmap_item);
1297         }
1298         return err ? NULL : page;
1299 }
1300
1301 static __always_inline
1302 bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1303 {
1304         VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1305         /*
1306          * Check that at least one mapping still exists, otherwise
1307          * there's no much point to merge and share with this
1308          * stable_node, as the underlying tree_page of the other
1309          * sharer is going to be freed soon.
1310          */
1311         return stable_node->rmap_hlist_len &&
1312                 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1313 }
1314
1315 static __always_inline
1316 bool is_page_sharing_candidate(struct stable_node *stable_node)
1317 {
1318         return __is_page_sharing_candidate(stable_node, 0);
1319 }
1320
1321 static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1322                                     struct stable_node **_stable_node,
1323                                     struct rb_root *root,
1324                                     bool prune_stale_stable_nodes)
1325 {
1326         struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
1327         struct hlist_node *hlist_safe;
1328         struct page *_tree_page, *tree_page = NULL;
1329         int nr = 0;
1330         int found_rmap_hlist_len;
1331
1332         if (!prune_stale_stable_nodes ||
1333             time_before(jiffies, stable_node->chain_prune_time +
1334                         msecs_to_jiffies(
1335                                 ksm_stable_node_chains_prune_millisecs)))
1336                 prune_stale_stable_nodes = false;
1337         else
1338                 stable_node->chain_prune_time = jiffies;
1339
1340         hlist_for_each_entry_safe(dup, hlist_safe,
1341                                   &stable_node->hlist, hlist_dup) {
1342                 cond_resched();
1343                 /*
1344                  * We must walk all stable_node_dup to prune the stale
1345                  * stable nodes during lookup.
1346                  *
1347                  * get_ksm_page can drop the nodes from the
1348                  * stable_node->hlist if they point to freed pages
1349                  * (that's why we do a _safe walk). The "dup"
1350                  * stable_node parameter itself will be freed from
1351                  * under us if it returns NULL.
1352                  */
1353                 _tree_page = get_ksm_page(dup, false);
1354                 if (!_tree_page)
1355                         continue;
1356                 nr += 1;
1357                 if (is_page_sharing_candidate(dup)) {
1358                         if (!found ||
1359                             dup->rmap_hlist_len > found_rmap_hlist_len) {
1360                                 if (found)
1361                                         put_page(tree_page);
1362                                 found = dup;
1363                                 found_rmap_hlist_len = found->rmap_hlist_len;
1364                                 tree_page = _tree_page;
1365
1366                                 /* skip put_page for found dup */
1367                                 if (!prune_stale_stable_nodes)
1368                                         break;
1369                                 continue;
1370                         }
1371                 }
1372                 put_page(_tree_page);
1373         }
1374
1375         if (found) {
1376                 /*
1377                  * nr is counting all dups in the chain only if
1378                  * prune_stale_stable_nodes is true, otherwise we may
1379                  * break the loop at nr == 1 even if there are
1380                  * multiple entries.
1381                  */
1382                 if (prune_stale_stable_nodes && nr == 1) {
1383                         /*
1384                          * If there's not just one entry it would
1385                          * corrupt memory, better BUG_ON. In KSM
1386                          * context with no lock held it's not even
1387                          * fatal.
1388                          */
1389                         BUG_ON(stable_node->hlist.first->next);
1390
1391                         /*
1392                          * There's just one entry and it is below the
1393                          * deduplication limit so drop the chain.
1394                          */
1395                         rb_replace_node(&stable_node->node, &found->node,
1396                                         root);
1397                         free_stable_node(stable_node);
1398                         ksm_stable_node_chains--;
1399                         ksm_stable_node_dups--;
1400                         /*
1401                          * NOTE: the caller depends on the stable_node
1402                          * to be equal to stable_node_dup if the chain
1403                          * was collapsed.
1404                          */
1405                         *_stable_node = found;
1406                         /*
1407                          * Just for robustneess as stable_node is
1408                          * otherwise left as a stable pointer, the
1409                          * compiler shall optimize it away at build
1410                          * time.
1411                          */
1412                         stable_node = NULL;
1413                 } else if (stable_node->hlist.first != &found->hlist_dup &&
1414                            __is_page_sharing_candidate(found, 1)) {
1415                         /*
1416                          * If the found stable_node dup can accept one
1417                          * more future merge (in addition to the one
1418                          * that is underway) and is not at the head of
1419                          * the chain, put it there so next search will
1420                          * be quicker in the !prune_stale_stable_nodes
1421                          * case.
1422                          *
1423                          * NOTE: it would be inaccurate to use nr > 1
1424                          * instead of checking the hlist.first pointer
1425                          * directly, because in the
1426                          * prune_stale_stable_nodes case "nr" isn't
1427                          * the position of the found dup in the chain,
1428                          * but the total number of dups in the chain.
1429                          */
1430                         hlist_del(&found->hlist_dup);
1431                         hlist_add_head(&found->hlist_dup,
1432                                        &stable_node->hlist);
1433                 }
1434         }
1435
1436         *_stable_node_dup = found;
1437         return tree_page;
1438 }
1439
1440 static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1441                                                struct rb_root *root)
1442 {
1443         if (!is_stable_node_chain(stable_node))
1444                 return stable_node;
1445         if (hlist_empty(&stable_node->hlist)) {
1446                 free_stable_node_chain(stable_node, root);
1447                 return NULL;
1448         }
1449         return hlist_entry(stable_node->hlist.first,
1450                            typeof(*stable_node), hlist_dup);
1451 }
1452
1453 /*
1454  * Like for get_ksm_page, this function can free the *_stable_node and
1455  * *_stable_node_dup if the returned tree_page is NULL.
1456  *
1457  * It can also free and overwrite *_stable_node with the found
1458  * stable_node_dup if the chain is collapsed (in which case
1459  * *_stable_node will be equal to *_stable_node_dup like if the chain
1460  * never existed). It's up to the caller to verify tree_page is not
1461  * NULL before dereferencing *_stable_node or *_stable_node_dup.
1462  *
1463  * *_stable_node_dup is really a second output parameter of this
1464  * function and will be overwritten in all cases, the caller doesn't
1465  * need to initialize it.
1466  */
1467 static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1468                                         struct stable_node **_stable_node,
1469                                         struct rb_root *root,
1470                                         bool prune_stale_stable_nodes)
1471 {
1472         struct stable_node *stable_node = *_stable_node;
1473         if (!is_stable_node_chain(stable_node)) {
1474                 if (is_page_sharing_candidate(stable_node)) {
1475                         *_stable_node_dup = stable_node;
1476                         return get_ksm_page(stable_node, false);
1477                 }
1478                 /*
1479                  * _stable_node_dup set to NULL means the stable_node
1480                  * reached the ksm_max_page_sharing limit.
1481                  */
1482                 *_stable_node_dup = NULL;
1483                 return NULL;
1484         }
1485         return stable_node_dup(_stable_node_dup, _stable_node, root,
1486                                prune_stale_stable_nodes);
1487 }
1488
1489 static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1490                                                 struct stable_node **s_n,
1491                                                 struct rb_root *root)
1492 {
1493         return __stable_node_chain(s_n_d, s_n, root, true);
1494 }
1495
1496 static __always_inline struct page *chain(struct stable_node **s_n_d,
1497                                           struct stable_node *s_n,
1498                                           struct rb_root *root)
1499 {
1500         struct stable_node *old_stable_node = s_n;
1501         struct page *tree_page;
1502
1503         tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1504         /* not pruning dups so s_n cannot have changed */
1505         VM_BUG_ON(s_n != old_stable_node);
1506         return tree_page;
1507 }
1508
1509 /*
1510  * stable_tree_search - search for page inside the stable tree
1511  *
1512  * This function checks if there is a page inside the stable tree
1513  * with identical content to the page that we are scanning right now.
1514  *
1515  * This function returns the stable tree node of identical content if found,
1516  * NULL otherwise.
1517  */
1518 static struct page *stable_tree_search(struct page *page)
1519 {
1520         int nid;
1521         struct rb_root *root;
1522         struct rb_node **new;
1523         struct rb_node *parent;
1524         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1525         struct stable_node *page_node;
1526
1527         page_node = page_stable_node(page);
1528         if (page_node && page_node->head != &migrate_nodes) {
1529                 /* ksm page forked */
1530                 get_page(page);
1531                 return page;
1532         }
1533
1534         nid = get_kpfn_nid(page_to_pfn(page));
1535         root = root_stable_tree + nid;
1536 again:
1537         new = &root->rb_node;
1538         parent = NULL;
1539
1540         while (*new) {
1541                 struct page *tree_page;
1542                 int ret;
1543
1544                 cond_resched();
1545                 stable_node = rb_entry(*new, struct stable_node, node);
1546                 stable_node_any = NULL;
1547                 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
1548                 /*
1549                  * NOTE: stable_node may have been freed by
1550                  * chain_prune() if the returned stable_node_dup is
1551                  * not NULL. stable_node_dup may have been inserted in
1552                  * the rbtree instead as a regular stable_node (in
1553                  * order to collapse the stable_node chain if a single
1554                  * stable_node dup was found in it). In such case the
1555                  * stable_node is overwritten by the calleee to point
1556                  * to the stable_node_dup that was collapsed in the
1557                  * stable rbtree and stable_node will be equal to
1558                  * stable_node_dup like if the chain never existed.
1559                  */
1560                 if (!stable_node_dup) {
1561                         /*
1562                          * Either all stable_node dups were full in
1563                          * this stable_node chain, or this chain was
1564                          * empty and should be rb_erased.
1565                          */
1566                         stable_node_any = stable_node_dup_any(stable_node,
1567                                                               root);
1568                         if (!stable_node_any) {
1569                                 /* rb_erase just run */
1570                                 goto again;
1571                         }
1572                         /*
1573                          * Take any of the stable_node dups page of
1574                          * this stable_node chain to let the tree walk
1575                          * continue. All KSM pages belonging to the
1576                          * stable_node dups in a stable_node chain
1577                          * have the same content and they're
1578                          * wrprotected at all times. Any will work
1579                          * fine to continue the walk.
1580                          */
1581                         tree_page = get_ksm_page(stable_node_any, false);
1582                 }
1583                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1584                 if (!tree_page) {
1585                         /*
1586                          * If we walked over a stale stable_node,
1587                          * get_ksm_page() will call rb_erase() and it
1588                          * may rebalance the tree from under us. So
1589                          * restart the search from scratch. Returning
1590                          * NULL would be safe too, but we'd generate
1591                          * false negative insertions just because some
1592                          * stable_node was stale.
1593                          */
1594                         goto again;
1595                 }
1596
1597                 ret = memcmp_pages(page, tree_page);
1598                 put_page(tree_page);
1599
1600                 parent = *new;
1601                 if (ret < 0)
1602                         new = &parent->rb_left;
1603                 else if (ret > 0)
1604                         new = &parent->rb_right;
1605                 else {
1606                         if (page_node) {
1607                                 VM_BUG_ON(page_node->head != &migrate_nodes);
1608                                 /*
1609                                  * Test if the migrated page should be merged
1610                                  * into a stable node dup. If the mapcount is
1611                                  * 1 we can migrate it with another KSM page
1612                                  * without adding it to the chain.
1613                                  */
1614                                 if (page_mapcount(page) > 1)
1615                                         goto chain_append;
1616                         }
1617
1618                         if (!stable_node_dup) {
1619                                 /*
1620                                  * If the stable_node is a chain and
1621                                  * we got a payload match in memcmp
1622                                  * but we cannot merge the scanned
1623                                  * page in any of the existing
1624                                  * stable_node dups because they're
1625                                  * all full, we need to wait the
1626                                  * scanned page to find itself a match
1627                                  * in the unstable tree to create a
1628                                  * brand new KSM page to add later to
1629                                  * the dups of this stable_node.
1630                                  */
1631                                 return NULL;
1632                         }
1633
1634                         /*
1635                          * Lock and unlock the stable_node's page (which
1636                          * might already have been migrated) so that page
1637                          * migration is sure to notice its raised count.
1638                          * It would be more elegant to return stable_node
1639                          * than kpage, but that involves more changes.
1640                          */
1641                         tree_page = get_ksm_page(stable_node_dup, true);
1642                         if (unlikely(!tree_page))
1643                                 /*
1644                                  * The tree may have been rebalanced,
1645                                  * so re-evaluate parent and new.
1646                                  */
1647                                 goto again;
1648                         unlock_page(tree_page);
1649
1650                         if (get_kpfn_nid(stable_node_dup->kpfn) !=
1651                             NUMA(stable_node_dup->nid)) {
1652                                 put_page(tree_page);
1653                                 goto replace;
1654                         }
1655                         return tree_page;
1656                 }
1657         }
1658
1659         if (!page_node)
1660                 return NULL;
1661
1662         list_del(&page_node->list);
1663         DO_NUMA(page_node->nid = nid);
1664         rb_link_node(&page_node->node, parent, new);
1665         rb_insert_color(&page_node->node, root);
1666 out:
1667         if (is_page_sharing_candidate(page_node)) {
1668                 get_page(page);
1669                 return page;
1670         } else
1671                 return NULL;
1672
1673 replace:
1674         /*
1675          * If stable_node was a chain and chain_prune collapsed it,
1676          * stable_node has been updated to be the new regular
1677          * stable_node. A collapse of the chain is indistinguishable
1678          * from the case there was no chain in the stable
1679          * rbtree. Otherwise stable_node is the chain and
1680          * stable_node_dup is the dup to replace.
1681          */
1682         if (stable_node_dup == stable_node) {
1683                 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1684                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1685                 /* there is no chain */
1686                 if (page_node) {
1687                         VM_BUG_ON(page_node->head != &migrate_nodes);
1688                         list_del(&page_node->list);
1689                         DO_NUMA(page_node->nid = nid);
1690                         rb_replace_node(&stable_node_dup->node,
1691                                         &page_node->node,
1692                                         root);
1693                         if (is_page_sharing_candidate(page_node))
1694                                 get_page(page);
1695                         else
1696                                 page = NULL;
1697                 } else {
1698                         rb_erase(&stable_node_dup->node, root);
1699                         page = NULL;
1700                 }
1701         } else {
1702                 VM_BUG_ON(!is_stable_node_chain(stable_node));
1703                 __stable_node_dup_del(stable_node_dup);
1704                 if (page_node) {
1705                         VM_BUG_ON(page_node->head != &migrate_nodes);
1706                         list_del(&page_node->list);
1707                         DO_NUMA(page_node->nid = nid);
1708                         stable_node_chain_add_dup(page_node, stable_node);
1709                         if (is_page_sharing_candidate(page_node))
1710                                 get_page(page);
1711                         else
1712                                 page = NULL;
1713                 } else {
1714                         page = NULL;
1715                 }
1716         }
1717         stable_node_dup->head = &migrate_nodes;
1718         list_add(&stable_node_dup->list, stable_node_dup->head);
1719         return page;
1720
1721 chain_append:
1722         /* stable_node_dup could be null if it reached the limit */
1723         if (!stable_node_dup)
1724                 stable_node_dup = stable_node_any;
1725         /*
1726          * If stable_node was a chain and chain_prune collapsed it,
1727          * stable_node has been updated to be the new regular
1728          * stable_node. A collapse of the chain is indistinguishable
1729          * from the case there was no chain in the stable
1730          * rbtree. Otherwise stable_node is the chain and
1731          * stable_node_dup is the dup to replace.
1732          */
1733         if (stable_node_dup == stable_node) {
1734                 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1735                 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
1736                 /* chain is missing so create it */
1737                 stable_node = alloc_stable_node_chain(stable_node_dup,
1738                                                       root);
1739                 if (!stable_node)
1740                         return NULL;
1741         }
1742         /*
1743          * Add this stable_node dup that was
1744          * migrated to the stable_node chain
1745          * of the current nid for this page
1746          * content.
1747          */
1748         VM_BUG_ON(!is_stable_node_chain(stable_node));
1749         VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
1750         VM_BUG_ON(page_node->head != &migrate_nodes);
1751         list_del(&page_node->list);
1752         DO_NUMA(page_node->nid = nid);
1753         stable_node_chain_add_dup(page_node, stable_node);
1754         goto out;
1755 }
1756
1757 /*
1758  * stable_tree_insert - insert stable tree node pointing to new ksm page
1759  * into the stable tree.
1760  *
1761  * This function returns the stable tree node just allocated on success,
1762  * NULL otherwise.
1763  */
1764 static struct stable_node *stable_tree_insert(struct page *kpage)
1765 {
1766         int nid;
1767         unsigned long kpfn;
1768         struct rb_root *root;
1769         struct rb_node **new;
1770         struct rb_node *parent;
1771         struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1772         bool need_chain = false;
1773
1774         kpfn = page_to_pfn(kpage);
1775         nid = get_kpfn_nid(kpfn);
1776         root = root_stable_tree + nid;
1777 again:
1778         parent = NULL;
1779         new = &root->rb_node;
1780
1781         while (*new) {
1782                 struct page *tree_page;
1783                 int ret;
1784
1785                 cond_resched();
1786                 stable_node = rb_entry(*new, struct stable_node, node);
1787                 stable_node_any = NULL;
1788                 tree_page = chain(&stable_node_dup, stable_node, root);
1789                 if (!stable_node_dup) {
1790                         /*
1791                          * Either all stable_node dups were full in
1792                          * this stable_node chain, or this chain was
1793                          * empty and should be rb_erased.
1794                          */
1795                         stable_node_any = stable_node_dup_any(stable_node,
1796                                                               root);
1797                         if (!stable_node_any) {
1798                                 /* rb_erase just run */
1799                                 goto again;
1800                         }
1801                         /*
1802                          * Take any of the stable_node dups page of
1803                          * this stable_node chain to let the tree walk
1804                          * continue. All KSM pages belonging to the
1805                          * stable_node dups in a stable_node chain
1806                          * have the same content and they're
1807                          * wrprotected at all times. Any will work
1808                          * fine to continue the walk.
1809                          */
1810                         tree_page = get_ksm_page(stable_node_any, false);
1811                 }
1812                 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
1813                 if (!tree_page) {
1814                         /*
1815                          * If we walked over a stale stable_node,
1816                          * get_ksm_page() will call rb_erase() and it
1817                          * may rebalance the tree from under us. So
1818                          * restart the search from scratch. Returning
1819                          * NULL would be safe too, but we'd generate
1820                          * false negative insertions just because some
1821                          * stable_node was stale.
1822                          */
1823                         goto again;
1824                 }
1825
1826                 ret = memcmp_pages(kpage, tree_page);
1827                 put_page(tree_page);
1828
1829                 parent = *new;
1830                 if (ret < 0)
1831                         new = &parent->rb_left;
1832                 else if (ret > 0)
1833                         new = &parent->rb_right;
1834                 else {
1835                         need_chain = true;
1836                         break;
1837                 }
1838         }
1839
1840         stable_node_dup = alloc_stable_node();
1841         if (!stable_node_dup)
1842                 return NULL;
1843
1844         INIT_HLIST_HEAD(&stable_node_dup->hlist);
1845         stable_node_dup->kpfn = kpfn;
1846         set_page_stable_node(kpage, stable_node_dup);
1847         stable_node_dup->rmap_hlist_len = 0;
1848         DO_NUMA(stable_node_dup->nid = nid);
1849         if (!need_chain) {
1850                 rb_link_node(&stable_node_dup->node, parent, new);
1851                 rb_insert_color(&stable_node_dup->node, root);
1852         } else {
1853                 if (!is_stable_node_chain(stable_node)) {
1854                         struct stable_node *orig = stable_node;
1855                         /* chain is missing so create it */
1856                         stable_node = alloc_stable_node_chain(orig, root);
1857                         if (!stable_node) {
1858                                 free_stable_node(stable_node_dup);
1859                                 return NULL;
1860                         }
1861                 }
1862                 stable_node_chain_add_dup(stable_node_dup, stable_node);
1863         }
1864
1865         return stable_node_dup;
1866 }
1867
1868 /*
1869  * unstable_tree_search_insert - search for identical page,
1870  * else insert rmap_item into the unstable tree.
1871  *
1872  * This function searches for a page in the unstable tree identical to the
1873  * page currently being scanned; and if no identical page is found in the
1874  * tree, we insert rmap_item as a new object into the unstable tree.
1875  *
1876  * This function returns pointer to rmap_item found to be identical
1877  * to the currently scanned page, NULL otherwise.
1878  *
1879  * This function does both searching and inserting, because they share
1880  * the same walking algorithm in an rbtree.
1881  */
1882 static
1883 struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1884                                               struct page *page,
1885                                               struct page **tree_pagep)
1886 {
1887         struct rb_node **new;
1888         struct rb_root *root;
1889         struct rb_node *parent = NULL;
1890         int nid;
1891
1892         nid = get_kpfn_nid(page_to_pfn(page));
1893         root = root_unstable_tree + nid;
1894         new = &root->rb_node;
1895
1896         while (*new) {
1897                 struct rmap_item *tree_rmap_item;
1898                 struct page *tree_page;
1899                 int ret;
1900
1901                 cond_resched();
1902                 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
1903                 tree_page = get_mergeable_page(tree_rmap_item);
1904                 if (!tree_page)
1905                         return NULL;
1906
1907                 /*
1908                  * Don't substitute a ksm page for a forked page.
1909                  */
1910                 if (page == tree_page) {
1911                         put_page(tree_page);
1912                         return NULL;
1913                 }
1914
1915                 ret = memcmp_pages(page, tree_page);
1916
1917                 parent = *new;
1918                 if (ret < 0) {
1919                         put_page(tree_page);
1920                         new = &parent->rb_left;
1921                 } else if (ret > 0) {
1922                         put_page(tree_page);
1923                         new = &parent->rb_right;
1924                 } else if (!ksm_merge_across_nodes &&
1925                            page_to_nid(tree_page) != nid) {
1926                         /*
1927                          * If tree_page has been migrated to another NUMA node,
1928                          * it will be flushed out and put in the right unstable
1929                          * tree next time: only merge with it when across_nodes.
1930                          */
1931                         put_page(tree_page);
1932                         return NULL;
1933                 } else {
1934                         *tree_pagep = tree_page;
1935                         return tree_rmap_item;
1936                 }
1937         }
1938
1939         rmap_item->address |= UNSTABLE_FLAG;
1940         rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1941         DO_NUMA(rmap_item->nid = nid);
1942         rb_link_node(&rmap_item->node, parent, new);
1943         rb_insert_color(&rmap_item->node, root);
1944
1945         ksm_pages_unshared++;
1946         return NULL;
1947 }
1948
1949 /*
1950  * stable_tree_append - add another rmap_item to the linked list of
1951  * rmap_items hanging off a given node of the stable tree, all sharing
1952  * the same ksm page.
1953  */
1954 static void stable_tree_append(struct rmap_item *rmap_item,
1955                                struct stable_node *stable_node,
1956                                bool max_page_sharing_bypass)
1957 {
1958         /*
1959          * rmap won't find this mapping if we don't insert the
1960          * rmap_item in the right stable_node
1961          * duplicate. page_migration could break later if rmap breaks,
1962          * so we can as well crash here. We really need to check for
1963          * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
1964          * for other negative values as an undeflow if detected here
1965          * for the first time (and not when decreasing rmap_hlist_len)
1966          * would be sign of memory corruption in the stable_node.
1967          */
1968         BUG_ON(stable_node->rmap_hlist_len < 0);
1969
1970         stable_node->rmap_hlist_len++;
1971         if (!max_page_sharing_bypass)
1972                 /* possibly non fatal but unexpected overflow, only warn */
1973                 WARN_ON_ONCE(stable_node->rmap_hlist_len >
1974                              ksm_max_page_sharing);
1975
1976         rmap_item->head = stable_node;
1977         rmap_item->address |= STABLE_FLAG;
1978         hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
1979
1980         if (rmap_item->hlist.next)
1981                 ksm_pages_sharing++;
1982         else
1983                 ksm_pages_shared++;
1984 }
1985
1986 /*
1987  * cmp_and_merge_page - first see if page can be merged into the stable tree;
1988  * if not, compare checksum to previous and if it's the same, see if page can
1989  * be inserted into the unstable tree, or merged with a page already there and
1990  * both transferred to the stable tree.
1991  *
1992  * @page: the page that we are searching identical page to.
1993  * @rmap_item: the reverse mapping into the virtual address of this page
1994  */
1995 static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1996 {
1997         struct mm_struct *mm = rmap_item->mm;
1998         struct rmap_item *tree_rmap_item;
1999         struct page *tree_page = NULL;
2000         struct stable_node *stable_node;
2001         struct page *kpage;
2002         unsigned int checksum;
2003         int err;
2004         bool max_page_sharing_bypass = false;
2005
2006         stable_node = page_stable_node(page);
2007         if (stable_node) {
2008                 if (stable_node->head != &migrate_nodes &&
2009                     get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2010                     NUMA(stable_node->nid)) {
2011                         stable_node_dup_del(stable_node);
2012                         stable_node->head = &migrate_nodes;
2013                         list_add(&stable_node->list, stable_node->head);
2014                 }
2015                 if (stable_node->head != &migrate_nodes &&
2016                     rmap_item->head == stable_node)
2017                         return;
2018                 /*
2019                  * If it's a KSM fork, allow it to go over the sharing limit
2020                  * without warnings.
2021                  */
2022                 if (!is_page_sharing_candidate(stable_node))
2023                         max_page_sharing_bypass = true;
2024         }
2025
2026         /* We first start with searching the page inside the stable tree */
2027         kpage = stable_tree_search(page);
2028         if (kpage == page && rmap_item->head == stable_node) {
2029                 put_page(kpage);
2030                 return;
2031         }
2032
2033         remove_rmap_item_from_tree(rmap_item);
2034
2035         if (kpage) {
2036                 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
2037                 if (!err) {
2038                         /*
2039                          * The page was successfully merged:
2040                          * add its rmap_item to the stable tree.
2041                          */
2042                         lock_page(kpage);
2043                         stable_tree_append(rmap_item, page_stable_node(kpage),
2044                                            max_page_sharing_bypass);
2045                         unlock_page(kpage);
2046                 }
2047                 put_page(kpage);
2048                 return;
2049         }
2050
2051         /*
2052          * If the hash value of the page has changed from the last time
2053          * we calculated it, this page is changing frequently: therefore we
2054          * don't want to insert it in the unstable tree, and we don't want
2055          * to waste our time searching for something identical to it there.
2056          */
2057         checksum = calc_checksum(page);
2058         if (rmap_item->oldchecksum != checksum) {
2059                 rmap_item->oldchecksum = checksum;
2060                 return;
2061         }
2062
2063         /*
2064          * Same checksum as an empty page. We attempt to merge it with the
2065          * appropriate zero page if the user enabled this via sysfs.
2066          */
2067         if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2068                 struct vm_area_struct *vma;
2069
2070                 down_read(&mm->mmap_sem);
2071                 vma = find_mergeable_vma(mm, rmap_item->address);
2072                 err = try_to_merge_one_page(vma, page,
2073                                             ZERO_PAGE(rmap_item->address));
2074                 up_read(&mm->mmap_sem);
2075                 /*
2076                  * In case of failure, the page was not really empty, so we
2077                  * need to continue. Otherwise we're done.
2078                  */
2079                 if (!err)
2080                         return;
2081         }
2082         tree_rmap_item =
2083                 unstable_tree_search_insert(rmap_item, page, &tree_page);
2084         if (tree_rmap_item) {
2085                 kpage = try_to_merge_two_pages(rmap_item, page,
2086                                                 tree_rmap_item, tree_page);
2087                 put_page(tree_page);
2088                 if (kpage) {
2089                         /*
2090                          * The pages were successfully merged: insert new
2091                          * node in the stable tree and add both rmap_items.
2092                          */
2093                         lock_page(kpage);
2094                         stable_node = stable_tree_insert(kpage);
2095                         if (stable_node) {
2096                                 stable_tree_append(tree_rmap_item, stable_node,
2097                                                    false);
2098                                 stable_tree_append(rmap_item, stable_node,
2099                                                    false);
2100                         }
2101                         unlock_page(kpage);
2102
2103                         /*
2104                          * If we fail to insert the page into the stable tree,
2105                          * we will have 2 virtual addresses that are pointing
2106                          * to a ksm page left outside the stable tree,
2107                          * in which case we need to break_cow on both.
2108                          */
2109                         if (!stable_node) {
2110                                 break_cow(tree_rmap_item);
2111                                 break_cow(rmap_item);
2112                         }
2113                 }
2114         }
2115 }
2116
2117 static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
2118                                             struct rmap_item **rmap_list,
2119                                             unsigned long addr)
2120 {
2121         struct rmap_item *rmap_item;
2122
2123         while (*rmap_list) {
2124                 rmap_item = *rmap_list;
2125                 if ((rmap_item->address & PAGE_MASK) == addr)
2126                         return rmap_item;
2127                 if (rmap_item->address > addr)
2128                         break;
2129                 *rmap_list = rmap_item->rmap_list;
2130                 remove_rmap_item_from_tree(rmap_item);
2131                 free_rmap_item(rmap_item);
2132         }
2133
2134         rmap_item = alloc_rmap_item();
2135         if (rmap_item) {
2136                 /* It has already been zeroed */
2137                 rmap_item->mm = mm_slot->mm;
2138                 rmap_item->address = addr;
2139                 rmap_item->rmap_list = *rmap_list;
2140                 *rmap_list = rmap_item;
2141         }
2142         return rmap_item;
2143 }
2144
2145 static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2146 {
2147         struct mm_struct *mm;
2148         struct mm_slot *slot;
2149         struct vm_area_struct *vma;
2150         struct rmap_item *rmap_item;
2151         int nid;
2152
2153         if (list_empty(&ksm_mm_head.mm_list))
2154                 return NULL;
2155
2156         slot = ksm_scan.mm_slot;
2157         if (slot == &ksm_mm_head) {
2158                 /*
2159                  * A number of pages can hang around indefinitely on per-cpu
2160                  * pagevecs, raised page count preventing write_protect_page
2161                  * from merging them.  Though it doesn't really matter much,
2162                  * it is puzzling to see some stuck in pages_volatile until
2163                  * other activity jostles them out, and they also prevented
2164                  * LTP's KSM test from succeeding deterministically; so drain
2165                  * them here (here rather than on entry to ksm_do_scan(),
2166                  * so we don't IPI too often when pages_to_scan is set low).
2167                  */
2168                 lru_add_drain_all();
2169
2170                 /*
2171                  * Whereas stale stable_nodes on the stable_tree itself
2172                  * get pruned in the regular course of stable_tree_search(),
2173                  * those moved out to the migrate_nodes list can accumulate:
2174                  * so prune them once before each full scan.
2175                  */
2176                 if (!ksm_merge_across_nodes) {
2177                         struct stable_node *stable_node, *next;
2178                         struct page *page;
2179
2180                         list_for_each_entry_safe(stable_node, next,
2181                                                  &migrate_nodes, list) {
2182                                 page = get_ksm_page(stable_node, false);
2183                                 if (page)
2184                                         put_page(page);
2185                                 cond_resched();
2186                         }
2187                 }
2188
2189                 for (nid = 0; nid < ksm_nr_node_ids; nid++)
2190                         root_unstable_tree[nid] = RB_ROOT;
2191
2192                 spin_lock(&ksm_mmlist_lock);
2193                 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2194                 ksm_scan.mm_slot = slot;
2195                 spin_unlock(&ksm_mmlist_lock);
2196                 /*
2197                  * Although we tested list_empty() above, a racing __ksm_exit
2198                  * of the last mm on the list may have removed it since then.
2199                  */
2200                 if (slot == &ksm_mm_head)
2201                         return NULL;
2202 next_mm:
2203                 ksm_scan.address = 0;
2204                 ksm_scan.rmap_list = &slot->rmap_list;
2205         }
2206
2207         mm = slot->mm;
2208         down_read(&mm->mmap_sem);
2209         if (ksm_test_exit(mm))
2210                 vma = NULL;
2211         else
2212                 vma = find_vma(mm, ksm_scan.address);
2213
2214         for (; vma; vma = vma->vm_next) {
2215                 if (!(vma->vm_flags & VM_MERGEABLE))
2216                         continue;
2217                 if (ksm_scan.address < vma->vm_start)
2218                         ksm_scan.address = vma->vm_start;
2219                 if (!vma->anon_vma)
2220                         ksm_scan.address = vma->vm_end;
2221
2222                 while (ksm_scan.address < vma->vm_end) {
2223                         if (ksm_test_exit(mm))
2224                                 break;
2225                         *page = follow_page(vma, ksm_scan.address, FOLL_GET);
2226                         if (IS_ERR_OR_NULL(*page)) {
2227                                 ksm_scan.address += PAGE_SIZE;
2228                                 cond_resched();
2229                                 continue;
2230                         }
2231                         if (PageAnon(*page)) {
2232                                 flush_anon_page(vma, *page, ksm_scan.address);
2233                                 flush_dcache_page(*page);
2234                                 rmap_item = get_next_rmap_item(slot,
2235                                         ksm_scan.rmap_list, ksm_scan.address);
2236                                 if (rmap_item) {
2237                                         ksm_scan.rmap_list =
2238                                                         &rmap_item->rmap_list;
2239                                         ksm_scan.address += PAGE_SIZE;
2240                                 } else
2241                                         put_page(*page);
2242                                 up_read(&mm->mmap_sem);
2243                                 return rmap_item;
2244                         }
2245                         put_page(*page);
2246                         ksm_scan.address += PAGE_SIZE;
2247                         cond_resched();
2248                 }
2249         }
2250
2251         if (ksm_test_exit(mm)) {
2252                 ksm_scan.address = 0;
2253                 ksm_scan.rmap_list = &slot->rmap_list;
2254         }
2255         /*
2256          * Nuke all the rmap_items that are above this current rmap:
2257          * because there were no VM_MERGEABLE vmas with such addresses.
2258          */
2259         remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
2260
2261         spin_lock(&ksm_mmlist_lock);
2262         ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2263                                                 struct mm_slot, mm_list);
2264         if (ksm_scan.address == 0) {
2265                 /*
2266                  * We've completed a full scan of all vmas, holding mmap_sem
2267                  * throughout, and found no VM_MERGEABLE: so do the same as
2268                  * __ksm_exit does to remove this mm from all our lists now.
2269                  * This applies either when cleaning up after __ksm_exit
2270                  * (but beware: we can reach here even before __ksm_exit),
2271                  * or when all VM_MERGEABLE areas have been unmapped (and
2272                  * mmap_sem then protects against race with MADV_MERGEABLE).
2273                  */
2274                 hash_del(&slot->link);
2275                 list_del(&slot->mm_list);
2276                 spin_unlock(&ksm_mmlist_lock);
2277
2278                 free_mm_slot(slot);
2279                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2280                 up_read(&mm->mmap_sem);
2281                 mmdrop(mm);
2282         } else {
2283                 up_read(&mm->mmap_sem);
2284                 /*
2285                  * up_read(&mm->mmap_sem) first because after
2286                  * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2287                  * already have been freed under us by __ksm_exit()
2288                  * because the "mm_slot" is still hashed and
2289                  * ksm_scan.mm_slot doesn't point to it anymore.
2290                  */
2291                 spin_unlock(&ksm_mmlist_lock);
2292         }
2293
2294         /* Repeat until we've completed scanning the whole list */
2295         slot = ksm_scan.mm_slot;
2296         if (slot != &ksm_mm_head)
2297                 goto next_mm;
2298
2299         ksm_scan.seqnr++;
2300         return NULL;
2301 }
2302
2303 /**
2304  * ksm_do_scan  - the ksm scanner main worker function.
2305  * @scan_npages:  number of pages we want to scan before we return.
2306  */
2307 static void ksm_do_scan(unsigned int scan_npages)
2308 {
2309         struct rmap_item *rmap_item;
2310         struct page *uninitialized_var(page);
2311
2312         while (scan_npages-- && likely(!freezing(current))) {
2313                 cond_resched();
2314                 rmap_item = scan_get_next_rmap_item(&page);
2315                 if (!rmap_item)
2316                         return;
2317                 cmp_and_merge_page(page, rmap_item);
2318                 put_page(page);
2319         }
2320 }
2321
2322 static int ksmd_should_run(void)
2323 {
2324         return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2325 }
2326
2327 static int ksm_scan_thread(void *nothing)
2328 {
2329         set_freezable();
2330         set_user_nice(current, 5);
2331
2332         while (!kthread_should_stop()) {
2333                 mutex_lock(&ksm_thread_mutex);
2334                 wait_while_offlining();
2335                 if (ksmd_should_run())
2336                         ksm_do_scan(ksm_thread_pages_to_scan);
2337                 mutex_unlock(&ksm_thread_mutex);
2338
2339                 try_to_freeze();
2340
2341                 if (ksmd_should_run()) {
2342                         schedule_timeout_interruptible(
2343                                 msecs_to_jiffies(ksm_thread_sleep_millisecs));
2344                 } else {
2345                         wait_event_freezable(ksm_thread_wait,
2346                                 ksmd_should_run() || kthread_should_stop());
2347                 }
2348         }
2349         return 0;
2350 }
2351
2352 int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2353                 unsigned long end, int advice, unsigned long *vm_flags)
2354 {
2355         struct mm_struct *mm = vma->vm_mm;
2356         int err;
2357
2358         switch (advice) {
2359         case MADV_MERGEABLE:
2360                 /*
2361                  * Be somewhat over-protective for now!
2362                  */
2363                 if (*vm_flags & (VM_MERGEABLE | VM_SHARED  | VM_MAYSHARE   |
2364                                  VM_PFNMAP    | VM_IO      | VM_DONTEXPAND |
2365                                  VM_HUGETLB | VM_MIXEDMAP))
2366                         return 0;               /* just ignore the advice */
2367
2368 #ifdef VM_SAO
2369                 if (*vm_flags & VM_SAO)
2370                         return 0;
2371 #endif
2372 #ifdef VM_SPARC_ADI
2373                 if (*vm_flags & VM_SPARC_ADI)
2374                         return 0;
2375 #endif
2376
2377                 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2378                         err = __ksm_enter(mm);
2379                         if (err)
2380                                 return err;
2381                 }
2382
2383                 *vm_flags |= VM_MERGEABLE;
2384                 break;
2385
2386         case MADV_UNMERGEABLE:
2387                 if (!(*vm_flags & VM_MERGEABLE))
2388                         return 0;               /* just ignore the advice */
2389
2390                 if (vma->anon_vma) {
2391                         err = unmerge_ksm_pages(vma, start, end);
2392                         if (err)
2393                                 return err;
2394                 }
2395
2396                 *vm_flags &= ~VM_MERGEABLE;
2397                 break;
2398         }
2399
2400         return 0;
2401 }
2402
2403 int __ksm_enter(struct mm_struct *mm)
2404 {
2405         struct mm_slot *mm_slot;
2406         int needs_wakeup;
2407
2408         mm_slot = alloc_mm_slot();
2409         if (!mm_slot)
2410                 return -ENOMEM;
2411
2412         /* Check ksm_run too?  Would need tighter locking */
2413         needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2414
2415         spin_lock(&ksm_mmlist_lock);
2416         insert_to_mm_slots_hash(mm, mm_slot);
2417         /*
2418          * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2419          * insert just behind the scanning cursor, to let the area settle
2420          * down a little; when fork is followed by immediate exec, we don't
2421          * want ksmd to waste time setting up and tearing down an rmap_list.
2422          *
2423          * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2424          * scanning cursor, otherwise KSM pages in newly forked mms will be
2425          * missed: then we might as well insert at the end of the list.
2426          */
2427         if (ksm_run & KSM_RUN_UNMERGE)
2428                 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2429         else
2430                 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
2431         spin_unlock(&ksm_mmlist_lock);
2432
2433         set_bit(MMF_VM_MERGEABLE, &mm->flags);
2434         mmgrab(mm);
2435
2436         if (needs_wakeup)
2437                 wake_up_interruptible(&ksm_thread_wait);
2438
2439         return 0;
2440 }
2441
2442 void __ksm_exit(struct mm_struct *mm)
2443 {
2444         struct mm_slot *mm_slot;
2445         int easy_to_free = 0;
2446
2447         /*
2448          * This process is exiting: if it's straightforward (as is the
2449          * case when ksmd was never running), free mm_slot immediately.
2450          * But if it's at the cursor or has rmap_items linked to it, use
2451          * mmap_sem to synchronize with any break_cows before pagetables
2452          * are freed, and leave the mm_slot on the list for ksmd to free.
2453          * Beware: ksm may already have noticed it exiting and freed the slot.
2454          */
2455
2456         spin_lock(&ksm_mmlist_lock);
2457         mm_slot = get_mm_slot(mm);
2458         if (mm_slot && ksm_scan.mm_slot != mm_slot) {
2459                 if (!mm_slot->rmap_list) {
2460                         hash_del(&mm_slot->link);
2461                         list_del(&mm_slot->mm_list);
2462                         easy_to_free = 1;
2463                 } else {
2464                         list_move(&mm_slot->mm_list,
2465                                   &ksm_scan.mm_slot->mm_list);
2466                 }
2467         }
2468         spin_unlock(&ksm_mmlist_lock);
2469
2470         if (easy_to_free) {
2471                 free_mm_slot(mm_slot);
2472                 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2473                 mmdrop(mm);
2474         } else if (mm_slot) {
2475                 down_write(&mm->mmap_sem);
2476                 up_write(&mm->mmap_sem);
2477         }
2478 }
2479
2480 struct page *ksm_might_need_to_copy(struct page *page,
2481                         struct vm_area_struct *vma, unsigned long address)
2482 {
2483         struct anon_vma *anon_vma = page_anon_vma(page);
2484         struct page *new_page;
2485
2486         if (PageKsm(page)) {
2487                 if (page_stable_node(page) &&
2488                     !(ksm_run & KSM_RUN_UNMERGE))
2489                         return page;    /* no need to copy it */
2490         } else if (!anon_vma) {
2491                 return page;            /* no need to copy it */
2492         } else if (anon_vma->root == vma->anon_vma->root &&
2493                  page->index == linear_page_index(vma, address)) {
2494                 return page;            /* still no need to copy it */
2495         }
2496         if (!PageUptodate(page))
2497                 return page;            /* let do_swap_page report the error */
2498
2499         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2500         if (new_page) {
2501                 copy_user_highpage(new_page, page, address, vma);
2502
2503                 SetPageDirty(new_page);
2504                 __SetPageUptodate(new_page);
2505                 __SetPageLocked(new_page);
2506         }
2507
2508         return new_page;
2509 }
2510
2511 void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
2512 {
2513         struct stable_node *stable_node;
2514         struct rmap_item *rmap_item;
2515         int search_new_forks = 0;
2516
2517         VM_BUG_ON_PAGE(!PageKsm(page), page);
2518
2519         /*
2520          * Rely on the page lock to protect against concurrent modifications
2521          * to that page's node of the stable tree.
2522          */
2523         VM_BUG_ON_PAGE(!PageLocked(page), page);
2524
2525         stable_node = page_stable_node(page);
2526         if (!stable_node)
2527                 return;
2528 again:
2529         hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
2530                 struct anon_vma *anon_vma = rmap_item->anon_vma;
2531                 struct anon_vma_chain *vmac;
2532                 struct vm_area_struct *vma;
2533
2534                 cond_resched();
2535                 anon_vma_lock_read(anon_vma);
2536                 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2537                                                0, ULONG_MAX) {
2538                         cond_resched();
2539                         vma = vmac->vma;
2540                         if (rmap_item->address < vma->vm_start ||
2541                             rmap_item->address >= vma->vm_end)
2542                                 continue;
2543                         /*
2544                          * Initially we examine only the vma which covers this
2545                          * rmap_item; but later, if there is still work to do,
2546                          * we examine covering vmas in other mms: in case they
2547                          * were forked from the original since ksmd passed.
2548                          */
2549                         if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2550                                 continue;
2551
2552                         if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2553                                 continue;
2554
2555                         if (!rwc->rmap_one(page, vma,
2556                                         rmap_item->address, rwc->arg)) {
2557                                 anon_vma_unlock_read(anon_vma);
2558                                 return;
2559                         }
2560                         if (rwc->done && rwc->done(page)) {
2561                                 anon_vma_unlock_read(anon_vma);
2562                                 return;
2563                         }
2564                 }
2565                 anon_vma_unlock_read(anon_vma);
2566         }
2567         if (!search_new_forks++)
2568                 goto again;
2569 }
2570
2571 #ifdef CONFIG_MIGRATION
2572 void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2573 {
2574         struct stable_node *stable_node;
2575
2576         VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2577         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2578         VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
2579
2580         stable_node = page_stable_node(newpage);
2581         if (stable_node) {
2582                 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
2583                 stable_node->kpfn = page_to_pfn(newpage);
2584                 /*
2585                  * newpage->mapping was set in advance; now we need smp_wmb()
2586                  * to make sure that the new stable_node->kpfn is visible
2587                  * to get_ksm_page() before it can see that oldpage->mapping
2588                  * has gone stale (or that PageSwapCache has been cleared).
2589                  */
2590                 smp_wmb();
2591                 set_page_stable_node(oldpage, NULL);
2592         }
2593 }
2594 #endif /* CONFIG_MIGRATION */
2595
2596 #ifdef CONFIG_MEMORY_HOTREMOVE
2597 static void wait_while_offlining(void)
2598 {
2599         while (ksm_run & KSM_RUN_OFFLINE) {
2600                 mutex_unlock(&ksm_thread_mutex);
2601                 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2602                             TASK_UNINTERRUPTIBLE);
2603                 mutex_lock(&ksm_thread_mutex);
2604         }
2605 }
2606
2607 static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2608                                          unsigned long start_pfn,
2609                                          unsigned long end_pfn)
2610 {
2611         if (stable_node->kpfn >= start_pfn &&
2612             stable_node->kpfn < end_pfn) {
2613                 /*
2614                  * Don't get_ksm_page, page has already gone:
2615                  * which is why we keep kpfn instead of page*
2616                  */
2617                 remove_node_from_stable_tree(stable_node);
2618                 return true;
2619         }
2620         return false;
2621 }
2622
2623 static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2624                                            unsigned long start_pfn,
2625                                            unsigned long end_pfn,
2626                                            struct rb_root *root)
2627 {
2628         struct stable_node *dup;
2629         struct hlist_node *hlist_safe;
2630
2631         if (!is_stable_node_chain(stable_node)) {
2632                 VM_BUG_ON(is_stable_node_dup(stable_node));
2633                 return stable_node_dup_remove_range(stable_node, start_pfn,
2634                                                     end_pfn);
2635         }
2636
2637         hlist_for_each_entry_safe(dup, hlist_safe,
2638                                   &stable_node->hlist, hlist_dup) {
2639                 VM_BUG_ON(!is_stable_node_dup(dup));
2640                 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2641         }
2642         if (hlist_empty(&stable_node->hlist)) {
2643                 free_stable_node_chain(stable_node, root);
2644                 return true; /* notify caller that tree was rebalanced */
2645         } else
2646                 return false;
2647 }
2648
2649 static void ksm_check_stable_tree(unsigned long start_pfn,
2650                                   unsigned long end_pfn)
2651 {
2652         struct stable_node *stable_node, *next;
2653         struct rb_node *node;
2654         int nid;
2655
2656         for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2657                 node = rb_first(root_stable_tree + nid);
2658                 while (node) {
2659                         stable_node = rb_entry(node, struct stable_node, node);
2660                         if (stable_node_chain_remove_range(stable_node,
2661                                                            start_pfn, end_pfn,
2662                                                            root_stable_tree +
2663                                                            nid))
2664                                 node = rb_first(root_stable_tree + nid);
2665                         else
2666                                 node = rb_next(node);
2667                         cond_resched();
2668                 }
2669         }
2670         list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
2671                 if (stable_node->kpfn >= start_pfn &&
2672                     stable_node->kpfn < end_pfn)
2673                         remove_node_from_stable_tree(stable_node);
2674                 cond_resched();
2675         }
2676 }
2677
2678 static int ksm_memory_callback(struct notifier_block *self,
2679                                unsigned long action, void *arg)
2680 {
2681         struct memory_notify *mn = arg;
2682
2683         switch (action) {
2684         case MEM_GOING_OFFLINE:
2685                 /*
2686                  * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2687                  * and remove_all_stable_nodes() while memory is going offline:
2688                  * it is unsafe for them to touch the stable tree at this time.
2689                  * But unmerge_ksm_pages(), rmap lookups and other entry points
2690                  * which do not need the ksm_thread_mutex are all safe.
2691                  */
2692                 mutex_lock(&ksm_thread_mutex);
2693                 ksm_run |= KSM_RUN_OFFLINE;
2694                 mutex_unlock(&ksm_thread_mutex);
2695                 break;
2696
2697         case MEM_OFFLINE:
2698                 /*
2699                  * Most of the work is done by page migration; but there might
2700                  * be a few stable_nodes left over, still pointing to struct
2701                  * pages which have been offlined: prune those from the tree,
2702                  * otherwise get_ksm_page() might later try to access a
2703                  * non-existent struct page.
2704                  */
2705                 ksm_check_stable_tree(mn->start_pfn,
2706                                       mn->start_pfn + mn->nr_pages);
2707                 /* fallthrough */
2708
2709         case MEM_CANCEL_OFFLINE:
2710                 mutex_lock(&ksm_thread_mutex);
2711                 ksm_run &= ~KSM_RUN_OFFLINE;
2712                 mutex_unlock(&ksm_thread_mutex);
2713
2714                 smp_mb();       /* wake_up_bit advises this */
2715                 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
2716                 break;
2717         }
2718         return NOTIFY_OK;
2719 }
2720 #else
2721 static void wait_while_offlining(void)
2722 {
2723 }
2724 #endif /* CONFIG_MEMORY_HOTREMOVE */
2725
2726 #ifdef CONFIG_SYSFS
2727 /*
2728  * This all compiles without CONFIG_SYSFS, but is a waste of space.
2729  */
2730
2731 #define KSM_ATTR_RO(_name) \
2732         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2733 #define KSM_ATTR(_name) \
2734         static struct kobj_attribute _name##_attr = \
2735                 __ATTR(_name, 0644, _name##_show, _name##_store)
2736
2737 static ssize_t sleep_millisecs_show(struct kobject *kobj,
2738                                     struct kobj_attribute *attr, char *buf)
2739 {
2740         return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2741 }
2742
2743 static ssize_t sleep_millisecs_store(struct kobject *kobj,
2744                                      struct kobj_attribute *attr,
2745                                      const char *buf, size_t count)
2746 {
2747         unsigned long msecs;
2748         int err;
2749
2750         err = kstrtoul(buf, 10, &msecs);
2751         if (err || msecs > UINT_MAX)
2752                 return -EINVAL;
2753
2754         ksm_thread_sleep_millisecs = msecs;
2755
2756         return count;
2757 }
2758 KSM_ATTR(sleep_millisecs);
2759
2760 static ssize_t pages_to_scan_show(struct kobject *kobj,
2761                                   struct kobj_attribute *attr, char *buf)
2762 {
2763         return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2764 }
2765
2766 static ssize_t pages_to_scan_store(struct kobject *kobj,
2767                                    struct kobj_attribute *attr,
2768                                    const char *buf, size_t count)
2769 {
2770         int err;
2771         unsigned long nr_pages;
2772
2773         err = kstrtoul(buf, 10, &nr_pages);
2774         if (err || nr_pages > UINT_MAX)
2775                 return -EINVAL;
2776
2777         ksm_thread_pages_to_scan = nr_pages;
2778
2779         return count;
2780 }
2781 KSM_ATTR(pages_to_scan);
2782
2783 static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2784                         char *buf)
2785 {
2786         return sprintf(buf, "%lu\n", ksm_run);
2787 }
2788
2789 static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2790                          const char *buf, size_t count)
2791 {
2792         int err;
2793         unsigned long flags;
2794
2795         err = kstrtoul(buf, 10, &flags);
2796         if (err || flags > UINT_MAX)
2797                 return -EINVAL;
2798         if (flags > KSM_RUN_UNMERGE)
2799                 return -EINVAL;
2800
2801         /*
2802          * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2803          * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
2804          * breaking COW to free the pages_shared (but leaves mm_slots
2805          * on the list for when ksmd may be set running again).
2806          */
2807
2808         mutex_lock(&ksm_thread_mutex);
2809         wait_while_offlining();
2810         if (ksm_run != flags) {
2811                 ksm_run = flags;
2812                 if (flags & KSM_RUN_UNMERGE) {
2813                         set_current_oom_origin();
2814                         err = unmerge_and_remove_all_rmap_items();
2815                         clear_current_oom_origin();
2816                         if (err) {
2817                                 ksm_run = KSM_RUN_STOP;
2818                                 count = err;
2819                         }
2820                 }
2821         }
2822         mutex_unlock(&ksm_thread_mutex);
2823
2824         if (flags & KSM_RUN_MERGE)
2825                 wake_up_interruptible(&ksm_thread_wait);
2826
2827         return count;
2828 }
2829 KSM_ATTR(run);
2830
2831 #ifdef CONFIG_NUMA
2832 static ssize_t merge_across_nodes_show(struct kobject *kobj,
2833                                 struct kobj_attribute *attr, char *buf)
2834 {
2835         return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2836 }
2837
2838 static ssize_t merge_across_nodes_store(struct kobject *kobj,
2839                                    struct kobj_attribute *attr,
2840                                    const char *buf, size_t count)
2841 {
2842         int err;
2843         unsigned long knob;
2844
2845         err = kstrtoul(buf, 10, &knob);
2846         if (err)
2847                 return err;
2848         if (knob > 1)
2849                 return -EINVAL;
2850
2851         mutex_lock(&ksm_thread_mutex);
2852         wait_while_offlining();
2853         if (ksm_merge_across_nodes != knob) {
2854                 if (ksm_pages_shared || remove_all_stable_nodes())
2855                         err = -EBUSY;
2856                 else if (root_stable_tree == one_stable_tree) {
2857                         struct rb_root *buf;
2858                         /*
2859                          * This is the first time that we switch away from the
2860                          * default of merging across nodes: must now allocate
2861                          * a buffer to hold as many roots as may be needed.
2862                          * Allocate stable and unstable together:
2863                          * MAXSMP NODES_SHIFT 10 will use 16kB.
2864                          */
2865                         buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2866                                       GFP_KERNEL);
2867                         /* Let us assume that RB_ROOT is NULL is zero */
2868                         if (!buf)
2869                                 err = -ENOMEM;
2870                         else {
2871                                 root_stable_tree = buf;
2872                                 root_unstable_tree = buf + nr_node_ids;
2873                                 /* Stable tree is empty but not the unstable */
2874                                 root_unstable_tree[0] = one_unstable_tree[0];
2875                         }
2876                 }
2877                 if (!err) {
2878                         ksm_merge_across_nodes = knob;
2879                         ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2880                 }
2881         }
2882         mutex_unlock(&ksm_thread_mutex);
2883
2884         return err ? err : count;
2885 }
2886 KSM_ATTR(merge_across_nodes);
2887 #endif
2888
2889 static ssize_t use_zero_pages_show(struct kobject *kobj,
2890                                 struct kobj_attribute *attr, char *buf)
2891 {
2892         return sprintf(buf, "%u\n", ksm_use_zero_pages);
2893 }
2894 static ssize_t use_zero_pages_store(struct kobject *kobj,
2895                                    struct kobj_attribute *attr,
2896                                    const char *buf, size_t count)
2897 {
2898         int err;
2899         bool value;
2900
2901         err = kstrtobool(buf, &value);
2902         if (err)
2903                 return -EINVAL;
2904
2905         ksm_use_zero_pages = value;
2906
2907         return count;
2908 }
2909 KSM_ATTR(use_zero_pages);
2910
2911 static ssize_t max_page_sharing_show(struct kobject *kobj,
2912                                      struct kobj_attribute *attr, char *buf)
2913 {
2914         return sprintf(buf, "%u\n", ksm_max_page_sharing);
2915 }
2916
2917 static ssize_t max_page_sharing_store(struct kobject *kobj,
2918                                       struct kobj_attribute *attr,
2919                                       const char *buf, size_t count)
2920 {
2921         int err;
2922         int knob;
2923
2924         err = kstrtoint(buf, 10, &knob);
2925         if (err)
2926                 return err;
2927         /*
2928          * When a KSM page is created it is shared by 2 mappings. This
2929          * being a signed comparison, it implicitly verifies it's not
2930          * negative.
2931          */
2932         if (knob < 2)
2933                 return -EINVAL;
2934
2935         if (READ_ONCE(ksm_max_page_sharing) == knob)
2936                 return count;
2937
2938         mutex_lock(&ksm_thread_mutex);
2939         wait_while_offlining();
2940         if (ksm_max_page_sharing != knob) {
2941                 if (ksm_pages_shared || remove_all_stable_nodes())
2942                         err = -EBUSY;
2943                 else
2944                         ksm_max_page_sharing = knob;
2945         }
2946         mutex_unlock(&ksm_thread_mutex);
2947
2948         return err ? err : count;
2949 }
2950 KSM_ATTR(max_page_sharing);
2951
2952 static ssize_t pages_shared_show(struct kobject *kobj,
2953                                  struct kobj_attribute *attr, char *buf)
2954 {
2955         return sprintf(buf, "%lu\n", ksm_pages_shared);
2956 }
2957 KSM_ATTR_RO(pages_shared);
2958
2959 static ssize_t pages_sharing_show(struct kobject *kobj,
2960                                   struct kobj_attribute *attr, char *buf)
2961 {
2962         return sprintf(buf, "%lu\n", ksm_pages_sharing);
2963 }
2964 KSM_ATTR_RO(pages_sharing);
2965
2966 static ssize_t pages_unshared_show(struct kobject *kobj,
2967                                    struct kobj_attribute *attr, char *buf)
2968 {
2969         return sprintf(buf, "%lu\n", ksm_pages_unshared);
2970 }
2971 KSM_ATTR_RO(pages_unshared);
2972
2973 static ssize_t pages_volatile_show(struct kobject *kobj,
2974                                    struct kobj_attribute *attr, char *buf)
2975 {
2976         long ksm_pages_volatile;
2977
2978         ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2979                                 - ksm_pages_sharing - ksm_pages_unshared;
2980         /*
2981          * It was not worth any locking to calculate that statistic,
2982          * but it might therefore sometimes be negative: conceal that.
2983          */
2984         if (ksm_pages_volatile < 0)
2985                 ksm_pages_volatile = 0;
2986         return sprintf(buf, "%ld\n", ksm_pages_volatile);
2987 }
2988 KSM_ATTR_RO(pages_volatile);
2989
2990 static ssize_t stable_node_dups_show(struct kobject *kobj,
2991                                      struct kobj_attribute *attr, char *buf)
2992 {
2993         return sprintf(buf, "%lu\n", ksm_stable_node_dups);
2994 }
2995 KSM_ATTR_RO(stable_node_dups);
2996
2997 static ssize_t stable_node_chains_show(struct kobject *kobj,
2998                                        struct kobj_attribute *attr, char *buf)
2999 {
3000         return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3001 }
3002 KSM_ATTR_RO(stable_node_chains);
3003
3004 static ssize_t
3005 stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3006                                         struct kobj_attribute *attr,
3007                                         char *buf)
3008 {
3009         return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3010 }
3011
3012 static ssize_t
3013 stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3014                                          struct kobj_attribute *attr,
3015                                          const char *buf, size_t count)
3016 {
3017         unsigned long msecs;
3018         int err;
3019
3020         err = kstrtoul(buf, 10, &msecs);
3021         if (err || msecs > UINT_MAX)
3022                 return -EINVAL;
3023
3024         ksm_stable_node_chains_prune_millisecs = msecs;
3025
3026         return count;
3027 }
3028 KSM_ATTR(stable_node_chains_prune_millisecs);
3029
3030 static ssize_t full_scans_show(struct kobject *kobj,
3031                                struct kobj_attribute *attr, char *buf)
3032 {
3033         return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3034 }
3035 KSM_ATTR_RO(full_scans);
3036
3037 static struct attribute *ksm_attrs[] = {
3038         &sleep_millisecs_attr.attr,
3039         &pages_to_scan_attr.attr,
3040         &run_attr.attr,
3041         &pages_shared_attr.attr,
3042         &pages_sharing_attr.attr,
3043         &pages_unshared_attr.attr,
3044         &pages_volatile_attr.attr,
3045         &full_scans_attr.attr,
3046 #ifdef CONFIG_NUMA
3047         &merge_across_nodes_attr.attr,
3048 #endif
3049         &max_page_sharing_attr.attr,
3050         &stable_node_chains_attr.attr,
3051         &stable_node_dups_attr.attr,
3052         &stable_node_chains_prune_millisecs_attr.attr,
3053         &use_zero_pages_attr.attr,
3054         NULL,
3055 };
3056
3057 static const struct attribute_group ksm_attr_group = {
3058         .attrs = ksm_attrs,
3059         .name = "ksm",
3060 };
3061 #endif /* CONFIG_SYSFS */
3062
3063 static int __init ksm_init(void)
3064 {
3065         struct task_struct *ksm_thread;
3066         int err;
3067
3068         /* The correct value depends on page size and endianness */
3069         zero_checksum = calc_checksum(ZERO_PAGE(0));
3070         /* Default to false for backwards compatibility */
3071         ksm_use_zero_pages = false;
3072
3073         err = ksm_slab_init();
3074         if (err)
3075                 goto out;
3076
3077         ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3078         if (IS_ERR(ksm_thread)) {
3079                 pr_err("ksm: creating kthread failed\n");
3080                 err = PTR_ERR(ksm_thread);
3081                 goto out_free;
3082         }
3083
3084 #ifdef CONFIG_SYSFS
3085         err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3086         if (err) {
3087                 pr_err("ksm: register sysfs failed\n");
3088                 kthread_stop(ksm_thread);
3089                 goto out_free;
3090         }
3091 #else
3092         ksm_run = KSM_RUN_MERGE;        /* no way for user to start it */
3093
3094 #endif /* CONFIG_SYSFS */
3095
3096 #ifdef CONFIG_MEMORY_HOTREMOVE
3097         /* There is no significance to this priority 100 */
3098         hotplug_memory_notifier(ksm_memory_callback, 100);
3099 #endif
3100         return 0;
3101
3102 out_free:
3103         ksm_slab_free();
3104 out:
3105         return err;
3106 }
3107 subsys_initcall(ksm_init);