]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blame - mm/ksm.c
ksm: cleanup stable_node chain collapse case
[mirror_ubuntu-artful-kernel.git] / mm / ksm.c
CommitLineData
f8af4da3 1/*
31dbd01f
IE
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 *
36b2528d 7 * Copyright (C) 2008-2009 Red Hat, Inc.
31dbd01f
IE
8 * Authors:
9 * Izik Eidus
10 * Andrea Arcangeli
11 * Chris Wright
36b2528d 12 * Hugh Dickins
31dbd01f
IE
13 *
14 * This work is licensed under the terms of the GNU GPL, version 2.
f8af4da3
HD
15 */
16
17#include <linux/errno.h>
31dbd01f
IE
18#include <linux/mm.h>
19#include <linux/fs.h>
f8af4da3 20#include <linux/mman.h>
31dbd01f 21#include <linux/sched.h>
6e84f315 22#include <linux/sched/mm.h>
f7ccbae4 23#include <linux/sched/coredump.h>
31dbd01f
IE
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>
62b61f61 34#include <linux/memory.h>
31dbd01f 35#include <linux/mmu_notifier.h>
2c6854fd 36#include <linux/swap.h>
f8af4da3 37#include <linux/ksm.h>
4ca3a69b 38#include <linux/hashtable.h>
878aee7d 39#include <linux/freezer.h>
72788c38 40#include <linux/oom.h>
90bd6fd3 41#include <linux/numa.h>
f8af4da3 42
31dbd01f 43#include <asm/tlbflush.h>
73848b46 44#include "internal.h"
31dbd01f 45
e850dcf5
HD
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
31dbd01f
IE
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.)
8fdb3dbf
HD
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.
31dbd01f
IE
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
6514d511 101 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
31dbd01f
IE
102 * @mm: the mm that this information is valid for
103 */
104struct mm_slot {
105 struct hlist_node link;
106 struct list_head mm_list;
6514d511 107 struct rmap_item *rmap_list;
31dbd01f
IE
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
6514d511 115 * @rmap_list: link to the next rmap to be scanned in the rmap_list
31dbd01f
IE
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 */
120struct ksm_scan {
121 struct mm_slot *mm_slot;
122 unsigned long address;
6514d511 123 struct rmap_item **rmap_list;
31dbd01f
IE
124 unsigned long seqnr;
125};
126
7b6ba2c7
HD
127/**
128 * struct stable_node - node of the stable rbtree
129 * @node: rb node of this ksm page in the stable tree
4146d2d6 130 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
2c653d0e 131 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
4146d2d6 132 * @list: linked into migrate_nodes, pending placement in the proper node tree
7b6ba2c7 133 * @hlist: hlist head of rmap_items using this ksm page
4146d2d6 134 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
2c653d0e
AA
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
4146d2d6 137 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
7b6ba2c7
HD
138 */
139struct stable_node {
4146d2d6
HD
140 union {
141 struct rb_node node; /* when node of stable tree */
142 struct { /* when listed for migration */
143 struct list_head *head;
2c653d0e
AA
144 struct {
145 struct hlist_node hlist_dup;
146 struct list_head list;
147 };
4146d2d6
HD
148 };
149 };
7b6ba2c7 150 struct hlist_head hlist;
2c653d0e
AA
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;
4146d2d6
HD
162#ifdef CONFIG_NUMA
163 int nid;
164#endif
7b6ba2c7
HD
165};
166
31dbd01f
IE
167/**
168 * struct rmap_item - reverse mapping item for virtual addresses
6514d511 169 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
db114b83 170 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
bc56620b 171 * @nid: NUMA node id of unstable tree in which linked (may not match page)
31dbd01f
IE
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
7b6ba2c7
HD
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
31dbd01f
IE
178 */
179struct rmap_item {
6514d511 180 struct rmap_item *rmap_list;
bc56620b
HD
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 };
31dbd01f
IE
187 struct mm_struct *mm;
188 unsigned long address; /* + low bits used for flags below */
7b6ba2c7 189 unsigned int oldchecksum; /* when unstable */
31dbd01f 190 union {
7b6ba2c7
HD
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 };
31dbd01f
IE
196 };
197};
198
199#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
7b6ba2c7
HD
200#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
201#define STABLE_FLAG 0x200 /* is listed from the stable tree */
31dbd01f
IE
202
203/* The stable and unstable tree heads */
ef53d16c
HD
204static struct rb_root one_stable_tree[1] = { RB_ROOT };
205static struct rb_root one_unstable_tree[1] = { RB_ROOT };
206static struct rb_root *root_stable_tree = one_stable_tree;
207static struct rb_root *root_unstable_tree = one_unstable_tree;
31dbd01f 208
4146d2d6
HD
209/* Recently migrated nodes of stable tree, pending proper placement */
210static LIST_HEAD(migrate_nodes);
2c653d0e 211#define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
4146d2d6 212
4ca3a69b
SL
213#define MM_SLOTS_HASH_BITS 10
214static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
31dbd01f
IE
215
216static struct mm_slot ksm_mm_head = {
217 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
218};
219static struct ksm_scan ksm_scan = {
220 .mm_slot = &ksm_mm_head,
221};
222
223static struct kmem_cache *rmap_item_cache;
7b6ba2c7 224static struct kmem_cache *stable_node_cache;
31dbd01f
IE
225static struct kmem_cache *mm_slot_cache;
226
227/* The number of nodes in the stable tree */
b4028260 228static unsigned long ksm_pages_shared;
31dbd01f 229
e178dfde 230/* The number of page slots additionally sharing those nodes */
b4028260 231static unsigned long ksm_pages_sharing;
31dbd01f 232
473b0ce4
HD
233/* The number of nodes in the unstable tree */
234static unsigned long ksm_pages_unshared;
235
236/* The number of rmap_items in use: to calculate pages_volatile */
237static unsigned long ksm_rmap_items;
238
2c653d0e
AA
239/* The number of stable_node chains */
240static unsigned long ksm_stable_node_chains;
241
242/* The number of stable_node dups linked to the stable_node chains */
243static unsigned long ksm_stable_node_dups;
244
245/* Delay in pruning stale stable_node_dups in the stable_node_chains */
246static int ksm_stable_node_chains_prune_millisecs = 2000;
247
248/* Maximum number of page slots sharing a stable node */
249static int ksm_max_page_sharing = 256;
250
31dbd01f 251/* Number of pages ksmd should scan in one batch */
2c6854fd 252static unsigned int ksm_thread_pages_to_scan = 100;
31dbd01f
IE
253
254/* Milliseconds ksmd should sleep between batches */
2ffd8679 255static unsigned int ksm_thread_sleep_millisecs = 20;
31dbd01f 256
e86c59b1
CI
257/* Checksum of an empty (zeroed) page */
258static unsigned int zero_checksum __read_mostly;
259
260/* Whether to merge empty (zeroed) pages with actual zero pages */
261static bool ksm_use_zero_pages __read_mostly;
262
e850dcf5 263#ifdef CONFIG_NUMA
90bd6fd3
PH
264/* Zeroed when merging across nodes is not allowed */
265static unsigned int ksm_merge_across_nodes = 1;
ef53d16c 266static int ksm_nr_node_ids = 1;
e850dcf5
HD
267#else
268#define ksm_merge_across_nodes 1U
ef53d16c 269#define ksm_nr_node_ids 1
e850dcf5 270#endif
90bd6fd3 271
31dbd01f
IE
272#define KSM_RUN_STOP 0
273#define KSM_RUN_MERGE 1
274#define KSM_RUN_UNMERGE 2
ef4d43a8
HD
275#define KSM_RUN_OFFLINE 4
276static unsigned long ksm_run = KSM_RUN_STOP;
277static void wait_while_offlining(void);
31dbd01f
IE
278
279static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
280static DEFINE_MUTEX(ksm_thread_mutex);
281static 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
287static 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
7b6ba2c7
HD
293 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
294 if (!stable_node_cache)
295 goto out_free1;
296
31dbd01f
IE
297 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
298 if (!mm_slot_cache)
7b6ba2c7 299 goto out_free2;
31dbd01f
IE
300
301 return 0;
302
7b6ba2c7
HD
303out_free2:
304 kmem_cache_destroy(stable_node_cache);
305out_free1:
31dbd01f
IE
306 kmem_cache_destroy(rmap_item_cache);
307out:
308 return -ENOMEM;
309}
310
311static void __init ksm_slab_free(void)
312{
313 kmem_cache_destroy(mm_slot_cache);
7b6ba2c7 314 kmem_cache_destroy(stable_node_cache);
31dbd01f
IE
315 kmem_cache_destroy(rmap_item_cache);
316 mm_slot_cache = NULL;
317}
318
2c653d0e
AA
319static __always_inline bool is_stable_node_chain(struct stable_node *chain)
320{
321 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
322}
323
324static __always_inline bool is_stable_node_dup(struct stable_node *dup)
325{
326 return dup->head == STABLE_NODE_DUP_HEAD;
327}
328
329static 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
339static inline void __stable_node_dup_del(struct stable_node *dup)
340{
b4fecc67 341 VM_BUG_ON(!is_stable_node_dup(dup));
2c653d0e
AA
342 hlist_del(&dup->hlist_dup);
343 ksm_stable_node_dups--;
344}
345
346static 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
31dbd01f
IE
358static inline struct rmap_item *alloc_rmap_item(void)
359{
473b0ce4
HD
360 struct rmap_item *rmap_item;
361
5b398e41 362 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
363 __GFP_NORETRY | __GFP_NOWARN);
473b0ce4
HD
364 if (rmap_item)
365 ksm_rmap_items++;
366 return rmap_item;
31dbd01f
IE
367}
368
369static inline void free_rmap_item(struct rmap_item *rmap_item)
370{
473b0ce4 371 ksm_rmap_items--;
31dbd01f
IE
372 rmap_item->mm = NULL; /* debug safety */
373 kmem_cache_free(rmap_item_cache, rmap_item);
374}
375
7b6ba2c7
HD
376static inline struct stable_node *alloc_stable_node(void)
377{
6213055f 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);
7b6ba2c7
HD
384}
385
386static inline void free_stable_node(struct stable_node *stable_node)
387{
2c653d0e
AA
388 VM_BUG_ON(stable_node->rmap_hlist_len &&
389 !is_stable_node_chain(stable_node));
7b6ba2c7
HD
390 kmem_cache_free(stable_node_cache, stable_node);
391}
392
31dbd01f
IE
393static 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
400static inline void free_mm_slot(struct mm_slot *mm_slot)
401{
402 kmem_cache_free(mm_slot_cache, mm_slot);
403}
404
31dbd01f
IE
405static struct mm_slot *get_mm_slot(struct mm_struct *mm)
406{
4ca3a69b
SL
407 struct mm_slot *slot;
408
b67bfe0d 409 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
4ca3a69b
SL
410 if (slot->mm == mm)
411 return slot;
31dbd01f 412
31dbd01f
IE
413 return NULL;
414}
415
416static void insert_to_mm_slots_hash(struct mm_struct *mm,
417 struct mm_slot *mm_slot)
418{
31dbd01f 419 mm_slot->mm = mm;
4ca3a69b 420 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
31dbd01f
IE
421}
422
a913e182
HD
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 */
431static inline bool ksm_test_exit(struct mm_struct *mm)
432{
433 return atomic_read(&mm->mm_users) == 0;
434}
435
31dbd01f
IE
436/*
437 * We use break_ksm to break COW on a ksm page: it's a stripped down
438 *
d4edcf0d 439 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
31dbd01f
IE
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.
1b2ee126
DH
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.
31dbd01f 450 */
d952b791 451static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
31dbd01f
IE
452{
453 struct page *page;
d952b791 454 int ret = 0;
31dbd01f
IE
455
456 do {
457 cond_resched();
1b2ee126
DH
458 page = follow_page(vma, addr,
459 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
22eccdd7 460 if (IS_ERR_OR_NULL(page))
31dbd01f
IE
461 break;
462 if (PageKsm(page))
dcddffd4
KS
463 ret = handle_mm_fault(vma, addr,
464 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
31dbd01f
IE
465 else
466 ret = VM_FAULT_WRITE;
467 put_page(page);
33692f27 468 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
d952b791
HD
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;
31dbd01f
IE
498}
499
ef694222
BL
500static 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
8dd3557a 514static void break_cow(struct rmap_item *rmap_item)
31dbd01f 515{
8dd3557a
HD
516 struct mm_struct *mm = rmap_item->mm;
517 unsigned long addr = rmap_item->address;
31dbd01f
IE
518 struct vm_area_struct *vma;
519
4035c07a
HD
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 */
9e60109f 524 put_anon_vma(rmap_item->anon_vma);
4035c07a 525
81464e30 526 down_read(&mm->mmap_sem);
ef694222
BL
527 vma = find_mergeable_vma(mm, addr);
528 if (vma)
529 break_ksm(vma, addr);
31dbd01f
IE
530 up_read(&mm->mmap_sem);
531}
532
533static 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);
ef694222
BL
541 vma = find_mergeable_vma(mm, addr);
542 if (!vma)
31dbd01f
IE
543 goto out;
544
545 page = follow_page(vma, addr, FOLL_GET);
22eccdd7 546 if (IS_ERR_OR_NULL(page))
31dbd01f 547 goto out;
f765f540 548 if (PageAnon(page)) {
31dbd01f
IE
549 flush_anon_page(vma, page, addr);
550 flush_dcache_page(page);
551 } else {
552 put_page(page);
c8f95ed1
AA
553out:
554 page = NULL;
31dbd01f
IE
555 }
556 up_read(&mm->mmap_sem);
557 return page;
558}
559
90bd6fd3
PH
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 */
566static inline int get_kpfn_nid(unsigned long kpfn)
567{
d8fc16a8 568 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
90bd6fd3
PH
569}
570
2c653d0e
AA
571static 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
604static 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
4035c07a
HD
612static void remove_node_from_stable_tree(struct stable_node *stable_node)
613{
614 struct rmap_item *rmap_item;
4035c07a 615
2c653d0e
AA
616 /* check it's not STABLE_NODE_CHAIN or negative */
617 BUG_ON(stable_node->rmap_hlist_len < 0);
618
b67bfe0d 619 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
4035c07a
HD
620 if (rmap_item->hlist.next)
621 ksm_pages_sharing--;
622 else
623 ksm_pages_shared--;
2c653d0e
AA
624 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
625 stable_node->rmap_hlist_len--;
9e60109f 626 put_anon_vma(rmap_item->anon_vma);
4035c07a
HD
627 rmap_item->address &= PAGE_MASK;
628 cond_resched();
629 }
630
2c653d0e
AA
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
4146d2d6
HD
643 if (stable_node->head == &migrate_nodes)
644 list_del(&stable_node->list);
645 else
2c653d0e 646 stable_node_dup_del(stable_node);
4035c07a
HD
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.
c8d6553b 656 * But beware, the stable node's page might be being migrated.
4035c07a
HD
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.
4035c07a
HD
667 * is on its way to being freed; but it is an anomaly to bear in mind.
668 */
8fdb3dbf 669static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
4035c07a
HD
670{
671 struct page *page;
672 void *expected_mapping;
c8d6553b 673 unsigned long kpfn;
4035c07a 674
bda807d4
MK
675 expected_mapping = (void *)((unsigned long)stable_node |
676 PAGE_MAPPING_KSM);
c8d6553b 677again:
4db0c3c2 678 kpfn = READ_ONCE(stable_node->kpfn);
c8d6553b
HD
679 page = pfn_to_page(kpfn);
680
681 /*
682 * page is computed from kpfn, so on most architectures reading
683 * page->mapping is naturally ordered after reading node->kpfn,
684 * but on Alpha we need to be more careful.
685 */
686 smp_read_barrier_depends();
4db0c3c2 687 if (READ_ONCE(page->mapping) != expected_mapping)
4035c07a 688 goto stale;
c8d6553b
HD
689
690 /*
691 * We cannot do anything with the page while its refcount is 0.
692 * Usually 0 means free, or tail of a higher-order page: in which
693 * case this node is no longer referenced, and should be freed;
694 * however, it might mean that the page is under page_freeze_refs().
695 * The __remove_mapping() case is easy, again the node is now stale;
696 * but if page is swapcache in migrate_page_move_mapping(), it might
697 * still be our page, in which case it's essential to keep the node.
698 */
699 while (!get_page_unless_zero(page)) {
700 /*
701 * Another check for page->mapping != expected_mapping would
702 * work here too. We have chosen the !PageSwapCache test to
703 * optimize the common case, when the page is or is about to
704 * be freed: PageSwapCache is cleared (under spin_lock_irq)
705 * in the freeze_refs section of __remove_mapping(); but Anon
706 * page->mapping reset to NULL later, in free_pages_prepare().
707 */
708 if (!PageSwapCache(page))
709 goto stale;
710 cpu_relax();
711 }
712
4db0c3c2 713 if (READ_ONCE(page->mapping) != expected_mapping) {
4035c07a
HD
714 put_page(page);
715 goto stale;
716 }
c8d6553b 717
8fdb3dbf 718 if (lock_it) {
8aafa6a4 719 lock_page(page);
4db0c3c2 720 if (READ_ONCE(page->mapping) != expected_mapping) {
8aafa6a4
HD
721 unlock_page(page);
722 put_page(page);
723 goto stale;
724 }
725 }
4035c07a 726 return page;
c8d6553b 727
4035c07a 728stale:
c8d6553b
HD
729 /*
730 * We come here from above when page->mapping or !PageSwapCache
731 * suggests that the node is stale; but it might be under migration.
732 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
733 * before checking whether node->kpfn has been changed.
734 */
735 smp_rmb();
4db0c3c2 736 if (READ_ONCE(stable_node->kpfn) != kpfn)
c8d6553b 737 goto again;
4035c07a
HD
738 remove_node_from_stable_tree(stable_node);
739 return NULL;
740}
741
31dbd01f
IE
742/*
743 * Removing rmap_item from stable or unstable tree.
744 * This function will clean the information from the stable/unstable tree.
745 */
746static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
747{
7b6ba2c7
HD
748 if (rmap_item->address & STABLE_FLAG) {
749 struct stable_node *stable_node;
5ad64688 750 struct page *page;
31dbd01f 751
7b6ba2c7 752 stable_node = rmap_item->head;
8aafa6a4 753 page = get_ksm_page(stable_node, true);
4035c07a
HD
754 if (!page)
755 goto out;
5ad64688 756
7b6ba2c7 757 hlist_del(&rmap_item->hlist);
4035c07a
HD
758 unlock_page(page);
759 put_page(page);
08beca44 760
98666f8a 761 if (!hlist_empty(&stable_node->hlist))
4035c07a
HD
762 ksm_pages_sharing--;
763 else
7b6ba2c7 764 ksm_pages_shared--;
2c653d0e
AA
765 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
766 stable_node->rmap_hlist_len--;
31dbd01f 767
9e60109f 768 put_anon_vma(rmap_item->anon_vma);
93d17715 769 rmap_item->address &= PAGE_MASK;
31dbd01f 770
7b6ba2c7 771 } else if (rmap_item->address & UNSTABLE_FLAG) {
31dbd01f
IE
772 unsigned char age;
773 /*
9ba69294 774 * Usually ksmd can and must skip the rb_erase, because
31dbd01f 775 * root_unstable_tree was already reset to RB_ROOT.
9ba69294
HD
776 * But be careful when an mm is exiting: do the rb_erase
777 * if this rmap_item was inserted by this scan, rather
778 * than left over from before.
31dbd01f
IE
779 */
780 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
cd551f97 781 BUG_ON(age > 1);
31dbd01f 782 if (!age)
90bd6fd3 783 rb_erase(&rmap_item->node,
ef53d16c 784 root_unstable_tree + NUMA(rmap_item->nid));
473b0ce4 785 ksm_pages_unshared--;
93d17715 786 rmap_item->address &= PAGE_MASK;
31dbd01f 787 }
4035c07a 788out:
31dbd01f
IE
789 cond_resched(); /* we're called from many long loops */
790}
791
31dbd01f 792static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
6514d511 793 struct rmap_item **rmap_list)
31dbd01f 794{
6514d511
HD
795 while (*rmap_list) {
796 struct rmap_item *rmap_item = *rmap_list;
797 *rmap_list = rmap_item->rmap_list;
31dbd01f 798 remove_rmap_item_from_tree(rmap_item);
31dbd01f
IE
799 free_rmap_item(rmap_item);
800 }
801}
802
803/*
e850dcf5 804 * Though it's very tempting to unmerge rmap_items from stable tree rather
31dbd01f
IE
805 * than check every pte of a given vma, the locking doesn't quite work for
806 * that - an rmap_item is assigned to the stable tree after inserting ksm
807 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
808 * rmap_items from parent to child at fork time (so as not to waste time
809 * if exit comes before the next scan reaches it).
81464e30
HD
810 *
811 * Similarly, although we'd like to remove rmap_items (so updating counts
812 * and freeing memory) when unmerging an area, it's easier to leave that
813 * to the next pass of ksmd - consider, for example, how ksmd might be
814 * in cmp_and_merge_page on one of the rmap_items we would be removing.
31dbd01f 815 */
d952b791
HD
816static int unmerge_ksm_pages(struct vm_area_struct *vma,
817 unsigned long start, unsigned long end)
31dbd01f
IE
818{
819 unsigned long addr;
d952b791 820 int err = 0;
31dbd01f 821
d952b791 822 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
9ba69294
HD
823 if (ksm_test_exit(vma->vm_mm))
824 break;
d952b791
HD
825 if (signal_pending(current))
826 err = -ERESTARTSYS;
827 else
828 err = break_ksm(vma, addr);
829 }
830 return err;
31dbd01f
IE
831}
832
2ffd8679
HD
833#ifdef CONFIG_SYSFS
834/*
835 * Only called through the sysfs control interface:
836 */
cbf86cfe
HD
837static int remove_stable_node(struct stable_node *stable_node)
838{
839 struct page *page;
840 int err;
841
842 page = get_ksm_page(stable_node, true);
843 if (!page) {
844 /*
845 * get_ksm_page did remove_node_from_stable_tree itself.
846 */
847 return 0;
848 }
849
8fdb3dbf
HD
850 if (WARN_ON_ONCE(page_mapped(page))) {
851 /*
852 * This should not happen: but if it does, just refuse to let
853 * merge_across_nodes be switched - there is no need to panic.
854 */
cbf86cfe 855 err = -EBUSY;
8fdb3dbf 856 } else {
cbf86cfe 857 /*
8fdb3dbf
HD
858 * The stable node did not yet appear stale to get_ksm_page(),
859 * since that allows for an unmapped ksm page to be recognized
860 * right up until it is freed; but the node is safe to remove.
cbf86cfe
HD
861 * This page might be in a pagevec waiting to be freed,
862 * or it might be PageSwapCache (perhaps under writeback),
863 * or it might have been removed from swapcache a moment ago.
864 */
865 set_page_stable_node(page, NULL);
866 remove_node_from_stable_tree(stable_node);
867 err = 0;
868 }
869
870 unlock_page(page);
871 put_page(page);
872 return err;
873}
874
2c653d0e
AA
875static int remove_stable_node_chain(struct stable_node *stable_node,
876 struct rb_root *root)
877{
878 struct stable_node *dup;
879 struct hlist_node *hlist_safe;
880
881 if (!is_stable_node_chain(stable_node)) {
882 VM_BUG_ON(is_stable_node_dup(stable_node));
883 if (remove_stable_node(stable_node))
884 return true;
885 else
886 return false;
887 }
888
889 hlist_for_each_entry_safe(dup, hlist_safe,
890 &stable_node->hlist, hlist_dup) {
891 VM_BUG_ON(!is_stable_node_dup(dup));
892 if (remove_stable_node(dup))
893 return true;
894 }
895 BUG_ON(!hlist_empty(&stable_node->hlist));
896 free_stable_node_chain(stable_node, root);
897 return false;
898}
899
cbf86cfe
HD
900static int remove_all_stable_nodes(void)
901{
03640418 902 struct stable_node *stable_node, *next;
cbf86cfe
HD
903 int nid;
904 int err = 0;
905
ef53d16c 906 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
cbf86cfe
HD
907 while (root_stable_tree[nid].rb_node) {
908 stable_node = rb_entry(root_stable_tree[nid].rb_node,
909 struct stable_node, node);
2c653d0e
AA
910 if (remove_stable_node_chain(stable_node,
911 root_stable_tree + nid)) {
cbf86cfe
HD
912 err = -EBUSY;
913 break; /* proceed to next nid */
914 }
915 cond_resched();
916 }
917 }
03640418 918 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
4146d2d6
HD
919 if (remove_stable_node(stable_node))
920 err = -EBUSY;
921 cond_resched();
922 }
cbf86cfe
HD
923 return err;
924}
925
d952b791 926static int unmerge_and_remove_all_rmap_items(void)
31dbd01f
IE
927{
928 struct mm_slot *mm_slot;
929 struct mm_struct *mm;
930 struct vm_area_struct *vma;
d952b791
HD
931 int err = 0;
932
933 spin_lock(&ksm_mmlist_lock);
9ba69294 934 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
d952b791
HD
935 struct mm_slot, mm_list);
936 spin_unlock(&ksm_mmlist_lock);
31dbd01f 937
9ba69294
HD
938 for (mm_slot = ksm_scan.mm_slot;
939 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
31dbd01f
IE
940 mm = mm_slot->mm;
941 down_read(&mm->mmap_sem);
942 for (vma = mm->mmap; vma; vma = vma->vm_next) {
9ba69294
HD
943 if (ksm_test_exit(mm))
944 break;
31dbd01f
IE
945 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
946 continue;
d952b791
HD
947 err = unmerge_ksm_pages(vma,
948 vma->vm_start, vma->vm_end);
9ba69294
HD
949 if (err)
950 goto error;
31dbd01f 951 }
9ba69294 952
6514d511 953 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
7496fea9 954 up_read(&mm->mmap_sem);
d952b791
HD
955
956 spin_lock(&ksm_mmlist_lock);
9ba69294 957 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
d952b791 958 struct mm_slot, mm_list);
9ba69294 959 if (ksm_test_exit(mm)) {
4ca3a69b 960 hash_del(&mm_slot->link);
9ba69294
HD
961 list_del(&mm_slot->mm_list);
962 spin_unlock(&ksm_mmlist_lock);
963
964 free_mm_slot(mm_slot);
965 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
9ba69294 966 mmdrop(mm);
7496fea9 967 } else
9ba69294 968 spin_unlock(&ksm_mmlist_lock);
31dbd01f
IE
969 }
970
cbf86cfe
HD
971 /* Clean up stable nodes, but don't worry if some are still busy */
972 remove_all_stable_nodes();
d952b791 973 ksm_scan.seqnr = 0;
9ba69294
HD
974 return 0;
975
976error:
977 up_read(&mm->mmap_sem);
31dbd01f 978 spin_lock(&ksm_mmlist_lock);
d952b791 979 ksm_scan.mm_slot = &ksm_mm_head;
31dbd01f 980 spin_unlock(&ksm_mmlist_lock);
d952b791 981 return err;
31dbd01f 982}
2ffd8679 983#endif /* CONFIG_SYSFS */
31dbd01f 984
31dbd01f
IE
985static u32 calc_checksum(struct page *page)
986{
987 u32 checksum;
9b04c5fe 988 void *addr = kmap_atomic(page);
31dbd01f 989 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
9b04c5fe 990 kunmap_atomic(addr);
31dbd01f
IE
991 return checksum;
992}
993
994static int memcmp_pages(struct page *page1, struct page *page2)
995{
996 char *addr1, *addr2;
997 int ret;
998
9b04c5fe
CW
999 addr1 = kmap_atomic(page1);
1000 addr2 = kmap_atomic(page2);
31dbd01f 1001 ret = memcmp(addr1, addr2, PAGE_SIZE);
9b04c5fe
CW
1002 kunmap_atomic(addr2);
1003 kunmap_atomic(addr1);
31dbd01f
IE
1004 return ret;
1005}
1006
1007static inline int pages_identical(struct page *page1, struct page *page2)
1008{
1009 return !memcmp_pages(page1, page2);
1010}
1011
1012static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1013 pte_t *orig_pte)
1014{
1015 struct mm_struct *mm = vma->vm_mm;
36eaff33
KS
1016 struct page_vma_mapped_walk pvmw = {
1017 .page = page,
1018 .vma = vma,
1019 };
31dbd01f
IE
1020 int swapped;
1021 int err = -EFAULT;
6bdb913f
HE
1022 unsigned long mmun_start; /* For mmu_notifiers */
1023 unsigned long mmun_end; /* For mmu_notifiers */
31dbd01f 1024
36eaff33
KS
1025 pvmw.address = page_address_in_vma(page, vma);
1026 if (pvmw.address == -EFAULT)
31dbd01f
IE
1027 goto out;
1028
29ad768c 1029 BUG_ON(PageTransCompound(page));
6bdb913f 1030
36eaff33
KS
1031 mmun_start = pvmw.address;
1032 mmun_end = pvmw.address + PAGE_SIZE;
6bdb913f
HE
1033 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1034
36eaff33 1035 if (!page_vma_mapped_walk(&pvmw))
6bdb913f 1036 goto out_mn;
36eaff33
KS
1037 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1038 goto out_unlock;
31dbd01f 1039
595cd8f2
AK
1040 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
1041 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte))) {
31dbd01f
IE
1042 pte_t entry;
1043
1044 swapped = PageSwapCache(page);
36eaff33 1045 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
31dbd01f 1046 /*
25985edc 1047 * Ok this is tricky, when get_user_pages_fast() run it doesn't
31dbd01f
IE
1048 * take any lock, therefore the check that we are going to make
1049 * with the pagecount against the mapcount is racey and
1050 * O_DIRECT can happen right after the check.
1051 * So we clear the pte and flush the tlb before the check
1052 * this assure us that no O_DIRECT can happen after the check
1053 * or in the middle of the check.
1054 */
36eaff33 1055 entry = ptep_clear_flush_notify(vma, pvmw.address, pvmw.pte);
31dbd01f
IE
1056 /*
1057 * Check that no O_DIRECT or similar I/O is in progress on the
1058 * page
1059 */
31e855ea 1060 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
36eaff33 1061 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
31dbd01f
IE
1062 goto out_unlock;
1063 }
4e31635c
HD
1064 if (pte_dirty(entry))
1065 set_page_dirty(page);
595cd8f2
AK
1066
1067 if (pte_protnone(entry))
1068 entry = pte_mkclean(pte_clear_savedwrite(entry));
1069 else
1070 entry = pte_mkclean(pte_wrprotect(entry));
36eaff33 1071 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
31dbd01f 1072 }
36eaff33 1073 *orig_pte = *pvmw.pte;
31dbd01f
IE
1074 err = 0;
1075
1076out_unlock:
36eaff33 1077 page_vma_mapped_walk_done(&pvmw);
6bdb913f
HE
1078out_mn:
1079 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
31dbd01f
IE
1080out:
1081 return err;
1082}
1083
1084/**
1085 * replace_page - replace page in vma by new ksm page
8dd3557a
HD
1086 * @vma: vma that holds the pte pointing to page
1087 * @page: the page we are replacing by kpage
1088 * @kpage: the ksm page we replace page by
31dbd01f
IE
1089 * @orig_pte: the original value of the pte
1090 *
1091 * Returns 0 on success, -EFAULT on failure.
1092 */
8dd3557a
HD
1093static int replace_page(struct vm_area_struct *vma, struct page *page,
1094 struct page *kpage, pte_t orig_pte)
31dbd01f
IE
1095{
1096 struct mm_struct *mm = vma->vm_mm;
31dbd01f
IE
1097 pmd_t *pmd;
1098 pte_t *ptep;
e86c59b1 1099 pte_t newpte;
31dbd01f
IE
1100 spinlock_t *ptl;
1101 unsigned long addr;
31dbd01f 1102 int err = -EFAULT;
6bdb913f
HE
1103 unsigned long mmun_start; /* For mmu_notifiers */
1104 unsigned long mmun_end; /* For mmu_notifiers */
31dbd01f 1105
8dd3557a 1106 addr = page_address_in_vma(page, vma);
31dbd01f
IE
1107 if (addr == -EFAULT)
1108 goto out;
1109
6219049a
BL
1110 pmd = mm_find_pmd(mm, addr);
1111 if (!pmd)
31dbd01f 1112 goto out;
31dbd01f 1113
6bdb913f
HE
1114 mmun_start = addr;
1115 mmun_end = addr + PAGE_SIZE;
1116 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1117
31dbd01f
IE
1118 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1119 if (!pte_same(*ptep, orig_pte)) {
1120 pte_unmap_unlock(ptep, ptl);
6bdb913f 1121 goto out_mn;
31dbd01f
IE
1122 }
1123
e86c59b1
CI
1124 /*
1125 * No need to check ksm_use_zero_pages here: we can only have a
1126 * zero_page here if ksm_use_zero_pages was enabled alreaady.
1127 */
1128 if (!is_zero_pfn(page_to_pfn(kpage))) {
1129 get_page(kpage);
1130 page_add_anon_rmap(kpage, vma, addr, false);
1131 newpte = mk_pte(kpage, vma->vm_page_prot);
1132 } else {
1133 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1134 vma->vm_page_prot));
1135 }
31dbd01f
IE
1136
1137 flush_cache_page(vma, addr, pte_pfn(*ptep));
34ee645e 1138 ptep_clear_flush_notify(vma, addr, ptep);
e86c59b1 1139 set_pte_at_notify(mm, addr, ptep, newpte);
31dbd01f 1140
d281ee61 1141 page_remove_rmap(page, false);
ae52a2ad
HD
1142 if (!page_mapped(page))
1143 try_to_free_swap(page);
8dd3557a 1144 put_page(page);
31dbd01f
IE
1145
1146 pte_unmap_unlock(ptep, ptl);
1147 err = 0;
6bdb913f
HE
1148out_mn:
1149 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
31dbd01f
IE
1150out:
1151 return err;
1152}
1153
1154/*
1155 * try_to_merge_one_page - take two pages and merge them into one
8dd3557a
HD
1156 * @vma: the vma that holds the pte pointing to page
1157 * @page: the PageAnon page that we want to replace with kpage
80e14822
HD
1158 * @kpage: the PageKsm page that we want to map instead of page,
1159 * or NULL the first time when we want to use page as kpage.
31dbd01f
IE
1160 *
1161 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1162 */
1163static int try_to_merge_one_page(struct vm_area_struct *vma,
8dd3557a 1164 struct page *page, struct page *kpage)
31dbd01f
IE
1165{
1166 pte_t orig_pte = __pte(0);
1167 int err = -EFAULT;
1168
db114b83
HD
1169 if (page == kpage) /* ksm page forked */
1170 return 0;
1171
8dd3557a 1172 if (!PageAnon(page))
31dbd01f
IE
1173 goto out;
1174
31dbd01f
IE
1175 /*
1176 * We need the page lock to read a stable PageSwapCache in
1177 * write_protect_page(). We use trylock_page() instead of
1178 * lock_page() because we don't want to wait here - we
1179 * prefer to continue scanning and merging different pages,
1180 * then come back to this page when it is unlocked.
1181 */
8dd3557a 1182 if (!trylock_page(page))
31e855ea 1183 goto out;
f765f540
KS
1184
1185 if (PageTransCompound(page)) {
a7306c34 1186 if (split_huge_page(page))
f765f540
KS
1187 goto out_unlock;
1188 }
1189
31dbd01f
IE
1190 /*
1191 * If this anonymous page is mapped only here, its pte may need
1192 * to be write-protected. If it's mapped elsewhere, all of its
1193 * ptes are necessarily already write-protected. But in either
1194 * case, we need to lock and check page_count is not raised.
1195 */
80e14822
HD
1196 if (write_protect_page(vma, page, &orig_pte) == 0) {
1197 if (!kpage) {
1198 /*
1199 * While we hold page lock, upgrade page from
1200 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1201 * stable_tree_insert() will update stable_node.
1202 */
1203 set_page_stable_node(page, NULL);
1204 mark_page_accessed(page);
337ed7eb
MK
1205 /*
1206 * Page reclaim just frees a clean page with no dirty
1207 * ptes: make sure that the ksm page would be swapped.
1208 */
1209 if (!PageDirty(page))
1210 SetPageDirty(page);
80e14822
HD
1211 err = 0;
1212 } else if (pages_identical(page, kpage))
1213 err = replace_page(vma, page, kpage, orig_pte);
1214 }
31dbd01f 1215
80e14822 1216 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
73848b46 1217 munlock_vma_page(page);
5ad64688
HD
1218 if (!PageMlocked(kpage)) {
1219 unlock_page(page);
5ad64688
HD
1220 lock_page(kpage);
1221 mlock_vma_page(kpage);
1222 page = kpage; /* for final unlock */
1223 }
1224 }
73848b46 1225
f765f540 1226out_unlock:
8dd3557a 1227 unlock_page(page);
31dbd01f
IE
1228out:
1229 return err;
1230}
1231
81464e30
HD
1232/*
1233 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1234 * but no new kernel page is allocated: kpage must already be a ksm page.
8dd3557a
HD
1235 *
1236 * This function returns 0 if the pages were merged, -EFAULT otherwise.
81464e30 1237 */
8dd3557a
HD
1238static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1239 struct page *page, struct page *kpage)
81464e30 1240{
8dd3557a 1241 struct mm_struct *mm = rmap_item->mm;
81464e30
HD
1242 struct vm_area_struct *vma;
1243 int err = -EFAULT;
1244
8dd3557a 1245 down_read(&mm->mmap_sem);
85c6e8dd
AA
1246 vma = find_mergeable_vma(mm, rmap_item->address);
1247 if (!vma)
81464e30
HD
1248 goto out;
1249
8dd3557a 1250 err = try_to_merge_one_page(vma, page, kpage);
db114b83
HD
1251 if (err)
1252 goto out;
1253
bc56620b
HD
1254 /* Unstable nid is in union with stable anon_vma: remove first */
1255 remove_rmap_item_from_tree(rmap_item);
1256
db114b83 1257 /* Must get reference to anon_vma while still holding mmap_sem */
9e60109f
PZ
1258 rmap_item->anon_vma = vma->anon_vma;
1259 get_anon_vma(vma->anon_vma);
81464e30 1260out:
8dd3557a 1261 up_read(&mm->mmap_sem);
81464e30
HD
1262 return err;
1263}
1264
31dbd01f
IE
1265/*
1266 * try_to_merge_two_pages - take two identical pages and prepare them
1267 * to be merged into one page.
1268 *
8dd3557a
HD
1269 * This function returns the kpage if we successfully merged two identical
1270 * pages into one ksm page, NULL otherwise.
31dbd01f 1271 *
80e14822 1272 * Note that this function upgrades page to ksm page: if one of the pages
31dbd01f
IE
1273 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1274 */
8dd3557a
HD
1275static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1276 struct page *page,
1277 struct rmap_item *tree_rmap_item,
1278 struct page *tree_page)
31dbd01f 1279{
80e14822 1280 int err;
31dbd01f 1281
80e14822 1282 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
31dbd01f 1283 if (!err) {
8dd3557a 1284 err = try_to_merge_with_ksm_page(tree_rmap_item,
80e14822 1285 tree_page, page);
31dbd01f 1286 /*
81464e30
HD
1287 * If that fails, we have a ksm page with only one pte
1288 * pointing to it: so break it.
31dbd01f 1289 */
4035c07a 1290 if (err)
8dd3557a 1291 break_cow(rmap_item);
31dbd01f 1292 }
80e14822 1293 return err ? NULL : page;
31dbd01f
IE
1294}
1295
2c653d0e
AA
1296static __always_inline
1297bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1298{
1299 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1300 /*
1301 * Check that at least one mapping still exists, otherwise
1302 * there's no much point to merge and share with this
1303 * stable_node, as the underlying tree_page of the other
1304 * sharer is going to be freed soon.
1305 */
1306 return stable_node->rmap_hlist_len &&
1307 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1308}
1309
1310static __always_inline
1311bool is_page_sharing_candidate(struct stable_node *stable_node)
1312{
1313 return __is_page_sharing_candidate(stable_node, 0);
1314}
1315
b4fecc67 1316static struct stable_node *stable_node_dup(struct stable_node **_stable_node,
2c653d0e
AA
1317 struct page **tree_page,
1318 struct rb_root *root,
1319 bool prune_stale_stable_nodes)
1320{
b4fecc67 1321 struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
2c653d0e
AA
1322 struct hlist_node *hlist_safe;
1323 struct page *_tree_page;
1324 int nr = 0;
1325 int found_rmap_hlist_len;
1326
1327 if (!prune_stale_stable_nodes ||
1328 time_before(jiffies, stable_node->chain_prune_time +
1329 msecs_to_jiffies(
1330 ksm_stable_node_chains_prune_millisecs)))
1331 prune_stale_stable_nodes = false;
1332 else
1333 stable_node->chain_prune_time = jiffies;
1334
1335 hlist_for_each_entry_safe(dup, hlist_safe,
1336 &stable_node->hlist, hlist_dup) {
1337 cond_resched();
1338 /*
1339 * We must walk all stable_node_dup to prune the stale
1340 * stable nodes during lookup.
1341 *
1342 * get_ksm_page can drop the nodes from the
1343 * stable_node->hlist if they point to freed pages
1344 * (that's why we do a _safe walk). The "dup"
1345 * stable_node parameter itself will be freed from
1346 * under us if it returns NULL.
1347 */
1348 _tree_page = get_ksm_page(dup, false);
1349 if (!_tree_page)
1350 continue;
1351 nr += 1;
1352 if (is_page_sharing_candidate(dup)) {
1353 if (!found ||
1354 dup->rmap_hlist_len > found_rmap_hlist_len) {
1355 if (found)
1356 put_page(*tree_page);
1357 found = dup;
1358 found_rmap_hlist_len = found->rmap_hlist_len;
1359 *tree_page = _tree_page;
1360
1361 if (!prune_stale_stable_nodes)
1362 break;
1363 /* skip put_page */
1364 continue;
1365 }
1366 }
1367 put_page(_tree_page);
1368 }
1369
1370 /*
1371 * nr is relevant only if prune_stale_stable_nodes is true,
1372 * otherwise we may break the loop at nr == 1 even if there
1373 * are multiple entries.
1374 */
1375 if (prune_stale_stable_nodes && found) {
1376 if (nr == 1) {
1377 /*
1378 * If there's not just one entry it would
1379 * corrupt memory, better BUG_ON. In KSM
1380 * context with no lock held it's not even
1381 * fatal.
1382 */
1383 BUG_ON(stable_node->hlist.first->next);
1384
1385 /*
1386 * There's just one entry and it is below the
1387 * deduplication limit so drop the chain.
1388 */
1389 rb_replace_node(&stable_node->node, &found->node,
1390 root);
1391 free_stable_node(stable_node);
1392 ksm_stable_node_chains--;
1393 ksm_stable_node_dups--;
b4fecc67 1394 /*
0ba1d0f7
AA
1395 * NOTE: the caller depends on the stable_node
1396 * to be equal to stable_node_dup if the chain
1397 * was collapsed.
b4fecc67 1398 */
0ba1d0f7
AA
1399 *_stable_node = found;
1400 /*
1401 * Just for robustneess as stable_node is
1402 * otherwise left as a stable pointer, the
1403 * compiler shall optimize it away at build
1404 * time.
1405 */
1406 stable_node = NULL;
2c653d0e
AA
1407 } else if (__is_page_sharing_candidate(found, 1)) {
1408 /*
1409 * Refile our candidate at the head
1410 * after the prune if our candidate
1411 * can accept one more future sharing
1412 * in addition to the one underway.
1413 */
1414 hlist_del(&found->hlist_dup);
1415 hlist_add_head(&found->hlist_dup,
1416 &stable_node->hlist);
1417 }
1418 }
1419
1420 return found;
1421}
1422
1423static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1424 struct rb_root *root)
1425{
1426 if (!is_stable_node_chain(stable_node))
1427 return stable_node;
1428 if (hlist_empty(&stable_node->hlist)) {
1429 free_stable_node_chain(stable_node, root);
1430 return NULL;
1431 }
1432 return hlist_entry(stable_node->hlist.first,
1433 typeof(*stable_node), hlist_dup);
1434}
1435
b4fecc67 1436static struct stable_node *__stable_node_chain(struct stable_node **_stable_node,
2c653d0e
AA
1437 struct page **tree_page,
1438 struct rb_root *root,
1439 bool prune_stale_stable_nodes)
1440{
b4fecc67 1441 struct stable_node *stable_node = *_stable_node;
2c653d0e
AA
1442 if (!is_stable_node_chain(stable_node)) {
1443 if (is_page_sharing_candidate(stable_node)) {
1444 *tree_page = get_ksm_page(stable_node, false);
1445 return stable_node;
1446 }
1447 return NULL;
1448 }
b4fecc67 1449 return stable_node_dup(_stable_node, tree_page, root,
2c653d0e
AA
1450 prune_stale_stable_nodes);
1451}
1452
b4fecc67 1453static __always_inline struct stable_node *chain_prune(struct stable_node **s_n,
2c653d0e
AA
1454 struct page **t_p,
1455 struct rb_root *root)
1456{
1457 return __stable_node_chain(s_n, t_p, root, true);
1458}
1459
1460static __always_inline struct stable_node *chain(struct stable_node *s_n,
1461 struct page **t_p,
1462 struct rb_root *root)
1463{
b4fecc67 1464 return __stable_node_chain(&s_n, t_p, root, false);
2c653d0e
AA
1465}
1466
31dbd01f 1467/*
8dd3557a 1468 * stable_tree_search - search for page inside the stable tree
31dbd01f
IE
1469 *
1470 * This function checks if there is a page inside the stable tree
1471 * with identical content to the page that we are scanning right now.
1472 *
7b6ba2c7 1473 * This function returns the stable tree node of identical content if found,
31dbd01f
IE
1474 * NULL otherwise.
1475 */
62b61f61 1476static struct page *stable_tree_search(struct page *page)
31dbd01f 1477{
90bd6fd3 1478 int nid;
ef53d16c 1479 struct rb_root *root;
4146d2d6
HD
1480 struct rb_node **new;
1481 struct rb_node *parent;
2c653d0e 1482 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
4146d2d6 1483 struct stable_node *page_node;
31dbd01f 1484
4146d2d6
HD
1485 page_node = page_stable_node(page);
1486 if (page_node && page_node->head != &migrate_nodes) {
1487 /* ksm page forked */
08beca44 1488 get_page(page);
62b61f61 1489 return page;
08beca44
HD
1490 }
1491
90bd6fd3 1492 nid = get_kpfn_nid(page_to_pfn(page));
ef53d16c 1493 root = root_stable_tree + nid;
4146d2d6 1494again:
ef53d16c 1495 new = &root->rb_node;
4146d2d6 1496 parent = NULL;
90bd6fd3 1497
4146d2d6 1498 while (*new) {
4035c07a 1499 struct page *tree_page;
31dbd01f
IE
1500 int ret;
1501
08beca44 1502 cond_resched();
4146d2d6 1503 stable_node = rb_entry(*new, struct stable_node, node);
2c653d0e 1504 stable_node_any = NULL;
b4fecc67
AA
1505 stable_node_dup = chain_prune(&stable_node, &tree_page, root);
1506 /*
1507 * NOTE: stable_node may have been freed by
1508 * chain_prune() if the returned stable_node_dup is
1509 * not NULL. stable_node_dup may have been inserted in
1510 * the rbtree instead as a regular stable_node (in
1511 * order to collapse the stable_node chain if a single
0ba1d0f7
AA
1512 * stable_node dup was found in it). In such case the
1513 * stable_node is overwritten by the calleee to point
1514 * to the stable_node_dup that was collapsed in the
1515 * stable rbtree and stable_node will be equal to
1516 * stable_node_dup like if the chain never existed.
b4fecc67 1517 */
2c653d0e
AA
1518 if (!stable_node_dup) {
1519 /*
1520 * Either all stable_node dups were full in
1521 * this stable_node chain, or this chain was
1522 * empty and should be rb_erased.
1523 */
1524 stable_node_any = stable_node_dup_any(stable_node,
1525 root);
1526 if (!stable_node_any) {
1527 /* rb_erase just run */
1528 goto again;
1529 }
1530 /*
1531 * Take any of the stable_node dups page of
1532 * this stable_node chain to let the tree walk
1533 * continue. All KSM pages belonging to the
1534 * stable_node dups in a stable_node chain
1535 * have the same content and they're
1536 * wrprotected at all times. Any will work
1537 * fine to continue the walk.
1538 */
1539 tree_page = get_ksm_page(stable_node_any, false);
1540 }
1541 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
f2e5ff85
AA
1542 if (!tree_page) {
1543 /*
1544 * If we walked over a stale stable_node,
1545 * get_ksm_page() will call rb_erase() and it
1546 * may rebalance the tree from under us. So
1547 * restart the search from scratch. Returning
1548 * NULL would be safe too, but we'd generate
1549 * false negative insertions just because some
1550 * stable_node was stale.
1551 */
1552 goto again;
1553 }
31dbd01f 1554
4035c07a 1555 ret = memcmp_pages(page, tree_page);
c8d6553b 1556 put_page(tree_page);
31dbd01f 1557
4146d2d6 1558 parent = *new;
c8d6553b 1559 if (ret < 0)
4146d2d6 1560 new = &parent->rb_left;
c8d6553b 1561 else if (ret > 0)
4146d2d6 1562 new = &parent->rb_right;
c8d6553b 1563 else {
2c653d0e
AA
1564 if (page_node) {
1565 VM_BUG_ON(page_node->head != &migrate_nodes);
1566 /*
1567 * Test if the migrated page should be merged
1568 * into a stable node dup. If the mapcount is
1569 * 1 we can migrate it with another KSM page
1570 * without adding it to the chain.
1571 */
1572 if (page_mapcount(page) > 1)
1573 goto chain_append;
1574 }
1575
1576 if (!stable_node_dup) {
1577 /*
1578 * If the stable_node is a chain and
1579 * we got a payload match in memcmp
1580 * but we cannot merge the scanned
1581 * page in any of the existing
1582 * stable_node dups because they're
1583 * all full, we need to wait the
1584 * scanned page to find itself a match
1585 * in the unstable tree to create a
1586 * brand new KSM page to add later to
1587 * the dups of this stable_node.
1588 */
1589 return NULL;
1590 }
1591
c8d6553b
HD
1592 /*
1593 * Lock and unlock the stable_node's page (which
1594 * might already have been migrated) so that page
1595 * migration is sure to notice its raised count.
1596 * It would be more elegant to return stable_node
1597 * than kpage, but that involves more changes.
1598 */
2c653d0e
AA
1599 tree_page = get_ksm_page(stable_node_dup, true);
1600 if (unlikely(!tree_page))
1601 /*
1602 * The tree may have been rebalanced,
1603 * so re-evaluate parent and new.
1604 */
4146d2d6 1605 goto again;
2c653d0e
AA
1606 unlock_page(tree_page);
1607
1608 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1609 NUMA(stable_node_dup->nid)) {
1610 put_page(tree_page);
1611 goto replace;
1612 }
1613 return tree_page;
c8d6553b 1614 }
31dbd01f
IE
1615 }
1616
4146d2d6
HD
1617 if (!page_node)
1618 return NULL;
1619
1620 list_del(&page_node->list);
1621 DO_NUMA(page_node->nid = nid);
1622 rb_link_node(&page_node->node, parent, new);
ef53d16c 1623 rb_insert_color(&page_node->node, root);
2c653d0e
AA
1624out:
1625 if (is_page_sharing_candidate(page_node)) {
1626 get_page(page);
1627 return page;
1628 } else
1629 return NULL;
4146d2d6
HD
1630
1631replace:
b4fecc67
AA
1632 /*
1633 * If stable_node was a chain and chain_prune collapsed it,
0ba1d0f7
AA
1634 * stable_node has been updated to be the new regular
1635 * stable_node. A collapse of the chain is indistinguishable
1636 * from the case there was no chain in the stable
1637 * rbtree. Otherwise stable_node is the chain and
1638 * stable_node_dup is the dup to replace.
b4fecc67 1639 */
0ba1d0f7 1640 if (stable_node_dup == stable_node) {
b4fecc67
AA
1641 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1642 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
2c653d0e
AA
1643 /* there is no chain */
1644 if (page_node) {
1645 VM_BUG_ON(page_node->head != &migrate_nodes);
1646 list_del(&page_node->list);
1647 DO_NUMA(page_node->nid = nid);
b4fecc67
AA
1648 rb_replace_node(&stable_node_dup->node,
1649 &page_node->node,
2c653d0e
AA
1650 root);
1651 if (is_page_sharing_candidate(page_node))
1652 get_page(page);
1653 else
1654 page = NULL;
1655 } else {
b4fecc67 1656 rb_erase(&stable_node_dup->node, root);
2c653d0e
AA
1657 page = NULL;
1658 }
4146d2d6 1659 } else {
2c653d0e
AA
1660 VM_BUG_ON(!is_stable_node_chain(stable_node));
1661 __stable_node_dup_del(stable_node_dup);
1662 if (page_node) {
1663 VM_BUG_ON(page_node->head != &migrate_nodes);
1664 list_del(&page_node->list);
1665 DO_NUMA(page_node->nid = nid);
1666 stable_node_chain_add_dup(page_node, stable_node);
1667 if (is_page_sharing_candidate(page_node))
1668 get_page(page);
1669 else
1670 page = NULL;
1671 } else {
1672 page = NULL;
1673 }
4146d2d6 1674 }
2c653d0e
AA
1675 stable_node_dup->head = &migrate_nodes;
1676 list_add(&stable_node_dup->list, stable_node_dup->head);
4146d2d6 1677 return page;
2c653d0e
AA
1678
1679chain_append:
1680 /* stable_node_dup could be null if it reached the limit */
1681 if (!stable_node_dup)
1682 stable_node_dup = stable_node_any;
b4fecc67
AA
1683 /*
1684 * If stable_node was a chain and chain_prune collapsed it,
0ba1d0f7
AA
1685 * stable_node has been updated to be the new regular
1686 * stable_node. A collapse of the chain is indistinguishable
1687 * from the case there was no chain in the stable
1688 * rbtree. Otherwise stable_node is the chain and
1689 * stable_node_dup is the dup to replace.
b4fecc67 1690 */
0ba1d0f7 1691 if (stable_node_dup == stable_node) {
b4fecc67
AA
1692 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1693 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
2c653d0e
AA
1694 /* chain is missing so create it */
1695 stable_node = alloc_stable_node_chain(stable_node_dup,
1696 root);
1697 if (!stable_node)
1698 return NULL;
1699 }
1700 /*
1701 * Add this stable_node dup that was
1702 * migrated to the stable_node chain
1703 * of the current nid for this page
1704 * content.
1705 */
b4fecc67
AA
1706 VM_BUG_ON(!is_stable_node_chain(stable_node));
1707 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
2c653d0e
AA
1708 VM_BUG_ON(page_node->head != &migrate_nodes);
1709 list_del(&page_node->list);
1710 DO_NUMA(page_node->nid = nid);
1711 stable_node_chain_add_dup(page_node, stable_node);
1712 goto out;
31dbd01f
IE
1713}
1714
1715/*
e850dcf5 1716 * stable_tree_insert - insert stable tree node pointing to new ksm page
31dbd01f
IE
1717 * into the stable tree.
1718 *
7b6ba2c7
HD
1719 * This function returns the stable tree node just allocated on success,
1720 * NULL otherwise.
31dbd01f 1721 */
7b6ba2c7 1722static struct stable_node *stable_tree_insert(struct page *kpage)
31dbd01f 1723{
90bd6fd3
PH
1724 int nid;
1725 unsigned long kpfn;
ef53d16c 1726 struct rb_root *root;
90bd6fd3 1727 struct rb_node **new;
f2e5ff85 1728 struct rb_node *parent;
2c653d0e
AA
1729 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1730 bool need_chain = false;
31dbd01f 1731
90bd6fd3
PH
1732 kpfn = page_to_pfn(kpage);
1733 nid = get_kpfn_nid(kpfn);
ef53d16c 1734 root = root_stable_tree + nid;
f2e5ff85
AA
1735again:
1736 parent = NULL;
ef53d16c 1737 new = &root->rb_node;
90bd6fd3 1738
31dbd01f 1739 while (*new) {
4035c07a 1740 struct page *tree_page;
31dbd01f
IE
1741 int ret;
1742
08beca44 1743 cond_resched();
7b6ba2c7 1744 stable_node = rb_entry(*new, struct stable_node, node);
2c653d0e
AA
1745 stable_node_any = NULL;
1746 stable_node_dup = chain(stable_node, &tree_page, root);
1747 if (!stable_node_dup) {
1748 /*
1749 * Either all stable_node dups were full in
1750 * this stable_node chain, or this chain was
1751 * empty and should be rb_erased.
1752 */
1753 stable_node_any = stable_node_dup_any(stable_node,
1754 root);
1755 if (!stable_node_any) {
1756 /* rb_erase just run */
1757 goto again;
1758 }
1759 /*
1760 * Take any of the stable_node dups page of
1761 * this stable_node chain to let the tree walk
1762 * continue. All KSM pages belonging to the
1763 * stable_node dups in a stable_node chain
1764 * have the same content and they're
1765 * wrprotected at all times. Any will work
1766 * fine to continue the walk.
1767 */
1768 tree_page = get_ksm_page(stable_node_any, false);
1769 }
1770 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
f2e5ff85
AA
1771 if (!tree_page) {
1772 /*
1773 * If we walked over a stale stable_node,
1774 * get_ksm_page() will call rb_erase() and it
1775 * may rebalance the tree from under us. So
1776 * restart the search from scratch. Returning
1777 * NULL would be safe too, but we'd generate
1778 * false negative insertions just because some
1779 * stable_node was stale.
1780 */
1781 goto again;
1782 }
31dbd01f 1783
4035c07a
HD
1784 ret = memcmp_pages(kpage, tree_page);
1785 put_page(tree_page);
31dbd01f
IE
1786
1787 parent = *new;
1788 if (ret < 0)
1789 new = &parent->rb_left;
1790 else if (ret > 0)
1791 new = &parent->rb_right;
1792 else {
2c653d0e
AA
1793 need_chain = true;
1794 break;
31dbd01f
IE
1795 }
1796 }
1797
2c653d0e
AA
1798 stable_node_dup = alloc_stable_node();
1799 if (!stable_node_dup)
7b6ba2c7 1800 return NULL;
31dbd01f 1801
2c653d0e
AA
1802 INIT_HLIST_HEAD(&stable_node_dup->hlist);
1803 stable_node_dup->kpfn = kpfn;
1804 set_page_stable_node(kpage, stable_node_dup);
1805 stable_node_dup->rmap_hlist_len = 0;
1806 DO_NUMA(stable_node_dup->nid = nid);
1807 if (!need_chain) {
1808 rb_link_node(&stable_node_dup->node, parent, new);
1809 rb_insert_color(&stable_node_dup->node, root);
1810 } else {
1811 if (!is_stable_node_chain(stable_node)) {
1812 struct stable_node *orig = stable_node;
1813 /* chain is missing so create it */
1814 stable_node = alloc_stable_node_chain(orig, root);
1815 if (!stable_node) {
1816 free_stable_node(stable_node_dup);
1817 return NULL;
1818 }
1819 }
1820 stable_node_chain_add_dup(stable_node_dup, stable_node);
1821 }
08beca44 1822
2c653d0e 1823 return stable_node_dup;
31dbd01f
IE
1824}
1825
1826/*
8dd3557a
HD
1827 * unstable_tree_search_insert - search for identical page,
1828 * else insert rmap_item into the unstable tree.
31dbd01f
IE
1829 *
1830 * This function searches for a page in the unstable tree identical to the
1831 * page currently being scanned; and if no identical page is found in the
1832 * tree, we insert rmap_item as a new object into the unstable tree.
1833 *
1834 * This function returns pointer to rmap_item found to be identical
1835 * to the currently scanned page, NULL otherwise.
1836 *
1837 * This function does both searching and inserting, because they share
1838 * the same walking algorithm in an rbtree.
1839 */
8dd3557a
HD
1840static
1841struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1842 struct page *page,
1843 struct page **tree_pagep)
31dbd01f 1844{
90bd6fd3
PH
1845 struct rb_node **new;
1846 struct rb_root *root;
31dbd01f 1847 struct rb_node *parent = NULL;
90bd6fd3
PH
1848 int nid;
1849
1850 nid = get_kpfn_nid(page_to_pfn(page));
ef53d16c 1851 root = root_unstable_tree + nid;
90bd6fd3 1852 new = &root->rb_node;
31dbd01f
IE
1853
1854 while (*new) {
1855 struct rmap_item *tree_rmap_item;
8dd3557a 1856 struct page *tree_page;
31dbd01f
IE
1857 int ret;
1858
d178f27f 1859 cond_resched();
31dbd01f 1860 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
8dd3557a 1861 tree_page = get_mergeable_page(tree_rmap_item);
c8f95ed1 1862 if (!tree_page)
31dbd01f
IE
1863 return NULL;
1864
1865 /*
8dd3557a 1866 * Don't substitute a ksm page for a forked page.
31dbd01f 1867 */
8dd3557a
HD
1868 if (page == tree_page) {
1869 put_page(tree_page);
31dbd01f
IE
1870 return NULL;
1871 }
1872
8dd3557a 1873 ret = memcmp_pages(page, tree_page);
31dbd01f
IE
1874
1875 parent = *new;
1876 if (ret < 0) {
8dd3557a 1877 put_page(tree_page);
31dbd01f
IE
1878 new = &parent->rb_left;
1879 } else if (ret > 0) {
8dd3557a 1880 put_page(tree_page);
31dbd01f 1881 new = &parent->rb_right;
b599cbdf
HD
1882 } else if (!ksm_merge_across_nodes &&
1883 page_to_nid(tree_page) != nid) {
1884 /*
1885 * If tree_page has been migrated to another NUMA node,
1886 * it will be flushed out and put in the right unstable
1887 * tree next time: only merge with it when across_nodes.
1888 */
1889 put_page(tree_page);
1890 return NULL;
31dbd01f 1891 } else {
8dd3557a 1892 *tree_pagep = tree_page;
31dbd01f
IE
1893 return tree_rmap_item;
1894 }
1895 }
1896
7b6ba2c7 1897 rmap_item->address |= UNSTABLE_FLAG;
31dbd01f 1898 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
e850dcf5 1899 DO_NUMA(rmap_item->nid = nid);
31dbd01f 1900 rb_link_node(&rmap_item->node, parent, new);
90bd6fd3 1901 rb_insert_color(&rmap_item->node, root);
31dbd01f 1902
473b0ce4 1903 ksm_pages_unshared++;
31dbd01f
IE
1904 return NULL;
1905}
1906
1907/*
1908 * stable_tree_append - add another rmap_item to the linked list of
1909 * rmap_items hanging off a given node of the stable tree, all sharing
1910 * the same ksm page.
1911 */
1912static void stable_tree_append(struct rmap_item *rmap_item,
2c653d0e
AA
1913 struct stable_node *stable_node,
1914 bool max_page_sharing_bypass)
31dbd01f 1915{
2c653d0e
AA
1916 /*
1917 * rmap won't find this mapping if we don't insert the
1918 * rmap_item in the right stable_node
1919 * duplicate. page_migration could break later if rmap breaks,
1920 * so we can as well crash here. We really need to check for
1921 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
1922 * for other negative values as an undeflow if detected here
1923 * for the first time (and not when decreasing rmap_hlist_len)
1924 * would be sign of memory corruption in the stable_node.
1925 */
1926 BUG_ON(stable_node->rmap_hlist_len < 0);
1927
1928 stable_node->rmap_hlist_len++;
1929 if (!max_page_sharing_bypass)
1930 /* possibly non fatal but unexpected overflow, only warn */
1931 WARN_ON_ONCE(stable_node->rmap_hlist_len >
1932 ksm_max_page_sharing);
1933
7b6ba2c7 1934 rmap_item->head = stable_node;
31dbd01f 1935 rmap_item->address |= STABLE_FLAG;
7b6ba2c7 1936 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
e178dfde 1937
7b6ba2c7
HD
1938 if (rmap_item->hlist.next)
1939 ksm_pages_sharing++;
1940 else
1941 ksm_pages_shared++;
31dbd01f
IE
1942}
1943
1944/*
81464e30
HD
1945 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1946 * if not, compare checksum to previous and if it's the same, see if page can
1947 * be inserted into the unstable tree, or merged with a page already there and
1948 * both transferred to the stable tree.
31dbd01f
IE
1949 *
1950 * @page: the page that we are searching identical page to.
1951 * @rmap_item: the reverse mapping into the virtual address of this page
1952 */
1953static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1954{
31dbd01f 1955 struct rmap_item *tree_rmap_item;
8dd3557a 1956 struct page *tree_page = NULL;
7b6ba2c7 1957 struct stable_node *stable_node;
8dd3557a 1958 struct page *kpage;
31dbd01f
IE
1959 unsigned int checksum;
1960 int err;
2c653d0e 1961 bool max_page_sharing_bypass = false;
31dbd01f 1962
4146d2d6
HD
1963 stable_node = page_stable_node(page);
1964 if (stable_node) {
1965 if (stable_node->head != &migrate_nodes &&
2c653d0e
AA
1966 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
1967 NUMA(stable_node->nid)) {
1968 stable_node_dup_del(stable_node);
4146d2d6
HD
1969 stable_node->head = &migrate_nodes;
1970 list_add(&stable_node->list, stable_node->head);
1971 }
1972 if (stable_node->head != &migrate_nodes &&
1973 rmap_item->head == stable_node)
1974 return;
2c653d0e
AA
1975 /*
1976 * If it's a KSM fork, allow it to go over the sharing limit
1977 * without warnings.
1978 */
1979 if (!is_page_sharing_candidate(stable_node))
1980 max_page_sharing_bypass = true;
4146d2d6 1981 }
31dbd01f
IE
1982
1983 /* We first start with searching the page inside the stable tree */
62b61f61 1984 kpage = stable_tree_search(page);
4146d2d6
HD
1985 if (kpage == page && rmap_item->head == stable_node) {
1986 put_page(kpage);
1987 return;
1988 }
1989
1990 remove_rmap_item_from_tree(rmap_item);
1991
62b61f61 1992 if (kpage) {
08beca44 1993 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
31dbd01f
IE
1994 if (!err) {
1995 /*
1996 * The page was successfully merged:
1997 * add its rmap_item to the stable tree.
1998 */
5ad64688 1999 lock_page(kpage);
2c653d0e
AA
2000 stable_tree_append(rmap_item, page_stable_node(kpage),
2001 max_page_sharing_bypass);
5ad64688 2002 unlock_page(kpage);
31dbd01f 2003 }
8dd3557a 2004 put_page(kpage);
31dbd01f
IE
2005 return;
2006 }
2007
2008 /*
4035c07a
HD
2009 * If the hash value of the page has changed from the last time
2010 * we calculated it, this page is changing frequently: therefore we
2011 * don't want to insert it in the unstable tree, and we don't want
2012 * to waste our time searching for something identical to it there.
31dbd01f
IE
2013 */
2014 checksum = calc_checksum(page);
2015 if (rmap_item->oldchecksum != checksum) {
2016 rmap_item->oldchecksum = checksum;
2017 return;
2018 }
2019
e86c59b1
CI
2020 /*
2021 * Same checksum as an empty page. We attempt to merge it with the
2022 * appropriate zero page if the user enabled this via sysfs.
2023 */
2024 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2025 struct vm_area_struct *vma;
2026
2027 vma = find_mergeable_vma(rmap_item->mm, rmap_item->address);
2028 err = try_to_merge_one_page(vma, page,
2029 ZERO_PAGE(rmap_item->address));
2030 /*
2031 * In case of failure, the page was not really empty, so we
2032 * need to continue. Otherwise we're done.
2033 */
2034 if (!err)
2035 return;
2036 }
8dd3557a
HD
2037 tree_rmap_item =
2038 unstable_tree_search_insert(rmap_item, page, &tree_page);
31dbd01f 2039 if (tree_rmap_item) {
8dd3557a
HD
2040 kpage = try_to_merge_two_pages(rmap_item, page,
2041 tree_rmap_item, tree_page);
2042 put_page(tree_page);
8dd3557a 2043 if (kpage) {
bc56620b
HD
2044 /*
2045 * The pages were successfully merged: insert new
2046 * node in the stable tree and add both rmap_items.
2047 */
5ad64688 2048 lock_page(kpage);
7b6ba2c7
HD
2049 stable_node = stable_tree_insert(kpage);
2050 if (stable_node) {
2c653d0e
AA
2051 stable_tree_append(tree_rmap_item, stable_node,
2052 false);
2053 stable_tree_append(rmap_item, stable_node,
2054 false);
7b6ba2c7 2055 }
5ad64688 2056 unlock_page(kpage);
7b6ba2c7 2057
31dbd01f
IE
2058 /*
2059 * If we fail to insert the page into the stable tree,
2060 * we will have 2 virtual addresses that are pointing
2061 * to a ksm page left outside the stable tree,
2062 * in which case we need to break_cow on both.
2063 */
7b6ba2c7 2064 if (!stable_node) {
8dd3557a
HD
2065 break_cow(tree_rmap_item);
2066 break_cow(rmap_item);
31dbd01f
IE
2067 }
2068 }
31dbd01f
IE
2069 }
2070}
2071
2072static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
6514d511 2073 struct rmap_item **rmap_list,
31dbd01f
IE
2074 unsigned long addr)
2075{
2076 struct rmap_item *rmap_item;
2077
6514d511
HD
2078 while (*rmap_list) {
2079 rmap_item = *rmap_list;
93d17715 2080 if ((rmap_item->address & PAGE_MASK) == addr)
31dbd01f 2081 return rmap_item;
31dbd01f
IE
2082 if (rmap_item->address > addr)
2083 break;
6514d511 2084 *rmap_list = rmap_item->rmap_list;
31dbd01f 2085 remove_rmap_item_from_tree(rmap_item);
31dbd01f
IE
2086 free_rmap_item(rmap_item);
2087 }
2088
2089 rmap_item = alloc_rmap_item();
2090 if (rmap_item) {
2091 /* It has already been zeroed */
2092 rmap_item->mm = mm_slot->mm;
2093 rmap_item->address = addr;
6514d511
HD
2094 rmap_item->rmap_list = *rmap_list;
2095 *rmap_list = rmap_item;
31dbd01f
IE
2096 }
2097 return rmap_item;
2098}
2099
2100static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2101{
2102 struct mm_struct *mm;
2103 struct mm_slot *slot;
2104 struct vm_area_struct *vma;
2105 struct rmap_item *rmap_item;
90bd6fd3 2106 int nid;
31dbd01f
IE
2107
2108 if (list_empty(&ksm_mm_head.mm_list))
2109 return NULL;
2110
2111 slot = ksm_scan.mm_slot;
2112 if (slot == &ksm_mm_head) {
2919bfd0
HD
2113 /*
2114 * A number of pages can hang around indefinitely on per-cpu
2115 * pagevecs, raised page count preventing write_protect_page
2116 * from merging them. Though it doesn't really matter much,
2117 * it is puzzling to see some stuck in pages_volatile until
2118 * other activity jostles them out, and they also prevented
2119 * LTP's KSM test from succeeding deterministically; so drain
2120 * them here (here rather than on entry to ksm_do_scan(),
2121 * so we don't IPI too often when pages_to_scan is set low).
2122 */
2123 lru_add_drain_all();
2124
4146d2d6
HD
2125 /*
2126 * Whereas stale stable_nodes on the stable_tree itself
2127 * get pruned in the regular course of stable_tree_search(),
2128 * those moved out to the migrate_nodes list can accumulate:
2129 * so prune them once before each full scan.
2130 */
2131 if (!ksm_merge_across_nodes) {
03640418 2132 struct stable_node *stable_node, *next;
4146d2d6
HD
2133 struct page *page;
2134
03640418
GT
2135 list_for_each_entry_safe(stable_node, next,
2136 &migrate_nodes, list) {
4146d2d6
HD
2137 page = get_ksm_page(stable_node, false);
2138 if (page)
2139 put_page(page);
2140 cond_resched();
2141 }
2142 }
2143
ef53d16c 2144 for (nid = 0; nid < ksm_nr_node_ids; nid++)
90bd6fd3 2145 root_unstable_tree[nid] = RB_ROOT;
31dbd01f
IE
2146
2147 spin_lock(&ksm_mmlist_lock);
2148 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2149 ksm_scan.mm_slot = slot;
2150 spin_unlock(&ksm_mmlist_lock);
2b472611
HD
2151 /*
2152 * Although we tested list_empty() above, a racing __ksm_exit
2153 * of the last mm on the list may have removed it since then.
2154 */
2155 if (slot == &ksm_mm_head)
2156 return NULL;
31dbd01f
IE
2157next_mm:
2158 ksm_scan.address = 0;
6514d511 2159 ksm_scan.rmap_list = &slot->rmap_list;
31dbd01f
IE
2160 }
2161
2162 mm = slot->mm;
2163 down_read(&mm->mmap_sem);
9ba69294
HD
2164 if (ksm_test_exit(mm))
2165 vma = NULL;
2166 else
2167 vma = find_vma(mm, ksm_scan.address);
2168
2169 for (; vma; vma = vma->vm_next) {
31dbd01f
IE
2170 if (!(vma->vm_flags & VM_MERGEABLE))
2171 continue;
2172 if (ksm_scan.address < vma->vm_start)
2173 ksm_scan.address = vma->vm_start;
2174 if (!vma->anon_vma)
2175 ksm_scan.address = vma->vm_end;
2176
2177 while (ksm_scan.address < vma->vm_end) {
9ba69294
HD
2178 if (ksm_test_exit(mm))
2179 break;
31dbd01f 2180 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
21ae5b01
AA
2181 if (IS_ERR_OR_NULL(*page)) {
2182 ksm_scan.address += PAGE_SIZE;
2183 cond_resched();
2184 continue;
2185 }
f765f540 2186 if (PageAnon(*page)) {
31dbd01f
IE
2187 flush_anon_page(vma, *page, ksm_scan.address);
2188 flush_dcache_page(*page);
2189 rmap_item = get_next_rmap_item(slot,
6514d511 2190 ksm_scan.rmap_list, ksm_scan.address);
31dbd01f 2191 if (rmap_item) {
6514d511
HD
2192 ksm_scan.rmap_list =
2193 &rmap_item->rmap_list;
31dbd01f
IE
2194 ksm_scan.address += PAGE_SIZE;
2195 } else
2196 put_page(*page);
2197 up_read(&mm->mmap_sem);
2198 return rmap_item;
2199 }
21ae5b01 2200 put_page(*page);
31dbd01f
IE
2201 ksm_scan.address += PAGE_SIZE;
2202 cond_resched();
2203 }
2204 }
2205
9ba69294
HD
2206 if (ksm_test_exit(mm)) {
2207 ksm_scan.address = 0;
6514d511 2208 ksm_scan.rmap_list = &slot->rmap_list;
9ba69294 2209 }
31dbd01f
IE
2210 /*
2211 * Nuke all the rmap_items that are above this current rmap:
2212 * because there were no VM_MERGEABLE vmas with such addresses.
2213 */
6514d511 2214 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
31dbd01f
IE
2215
2216 spin_lock(&ksm_mmlist_lock);
cd551f97
HD
2217 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2218 struct mm_slot, mm_list);
2219 if (ksm_scan.address == 0) {
2220 /*
2221 * We've completed a full scan of all vmas, holding mmap_sem
2222 * throughout, and found no VM_MERGEABLE: so do the same as
2223 * __ksm_exit does to remove this mm from all our lists now.
9ba69294
HD
2224 * This applies either when cleaning up after __ksm_exit
2225 * (but beware: we can reach here even before __ksm_exit),
2226 * or when all VM_MERGEABLE areas have been unmapped (and
2227 * mmap_sem then protects against race with MADV_MERGEABLE).
cd551f97 2228 */
4ca3a69b 2229 hash_del(&slot->link);
cd551f97 2230 list_del(&slot->mm_list);
9ba69294
HD
2231 spin_unlock(&ksm_mmlist_lock);
2232
cd551f97
HD
2233 free_mm_slot(slot);
2234 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
9ba69294
HD
2235 up_read(&mm->mmap_sem);
2236 mmdrop(mm);
2237 } else {
9ba69294 2238 up_read(&mm->mmap_sem);
7496fea9
ZC
2239 /*
2240 * up_read(&mm->mmap_sem) first because after
2241 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2242 * already have been freed under us by __ksm_exit()
2243 * because the "mm_slot" is still hashed and
2244 * ksm_scan.mm_slot doesn't point to it anymore.
2245 */
2246 spin_unlock(&ksm_mmlist_lock);
cd551f97 2247 }
31dbd01f
IE
2248
2249 /* Repeat until we've completed scanning the whole list */
cd551f97 2250 slot = ksm_scan.mm_slot;
31dbd01f
IE
2251 if (slot != &ksm_mm_head)
2252 goto next_mm;
2253
31dbd01f
IE
2254 ksm_scan.seqnr++;
2255 return NULL;
2256}
2257
2258/**
2259 * ksm_do_scan - the ksm scanner main worker function.
2260 * @scan_npages - number of pages we want to scan before we return.
2261 */
2262static void ksm_do_scan(unsigned int scan_npages)
2263{
2264 struct rmap_item *rmap_item;
22eccdd7 2265 struct page *uninitialized_var(page);
31dbd01f 2266
878aee7d 2267 while (scan_npages-- && likely(!freezing(current))) {
31dbd01f
IE
2268 cond_resched();
2269 rmap_item = scan_get_next_rmap_item(&page);
2270 if (!rmap_item)
2271 return;
4146d2d6 2272 cmp_and_merge_page(page, rmap_item);
31dbd01f
IE
2273 put_page(page);
2274 }
2275}
2276
6e158384
HD
2277static int ksmd_should_run(void)
2278{
2279 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2280}
2281
31dbd01f
IE
2282static int ksm_scan_thread(void *nothing)
2283{
878aee7d 2284 set_freezable();
339aa624 2285 set_user_nice(current, 5);
31dbd01f
IE
2286
2287 while (!kthread_should_stop()) {
6e158384 2288 mutex_lock(&ksm_thread_mutex);
ef4d43a8 2289 wait_while_offlining();
6e158384 2290 if (ksmd_should_run())
31dbd01f 2291 ksm_do_scan(ksm_thread_pages_to_scan);
6e158384
HD
2292 mutex_unlock(&ksm_thread_mutex);
2293
878aee7d
AA
2294 try_to_freeze();
2295
6e158384 2296 if (ksmd_should_run()) {
31dbd01f
IE
2297 schedule_timeout_interruptible(
2298 msecs_to_jiffies(ksm_thread_sleep_millisecs));
2299 } else {
878aee7d 2300 wait_event_freezable(ksm_thread_wait,
6e158384 2301 ksmd_should_run() || kthread_should_stop());
31dbd01f
IE
2302 }
2303 }
2304 return 0;
2305}
2306
f8af4da3
HD
2307int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2308 unsigned long end, int advice, unsigned long *vm_flags)
2309{
2310 struct mm_struct *mm = vma->vm_mm;
d952b791 2311 int err;
f8af4da3
HD
2312
2313 switch (advice) {
2314 case MADV_MERGEABLE:
2315 /*
2316 * Be somewhat over-protective for now!
2317 */
2318 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
2319 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
0661a336 2320 VM_HUGETLB | VM_MIXEDMAP))
f8af4da3
HD
2321 return 0; /* just ignore the advice */
2322
cc2383ec
KK
2323#ifdef VM_SAO
2324 if (*vm_flags & VM_SAO)
2325 return 0;
2326#endif
2327
d952b791
HD
2328 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2329 err = __ksm_enter(mm);
2330 if (err)
2331 return err;
2332 }
f8af4da3
HD
2333
2334 *vm_flags |= VM_MERGEABLE;
2335 break;
2336
2337 case MADV_UNMERGEABLE:
2338 if (!(*vm_flags & VM_MERGEABLE))
2339 return 0; /* just ignore the advice */
2340
d952b791
HD
2341 if (vma->anon_vma) {
2342 err = unmerge_ksm_pages(vma, start, end);
2343 if (err)
2344 return err;
2345 }
f8af4da3
HD
2346
2347 *vm_flags &= ~VM_MERGEABLE;
2348 break;
2349 }
2350
2351 return 0;
2352}
2353
2354int __ksm_enter(struct mm_struct *mm)
2355{
6e158384
HD
2356 struct mm_slot *mm_slot;
2357 int needs_wakeup;
2358
2359 mm_slot = alloc_mm_slot();
31dbd01f
IE
2360 if (!mm_slot)
2361 return -ENOMEM;
2362
6e158384
HD
2363 /* Check ksm_run too? Would need tighter locking */
2364 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2365
31dbd01f
IE
2366 spin_lock(&ksm_mmlist_lock);
2367 insert_to_mm_slots_hash(mm, mm_slot);
2368 /*
cbf86cfe
HD
2369 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2370 * insert just behind the scanning cursor, to let the area settle
31dbd01f
IE
2371 * down a little; when fork is followed by immediate exec, we don't
2372 * want ksmd to waste time setting up and tearing down an rmap_list.
cbf86cfe
HD
2373 *
2374 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2375 * scanning cursor, otherwise KSM pages in newly forked mms will be
2376 * missed: then we might as well insert at the end of the list.
31dbd01f 2377 */
cbf86cfe
HD
2378 if (ksm_run & KSM_RUN_UNMERGE)
2379 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2380 else
2381 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
31dbd01f
IE
2382 spin_unlock(&ksm_mmlist_lock);
2383
f8af4da3 2384 set_bit(MMF_VM_MERGEABLE, &mm->flags);
f1f10076 2385 mmgrab(mm);
6e158384
HD
2386
2387 if (needs_wakeup)
2388 wake_up_interruptible(&ksm_thread_wait);
2389
f8af4da3
HD
2390 return 0;
2391}
2392
1c2fb7a4 2393void __ksm_exit(struct mm_struct *mm)
f8af4da3 2394{
cd551f97 2395 struct mm_slot *mm_slot;
9ba69294 2396 int easy_to_free = 0;
cd551f97 2397
31dbd01f 2398 /*
9ba69294
HD
2399 * This process is exiting: if it's straightforward (as is the
2400 * case when ksmd was never running), free mm_slot immediately.
2401 * But if it's at the cursor or has rmap_items linked to it, use
2402 * mmap_sem to synchronize with any break_cows before pagetables
2403 * are freed, and leave the mm_slot on the list for ksmd to free.
2404 * Beware: ksm may already have noticed it exiting and freed the slot.
31dbd01f 2405 */
9ba69294 2406
cd551f97
HD
2407 spin_lock(&ksm_mmlist_lock);
2408 mm_slot = get_mm_slot(mm);
9ba69294 2409 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
6514d511 2410 if (!mm_slot->rmap_list) {
4ca3a69b 2411 hash_del(&mm_slot->link);
9ba69294
HD
2412 list_del(&mm_slot->mm_list);
2413 easy_to_free = 1;
2414 } else {
2415 list_move(&mm_slot->mm_list,
2416 &ksm_scan.mm_slot->mm_list);
2417 }
cd551f97 2418 }
cd551f97
HD
2419 spin_unlock(&ksm_mmlist_lock);
2420
9ba69294
HD
2421 if (easy_to_free) {
2422 free_mm_slot(mm_slot);
2423 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2424 mmdrop(mm);
2425 } else if (mm_slot) {
9ba69294
HD
2426 down_write(&mm->mmap_sem);
2427 up_write(&mm->mmap_sem);
9ba69294 2428 }
31dbd01f
IE
2429}
2430
cbf86cfe 2431struct page *ksm_might_need_to_copy(struct page *page,
5ad64688
HD
2432 struct vm_area_struct *vma, unsigned long address)
2433{
cbf86cfe 2434 struct anon_vma *anon_vma = page_anon_vma(page);
5ad64688
HD
2435 struct page *new_page;
2436
cbf86cfe
HD
2437 if (PageKsm(page)) {
2438 if (page_stable_node(page) &&
2439 !(ksm_run & KSM_RUN_UNMERGE))
2440 return page; /* no need to copy it */
2441 } else if (!anon_vma) {
2442 return page; /* no need to copy it */
2443 } else if (anon_vma->root == vma->anon_vma->root &&
2444 page->index == linear_page_index(vma, address)) {
2445 return page; /* still no need to copy it */
2446 }
2447 if (!PageUptodate(page))
2448 return page; /* let do_swap_page report the error */
2449
5ad64688
HD
2450 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2451 if (new_page) {
2452 copy_user_highpage(new_page, page, address, vma);
2453
2454 SetPageDirty(new_page);
2455 __SetPageUptodate(new_page);
48c935ad 2456 __SetPageLocked(new_page);
5ad64688
HD
2457 }
2458
5ad64688
HD
2459 return new_page;
2460}
2461
1df631ae 2462void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
e9995ef9
HD
2463{
2464 struct stable_node *stable_node;
e9995ef9 2465 struct rmap_item *rmap_item;
e9995ef9
HD
2466 int search_new_forks = 0;
2467
309381fe 2468 VM_BUG_ON_PAGE(!PageKsm(page), page);
9f32624b
JK
2469
2470 /*
2471 * Rely on the page lock to protect against concurrent modifications
2472 * to that page's node of the stable tree.
2473 */
309381fe 2474 VM_BUG_ON_PAGE(!PageLocked(page), page);
e9995ef9
HD
2475
2476 stable_node = page_stable_node(page);
2477 if (!stable_node)
1df631ae 2478 return;
e9995ef9 2479again:
b67bfe0d 2480 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
e9995ef9 2481 struct anon_vma *anon_vma = rmap_item->anon_vma;
5beb4930 2482 struct anon_vma_chain *vmac;
e9995ef9
HD
2483 struct vm_area_struct *vma;
2484
ad12695f 2485 cond_resched();
b6b19f25 2486 anon_vma_lock_read(anon_vma);
bf181b9f
ML
2487 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2488 0, ULONG_MAX) {
ad12695f 2489 cond_resched();
5beb4930 2490 vma = vmac->vma;
e9995ef9
HD
2491 if (rmap_item->address < vma->vm_start ||
2492 rmap_item->address >= vma->vm_end)
2493 continue;
2494 /*
2495 * Initially we examine only the vma which covers this
2496 * rmap_item; but later, if there is still work to do,
2497 * we examine covering vmas in other mms: in case they
2498 * were forked from the original since ksmd passed.
2499 */
2500 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2501 continue;
2502
0dd1c7bb
JK
2503 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2504 continue;
2505
e4b82222 2506 if (!rwc->rmap_one(page, vma,
1df631ae 2507 rmap_item->address, rwc->arg)) {
b6b19f25 2508 anon_vma_unlock_read(anon_vma);
1df631ae 2509 return;
e9995ef9 2510 }
0dd1c7bb
JK
2511 if (rwc->done && rwc->done(page)) {
2512 anon_vma_unlock_read(anon_vma);
1df631ae 2513 return;
0dd1c7bb 2514 }
e9995ef9 2515 }
b6b19f25 2516 anon_vma_unlock_read(anon_vma);
e9995ef9
HD
2517 }
2518 if (!search_new_forks++)
2519 goto again;
e9995ef9
HD
2520}
2521
52629506 2522#ifdef CONFIG_MIGRATION
e9995ef9
HD
2523void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2524{
2525 struct stable_node *stable_node;
2526
309381fe
SL
2527 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2528 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2529 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
e9995ef9
HD
2530
2531 stable_node = page_stable_node(newpage);
2532 if (stable_node) {
309381fe 2533 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
62b61f61 2534 stable_node->kpfn = page_to_pfn(newpage);
c8d6553b
HD
2535 /*
2536 * newpage->mapping was set in advance; now we need smp_wmb()
2537 * to make sure that the new stable_node->kpfn is visible
2538 * to get_ksm_page() before it can see that oldpage->mapping
2539 * has gone stale (or that PageSwapCache has been cleared).
2540 */
2541 smp_wmb();
2542 set_page_stable_node(oldpage, NULL);
e9995ef9
HD
2543 }
2544}
2545#endif /* CONFIG_MIGRATION */
2546
62b61f61 2547#ifdef CONFIG_MEMORY_HOTREMOVE
ef4d43a8
HD
2548static void wait_while_offlining(void)
2549{
2550 while (ksm_run & KSM_RUN_OFFLINE) {
2551 mutex_unlock(&ksm_thread_mutex);
2552 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
74316201 2553 TASK_UNINTERRUPTIBLE);
ef4d43a8
HD
2554 mutex_lock(&ksm_thread_mutex);
2555 }
2556}
2557
2c653d0e
AA
2558static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2559 unsigned long start_pfn,
2560 unsigned long end_pfn)
2561{
2562 if (stable_node->kpfn >= start_pfn &&
2563 stable_node->kpfn < end_pfn) {
2564 /*
2565 * Don't get_ksm_page, page has already gone:
2566 * which is why we keep kpfn instead of page*
2567 */
2568 remove_node_from_stable_tree(stable_node);
2569 return true;
2570 }
2571 return false;
2572}
2573
2574static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2575 unsigned long start_pfn,
2576 unsigned long end_pfn,
2577 struct rb_root *root)
2578{
2579 struct stable_node *dup;
2580 struct hlist_node *hlist_safe;
2581
2582 if (!is_stable_node_chain(stable_node)) {
2583 VM_BUG_ON(is_stable_node_dup(stable_node));
2584 return stable_node_dup_remove_range(stable_node, start_pfn,
2585 end_pfn);
2586 }
2587
2588 hlist_for_each_entry_safe(dup, hlist_safe,
2589 &stable_node->hlist, hlist_dup) {
2590 VM_BUG_ON(!is_stable_node_dup(dup));
2591 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2592 }
2593 if (hlist_empty(&stable_node->hlist)) {
2594 free_stable_node_chain(stable_node, root);
2595 return true; /* notify caller that tree was rebalanced */
2596 } else
2597 return false;
2598}
2599
ee0ea59c
HD
2600static void ksm_check_stable_tree(unsigned long start_pfn,
2601 unsigned long end_pfn)
62b61f61 2602{
03640418 2603 struct stable_node *stable_node, *next;
62b61f61 2604 struct rb_node *node;
90bd6fd3 2605 int nid;
62b61f61 2606
ef53d16c
HD
2607 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2608 node = rb_first(root_stable_tree + nid);
ee0ea59c 2609 while (node) {
90bd6fd3 2610 stable_node = rb_entry(node, struct stable_node, node);
2c653d0e
AA
2611 if (stable_node_chain_remove_range(stable_node,
2612 start_pfn, end_pfn,
2613 root_stable_tree +
2614 nid))
ef53d16c 2615 node = rb_first(root_stable_tree + nid);
2c653d0e 2616 else
ee0ea59c
HD
2617 node = rb_next(node);
2618 cond_resched();
90bd6fd3 2619 }
ee0ea59c 2620 }
03640418 2621 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
4146d2d6
HD
2622 if (stable_node->kpfn >= start_pfn &&
2623 stable_node->kpfn < end_pfn)
2624 remove_node_from_stable_tree(stable_node);
2625 cond_resched();
2626 }
62b61f61
HD
2627}
2628
2629static int ksm_memory_callback(struct notifier_block *self,
2630 unsigned long action, void *arg)
2631{
2632 struct memory_notify *mn = arg;
62b61f61
HD
2633
2634 switch (action) {
2635 case MEM_GOING_OFFLINE:
2636 /*
ef4d43a8
HD
2637 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2638 * and remove_all_stable_nodes() while memory is going offline:
2639 * it is unsafe for them to touch the stable tree at this time.
2640 * But unmerge_ksm_pages(), rmap lookups and other entry points
2641 * which do not need the ksm_thread_mutex are all safe.
62b61f61 2642 */
ef4d43a8
HD
2643 mutex_lock(&ksm_thread_mutex);
2644 ksm_run |= KSM_RUN_OFFLINE;
2645 mutex_unlock(&ksm_thread_mutex);
62b61f61
HD
2646 break;
2647
2648 case MEM_OFFLINE:
2649 /*
2650 * Most of the work is done by page migration; but there might
2651 * be a few stable_nodes left over, still pointing to struct
ee0ea59c
HD
2652 * pages which have been offlined: prune those from the tree,
2653 * otherwise get_ksm_page() might later try to access a
2654 * non-existent struct page.
62b61f61 2655 */
ee0ea59c
HD
2656 ksm_check_stable_tree(mn->start_pfn,
2657 mn->start_pfn + mn->nr_pages);
62b61f61
HD
2658 /* fallthrough */
2659
2660 case MEM_CANCEL_OFFLINE:
ef4d43a8
HD
2661 mutex_lock(&ksm_thread_mutex);
2662 ksm_run &= ~KSM_RUN_OFFLINE;
62b61f61 2663 mutex_unlock(&ksm_thread_mutex);
ef4d43a8
HD
2664
2665 smp_mb(); /* wake_up_bit advises this */
2666 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
62b61f61
HD
2667 break;
2668 }
2669 return NOTIFY_OK;
2670}
ef4d43a8
HD
2671#else
2672static void wait_while_offlining(void)
2673{
2674}
62b61f61
HD
2675#endif /* CONFIG_MEMORY_HOTREMOVE */
2676
2ffd8679
HD
2677#ifdef CONFIG_SYSFS
2678/*
2679 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2680 */
2681
31dbd01f
IE
2682#define KSM_ATTR_RO(_name) \
2683 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2684#define KSM_ATTR(_name) \
2685 static struct kobj_attribute _name##_attr = \
2686 __ATTR(_name, 0644, _name##_show, _name##_store)
2687
2688static ssize_t sleep_millisecs_show(struct kobject *kobj,
2689 struct kobj_attribute *attr, char *buf)
2690{
2691 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2692}
2693
2694static ssize_t sleep_millisecs_store(struct kobject *kobj,
2695 struct kobj_attribute *attr,
2696 const char *buf, size_t count)
2697{
2698 unsigned long msecs;
2699 int err;
2700
3dbb95f7 2701 err = kstrtoul(buf, 10, &msecs);
31dbd01f
IE
2702 if (err || msecs > UINT_MAX)
2703 return -EINVAL;
2704
2705 ksm_thread_sleep_millisecs = msecs;
2706
2707 return count;
2708}
2709KSM_ATTR(sleep_millisecs);
2710
2711static ssize_t pages_to_scan_show(struct kobject *kobj,
2712 struct kobj_attribute *attr, char *buf)
2713{
2714 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2715}
2716
2717static ssize_t pages_to_scan_store(struct kobject *kobj,
2718 struct kobj_attribute *attr,
2719 const char *buf, size_t count)
2720{
2721 int err;
2722 unsigned long nr_pages;
2723
3dbb95f7 2724 err = kstrtoul(buf, 10, &nr_pages);
31dbd01f
IE
2725 if (err || nr_pages > UINT_MAX)
2726 return -EINVAL;
2727
2728 ksm_thread_pages_to_scan = nr_pages;
2729
2730 return count;
2731}
2732KSM_ATTR(pages_to_scan);
2733
2734static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2735 char *buf)
2736{
ef4d43a8 2737 return sprintf(buf, "%lu\n", ksm_run);
31dbd01f
IE
2738}
2739
2740static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2741 const char *buf, size_t count)
2742{
2743 int err;
2744 unsigned long flags;
2745
3dbb95f7 2746 err = kstrtoul(buf, 10, &flags);
31dbd01f
IE
2747 if (err || flags > UINT_MAX)
2748 return -EINVAL;
2749 if (flags > KSM_RUN_UNMERGE)
2750 return -EINVAL;
2751
2752 /*
2753 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2754 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
d0f209f6
HD
2755 * breaking COW to free the pages_shared (but leaves mm_slots
2756 * on the list for when ksmd may be set running again).
31dbd01f
IE
2757 */
2758
2759 mutex_lock(&ksm_thread_mutex);
ef4d43a8 2760 wait_while_offlining();
31dbd01f
IE
2761 if (ksm_run != flags) {
2762 ksm_run = flags;
d952b791 2763 if (flags & KSM_RUN_UNMERGE) {
e1e12d2f 2764 set_current_oom_origin();
d952b791 2765 err = unmerge_and_remove_all_rmap_items();
e1e12d2f 2766 clear_current_oom_origin();
d952b791
HD
2767 if (err) {
2768 ksm_run = KSM_RUN_STOP;
2769 count = err;
2770 }
2771 }
31dbd01f
IE
2772 }
2773 mutex_unlock(&ksm_thread_mutex);
2774
2775 if (flags & KSM_RUN_MERGE)
2776 wake_up_interruptible(&ksm_thread_wait);
2777
2778 return count;
2779}
2780KSM_ATTR(run);
2781
90bd6fd3
PH
2782#ifdef CONFIG_NUMA
2783static ssize_t merge_across_nodes_show(struct kobject *kobj,
2784 struct kobj_attribute *attr, char *buf)
2785{
2786 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2787}
2788
2789static ssize_t merge_across_nodes_store(struct kobject *kobj,
2790 struct kobj_attribute *attr,
2791 const char *buf, size_t count)
2792{
2793 int err;
2794 unsigned long knob;
2795
2796 err = kstrtoul(buf, 10, &knob);
2797 if (err)
2798 return err;
2799 if (knob > 1)
2800 return -EINVAL;
2801
2802 mutex_lock(&ksm_thread_mutex);
ef4d43a8 2803 wait_while_offlining();
90bd6fd3 2804 if (ksm_merge_across_nodes != knob) {
cbf86cfe 2805 if (ksm_pages_shared || remove_all_stable_nodes())
90bd6fd3 2806 err = -EBUSY;
ef53d16c
HD
2807 else if (root_stable_tree == one_stable_tree) {
2808 struct rb_root *buf;
2809 /*
2810 * This is the first time that we switch away from the
2811 * default of merging across nodes: must now allocate
2812 * a buffer to hold as many roots as may be needed.
2813 * Allocate stable and unstable together:
2814 * MAXSMP NODES_SHIFT 10 will use 16kB.
2815 */
bafe1e14
JP
2816 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2817 GFP_KERNEL);
ef53d16c
HD
2818 /* Let us assume that RB_ROOT is NULL is zero */
2819 if (!buf)
2820 err = -ENOMEM;
2821 else {
2822 root_stable_tree = buf;
2823 root_unstable_tree = buf + nr_node_ids;
2824 /* Stable tree is empty but not the unstable */
2825 root_unstable_tree[0] = one_unstable_tree[0];
2826 }
2827 }
2828 if (!err) {
90bd6fd3 2829 ksm_merge_across_nodes = knob;
ef53d16c
HD
2830 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2831 }
90bd6fd3
PH
2832 }
2833 mutex_unlock(&ksm_thread_mutex);
2834
2835 return err ? err : count;
2836}
2837KSM_ATTR(merge_across_nodes);
2838#endif
2839
e86c59b1
CI
2840static ssize_t use_zero_pages_show(struct kobject *kobj,
2841 struct kobj_attribute *attr, char *buf)
2842{
2843 return sprintf(buf, "%u\n", ksm_use_zero_pages);
2844}
2845static ssize_t use_zero_pages_store(struct kobject *kobj,
2846 struct kobj_attribute *attr,
2847 const char *buf, size_t count)
2848{
2849 int err;
2850 bool value;
2851
2852 err = kstrtobool(buf, &value);
2853 if (err)
2854 return -EINVAL;
2855
2856 ksm_use_zero_pages = value;
2857
2858 return count;
2859}
2860KSM_ATTR(use_zero_pages);
2861
2c653d0e
AA
2862static ssize_t max_page_sharing_show(struct kobject *kobj,
2863 struct kobj_attribute *attr, char *buf)
2864{
2865 return sprintf(buf, "%u\n", ksm_max_page_sharing);
2866}
2867
2868static ssize_t max_page_sharing_store(struct kobject *kobj,
2869 struct kobj_attribute *attr,
2870 const char *buf, size_t count)
2871{
2872 int err;
2873 int knob;
2874
2875 err = kstrtoint(buf, 10, &knob);
2876 if (err)
2877 return err;
2878 /*
2879 * When a KSM page is created it is shared by 2 mappings. This
2880 * being a signed comparison, it implicitly verifies it's not
2881 * negative.
2882 */
2883 if (knob < 2)
2884 return -EINVAL;
2885
2886 if (READ_ONCE(ksm_max_page_sharing) == knob)
2887 return count;
2888
2889 mutex_lock(&ksm_thread_mutex);
2890 wait_while_offlining();
2891 if (ksm_max_page_sharing != knob) {
2892 if (ksm_pages_shared || remove_all_stable_nodes())
2893 err = -EBUSY;
2894 else
2895 ksm_max_page_sharing = knob;
2896 }
2897 mutex_unlock(&ksm_thread_mutex);
2898
2899 return err ? err : count;
2900}
2901KSM_ATTR(max_page_sharing);
2902
b4028260
HD
2903static ssize_t pages_shared_show(struct kobject *kobj,
2904 struct kobj_attribute *attr, char *buf)
2905{
2906 return sprintf(buf, "%lu\n", ksm_pages_shared);
2907}
2908KSM_ATTR_RO(pages_shared);
2909
2910static ssize_t pages_sharing_show(struct kobject *kobj,
2911 struct kobj_attribute *attr, char *buf)
2912{
e178dfde 2913 return sprintf(buf, "%lu\n", ksm_pages_sharing);
b4028260
HD
2914}
2915KSM_ATTR_RO(pages_sharing);
2916
473b0ce4
HD
2917static ssize_t pages_unshared_show(struct kobject *kobj,
2918 struct kobj_attribute *attr, char *buf)
2919{
2920 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2921}
2922KSM_ATTR_RO(pages_unshared);
2923
2924static ssize_t pages_volatile_show(struct kobject *kobj,
2925 struct kobj_attribute *attr, char *buf)
2926{
2927 long ksm_pages_volatile;
2928
2929 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2930 - ksm_pages_sharing - ksm_pages_unshared;
2931 /*
2932 * It was not worth any locking to calculate that statistic,
2933 * but it might therefore sometimes be negative: conceal that.
2934 */
2935 if (ksm_pages_volatile < 0)
2936 ksm_pages_volatile = 0;
2937 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2938}
2939KSM_ATTR_RO(pages_volatile);
2940
2c653d0e
AA
2941static ssize_t stable_node_dups_show(struct kobject *kobj,
2942 struct kobj_attribute *attr, char *buf)
2943{
2944 return sprintf(buf, "%lu\n", ksm_stable_node_dups);
2945}
2946KSM_ATTR_RO(stable_node_dups);
2947
2948static ssize_t stable_node_chains_show(struct kobject *kobj,
2949 struct kobj_attribute *attr, char *buf)
2950{
2951 return sprintf(buf, "%lu\n", ksm_stable_node_chains);
2952}
2953KSM_ATTR_RO(stable_node_chains);
2954
2955static ssize_t
2956stable_node_chains_prune_millisecs_show(struct kobject *kobj,
2957 struct kobj_attribute *attr,
2958 char *buf)
2959{
2960 return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
2961}
2962
2963static ssize_t
2964stable_node_chains_prune_millisecs_store(struct kobject *kobj,
2965 struct kobj_attribute *attr,
2966 const char *buf, size_t count)
2967{
2968 unsigned long msecs;
2969 int err;
2970
2971 err = kstrtoul(buf, 10, &msecs);
2972 if (err || msecs > UINT_MAX)
2973 return -EINVAL;
2974
2975 ksm_stable_node_chains_prune_millisecs = msecs;
2976
2977 return count;
2978}
2979KSM_ATTR(stable_node_chains_prune_millisecs);
2980
473b0ce4
HD
2981static ssize_t full_scans_show(struct kobject *kobj,
2982 struct kobj_attribute *attr, char *buf)
2983{
2984 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2985}
2986KSM_ATTR_RO(full_scans);
2987
31dbd01f
IE
2988static struct attribute *ksm_attrs[] = {
2989 &sleep_millisecs_attr.attr,
2990 &pages_to_scan_attr.attr,
2991 &run_attr.attr,
b4028260
HD
2992 &pages_shared_attr.attr,
2993 &pages_sharing_attr.attr,
473b0ce4
HD
2994 &pages_unshared_attr.attr,
2995 &pages_volatile_attr.attr,
2996 &full_scans_attr.attr,
90bd6fd3
PH
2997#ifdef CONFIG_NUMA
2998 &merge_across_nodes_attr.attr,
2999#endif
2c653d0e
AA
3000 &max_page_sharing_attr.attr,
3001 &stable_node_chains_attr.attr,
3002 &stable_node_dups_attr.attr,
3003 &stable_node_chains_prune_millisecs_attr.attr,
e86c59b1 3004 &use_zero_pages_attr.attr,
31dbd01f
IE
3005 NULL,
3006};
3007
3008static struct attribute_group ksm_attr_group = {
3009 .attrs = ksm_attrs,
3010 .name = "ksm",
3011};
2ffd8679 3012#endif /* CONFIG_SYSFS */
31dbd01f
IE
3013
3014static int __init ksm_init(void)
3015{
3016 struct task_struct *ksm_thread;
3017 int err;
3018
e86c59b1
CI
3019 /* The correct value depends on page size and endianness */
3020 zero_checksum = calc_checksum(ZERO_PAGE(0));
3021 /* Default to false for backwards compatibility */
3022 ksm_use_zero_pages = false;
3023
31dbd01f
IE
3024 err = ksm_slab_init();
3025 if (err)
3026 goto out;
3027
31dbd01f
IE
3028 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3029 if (IS_ERR(ksm_thread)) {
25acde31 3030 pr_err("ksm: creating kthread failed\n");
31dbd01f 3031 err = PTR_ERR(ksm_thread);
d9f8984c 3032 goto out_free;
31dbd01f
IE
3033 }
3034
2ffd8679 3035#ifdef CONFIG_SYSFS
31dbd01f
IE
3036 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3037 if (err) {
25acde31 3038 pr_err("ksm: register sysfs failed\n");
2ffd8679 3039 kthread_stop(ksm_thread);
d9f8984c 3040 goto out_free;
31dbd01f 3041 }
c73602ad
HD
3042#else
3043 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3044
2ffd8679 3045#endif /* CONFIG_SYSFS */
31dbd01f 3046
62b61f61 3047#ifdef CONFIG_MEMORY_HOTREMOVE
ef4d43a8 3048 /* There is no significance to this priority 100 */
62b61f61
HD
3049 hotplug_memory_notifier(ksm_memory_callback, 100);
3050#endif
31dbd01f
IE
3051 return 0;
3052
d9f8984c 3053out_free:
31dbd01f
IE
3054 ksm_slab_free();
3055out:
3056 return err;
f8af4da3 3057}
a64fb3cd 3058subsys_initcall(ksm_init);