Merge tag 'char-misc-5.4-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh...
[linux-2.6-block.git] / mm / ksm.c
CommitLineData
7a338472 1// SPDX-License-Identifier: GPL-2.0-only
f8af4da3 2/*
31dbd01f
IE
3 * Memory merging support.
4 *
5 * This code enables dynamic sharing of identical pages found in different
6 * memory areas, even if they are not shared by fork()
7 *
36b2528d 8 * Copyright (C) 2008-2009 Red Hat, Inc.
31dbd01f
IE
9 * Authors:
10 * Izik Eidus
11 * Andrea Arcangeli
12 * Chris Wright
36b2528d 13 * Hugh Dickins
f8af4da3
HD
14 */
15
16#include <linux/errno.h>
31dbd01f
IE
17#include <linux/mm.h>
18#include <linux/fs.h>
f8af4da3 19#include <linux/mman.h>
31dbd01f 20#include <linux/sched.h>
6e84f315 21#include <linux/sched/mm.h>
f7ccbae4 22#include <linux/sched/coredump.h>
31dbd01f
IE
23#include <linux/rwsem.h>
24#include <linux/pagemap.h>
25#include <linux/rmap.h>
26#include <linux/spinlock.h>
59e1a2f4 27#include <linux/xxhash.h>
31dbd01f
IE
28#include <linux/delay.h>
29#include <linux/kthread.h>
30#include <linux/wait.h>
31#include <linux/slab.h>
32#include <linux/rbtree.h>
62b61f61 33#include <linux/memory.h>
31dbd01f 34#include <linux/mmu_notifier.h>
2c6854fd 35#include <linux/swap.h>
f8af4da3 36#include <linux/ksm.h>
4ca3a69b 37#include <linux/hashtable.h>
878aee7d 38#include <linux/freezer.h>
72788c38 39#include <linux/oom.h>
90bd6fd3 40#include <linux/numa.h>
f8af4da3 41
31dbd01f 42#include <asm/tlbflush.h>
73848b46 43#include "internal.h"
31dbd01f 44
e850dcf5
HD
45#ifdef CONFIG_NUMA
46#define NUMA(x) (x)
47#define DO_NUMA(x) do { (x); } while (0)
48#else
49#define NUMA(x) (0)
50#define DO_NUMA(x) do { } while (0)
51#endif
52
5a2ca3ef
MR
53/**
54 * DOC: Overview
55 *
31dbd01f
IE
56 * A few notes about the KSM scanning process,
57 * to make it easier to understand the data structures below:
58 *
59 * In order to reduce excessive scanning, KSM sorts the memory pages by their
60 * contents into a data structure that holds pointers to the pages' locations.
61 *
62 * Since the contents of the pages may change at any moment, KSM cannot just
63 * insert the pages into a normal sorted tree and expect it to find anything.
64 * Therefore KSM uses two data structures - the stable and the unstable tree.
65 *
66 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
67 * by their contents. Because each such page is write-protected, searching on
68 * this tree is fully assured to be working (except when pages are unmapped),
69 * and therefore this tree is called the stable tree.
70 *
5a2ca3ef
MR
71 * The stable tree node includes information required for reverse
72 * mapping from a KSM page to virtual addresses that map this page.
73 *
74 * In order to avoid large latencies of the rmap walks on KSM pages,
75 * KSM maintains two types of nodes in the stable tree:
76 *
77 * * the regular nodes that keep the reverse mapping structures in a
78 * linked list
79 * * the "chains" that link nodes ("dups") that represent the same
80 * write protected memory content, but each "dup" corresponds to a
81 * different KSM page copy of that content
82 *
83 * Internally, the regular nodes, "dups" and "chains" are represented
84 * using the same :c:type:`struct stable_node` structure.
85 *
31dbd01f
IE
86 * In addition to the stable tree, KSM uses a second data structure called the
87 * unstable tree: this tree holds pointers to pages which have been found to
88 * be "unchanged for a period of time". The unstable tree sorts these pages
89 * by their contents, but since they are not write-protected, KSM cannot rely
90 * upon the unstable tree to work correctly - the unstable tree is liable to
91 * be corrupted as its contents are modified, and so it is called unstable.
92 *
93 * KSM solves this problem by several techniques:
94 *
95 * 1) The unstable tree is flushed every time KSM completes scanning all
96 * memory areas, and then the tree is rebuilt again from the beginning.
97 * 2) KSM will only insert into the unstable tree, pages whose hash value
98 * has not changed since the previous scan of all memory areas.
99 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
100 * colors of the nodes and not on their contents, assuring that even when
101 * the tree gets "corrupted" it won't get out of balance, so scanning time
102 * remains the same (also, searching and inserting nodes in an rbtree uses
103 * the same algorithm, so we have no overhead when we flush and rebuild).
104 * 4) KSM never flushes the stable tree, which means that even if it were to
105 * take 10 attempts to find a page in the unstable tree, once it is found,
106 * it is secured in the stable tree. (When we scan a new page, we first
107 * compare it against the stable tree, and then against the unstable tree.)
8fdb3dbf
HD
108 *
109 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
110 * stable trees and multiple unstable trees: one of each for each NUMA node.
31dbd01f
IE
111 */
112
113/**
114 * struct mm_slot - ksm information per mm that is being scanned
115 * @link: link to the mm_slots hash list
116 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
6514d511 117 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
31dbd01f
IE
118 * @mm: the mm that this information is valid for
119 */
120struct mm_slot {
121 struct hlist_node link;
122 struct list_head mm_list;
6514d511 123 struct rmap_item *rmap_list;
31dbd01f
IE
124 struct mm_struct *mm;
125};
126
127/**
128 * struct ksm_scan - cursor for scanning
129 * @mm_slot: the current mm_slot we are scanning
130 * @address: the next address inside that to be scanned
6514d511 131 * @rmap_list: link to the next rmap to be scanned in the rmap_list
31dbd01f
IE
132 * @seqnr: count of completed full scans (needed when removing unstable node)
133 *
134 * There is only the one ksm_scan instance of this cursor structure.
135 */
136struct ksm_scan {
137 struct mm_slot *mm_slot;
138 unsigned long address;
6514d511 139 struct rmap_item **rmap_list;
31dbd01f
IE
140 unsigned long seqnr;
141};
142
7b6ba2c7
HD
143/**
144 * struct stable_node - node of the stable rbtree
145 * @node: rb node of this ksm page in the stable tree
4146d2d6 146 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
2c653d0e 147 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
4146d2d6 148 * @list: linked into migrate_nodes, pending placement in the proper node tree
7b6ba2c7 149 * @hlist: hlist head of rmap_items using this ksm page
4146d2d6 150 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
2c653d0e
AA
151 * @chain_prune_time: time of the last full garbage collection
152 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
4146d2d6 153 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
7b6ba2c7
HD
154 */
155struct stable_node {
4146d2d6
HD
156 union {
157 struct rb_node node; /* when node of stable tree */
158 struct { /* when listed for migration */
159 struct list_head *head;
2c653d0e
AA
160 struct {
161 struct hlist_node hlist_dup;
162 struct list_head list;
163 };
4146d2d6
HD
164 };
165 };
7b6ba2c7 166 struct hlist_head hlist;
2c653d0e
AA
167 union {
168 unsigned long kpfn;
169 unsigned long chain_prune_time;
170 };
171 /*
172 * STABLE_NODE_CHAIN can be any negative number in
173 * rmap_hlist_len negative range, but better not -1 to be able
174 * to reliably detect underflows.
175 */
176#define STABLE_NODE_CHAIN -1024
177 int rmap_hlist_len;
4146d2d6
HD
178#ifdef CONFIG_NUMA
179 int nid;
180#endif
7b6ba2c7
HD
181};
182
31dbd01f
IE
183/**
184 * struct rmap_item - reverse mapping item for virtual addresses
6514d511 185 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
db114b83 186 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
bc56620b 187 * @nid: NUMA node id of unstable tree in which linked (may not match page)
31dbd01f
IE
188 * @mm: the memory structure this rmap_item is pointing into
189 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
190 * @oldchecksum: previous checksum of the page at that virtual address
7b6ba2c7
HD
191 * @node: rb node of this rmap_item in the unstable tree
192 * @head: pointer to stable_node heading this list in the stable tree
193 * @hlist: link into hlist of rmap_items hanging off that stable_node
31dbd01f
IE
194 */
195struct rmap_item {
6514d511 196 struct rmap_item *rmap_list;
bc56620b
HD
197 union {
198 struct anon_vma *anon_vma; /* when stable */
199#ifdef CONFIG_NUMA
200 int nid; /* when node of unstable tree */
201#endif
202 };
31dbd01f
IE
203 struct mm_struct *mm;
204 unsigned long address; /* + low bits used for flags below */
7b6ba2c7 205 unsigned int oldchecksum; /* when unstable */
31dbd01f 206 union {
7b6ba2c7
HD
207 struct rb_node node; /* when node of unstable tree */
208 struct { /* when listed from stable tree */
209 struct stable_node *head;
210 struct hlist_node hlist;
211 };
31dbd01f
IE
212 };
213};
214
215#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
7b6ba2c7
HD
216#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
217#define STABLE_FLAG 0x200 /* is listed from the stable tree */
1105a2fc
JH
218#define KSM_FLAG_MASK (SEQNR_MASK|UNSTABLE_FLAG|STABLE_FLAG)
219 /* to mask all the flags */
31dbd01f
IE
220
221/* The stable and unstable tree heads */
ef53d16c
HD
222static struct rb_root one_stable_tree[1] = { RB_ROOT };
223static struct rb_root one_unstable_tree[1] = { RB_ROOT };
224static struct rb_root *root_stable_tree = one_stable_tree;
225static struct rb_root *root_unstable_tree = one_unstable_tree;
31dbd01f 226
4146d2d6
HD
227/* Recently migrated nodes of stable tree, pending proper placement */
228static LIST_HEAD(migrate_nodes);
2c653d0e 229#define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
4146d2d6 230
4ca3a69b
SL
231#define MM_SLOTS_HASH_BITS 10
232static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
31dbd01f
IE
233
234static struct mm_slot ksm_mm_head = {
235 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
236};
237static struct ksm_scan ksm_scan = {
238 .mm_slot = &ksm_mm_head,
239};
240
241static struct kmem_cache *rmap_item_cache;
7b6ba2c7 242static struct kmem_cache *stable_node_cache;
31dbd01f
IE
243static struct kmem_cache *mm_slot_cache;
244
245/* The number of nodes in the stable tree */
b4028260 246static unsigned long ksm_pages_shared;
31dbd01f 247
e178dfde 248/* The number of page slots additionally sharing those nodes */
b4028260 249static unsigned long ksm_pages_sharing;
31dbd01f 250
473b0ce4
HD
251/* The number of nodes in the unstable tree */
252static unsigned long ksm_pages_unshared;
253
254/* The number of rmap_items in use: to calculate pages_volatile */
255static unsigned long ksm_rmap_items;
256
2c653d0e
AA
257/* The number of stable_node chains */
258static unsigned long ksm_stable_node_chains;
259
260/* The number of stable_node dups linked to the stable_node chains */
261static unsigned long ksm_stable_node_dups;
262
263/* Delay in pruning stale stable_node_dups in the stable_node_chains */
264static int ksm_stable_node_chains_prune_millisecs = 2000;
265
266/* Maximum number of page slots sharing a stable node */
267static int ksm_max_page_sharing = 256;
268
31dbd01f 269/* Number of pages ksmd should scan in one batch */
2c6854fd 270static unsigned int ksm_thread_pages_to_scan = 100;
31dbd01f
IE
271
272/* Milliseconds ksmd should sleep between batches */
2ffd8679 273static unsigned int ksm_thread_sleep_millisecs = 20;
31dbd01f 274
e86c59b1
CI
275/* Checksum of an empty (zeroed) page */
276static unsigned int zero_checksum __read_mostly;
277
278/* Whether to merge empty (zeroed) pages with actual zero pages */
279static bool ksm_use_zero_pages __read_mostly;
280
e850dcf5 281#ifdef CONFIG_NUMA
90bd6fd3
PH
282/* Zeroed when merging across nodes is not allowed */
283static unsigned int ksm_merge_across_nodes = 1;
ef53d16c 284static int ksm_nr_node_ids = 1;
e850dcf5
HD
285#else
286#define ksm_merge_across_nodes 1U
ef53d16c 287#define ksm_nr_node_ids 1
e850dcf5 288#endif
90bd6fd3 289
31dbd01f
IE
290#define KSM_RUN_STOP 0
291#define KSM_RUN_MERGE 1
292#define KSM_RUN_UNMERGE 2
ef4d43a8
HD
293#define KSM_RUN_OFFLINE 4
294static unsigned long ksm_run = KSM_RUN_STOP;
295static void wait_while_offlining(void);
31dbd01f
IE
296
297static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
fcf9a0ef 298static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait);
31dbd01f
IE
299static DEFINE_MUTEX(ksm_thread_mutex);
300static DEFINE_SPINLOCK(ksm_mmlist_lock);
301
302#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
303 sizeof(struct __struct), __alignof__(struct __struct),\
304 (__flags), NULL)
305
306static int __init ksm_slab_init(void)
307{
308 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
309 if (!rmap_item_cache)
310 goto out;
311
7b6ba2c7
HD
312 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
313 if (!stable_node_cache)
314 goto out_free1;
315
31dbd01f
IE
316 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
317 if (!mm_slot_cache)
7b6ba2c7 318 goto out_free2;
31dbd01f
IE
319
320 return 0;
321
7b6ba2c7
HD
322out_free2:
323 kmem_cache_destroy(stable_node_cache);
324out_free1:
31dbd01f
IE
325 kmem_cache_destroy(rmap_item_cache);
326out:
327 return -ENOMEM;
328}
329
330static void __init ksm_slab_free(void)
331{
332 kmem_cache_destroy(mm_slot_cache);
7b6ba2c7 333 kmem_cache_destroy(stable_node_cache);
31dbd01f
IE
334 kmem_cache_destroy(rmap_item_cache);
335 mm_slot_cache = NULL;
336}
337
2c653d0e
AA
338static __always_inline bool is_stable_node_chain(struct stable_node *chain)
339{
340 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
341}
342
343static __always_inline bool is_stable_node_dup(struct stable_node *dup)
344{
345 return dup->head == STABLE_NODE_DUP_HEAD;
346}
347
348static inline void stable_node_chain_add_dup(struct stable_node *dup,
349 struct stable_node *chain)
350{
351 VM_BUG_ON(is_stable_node_dup(dup));
352 dup->head = STABLE_NODE_DUP_HEAD;
353 VM_BUG_ON(!is_stable_node_chain(chain));
354 hlist_add_head(&dup->hlist_dup, &chain->hlist);
355 ksm_stable_node_dups++;
356}
357
358static inline void __stable_node_dup_del(struct stable_node *dup)
359{
b4fecc67 360 VM_BUG_ON(!is_stable_node_dup(dup));
2c653d0e
AA
361 hlist_del(&dup->hlist_dup);
362 ksm_stable_node_dups--;
363}
364
365static inline void stable_node_dup_del(struct stable_node *dup)
366{
367 VM_BUG_ON(is_stable_node_chain(dup));
368 if (is_stable_node_dup(dup))
369 __stable_node_dup_del(dup);
370 else
371 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
372#ifdef CONFIG_DEBUG_VM
373 dup->head = NULL;
374#endif
375}
376
31dbd01f
IE
377static inline struct rmap_item *alloc_rmap_item(void)
378{
473b0ce4
HD
379 struct rmap_item *rmap_item;
380
5b398e41 381 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
382 __GFP_NORETRY | __GFP_NOWARN);
473b0ce4
HD
383 if (rmap_item)
384 ksm_rmap_items++;
385 return rmap_item;
31dbd01f
IE
386}
387
388static inline void free_rmap_item(struct rmap_item *rmap_item)
389{
473b0ce4 390 ksm_rmap_items--;
31dbd01f
IE
391 rmap_item->mm = NULL; /* debug safety */
392 kmem_cache_free(rmap_item_cache, rmap_item);
393}
394
7b6ba2c7
HD
395static inline struct stable_node *alloc_stable_node(void)
396{
6213055f 397 /*
398 * The allocation can take too long with GFP_KERNEL when memory is under
399 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
400 * grants access to memory reserves, helping to avoid this problem.
401 */
402 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
7b6ba2c7
HD
403}
404
405static inline void free_stable_node(struct stable_node *stable_node)
406{
2c653d0e
AA
407 VM_BUG_ON(stable_node->rmap_hlist_len &&
408 !is_stable_node_chain(stable_node));
7b6ba2c7
HD
409 kmem_cache_free(stable_node_cache, stable_node);
410}
411
31dbd01f
IE
412static inline struct mm_slot *alloc_mm_slot(void)
413{
414 if (!mm_slot_cache) /* initialization failed */
415 return NULL;
416 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
417}
418
419static inline void free_mm_slot(struct mm_slot *mm_slot)
420{
421 kmem_cache_free(mm_slot_cache, mm_slot);
422}
423
31dbd01f
IE
424static struct mm_slot *get_mm_slot(struct mm_struct *mm)
425{
4ca3a69b
SL
426 struct mm_slot *slot;
427
b67bfe0d 428 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
4ca3a69b
SL
429 if (slot->mm == mm)
430 return slot;
31dbd01f 431
31dbd01f
IE
432 return NULL;
433}
434
435static void insert_to_mm_slots_hash(struct mm_struct *mm,
436 struct mm_slot *mm_slot)
437{
31dbd01f 438 mm_slot->mm = mm;
4ca3a69b 439 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
31dbd01f
IE
440}
441
a913e182
HD
442/*
443 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
444 * page tables after it has passed through ksm_exit() - which, if necessary,
445 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
446 * a special flag: they can just back out as soon as mm_users goes to zero.
447 * ksm_test_exit() is used throughout to make this test for exit: in some
448 * places for correctness, in some places just to avoid unnecessary work.
449 */
450static inline bool ksm_test_exit(struct mm_struct *mm)
451{
452 return atomic_read(&mm->mm_users) == 0;
453}
454
31dbd01f
IE
455/*
456 * We use break_ksm to break COW on a ksm page: it's a stripped down
457 *
d4edcf0d 458 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
31dbd01f
IE
459 * put_page(page);
460 *
461 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
462 * in case the application has unmapped and remapped mm,addr meanwhile.
463 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
464 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
1b2ee126
DH
465 *
466 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
467 * of the process that owns 'vma'. We also do not want to enforce
468 * protection keys here anyway.
31dbd01f 469 */
d952b791 470static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
31dbd01f
IE
471{
472 struct page *page;
50a7ca3c 473 vm_fault_t ret = 0;
31dbd01f
IE
474
475 do {
476 cond_resched();
1b2ee126
DH
477 page = follow_page(vma, addr,
478 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
22eccdd7 479 if (IS_ERR_OR_NULL(page))
31dbd01f
IE
480 break;
481 if (PageKsm(page))
dcddffd4
KS
482 ret = handle_mm_fault(vma, addr,
483 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
31dbd01f
IE
484 else
485 ret = VM_FAULT_WRITE;
486 put_page(page);
33692f27 487 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
d952b791
HD
488 /*
489 * We must loop because handle_mm_fault() may back out if there's
490 * any difficulty e.g. if pte accessed bit gets updated concurrently.
491 *
492 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
493 * COW has been broken, even if the vma does not permit VM_WRITE;
494 * but note that a concurrent fault might break PageKsm for us.
495 *
496 * VM_FAULT_SIGBUS could occur if we race with truncation of the
497 * backing file, which also invalidates anonymous pages: that's
498 * okay, that truncation will have unmapped the PageKsm for us.
499 *
500 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
501 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
502 * current task has TIF_MEMDIE set, and will be OOM killed on return
503 * to user; and ksmd, having no mm, would never be chosen for that.
504 *
505 * But if the mm is in a limited mem_cgroup, then the fault may fail
506 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
507 * even ksmd can fail in this way - though it's usually breaking ksm
508 * just to undo a merge it made a moment before, so unlikely to oom.
509 *
510 * That's a pity: we might therefore have more kernel pages allocated
511 * than we're counting as nodes in the stable tree; but ksm_do_scan
512 * will retry to break_cow on each pass, so should recover the page
513 * in due course. The important thing is to not let VM_MERGEABLE
514 * be cleared while any such pages might remain in the area.
515 */
516 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
31dbd01f
IE
517}
518
ef694222
BL
519static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
520 unsigned long addr)
521{
522 struct vm_area_struct *vma;
523 if (ksm_test_exit(mm))
524 return NULL;
525 vma = find_vma(mm, addr);
526 if (!vma || vma->vm_start > addr)
527 return NULL;
528 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
529 return NULL;
530 return vma;
531}
532
8dd3557a 533static void break_cow(struct rmap_item *rmap_item)
31dbd01f 534{
8dd3557a
HD
535 struct mm_struct *mm = rmap_item->mm;
536 unsigned long addr = rmap_item->address;
31dbd01f
IE
537 struct vm_area_struct *vma;
538
4035c07a
HD
539 /*
540 * It is not an accident that whenever we want to break COW
541 * to undo, we also need to drop a reference to the anon_vma.
542 */
9e60109f 543 put_anon_vma(rmap_item->anon_vma);
4035c07a 544
81464e30 545 down_read(&mm->mmap_sem);
ef694222
BL
546 vma = find_mergeable_vma(mm, addr);
547 if (vma)
548 break_ksm(vma, addr);
31dbd01f
IE
549 up_read(&mm->mmap_sem);
550}
551
552static struct page *get_mergeable_page(struct rmap_item *rmap_item)
553{
554 struct mm_struct *mm = rmap_item->mm;
555 unsigned long addr = rmap_item->address;
556 struct vm_area_struct *vma;
557 struct page *page;
558
559 down_read(&mm->mmap_sem);
ef694222
BL
560 vma = find_mergeable_vma(mm, addr);
561 if (!vma)
31dbd01f
IE
562 goto out;
563
564 page = follow_page(vma, addr, FOLL_GET);
22eccdd7 565 if (IS_ERR_OR_NULL(page))
31dbd01f 566 goto out;
f765f540 567 if (PageAnon(page)) {
31dbd01f
IE
568 flush_anon_page(vma, page, addr);
569 flush_dcache_page(page);
570 } else {
571 put_page(page);
c8f95ed1
AA
572out:
573 page = NULL;
31dbd01f
IE
574 }
575 up_read(&mm->mmap_sem);
576 return page;
577}
578
90bd6fd3
PH
579/*
580 * This helper is used for getting right index into array of tree roots.
581 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
582 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
583 * every node has its own stable and unstable tree.
584 */
585static inline int get_kpfn_nid(unsigned long kpfn)
586{
d8fc16a8 587 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
90bd6fd3
PH
588}
589
2c653d0e
AA
590static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
591 struct rb_root *root)
592{
593 struct stable_node *chain = alloc_stable_node();
594 VM_BUG_ON(is_stable_node_chain(dup));
595 if (likely(chain)) {
596 INIT_HLIST_HEAD(&chain->hlist);
597 chain->chain_prune_time = jiffies;
598 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
599#if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
98fa15f3 600 chain->nid = NUMA_NO_NODE; /* debug */
2c653d0e
AA
601#endif
602 ksm_stable_node_chains++;
603
604 /*
605 * Put the stable node chain in the first dimension of
606 * the stable tree and at the same time remove the old
607 * stable node.
608 */
609 rb_replace_node(&dup->node, &chain->node, root);
610
611 /*
612 * Move the old stable node to the second dimension
613 * queued in the hlist_dup. The invariant is that all
614 * dup stable_nodes in the chain->hlist point to pages
615 * that are wrprotected and have the exact same
616 * content.
617 */
618 stable_node_chain_add_dup(dup, chain);
619 }
620 return chain;
621}
622
623static inline void free_stable_node_chain(struct stable_node *chain,
624 struct rb_root *root)
625{
626 rb_erase(&chain->node, root);
627 free_stable_node(chain);
628 ksm_stable_node_chains--;
629}
630
4035c07a
HD
631static void remove_node_from_stable_tree(struct stable_node *stable_node)
632{
633 struct rmap_item *rmap_item;
4035c07a 634
2c653d0e
AA
635 /* check it's not STABLE_NODE_CHAIN or negative */
636 BUG_ON(stable_node->rmap_hlist_len < 0);
637
b67bfe0d 638 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
4035c07a
HD
639 if (rmap_item->hlist.next)
640 ksm_pages_sharing--;
641 else
642 ksm_pages_shared--;
2c653d0e
AA
643 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
644 stable_node->rmap_hlist_len--;
9e60109f 645 put_anon_vma(rmap_item->anon_vma);
4035c07a
HD
646 rmap_item->address &= PAGE_MASK;
647 cond_resched();
648 }
649
2c653d0e
AA
650 /*
651 * We need the second aligned pointer of the migrate_nodes
652 * list_head to stay clear from the rb_parent_color union
653 * (aligned and different than any node) and also different
654 * from &migrate_nodes. This will verify that future list.h changes
815f0ddb 655 * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
2c653d0e 656 */
815f0ddb 657#if defined(GCC_VERSION) && GCC_VERSION >= 40903
2c653d0e
AA
658 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
659 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
660#endif
661
4146d2d6
HD
662 if (stable_node->head == &migrate_nodes)
663 list_del(&stable_node->list);
664 else
2c653d0e 665 stable_node_dup_del(stable_node);
4035c07a
HD
666 free_stable_node(stable_node);
667}
668
2cee57d1
YS
669enum get_ksm_page_flags {
670 GET_KSM_PAGE_NOLOCK,
671 GET_KSM_PAGE_LOCK,
672 GET_KSM_PAGE_TRYLOCK
673};
674
4035c07a
HD
675/*
676 * get_ksm_page: checks if the page indicated by the stable node
677 * is still its ksm page, despite having held no reference to it.
678 * In which case we can trust the content of the page, and it
679 * returns the gotten page; but if the page has now been zapped,
680 * remove the stale node from the stable tree and return NULL.
c8d6553b 681 * But beware, the stable node's page might be being migrated.
4035c07a
HD
682 *
683 * You would expect the stable_node to hold a reference to the ksm page.
684 * But if it increments the page's count, swapping out has to wait for
685 * ksmd to come around again before it can free the page, which may take
686 * seconds or even minutes: much too unresponsive. So instead we use a
687 * "keyhole reference": access to the ksm page from the stable node peeps
688 * out through its keyhole to see if that page still holds the right key,
689 * pointing back to this stable node. This relies on freeing a PageAnon
690 * page to reset its page->mapping to NULL, and relies on no other use of
691 * a page to put something that might look like our key in page->mapping.
4035c07a
HD
692 * is on its way to being freed; but it is an anomaly to bear in mind.
693 */
2cee57d1
YS
694static struct page *get_ksm_page(struct stable_node *stable_node,
695 enum get_ksm_page_flags flags)
4035c07a
HD
696{
697 struct page *page;
698 void *expected_mapping;
c8d6553b 699 unsigned long kpfn;
4035c07a 700
bda807d4
MK
701 expected_mapping = (void *)((unsigned long)stable_node |
702 PAGE_MAPPING_KSM);
c8d6553b 703again:
08df4774 704 kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
c8d6553b 705 page = pfn_to_page(kpfn);
4db0c3c2 706 if (READ_ONCE(page->mapping) != expected_mapping)
4035c07a 707 goto stale;
c8d6553b
HD
708
709 /*
710 * We cannot do anything with the page while its refcount is 0.
711 * Usually 0 means free, or tail of a higher-order page: in which
712 * case this node is no longer referenced, and should be freed;
1c4c3b99 713 * however, it might mean that the page is under page_ref_freeze().
c8d6553b 714 * The __remove_mapping() case is easy, again the node is now stale;
52d1e606
KT
715 * the same is in reuse_ksm_page() case; but if page is swapcache
716 * in migrate_page_move_mapping(), it might still be our page,
717 * in which case it's essential to keep the node.
c8d6553b
HD
718 */
719 while (!get_page_unless_zero(page)) {
720 /*
721 * Another check for page->mapping != expected_mapping would
722 * work here too. We have chosen the !PageSwapCache test to
723 * optimize the common case, when the page is or is about to
724 * be freed: PageSwapCache is cleared (under spin_lock_irq)
1c4c3b99 725 * in the ref_freeze section of __remove_mapping(); but Anon
c8d6553b
HD
726 * page->mapping reset to NULL later, in free_pages_prepare().
727 */
728 if (!PageSwapCache(page))
729 goto stale;
730 cpu_relax();
731 }
732
4db0c3c2 733 if (READ_ONCE(page->mapping) != expected_mapping) {
4035c07a
HD
734 put_page(page);
735 goto stale;
736 }
c8d6553b 737
2cee57d1
YS
738 if (flags == GET_KSM_PAGE_TRYLOCK) {
739 if (!trylock_page(page)) {
740 put_page(page);
741 return ERR_PTR(-EBUSY);
742 }
743 } else if (flags == GET_KSM_PAGE_LOCK)
8aafa6a4 744 lock_page(page);
2cee57d1
YS
745
746 if (flags != GET_KSM_PAGE_NOLOCK) {
4db0c3c2 747 if (READ_ONCE(page->mapping) != expected_mapping) {
8aafa6a4
HD
748 unlock_page(page);
749 put_page(page);
750 goto stale;
751 }
752 }
4035c07a 753 return page;
c8d6553b 754
4035c07a 755stale:
c8d6553b
HD
756 /*
757 * We come here from above when page->mapping or !PageSwapCache
758 * suggests that the node is stale; but it might be under migration.
759 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
760 * before checking whether node->kpfn has been changed.
761 */
762 smp_rmb();
4db0c3c2 763 if (READ_ONCE(stable_node->kpfn) != kpfn)
c8d6553b 764 goto again;
4035c07a
HD
765 remove_node_from_stable_tree(stable_node);
766 return NULL;
767}
768
31dbd01f
IE
769/*
770 * Removing rmap_item from stable or unstable tree.
771 * This function will clean the information from the stable/unstable tree.
772 */
773static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
774{
7b6ba2c7
HD
775 if (rmap_item->address & STABLE_FLAG) {
776 struct stable_node *stable_node;
5ad64688 777 struct page *page;
31dbd01f 778
7b6ba2c7 779 stable_node = rmap_item->head;
2cee57d1 780 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
4035c07a
HD
781 if (!page)
782 goto out;
5ad64688 783
7b6ba2c7 784 hlist_del(&rmap_item->hlist);
4035c07a
HD
785 unlock_page(page);
786 put_page(page);
08beca44 787
98666f8a 788 if (!hlist_empty(&stable_node->hlist))
4035c07a
HD
789 ksm_pages_sharing--;
790 else
7b6ba2c7 791 ksm_pages_shared--;
2c653d0e
AA
792 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
793 stable_node->rmap_hlist_len--;
31dbd01f 794
9e60109f 795 put_anon_vma(rmap_item->anon_vma);
93d17715 796 rmap_item->address &= PAGE_MASK;
31dbd01f 797
7b6ba2c7 798 } else if (rmap_item->address & UNSTABLE_FLAG) {
31dbd01f
IE
799 unsigned char age;
800 /*
9ba69294 801 * Usually ksmd can and must skip the rb_erase, because
31dbd01f 802 * root_unstable_tree was already reset to RB_ROOT.
9ba69294
HD
803 * But be careful when an mm is exiting: do the rb_erase
804 * if this rmap_item was inserted by this scan, rather
805 * than left over from before.
31dbd01f
IE
806 */
807 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
cd551f97 808 BUG_ON(age > 1);
31dbd01f 809 if (!age)
90bd6fd3 810 rb_erase(&rmap_item->node,
ef53d16c 811 root_unstable_tree + NUMA(rmap_item->nid));
473b0ce4 812 ksm_pages_unshared--;
93d17715 813 rmap_item->address &= PAGE_MASK;
31dbd01f 814 }
4035c07a 815out:
31dbd01f
IE
816 cond_resched(); /* we're called from many long loops */
817}
818
31dbd01f 819static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
6514d511 820 struct rmap_item **rmap_list)
31dbd01f 821{
6514d511
HD
822 while (*rmap_list) {
823 struct rmap_item *rmap_item = *rmap_list;
824 *rmap_list = rmap_item->rmap_list;
31dbd01f 825 remove_rmap_item_from_tree(rmap_item);
31dbd01f
IE
826 free_rmap_item(rmap_item);
827 }
828}
829
830/*
e850dcf5 831 * Though it's very tempting to unmerge rmap_items from stable tree rather
31dbd01f
IE
832 * than check every pte of a given vma, the locking doesn't quite work for
833 * that - an rmap_item is assigned to the stable tree after inserting ksm
834 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
835 * rmap_items from parent to child at fork time (so as not to waste time
836 * if exit comes before the next scan reaches it).
81464e30
HD
837 *
838 * Similarly, although we'd like to remove rmap_items (so updating counts
839 * and freeing memory) when unmerging an area, it's easier to leave that
840 * to the next pass of ksmd - consider, for example, how ksmd might be
841 * in cmp_and_merge_page on one of the rmap_items we would be removing.
31dbd01f 842 */
d952b791
HD
843static int unmerge_ksm_pages(struct vm_area_struct *vma,
844 unsigned long start, unsigned long end)
31dbd01f
IE
845{
846 unsigned long addr;
d952b791 847 int err = 0;
31dbd01f 848
d952b791 849 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
9ba69294
HD
850 if (ksm_test_exit(vma->vm_mm))
851 break;
d952b791
HD
852 if (signal_pending(current))
853 err = -ERESTARTSYS;
854 else
855 err = break_ksm(vma, addr);
856 }
857 return err;
31dbd01f
IE
858}
859
88484826
MR
860static inline struct stable_node *page_stable_node(struct page *page)
861{
862 return PageKsm(page) ? page_rmapping(page) : NULL;
863}
864
865static inline void set_page_stable_node(struct page *page,
866 struct stable_node *stable_node)
867{
868 page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
869}
870
2ffd8679
HD
871#ifdef CONFIG_SYSFS
872/*
873 * Only called through the sysfs control interface:
874 */
cbf86cfe
HD
875static int remove_stable_node(struct stable_node *stable_node)
876{
877 struct page *page;
878 int err;
879
2cee57d1 880 page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
cbf86cfe
HD
881 if (!page) {
882 /*
883 * get_ksm_page did remove_node_from_stable_tree itself.
884 */
885 return 0;
886 }
887
8fdb3dbf
HD
888 if (WARN_ON_ONCE(page_mapped(page))) {
889 /*
890 * This should not happen: but if it does, just refuse to let
891 * merge_across_nodes be switched - there is no need to panic.
892 */
cbf86cfe 893 err = -EBUSY;
8fdb3dbf 894 } else {
cbf86cfe 895 /*
8fdb3dbf
HD
896 * The stable node did not yet appear stale to get_ksm_page(),
897 * since that allows for an unmapped ksm page to be recognized
898 * right up until it is freed; but the node is safe to remove.
cbf86cfe
HD
899 * This page might be in a pagevec waiting to be freed,
900 * or it might be PageSwapCache (perhaps under writeback),
901 * or it might have been removed from swapcache a moment ago.
902 */
903 set_page_stable_node(page, NULL);
904 remove_node_from_stable_tree(stable_node);
905 err = 0;
906 }
907
908 unlock_page(page);
909 put_page(page);
910 return err;
911}
912
2c653d0e
AA
913static int remove_stable_node_chain(struct stable_node *stable_node,
914 struct rb_root *root)
915{
916 struct stable_node *dup;
917 struct hlist_node *hlist_safe;
918
919 if (!is_stable_node_chain(stable_node)) {
920 VM_BUG_ON(is_stable_node_dup(stable_node));
921 if (remove_stable_node(stable_node))
922 return true;
923 else
924 return false;
925 }
926
927 hlist_for_each_entry_safe(dup, hlist_safe,
928 &stable_node->hlist, hlist_dup) {
929 VM_BUG_ON(!is_stable_node_dup(dup));
930 if (remove_stable_node(dup))
931 return true;
932 }
933 BUG_ON(!hlist_empty(&stable_node->hlist));
934 free_stable_node_chain(stable_node, root);
935 return false;
936}
937
cbf86cfe
HD
938static int remove_all_stable_nodes(void)
939{
03640418 940 struct stable_node *stable_node, *next;
cbf86cfe
HD
941 int nid;
942 int err = 0;
943
ef53d16c 944 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
cbf86cfe
HD
945 while (root_stable_tree[nid].rb_node) {
946 stable_node = rb_entry(root_stable_tree[nid].rb_node,
947 struct stable_node, node);
2c653d0e
AA
948 if (remove_stable_node_chain(stable_node,
949 root_stable_tree + nid)) {
cbf86cfe
HD
950 err = -EBUSY;
951 break; /* proceed to next nid */
952 }
953 cond_resched();
954 }
955 }
03640418 956 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
4146d2d6
HD
957 if (remove_stable_node(stable_node))
958 err = -EBUSY;
959 cond_resched();
960 }
cbf86cfe
HD
961 return err;
962}
963
d952b791 964static int unmerge_and_remove_all_rmap_items(void)
31dbd01f
IE
965{
966 struct mm_slot *mm_slot;
967 struct mm_struct *mm;
968 struct vm_area_struct *vma;
d952b791
HD
969 int err = 0;
970
971 spin_lock(&ksm_mmlist_lock);
9ba69294 972 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
d952b791
HD
973 struct mm_slot, mm_list);
974 spin_unlock(&ksm_mmlist_lock);
31dbd01f 975
9ba69294
HD
976 for (mm_slot = ksm_scan.mm_slot;
977 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
31dbd01f
IE
978 mm = mm_slot->mm;
979 down_read(&mm->mmap_sem);
980 for (vma = mm->mmap; vma; vma = vma->vm_next) {
9ba69294
HD
981 if (ksm_test_exit(mm))
982 break;
31dbd01f
IE
983 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
984 continue;
d952b791
HD
985 err = unmerge_ksm_pages(vma,
986 vma->vm_start, vma->vm_end);
9ba69294
HD
987 if (err)
988 goto error;
31dbd01f 989 }
9ba69294 990
6514d511 991 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
7496fea9 992 up_read(&mm->mmap_sem);
d952b791
HD
993
994 spin_lock(&ksm_mmlist_lock);
9ba69294 995 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
d952b791 996 struct mm_slot, mm_list);
9ba69294 997 if (ksm_test_exit(mm)) {
4ca3a69b 998 hash_del(&mm_slot->link);
9ba69294
HD
999 list_del(&mm_slot->mm_list);
1000 spin_unlock(&ksm_mmlist_lock);
1001
1002 free_mm_slot(mm_slot);
1003 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
9ba69294 1004 mmdrop(mm);
7496fea9 1005 } else
9ba69294 1006 spin_unlock(&ksm_mmlist_lock);
31dbd01f
IE
1007 }
1008
cbf86cfe
HD
1009 /* Clean up stable nodes, but don't worry if some are still busy */
1010 remove_all_stable_nodes();
d952b791 1011 ksm_scan.seqnr = 0;
9ba69294
HD
1012 return 0;
1013
1014error:
1015 up_read(&mm->mmap_sem);
31dbd01f 1016 spin_lock(&ksm_mmlist_lock);
d952b791 1017 ksm_scan.mm_slot = &ksm_mm_head;
31dbd01f 1018 spin_unlock(&ksm_mmlist_lock);
d952b791 1019 return err;
31dbd01f 1020}
2ffd8679 1021#endif /* CONFIG_SYSFS */
31dbd01f 1022
31dbd01f
IE
1023static u32 calc_checksum(struct page *page)
1024{
1025 u32 checksum;
9b04c5fe 1026 void *addr = kmap_atomic(page);
59e1a2f4 1027 checksum = xxhash(addr, PAGE_SIZE, 0);
9b04c5fe 1028 kunmap_atomic(addr);
31dbd01f
IE
1029 return checksum;
1030}
1031
1032static int memcmp_pages(struct page *page1, struct page *page2)
1033{
1034 char *addr1, *addr2;
1035 int ret;
1036
9b04c5fe
CW
1037 addr1 = kmap_atomic(page1);
1038 addr2 = kmap_atomic(page2);
31dbd01f 1039 ret = memcmp(addr1, addr2, PAGE_SIZE);
9b04c5fe
CW
1040 kunmap_atomic(addr2);
1041 kunmap_atomic(addr1);
31dbd01f
IE
1042 return ret;
1043}
1044
1045static inline int pages_identical(struct page *page1, struct page *page2)
1046{
1047 return !memcmp_pages(page1, page2);
1048}
1049
1050static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1051 pte_t *orig_pte)
1052{
1053 struct mm_struct *mm = vma->vm_mm;
36eaff33
KS
1054 struct page_vma_mapped_walk pvmw = {
1055 .page = page,
1056 .vma = vma,
1057 };
31dbd01f
IE
1058 int swapped;
1059 int err = -EFAULT;
ac46d4f3 1060 struct mmu_notifier_range range;
31dbd01f 1061
36eaff33
KS
1062 pvmw.address = page_address_in_vma(page, vma);
1063 if (pvmw.address == -EFAULT)
31dbd01f
IE
1064 goto out;
1065
29ad768c 1066 BUG_ON(PageTransCompound(page));
6bdb913f 1067
7269f999 1068 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
6f4f13e8 1069 pvmw.address,
ac46d4f3
JG
1070 pvmw.address + PAGE_SIZE);
1071 mmu_notifier_invalidate_range_start(&range);
6bdb913f 1072
36eaff33 1073 if (!page_vma_mapped_walk(&pvmw))
6bdb913f 1074 goto out_mn;
36eaff33
KS
1075 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1076 goto out_unlock;
31dbd01f 1077
595cd8f2 1078 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
b3a81d08
MK
1079 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1080 mm_tlb_flush_pending(mm)) {
31dbd01f
IE
1081 pte_t entry;
1082
1083 swapped = PageSwapCache(page);
36eaff33 1084 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
31dbd01f 1085 /*
25985edc 1086 * Ok this is tricky, when get_user_pages_fast() run it doesn't
31dbd01f
IE
1087 * take any lock, therefore the check that we are going to make
1088 * with the pagecount against the mapcount is racey and
1089 * O_DIRECT can happen right after the check.
1090 * So we clear the pte and flush the tlb before the check
1091 * this assure us that no O_DIRECT can happen after the check
1092 * or in the middle of the check.
0f10851e
JG
1093 *
1094 * No need to notify as we are downgrading page table to read
1095 * only not changing it to point to a new page.
1096 *
ad56b738 1097 * See Documentation/vm/mmu_notifier.rst
31dbd01f 1098 */
0f10851e 1099 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
31dbd01f
IE
1100 /*
1101 * Check that no O_DIRECT or similar I/O is in progress on the
1102 * page
1103 */
31e855ea 1104 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
36eaff33 1105 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
31dbd01f
IE
1106 goto out_unlock;
1107 }
4e31635c
HD
1108 if (pte_dirty(entry))
1109 set_page_dirty(page);
595cd8f2
AK
1110
1111 if (pte_protnone(entry))
1112 entry = pte_mkclean(pte_clear_savedwrite(entry));
1113 else
1114 entry = pte_mkclean(pte_wrprotect(entry));
36eaff33 1115 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
31dbd01f 1116 }
36eaff33 1117 *orig_pte = *pvmw.pte;
31dbd01f
IE
1118 err = 0;
1119
1120out_unlock:
36eaff33 1121 page_vma_mapped_walk_done(&pvmw);
6bdb913f 1122out_mn:
ac46d4f3 1123 mmu_notifier_invalidate_range_end(&range);
31dbd01f
IE
1124out:
1125 return err;
1126}
1127
1128/**
1129 * replace_page - replace page in vma by new ksm page
8dd3557a
HD
1130 * @vma: vma that holds the pte pointing to page
1131 * @page: the page we are replacing by kpage
1132 * @kpage: the ksm page we replace page by
31dbd01f
IE
1133 * @orig_pte: the original value of the pte
1134 *
1135 * Returns 0 on success, -EFAULT on failure.
1136 */
8dd3557a
HD
1137static int replace_page(struct vm_area_struct *vma, struct page *page,
1138 struct page *kpage, pte_t orig_pte)
31dbd01f
IE
1139{
1140 struct mm_struct *mm = vma->vm_mm;
31dbd01f
IE
1141 pmd_t *pmd;
1142 pte_t *ptep;
e86c59b1 1143 pte_t newpte;
31dbd01f
IE
1144 spinlock_t *ptl;
1145 unsigned long addr;
31dbd01f 1146 int err = -EFAULT;
ac46d4f3 1147 struct mmu_notifier_range range;
31dbd01f 1148
8dd3557a 1149 addr = page_address_in_vma(page, vma);
31dbd01f
IE
1150 if (addr == -EFAULT)
1151 goto out;
1152
6219049a
BL
1153 pmd = mm_find_pmd(mm, addr);
1154 if (!pmd)
31dbd01f 1155 goto out;
31dbd01f 1156
7269f999 1157 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr,
6f4f13e8 1158 addr + PAGE_SIZE);
ac46d4f3 1159 mmu_notifier_invalidate_range_start(&range);
6bdb913f 1160
31dbd01f
IE
1161 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1162 if (!pte_same(*ptep, orig_pte)) {
1163 pte_unmap_unlock(ptep, ptl);
6bdb913f 1164 goto out_mn;
31dbd01f
IE
1165 }
1166
e86c59b1
CI
1167 /*
1168 * No need to check ksm_use_zero_pages here: we can only have a
1169 * zero_page here if ksm_use_zero_pages was enabled alreaady.
1170 */
1171 if (!is_zero_pfn(page_to_pfn(kpage))) {
1172 get_page(kpage);
1173 page_add_anon_rmap(kpage, vma, addr, false);
1174 newpte = mk_pte(kpage, vma->vm_page_prot);
1175 } else {
1176 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1177 vma->vm_page_prot));
a38c015f
CI
1178 /*
1179 * We're replacing an anonymous page with a zero page, which is
1180 * not anonymous. We need to do proper accounting otherwise we
1181 * will get wrong values in /proc, and a BUG message in dmesg
1182 * when tearing down the mm.
1183 */
1184 dec_mm_counter(mm, MM_ANONPAGES);
e86c59b1 1185 }
31dbd01f
IE
1186
1187 flush_cache_page(vma, addr, pte_pfn(*ptep));
0f10851e
JG
1188 /*
1189 * No need to notify as we are replacing a read only page with another
1190 * read only page with the same content.
1191 *
ad56b738 1192 * See Documentation/vm/mmu_notifier.rst
0f10851e
JG
1193 */
1194 ptep_clear_flush(vma, addr, ptep);
e86c59b1 1195 set_pte_at_notify(mm, addr, ptep, newpte);
31dbd01f 1196
d281ee61 1197 page_remove_rmap(page, false);
ae52a2ad
HD
1198 if (!page_mapped(page))
1199 try_to_free_swap(page);
8dd3557a 1200 put_page(page);
31dbd01f
IE
1201
1202 pte_unmap_unlock(ptep, ptl);
1203 err = 0;
6bdb913f 1204out_mn:
ac46d4f3 1205 mmu_notifier_invalidate_range_end(&range);
31dbd01f
IE
1206out:
1207 return err;
1208}
1209
1210/*
1211 * try_to_merge_one_page - take two pages and merge them into one
8dd3557a
HD
1212 * @vma: the vma that holds the pte pointing to page
1213 * @page: the PageAnon page that we want to replace with kpage
80e14822
HD
1214 * @kpage: the PageKsm page that we want to map instead of page,
1215 * or NULL the first time when we want to use page as kpage.
31dbd01f
IE
1216 *
1217 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1218 */
1219static int try_to_merge_one_page(struct vm_area_struct *vma,
8dd3557a 1220 struct page *page, struct page *kpage)
31dbd01f
IE
1221{
1222 pte_t orig_pte = __pte(0);
1223 int err = -EFAULT;
1224
db114b83
HD
1225 if (page == kpage) /* ksm page forked */
1226 return 0;
1227
8dd3557a 1228 if (!PageAnon(page))
31dbd01f
IE
1229 goto out;
1230
31dbd01f
IE
1231 /*
1232 * We need the page lock to read a stable PageSwapCache in
1233 * write_protect_page(). We use trylock_page() instead of
1234 * lock_page() because we don't want to wait here - we
1235 * prefer to continue scanning and merging different pages,
1236 * then come back to this page when it is unlocked.
1237 */
8dd3557a 1238 if (!trylock_page(page))
31e855ea 1239 goto out;
f765f540
KS
1240
1241 if (PageTransCompound(page)) {
a7306c34 1242 if (split_huge_page(page))
f765f540
KS
1243 goto out_unlock;
1244 }
1245
31dbd01f
IE
1246 /*
1247 * If this anonymous page is mapped only here, its pte may need
1248 * to be write-protected. If it's mapped elsewhere, all of its
1249 * ptes are necessarily already write-protected. But in either
1250 * case, we need to lock and check page_count is not raised.
1251 */
80e14822
HD
1252 if (write_protect_page(vma, page, &orig_pte) == 0) {
1253 if (!kpage) {
1254 /*
1255 * While we hold page lock, upgrade page from
1256 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1257 * stable_tree_insert() will update stable_node.
1258 */
1259 set_page_stable_node(page, NULL);
1260 mark_page_accessed(page);
337ed7eb
MK
1261 /*
1262 * Page reclaim just frees a clean page with no dirty
1263 * ptes: make sure that the ksm page would be swapped.
1264 */
1265 if (!PageDirty(page))
1266 SetPageDirty(page);
80e14822
HD
1267 err = 0;
1268 } else if (pages_identical(page, kpage))
1269 err = replace_page(vma, page, kpage, orig_pte);
1270 }
31dbd01f 1271
80e14822 1272 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
73848b46 1273 munlock_vma_page(page);
5ad64688
HD
1274 if (!PageMlocked(kpage)) {
1275 unlock_page(page);
5ad64688
HD
1276 lock_page(kpage);
1277 mlock_vma_page(kpage);
1278 page = kpage; /* for final unlock */
1279 }
1280 }
73848b46 1281
f765f540 1282out_unlock:
8dd3557a 1283 unlock_page(page);
31dbd01f
IE
1284out:
1285 return err;
1286}
1287
81464e30
HD
1288/*
1289 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1290 * but no new kernel page is allocated: kpage must already be a ksm page.
8dd3557a
HD
1291 *
1292 * This function returns 0 if the pages were merged, -EFAULT otherwise.
81464e30 1293 */
8dd3557a
HD
1294static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1295 struct page *page, struct page *kpage)
81464e30 1296{
8dd3557a 1297 struct mm_struct *mm = rmap_item->mm;
81464e30
HD
1298 struct vm_area_struct *vma;
1299 int err = -EFAULT;
1300
8dd3557a 1301 down_read(&mm->mmap_sem);
85c6e8dd
AA
1302 vma = find_mergeable_vma(mm, rmap_item->address);
1303 if (!vma)
81464e30
HD
1304 goto out;
1305
8dd3557a 1306 err = try_to_merge_one_page(vma, page, kpage);
db114b83
HD
1307 if (err)
1308 goto out;
1309
bc56620b
HD
1310 /* Unstable nid is in union with stable anon_vma: remove first */
1311 remove_rmap_item_from_tree(rmap_item);
1312
db114b83 1313 /* Must get reference to anon_vma while still holding mmap_sem */
9e60109f
PZ
1314 rmap_item->anon_vma = vma->anon_vma;
1315 get_anon_vma(vma->anon_vma);
81464e30 1316out:
8dd3557a 1317 up_read(&mm->mmap_sem);
81464e30
HD
1318 return err;
1319}
1320
31dbd01f
IE
1321/*
1322 * try_to_merge_two_pages - take two identical pages and prepare them
1323 * to be merged into one page.
1324 *
8dd3557a
HD
1325 * This function returns the kpage if we successfully merged two identical
1326 * pages into one ksm page, NULL otherwise.
31dbd01f 1327 *
80e14822 1328 * Note that this function upgrades page to ksm page: if one of the pages
31dbd01f
IE
1329 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1330 */
8dd3557a
HD
1331static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1332 struct page *page,
1333 struct rmap_item *tree_rmap_item,
1334 struct page *tree_page)
31dbd01f 1335{
80e14822 1336 int err;
31dbd01f 1337
80e14822 1338 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
31dbd01f 1339 if (!err) {
8dd3557a 1340 err = try_to_merge_with_ksm_page(tree_rmap_item,
80e14822 1341 tree_page, page);
31dbd01f 1342 /*
81464e30
HD
1343 * If that fails, we have a ksm page with only one pte
1344 * pointing to it: so break it.
31dbd01f 1345 */
4035c07a 1346 if (err)
8dd3557a 1347 break_cow(rmap_item);
31dbd01f 1348 }
80e14822 1349 return err ? NULL : page;
31dbd01f
IE
1350}
1351
2c653d0e
AA
1352static __always_inline
1353bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1354{
1355 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1356 /*
1357 * Check that at least one mapping still exists, otherwise
1358 * there's no much point to merge and share with this
1359 * stable_node, as the underlying tree_page of the other
1360 * sharer is going to be freed soon.
1361 */
1362 return stable_node->rmap_hlist_len &&
1363 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1364}
1365
1366static __always_inline
1367bool is_page_sharing_candidate(struct stable_node *stable_node)
1368{
1369 return __is_page_sharing_candidate(stable_node, 0);
1370}
1371
c01f0b54
CIK
1372static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1373 struct stable_node **_stable_node,
1374 struct rb_root *root,
1375 bool prune_stale_stable_nodes)
2c653d0e 1376{
b4fecc67 1377 struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
2c653d0e 1378 struct hlist_node *hlist_safe;
8dc5ffcd 1379 struct page *_tree_page, *tree_page = NULL;
2c653d0e
AA
1380 int nr = 0;
1381 int found_rmap_hlist_len;
1382
1383 if (!prune_stale_stable_nodes ||
1384 time_before(jiffies, stable_node->chain_prune_time +
1385 msecs_to_jiffies(
1386 ksm_stable_node_chains_prune_millisecs)))
1387 prune_stale_stable_nodes = false;
1388 else
1389 stable_node->chain_prune_time = jiffies;
1390
1391 hlist_for_each_entry_safe(dup, hlist_safe,
1392 &stable_node->hlist, hlist_dup) {
1393 cond_resched();
1394 /*
1395 * We must walk all stable_node_dup to prune the stale
1396 * stable nodes during lookup.
1397 *
1398 * get_ksm_page can drop the nodes from the
1399 * stable_node->hlist if they point to freed pages
1400 * (that's why we do a _safe walk). The "dup"
1401 * stable_node parameter itself will be freed from
1402 * under us if it returns NULL.
1403 */
2cee57d1 1404 _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK);
2c653d0e
AA
1405 if (!_tree_page)
1406 continue;
1407 nr += 1;
1408 if (is_page_sharing_candidate(dup)) {
1409 if (!found ||
1410 dup->rmap_hlist_len > found_rmap_hlist_len) {
1411 if (found)
8dc5ffcd 1412 put_page(tree_page);
2c653d0e
AA
1413 found = dup;
1414 found_rmap_hlist_len = found->rmap_hlist_len;
8dc5ffcd 1415 tree_page = _tree_page;
2c653d0e 1416
8dc5ffcd 1417 /* skip put_page for found dup */
2c653d0e
AA
1418 if (!prune_stale_stable_nodes)
1419 break;
2c653d0e
AA
1420 continue;
1421 }
1422 }
1423 put_page(_tree_page);
1424 }
1425
80b18dfa
AA
1426 if (found) {
1427 /*
1428 * nr is counting all dups in the chain only if
1429 * prune_stale_stable_nodes is true, otherwise we may
1430 * break the loop at nr == 1 even if there are
1431 * multiple entries.
1432 */
1433 if (prune_stale_stable_nodes && nr == 1) {
2c653d0e
AA
1434 /*
1435 * If there's not just one entry it would
1436 * corrupt memory, better BUG_ON. In KSM
1437 * context with no lock held it's not even
1438 * fatal.
1439 */
1440 BUG_ON(stable_node->hlist.first->next);
1441
1442 /*
1443 * There's just one entry and it is below the
1444 * deduplication limit so drop the chain.
1445 */
1446 rb_replace_node(&stable_node->node, &found->node,
1447 root);
1448 free_stable_node(stable_node);
1449 ksm_stable_node_chains--;
1450 ksm_stable_node_dups--;
b4fecc67 1451 /*
0ba1d0f7
AA
1452 * NOTE: the caller depends on the stable_node
1453 * to be equal to stable_node_dup if the chain
1454 * was collapsed.
b4fecc67 1455 */
0ba1d0f7
AA
1456 *_stable_node = found;
1457 /*
1458 * Just for robustneess as stable_node is
1459 * otherwise left as a stable pointer, the
1460 * compiler shall optimize it away at build
1461 * time.
1462 */
1463 stable_node = NULL;
80b18dfa
AA
1464 } else if (stable_node->hlist.first != &found->hlist_dup &&
1465 __is_page_sharing_candidate(found, 1)) {
2c653d0e 1466 /*
80b18dfa
AA
1467 * If the found stable_node dup can accept one
1468 * more future merge (in addition to the one
1469 * that is underway) and is not at the head of
1470 * the chain, put it there so next search will
1471 * be quicker in the !prune_stale_stable_nodes
1472 * case.
1473 *
1474 * NOTE: it would be inaccurate to use nr > 1
1475 * instead of checking the hlist.first pointer
1476 * directly, because in the
1477 * prune_stale_stable_nodes case "nr" isn't
1478 * the position of the found dup in the chain,
1479 * but the total number of dups in the chain.
2c653d0e
AA
1480 */
1481 hlist_del(&found->hlist_dup);
1482 hlist_add_head(&found->hlist_dup,
1483 &stable_node->hlist);
1484 }
1485 }
1486
8dc5ffcd
AA
1487 *_stable_node_dup = found;
1488 return tree_page;
2c653d0e
AA
1489}
1490
1491static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1492 struct rb_root *root)
1493{
1494 if (!is_stable_node_chain(stable_node))
1495 return stable_node;
1496 if (hlist_empty(&stable_node->hlist)) {
1497 free_stable_node_chain(stable_node, root);
1498 return NULL;
1499 }
1500 return hlist_entry(stable_node->hlist.first,
1501 typeof(*stable_node), hlist_dup);
1502}
1503
8dc5ffcd
AA
1504/*
1505 * Like for get_ksm_page, this function can free the *_stable_node and
1506 * *_stable_node_dup if the returned tree_page is NULL.
1507 *
1508 * It can also free and overwrite *_stable_node with the found
1509 * stable_node_dup if the chain is collapsed (in which case
1510 * *_stable_node will be equal to *_stable_node_dup like if the chain
1511 * never existed). It's up to the caller to verify tree_page is not
1512 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1513 *
1514 * *_stable_node_dup is really a second output parameter of this
1515 * function and will be overwritten in all cases, the caller doesn't
1516 * need to initialize it.
1517 */
1518static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1519 struct stable_node **_stable_node,
1520 struct rb_root *root,
1521 bool prune_stale_stable_nodes)
2c653d0e 1522{
b4fecc67 1523 struct stable_node *stable_node = *_stable_node;
2c653d0e
AA
1524 if (!is_stable_node_chain(stable_node)) {
1525 if (is_page_sharing_candidate(stable_node)) {
8dc5ffcd 1526 *_stable_node_dup = stable_node;
2cee57d1 1527 return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
2c653d0e 1528 }
8dc5ffcd
AA
1529 /*
1530 * _stable_node_dup set to NULL means the stable_node
1531 * reached the ksm_max_page_sharing limit.
1532 */
1533 *_stable_node_dup = NULL;
2c653d0e
AA
1534 return NULL;
1535 }
8dc5ffcd 1536 return stable_node_dup(_stable_node_dup, _stable_node, root,
2c653d0e
AA
1537 prune_stale_stable_nodes);
1538}
1539
8dc5ffcd
AA
1540static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1541 struct stable_node **s_n,
1542 struct rb_root *root)
2c653d0e 1543{
8dc5ffcd 1544 return __stable_node_chain(s_n_d, s_n, root, true);
2c653d0e
AA
1545}
1546
8dc5ffcd
AA
1547static __always_inline struct page *chain(struct stable_node **s_n_d,
1548 struct stable_node *s_n,
1549 struct rb_root *root)
2c653d0e 1550{
8dc5ffcd
AA
1551 struct stable_node *old_stable_node = s_n;
1552 struct page *tree_page;
1553
1554 tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1555 /* not pruning dups so s_n cannot have changed */
1556 VM_BUG_ON(s_n != old_stable_node);
1557 return tree_page;
2c653d0e
AA
1558}
1559
31dbd01f 1560/*
8dd3557a 1561 * stable_tree_search - search for page inside the stable tree
31dbd01f
IE
1562 *
1563 * This function checks if there is a page inside the stable tree
1564 * with identical content to the page that we are scanning right now.
1565 *
7b6ba2c7 1566 * This function returns the stable tree node of identical content if found,
31dbd01f
IE
1567 * NULL otherwise.
1568 */
62b61f61 1569static struct page *stable_tree_search(struct page *page)
31dbd01f 1570{
90bd6fd3 1571 int nid;
ef53d16c 1572 struct rb_root *root;
4146d2d6
HD
1573 struct rb_node **new;
1574 struct rb_node *parent;
2c653d0e 1575 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
4146d2d6 1576 struct stable_node *page_node;
31dbd01f 1577
4146d2d6
HD
1578 page_node = page_stable_node(page);
1579 if (page_node && page_node->head != &migrate_nodes) {
1580 /* ksm page forked */
08beca44 1581 get_page(page);
62b61f61 1582 return page;
08beca44
HD
1583 }
1584
90bd6fd3 1585 nid = get_kpfn_nid(page_to_pfn(page));
ef53d16c 1586 root = root_stable_tree + nid;
4146d2d6 1587again:
ef53d16c 1588 new = &root->rb_node;
4146d2d6 1589 parent = NULL;
90bd6fd3 1590
4146d2d6 1591 while (*new) {
4035c07a 1592 struct page *tree_page;
31dbd01f
IE
1593 int ret;
1594
08beca44 1595 cond_resched();
4146d2d6 1596 stable_node = rb_entry(*new, struct stable_node, node);
2c653d0e 1597 stable_node_any = NULL;
8dc5ffcd 1598 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
b4fecc67
AA
1599 /*
1600 * NOTE: stable_node may have been freed by
1601 * chain_prune() if the returned stable_node_dup is
1602 * not NULL. stable_node_dup may have been inserted in
1603 * the rbtree instead as a regular stable_node (in
1604 * order to collapse the stable_node chain if a single
0ba1d0f7
AA
1605 * stable_node dup was found in it). In such case the
1606 * stable_node is overwritten by the calleee to point
1607 * to the stable_node_dup that was collapsed in the
1608 * stable rbtree and stable_node will be equal to
1609 * stable_node_dup like if the chain never existed.
b4fecc67 1610 */
2c653d0e
AA
1611 if (!stable_node_dup) {
1612 /*
1613 * Either all stable_node dups were full in
1614 * this stable_node chain, or this chain was
1615 * empty and should be rb_erased.
1616 */
1617 stable_node_any = stable_node_dup_any(stable_node,
1618 root);
1619 if (!stable_node_any) {
1620 /* rb_erase just run */
1621 goto again;
1622 }
1623 /*
1624 * Take any of the stable_node dups page of
1625 * this stable_node chain to let the tree walk
1626 * continue. All KSM pages belonging to the
1627 * stable_node dups in a stable_node chain
1628 * have the same content and they're
1629 * wrprotected at all times. Any will work
1630 * fine to continue the walk.
1631 */
2cee57d1
YS
1632 tree_page = get_ksm_page(stable_node_any,
1633 GET_KSM_PAGE_NOLOCK);
2c653d0e
AA
1634 }
1635 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
f2e5ff85
AA
1636 if (!tree_page) {
1637 /*
1638 * If we walked over a stale stable_node,
1639 * get_ksm_page() will call rb_erase() and it
1640 * may rebalance the tree from under us. So
1641 * restart the search from scratch. Returning
1642 * NULL would be safe too, but we'd generate
1643 * false negative insertions just because some
1644 * stable_node was stale.
1645 */
1646 goto again;
1647 }
31dbd01f 1648
4035c07a 1649 ret = memcmp_pages(page, tree_page);
c8d6553b 1650 put_page(tree_page);
31dbd01f 1651
4146d2d6 1652 parent = *new;
c8d6553b 1653 if (ret < 0)
4146d2d6 1654 new = &parent->rb_left;
c8d6553b 1655 else if (ret > 0)
4146d2d6 1656 new = &parent->rb_right;
c8d6553b 1657 else {
2c653d0e
AA
1658 if (page_node) {
1659 VM_BUG_ON(page_node->head != &migrate_nodes);
1660 /*
1661 * Test if the migrated page should be merged
1662 * into a stable node dup. If the mapcount is
1663 * 1 we can migrate it with another KSM page
1664 * without adding it to the chain.
1665 */
1666 if (page_mapcount(page) > 1)
1667 goto chain_append;
1668 }
1669
1670 if (!stable_node_dup) {
1671 /*
1672 * If the stable_node is a chain and
1673 * we got a payload match in memcmp
1674 * but we cannot merge the scanned
1675 * page in any of the existing
1676 * stable_node dups because they're
1677 * all full, we need to wait the
1678 * scanned page to find itself a match
1679 * in the unstable tree to create a
1680 * brand new KSM page to add later to
1681 * the dups of this stable_node.
1682 */
1683 return NULL;
1684 }
1685
c8d6553b
HD
1686 /*
1687 * Lock and unlock the stable_node's page (which
1688 * might already have been migrated) so that page
1689 * migration is sure to notice its raised count.
1690 * It would be more elegant to return stable_node
1691 * than kpage, but that involves more changes.
1692 */
2cee57d1
YS
1693 tree_page = get_ksm_page(stable_node_dup,
1694 GET_KSM_PAGE_TRYLOCK);
1695
1696 if (PTR_ERR(tree_page) == -EBUSY)
1697 return ERR_PTR(-EBUSY);
1698
2c653d0e
AA
1699 if (unlikely(!tree_page))
1700 /*
1701 * The tree may have been rebalanced,
1702 * so re-evaluate parent and new.
1703 */
4146d2d6 1704 goto again;
2c653d0e
AA
1705 unlock_page(tree_page);
1706
1707 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1708 NUMA(stable_node_dup->nid)) {
1709 put_page(tree_page);
1710 goto replace;
1711 }
1712 return tree_page;
c8d6553b 1713 }
31dbd01f
IE
1714 }
1715
4146d2d6
HD
1716 if (!page_node)
1717 return NULL;
1718
1719 list_del(&page_node->list);
1720 DO_NUMA(page_node->nid = nid);
1721 rb_link_node(&page_node->node, parent, new);
ef53d16c 1722 rb_insert_color(&page_node->node, root);
2c653d0e
AA
1723out:
1724 if (is_page_sharing_candidate(page_node)) {
1725 get_page(page);
1726 return page;
1727 } else
1728 return NULL;
4146d2d6
HD
1729
1730replace:
b4fecc67
AA
1731 /*
1732 * If stable_node was a chain and chain_prune collapsed it,
0ba1d0f7
AA
1733 * stable_node has been updated to be the new regular
1734 * stable_node. A collapse of the chain is indistinguishable
1735 * from the case there was no chain in the stable
1736 * rbtree. Otherwise stable_node is the chain and
1737 * stable_node_dup is the dup to replace.
b4fecc67 1738 */
0ba1d0f7 1739 if (stable_node_dup == stable_node) {
b4fecc67
AA
1740 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1741 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
2c653d0e
AA
1742 /* there is no chain */
1743 if (page_node) {
1744 VM_BUG_ON(page_node->head != &migrate_nodes);
1745 list_del(&page_node->list);
1746 DO_NUMA(page_node->nid = nid);
b4fecc67
AA
1747 rb_replace_node(&stable_node_dup->node,
1748 &page_node->node,
2c653d0e
AA
1749 root);
1750 if (is_page_sharing_candidate(page_node))
1751 get_page(page);
1752 else
1753 page = NULL;
1754 } else {
b4fecc67 1755 rb_erase(&stable_node_dup->node, root);
2c653d0e
AA
1756 page = NULL;
1757 }
4146d2d6 1758 } else {
2c653d0e
AA
1759 VM_BUG_ON(!is_stable_node_chain(stable_node));
1760 __stable_node_dup_del(stable_node_dup);
1761 if (page_node) {
1762 VM_BUG_ON(page_node->head != &migrate_nodes);
1763 list_del(&page_node->list);
1764 DO_NUMA(page_node->nid = nid);
1765 stable_node_chain_add_dup(page_node, stable_node);
1766 if (is_page_sharing_candidate(page_node))
1767 get_page(page);
1768 else
1769 page = NULL;
1770 } else {
1771 page = NULL;
1772 }
4146d2d6 1773 }
2c653d0e
AA
1774 stable_node_dup->head = &migrate_nodes;
1775 list_add(&stable_node_dup->list, stable_node_dup->head);
4146d2d6 1776 return page;
2c653d0e
AA
1777
1778chain_append:
1779 /* stable_node_dup could be null if it reached the limit */
1780 if (!stable_node_dup)
1781 stable_node_dup = stable_node_any;
b4fecc67
AA
1782 /*
1783 * If stable_node was a chain and chain_prune collapsed it,
0ba1d0f7
AA
1784 * stable_node has been updated to be the new regular
1785 * stable_node. A collapse of the chain is indistinguishable
1786 * from the case there was no chain in the stable
1787 * rbtree. Otherwise stable_node is the chain and
1788 * stable_node_dup is the dup to replace.
b4fecc67 1789 */
0ba1d0f7 1790 if (stable_node_dup == stable_node) {
b4fecc67
AA
1791 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1792 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
2c653d0e
AA
1793 /* chain is missing so create it */
1794 stable_node = alloc_stable_node_chain(stable_node_dup,
1795 root);
1796 if (!stable_node)
1797 return NULL;
1798 }
1799 /*
1800 * Add this stable_node dup that was
1801 * migrated to the stable_node chain
1802 * of the current nid for this page
1803 * content.
1804 */
b4fecc67
AA
1805 VM_BUG_ON(!is_stable_node_chain(stable_node));
1806 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
2c653d0e
AA
1807 VM_BUG_ON(page_node->head != &migrate_nodes);
1808 list_del(&page_node->list);
1809 DO_NUMA(page_node->nid = nid);
1810 stable_node_chain_add_dup(page_node, stable_node);
1811 goto out;
31dbd01f
IE
1812}
1813
1814/*
e850dcf5 1815 * stable_tree_insert - insert stable tree node pointing to new ksm page
31dbd01f
IE
1816 * into the stable tree.
1817 *
7b6ba2c7
HD
1818 * This function returns the stable tree node just allocated on success,
1819 * NULL otherwise.
31dbd01f 1820 */
7b6ba2c7 1821static struct stable_node *stable_tree_insert(struct page *kpage)
31dbd01f 1822{
90bd6fd3
PH
1823 int nid;
1824 unsigned long kpfn;
ef53d16c 1825 struct rb_root *root;
90bd6fd3 1826 struct rb_node **new;
f2e5ff85 1827 struct rb_node *parent;
2c653d0e
AA
1828 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1829 bool need_chain = false;
31dbd01f 1830
90bd6fd3
PH
1831 kpfn = page_to_pfn(kpage);
1832 nid = get_kpfn_nid(kpfn);
ef53d16c 1833 root = root_stable_tree + nid;
f2e5ff85
AA
1834again:
1835 parent = NULL;
ef53d16c 1836 new = &root->rb_node;
90bd6fd3 1837
31dbd01f 1838 while (*new) {
4035c07a 1839 struct page *tree_page;
31dbd01f
IE
1840 int ret;
1841
08beca44 1842 cond_resched();
7b6ba2c7 1843 stable_node = rb_entry(*new, struct stable_node, node);
2c653d0e 1844 stable_node_any = NULL;
8dc5ffcd 1845 tree_page = chain(&stable_node_dup, stable_node, root);
2c653d0e
AA
1846 if (!stable_node_dup) {
1847 /*
1848 * Either all stable_node dups were full in
1849 * this stable_node chain, or this chain was
1850 * empty and should be rb_erased.
1851 */
1852 stable_node_any = stable_node_dup_any(stable_node,
1853 root);
1854 if (!stable_node_any) {
1855 /* rb_erase just run */
1856 goto again;
1857 }
1858 /*
1859 * Take any of the stable_node dups page of
1860 * this stable_node chain to let the tree walk
1861 * continue. All KSM pages belonging to the
1862 * stable_node dups in a stable_node chain
1863 * have the same content and they're
1864 * wrprotected at all times. Any will work
1865 * fine to continue the walk.
1866 */
2cee57d1
YS
1867 tree_page = get_ksm_page(stable_node_any,
1868 GET_KSM_PAGE_NOLOCK);
2c653d0e
AA
1869 }
1870 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
f2e5ff85
AA
1871 if (!tree_page) {
1872 /*
1873 * If we walked over a stale stable_node,
1874 * get_ksm_page() will call rb_erase() and it
1875 * may rebalance the tree from under us. So
1876 * restart the search from scratch. Returning
1877 * NULL would be safe too, but we'd generate
1878 * false negative insertions just because some
1879 * stable_node was stale.
1880 */
1881 goto again;
1882 }
31dbd01f 1883
4035c07a
HD
1884 ret = memcmp_pages(kpage, tree_page);
1885 put_page(tree_page);
31dbd01f
IE
1886
1887 parent = *new;
1888 if (ret < 0)
1889 new = &parent->rb_left;
1890 else if (ret > 0)
1891 new = &parent->rb_right;
1892 else {
2c653d0e
AA
1893 need_chain = true;
1894 break;
31dbd01f
IE
1895 }
1896 }
1897
2c653d0e
AA
1898 stable_node_dup = alloc_stable_node();
1899 if (!stable_node_dup)
7b6ba2c7 1900 return NULL;
31dbd01f 1901
2c653d0e
AA
1902 INIT_HLIST_HEAD(&stable_node_dup->hlist);
1903 stable_node_dup->kpfn = kpfn;
1904 set_page_stable_node(kpage, stable_node_dup);
1905 stable_node_dup->rmap_hlist_len = 0;
1906 DO_NUMA(stable_node_dup->nid = nid);
1907 if (!need_chain) {
1908 rb_link_node(&stable_node_dup->node, parent, new);
1909 rb_insert_color(&stable_node_dup->node, root);
1910 } else {
1911 if (!is_stable_node_chain(stable_node)) {
1912 struct stable_node *orig = stable_node;
1913 /* chain is missing so create it */
1914 stable_node = alloc_stable_node_chain(orig, root);
1915 if (!stable_node) {
1916 free_stable_node(stable_node_dup);
1917 return NULL;
1918 }
1919 }
1920 stable_node_chain_add_dup(stable_node_dup, stable_node);
1921 }
08beca44 1922
2c653d0e 1923 return stable_node_dup;
31dbd01f
IE
1924}
1925
1926/*
8dd3557a
HD
1927 * unstable_tree_search_insert - search for identical page,
1928 * else insert rmap_item into the unstable tree.
31dbd01f
IE
1929 *
1930 * This function searches for a page in the unstable tree identical to the
1931 * page currently being scanned; and if no identical page is found in the
1932 * tree, we insert rmap_item as a new object into the unstable tree.
1933 *
1934 * This function returns pointer to rmap_item found to be identical
1935 * to the currently scanned page, NULL otherwise.
1936 *
1937 * This function does both searching and inserting, because they share
1938 * the same walking algorithm in an rbtree.
1939 */
8dd3557a
HD
1940static
1941struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1942 struct page *page,
1943 struct page **tree_pagep)
31dbd01f 1944{
90bd6fd3
PH
1945 struct rb_node **new;
1946 struct rb_root *root;
31dbd01f 1947 struct rb_node *parent = NULL;
90bd6fd3
PH
1948 int nid;
1949
1950 nid = get_kpfn_nid(page_to_pfn(page));
ef53d16c 1951 root = root_unstable_tree + nid;
90bd6fd3 1952 new = &root->rb_node;
31dbd01f
IE
1953
1954 while (*new) {
1955 struct rmap_item *tree_rmap_item;
8dd3557a 1956 struct page *tree_page;
31dbd01f
IE
1957 int ret;
1958
d178f27f 1959 cond_resched();
31dbd01f 1960 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
8dd3557a 1961 tree_page = get_mergeable_page(tree_rmap_item);
c8f95ed1 1962 if (!tree_page)
31dbd01f
IE
1963 return NULL;
1964
1965 /*
8dd3557a 1966 * Don't substitute a ksm page for a forked page.
31dbd01f 1967 */
8dd3557a
HD
1968 if (page == tree_page) {
1969 put_page(tree_page);
31dbd01f
IE
1970 return NULL;
1971 }
1972
8dd3557a 1973 ret = memcmp_pages(page, tree_page);
31dbd01f
IE
1974
1975 parent = *new;
1976 if (ret < 0) {
8dd3557a 1977 put_page(tree_page);
31dbd01f
IE
1978 new = &parent->rb_left;
1979 } else if (ret > 0) {
8dd3557a 1980 put_page(tree_page);
31dbd01f 1981 new = &parent->rb_right;
b599cbdf
HD
1982 } else if (!ksm_merge_across_nodes &&
1983 page_to_nid(tree_page) != nid) {
1984 /*
1985 * If tree_page has been migrated to another NUMA node,
1986 * it will be flushed out and put in the right unstable
1987 * tree next time: only merge with it when across_nodes.
1988 */
1989 put_page(tree_page);
1990 return NULL;
31dbd01f 1991 } else {
8dd3557a 1992 *tree_pagep = tree_page;
31dbd01f
IE
1993 return tree_rmap_item;
1994 }
1995 }
1996
7b6ba2c7 1997 rmap_item->address |= UNSTABLE_FLAG;
31dbd01f 1998 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
e850dcf5 1999 DO_NUMA(rmap_item->nid = nid);
31dbd01f 2000 rb_link_node(&rmap_item->node, parent, new);
90bd6fd3 2001 rb_insert_color(&rmap_item->node, root);
31dbd01f 2002
473b0ce4 2003 ksm_pages_unshared++;
31dbd01f
IE
2004 return NULL;
2005}
2006
2007/*
2008 * stable_tree_append - add another rmap_item to the linked list of
2009 * rmap_items hanging off a given node of the stable tree, all sharing
2010 * the same ksm page.
2011 */
2012static void stable_tree_append(struct rmap_item *rmap_item,
2c653d0e
AA
2013 struct stable_node *stable_node,
2014 bool max_page_sharing_bypass)
31dbd01f 2015{
2c653d0e
AA
2016 /*
2017 * rmap won't find this mapping if we don't insert the
2018 * rmap_item in the right stable_node
2019 * duplicate. page_migration could break later if rmap breaks,
2020 * so we can as well crash here. We really need to check for
2021 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2022 * for other negative values as an undeflow if detected here
2023 * for the first time (and not when decreasing rmap_hlist_len)
2024 * would be sign of memory corruption in the stable_node.
2025 */
2026 BUG_ON(stable_node->rmap_hlist_len < 0);
2027
2028 stable_node->rmap_hlist_len++;
2029 if (!max_page_sharing_bypass)
2030 /* possibly non fatal but unexpected overflow, only warn */
2031 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2032 ksm_max_page_sharing);
2033
7b6ba2c7 2034 rmap_item->head = stable_node;
31dbd01f 2035 rmap_item->address |= STABLE_FLAG;
7b6ba2c7 2036 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
e178dfde 2037
7b6ba2c7
HD
2038 if (rmap_item->hlist.next)
2039 ksm_pages_sharing++;
2040 else
2041 ksm_pages_shared++;
31dbd01f
IE
2042}
2043
2044/*
81464e30
HD
2045 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2046 * if not, compare checksum to previous and if it's the same, see if page can
2047 * be inserted into the unstable tree, or merged with a page already there and
2048 * both transferred to the stable tree.
31dbd01f
IE
2049 *
2050 * @page: the page that we are searching identical page to.
2051 * @rmap_item: the reverse mapping into the virtual address of this page
2052 */
2053static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2054{
4b22927f 2055 struct mm_struct *mm = rmap_item->mm;
31dbd01f 2056 struct rmap_item *tree_rmap_item;
8dd3557a 2057 struct page *tree_page = NULL;
7b6ba2c7 2058 struct stable_node *stable_node;
8dd3557a 2059 struct page *kpage;
31dbd01f
IE
2060 unsigned int checksum;
2061 int err;
2c653d0e 2062 bool max_page_sharing_bypass = false;
31dbd01f 2063
4146d2d6
HD
2064 stable_node = page_stable_node(page);
2065 if (stable_node) {
2066 if (stable_node->head != &migrate_nodes &&
2c653d0e
AA
2067 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2068 NUMA(stable_node->nid)) {
2069 stable_node_dup_del(stable_node);
4146d2d6
HD
2070 stable_node->head = &migrate_nodes;
2071 list_add(&stable_node->list, stable_node->head);
2072 }
2073 if (stable_node->head != &migrate_nodes &&
2074 rmap_item->head == stable_node)
2075 return;
2c653d0e
AA
2076 /*
2077 * If it's a KSM fork, allow it to go over the sharing limit
2078 * without warnings.
2079 */
2080 if (!is_page_sharing_candidate(stable_node))
2081 max_page_sharing_bypass = true;
4146d2d6 2082 }
31dbd01f
IE
2083
2084 /* We first start with searching the page inside the stable tree */
62b61f61 2085 kpage = stable_tree_search(page);
4146d2d6
HD
2086 if (kpage == page && rmap_item->head == stable_node) {
2087 put_page(kpage);
2088 return;
2089 }
2090
2091 remove_rmap_item_from_tree(rmap_item);
2092
62b61f61 2093 if (kpage) {
2cee57d1
YS
2094 if (PTR_ERR(kpage) == -EBUSY)
2095 return;
2096
08beca44 2097 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
31dbd01f
IE
2098 if (!err) {
2099 /*
2100 * The page was successfully merged:
2101 * add its rmap_item to the stable tree.
2102 */
5ad64688 2103 lock_page(kpage);
2c653d0e
AA
2104 stable_tree_append(rmap_item, page_stable_node(kpage),
2105 max_page_sharing_bypass);
5ad64688 2106 unlock_page(kpage);
31dbd01f 2107 }
8dd3557a 2108 put_page(kpage);
31dbd01f
IE
2109 return;
2110 }
2111
2112 /*
4035c07a
HD
2113 * If the hash value of the page has changed from the last time
2114 * we calculated it, this page is changing frequently: therefore we
2115 * don't want to insert it in the unstable tree, and we don't want
2116 * to waste our time searching for something identical to it there.
31dbd01f
IE
2117 */
2118 checksum = calc_checksum(page);
2119 if (rmap_item->oldchecksum != checksum) {
2120 rmap_item->oldchecksum = checksum;
2121 return;
2122 }
2123
e86c59b1
CI
2124 /*
2125 * Same checksum as an empty page. We attempt to merge it with the
2126 * appropriate zero page if the user enabled this via sysfs.
2127 */
2128 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2129 struct vm_area_struct *vma;
2130
4b22927f
KT
2131 down_read(&mm->mmap_sem);
2132 vma = find_mergeable_vma(mm, rmap_item->address);
e86c59b1
CI
2133 err = try_to_merge_one_page(vma, page,
2134 ZERO_PAGE(rmap_item->address));
4b22927f 2135 up_read(&mm->mmap_sem);
e86c59b1
CI
2136 /*
2137 * In case of failure, the page was not really empty, so we
2138 * need to continue. Otherwise we're done.
2139 */
2140 if (!err)
2141 return;
2142 }
8dd3557a
HD
2143 tree_rmap_item =
2144 unstable_tree_search_insert(rmap_item, page, &tree_page);
31dbd01f 2145 if (tree_rmap_item) {
77da2ba0
CI
2146 bool split;
2147
8dd3557a
HD
2148 kpage = try_to_merge_two_pages(rmap_item, page,
2149 tree_rmap_item, tree_page);
77da2ba0
CI
2150 /*
2151 * If both pages we tried to merge belong to the same compound
2152 * page, then we actually ended up increasing the reference
2153 * count of the same compound page twice, and split_huge_page
2154 * failed.
2155 * Here we set a flag if that happened, and we use it later to
2156 * try split_huge_page again. Since we call put_page right
2157 * afterwards, the reference count will be correct and
2158 * split_huge_page should succeed.
2159 */
2160 split = PageTransCompound(page)
2161 && compound_head(page) == compound_head(tree_page);
8dd3557a 2162 put_page(tree_page);
8dd3557a 2163 if (kpage) {
bc56620b
HD
2164 /*
2165 * The pages were successfully merged: insert new
2166 * node in the stable tree and add both rmap_items.
2167 */
5ad64688 2168 lock_page(kpage);
7b6ba2c7
HD
2169 stable_node = stable_tree_insert(kpage);
2170 if (stable_node) {
2c653d0e
AA
2171 stable_tree_append(tree_rmap_item, stable_node,
2172 false);
2173 stable_tree_append(rmap_item, stable_node,
2174 false);
7b6ba2c7 2175 }
5ad64688 2176 unlock_page(kpage);
7b6ba2c7 2177
31dbd01f
IE
2178 /*
2179 * If we fail to insert the page into the stable tree,
2180 * we will have 2 virtual addresses that are pointing
2181 * to a ksm page left outside the stable tree,
2182 * in which case we need to break_cow on both.
2183 */
7b6ba2c7 2184 if (!stable_node) {
8dd3557a
HD
2185 break_cow(tree_rmap_item);
2186 break_cow(rmap_item);
31dbd01f 2187 }
77da2ba0
CI
2188 } else if (split) {
2189 /*
2190 * We are here if we tried to merge two pages and
2191 * failed because they both belonged to the same
2192 * compound page. We will split the page now, but no
2193 * merging will take place.
2194 * We do not want to add the cost of a full lock; if
2195 * the page is locked, it is better to skip it and
2196 * perhaps try again later.
2197 */
2198 if (!trylock_page(page))
2199 return;
2200 split_huge_page(page);
2201 unlock_page(page);
31dbd01f 2202 }
31dbd01f
IE
2203 }
2204}
2205
2206static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
6514d511 2207 struct rmap_item **rmap_list,
31dbd01f
IE
2208 unsigned long addr)
2209{
2210 struct rmap_item *rmap_item;
2211
6514d511
HD
2212 while (*rmap_list) {
2213 rmap_item = *rmap_list;
93d17715 2214 if ((rmap_item->address & PAGE_MASK) == addr)
31dbd01f 2215 return rmap_item;
31dbd01f
IE
2216 if (rmap_item->address > addr)
2217 break;
6514d511 2218 *rmap_list = rmap_item->rmap_list;
31dbd01f 2219 remove_rmap_item_from_tree(rmap_item);
31dbd01f
IE
2220 free_rmap_item(rmap_item);
2221 }
2222
2223 rmap_item = alloc_rmap_item();
2224 if (rmap_item) {
2225 /* It has already been zeroed */
2226 rmap_item->mm = mm_slot->mm;
2227 rmap_item->address = addr;
6514d511
HD
2228 rmap_item->rmap_list = *rmap_list;
2229 *rmap_list = rmap_item;
31dbd01f
IE
2230 }
2231 return rmap_item;
2232}
2233
2234static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2235{
2236 struct mm_struct *mm;
2237 struct mm_slot *slot;
2238 struct vm_area_struct *vma;
2239 struct rmap_item *rmap_item;
90bd6fd3 2240 int nid;
31dbd01f
IE
2241
2242 if (list_empty(&ksm_mm_head.mm_list))
2243 return NULL;
2244
2245 slot = ksm_scan.mm_slot;
2246 if (slot == &ksm_mm_head) {
2919bfd0
HD
2247 /*
2248 * A number of pages can hang around indefinitely on per-cpu
2249 * pagevecs, raised page count preventing write_protect_page
2250 * from merging them. Though it doesn't really matter much,
2251 * it is puzzling to see some stuck in pages_volatile until
2252 * other activity jostles them out, and they also prevented
2253 * LTP's KSM test from succeeding deterministically; so drain
2254 * them here (here rather than on entry to ksm_do_scan(),
2255 * so we don't IPI too often when pages_to_scan is set low).
2256 */
2257 lru_add_drain_all();
2258
4146d2d6
HD
2259 /*
2260 * Whereas stale stable_nodes on the stable_tree itself
2261 * get pruned in the regular course of stable_tree_search(),
2262 * those moved out to the migrate_nodes list can accumulate:
2263 * so prune them once before each full scan.
2264 */
2265 if (!ksm_merge_across_nodes) {
03640418 2266 struct stable_node *stable_node, *next;
4146d2d6
HD
2267 struct page *page;
2268
03640418
GT
2269 list_for_each_entry_safe(stable_node, next,
2270 &migrate_nodes, list) {
2cee57d1
YS
2271 page = get_ksm_page(stable_node,
2272 GET_KSM_PAGE_NOLOCK);
4146d2d6
HD
2273 if (page)
2274 put_page(page);
2275 cond_resched();
2276 }
2277 }
2278
ef53d16c 2279 for (nid = 0; nid < ksm_nr_node_ids; nid++)
90bd6fd3 2280 root_unstable_tree[nid] = RB_ROOT;
31dbd01f
IE
2281
2282 spin_lock(&ksm_mmlist_lock);
2283 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2284 ksm_scan.mm_slot = slot;
2285 spin_unlock(&ksm_mmlist_lock);
2b472611
HD
2286 /*
2287 * Although we tested list_empty() above, a racing __ksm_exit
2288 * of the last mm on the list may have removed it since then.
2289 */
2290 if (slot == &ksm_mm_head)
2291 return NULL;
31dbd01f
IE
2292next_mm:
2293 ksm_scan.address = 0;
6514d511 2294 ksm_scan.rmap_list = &slot->rmap_list;
31dbd01f
IE
2295 }
2296
2297 mm = slot->mm;
2298 down_read(&mm->mmap_sem);
9ba69294
HD
2299 if (ksm_test_exit(mm))
2300 vma = NULL;
2301 else
2302 vma = find_vma(mm, ksm_scan.address);
2303
2304 for (; vma; vma = vma->vm_next) {
31dbd01f
IE
2305 if (!(vma->vm_flags & VM_MERGEABLE))
2306 continue;
2307 if (ksm_scan.address < vma->vm_start)
2308 ksm_scan.address = vma->vm_start;
2309 if (!vma->anon_vma)
2310 ksm_scan.address = vma->vm_end;
2311
2312 while (ksm_scan.address < vma->vm_end) {
9ba69294
HD
2313 if (ksm_test_exit(mm))
2314 break;
31dbd01f 2315 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
21ae5b01
AA
2316 if (IS_ERR_OR_NULL(*page)) {
2317 ksm_scan.address += PAGE_SIZE;
2318 cond_resched();
2319 continue;
2320 }
f765f540 2321 if (PageAnon(*page)) {
31dbd01f
IE
2322 flush_anon_page(vma, *page, ksm_scan.address);
2323 flush_dcache_page(*page);
2324 rmap_item = get_next_rmap_item(slot,
6514d511 2325 ksm_scan.rmap_list, ksm_scan.address);
31dbd01f 2326 if (rmap_item) {
6514d511
HD
2327 ksm_scan.rmap_list =
2328 &rmap_item->rmap_list;
31dbd01f
IE
2329 ksm_scan.address += PAGE_SIZE;
2330 } else
2331 put_page(*page);
2332 up_read(&mm->mmap_sem);
2333 return rmap_item;
2334 }
21ae5b01 2335 put_page(*page);
31dbd01f
IE
2336 ksm_scan.address += PAGE_SIZE;
2337 cond_resched();
2338 }
2339 }
2340
9ba69294
HD
2341 if (ksm_test_exit(mm)) {
2342 ksm_scan.address = 0;
6514d511 2343 ksm_scan.rmap_list = &slot->rmap_list;
9ba69294 2344 }
31dbd01f
IE
2345 /*
2346 * Nuke all the rmap_items that are above this current rmap:
2347 * because there were no VM_MERGEABLE vmas with such addresses.
2348 */
6514d511 2349 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
31dbd01f
IE
2350
2351 spin_lock(&ksm_mmlist_lock);
cd551f97
HD
2352 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2353 struct mm_slot, mm_list);
2354 if (ksm_scan.address == 0) {
2355 /*
2356 * We've completed a full scan of all vmas, holding mmap_sem
2357 * throughout, and found no VM_MERGEABLE: so do the same as
2358 * __ksm_exit does to remove this mm from all our lists now.
9ba69294
HD
2359 * This applies either when cleaning up after __ksm_exit
2360 * (but beware: we can reach here even before __ksm_exit),
2361 * or when all VM_MERGEABLE areas have been unmapped (and
2362 * mmap_sem then protects against race with MADV_MERGEABLE).
cd551f97 2363 */
4ca3a69b 2364 hash_del(&slot->link);
cd551f97 2365 list_del(&slot->mm_list);
9ba69294
HD
2366 spin_unlock(&ksm_mmlist_lock);
2367
cd551f97
HD
2368 free_mm_slot(slot);
2369 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
9ba69294
HD
2370 up_read(&mm->mmap_sem);
2371 mmdrop(mm);
2372 } else {
9ba69294 2373 up_read(&mm->mmap_sem);
7496fea9
ZC
2374 /*
2375 * up_read(&mm->mmap_sem) first because after
2376 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2377 * already have been freed under us by __ksm_exit()
2378 * because the "mm_slot" is still hashed and
2379 * ksm_scan.mm_slot doesn't point to it anymore.
2380 */
2381 spin_unlock(&ksm_mmlist_lock);
cd551f97 2382 }
31dbd01f
IE
2383
2384 /* Repeat until we've completed scanning the whole list */
cd551f97 2385 slot = ksm_scan.mm_slot;
31dbd01f
IE
2386 if (slot != &ksm_mm_head)
2387 goto next_mm;
2388
31dbd01f
IE
2389 ksm_scan.seqnr++;
2390 return NULL;
2391}
2392
2393/**
2394 * ksm_do_scan - the ksm scanner main worker function.
b7701a5f 2395 * @scan_npages: number of pages we want to scan before we return.
31dbd01f
IE
2396 */
2397static void ksm_do_scan(unsigned int scan_npages)
2398{
2399 struct rmap_item *rmap_item;
22eccdd7 2400 struct page *uninitialized_var(page);
31dbd01f 2401
878aee7d 2402 while (scan_npages-- && likely(!freezing(current))) {
31dbd01f
IE
2403 cond_resched();
2404 rmap_item = scan_get_next_rmap_item(&page);
2405 if (!rmap_item)
2406 return;
4146d2d6 2407 cmp_and_merge_page(page, rmap_item);
31dbd01f
IE
2408 put_page(page);
2409 }
2410}
2411
6e158384
HD
2412static int ksmd_should_run(void)
2413{
2414 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2415}
2416
31dbd01f
IE
2417static int ksm_scan_thread(void *nothing)
2418{
fcf9a0ef
KT
2419 unsigned int sleep_ms;
2420
878aee7d 2421 set_freezable();
339aa624 2422 set_user_nice(current, 5);
31dbd01f
IE
2423
2424 while (!kthread_should_stop()) {
6e158384 2425 mutex_lock(&ksm_thread_mutex);
ef4d43a8 2426 wait_while_offlining();
6e158384 2427 if (ksmd_should_run())
31dbd01f 2428 ksm_do_scan(ksm_thread_pages_to_scan);
6e158384
HD
2429 mutex_unlock(&ksm_thread_mutex);
2430
878aee7d
AA
2431 try_to_freeze();
2432
6e158384 2433 if (ksmd_should_run()) {
fcf9a0ef
KT
2434 sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs);
2435 wait_event_interruptible_timeout(ksm_iter_wait,
2436 sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs),
2437 msecs_to_jiffies(sleep_ms));
31dbd01f 2438 } else {
878aee7d 2439 wait_event_freezable(ksm_thread_wait,
6e158384 2440 ksmd_should_run() || kthread_should_stop());
31dbd01f
IE
2441 }
2442 }
2443 return 0;
2444}
2445
f8af4da3
HD
2446int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2447 unsigned long end, int advice, unsigned long *vm_flags)
2448{
2449 struct mm_struct *mm = vma->vm_mm;
d952b791 2450 int err;
f8af4da3
HD
2451
2452 switch (advice) {
2453 case MADV_MERGEABLE:
2454 /*
2455 * Be somewhat over-protective for now!
2456 */
2457 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
2458 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
0661a336 2459 VM_HUGETLB | VM_MIXEDMAP))
f8af4da3
HD
2460 return 0; /* just ignore the advice */
2461
e1fb4a08
DJ
2462 if (vma_is_dax(vma))
2463 return 0;
2464
cc2383ec
KK
2465#ifdef VM_SAO
2466 if (*vm_flags & VM_SAO)
2467 return 0;
2468#endif
74a04967
KA
2469#ifdef VM_SPARC_ADI
2470 if (*vm_flags & VM_SPARC_ADI)
2471 return 0;
2472#endif
cc2383ec 2473
d952b791
HD
2474 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2475 err = __ksm_enter(mm);
2476 if (err)
2477 return err;
2478 }
f8af4da3
HD
2479
2480 *vm_flags |= VM_MERGEABLE;
2481 break;
2482
2483 case MADV_UNMERGEABLE:
2484 if (!(*vm_flags & VM_MERGEABLE))
2485 return 0; /* just ignore the advice */
2486
d952b791
HD
2487 if (vma->anon_vma) {
2488 err = unmerge_ksm_pages(vma, start, end);
2489 if (err)
2490 return err;
2491 }
f8af4da3
HD
2492
2493 *vm_flags &= ~VM_MERGEABLE;
2494 break;
2495 }
2496
2497 return 0;
2498}
2499
2500int __ksm_enter(struct mm_struct *mm)
2501{
6e158384
HD
2502 struct mm_slot *mm_slot;
2503 int needs_wakeup;
2504
2505 mm_slot = alloc_mm_slot();
31dbd01f
IE
2506 if (!mm_slot)
2507 return -ENOMEM;
2508
6e158384
HD
2509 /* Check ksm_run too? Would need tighter locking */
2510 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2511
31dbd01f
IE
2512 spin_lock(&ksm_mmlist_lock);
2513 insert_to_mm_slots_hash(mm, mm_slot);
2514 /*
cbf86cfe
HD
2515 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2516 * insert just behind the scanning cursor, to let the area settle
31dbd01f
IE
2517 * down a little; when fork is followed by immediate exec, we don't
2518 * want ksmd to waste time setting up and tearing down an rmap_list.
cbf86cfe
HD
2519 *
2520 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2521 * scanning cursor, otherwise KSM pages in newly forked mms will be
2522 * missed: then we might as well insert at the end of the list.
31dbd01f 2523 */
cbf86cfe
HD
2524 if (ksm_run & KSM_RUN_UNMERGE)
2525 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2526 else
2527 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
31dbd01f
IE
2528 spin_unlock(&ksm_mmlist_lock);
2529
f8af4da3 2530 set_bit(MMF_VM_MERGEABLE, &mm->flags);
f1f10076 2531 mmgrab(mm);
6e158384
HD
2532
2533 if (needs_wakeup)
2534 wake_up_interruptible(&ksm_thread_wait);
2535
f8af4da3
HD
2536 return 0;
2537}
2538
1c2fb7a4 2539void __ksm_exit(struct mm_struct *mm)
f8af4da3 2540{
cd551f97 2541 struct mm_slot *mm_slot;
9ba69294 2542 int easy_to_free = 0;
cd551f97 2543
31dbd01f 2544 /*
9ba69294
HD
2545 * This process is exiting: if it's straightforward (as is the
2546 * case when ksmd was never running), free mm_slot immediately.
2547 * But if it's at the cursor or has rmap_items linked to it, use
2548 * mmap_sem to synchronize with any break_cows before pagetables
2549 * are freed, and leave the mm_slot on the list for ksmd to free.
2550 * Beware: ksm may already have noticed it exiting and freed the slot.
31dbd01f 2551 */
9ba69294 2552
cd551f97
HD
2553 spin_lock(&ksm_mmlist_lock);
2554 mm_slot = get_mm_slot(mm);
9ba69294 2555 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
6514d511 2556 if (!mm_slot->rmap_list) {
4ca3a69b 2557 hash_del(&mm_slot->link);
9ba69294
HD
2558 list_del(&mm_slot->mm_list);
2559 easy_to_free = 1;
2560 } else {
2561 list_move(&mm_slot->mm_list,
2562 &ksm_scan.mm_slot->mm_list);
2563 }
cd551f97 2564 }
cd551f97
HD
2565 spin_unlock(&ksm_mmlist_lock);
2566
9ba69294
HD
2567 if (easy_to_free) {
2568 free_mm_slot(mm_slot);
2569 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2570 mmdrop(mm);
2571 } else if (mm_slot) {
9ba69294
HD
2572 down_write(&mm->mmap_sem);
2573 up_write(&mm->mmap_sem);
9ba69294 2574 }
31dbd01f
IE
2575}
2576
cbf86cfe 2577struct page *ksm_might_need_to_copy(struct page *page,
5ad64688
HD
2578 struct vm_area_struct *vma, unsigned long address)
2579{
cbf86cfe 2580 struct anon_vma *anon_vma = page_anon_vma(page);
5ad64688
HD
2581 struct page *new_page;
2582
cbf86cfe
HD
2583 if (PageKsm(page)) {
2584 if (page_stable_node(page) &&
2585 !(ksm_run & KSM_RUN_UNMERGE))
2586 return page; /* no need to copy it */
2587 } else if (!anon_vma) {
2588 return page; /* no need to copy it */
2589 } else if (anon_vma->root == vma->anon_vma->root &&
2590 page->index == linear_page_index(vma, address)) {
2591 return page; /* still no need to copy it */
2592 }
2593 if (!PageUptodate(page))
2594 return page; /* let do_swap_page report the error */
2595
5ad64688
HD
2596 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2597 if (new_page) {
2598 copy_user_highpage(new_page, page, address, vma);
2599
2600 SetPageDirty(new_page);
2601 __SetPageUptodate(new_page);
48c935ad 2602 __SetPageLocked(new_page);
5ad64688
HD
2603 }
2604
5ad64688
HD
2605 return new_page;
2606}
2607
1df631ae 2608void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
e9995ef9
HD
2609{
2610 struct stable_node *stable_node;
e9995ef9 2611 struct rmap_item *rmap_item;
e9995ef9
HD
2612 int search_new_forks = 0;
2613
309381fe 2614 VM_BUG_ON_PAGE(!PageKsm(page), page);
9f32624b
JK
2615
2616 /*
2617 * Rely on the page lock to protect against concurrent modifications
2618 * to that page's node of the stable tree.
2619 */
309381fe 2620 VM_BUG_ON_PAGE(!PageLocked(page), page);
e9995ef9
HD
2621
2622 stable_node = page_stable_node(page);
2623 if (!stable_node)
1df631ae 2624 return;
e9995ef9 2625again:
b67bfe0d 2626 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
e9995ef9 2627 struct anon_vma *anon_vma = rmap_item->anon_vma;
5beb4930 2628 struct anon_vma_chain *vmac;
e9995ef9
HD
2629 struct vm_area_struct *vma;
2630
ad12695f 2631 cond_resched();
b6b19f25 2632 anon_vma_lock_read(anon_vma);
bf181b9f
ML
2633 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2634 0, ULONG_MAX) {
1105a2fc
JH
2635 unsigned long addr;
2636
ad12695f 2637 cond_resched();
5beb4930 2638 vma = vmac->vma;
1105a2fc
JH
2639
2640 /* Ignore the stable/unstable/sqnr flags */
2641 addr = rmap_item->address & ~KSM_FLAG_MASK;
2642
2643 if (addr < vma->vm_start || addr >= vma->vm_end)
e9995ef9
HD
2644 continue;
2645 /*
2646 * Initially we examine only the vma which covers this
2647 * rmap_item; but later, if there is still work to do,
2648 * we examine covering vmas in other mms: in case they
2649 * were forked from the original since ksmd passed.
2650 */
2651 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2652 continue;
2653
0dd1c7bb
JK
2654 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2655 continue;
2656
1105a2fc 2657 if (!rwc->rmap_one(page, vma, addr, rwc->arg)) {
b6b19f25 2658 anon_vma_unlock_read(anon_vma);
1df631ae 2659 return;
e9995ef9 2660 }
0dd1c7bb
JK
2661 if (rwc->done && rwc->done(page)) {
2662 anon_vma_unlock_read(anon_vma);
1df631ae 2663 return;
0dd1c7bb 2664 }
e9995ef9 2665 }
b6b19f25 2666 anon_vma_unlock_read(anon_vma);
e9995ef9
HD
2667 }
2668 if (!search_new_forks++)
2669 goto again;
e9995ef9
HD
2670}
2671
52d1e606
KT
2672bool reuse_ksm_page(struct page *page,
2673 struct vm_area_struct *vma,
2674 unsigned long address)
2675{
2676#ifdef CONFIG_DEBUG_VM
2677 if (WARN_ON(is_zero_pfn(page_to_pfn(page))) ||
2678 WARN_ON(!page_mapped(page)) ||
2679 WARN_ON(!PageLocked(page))) {
2680 dump_page(page, "reuse_ksm_page");
2681 return false;
2682 }
2683#endif
2684
2685 if (PageSwapCache(page) || !page_stable_node(page))
2686 return false;
2687 /* Prohibit parallel get_ksm_page() */
2688 if (!page_ref_freeze(page, 1))
2689 return false;
2690
2691 page_move_anon_rmap(page, vma);
2692 page->index = linear_page_index(vma, address);
2693 page_ref_unfreeze(page, 1);
2694
2695 return true;
2696}
52629506 2697#ifdef CONFIG_MIGRATION
e9995ef9
HD
2698void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2699{
2700 struct stable_node *stable_node;
2701
309381fe
SL
2702 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2703 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2704 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
e9995ef9
HD
2705
2706 stable_node = page_stable_node(newpage);
2707 if (stable_node) {
309381fe 2708 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
62b61f61 2709 stable_node->kpfn = page_to_pfn(newpage);
c8d6553b
HD
2710 /*
2711 * newpage->mapping was set in advance; now we need smp_wmb()
2712 * to make sure that the new stable_node->kpfn is visible
2713 * to get_ksm_page() before it can see that oldpage->mapping
2714 * has gone stale (or that PageSwapCache has been cleared).
2715 */
2716 smp_wmb();
2717 set_page_stable_node(oldpage, NULL);
e9995ef9
HD
2718 }
2719}
2720#endif /* CONFIG_MIGRATION */
2721
62b61f61 2722#ifdef CONFIG_MEMORY_HOTREMOVE
ef4d43a8
HD
2723static void wait_while_offlining(void)
2724{
2725 while (ksm_run & KSM_RUN_OFFLINE) {
2726 mutex_unlock(&ksm_thread_mutex);
2727 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
74316201 2728 TASK_UNINTERRUPTIBLE);
ef4d43a8
HD
2729 mutex_lock(&ksm_thread_mutex);
2730 }
2731}
2732
2c653d0e
AA
2733static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2734 unsigned long start_pfn,
2735 unsigned long end_pfn)
2736{
2737 if (stable_node->kpfn >= start_pfn &&
2738 stable_node->kpfn < end_pfn) {
2739 /*
2740 * Don't get_ksm_page, page has already gone:
2741 * which is why we keep kpfn instead of page*
2742 */
2743 remove_node_from_stable_tree(stable_node);
2744 return true;
2745 }
2746 return false;
2747}
2748
2749static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2750 unsigned long start_pfn,
2751 unsigned long end_pfn,
2752 struct rb_root *root)
2753{
2754 struct stable_node *dup;
2755 struct hlist_node *hlist_safe;
2756
2757 if (!is_stable_node_chain(stable_node)) {
2758 VM_BUG_ON(is_stable_node_dup(stable_node));
2759 return stable_node_dup_remove_range(stable_node, start_pfn,
2760 end_pfn);
2761 }
2762
2763 hlist_for_each_entry_safe(dup, hlist_safe,
2764 &stable_node->hlist, hlist_dup) {
2765 VM_BUG_ON(!is_stable_node_dup(dup));
2766 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2767 }
2768 if (hlist_empty(&stable_node->hlist)) {
2769 free_stable_node_chain(stable_node, root);
2770 return true; /* notify caller that tree was rebalanced */
2771 } else
2772 return false;
2773}
2774
ee0ea59c
HD
2775static void ksm_check_stable_tree(unsigned long start_pfn,
2776 unsigned long end_pfn)
62b61f61 2777{
03640418 2778 struct stable_node *stable_node, *next;
62b61f61 2779 struct rb_node *node;
90bd6fd3 2780 int nid;
62b61f61 2781
ef53d16c
HD
2782 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2783 node = rb_first(root_stable_tree + nid);
ee0ea59c 2784 while (node) {
90bd6fd3 2785 stable_node = rb_entry(node, struct stable_node, node);
2c653d0e
AA
2786 if (stable_node_chain_remove_range(stable_node,
2787 start_pfn, end_pfn,
2788 root_stable_tree +
2789 nid))
ef53d16c 2790 node = rb_first(root_stable_tree + nid);
2c653d0e 2791 else
ee0ea59c
HD
2792 node = rb_next(node);
2793 cond_resched();
90bd6fd3 2794 }
ee0ea59c 2795 }
03640418 2796 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
4146d2d6
HD
2797 if (stable_node->kpfn >= start_pfn &&
2798 stable_node->kpfn < end_pfn)
2799 remove_node_from_stable_tree(stable_node);
2800 cond_resched();
2801 }
62b61f61
HD
2802}
2803
2804static int ksm_memory_callback(struct notifier_block *self,
2805 unsigned long action, void *arg)
2806{
2807 struct memory_notify *mn = arg;
62b61f61
HD
2808
2809 switch (action) {
2810 case MEM_GOING_OFFLINE:
2811 /*
ef4d43a8
HD
2812 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2813 * and remove_all_stable_nodes() while memory is going offline:
2814 * it is unsafe for them to touch the stable tree at this time.
2815 * But unmerge_ksm_pages(), rmap lookups and other entry points
2816 * which do not need the ksm_thread_mutex are all safe.
62b61f61 2817 */
ef4d43a8
HD
2818 mutex_lock(&ksm_thread_mutex);
2819 ksm_run |= KSM_RUN_OFFLINE;
2820 mutex_unlock(&ksm_thread_mutex);
62b61f61
HD
2821 break;
2822
2823 case MEM_OFFLINE:
2824 /*
2825 * Most of the work is done by page migration; but there might
2826 * be a few stable_nodes left over, still pointing to struct
ee0ea59c
HD
2827 * pages which have been offlined: prune those from the tree,
2828 * otherwise get_ksm_page() might later try to access a
2829 * non-existent struct page.
62b61f61 2830 */
ee0ea59c
HD
2831 ksm_check_stable_tree(mn->start_pfn,
2832 mn->start_pfn + mn->nr_pages);
62b61f61
HD
2833 /* fallthrough */
2834
2835 case MEM_CANCEL_OFFLINE:
ef4d43a8
HD
2836 mutex_lock(&ksm_thread_mutex);
2837 ksm_run &= ~KSM_RUN_OFFLINE;
62b61f61 2838 mutex_unlock(&ksm_thread_mutex);
ef4d43a8
HD
2839
2840 smp_mb(); /* wake_up_bit advises this */
2841 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
62b61f61
HD
2842 break;
2843 }
2844 return NOTIFY_OK;
2845}
ef4d43a8
HD
2846#else
2847static void wait_while_offlining(void)
2848{
2849}
62b61f61
HD
2850#endif /* CONFIG_MEMORY_HOTREMOVE */
2851
2ffd8679
HD
2852#ifdef CONFIG_SYSFS
2853/*
2854 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2855 */
2856
31dbd01f
IE
2857#define KSM_ATTR_RO(_name) \
2858 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2859#define KSM_ATTR(_name) \
2860 static struct kobj_attribute _name##_attr = \
2861 __ATTR(_name, 0644, _name##_show, _name##_store)
2862
2863static ssize_t sleep_millisecs_show(struct kobject *kobj,
2864 struct kobj_attribute *attr, char *buf)
2865{
2866 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2867}
2868
2869static ssize_t sleep_millisecs_store(struct kobject *kobj,
2870 struct kobj_attribute *attr,
2871 const char *buf, size_t count)
2872{
2873 unsigned long msecs;
2874 int err;
2875
3dbb95f7 2876 err = kstrtoul(buf, 10, &msecs);
31dbd01f
IE
2877 if (err || msecs > UINT_MAX)
2878 return -EINVAL;
2879
2880 ksm_thread_sleep_millisecs = msecs;
fcf9a0ef 2881 wake_up_interruptible(&ksm_iter_wait);
31dbd01f
IE
2882
2883 return count;
2884}
2885KSM_ATTR(sleep_millisecs);
2886
2887static ssize_t pages_to_scan_show(struct kobject *kobj,
2888 struct kobj_attribute *attr, char *buf)
2889{
2890 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2891}
2892
2893static ssize_t pages_to_scan_store(struct kobject *kobj,
2894 struct kobj_attribute *attr,
2895 const char *buf, size_t count)
2896{
2897 int err;
2898 unsigned long nr_pages;
2899
3dbb95f7 2900 err = kstrtoul(buf, 10, &nr_pages);
31dbd01f
IE
2901 if (err || nr_pages > UINT_MAX)
2902 return -EINVAL;
2903
2904 ksm_thread_pages_to_scan = nr_pages;
2905
2906 return count;
2907}
2908KSM_ATTR(pages_to_scan);
2909
2910static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2911 char *buf)
2912{
ef4d43a8 2913 return sprintf(buf, "%lu\n", ksm_run);
31dbd01f
IE
2914}
2915
2916static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2917 const char *buf, size_t count)
2918{
2919 int err;
2920 unsigned long flags;
2921
3dbb95f7 2922 err = kstrtoul(buf, 10, &flags);
31dbd01f
IE
2923 if (err || flags > UINT_MAX)
2924 return -EINVAL;
2925 if (flags > KSM_RUN_UNMERGE)
2926 return -EINVAL;
2927
2928 /*
2929 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2930 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
d0f209f6
HD
2931 * breaking COW to free the pages_shared (but leaves mm_slots
2932 * on the list for when ksmd may be set running again).
31dbd01f
IE
2933 */
2934
2935 mutex_lock(&ksm_thread_mutex);
ef4d43a8 2936 wait_while_offlining();
31dbd01f
IE
2937 if (ksm_run != flags) {
2938 ksm_run = flags;
d952b791 2939 if (flags & KSM_RUN_UNMERGE) {
e1e12d2f 2940 set_current_oom_origin();
d952b791 2941 err = unmerge_and_remove_all_rmap_items();
e1e12d2f 2942 clear_current_oom_origin();
d952b791
HD
2943 if (err) {
2944 ksm_run = KSM_RUN_STOP;
2945 count = err;
2946 }
2947 }
31dbd01f
IE
2948 }
2949 mutex_unlock(&ksm_thread_mutex);
2950
2951 if (flags & KSM_RUN_MERGE)
2952 wake_up_interruptible(&ksm_thread_wait);
2953
2954 return count;
2955}
2956KSM_ATTR(run);
2957
90bd6fd3
PH
2958#ifdef CONFIG_NUMA
2959static ssize_t merge_across_nodes_show(struct kobject *kobj,
2960 struct kobj_attribute *attr, char *buf)
2961{
2962 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2963}
2964
2965static ssize_t merge_across_nodes_store(struct kobject *kobj,
2966 struct kobj_attribute *attr,
2967 const char *buf, size_t count)
2968{
2969 int err;
2970 unsigned long knob;
2971
2972 err = kstrtoul(buf, 10, &knob);
2973 if (err)
2974 return err;
2975 if (knob > 1)
2976 return -EINVAL;
2977
2978 mutex_lock(&ksm_thread_mutex);
ef4d43a8 2979 wait_while_offlining();
90bd6fd3 2980 if (ksm_merge_across_nodes != knob) {
cbf86cfe 2981 if (ksm_pages_shared || remove_all_stable_nodes())
90bd6fd3 2982 err = -EBUSY;
ef53d16c
HD
2983 else if (root_stable_tree == one_stable_tree) {
2984 struct rb_root *buf;
2985 /*
2986 * This is the first time that we switch away from the
2987 * default of merging across nodes: must now allocate
2988 * a buffer to hold as many roots as may be needed.
2989 * Allocate stable and unstable together:
2990 * MAXSMP NODES_SHIFT 10 will use 16kB.
2991 */
bafe1e14
JP
2992 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2993 GFP_KERNEL);
ef53d16c
HD
2994 /* Let us assume that RB_ROOT is NULL is zero */
2995 if (!buf)
2996 err = -ENOMEM;
2997 else {
2998 root_stable_tree = buf;
2999 root_unstable_tree = buf + nr_node_ids;
3000 /* Stable tree is empty but not the unstable */
3001 root_unstable_tree[0] = one_unstable_tree[0];
3002 }
3003 }
3004 if (!err) {
90bd6fd3 3005 ksm_merge_across_nodes = knob;
ef53d16c
HD
3006 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
3007 }
90bd6fd3
PH
3008 }
3009 mutex_unlock(&ksm_thread_mutex);
3010
3011 return err ? err : count;
3012}
3013KSM_ATTR(merge_across_nodes);
3014#endif
3015
e86c59b1
CI
3016static ssize_t use_zero_pages_show(struct kobject *kobj,
3017 struct kobj_attribute *attr, char *buf)
3018{
3019 return sprintf(buf, "%u\n", ksm_use_zero_pages);
3020}
3021static ssize_t use_zero_pages_store(struct kobject *kobj,
3022 struct kobj_attribute *attr,
3023 const char *buf, size_t count)
3024{
3025 int err;
3026 bool value;
3027
3028 err = kstrtobool(buf, &value);
3029 if (err)
3030 return -EINVAL;
3031
3032 ksm_use_zero_pages = value;
3033
3034 return count;
3035}
3036KSM_ATTR(use_zero_pages);
3037
2c653d0e
AA
3038static ssize_t max_page_sharing_show(struct kobject *kobj,
3039 struct kobj_attribute *attr, char *buf)
3040{
3041 return sprintf(buf, "%u\n", ksm_max_page_sharing);
3042}
3043
3044static ssize_t max_page_sharing_store(struct kobject *kobj,
3045 struct kobj_attribute *attr,
3046 const char *buf, size_t count)
3047{
3048 int err;
3049 int knob;
3050
3051 err = kstrtoint(buf, 10, &knob);
3052 if (err)
3053 return err;
3054 /*
3055 * When a KSM page is created it is shared by 2 mappings. This
3056 * being a signed comparison, it implicitly verifies it's not
3057 * negative.
3058 */
3059 if (knob < 2)
3060 return -EINVAL;
3061
3062 if (READ_ONCE(ksm_max_page_sharing) == knob)
3063 return count;
3064
3065 mutex_lock(&ksm_thread_mutex);
3066 wait_while_offlining();
3067 if (ksm_max_page_sharing != knob) {
3068 if (ksm_pages_shared || remove_all_stable_nodes())
3069 err = -EBUSY;
3070 else
3071 ksm_max_page_sharing = knob;
3072 }
3073 mutex_unlock(&ksm_thread_mutex);
3074
3075 return err ? err : count;
3076}
3077KSM_ATTR(max_page_sharing);
3078
b4028260
HD
3079static ssize_t pages_shared_show(struct kobject *kobj,
3080 struct kobj_attribute *attr, char *buf)
3081{
3082 return sprintf(buf, "%lu\n", ksm_pages_shared);
3083}
3084KSM_ATTR_RO(pages_shared);
3085
3086static ssize_t pages_sharing_show(struct kobject *kobj,
3087 struct kobj_attribute *attr, char *buf)
3088{
e178dfde 3089 return sprintf(buf, "%lu\n", ksm_pages_sharing);
b4028260
HD
3090}
3091KSM_ATTR_RO(pages_sharing);
3092
473b0ce4
HD
3093static ssize_t pages_unshared_show(struct kobject *kobj,
3094 struct kobj_attribute *attr, char *buf)
3095{
3096 return sprintf(buf, "%lu\n", ksm_pages_unshared);
3097}
3098KSM_ATTR_RO(pages_unshared);
3099
3100static ssize_t pages_volatile_show(struct kobject *kobj,
3101 struct kobj_attribute *attr, char *buf)
3102{
3103 long ksm_pages_volatile;
3104
3105 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3106 - ksm_pages_sharing - ksm_pages_unshared;
3107 /*
3108 * It was not worth any locking to calculate that statistic,
3109 * but it might therefore sometimes be negative: conceal that.
3110 */
3111 if (ksm_pages_volatile < 0)
3112 ksm_pages_volatile = 0;
3113 return sprintf(buf, "%ld\n", ksm_pages_volatile);
3114}
3115KSM_ATTR_RO(pages_volatile);
3116
2c653d0e
AA
3117static ssize_t stable_node_dups_show(struct kobject *kobj,
3118 struct kobj_attribute *attr, char *buf)
3119{
3120 return sprintf(buf, "%lu\n", ksm_stable_node_dups);
3121}
3122KSM_ATTR_RO(stable_node_dups);
3123
3124static ssize_t stable_node_chains_show(struct kobject *kobj,
3125 struct kobj_attribute *attr, char *buf)
3126{
3127 return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3128}
3129KSM_ATTR_RO(stable_node_chains);
3130
3131static ssize_t
3132stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3133 struct kobj_attribute *attr,
3134 char *buf)
3135{
3136 return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3137}
3138
3139static ssize_t
3140stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3141 struct kobj_attribute *attr,
3142 const char *buf, size_t count)
3143{
3144 unsigned long msecs;
3145 int err;
3146
3147 err = kstrtoul(buf, 10, &msecs);
3148 if (err || msecs > UINT_MAX)
3149 return -EINVAL;
3150
3151 ksm_stable_node_chains_prune_millisecs = msecs;
3152
3153 return count;
3154}
3155KSM_ATTR(stable_node_chains_prune_millisecs);
3156
473b0ce4
HD
3157static ssize_t full_scans_show(struct kobject *kobj,
3158 struct kobj_attribute *attr, char *buf)
3159{
3160 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3161}
3162KSM_ATTR_RO(full_scans);
3163
31dbd01f
IE
3164static struct attribute *ksm_attrs[] = {
3165 &sleep_millisecs_attr.attr,
3166 &pages_to_scan_attr.attr,
3167 &run_attr.attr,
b4028260
HD
3168 &pages_shared_attr.attr,
3169 &pages_sharing_attr.attr,
473b0ce4
HD
3170 &pages_unshared_attr.attr,
3171 &pages_volatile_attr.attr,
3172 &full_scans_attr.attr,
90bd6fd3
PH
3173#ifdef CONFIG_NUMA
3174 &merge_across_nodes_attr.attr,
3175#endif
2c653d0e
AA
3176 &max_page_sharing_attr.attr,
3177 &stable_node_chains_attr.attr,
3178 &stable_node_dups_attr.attr,
3179 &stable_node_chains_prune_millisecs_attr.attr,
e86c59b1 3180 &use_zero_pages_attr.attr,
31dbd01f
IE
3181 NULL,
3182};
3183
f907c26a 3184static const struct attribute_group ksm_attr_group = {
31dbd01f
IE
3185 .attrs = ksm_attrs,
3186 .name = "ksm",
3187};
2ffd8679 3188#endif /* CONFIG_SYSFS */
31dbd01f
IE
3189
3190static int __init ksm_init(void)
3191{
3192 struct task_struct *ksm_thread;
3193 int err;
3194
e86c59b1
CI
3195 /* The correct value depends on page size and endianness */
3196 zero_checksum = calc_checksum(ZERO_PAGE(0));
3197 /* Default to false for backwards compatibility */
3198 ksm_use_zero_pages = false;
3199
31dbd01f
IE
3200 err = ksm_slab_init();
3201 if (err)
3202 goto out;
3203
31dbd01f
IE
3204 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3205 if (IS_ERR(ksm_thread)) {
25acde31 3206 pr_err("ksm: creating kthread failed\n");
31dbd01f 3207 err = PTR_ERR(ksm_thread);
d9f8984c 3208 goto out_free;
31dbd01f
IE
3209 }
3210
2ffd8679 3211#ifdef CONFIG_SYSFS
31dbd01f
IE
3212 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3213 if (err) {
25acde31 3214 pr_err("ksm: register sysfs failed\n");
2ffd8679 3215 kthread_stop(ksm_thread);
d9f8984c 3216 goto out_free;
31dbd01f 3217 }
c73602ad
HD
3218#else
3219 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3220
2ffd8679 3221#endif /* CONFIG_SYSFS */
31dbd01f 3222
62b61f61 3223#ifdef CONFIG_MEMORY_HOTREMOVE
ef4d43a8 3224 /* There is no significance to this priority 100 */
62b61f61
HD
3225 hotplug_memory_notifier(ksm_memory_callback, 100);
3226#endif
31dbd01f
IE
3227 return 0;
3228
d9f8984c 3229out_free:
31dbd01f
IE
3230 ksm_slab_free();
3231out:
3232 return err;
f8af4da3 3233}
a64fb3cd 3234subsys_initcall(ksm_init);