blob: 927aa34f291f47a2c0396d878a66eb6071971ab2 [file] [log] [blame]
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001/*
Izik Eidus31dbd012009-09-21 17:02:03 -07002 * Memory merging support.
3 *
4 * This code enables dynamic sharing of identical pages found in different
5 * memory areas, even if they are not shared by fork()
6 *
Izik Eidus36b25282009-09-21 17:02:06 -07007 * Copyright (C) 2008-2009 Red Hat, Inc.
Izik Eidus31dbd012009-09-21 17:02:03 -07008 * Authors:
9 * Izik Eidus
10 * Andrea Arcangeli
11 * Chris Wright
Izik Eidus36b25282009-09-21 17:02:06 -070012 * Hugh Dickins
Izik Eidus31dbd012009-09-21 17:02:03 -070013 *
14 * This work is licensed under the terms of the GNU GPL, version 2.
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070015 */
16
17#include <linux/errno.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070018#include <linux/mm.h>
19#include <linux/fs.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070020#include <linux/mman.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070021#include <linux/sched.h>
22#include <linux/rwsem.h>
23#include <linux/pagemap.h>
24#include <linux/rmap.h>
25#include <linux/spinlock.h>
26#include <linux/jhash.h>
27#include <linux/delay.h>
28#include <linux/kthread.h>
29#include <linux/wait.h>
30#include <linux/slab.h>
31#include <linux/rbtree.h>
Hugh Dickins62b61f62009-12-14 17:59:33 -080032#include <linux/memory.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070033#include <linux/mmu_notifier.h>
Izik Eidus2c6854f2009-09-23 15:56:04 -070034#include <linux/swap.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070035#include <linux/ksm.h>
Sasha Levin4ca3a692013-02-22 16:32:28 -080036#include <linux/hashtable.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080037#include <linux/freezer.h>
David Rientjes72788c32011-05-24 17:11:40 -070038#include <linux/oom.h>
Petr Holasek90bd6fd2013-02-22 16:35:00 -080039#include <linux/numa.h>
Laura Abbott2b4be352014-09-24 15:53:57 -070040#include <linux/show_mem_notifier.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070041
Izik Eidus31dbd012009-09-21 17:02:03 -070042#include <asm/tlbflush.h>
Hugh Dickins73848b42009-12-14 17:59:22 -080043#include "internal.h"
Izik Eidus31dbd012009-09-21 17:02:03 -070044
Hugh Dickinse850dcf2013-02-22 16:35:03 -080045#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
Izik Eidus31dbd012009-09-21 17:02:03 -070053/*
54 * A few notes about the KSM scanning process,
55 * to make it easier to understand the data structures below:
56 *
57 * In order to reduce excessive scanning, KSM sorts the memory pages by their
58 * contents into a data structure that holds pointers to the pages' locations.
59 *
60 * Since the contents of the pages may change at any moment, KSM cannot just
61 * insert the pages into a normal sorted tree and expect it to find anything.
62 * Therefore KSM uses two data structures - the stable and the unstable tree.
63 *
64 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
65 * by their contents. Because each such page is write-protected, searching on
66 * this tree is fully assured to be working (except when pages are unmapped),
67 * and therefore this tree is called the stable tree.
68 *
69 * In addition to the stable tree, KSM uses a second data structure called the
70 * unstable tree: this tree holds pointers to pages which have been found to
71 * be "unchanged for a period of time". The unstable tree sorts these pages
72 * by their contents, but since they are not write-protected, KSM cannot rely
73 * upon the unstable tree to work correctly - the unstable tree is liable to
74 * be corrupted as its contents are modified, and so it is called unstable.
75 *
76 * KSM solves this problem by several techniques:
77 *
78 * 1) The unstable tree is flushed every time KSM completes scanning all
79 * memory areas, and then the tree is rebuilt again from the beginning.
80 * 2) KSM will only insert into the unstable tree, pages whose hash value
81 * has not changed since the previous scan of all memory areas.
82 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
83 * colors of the nodes and not on their contents, assuring that even when
84 * the tree gets "corrupted" it won't get out of balance, so scanning time
85 * remains the same (also, searching and inserting nodes in an rbtree uses
86 * the same algorithm, so we have no overhead when we flush and rebuild).
87 * 4) KSM never flushes the stable tree, which means that even if it were to
88 * take 10 attempts to find a page in the unstable tree, once it is found,
89 * it is secured in the stable tree. (When we scan a new page, we first
90 * compare it against the stable tree, and then against the unstable tree.)
Hugh Dickins8fdb3db2013-02-22 16:36:03 -080091 *
92 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
93 * stable trees and multiple unstable trees: one of each for each NUMA node.
Izik Eidus31dbd012009-09-21 17:02:03 -070094 */
95
96/**
97 * struct mm_slot - ksm information per mm that is being scanned
98 * @link: link to the mm_slots hash list
99 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
Hugh Dickins6514d512009-12-14 17:59:19 -0800100 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
Izik Eidus31dbd012009-09-21 17:02:03 -0700101 * @mm: the mm that this information is valid for
102 */
103struct mm_slot {
104 struct hlist_node link;
105 struct list_head mm_list;
Hugh Dickins6514d512009-12-14 17:59:19 -0800106 struct rmap_item *rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700107 struct mm_struct *mm;
108};
109
110/**
111 * struct ksm_scan - cursor for scanning
112 * @mm_slot: the current mm_slot we are scanning
113 * @address: the next address inside that to be scanned
Hugh Dickins6514d512009-12-14 17:59:19 -0800114 * @rmap_list: link to the next rmap to be scanned in the rmap_list
Izik Eidus31dbd012009-09-21 17:02:03 -0700115 * @seqnr: count of completed full scans (needed when removing unstable node)
116 *
117 * There is only the one ksm_scan instance of this cursor structure.
118 */
119struct ksm_scan {
120 struct mm_slot *mm_slot;
121 unsigned long address;
Hugh Dickins6514d512009-12-14 17:59:19 -0800122 struct rmap_item **rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700123 unsigned long seqnr;
124};
125
126/**
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800127 * struct stable_node - node of the stable rbtree
128 * @node: rb node of this ksm page in the stable tree
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800129 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
130 * @list: linked into migrate_nodes, pending placement in the proper node tree
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800131 * @hlist: hlist head of rmap_items using this ksm page
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800132 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
133 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800134 */
135struct stable_node {
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800136 union {
137 struct rb_node node; /* when node of stable tree */
138 struct { /* when listed for migration */
139 struct list_head *head;
140 struct list_head list;
141 };
142 };
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800143 struct hlist_head hlist;
Hugh Dickins62b61f62009-12-14 17:59:33 -0800144 unsigned long kpfn;
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800145#ifdef CONFIG_NUMA
146 int nid;
147#endif
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800148};
149
150/**
Izik Eidus31dbd012009-09-21 17:02:03 -0700151 * struct rmap_item - reverse mapping item for virtual addresses
Hugh Dickins6514d512009-12-14 17:59:19 -0800152 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800153 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
Hugh Dickinsbc566202013-02-22 16:36:06 -0800154 * @nid: NUMA node id of unstable tree in which linked (may not match page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700155 * @mm: the memory structure this rmap_item is pointing into
156 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
157 * @oldchecksum: previous checksum of the page at that virtual address
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800158 * @node: rb node of this rmap_item in the unstable tree
159 * @head: pointer to stable_node heading this list in the stable tree
160 * @hlist: link into hlist of rmap_items hanging off that stable_node
Izik Eidus31dbd012009-09-21 17:02:03 -0700161 */
162struct rmap_item {
Hugh Dickins6514d512009-12-14 17:59:19 -0800163 struct rmap_item *rmap_list;
Hugh Dickinsbc566202013-02-22 16:36:06 -0800164 union {
165 struct anon_vma *anon_vma; /* when stable */
166#ifdef CONFIG_NUMA
167 int nid; /* when node of unstable tree */
168#endif
169 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700170 struct mm_struct *mm;
171 unsigned long address; /* + low bits used for flags below */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800172 unsigned int oldchecksum; /* when unstable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700173 union {
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800174 struct rb_node node; /* when node of unstable tree */
175 struct { /* when listed from stable tree */
176 struct stable_node *head;
177 struct hlist_node hlist;
178 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700179 };
180};
181
182#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800183#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
184#define STABLE_FLAG 0x200 /* is listed from the stable tree */
Izik Eidus31dbd012009-09-21 17:02:03 -0700185
186/* The stable and unstable tree heads */
Hugh Dickinsef53d16c2013-02-22 16:36:12 -0800187static struct rb_root one_stable_tree[1] = { RB_ROOT };
188static struct rb_root one_unstable_tree[1] = { RB_ROOT };
189static struct rb_root *root_stable_tree = one_stable_tree;
190static struct rb_root *root_unstable_tree = one_unstable_tree;
Izik Eidus31dbd012009-09-21 17:02:03 -0700191
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800192/* Recently migrated nodes of stable tree, pending proper placement */
193static LIST_HEAD(migrate_nodes);
194
Sasha Levin4ca3a692013-02-22 16:32:28 -0800195#define MM_SLOTS_HASH_BITS 10
196static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
Izik Eidus31dbd012009-09-21 17:02:03 -0700197
198static struct mm_slot ksm_mm_head = {
199 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
200};
201static struct ksm_scan ksm_scan = {
202 .mm_slot = &ksm_mm_head,
203};
204
205static struct kmem_cache *rmap_item_cache;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800206static struct kmem_cache *stable_node_cache;
Izik Eidus31dbd012009-09-21 17:02:03 -0700207static struct kmem_cache *mm_slot_cache;
208
209/* The number of nodes in the stable tree */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700210static unsigned long ksm_pages_shared;
Izik Eidus31dbd012009-09-21 17:02:03 -0700211
Hugh Dickinse178dfd2009-09-21 17:02:10 -0700212/* The number of page slots additionally sharing those nodes */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700213static unsigned long ksm_pages_sharing;
Izik Eidus31dbd012009-09-21 17:02:03 -0700214
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700215/* The number of nodes in the unstable tree */
216static unsigned long ksm_pages_unshared;
217
218/* The number of rmap_items in use: to calculate pages_volatile */
219static unsigned long ksm_rmap_items;
220
Izik Eidus31dbd012009-09-21 17:02:03 -0700221/* Number of pages ksmd should scan in one batch */
Izik Eidus2c6854f2009-09-23 15:56:04 -0700222static unsigned int ksm_thread_pages_to_scan = 100;
Izik Eidus31dbd012009-09-21 17:02:03 -0700223
224/* Milliseconds ksmd should sleep between batches */
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700225static unsigned int ksm_thread_sleep_millisecs = 20;
Izik Eidus31dbd012009-09-21 17:02:03 -0700226
Chintan Pandya65180472014-02-18 13:40:01 +0530227/* Boolean to indicate whether to use deferred timer or not */
228static bool use_deferred_timer;
229
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800230#ifdef CONFIG_NUMA
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800231/* Zeroed when merging across nodes is not allowed */
232static unsigned int ksm_merge_across_nodes = 1;
Hugh Dickinsef53d16c2013-02-22 16:36:12 -0800233static int ksm_nr_node_ids = 1;
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800234#else
235#define ksm_merge_across_nodes 1U
Hugh Dickinsef53d16c2013-02-22 16:36:12 -0800236#define ksm_nr_node_ids 1
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800237#endif
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800238
Izik Eidus31dbd012009-09-21 17:02:03 -0700239#define KSM_RUN_STOP 0
240#define KSM_RUN_MERGE 1
241#define KSM_RUN_UNMERGE 2
Hugh Dickinsef4d43a2013-02-22 16:35:16 -0800242#define KSM_RUN_OFFLINE 4
Abhimanyu Gargbd46c352014-06-10 21:26:02 -0700243static unsigned long ksm_run = KSM_RUN_MERGE;
Hugh Dickinsef4d43a2013-02-22 16:35:16 -0800244static void wait_while_offlining(void);
Izik Eidus31dbd012009-09-21 17:02:03 -0700245
246static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
247static DEFINE_MUTEX(ksm_thread_mutex);
248static DEFINE_SPINLOCK(ksm_mmlist_lock);
249
250#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
251 sizeof(struct __struct), __alignof__(struct __struct),\
252 (__flags), NULL)
253
Laura Abbott2b4be352014-09-24 15:53:57 -0700254static int ksm_show_mem_notifier(struct notifier_block *nb,
255 unsigned long action,
256 void *data)
257{
258 pr_info("ksm_pages_sharing: %lu\n", ksm_pages_sharing);
259 pr_info("ksm_pages_shared: %lu\n", ksm_pages_shared);
260
261 return 0;
262}
263
264static struct notifier_block ksm_show_mem_notifier_block = {
265 .notifier_call = ksm_show_mem_notifier,
266};
267
Izik Eidus31dbd012009-09-21 17:02:03 -0700268static int __init ksm_slab_init(void)
269{
270 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
271 if (!rmap_item_cache)
272 goto out;
273
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800274 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
275 if (!stable_node_cache)
276 goto out_free1;
277
Izik Eidus31dbd012009-09-21 17:02:03 -0700278 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
279 if (!mm_slot_cache)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800280 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -0700281
282 return 0;
283
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800284out_free2:
285 kmem_cache_destroy(stable_node_cache);
286out_free1:
Izik Eidus31dbd012009-09-21 17:02:03 -0700287 kmem_cache_destroy(rmap_item_cache);
288out:
289 return -ENOMEM;
290}
291
292static void __init ksm_slab_free(void)
293{
294 kmem_cache_destroy(mm_slot_cache);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800295 kmem_cache_destroy(stable_node_cache);
Izik Eidus31dbd012009-09-21 17:02:03 -0700296 kmem_cache_destroy(rmap_item_cache);
297 mm_slot_cache = NULL;
298}
299
300static inline struct rmap_item *alloc_rmap_item(void)
301{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700302 struct rmap_item *rmap_item;
303
zhong jiang5b398e42016-09-28 15:22:30 -0700304 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
305 __GFP_NORETRY | __GFP_NOWARN);
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700306 if (rmap_item)
307 ksm_rmap_items++;
308 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -0700309}
310
311static inline void free_rmap_item(struct rmap_item *rmap_item)
312{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700313 ksm_rmap_items--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700314 rmap_item->mm = NULL; /* debug safety */
315 kmem_cache_free(rmap_item_cache, rmap_item);
316}
317
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800318static inline struct stable_node *alloc_stable_node(void)
319{
zhong jiang62130552016-10-07 17:01:19 -0700320 /*
321 * The allocation can take too long with GFP_KERNEL when memory is under
322 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
323 * grants access to memory reserves, helping to avoid this problem.
324 */
325 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800326}
327
328static inline void free_stable_node(struct stable_node *stable_node)
329{
330 kmem_cache_free(stable_node_cache, stable_node);
331}
332
Izik Eidus31dbd012009-09-21 17:02:03 -0700333static inline struct mm_slot *alloc_mm_slot(void)
334{
335 if (!mm_slot_cache) /* initialization failed */
336 return NULL;
337 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
338}
339
340static inline void free_mm_slot(struct mm_slot *mm_slot)
341{
342 kmem_cache_free(mm_slot_cache, mm_slot);
343}
344
Izik Eidus31dbd012009-09-21 17:02:03 -0700345static struct mm_slot *get_mm_slot(struct mm_struct *mm)
346{
Sasha Levin4ca3a692013-02-22 16:32:28 -0800347 struct mm_slot *slot;
Izik Eidus31dbd012009-09-21 17:02:03 -0700348
Sasha Levinb67bfe02013-02-27 17:06:00 -0800349 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
Sasha Levin4ca3a692013-02-22 16:32:28 -0800350 if (slot->mm == mm)
351 return slot;
352
Izik Eidus31dbd012009-09-21 17:02:03 -0700353 return NULL;
354}
355
356static void insert_to_mm_slots_hash(struct mm_struct *mm,
357 struct mm_slot *mm_slot)
358{
Izik Eidus31dbd012009-09-21 17:02:03 -0700359 mm_slot->mm = mm;
Sasha Levin4ca3a692013-02-22 16:32:28 -0800360 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
Izik Eidus31dbd012009-09-21 17:02:03 -0700361}
362
Izik Eidus31dbd012009-09-21 17:02:03 -0700363/*
Hugh Dickinsa913e182009-09-21 17:02:26 -0700364 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
365 * page tables after it has passed through ksm_exit() - which, if necessary,
366 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
367 * a special flag: they can just back out as soon as mm_users goes to zero.
368 * ksm_test_exit() is used throughout to make this test for exit: in some
369 * places for correctness, in some places just to avoid unnecessary work.
370 */
371static inline bool ksm_test_exit(struct mm_struct *mm)
372{
373 return atomic_read(&mm->mm_users) == 0;
374}
375
376/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700377 * We use break_ksm to break COW on a ksm page: it's a stripped down
378 *
Dave Hansend4edcf02016-02-12 13:01:56 -0800379 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
Izik Eidus31dbd012009-09-21 17:02:03 -0700380 * put_page(page);
381 *
382 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
383 * in case the application has unmapped and remapped mm,addr meanwhile.
384 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
385 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
Dave Hansen1b2ee122016-02-12 13:02:21 -0800386 *
387 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
388 * of the process that owns 'vma'. We also do not want to enforce
389 * protection keys here anyway.
Izik Eidus31dbd012009-09-21 17:02:03 -0700390 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700391static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
Izik Eidus31dbd012009-09-21 17:02:03 -0700392{
393 struct page *page;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700394 int ret = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700395
396 do {
397 cond_resched();
Dave Hansen1b2ee122016-02-12 13:02:21 -0800398 page = follow_page(vma, addr,
399 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400400 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700401 break;
402 if (PageKsm(page))
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700403 ret = handle_mm_fault(vma, addr,
404 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
Izik Eidus31dbd012009-09-21 17:02:03 -0700405 else
406 ret = VM_FAULT_WRITE;
407 put_page(page);
Linus Torvalds33692f22015-01-29 10:51:32 -0800408 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
Hugh Dickinsd952b792009-09-21 17:02:16 -0700409 /*
410 * We must loop because handle_mm_fault() may back out if there's
411 * any difficulty e.g. if pte accessed bit gets updated concurrently.
412 *
413 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
414 * COW has been broken, even if the vma does not permit VM_WRITE;
415 * but note that a concurrent fault might break PageKsm for us.
416 *
417 * VM_FAULT_SIGBUS could occur if we race with truncation of the
418 * backing file, which also invalidates anonymous pages: that's
419 * okay, that truncation will have unmapped the PageKsm for us.
420 *
421 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
422 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
423 * current task has TIF_MEMDIE set, and will be OOM killed on return
424 * to user; and ksmd, having no mm, would never be chosen for that.
425 *
426 * But if the mm is in a limited mem_cgroup, then the fault may fail
427 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
428 * even ksmd can fail in this way - though it's usually breaking ksm
429 * just to undo a merge it made a moment before, so unlikely to oom.
430 *
431 * That's a pity: we might therefore have more kernel pages allocated
432 * than we're counting as nodes in the stable tree; but ksm_do_scan
433 * will retry to break_cow on each pass, so should recover the page
434 * in due course. The important thing is to not let VM_MERGEABLE
435 * be cleared while any such pages might remain in the area.
436 */
437 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700438}
439
Bob Liuef694222012-03-21 16:34:11 -0700440static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
441 unsigned long addr)
442{
443 struct vm_area_struct *vma;
444 if (ksm_test_exit(mm))
445 return NULL;
446 vma = find_vma(mm, addr);
447 if (!vma || vma->vm_start > addr)
448 return NULL;
449 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
450 return NULL;
451 return vma;
452}
453
Hugh Dickins8dd35572009-12-14 17:59:18 -0800454static void break_cow(struct rmap_item *rmap_item)
Izik Eidus31dbd012009-09-21 17:02:03 -0700455{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800456 struct mm_struct *mm = rmap_item->mm;
457 unsigned long addr = rmap_item->address;
Izik Eidus31dbd012009-09-21 17:02:03 -0700458 struct vm_area_struct *vma;
459
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800460 /*
461 * It is not an accident that whenever we want to break COW
462 * to undo, we also need to drop a reference to the anon_vma.
463 */
Peter Zijlstra9e601092011-03-22 16:32:46 -0700464 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800465
Hugh Dickins81464e302009-09-21 17:02:15 -0700466 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700467 vma = find_mergeable_vma(mm, addr);
468 if (vma)
469 break_ksm(vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700470 up_read(&mm->mmap_sem);
471}
472
473static struct page *get_mergeable_page(struct rmap_item *rmap_item)
474{
475 struct mm_struct *mm = rmap_item->mm;
476 unsigned long addr = rmap_item->address;
477 struct vm_area_struct *vma;
478 struct page *page;
479
480 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700481 vma = find_mergeable_vma(mm, addr);
482 if (!vma)
Izik Eidus31dbd012009-09-21 17:02:03 -0700483 goto out;
484
485 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400486 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700487 goto out;
Kirill A. Shutemovf765f542016-01-15 16:53:03 -0800488 if (PageAnon(page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700489 flush_anon_page(vma, page, addr);
490 flush_dcache_page(page);
491 } else {
492 put_page(page);
Andrea Arcangelic8f95ed2015-11-05 18:49:19 -0800493out:
494 page = NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700495 }
496 up_read(&mm->mmap_sem);
497 return page;
498}
499
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800500/*
501 * This helper is used for getting right index into array of tree roots.
502 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
503 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
504 * every node has its own stable and unstable tree.
505 */
506static inline int get_kpfn_nid(unsigned long kpfn)
507{
Hugh Dickinsd8fc16a2013-03-08 12:43:34 -0800508 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800509}
510
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800511static void remove_node_from_stable_tree(struct stable_node *stable_node)
512{
513 struct rmap_item *rmap_item;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800514
Sasha Levinb67bfe02013-02-27 17:06:00 -0800515 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800516 if (rmap_item->hlist.next)
517 ksm_pages_sharing--;
518 else
519 ksm_pages_shared--;
Peter Zijlstra9e601092011-03-22 16:32:46 -0700520 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800521 rmap_item->address &= PAGE_MASK;
522 cond_resched();
523 }
524
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800525 if (stable_node->head == &migrate_nodes)
526 list_del(&stable_node->list);
527 else
528 rb_erase(&stable_node->node,
Hugh Dickinsef53d16c2013-02-22 16:36:12 -0800529 root_stable_tree + NUMA(stable_node->nid));
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800530 free_stable_node(stable_node);
531}
532
533/*
534 * get_ksm_page: checks if the page indicated by the stable node
535 * is still its ksm page, despite having held no reference to it.
536 * In which case we can trust the content of the page, and it
537 * returns the gotten page; but if the page has now been zapped,
538 * remove the stale node from the stable tree and return NULL.
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800539 * But beware, the stable node's page might be being migrated.
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800540 *
541 * You would expect the stable_node to hold a reference to the ksm page.
542 * But if it increments the page's count, swapping out has to wait for
543 * ksmd to come around again before it can free the page, which may take
544 * seconds or even minutes: much too unresponsive. So instead we use a
545 * "keyhole reference": access to the ksm page from the stable node peeps
546 * out through its keyhole to see if that page still holds the right key,
547 * pointing back to this stable node. This relies on freeing a PageAnon
548 * page to reset its page->mapping to NULL, and relies on no other use of
549 * a page to put something that might look like our key in page->mapping.
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800550 * is on its way to being freed; but it is an anomaly to bear in mind.
551 */
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800552static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800553{
554 struct page *page;
555 void *expected_mapping;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800556 unsigned long kpfn;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800557
Minchan Kimbda807d2016-07-26 15:23:05 -0700558 expected_mapping = (void *)((unsigned long)stable_node |
559 PAGE_MAPPING_KSM);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800560again:
Jason Low4db0c3c2015-04-15 16:14:08 -0700561 kpfn = READ_ONCE(stable_node->kpfn);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800562 page = pfn_to_page(kpfn);
563
564 /*
565 * page is computed from kpfn, so on most architectures reading
566 * page->mapping is naturally ordered after reading node->kpfn,
567 * but on Alpha we need to be more careful.
568 */
569 smp_read_barrier_depends();
Jason Low4db0c3c2015-04-15 16:14:08 -0700570 if (READ_ONCE(page->mapping) != expected_mapping)
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800571 goto stale;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800572
573 /*
574 * We cannot do anything with the page while its refcount is 0.
575 * Usually 0 means free, or tail of a higher-order page: in which
576 * case this node is no longer referenced, and should be freed;
577 * however, it might mean that the page is under page_freeze_refs().
578 * The __remove_mapping() case is easy, again the node is now stale;
579 * but if page is swapcache in migrate_page_move_mapping(), it might
580 * still be our page, in which case it's essential to keep the node.
581 */
582 while (!get_page_unless_zero(page)) {
583 /*
584 * Another check for page->mapping != expected_mapping would
585 * work here too. We have chosen the !PageSwapCache test to
586 * optimize the common case, when the page is or is about to
587 * be freed: PageSwapCache is cleared (under spin_lock_irq)
588 * in the freeze_refs section of __remove_mapping(); but Anon
589 * page->mapping reset to NULL later, in free_pages_prepare().
590 */
591 if (!PageSwapCache(page))
592 goto stale;
593 cpu_relax();
594 }
595
Jason Low4db0c3c2015-04-15 16:14:08 -0700596 if (READ_ONCE(page->mapping) != expected_mapping) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800597 put_page(page);
598 goto stale;
599 }
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800600
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800601 if (lock_it) {
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800602 lock_page(page);
Jason Low4db0c3c2015-04-15 16:14:08 -0700603 if (READ_ONCE(page->mapping) != expected_mapping) {
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800604 unlock_page(page);
605 put_page(page);
606 goto stale;
607 }
608 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800609 return page;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800610
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800611stale:
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800612 /*
613 * We come here from above when page->mapping or !PageSwapCache
614 * suggests that the node is stale; but it might be under migration.
615 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
616 * before checking whether node->kpfn has been changed.
617 */
618 smp_rmb();
Jason Low4db0c3c2015-04-15 16:14:08 -0700619 if (READ_ONCE(stable_node->kpfn) != kpfn)
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800620 goto again;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800621 remove_node_from_stable_tree(stable_node);
622 return NULL;
623}
624
Izik Eidus31dbd012009-09-21 17:02:03 -0700625/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700626 * Removing rmap_item from stable or unstable tree.
627 * This function will clean the information from the stable/unstable tree.
628 */
629static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
630{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800631 if (rmap_item->address & STABLE_FLAG) {
632 struct stable_node *stable_node;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800633 struct page *page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700634
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800635 stable_node = rmap_item->head;
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800636 page = get_ksm_page(stable_node, true);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800637 if (!page)
638 goto out;
639
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800640 hlist_del(&rmap_item->hlist);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800641 unlock_page(page);
642 put_page(page);
Hugh Dickins08beca42009-12-14 17:59:21 -0800643
Andrea Arcangeli98666f8a2015-11-05 18:49:13 -0800644 if (!hlist_empty(&stable_node->hlist))
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800645 ksm_pages_sharing--;
646 else
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800647 ksm_pages_shared--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700648
Peter Zijlstra9e601092011-03-22 16:32:46 -0700649 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins93d17712009-12-14 17:59:16 -0800650 rmap_item->address &= PAGE_MASK;
Izik Eidus31dbd012009-09-21 17:02:03 -0700651
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800652 } else if (rmap_item->address & UNSTABLE_FLAG) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700653 unsigned char age;
654 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -0700655 * Usually ksmd can and must skip the rb_erase, because
Izik Eidus31dbd012009-09-21 17:02:03 -0700656 * root_unstable_tree was already reset to RB_ROOT.
Hugh Dickins9ba69292009-09-21 17:02:20 -0700657 * But be careful when an mm is exiting: do the rb_erase
658 * if this rmap_item was inserted by this scan, rather
659 * than left over from before.
Izik Eidus31dbd012009-09-21 17:02:03 -0700660 */
661 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
Hugh Dickinscd551f92009-09-21 17:02:17 -0700662 BUG_ON(age > 1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700663 if (!age)
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800664 rb_erase(&rmap_item->node,
Hugh Dickinsef53d16c2013-02-22 16:36:12 -0800665 root_unstable_tree + NUMA(rmap_item->nid));
Hugh Dickins93d17712009-12-14 17:59:16 -0800666 ksm_pages_unshared--;
667 rmap_item->address &= PAGE_MASK;
668 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800669out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700670 cond_resched(); /* we're called from many long loops */
671}
672
Izik Eidus31dbd012009-09-21 17:02:03 -0700673static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -0800674 struct rmap_item **rmap_list)
Izik Eidus31dbd012009-09-21 17:02:03 -0700675{
Hugh Dickins6514d512009-12-14 17:59:19 -0800676 while (*rmap_list) {
677 struct rmap_item *rmap_item = *rmap_list;
678 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700679 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700680 free_rmap_item(rmap_item);
681 }
682}
683
684/*
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800685 * Though it's very tempting to unmerge rmap_items from stable tree rather
Izik Eidus31dbd012009-09-21 17:02:03 -0700686 * than check every pte of a given vma, the locking doesn't quite work for
687 * that - an rmap_item is assigned to the stable tree after inserting ksm
688 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
689 * rmap_items from parent to child at fork time (so as not to waste time
690 * if exit comes before the next scan reaches it).
Hugh Dickins81464e302009-09-21 17:02:15 -0700691 *
692 * Similarly, although we'd like to remove rmap_items (so updating counts
693 * and freeing memory) when unmerging an area, it's easier to leave that
694 * to the next pass of ksmd - consider, for example, how ksmd might be
695 * in cmp_and_merge_page on one of the rmap_items we would be removing.
Izik Eidus31dbd012009-09-21 17:02:03 -0700696 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700697static int unmerge_ksm_pages(struct vm_area_struct *vma,
698 unsigned long start, unsigned long end)
Izik Eidus31dbd012009-09-21 17:02:03 -0700699{
700 unsigned long addr;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700701 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700702
Hugh Dickinsd952b792009-09-21 17:02:16 -0700703 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700704 if (ksm_test_exit(vma->vm_mm))
705 break;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700706 if (signal_pending(current))
707 err = -ERESTARTSYS;
708 else
709 err = break_ksm(vma, addr);
710 }
711 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700712}
713
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700714#ifdef CONFIG_SYSFS
715/*
716 * Only called through the sysfs control interface:
717 */
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800718static int remove_stable_node(struct stable_node *stable_node)
719{
720 struct page *page;
721 int err;
722
723 page = get_ksm_page(stable_node, true);
724 if (!page) {
725 /*
726 * get_ksm_page did remove_node_from_stable_tree itself.
727 */
728 return 0;
729 }
730
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800731 if (WARN_ON_ONCE(page_mapped(page))) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800732 /*
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800733 * This should not happen: but if it does, just refuse to let
734 * merge_across_nodes be switched - there is no need to panic.
735 */
736 err = -EBUSY;
737 } else {
738 /*
739 * The stable node did not yet appear stale to get_ksm_page(),
740 * since that allows for an unmapped ksm page to be recognized
741 * right up until it is freed; but the node is safe to remove.
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800742 * This page might be in a pagevec waiting to be freed,
743 * or it might be PageSwapCache (perhaps under writeback),
744 * or it might have been removed from swapcache a moment ago.
745 */
746 set_page_stable_node(page, NULL);
747 remove_node_from_stable_tree(stable_node);
748 err = 0;
749 }
750
751 unlock_page(page);
752 put_page(page);
753 return err;
754}
755
756static int remove_all_stable_nodes(void)
757{
Geliang Tang03640412016-01-14 15:20:54 -0800758 struct stable_node *stable_node, *next;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800759 int nid;
760 int err = 0;
761
Hugh Dickinsef53d16c2013-02-22 16:36:12 -0800762 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800763 while (root_stable_tree[nid].rb_node) {
764 stable_node = rb_entry(root_stable_tree[nid].rb_node,
765 struct stable_node, node);
766 if (remove_stable_node(stable_node)) {
767 err = -EBUSY;
768 break; /* proceed to next nid */
769 }
770 cond_resched();
771 }
772 }
Geliang Tang03640412016-01-14 15:20:54 -0800773 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800774 if (remove_stable_node(stable_node))
775 err = -EBUSY;
776 cond_resched();
777 }
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800778 return err;
779}
780
Hugh Dickinsd952b792009-09-21 17:02:16 -0700781static int unmerge_and_remove_all_rmap_items(void)
Izik Eidus31dbd012009-09-21 17:02:03 -0700782{
783 struct mm_slot *mm_slot;
784 struct mm_struct *mm;
785 struct vm_area_struct *vma;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700786 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700787
Hugh Dickinsd952b792009-09-21 17:02:16 -0700788 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700789 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700790 struct mm_slot, mm_list);
791 spin_unlock(&ksm_mmlist_lock);
792
Hugh Dickins9ba69292009-09-21 17:02:20 -0700793 for (mm_slot = ksm_scan.mm_slot;
794 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700795 mm = mm_slot->mm;
796 down_read(&mm->mmap_sem);
797 for (vma = mm->mmap; vma; vma = vma->vm_next) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700798 if (ksm_test_exit(mm))
799 break;
Izik Eidus31dbd012009-09-21 17:02:03 -0700800 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
801 continue;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700802 err = unmerge_ksm_pages(vma,
803 vma->vm_start, vma->vm_end);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700804 if (err)
805 goto error;
Izik Eidus31dbd012009-09-21 17:02:03 -0700806 }
Hugh Dickins9ba69292009-09-21 17:02:20 -0700807
Hugh Dickins6514d512009-12-14 17:59:19 -0800808 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
Zhou Chengming7496fea2016-05-12 15:42:21 -0700809 up_read(&mm->mmap_sem);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700810
811 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700812 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700813 struct mm_slot, mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700814 if (ksm_test_exit(mm)) {
Sasha Levin4ca3a692013-02-22 16:32:28 -0800815 hash_del(&mm_slot->link);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700816 list_del(&mm_slot->mm_list);
817 spin_unlock(&ksm_mmlist_lock);
818
819 free_mm_slot(mm_slot);
820 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700821 mmdrop(mm);
Zhou Chengming7496fea2016-05-12 15:42:21 -0700822 } else
Hugh Dickins9ba69292009-09-21 17:02:20 -0700823 spin_unlock(&ksm_mmlist_lock);
Izik Eidus31dbd012009-09-21 17:02:03 -0700824 }
825
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800826 /* Clean up stable nodes, but don't worry if some are still busy */
827 remove_all_stable_nodes();
Hugh Dickinsd952b792009-09-21 17:02:16 -0700828 ksm_scan.seqnr = 0;
Hugh Dickins9ba69292009-09-21 17:02:20 -0700829 return 0;
830
831error:
832 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700833 spin_lock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700834 ksm_scan.mm_slot = &ksm_mm_head;
Izik Eidus31dbd012009-09-21 17:02:03 -0700835 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700836 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700837}
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700838#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -0700839
Izik Eidus31dbd012009-09-21 17:02:03 -0700840static u32 calc_checksum(struct page *page)
841{
842 u32 checksum;
Cong Wang9b04c5f2011-11-25 23:14:39 +0800843 void *addr = kmap_atomic(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700844 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
Cong Wang9b04c5f2011-11-25 23:14:39 +0800845 kunmap_atomic(addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700846 return checksum;
847}
848
849static int memcmp_pages(struct page *page1, struct page *page2)
850{
851 char *addr1, *addr2;
852 int ret;
853
Cong Wang9b04c5f2011-11-25 23:14:39 +0800854 addr1 = kmap_atomic(page1);
855 addr2 = kmap_atomic(page2);
Izik Eidus31dbd012009-09-21 17:02:03 -0700856 ret = memcmp(addr1, addr2, PAGE_SIZE);
Cong Wang9b04c5f2011-11-25 23:14:39 +0800857 kunmap_atomic(addr2);
858 kunmap_atomic(addr1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700859 return ret;
860}
861
862static inline int pages_identical(struct page *page1, struct page *page2)
863{
864 return !memcmp_pages(page1, page2);
865}
866
867static int write_protect_page(struct vm_area_struct *vma, struct page *page,
868 pte_t *orig_pte)
869{
870 struct mm_struct *mm = vma->vm_mm;
871 unsigned long addr;
872 pte_t *ptep;
873 spinlock_t *ptl;
874 int swapped;
875 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700876 unsigned long mmun_start; /* For mmu_notifiers */
877 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -0700878
879 addr = page_address_in_vma(page, vma);
880 if (addr == -EFAULT)
881 goto out;
882
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800883 BUG_ON(PageTransCompound(page));
Haggai Eran6bdb9132012-10-08 16:33:35 -0700884
885 mmun_start = addr;
886 mmun_end = addr + PAGE_SIZE;
887 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
888
Izik Eidus31dbd012009-09-21 17:02:03 -0700889 ptep = page_check_address(page, mm, addr, &ptl, 0);
890 if (!ptep)
Haggai Eran6bdb9132012-10-08 16:33:35 -0700891 goto out_mn;
Izik Eidus31dbd012009-09-21 17:02:03 -0700892
Hugh Dickins4e316352010-10-02 17:49:08 -0700893 if (pte_write(*ptep) || pte_dirty(*ptep)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700894 pte_t entry;
895
896 swapped = PageSwapCache(page);
897 flush_cache_page(vma, addr, page_to_pfn(page));
898 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300899 * Ok this is tricky, when get_user_pages_fast() run it doesn't
Izik Eidus31dbd012009-09-21 17:02:03 -0700900 * take any lock, therefore the check that we are going to make
901 * with the pagecount against the mapcount is racey and
902 * O_DIRECT can happen right after the check.
903 * So we clear the pte and flush the tlb before the check
904 * this assure us that no O_DIRECT can happen after the check
905 * or in the middle of the check.
906 */
Joerg Roedel34ee6452014-11-13 13:46:09 +1100907 entry = ptep_clear_flush_notify(vma, addr, ptep);
Izik Eidus31dbd012009-09-21 17:02:03 -0700908 /*
909 * Check that no O_DIRECT or similar I/O is in progress on the
910 * page
911 */
Hugh Dickins31e855e2009-12-14 17:59:17 -0800912 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
Robin Holtcb532372010-03-23 13:35:26 -0700913 set_pte_at(mm, addr, ptep, entry);
Izik Eidus31dbd012009-09-21 17:02:03 -0700914 goto out_unlock;
915 }
Hugh Dickins4e316352010-10-02 17:49:08 -0700916 if (pte_dirty(entry))
917 set_page_dirty(page);
918 entry = pte_mkclean(pte_wrprotect(entry));
Izik Eidus31dbd012009-09-21 17:02:03 -0700919 set_pte_at_notify(mm, addr, ptep, entry);
920 }
921 *orig_pte = *ptep;
922 err = 0;
923
924out_unlock:
925 pte_unmap_unlock(ptep, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -0700926out_mn:
927 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -0700928out:
929 return err;
930}
931
932/**
933 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -0800934 * @vma: vma that holds the pte pointing to page
935 * @page: the page we are replacing by kpage
936 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -0700937 * @orig_pte: the original value of the pte
938 *
939 * Returns 0 on success, -EFAULT on failure.
940 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800941static int replace_page(struct vm_area_struct *vma, struct page *page,
942 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -0700943{
944 struct mm_struct *mm = vma->vm_mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700945 pmd_t *pmd;
946 pte_t *ptep;
947 spinlock_t *ptl;
948 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -0700949 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700950 unsigned long mmun_start; /* For mmu_notifiers */
951 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -0700952
Hugh Dickins8dd35572009-12-14 17:59:18 -0800953 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -0700954 if (addr == -EFAULT)
955 goto out;
956
Bob Liu62190492012-12-11 16:00:37 -0800957 pmd = mm_find_pmd(mm, addr);
958 if (!pmd)
Izik Eidus31dbd012009-09-21 17:02:03 -0700959 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700960
Haggai Eran6bdb9132012-10-08 16:33:35 -0700961 mmun_start = addr;
962 mmun_end = addr + PAGE_SIZE;
963 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
964
Izik Eidus31dbd012009-09-21 17:02:03 -0700965 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
966 if (!pte_same(*ptep, orig_pte)) {
967 pte_unmap_unlock(ptep, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -0700968 goto out_mn;
Izik Eidus31dbd012009-09-21 17:02:03 -0700969 }
970
Hugh Dickins8dd35572009-12-14 17:59:18 -0800971 get_page(kpage);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800972 page_add_anon_rmap(kpage, vma, addr, false);
Izik Eidus31dbd012009-09-21 17:02:03 -0700973
974 flush_cache_page(vma, addr, pte_pfn(*ptep));
Joerg Roedel34ee6452014-11-13 13:46:09 +1100975 ptep_clear_flush_notify(vma, addr, ptep);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800976 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
Izik Eidus31dbd012009-09-21 17:02:03 -0700977
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800978 page_remove_rmap(page, false);
Hugh Dickinsae52a2a2011-01-13 15:46:28 -0800979 if (!page_mapped(page))
980 try_to_free_swap(page);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800981 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700982
983 pte_unmap_unlock(ptep, ptl);
984 err = 0;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700985out_mn:
986 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -0700987out:
988 return err;
989}
990
991/*
992 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -0800993 * @vma: the vma that holds the pte pointing to page
994 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins80e148222009-12-14 17:59:29 -0800995 * @kpage: the PageKsm page that we want to map instead of page,
996 * or NULL the first time when we want to use page as kpage.
Izik Eidus31dbd012009-09-21 17:02:03 -0700997 *
998 * This function returns 0 if the pages were merged, -EFAULT otherwise.
999 */
1000static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -08001001 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001002{
1003 pte_t orig_pte = __pte(0);
1004 int err = -EFAULT;
1005
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001006 if (page == kpage) /* ksm page forked */
1007 return 0;
1008
Hugh Dickins8dd35572009-12-14 17:59:18 -08001009 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -07001010 goto out;
1011
Izik Eidus31dbd012009-09-21 17:02:03 -07001012 /*
1013 * We need the page lock to read a stable PageSwapCache in
1014 * write_protect_page(). We use trylock_page() instead of
1015 * lock_page() because we don't want to wait here - we
1016 * prefer to continue scanning and merging different pages,
1017 * then come back to this page when it is unlocked.
1018 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001019 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -08001020 goto out;
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001021
1022 if (PageTransCompound(page)) {
Andrea Arcangeli873f3b02017-06-02 14:46:11 -07001023 if (split_huge_page(page))
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001024 goto out_unlock;
1025 }
1026
Izik Eidus31dbd012009-09-21 17:02:03 -07001027 /*
1028 * If this anonymous page is mapped only here, its pte may need
1029 * to be write-protected. If it's mapped elsewhere, all of its
1030 * ptes are necessarily already write-protected. But in either
1031 * case, we need to lock and check page_count is not raised.
1032 */
Hugh Dickins80e148222009-12-14 17:59:29 -08001033 if (write_protect_page(vma, page, &orig_pte) == 0) {
1034 if (!kpage) {
1035 /*
1036 * While we hold page lock, upgrade page from
1037 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1038 * stable_tree_insert() will update stable_node.
1039 */
1040 set_page_stable_node(page, NULL);
1041 mark_page_accessed(page);
Minchan Kim337ed7e2016-01-15 16:55:15 -08001042 /*
1043 * Page reclaim just frees a clean page with no dirty
1044 * ptes: make sure that the ksm page would be swapped.
1045 */
1046 if (!PageDirty(page))
1047 SetPageDirty(page);
Hugh Dickins80e148222009-12-14 17:59:29 -08001048 err = 0;
1049 } else if (pages_identical(page, kpage))
1050 err = replace_page(vma, page, kpage, orig_pte);
1051 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001052
Hugh Dickins80e148222009-12-14 17:59:29 -08001053 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
Hugh Dickins73848b42009-12-14 17:59:22 -08001054 munlock_vma_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001055 if (!PageMlocked(kpage)) {
1056 unlock_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001057 lock_page(kpage);
1058 mlock_vma_page(kpage);
1059 page = kpage; /* for final unlock */
1060 }
1061 }
Hugh Dickins73848b42009-12-14 17:59:22 -08001062
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001063out_unlock:
Hugh Dickins8dd35572009-12-14 17:59:18 -08001064 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001065out:
1066 return err;
1067}
1068
1069/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001070 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1071 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -08001072 *
1073 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -07001074 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001075static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1076 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -07001077{
Hugh Dickins8dd35572009-12-14 17:59:18 -08001078 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -07001079 struct vm_area_struct *vma;
1080 int err = -EFAULT;
1081
Hugh Dickins8dd35572009-12-14 17:59:18 -08001082 down_read(&mm->mmap_sem);
Andrea Arcangeli85c6e8d2015-11-05 18:49:16 -08001083 vma = find_mergeable_vma(mm, rmap_item->address);
1084 if (!vma)
Hugh Dickins9ba69292009-09-21 17:02:20 -07001085 goto out;
1086
Hugh Dickins8dd35572009-12-14 17:59:18 -08001087 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001088 if (err)
1089 goto out;
1090
Hugh Dickinsbc566202013-02-22 16:36:06 -08001091 /* Unstable nid is in union with stable anon_vma: remove first */
1092 remove_rmap_item_from_tree(rmap_item);
1093
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001094 /* Must get reference to anon_vma while still holding mmap_sem */
Peter Zijlstra9e601092011-03-22 16:32:46 -07001095 rmap_item->anon_vma = vma->anon_vma;
1096 get_anon_vma(vma->anon_vma);
Hugh Dickins81464e302009-09-21 17:02:15 -07001097out:
Hugh Dickins8dd35572009-12-14 17:59:18 -08001098 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -07001099 return err;
1100}
1101
1102/*
Izik Eidus31dbd012009-09-21 17:02:03 -07001103 * try_to_merge_two_pages - take two identical pages and prepare them
1104 * to be merged into one page.
1105 *
Hugh Dickins8dd35572009-12-14 17:59:18 -08001106 * This function returns the kpage if we successfully merged two identical
1107 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001108 *
Hugh Dickins80e148222009-12-14 17:59:29 -08001109 * Note that this function upgrades page to ksm page: if one of the pages
Izik Eidus31dbd012009-09-21 17:02:03 -07001110 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1111 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001112static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1113 struct page *page,
1114 struct rmap_item *tree_rmap_item,
1115 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001116{
Hugh Dickins80e148222009-12-14 17:59:29 -08001117 int err;
Izik Eidus31dbd012009-09-21 17:02:03 -07001118
Hugh Dickins80e148222009-12-14 17:59:29 -08001119 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
Izik Eidus31dbd012009-09-21 17:02:03 -07001120 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001121 err = try_to_merge_with_ksm_page(tree_rmap_item,
Hugh Dickins80e148222009-12-14 17:59:29 -08001122 tree_page, page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001123 /*
Hugh Dickins81464e302009-09-21 17:02:15 -07001124 * If that fails, we have a ksm page with only one pte
1125 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -07001126 */
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001127 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -08001128 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001129 }
Hugh Dickins80e148222009-12-14 17:59:29 -08001130 return err ? NULL : page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001131}
1132
1133/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001134 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -07001135 *
1136 * This function checks if there is a page inside the stable tree
1137 * with identical content to the page that we are scanning right now.
1138 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001139 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -07001140 * NULL otherwise.
1141 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001142static struct page *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001143{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001144 int nid;
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001145 struct rb_root *root;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001146 struct rb_node **new;
1147 struct rb_node *parent;
1148 struct stable_node *stable_node;
1149 struct stable_node *page_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001150
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001151 page_node = page_stable_node(page);
1152 if (page_node && page_node->head != &migrate_nodes) {
1153 /* ksm page forked */
Hugh Dickins08beca42009-12-14 17:59:21 -08001154 get_page(page);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001155 return page;
Hugh Dickins08beca42009-12-14 17:59:21 -08001156 }
1157
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001158 nid = get_kpfn_nid(page_to_pfn(page));
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001159 root = root_stable_tree + nid;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001160again:
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001161 new = &root->rb_node;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001162 parent = NULL;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001163
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001164 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001165 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001166 int ret;
1167
Hugh Dickins08beca42009-12-14 17:59:21 -08001168 cond_resched();
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001169 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins8aafa6a2013-02-22 16:35:06 -08001170 tree_page = get_ksm_page(stable_node, false);
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001171 if (!tree_page) {
1172 /*
1173 * If we walked over a stale stable_node,
1174 * get_ksm_page() will call rb_erase() and it
1175 * may rebalance the tree from under us. So
1176 * restart the search from scratch. Returning
1177 * NULL would be safe too, but we'd generate
1178 * false negative insertions just because some
1179 * stable_node was stale.
1180 */
1181 goto again;
1182 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001183
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001184 ret = memcmp_pages(page, tree_page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001185 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001186
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001187 parent = *new;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001188 if (ret < 0)
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001189 new = &parent->rb_left;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001190 else if (ret > 0)
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001191 new = &parent->rb_right;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001192 else {
1193 /*
1194 * Lock and unlock the stable_node's page (which
1195 * might already have been migrated) so that page
1196 * migration is sure to notice its raised count.
1197 * It would be more elegant to return stable_node
1198 * than kpage, but that involves more changes.
1199 */
1200 tree_page = get_ksm_page(stable_node, true);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001201 if (tree_page) {
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001202 unlock_page(tree_page);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001203 if (get_kpfn_nid(stable_node->kpfn) !=
1204 NUMA(stable_node->nid)) {
1205 put_page(tree_page);
1206 goto replace;
1207 }
1208 return tree_page;
1209 }
1210 /*
1211 * There is now a place for page_node, but the tree may
1212 * have been rebalanced, so re-evaluate parent and new.
1213 */
1214 if (page_node)
1215 goto again;
1216 return NULL;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001217 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001218 }
1219
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001220 if (!page_node)
1221 return NULL;
1222
1223 list_del(&page_node->list);
1224 DO_NUMA(page_node->nid = nid);
1225 rb_link_node(&page_node->node, parent, new);
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001226 rb_insert_color(&page_node->node, root);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001227 get_page(page);
1228 return page;
1229
1230replace:
1231 if (page_node) {
1232 list_del(&page_node->list);
1233 DO_NUMA(page_node->nid = nid);
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001234 rb_replace_node(&stable_node->node, &page_node->node, root);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001235 get_page(page);
1236 } else {
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001237 rb_erase(&stable_node->node, root);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001238 page = NULL;
1239 }
1240 stable_node->head = &migrate_nodes;
1241 list_add(&stable_node->list, stable_node->head);
1242 return page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001243}
1244
1245/*
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001246 * stable_tree_insert - insert stable tree node pointing to new ksm page
Izik Eidus31dbd012009-09-21 17:02:03 -07001247 * into the stable tree.
1248 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001249 * This function returns the stable tree node just allocated on success,
1250 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001251 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001252static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001253{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001254 int nid;
1255 unsigned long kpfn;
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001256 struct rb_root *root;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001257 struct rb_node **new;
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001258 struct rb_node *parent;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001259 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001260
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001261 kpfn = page_to_pfn(kpage);
1262 nid = get_kpfn_nid(kpfn);
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001263 root = root_stable_tree + nid;
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001264again:
1265 parent = NULL;
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001266 new = &root->rb_node;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001267
Izik Eidus31dbd012009-09-21 17:02:03 -07001268 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001269 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001270 int ret;
1271
Hugh Dickins08beca42009-12-14 17:59:21 -08001272 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001273 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins8aafa6a2013-02-22 16:35:06 -08001274 tree_page = get_ksm_page(stable_node, false);
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001275 if (!tree_page) {
1276 /*
1277 * If we walked over a stale stable_node,
1278 * get_ksm_page() will call rb_erase() and it
1279 * may rebalance the tree from under us. So
1280 * restart the search from scratch. Returning
1281 * NULL would be safe too, but we'd generate
1282 * false negative insertions just because some
1283 * stable_node was stale.
1284 */
1285 goto again;
1286 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001287
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001288 ret = memcmp_pages(kpage, tree_page);
1289 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001290
1291 parent = *new;
1292 if (ret < 0)
1293 new = &parent->rb_left;
1294 else if (ret > 0)
1295 new = &parent->rb_right;
1296 else {
1297 /*
1298 * It is not a bug that stable_tree_search() didn't
1299 * find this node: because at that time our page was
1300 * not yet write-protected, so may have changed since.
1301 */
1302 return NULL;
1303 }
1304 }
1305
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001306 stable_node = alloc_stable_node();
1307 if (!stable_node)
1308 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001309
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001310 INIT_HLIST_HEAD(&stable_node->hlist);
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001311 stable_node->kpfn = kpfn;
Hugh Dickins08beca42009-12-14 17:59:21 -08001312 set_page_stable_node(kpage, stable_node);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001313 DO_NUMA(stable_node->nid = nid);
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001314 rb_link_node(&stable_node->node, parent, new);
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001315 rb_insert_color(&stable_node->node, root);
Hugh Dickins08beca42009-12-14 17:59:21 -08001316
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001317 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001318}
1319
1320/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001321 * unstable_tree_search_insert - search for identical page,
1322 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001323 *
1324 * This function searches for a page in the unstable tree identical to the
1325 * page currently being scanned; and if no identical page is found in the
1326 * tree, we insert rmap_item as a new object into the unstable tree.
1327 *
1328 * This function returns pointer to rmap_item found to be identical
1329 * to the currently scanned page, NULL otherwise.
1330 *
1331 * This function does both searching and inserting, because they share
1332 * the same walking algorithm in an rbtree.
1333 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001334static
1335struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1336 struct page *page,
1337 struct page **tree_pagep)
Izik Eidus31dbd012009-09-21 17:02:03 -07001338{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001339 struct rb_node **new;
1340 struct rb_root *root;
Izik Eidus31dbd012009-09-21 17:02:03 -07001341 struct rb_node *parent = NULL;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001342 int nid;
1343
1344 nid = get_kpfn_nid(page_to_pfn(page));
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001345 root = root_unstable_tree + nid;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001346 new = &root->rb_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001347
1348 while (*new) {
1349 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001350 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001351 int ret;
1352
Hugh Dickinsd178f272009-11-09 15:58:23 +00001353 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -07001354 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001355 tree_page = get_mergeable_page(tree_rmap_item);
Andrea Arcangelic8f95ed2015-11-05 18:49:19 -08001356 if (!tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001357 return NULL;
1358
1359 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001360 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -07001361 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001362 if (page == tree_page) {
1363 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001364 return NULL;
1365 }
1366
Hugh Dickins8dd35572009-12-14 17:59:18 -08001367 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001368
1369 parent = *new;
1370 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001371 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001372 new = &parent->rb_left;
1373 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001374 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001375 new = &parent->rb_right;
Hugh Dickinsb599cbd2013-02-22 16:36:05 -08001376 } else if (!ksm_merge_across_nodes &&
1377 page_to_nid(tree_page) != nid) {
1378 /*
1379 * If tree_page has been migrated to another NUMA node,
1380 * it will be flushed out and put in the right unstable
1381 * tree next time: only merge with it when across_nodes.
1382 */
1383 put_page(tree_page);
1384 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001385 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001386 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001387 return tree_rmap_item;
1388 }
1389 }
1390
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001391 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001392 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001393 DO_NUMA(rmap_item->nid = nid);
Izik Eidus31dbd012009-09-21 17:02:03 -07001394 rb_link_node(&rmap_item->node, parent, new);
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001395 rb_insert_color(&rmap_item->node, root);
Izik Eidus31dbd012009-09-21 17:02:03 -07001396
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001397 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001398 return NULL;
1399}
1400
1401/*
1402 * stable_tree_append - add another rmap_item to the linked list of
1403 * rmap_items hanging off a given node of the stable tree, all sharing
1404 * the same ksm page.
1405 */
1406static void stable_tree_append(struct rmap_item *rmap_item,
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001407 struct stable_node *stable_node)
Izik Eidus31dbd012009-09-21 17:02:03 -07001408{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001409 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001410 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001411 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001412
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001413 if (rmap_item->hlist.next)
1414 ksm_pages_sharing++;
1415 else
1416 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001417}
1418
1419/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001420 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1421 * if not, compare checksum to previous and if it's the same, see if page can
1422 * be inserted into the unstable tree, or merged with a page already there and
1423 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001424 *
1425 * @page: the page that we are searching identical page to.
1426 * @rmap_item: the reverse mapping into the virtual address of this page
1427 */
1428static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1429{
Izik Eidus31dbd012009-09-21 17:02:03 -07001430 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001431 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001432 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001433 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07001434 unsigned int checksum;
1435 int err;
1436
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001437 stable_node = page_stable_node(page);
1438 if (stable_node) {
1439 if (stable_node->head != &migrate_nodes &&
1440 get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1441 rb_erase(&stable_node->node,
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001442 root_stable_tree + NUMA(stable_node->nid));
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001443 stable_node->head = &migrate_nodes;
1444 list_add(&stable_node->list, stable_node->head);
1445 }
1446 if (stable_node->head != &migrate_nodes &&
1447 rmap_item->head == stable_node)
1448 return;
1449 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001450
1451 /* We first start with searching the page inside the stable tree */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001452 kpage = stable_tree_search(page);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001453 if (kpage == page && rmap_item->head == stable_node) {
1454 put_page(kpage);
1455 return;
1456 }
1457
1458 remove_rmap_item_from_tree(rmap_item);
1459
Hugh Dickins62b61f62009-12-14 17:59:33 -08001460 if (kpage) {
Hugh Dickins08beca42009-12-14 17:59:21 -08001461 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001462 if (!err) {
1463 /*
1464 * The page was successfully merged:
1465 * add its rmap_item to the stable tree.
1466 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001467 lock_page(kpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001468 stable_tree_append(rmap_item, page_stable_node(kpage));
Hugh Dickins5ad64682009-12-14 17:59:24 -08001469 unlock_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001470 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001471 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001472 return;
1473 }
1474
1475 /*
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001476 * If the hash value of the page has changed from the last time
1477 * we calculated it, this page is changing frequently: therefore we
1478 * don't want to insert it in the unstable tree, and we don't want
1479 * to waste our time searching for something identical to it there.
Izik Eidus31dbd012009-09-21 17:02:03 -07001480 */
1481 checksum = calc_checksum(page);
1482 if (rmap_item->oldchecksum != checksum) {
1483 rmap_item->oldchecksum = checksum;
1484 return;
1485 }
1486
Hugh Dickins8dd35572009-12-14 17:59:18 -08001487 tree_rmap_item =
1488 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001489 if (tree_rmap_item) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001490 kpage = try_to_merge_two_pages(rmap_item, page,
1491 tree_rmap_item, tree_page);
1492 put_page(tree_page);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001493 if (kpage) {
Hugh Dickinsbc566202013-02-22 16:36:06 -08001494 /*
1495 * The pages were successfully merged: insert new
1496 * node in the stable tree and add both rmap_items.
1497 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001498 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001499 stable_node = stable_tree_insert(kpage);
1500 if (stable_node) {
1501 stable_tree_append(tree_rmap_item, stable_node);
1502 stable_tree_append(rmap_item, stable_node);
1503 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001504 unlock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001505
Izik Eidus31dbd012009-09-21 17:02:03 -07001506 /*
1507 * If we fail to insert the page into the stable tree,
1508 * we will have 2 virtual addresses that are pointing
1509 * to a ksm page left outside the stable tree,
1510 * in which case we need to break_cow on both.
1511 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001512 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001513 break_cow(tree_rmap_item);
1514 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001515 }
1516 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001517 }
1518}
1519
1520static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001521 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07001522 unsigned long addr)
1523{
1524 struct rmap_item *rmap_item;
1525
Hugh Dickins6514d512009-12-14 17:59:19 -08001526 while (*rmap_list) {
1527 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08001528 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07001529 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001530 if (rmap_item->address > addr)
1531 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08001532 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001533 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001534 free_rmap_item(rmap_item);
1535 }
1536
1537 rmap_item = alloc_rmap_item();
1538 if (rmap_item) {
1539 /* It has already been zeroed */
1540 rmap_item->mm = mm_slot->mm;
1541 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08001542 rmap_item->rmap_list = *rmap_list;
1543 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001544 }
1545 return rmap_item;
1546}
1547
1548static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1549{
1550 struct mm_struct *mm;
1551 struct mm_slot *slot;
1552 struct vm_area_struct *vma;
1553 struct rmap_item *rmap_item;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001554 int nid;
Izik Eidus31dbd012009-09-21 17:02:03 -07001555
1556 if (list_empty(&ksm_mm_head.mm_list))
1557 return NULL;
1558
1559 slot = ksm_scan.mm_slot;
1560 if (slot == &ksm_mm_head) {
Hugh Dickins2919bfd2011-01-13 15:47:29 -08001561 /*
1562 * A number of pages can hang around indefinitely on per-cpu
1563 * pagevecs, raised page count preventing write_protect_page
1564 * from merging them. Though it doesn't really matter much,
1565 * it is puzzling to see some stuck in pages_volatile until
1566 * other activity jostles them out, and they also prevented
1567 * LTP's KSM test from succeeding deterministically; so drain
1568 * them here (here rather than on entry to ksm_do_scan(),
1569 * so we don't IPI too often when pages_to_scan is set low).
1570 */
1571 lru_add_drain_all();
1572
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001573 /*
1574 * Whereas stale stable_nodes on the stable_tree itself
1575 * get pruned in the regular course of stable_tree_search(),
1576 * those moved out to the migrate_nodes list can accumulate:
1577 * so prune them once before each full scan.
1578 */
1579 if (!ksm_merge_across_nodes) {
Geliang Tang03640412016-01-14 15:20:54 -08001580 struct stable_node *stable_node, *next;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001581 struct page *page;
1582
Geliang Tang03640412016-01-14 15:20:54 -08001583 list_for_each_entry_safe(stable_node, next,
1584 &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001585 page = get_ksm_page(stable_node, false);
1586 if (page)
1587 put_page(page);
1588 cond_resched();
1589 }
1590 }
1591
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08001592 for (nid = 0; nid < ksm_nr_node_ids; nid++)
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001593 root_unstable_tree[nid] = RB_ROOT;
Izik Eidus31dbd012009-09-21 17:02:03 -07001594
1595 spin_lock(&ksm_mmlist_lock);
1596 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1597 ksm_scan.mm_slot = slot;
1598 spin_unlock(&ksm_mmlist_lock);
Hugh Dickins2b472612011-06-15 15:08:58 -07001599 /*
1600 * Although we tested list_empty() above, a racing __ksm_exit
1601 * of the last mm on the list may have removed it since then.
1602 */
1603 if (slot == &ksm_mm_head)
1604 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001605next_mm:
1606 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001607 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001608 }
1609
1610 mm = slot->mm;
1611 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001612 if (ksm_test_exit(mm))
1613 vma = NULL;
1614 else
1615 vma = find_vma(mm, ksm_scan.address);
1616
1617 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001618 if (!(vma->vm_flags & VM_MERGEABLE))
1619 continue;
1620 if (ksm_scan.address < vma->vm_start)
1621 ksm_scan.address = vma->vm_start;
1622 if (!vma->anon_vma)
1623 ksm_scan.address = vma->vm_end;
1624
1625 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001626 if (ksm_test_exit(mm))
1627 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001628 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001629 if (IS_ERR_OR_NULL(*page)) {
1630 ksm_scan.address += PAGE_SIZE;
1631 cond_resched();
1632 continue;
1633 }
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001634 if (PageAnon(*page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001635 flush_anon_page(vma, *page, ksm_scan.address);
1636 flush_dcache_page(*page);
1637 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001638 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07001639 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001640 ksm_scan.rmap_list =
1641 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001642 ksm_scan.address += PAGE_SIZE;
1643 } else
1644 put_page(*page);
1645 up_read(&mm->mmap_sem);
1646 return rmap_item;
1647 }
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001648 put_page(*page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001649 ksm_scan.address += PAGE_SIZE;
1650 cond_resched();
1651 }
1652 }
1653
Hugh Dickins9ba69292009-09-21 17:02:20 -07001654 if (ksm_test_exit(mm)) {
1655 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001656 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001657 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001658 /*
1659 * Nuke all the rmap_items that are above this current rmap:
1660 * because there were no VM_MERGEABLE vmas with such addresses.
1661 */
Hugh Dickins6514d512009-12-14 17:59:19 -08001662 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001663
1664 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001665 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1666 struct mm_slot, mm_list);
1667 if (ksm_scan.address == 0) {
1668 /*
1669 * We've completed a full scan of all vmas, holding mmap_sem
1670 * throughout, and found no VM_MERGEABLE: so do the same as
1671 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07001672 * This applies either when cleaning up after __ksm_exit
1673 * (but beware: we can reach here even before __ksm_exit),
1674 * or when all VM_MERGEABLE areas have been unmapped (and
1675 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07001676 */
Sasha Levin4ca3a692013-02-22 16:32:28 -08001677 hash_del(&slot->link);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001678 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001679 spin_unlock(&ksm_mmlist_lock);
1680
Hugh Dickinscd551f92009-09-21 17:02:17 -07001681 free_mm_slot(slot);
1682 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001683 up_read(&mm->mmap_sem);
1684 mmdrop(mm);
1685 } else {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001686 up_read(&mm->mmap_sem);
Zhou Chengming7496fea2016-05-12 15:42:21 -07001687 /*
1688 * up_read(&mm->mmap_sem) first because after
1689 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
1690 * already have been freed under us by __ksm_exit()
1691 * because the "mm_slot" is still hashed and
1692 * ksm_scan.mm_slot doesn't point to it anymore.
1693 */
1694 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001695 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001696
1697 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07001698 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07001699 if (slot != &ksm_mm_head)
1700 goto next_mm;
1701
Izik Eidus31dbd012009-09-21 17:02:03 -07001702 ksm_scan.seqnr++;
1703 return NULL;
1704}
1705
1706/**
1707 * ksm_do_scan - the ksm scanner main worker function.
1708 * @scan_npages - number of pages we want to scan before we return.
1709 */
1710static void ksm_do_scan(unsigned int scan_npages)
1711{
1712 struct rmap_item *rmap_item;
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001713 struct page *uninitialized_var(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001714
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001715 while (scan_npages-- && likely(!freezing(current))) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001716 cond_resched();
1717 rmap_item = scan_get_next_rmap_item(&page);
1718 if (!rmap_item)
1719 return;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001720 cmp_and_merge_page(page, rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001721 put_page(page);
1722 }
1723}
1724
Chintan Pandya65180472014-02-18 13:40:01 +05301725static void process_timeout(unsigned long __data)
1726{
1727 wake_up_process((struct task_struct *)__data);
1728}
1729
1730static signed long __sched deferred_schedule_timeout(signed long timeout)
1731{
1732 struct timer_list timer;
1733 unsigned long expire;
1734
1735 __set_current_state(TASK_INTERRUPTIBLE);
1736 if (timeout < 0) {
1737 pr_err("schedule_timeout: wrong timeout value %lx\n",
1738 timeout);
1739 __set_current_state(TASK_RUNNING);
1740 goto out;
1741 }
1742
1743 expire = timeout + jiffies;
1744
1745 setup_deferrable_timer_on_stack(&timer, process_timeout,
1746 (unsigned long)current);
1747 mod_timer(&timer, expire);
1748 schedule();
1749 del_singleshot_timer_sync(&timer);
1750
1751 /* Remove the timer from the object tracker */
1752 destroy_timer_on_stack(&timer);
1753
1754 timeout = expire - jiffies;
1755
1756out:
1757 return timeout < 0 ? 0 : timeout;
1758}
1759
Hugh Dickins6e1583842009-09-21 17:02:14 -07001760static int ksmd_should_run(void)
1761{
1762 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1763}
1764
Izik Eidus31dbd012009-09-21 17:02:03 -07001765static int ksm_scan_thread(void *nothing)
1766{
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001767 set_freezable();
Izik Eidus339aa622009-09-21 17:02:07 -07001768 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07001769
1770 while (!kthread_should_stop()) {
Hugh Dickins6e1583842009-09-21 17:02:14 -07001771 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08001772 wait_while_offlining();
Hugh Dickins6e1583842009-09-21 17:02:14 -07001773 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07001774 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001775 mutex_unlock(&ksm_thread_mutex);
1776
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001777 try_to_freeze();
1778
Hugh Dickins6e1583842009-09-21 17:02:14 -07001779 if (ksmd_should_run()) {
Chintan Pandya65180472014-02-18 13:40:01 +05301780 if (use_deferred_timer)
1781 deferred_schedule_timeout(
1782 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1783 else
1784 schedule_timeout_interruptible(
Izik Eidus31dbd012009-09-21 17:02:03 -07001785 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1786 } else {
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001787 wait_event_freezable(ksm_thread_wait,
Hugh Dickins6e1583842009-09-21 17:02:14 -07001788 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07001789 }
1790 }
1791 return 0;
1792}
1793
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001794int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1795 unsigned long end, int advice, unsigned long *vm_flags)
1796{
1797 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001798 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001799
1800 switch (advice) {
1801 case MADV_MERGEABLE:
1802 /*
1803 * Be somewhat over-protective for now!
1804 */
1805 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1806 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
Kirill A. Shutemov0661a332015-02-10 14:10:04 -08001807 VM_HUGETLB | VM_MIXEDMAP))
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001808 return 0; /* just ignore the advice */
1809
Konstantin Khlebnikovcc2383e2012-10-08 16:28:37 -07001810#ifdef VM_SAO
1811 if (*vm_flags & VM_SAO)
1812 return 0;
1813#endif
1814
Hugh Dickinsd952b792009-09-21 17:02:16 -07001815 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1816 err = __ksm_enter(mm);
1817 if (err)
1818 return err;
1819 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001820
1821 *vm_flags |= VM_MERGEABLE;
1822 break;
1823
1824 case MADV_UNMERGEABLE:
1825 if (!(*vm_flags & VM_MERGEABLE))
1826 return 0; /* just ignore the advice */
1827
Hugh Dickinsd952b792009-09-21 17:02:16 -07001828 if (vma->anon_vma) {
1829 err = unmerge_ksm_pages(vma, start, end);
1830 if (err)
1831 return err;
1832 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001833
1834 *vm_flags &= ~VM_MERGEABLE;
1835 break;
1836 }
1837
1838 return 0;
1839}
1840
1841int __ksm_enter(struct mm_struct *mm)
1842{
Hugh Dickins6e1583842009-09-21 17:02:14 -07001843 struct mm_slot *mm_slot;
1844 int needs_wakeup;
1845
1846 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07001847 if (!mm_slot)
1848 return -ENOMEM;
1849
Hugh Dickins6e1583842009-09-21 17:02:14 -07001850 /* Check ksm_run too? Would need tighter locking */
1851 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1852
Izik Eidus31dbd012009-09-21 17:02:03 -07001853 spin_lock(&ksm_mmlist_lock);
1854 insert_to_mm_slots_hash(mm, mm_slot);
1855 /*
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001856 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1857 * insert just behind the scanning cursor, to let the area settle
Izik Eidus31dbd012009-09-21 17:02:03 -07001858 * down a little; when fork is followed by immediate exec, we don't
1859 * want ksmd to waste time setting up and tearing down an rmap_list.
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001860 *
1861 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1862 * scanning cursor, otherwise KSM pages in newly forked mms will be
1863 * missed: then we might as well insert at the end of the list.
Izik Eidus31dbd012009-09-21 17:02:03 -07001864 */
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001865 if (ksm_run & KSM_RUN_UNMERGE)
1866 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1867 else
1868 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001869 spin_unlock(&ksm_mmlist_lock);
1870
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001871 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001872 atomic_inc(&mm->mm_count);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001873
1874 if (needs_wakeup)
1875 wake_up_interruptible(&ksm_thread_wait);
1876
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001877 return 0;
1878}
1879
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07001880void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001881{
Hugh Dickinscd551f92009-09-21 17:02:17 -07001882 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001883 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07001884
Izik Eidus31dbd012009-09-21 17:02:03 -07001885 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07001886 * This process is exiting: if it's straightforward (as is the
1887 * case when ksmd was never running), free mm_slot immediately.
1888 * But if it's at the cursor or has rmap_items linked to it, use
1889 * mmap_sem to synchronize with any break_cows before pagetables
1890 * are freed, and leave the mm_slot on the list for ksmd to free.
1891 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07001892 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07001893
Hugh Dickinscd551f92009-09-21 17:02:17 -07001894 spin_lock(&ksm_mmlist_lock);
1895 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001896 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001897 if (!mm_slot->rmap_list) {
Sasha Levin4ca3a692013-02-22 16:32:28 -08001898 hash_del(&mm_slot->link);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001899 list_del(&mm_slot->mm_list);
1900 easy_to_free = 1;
1901 } else {
1902 list_move(&mm_slot->mm_list,
1903 &ksm_scan.mm_slot->mm_list);
1904 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001905 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001906 spin_unlock(&ksm_mmlist_lock);
1907
Hugh Dickins9ba69292009-09-21 17:02:20 -07001908 if (easy_to_free) {
1909 free_mm_slot(mm_slot);
1910 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1911 mmdrop(mm);
1912 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001913 down_write(&mm->mmap_sem);
1914 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001915 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001916}
Izik Eidus31dbd012009-09-21 17:02:03 -07001917
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001918struct page *ksm_might_need_to_copy(struct page *page,
Hugh Dickins5ad64682009-12-14 17:59:24 -08001919 struct vm_area_struct *vma, unsigned long address)
1920{
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001921 struct anon_vma *anon_vma = page_anon_vma(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001922 struct page *new_page;
1923
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001924 if (PageKsm(page)) {
1925 if (page_stable_node(page) &&
1926 !(ksm_run & KSM_RUN_UNMERGE))
1927 return page; /* no need to copy it */
1928 } else if (!anon_vma) {
1929 return page; /* no need to copy it */
1930 } else if (anon_vma->root == vma->anon_vma->root &&
1931 page->index == linear_page_index(vma, address)) {
1932 return page; /* still no need to copy it */
1933 }
1934 if (!PageUptodate(page))
1935 return page; /* let do_swap_page report the error */
1936
Hugh Dickins5ad64682009-12-14 17:59:24 -08001937 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1938 if (new_page) {
1939 copy_user_highpage(new_page, page, address, vma);
1940
1941 SetPageDirty(new_page);
1942 __SetPageUptodate(new_page);
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001943 __SetPageLocked(new_page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001944 }
1945
Hugh Dickins5ad64682009-12-14 17:59:24 -08001946 return new_page;
1947}
1948
Joonsoo Kim051ac832014-01-21 15:49:48 -08001949int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001950{
1951 struct stable_node *stable_node;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001952 struct rmap_item *rmap_item;
1953 int ret = SWAP_AGAIN;
1954 int search_new_forks = 0;
1955
Sasha Levin309381fea2014-01-23 15:52:54 -08001956 VM_BUG_ON_PAGE(!PageKsm(page), page);
Joonsoo Kim9f326242014-01-21 15:49:53 -08001957
1958 /*
1959 * Rely on the page lock to protect against concurrent modifications
1960 * to that page's node of the stable tree.
1961 */
Sasha Levin309381fea2014-01-23 15:52:54 -08001962 VM_BUG_ON_PAGE(!PageLocked(page), page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001963
1964 stable_node = page_stable_node(page);
1965 if (!stable_node)
1966 return ret;
Minchan Kimcd256bf2013-05-09 16:21:27 +09001967
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001968again:
Sasha Levinb67bfe02013-02-27 17:06:00 -08001969 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001970 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001971 struct anon_vma_chain *vmac;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001972 struct vm_area_struct *vma;
1973
Andrea Arcangeliad126952015-11-05 18:49:07 -08001974 cond_resched();
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001975 anon_vma_lock_read(anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -07001976 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1977 0, ULONG_MAX) {
Andrea Arcangeliad126952015-11-05 18:49:07 -08001978 cond_resched();
Rik van Riel5beb4932010-03-05 13:42:07 -08001979 vma = vmac->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001980 if (rmap_item->address < vma->vm_start ||
1981 rmap_item->address >= vma->vm_end)
1982 continue;
1983 /*
1984 * Initially we examine only the vma which covers this
1985 * rmap_item; but later, if there is still work to do,
1986 * we examine covering vmas in other mms: in case they
1987 * were forked from the original since ksmd passed.
1988 */
1989 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1990 continue;
1991
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001992 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1993 continue;
1994
Joonsoo Kim051ac832014-01-21 15:49:48 -08001995 ret = rwc->rmap_one(page, vma,
1996 rmap_item->address, rwc->arg);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001997 if (ret != SWAP_AGAIN) {
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001998 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001999 goto out;
2000 }
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08002001 if (rwc->done && rwc->done(page)) {
2002 anon_vma_unlock_read(anon_vma);
2003 goto out;
2004 }
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002005 }
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08002006 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002007 }
2008 if (!search_new_forks++)
2009 goto again;
2010out:
2011 return ret;
2012}
2013
Joonsoo Kim52629502014-01-21 15:49:50 -08002014#ifdef CONFIG_MIGRATION
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002015void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2016{
2017 struct stable_node *stable_node;
2018
Sasha Levin309381fea2014-01-23 15:52:54 -08002019 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2020 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2021 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002022
2023 stable_node = page_stable_node(newpage);
2024 if (stable_node) {
Sasha Levin309381fea2014-01-23 15:52:54 -08002025 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002026 stable_node->kpfn = page_to_pfn(newpage);
Hugh Dickinsc8d65532013-02-22 16:35:10 -08002027 /*
2028 * newpage->mapping was set in advance; now we need smp_wmb()
2029 * to make sure that the new stable_node->kpfn is visible
2030 * to get_ksm_page() before it can see that oldpage->mapping
2031 * has gone stale (or that PageSwapCache has been cleared).
2032 */
2033 smp_wmb();
2034 set_page_stable_node(oldpage, NULL);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002035 }
2036}
2037#endif /* CONFIG_MIGRATION */
2038
Hugh Dickins62b61f62009-12-14 17:59:33 -08002039#ifdef CONFIG_MEMORY_HOTREMOVE
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002040static void wait_while_offlining(void)
2041{
2042 while (ksm_run & KSM_RUN_OFFLINE) {
2043 mutex_unlock(&ksm_thread_mutex);
2044 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
NeilBrown74316202014-07-07 15:16:04 +10002045 TASK_UNINTERRUPTIBLE);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002046 mutex_lock(&ksm_thread_mutex);
2047 }
2048}
2049
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002050static void ksm_check_stable_tree(unsigned long start_pfn,
2051 unsigned long end_pfn)
Hugh Dickins62b61f62009-12-14 17:59:33 -08002052{
Geliang Tang03640412016-01-14 15:20:54 -08002053 struct stable_node *stable_node, *next;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002054 struct rb_node *node;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002055 int nid;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002056
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08002057 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2058 node = rb_first(root_stable_tree + nid);
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002059 while (node) {
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002060 stable_node = rb_entry(node, struct stable_node, node);
2061 if (stable_node->kpfn >= start_pfn &&
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002062 stable_node->kpfn < end_pfn) {
2063 /*
2064 * Don't get_ksm_page, page has already gone:
2065 * which is why we keep kpfn instead of page*
2066 */
2067 remove_node_from_stable_tree(stable_node);
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08002068 node = rb_first(root_stable_tree + nid);
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002069 } else
2070 node = rb_next(node);
2071 cond_resched();
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002072 }
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002073 }
Geliang Tang03640412016-01-14 15:20:54 -08002074 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002075 if (stable_node->kpfn >= start_pfn &&
2076 stable_node->kpfn < end_pfn)
2077 remove_node_from_stable_tree(stable_node);
2078 cond_resched();
2079 }
Hugh Dickins62b61f62009-12-14 17:59:33 -08002080}
2081
2082static int ksm_memory_callback(struct notifier_block *self,
2083 unsigned long action, void *arg)
2084{
2085 struct memory_notify *mn = arg;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002086
2087 switch (action) {
2088 case MEM_GOING_OFFLINE:
2089 /*
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002090 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2091 * and remove_all_stable_nodes() while memory is going offline:
2092 * it is unsafe for them to touch the stable tree at this time.
2093 * But unmerge_ksm_pages(), rmap lookups and other entry points
2094 * which do not need the ksm_thread_mutex are all safe.
Hugh Dickins62b61f62009-12-14 17:59:33 -08002095 */
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002096 mutex_lock(&ksm_thread_mutex);
2097 ksm_run |= KSM_RUN_OFFLINE;
2098 mutex_unlock(&ksm_thread_mutex);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002099 break;
2100
2101 case MEM_OFFLINE:
2102 /*
2103 * Most of the work is done by page migration; but there might
2104 * be a few stable_nodes left over, still pointing to struct
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002105 * pages which have been offlined: prune those from the tree,
2106 * otherwise get_ksm_page() might later try to access a
2107 * non-existent struct page.
Hugh Dickins62b61f62009-12-14 17:59:33 -08002108 */
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002109 ksm_check_stable_tree(mn->start_pfn,
2110 mn->start_pfn + mn->nr_pages);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002111 /* fallthrough */
2112
2113 case MEM_CANCEL_OFFLINE:
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002114 mutex_lock(&ksm_thread_mutex);
2115 ksm_run &= ~KSM_RUN_OFFLINE;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002116 mutex_unlock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002117
2118 smp_mb(); /* wake_up_bit advises this */
2119 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
Hugh Dickins62b61f62009-12-14 17:59:33 -08002120 break;
2121 }
2122 return NOTIFY_OK;
2123}
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002124#else
2125static void wait_while_offlining(void)
2126{
2127}
Hugh Dickins62b61f62009-12-14 17:59:33 -08002128#endif /* CONFIG_MEMORY_HOTREMOVE */
2129
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002130#ifdef CONFIG_SYSFS
2131/*
2132 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2133 */
2134
Izik Eidus31dbd012009-09-21 17:02:03 -07002135#define KSM_ATTR_RO(_name) \
2136 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2137#define KSM_ATTR(_name) \
2138 static struct kobj_attribute _name##_attr = \
2139 __ATTR(_name, 0644, _name##_show, _name##_store)
2140
2141static ssize_t sleep_millisecs_show(struct kobject *kobj,
2142 struct kobj_attribute *attr, char *buf)
2143{
2144 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2145}
2146
2147static ssize_t sleep_millisecs_store(struct kobject *kobj,
2148 struct kobj_attribute *attr,
2149 const char *buf, size_t count)
2150{
2151 unsigned long msecs;
2152 int err;
2153
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002154 err = kstrtoul(buf, 10, &msecs);
Izik Eidus31dbd012009-09-21 17:02:03 -07002155 if (err || msecs > UINT_MAX)
2156 return -EINVAL;
2157
2158 ksm_thread_sleep_millisecs = msecs;
2159
2160 return count;
2161}
2162KSM_ATTR(sleep_millisecs);
2163
2164static ssize_t pages_to_scan_show(struct kobject *kobj,
2165 struct kobj_attribute *attr, char *buf)
2166{
2167 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2168}
2169
2170static ssize_t pages_to_scan_store(struct kobject *kobj,
2171 struct kobj_attribute *attr,
2172 const char *buf, size_t count)
2173{
2174 int err;
2175 unsigned long nr_pages;
2176
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002177 err = kstrtoul(buf, 10, &nr_pages);
Izik Eidus31dbd012009-09-21 17:02:03 -07002178 if (err || nr_pages > UINT_MAX)
2179 return -EINVAL;
2180
2181 ksm_thread_pages_to_scan = nr_pages;
2182
2183 return count;
2184}
2185KSM_ATTR(pages_to_scan);
2186
2187static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2188 char *buf)
2189{
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002190 return sprintf(buf, "%lu\n", ksm_run);
Izik Eidus31dbd012009-09-21 17:02:03 -07002191}
2192
2193static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2194 const char *buf, size_t count)
2195{
2196 int err;
2197 unsigned long flags;
2198
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002199 err = kstrtoul(buf, 10, &flags);
Izik Eidus31dbd012009-09-21 17:02:03 -07002200 if (err || flags > UINT_MAX)
2201 return -EINVAL;
2202 if (flags > KSM_RUN_UNMERGE)
2203 return -EINVAL;
2204
2205 /*
2206 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2207 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsd0f209f2009-12-14 17:59:34 -08002208 * breaking COW to free the pages_shared (but leaves mm_slots
2209 * on the list for when ksmd may be set running again).
Izik Eidus31dbd012009-09-21 17:02:03 -07002210 */
2211
2212 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002213 wait_while_offlining();
Izik Eidus31dbd012009-09-21 17:02:03 -07002214 if (ksm_run != flags) {
2215 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07002216 if (flags & KSM_RUN_UNMERGE) {
David Rientjese1e12d22012-12-11 16:02:56 -08002217 set_current_oom_origin();
Hugh Dickinsd952b792009-09-21 17:02:16 -07002218 err = unmerge_and_remove_all_rmap_items();
David Rientjese1e12d22012-12-11 16:02:56 -08002219 clear_current_oom_origin();
Hugh Dickinsd952b792009-09-21 17:02:16 -07002220 if (err) {
2221 ksm_run = KSM_RUN_STOP;
2222 count = err;
2223 }
2224 }
Izik Eidus31dbd012009-09-21 17:02:03 -07002225 }
2226 mutex_unlock(&ksm_thread_mutex);
2227
2228 if (flags & KSM_RUN_MERGE)
2229 wake_up_interruptible(&ksm_thread_wait);
2230
2231 return count;
2232}
2233KSM_ATTR(run);
2234
Chintan Pandya65180472014-02-18 13:40:01 +05302235static ssize_t deferred_timer_show(struct kobject *kobj,
2236 struct kobj_attribute *attr, char *buf)
2237{
2238 return snprintf(buf, 8, "%d\n", use_deferred_timer);
2239}
2240
2241static ssize_t deferred_timer_store(struct kobject *kobj,
2242 struct kobj_attribute *attr,
2243 const char *buf, size_t count)
2244{
2245 unsigned long enable;
2246 int err;
2247
2248 err = kstrtoul(buf, 10, &enable);
2249 use_deferred_timer = enable;
2250
2251 return count;
2252}
2253KSM_ATTR(deferred_timer);
2254
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002255#ifdef CONFIG_NUMA
2256static ssize_t merge_across_nodes_show(struct kobject *kobj,
2257 struct kobj_attribute *attr, char *buf)
2258{
2259 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2260}
2261
2262static ssize_t merge_across_nodes_store(struct kobject *kobj,
2263 struct kobj_attribute *attr,
2264 const char *buf, size_t count)
2265{
2266 int err;
2267 unsigned long knob;
2268
2269 err = kstrtoul(buf, 10, &knob);
2270 if (err)
2271 return err;
2272 if (knob > 1)
2273 return -EINVAL;
2274
2275 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002276 wait_while_offlining();
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002277 if (ksm_merge_across_nodes != knob) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002278 if (ksm_pages_shared || remove_all_stable_nodes())
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002279 err = -EBUSY;
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08002280 else if (root_stable_tree == one_stable_tree) {
2281 struct rb_root *buf;
2282 /*
2283 * This is the first time that we switch away from the
2284 * default of merging across nodes: must now allocate
2285 * a buffer to hold as many roots as may be needed.
2286 * Allocate stable and unstable together:
2287 * MAXSMP NODES_SHIFT 10 will use 16kB.
2288 */
Joe Perchesbafe1e12013-11-12 15:07:10 -08002289 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2290 GFP_KERNEL);
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08002291 /* Let us assume that RB_ROOT is NULL is zero */
2292 if (!buf)
2293 err = -ENOMEM;
2294 else {
2295 root_stable_tree = buf;
2296 root_unstable_tree = buf + nr_node_ids;
2297 /* Stable tree is empty but not the unstable */
2298 root_unstable_tree[0] = one_unstable_tree[0];
2299 }
2300 }
2301 if (!err) {
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002302 ksm_merge_across_nodes = knob;
Hugh Dickinsef53d16c2013-02-22 16:36:12 -08002303 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2304 }
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002305 }
2306 mutex_unlock(&ksm_thread_mutex);
2307
2308 return err ? err : count;
2309}
2310KSM_ATTR(merge_across_nodes);
2311#endif
2312
Hugh Dickinsb4028262009-09-21 17:02:09 -07002313static ssize_t pages_shared_show(struct kobject *kobj,
2314 struct kobj_attribute *attr, char *buf)
2315{
2316 return sprintf(buf, "%lu\n", ksm_pages_shared);
2317}
2318KSM_ATTR_RO(pages_shared);
2319
2320static ssize_t pages_sharing_show(struct kobject *kobj,
2321 struct kobj_attribute *attr, char *buf)
2322{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07002323 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07002324}
2325KSM_ATTR_RO(pages_sharing);
2326
Hugh Dickins473b0ce2009-09-21 17:02:11 -07002327static ssize_t pages_unshared_show(struct kobject *kobj,
2328 struct kobj_attribute *attr, char *buf)
2329{
2330 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2331}
2332KSM_ATTR_RO(pages_unshared);
2333
2334static ssize_t pages_volatile_show(struct kobject *kobj,
2335 struct kobj_attribute *attr, char *buf)
2336{
2337 long ksm_pages_volatile;
2338
2339 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2340 - ksm_pages_sharing - ksm_pages_unshared;
2341 /*
2342 * It was not worth any locking to calculate that statistic,
2343 * but it might therefore sometimes be negative: conceal that.
2344 */
2345 if (ksm_pages_volatile < 0)
2346 ksm_pages_volatile = 0;
2347 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2348}
2349KSM_ATTR_RO(pages_volatile);
2350
2351static ssize_t full_scans_show(struct kobject *kobj,
2352 struct kobj_attribute *attr, char *buf)
2353{
2354 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2355}
2356KSM_ATTR_RO(full_scans);
2357
Izik Eidus31dbd012009-09-21 17:02:03 -07002358static struct attribute *ksm_attrs[] = {
2359 &sleep_millisecs_attr.attr,
2360 &pages_to_scan_attr.attr,
2361 &run_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07002362 &pages_shared_attr.attr,
2363 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07002364 &pages_unshared_attr.attr,
2365 &pages_volatile_attr.attr,
2366 &full_scans_attr.attr,
Chintan Pandya65180472014-02-18 13:40:01 +05302367 &deferred_timer_attr.attr,
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002368#ifdef CONFIG_NUMA
2369 &merge_across_nodes_attr.attr,
2370#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07002371 NULL,
2372};
2373
2374static struct attribute_group ksm_attr_group = {
2375 .attrs = ksm_attrs,
2376 .name = "ksm",
2377};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002378#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002379
2380static int __init ksm_init(void)
2381{
2382 struct task_struct *ksm_thread;
2383 int err;
2384
2385 err = ksm_slab_init();
2386 if (err)
2387 goto out;
2388
Izik Eidus31dbd012009-09-21 17:02:03 -07002389 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2390 if (IS_ERR(ksm_thread)) {
Paul McQuade25acde32014-10-09 15:29:09 -07002391 pr_err("ksm: creating kthread failed\n");
Izik Eidus31dbd012009-09-21 17:02:03 -07002392 err = PTR_ERR(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002393 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002394 }
2395
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002396#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07002397 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2398 if (err) {
Paul McQuade25acde32014-10-09 15:29:09 -07002399 pr_err("ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002400 kthread_stop(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002401 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002402 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07002403#else
2404 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2405
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002406#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002407
Hugh Dickins62b61f62009-12-14 17:59:33 -08002408#ifdef CONFIG_MEMORY_HOTREMOVE
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002409 /* There is no significance to this priority 100 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08002410 hotplug_memory_notifier(ksm_memory_callback, 100);
2411#endif
Laura Abbott2b4be352014-09-24 15:53:57 -07002412
2413 show_mem_notifier_register(&ksm_show_mem_notifier_block);
Izik Eidus31dbd012009-09-21 17:02:03 -07002414 return 0;
2415
Lai Jiangshand9f89842010-08-09 17:20:02 -07002416out_free:
Izik Eidus31dbd012009-09-21 17:02:03 -07002417 ksm_slab_free();
2418out:
2419 return err;
2420}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -08002421subsys_initcall(ksm_init);