blob: 9227c0bcd2998aab82550e07c78b9d64805f2d1e [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>
Ingo Molnar6e84f312017-02-08 18:51:29 +010022#include <linux/sched/mm.h>
Ingo Molnarf7ccbae2017-02-08 18:51:30 +010023#include <linux/sched/coredump.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070024#include <linux/rwsem.h>
25#include <linux/pagemap.h>
26#include <linux/rmap.h>
27#include <linux/spinlock.h>
28#include <linux/jhash.h>
29#include <linux/delay.h>
30#include <linux/kthread.h>
31#include <linux/wait.h>
32#include <linux/slab.h>
33#include <linux/rbtree.h>
Hugh Dickins62b61f62009-12-14 17:59:33 -080034#include <linux/memory.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070035#include <linux/mmu_notifier.h>
Izik Eidus2c6854f2009-09-23 15:56:04 -070036#include <linux/swap.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070037#include <linux/ksm.h>
Sasha Levin4ca3a692013-02-22 16:32:28 -080038#include <linux/hashtable.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080039#include <linux/freezer.h>
David Rientjes72788c32011-05-24 17:11:40 -070040#include <linux/oom.h>
Petr Holasek90bd6fd2013-02-22 16:35:00 -080041#include <linux/numa.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070042
Izik Eidus31dbd012009-09-21 17:02:03 -070043#include <asm/tlbflush.h>
Hugh Dickins73848b42009-12-14 17:59:22 -080044#include "internal.h"
Izik Eidus31dbd012009-09-21 17:02:03 -070045
Hugh Dickinse850dcf2013-02-22 16:35:03 -080046#ifdef CONFIG_NUMA
47#define NUMA(x) (x)
48#define DO_NUMA(x) do { (x); } while (0)
49#else
50#define NUMA(x) (0)
51#define DO_NUMA(x) do { } while (0)
52#endif
53
Mike Rapoport5a2ca3e2018-04-24 09:40:22 +030054/**
55 * DOC: Overview
56 *
Izik Eidus31dbd012009-09-21 17:02:03 -070057 * A few notes about the KSM scanning process,
58 * to make it easier to understand the data structures below:
59 *
60 * In order to reduce excessive scanning, KSM sorts the memory pages by their
61 * contents into a data structure that holds pointers to the pages' locations.
62 *
63 * Since the contents of the pages may change at any moment, KSM cannot just
64 * insert the pages into a normal sorted tree and expect it to find anything.
65 * Therefore KSM uses two data structures - the stable and the unstable tree.
66 *
67 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
68 * by their contents. Because each such page is write-protected, searching on
69 * this tree is fully assured to be working (except when pages are unmapped),
70 * and therefore this tree is called the stable tree.
71 *
Mike Rapoport5a2ca3e2018-04-24 09:40:22 +030072 * The stable tree node includes information required for reverse
73 * mapping from a KSM page to virtual addresses that map this page.
74 *
75 * In order to avoid large latencies of the rmap walks on KSM pages,
76 * KSM maintains two types of nodes in the stable tree:
77 *
78 * * the regular nodes that keep the reverse mapping structures in a
79 * linked list
80 * * the "chains" that link nodes ("dups") that represent the same
81 * write protected memory content, but each "dup" corresponds to a
82 * different KSM page copy of that content
83 *
84 * Internally, the regular nodes, "dups" and "chains" are represented
85 * using the same :c:type:`struct stable_node` structure.
86 *
Izik Eidus31dbd012009-09-21 17:02:03 -070087 * In addition to the stable tree, KSM uses a second data structure called the
88 * unstable tree: this tree holds pointers to pages which have been found to
89 * be "unchanged for a period of time". The unstable tree sorts these pages
90 * by their contents, but since they are not write-protected, KSM cannot rely
91 * upon the unstable tree to work correctly - the unstable tree is liable to
92 * be corrupted as its contents are modified, and so it is called unstable.
93 *
94 * KSM solves this problem by several techniques:
95 *
96 * 1) The unstable tree is flushed every time KSM completes scanning all
97 * memory areas, and then the tree is rebuilt again from the beginning.
98 * 2) KSM will only insert into the unstable tree, pages whose hash value
99 * has not changed since the previous scan of all memory areas.
100 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
101 * colors of the nodes and not on their contents, assuring that even when
102 * the tree gets "corrupted" it won't get out of balance, so scanning time
103 * remains the same (also, searching and inserting nodes in an rbtree uses
104 * the same algorithm, so we have no overhead when we flush and rebuild).
105 * 4) KSM never flushes the stable tree, which means that even if it were to
106 * take 10 attempts to find a page in the unstable tree, once it is found,
107 * it is secured in the stable tree. (When we scan a new page, we first
108 * compare it against the stable tree, and then against the unstable tree.)
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800109 *
110 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
111 * stable trees and multiple unstable trees: one of each for each NUMA node.
Izik Eidus31dbd012009-09-21 17:02:03 -0700112 */
113
114/**
115 * struct mm_slot - ksm information per mm that is being scanned
116 * @link: link to the mm_slots hash list
117 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
Hugh Dickins6514d512009-12-14 17:59:19 -0800118 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
Izik Eidus31dbd012009-09-21 17:02:03 -0700119 * @mm: the mm that this information is valid for
120 */
121struct mm_slot {
122 struct hlist_node link;
123 struct list_head mm_list;
Hugh Dickins6514d512009-12-14 17:59:19 -0800124 struct rmap_item *rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700125 struct mm_struct *mm;
126};
127
128/**
129 * struct ksm_scan - cursor for scanning
130 * @mm_slot: the current mm_slot we are scanning
131 * @address: the next address inside that to be scanned
Hugh Dickins6514d512009-12-14 17:59:19 -0800132 * @rmap_list: link to the next rmap to be scanned in the rmap_list
Izik Eidus31dbd012009-09-21 17:02:03 -0700133 * @seqnr: count of completed full scans (needed when removing unstable node)
134 *
135 * There is only the one ksm_scan instance of this cursor structure.
136 */
137struct ksm_scan {
138 struct mm_slot *mm_slot;
139 unsigned long address;
Hugh Dickins6514d512009-12-14 17:59:19 -0800140 struct rmap_item **rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700141 unsigned long seqnr;
142};
143
144/**
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800145 * struct stable_node - node of the stable rbtree
146 * @node: rb node of this ksm page in the stable tree
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800147 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700148 * @hlist_dup: linked into the stable_node->hlist with a stable_node chain
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800149 * @list: linked into migrate_nodes, pending placement in the proper node tree
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800150 * @hlist: hlist head of rmap_items using this ksm page
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800151 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700152 * @chain_prune_time: time of the last full garbage collection
153 * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800154 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800155 */
156struct stable_node {
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800157 union {
158 struct rb_node node; /* when node of stable tree */
159 struct { /* when listed for migration */
160 struct list_head *head;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700161 struct {
162 struct hlist_node hlist_dup;
163 struct list_head list;
164 };
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800165 };
166 };
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800167 struct hlist_head hlist;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700168 union {
169 unsigned long kpfn;
170 unsigned long chain_prune_time;
171 };
172 /*
173 * STABLE_NODE_CHAIN can be any negative number in
174 * rmap_hlist_len negative range, but better not -1 to be able
175 * to reliably detect underflows.
176 */
177#define STABLE_NODE_CHAIN -1024
178 int rmap_hlist_len;
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800179#ifdef CONFIG_NUMA
180 int nid;
181#endif
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800182};
183
184/**
Izik Eidus31dbd012009-09-21 17:02:03 -0700185 * struct rmap_item - reverse mapping item for virtual addresses
Hugh Dickins6514d512009-12-14 17:59:19 -0800186 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800187 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
Hugh Dickinsbc566202013-02-22 16:36:06 -0800188 * @nid: NUMA node id of unstable tree in which linked (may not match page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700189 * @mm: the memory structure this rmap_item is pointing into
190 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
191 * @oldchecksum: previous checksum of the page at that virtual address
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800192 * @node: rb node of this rmap_item in the unstable tree
193 * @head: pointer to stable_node heading this list in the stable tree
194 * @hlist: link into hlist of rmap_items hanging off that stable_node
Izik Eidus31dbd012009-09-21 17:02:03 -0700195 */
196struct rmap_item {
Hugh Dickins6514d512009-12-14 17:59:19 -0800197 struct rmap_item *rmap_list;
Hugh Dickinsbc566202013-02-22 16:36:06 -0800198 union {
199 struct anon_vma *anon_vma; /* when stable */
200#ifdef CONFIG_NUMA
201 int nid; /* when node of unstable tree */
202#endif
203 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700204 struct mm_struct *mm;
205 unsigned long address; /* + low bits used for flags below */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800206 unsigned int oldchecksum; /* when unstable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700207 union {
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800208 struct rb_node node; /* when node of unstable tree */
209 struct { /* when listed from stable tree */
210 struct stable_node *head;
211 struct hlist_node hlist;
212 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700213 };
214};
215
216#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800217#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
218#define STABLE_FLAG 0x200 /* is listed from the stable tree */
Jia He1105a2f2018-06-14 15:26:14 -0700219#define KSM_FLAG_MASK (SEQNR_MASK|UNSTABLE_FLAG|STABLE_FLAG)
220 /* to mask all the flags */
Izik Eidus31dbd012009-09-21 17:02:03 -0700221
222/* The stable and unstable tree heads */
Hugh Dickinsef53d162013-02-22 16:36:12 -0800223static struct rb_root one_stable_tree[1] = { RB_ROOT };
224static struct rb_root one_unstable_tree[1] = { RB_ROOT };
225static struct rb_root *root_stable_tree = one_stable_tree;
226static struct rb_root *root_unstable_tree = one_unstable_tree;
Izik Eidus31dbd012009-09-21 17:02:03 -0700227
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800228/* Recently migrated nodes of stable tree, pending proper placement */
229static LIST_HEAD(migrate_nodes);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700230#define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev)
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800231
Sasha Levin4ca3a692013-02-22 16:32:28 -0800232#define MM_SLOTS_HASH_BITS 10
233static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
Izik Eidus31dbd012009-09-21 17:02:03 -0700234
235static struct mm_slot ksm_mm_head = {
236 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
237};
238static struct ksm_scan ksm_scan = {
239 .mm_slot = &ksm_mm_head,
240};
241
242static struct kmem_cache *rmap_item_cache;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800243static struct kmem_cache *stable_node_cache;
Izik Eidus31dbd012009-09-21 17:02:03 -0700244static struct kmem_cache *mm_slot_cache;
245
246/* The number of nodes in the stable tree */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700247static unsigned long ksm_pages_shared;
Izik Eidus31dbd012009-09-21 17:02:03 -0700248
Hugh Dickinse178dfd2009-09-21 17:02:10 -0700249/* The number of page slots additionally sharing those nodes */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700250static unsigned long ksm_pages_sharing;
Izik Eidus31dbd012009-09-21 17:02:03 -0700251
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700252/* The number of nodes in the unstable tree */
253static unsigned long ksm_pages_unshared;
254
255/* The number of rmap_items in use: to calculate pages_volatile */
256static unsigned long ksm_rmap_items;
257
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700258/* The number of stable_node chains */
259static unsigned long ksm_stable_node_chains;
260
261/* The number of stable_node dups linked to the stable_node chains */
262static unsigned long ksm_stable_node_dups;
263
264/* Delay in pruning stale stable_node_dups in the stable_node_chains */
265static int ksm_stable_node_chains_prune_millisecs = 2000;
266
267/* Maximum number of page slots sharing a stable node */
268static int ksm_max_page_sharing = 256;
269
Izik Eidus31dbd012009-09-21 17:02:03 -0700270/* Number of pages ksmd should scan in one batch */
Izik Eidus2c6854f2009-09-23 15:56:04 -0700271static unsigned int ksm_thread_pages_to_scan = 100;
Izik Eidus31dbd012009-09-21 17:02:03 -0700272
273/* Milliseconds ksmd should sleep between batches */
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700274static unsigned int ksm_thread_sleep_millisecs = 20;
Izik Eidus31dbd012009-09-21 17:02:03 -0700275
Claudio Imbrendae86c59b2017-02-24 14:55:39 -0800276/* Checksum of an empty (zeroed) page */
277static unsigned int zero_checksum __read_mostly;
278
279/* Whether to merge empty (zeroed) pages with actual zero pages */
280static bool ksm_use_zero_pages __read_mostly;
281
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800282#ifdef CONFIG_NUMA
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800283/* Zeroed when merging across nodes is not allowed */
284static unsigned int ksm_merge_across_nodes = 1;
Hugh Dickinsef53d162013-02-22 16:36:12 -0800285static int ksm_nr_node_ids = 1;
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800286#else
287#define ksm_merge_across_nodes 1U
Hugh Dickinsef53d162013-02-22 16:36:12 -0800288#define ksm_nr_node_ids 1
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800289#endif
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800290
Izik Eidus31dbd012009-09-21 17:02:03 -0700291#define KSM_RUN_STOP 0
292#define KSM_RUN_MERGE 1
293#define KSM_RUN_UNMERGE 2
Hugh Dickinsef4d43a2013-02-22 16:35:16 -0800294#define KSM_RUN_OFFLINE 4
295static unsigned long ksm_run = KSM_RUN_STOP;
296static void wait_while_offlining(void);
Izik Eidus31dbd012009-09-21 17:02:03 -0700297
298static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
299static DEFINE_MUTEX(ksm_thread_mutex);
300static DEFINE_SPINLOCK(ksm_mmlist_lock);
301
302#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
303 sizeof(struct __struct), __alignof__(struct __struct),\
304 (__flags), NULL)
305
306static int __init ksm_slab_init(void)
307{
308 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
309 if (!rmap_item_cache)
310 goto out;
311
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800312 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
313 if (!stable_node_cache)
314 goto out_free1;
315
Izik Eidus31dbd012009-09-21 17:02:03 -0700316 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
317 if (!mm_slot_cache)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800318 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -0700319
320 return 0;
321
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800322out_free2:
323 kmem_cache_destroy(stable_node_cache);
324out_free1:
Izik Eidus31dbd012009-09-21 17:02:03 -0700325 kmem_cache_destroy(rmap_item_cache);
326out:
327 return -ENOMEM;
328}
329
330static void __init ksm_slab_free(void)
331{
332 kmem_cache_destroy(mm_slot_cache);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800333 kmem_cache_destroy(stable_node_cache);
Izik Eidus31dbd012009-09-21 17:02:03 -0700334 kmem_cache_destroy(rmap_item_cache);
335 mm_slot_cache = NULL;
336}
337
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700338static __always_inline bool is_stable_node_chain(struct stable_node *chain)
339{
340 return chain->rmap_hlist_len == STABLE_NODE_CHAIN;
341}
342
343static __always_inline bool is_stable_node_dup(struct stable_node *dup)
344{
345 return dup->head == STABLE_NODE_DUP_HEAD;
346}
347
348static inline void stable_node_chain_add_dup(struct stable_node *dup,
349 struct stable_node *chain)
350{
351 VM_BUG_ON(is_stable_node_dup(dup));
352 dup->head = STABLE_NODE_DUP_HEAD;
353 VM_BUG_ON(!is_stable_node_chain(chain));
354 hlist_add_head(&dup->hlist_dup, &chain->hlist);
355 ksm_stable_node_dups++;
356}
357
358static inline void __stable_node_dup_del(struct stable_node *dup)
359{
Andrea Arcangelib4fecc62017-07-06 15:36:59 -0700360 VM_BUG_ON(!is_stable_node_dup(dup));
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700361 hlist_del(&dup->hlist_dup);
362 ksm_stable_node_dups--;
363}
364
365static inline void stable_node_dup_del(struct stable_node *dup)
366{
367 VM_BUG_ON(is_stable_node_chain(dup));
368 if (is_stable_node_dup(dup))
369 __stable_node_dup_del(dup);
370 else
371 rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid));
372#ifdef CONFIG_DEBUG_VM
373 dup->head = NULL;
374#endif
375}
376
Izik Eidus31dbd012009-09-21 17:02:03 -0700377static inline struct rmap_item *alloc_rmap_item(void)
378{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700379 struct rmap_item *rmap_item;
380
zhong jiang5b398e42016-09-28 15:22:30 -0700381 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
382 __GFP_NORETRY | __GFP_NOWARN);
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700383 if (rmap_item)
384 ksm_rmap_items++;
385 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -0700386}
387
388static inline void free_rmap_item(struct rmap_item *rmap_item)
389{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700390 ksm_rmap_items--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700391 rmap_item->mm = NULL; /* debug safety */
392 kmem_cache_free(rmap_item_cache, rmap_item);
393}
394
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800395static inline struct stable_node *alloc_stable_node(void)
396{
zhong jiang62130552016-10-07 17:01:19 -0700397 /*
398 * The allocation can take too long with GFP_KERNEL when memory is under
399 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
400 * grants access to memory reserves, helping to avoid this problem.
401 */
402 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800403}
404
405static inline void free_stable_node(struct stable_node *stable_node)
406{
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700407 VM_BUG_ON(stable_node->rmap_hlist_len &&
408 !is_stable_node_chain(stable_node));
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800409 kmem_cache_free(stable_node_cache, stable_node);
410}
411
Izik Eidus31dbd012009-09-21 17:02:03 -0700412static inline struct mm_slot *alloc_mm_slot(void)
413{
414 if (!mm_slot_cache) /* initialization failed */
415 return NULL;
416 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
417}
418
419static inline void free_mm_slot(struct mm_slot *mm_slot)
420{
421 kmem_cache_free(mm_slot_cache, mm_slot);
422}
423
Izik Eidus31dbd012009-09-21 17:02:03 -0700424static struct mm_slot *get_mm_slot(struct mm_struct *mm)
425{
Sasha Levin4ca3a692013-02-22 16:32:28 -0800426 struct mm_slot *slot;
Izik Eidus31dbd012009-09-21 17:02:03 -0700427
Sasha Levinb67bfe02013-02-27 17:06:00 -0800428 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
Sasha Levin4ca3a692013-02-22 16:32:28 -0800429 if (slot->mm == mm)
430 return slot;
431
Izik Eidus31dbd012009-09-21 17:02:03 -0700432 return NULL;
433}
434
435static void insert_to_mm_slots_hash(struct mm_struct *mm,
436 struct mm_slot *mm_slot)
437{
Izik Eidus31dbd012009-09-21 17:02:03 -0700438 mm_slot->mm = mm;
Sasha Levin4ca3a692013-02-22 16:32:28 -0800439 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
Izik Eidus31dbd012009-09-21 17:02:03 -0700440}
441
Izik Eidus31dbd012009-09-21 17:02:03 -0700442/*
Hugh Dickinsa913e182009-09-21 17:02:26 -0700443 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
444 * page tables after it has passed through ksm_exit() - which, if necessary,
445 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
446 * a special flag: they can just back out as soon as mm_users goes to zero.
447 * ksm_test_exit() is used throughout to make this test for exit: in some
448 * places for correctness, in some places just to avoid unnecessary work.
449 */
450static inline bool ksm_test_exit(struct mm_struct *mm)
451{
452 return atomic_read(&mm->mm_users) == 0;
453}
454
455/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700456 * We use break_ksm to break COW on a ksm page: it's a stripped down
457 *
Dave Hansend4edcf02016-02-12 13:01:56 -0800458 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
Izik Eidus31dbd012009-09-21 17:02:03 -0700459 * put_page(page);
460 *
461 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
462 * in case the application has unmapped and remapped mm,addr meanwhile.
463 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
464 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
Dave Hansen1b2ee122016-02-12 13:02:21 -0800465 *
466 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
467 * of the process that owns 'vma'. We also do not want to enforce
468 * protection keys here anyway.
Izik Eidus31dbd012009-09-21 17:02:03 -0700469 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700470static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
Izik Eidus31dbd012009-09-21 17:02:03 -0700471{
472 struct page *page;
Souptick Joarder50a7ca32018-08-17 15:44:47 -0700473 vm_fault_t ret = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700474
475 do {
476 cond_resched();
Dave Hansen1b2ee122016-02-12 13:02:21 -0800477 page = follow_page(vma, addr,
478 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400479 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700480 break;
481 if (PageKsm(page))
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700482 ret = handle_mm_fault(vma, addr,
483 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
Izik Eidus31dbd012009-09-21 17:02:03 -0700484 else
485 ret = VM_FAULT_WRITE;
486 put_page(page);
Linus Torvalds33692f22015-01-29 10:51:32 -0800487 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
Hugh Dickinsd952b792009-09-21 17:02:16 -0700488 /*
489 * We must loop because handle_mm_fault() may back out if there's
490 * any difficulty e.g. if pte accessed bit gets updated concurrently.
491 *
492 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
493 * COW has been broken, even if the vma does not permit VM_WRITE;
494 * but note that a concurrent fault might break PageKsm for us.
495 *
496 * VM_FAULT_SIGBUS could occur if we race with truncation of the
497 * backing file, which also invalidates anonymous pages: that's
498 * okay, that truncation will have unmapped the PageKsm for us.
499 *
500 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
501 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
502 * current task has TIF_MEMDIE set, and will be OOM killed on return
503 * to user; and ksmd, having no mm, would never be chosen for that.
504 *
505 * But if the mm is in a limited mem_cgroup, then the fault may fail
506 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
507 * even ksmd can fail in this way - though it's usually breaking ksm
508 * just to undo a merge it made a moment before, so unlikely to oom.
509 *
510 * That's a pity: we might therefore have more kernel pages allocated
511 * than we're counting as nodes in the stable tree; but ksm_do_scan
512 * will retry to break_cow on each pass, so should recover the page
513 * in due course. The important thing is to not let VM_MERGEABLE
514 * be cleared while any such pages might remain in the area.
515 */
516 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700517}
518
Bob Liuef694222012-03-21 16:34:11 -0700519static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
520 unsigned long addr)
521{
522 struct vm_area_struct *vma;
523 if (ksm_test_exit(mm))
524 return NULL;
525 vma = find_vma(mm, addr);
526 if (!vma || vma->vm_start > addr)
527 return NULL;
528 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
529 return NULL;
530 return vma;
531}
532
Hugh Dickins8dd35572009-12-14 17:59:18 -0800533static void break_cow(struct rmap_item *rmap_item)
Izik Eidus31dbd012009-09-21 17:02:03 -0700534{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800535 struct mm_struct *mm = rmap_item->mm;
536 unsigned long addr = rmap_item->address;
Izik Eidus31dbd012009-09-21 17:02:03 -0700537 struct vm_area_struct *vma;
538
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800539 /*
540 * It is not an accident that whenever we want to break COW
541 * to undo, we also need to drop a reference to the anon_vma.
542 */
Peter Zijlstra9e601092011-03-22 16:32:46 -0700543 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800544
Hugh Dickins81464e302009-09-21 17:02:15 -0700545 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700546 vma = find_mergeable_vma(mm, addr);
547 if (vma)
548 break_ksm(vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700549 up_read(&mm->mmap_sem);
550}
551
552static struct page *get_mergeable_page(struct rmap_item *rmap_item)
553{
554 struct mm_struct *mm = rmap_item->mm;
555 unsigned long addr = rmap_item->address;
556 struct vm_area_struct *vma;
557 struct page *page;
558
559 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700560 vma = find_mergeable_vma(mm, addr);
561 if (!vma)
Izik Eidus31dbd012009-09-21 17:02:03 -0700562 goto out;
563
564 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400565 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700566 goto out;
Kirill A. Shutemovf765f542016-01-15 16:53:03 -0800567 if (PageAnon(page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700568 flush_anon_page(vma, page, addr);
569 flush_dcache_page(page);
570 } else {
571 put_page(page);
Andrea Arcangelic8f95ed2015-11-05 18:49:19 -0800572out:
573 page = NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700574 }
575 up_read(&mm->mmap_sem);
576 return page;
577}
578
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800579/*
580 * This helper is used for getting right index into array of tree roots.
581 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
582 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
583 * every node has its own stable and unstable tree.
584 */
585static inline int get_kpfn_nid(unsigned long kpfn)
586{
Hugh Dickinsd8fc16a2013-03-08 12:43:34 -0800587 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800588}
589
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700590static struct stable_node *alloc_stable_node_chain(struct stable_node *dup,
591 struct rb_root *root)
592{
593 struct stable_node *chain = alloc_stable_node();
594 VM_BUG_ON(is_stable_node_chain(dup));
595 if (likely(chain)) {
596 INIT_HLIST_HEAD(&chain->hlist);
597 chain->chain_prune_time = jiffies;
598 chain->rmap_hlist_len = STABLE_NODE_CHAIN;
599#if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA)
600 chain->nid = -1; /* debug */
601#endif
602 ksm_stable_node_chains++;
603
604 /*
605 * Put the stable node chain in the first dimension of
606 * the stable tree and at the same time remove the old
607 * stable node.
608 */
609 rb_replace_node(&dup->node, &chain->node, root);
610
611 /*
612 * Move the old stable node to the second dimension
613 * queued in the hlist_dup. The invariant is that all
614 * dup stable_nodes in the chain->hlist point to pages
615 * that are wrprotected and have the exact same
616 * content.
617 */
618 stable_node_chain_add_dup(dup, chain);
619 }
620 return chain;
621}
622
623static inline void free_stable_node_chain(struct stable_node *chain,
624 struct rb_root *root)
625{
626 rb_erase(&chain->node, root);
627 free_stable_node(chain);
628 ksm_stable_node_chains--;
629}
630
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800631static void remove_node_from_stable_tree(struct stable_node *stable_node)
632{
633 struct rmap_item *rmap_item;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800634
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700635 /* check it's not STABLE_NODE_CHAIN or negative */
636 BUG_ON(stable_node->rmap_hlist_len < 0);
637
Sasha Levinb67bfe02013-02-27 17:06:00 -0800638 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800639 if (rmap_item->hlist.next)
640 ksm_pages_sharing--;
641 else
642 ksm_pages_shared--;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700643 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
644 stable_node->rmap_hlist_len--;
Peter Zijlstra9e601092011-03-22 16:32:46 -0700645 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800646 rmap_item->address &= PAGE_MASK;
647 cond_resched();
648 }
649
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700650 /*
651 * We need the second aligned pointer of the migrate_nodes
652 * list_head to stay clear from the rb_parent_color union
653 * (aligned and different than any node) and also different
654 * from &migrate_nodes. This will verify that future list.h changes
Nick Desaulniers815f0dd2018-08-22 16:37:24 -0700655 * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it.
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700656 */
Nick Desaulniers815f0dd2018-08-22 16:37:24 -0700657#if defined(GCC_VERSION) && GCC_VERSION >= 40903
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700658 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes);
659 BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1);
660#endif
661
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800662 if (stable_node->head == &migrate_nodes)
663 list_del(&stable_node->list);
664 else
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700665 stable_node_dup_del(stable_node);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800666 free_stable_node(stable_node);
667}
668
669/*
670 * get_ksm_page: checks if the page indicated by the stable node
671 * is still its ksm page, despite having held no reference to it.
672 * In which case we can trust the content of the page, and it
673 * returns the gotten page; but if the page has now been zapped,
674 * remove the stale node from the stable tree and return NULL.
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800675 * But beware, the stable node's page might be being migrated.
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800676 *
677 * You would expect the stable_node to hold a reference to the ksm page.
678 * But if it increments the page's count, swapping out has to wait for
679 * ksmd to come around again before it can free the page, which may take
680 * seconds or even minutes: much too unresponsive. So instead we use a
681 * "keyhole reference": access to the ksm page from the stable node peeps
682 * out through its keyhole to see if that page still holds the right key,
683 * pointing back to this stable node. This relies on freeing a PageAnon
684 * page to reset its page->mapping to NULL, and relies on no other use of
685 * a page to put something that might look like our key in page->mapping.
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800686 * is on its way to being freed; but it is an anomaly to bear in mind.
687 */
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800688static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800689{
690 struct page *page;
691 void *expected_mapping;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800692 unsigned long kpfn;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800693
Minchan Kimbda807d2016-07-26 15:23:05 -0700694 expected_mapping = (void *)((unsigned long)stable_node |
695 PAGE_MAPPING_KSM);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800696again:
Paul E. McKenney08df4772017-10-09 11:51:45 -0700697 kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800698 page = pfn_to_page(kpfn);
Jason Low4db0c3c2015-04-15 16:14:08 -0700699 if (READ_ONCE(page->mapping) != expected_mapping)
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800700 goto stale;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800701
702 /*
703 * We cannot do anything with the page while its refcount is 0.
704 * Usually 0 means free, or tail of a higher-order page: in which
705 * case this node is no longer referenced, and should be freed;
Jiang Biao1c4c3b92018-08-21 21:53:13 -0700706 * however, it might mean that the page is under page_ref_freeze().
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800707 * The __remove_mapping() case is easy, again the node is now stale;
708 * but if page is swapcache in migrate_page_move_mapping(), it might
709 * still be our page, in which case it's essential to keep the node.
710 */
711 while (!get_page_unless_zero(page)) {
712 /*
713 * Another check for page->mapping != expected_mapping would
714 * work here too. We have chosen the !PageSwapCache test to
715 * optimize the common case, when the page is or is about to
716 * be freed: PageSwapCache is cleared (under spin_lock_irq)
Jiang Biao1c4c3b92018-08-21 21:53:13 -0700717 * in the ref_freeze section of __remove_mapping(); but Anon
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800718 * page->mapping reset to NULL later, in free_pages_prepare().
719 */
720 if (!PageSwapCache(page))
721 goto stale;
722 cpu_relax();
723 }
724
Jason Low4db0c3c2015-04-15 16:14:08 -0700725 if (READ_ONCE(page->mapping) != expected_mapping) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800726 put_page(page);
727 goto stale;
728 }
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800729
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800730 if (lock_it) {
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800731 lock_page(page);
Jason Low4db0c3c2015-04-15 16:14:08 -0700732 if (READ_ONCE(page->mapping) != expected_mapping) {
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800733 unlock_page(page);
734 put_page(page);
735 goto stale;
736 }
737 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800738 return page;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800739
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800740stale:
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800741 /*
742 * We come here from above when page->mapping or !PageSwapCache
743 * suggests that the node is stale; but it might be under migration.
744 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
745 * before checking whether node->kpfn has been changed.
746 */
747 smp_rmb();
Jason Low4db0c3c2015-04-15 16:14:08 -0700748 if (READ_ONCE(stable_node->kpfn) != kpfn)
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800749 goto again;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800750 remove_node_from_stable_tree(stable_node);
751 return NULL;
752}
753
Izik Eidus31dbd012009-09-21 17:02:03 -0700754/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700755 * Removing rmap_item from stable or unstable tree.
756 * This function will clean the information from the stable/unstable tree.
757 */
758static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
759{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800760 if (rmap_item->address & STABLE_FLAG) {
761 struct stable_node *stable_node;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800762 struct page *page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700763
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800764 stable_node = rmap_item->head;
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800765 page = get_ksm_page(stable_node, true);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800766 if (!page)
767 goto out;
768
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800769 hlist_del(&rmap_item->hlist);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800770 unlock_page(page);
771 put_page(page);
Hugh Dickins08beca42009-12-14 17:59:21 -0800772
Andrea Arcangeli98666f8a2015-11-05 18:49:13 -0800773 if (!hlist_empty(&stable_node->hlist))
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800774 ksm_pages_sharing--;
775 else
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800776 ksm_pages_shared--;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700777 VM_BUG_ON(stable_node->rmap_hlist_len <= 0);
778 stable_node->rmap_hlist_len--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700779
Peter Zijlstra9e601092011-03-22 16:32:46 -0700780 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins93d17712009-12-14 17:59:16 -0800781 rmap_item->address &= PAGE_MASK;
Izik Eidus31dbd012009-09-21 17:02:03 -0700782
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800783 } else if (rmap_item->address & UNSTABLE_FLAG) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700784 unsigned char age;
785 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -0700786 * Usually ksmd can and must skip the rb_erase, because
Izik Eidus31dbd012009-09-21 17:02:03 -0700787 * root_unstable_tree was already reset to RB_ROOT.
Hugh Dickins9ba69292009-09-21 17:02:20 -0700788 * But be careful when an mm is exiting: do the rb_erase
789 * if this rmap_item was inserted by this scan, rather
790 * than left over from before.
Izik Eidus31dbd012009-09-21 17:02:03 -0700791 */
792 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
Hugh Dickinscd551f92009-09-21 17:02:17 -0700793 BUG_ON(age > 1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700794 if (!age)
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800795 rb_erase(&rmap_item->node,
Hugh Dickinsef53d162013-02-22 16:36:12 -0800796 root_unstable_tree + NUMA(rmap_item->nid));
Hugh Dickins93d17712009-12-14 17:59:16 -0800797 ksm_pages_unshared--;
798 rmap_item->address &= PAGE_MASK;
799 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800800out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700801 cond_resched(); /* we're called from many long loops */
802}
803
Izik Eidus31dbd012009-09-21 17:02:03 -0700804static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -0800805 struct rmap_item **rmap_list)
Izik Eidus31dbd012009-09-21 17:02:03 -0700806{
Hugh Dickins6514d512009-12-14 17:59:19 -0800807 while (*rmap_list) {
808 struct rmap_item *rmap_item = *rmap_list;
809 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700810 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700811 free_rmap_item(rmap_item);
812 }
813}
814
815/*
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800816 * Though it's very tempting to unmerge rmap_items from stable tree rather
Izik Eidus31dbd012009-09-21 17:02:03 -0700817 * than check every pte of a given vma, the locking doesn't quite work for
818 * that - an rmap_item is assigned to the stable tree after inserting ksm
819 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
820 * rmap_items from parent to child at fork time (so as not to waste time
821 * if exit comes before the next scan reaches it).
Hugh Dickins81464e302009-09-21 17:02:15 -0700822 *
823 * Similarly, although we'd like to remove rmap_items (so updating counts
824 * and freeing memory) when unmerging an area, it's easier to leave that
825 * to the next pass of ksmd - consider, for example, how ksmd might be
826 * in cmp_and_merge_page on one of the rmap_items we would be removing.
Izik Eidus31dbd012009-09-21 17:02:03 -0700827 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700828static int unmerge_ksm_pages(struct vm_area_struct *vma,
829 unsigned long start, unsigned long end)
Izik Eidus31dbd012009-09-21 17:02:03 -0700830{
831 unsigned long addr;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700832 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700833
Hugh Dickinsd952b792009-09-21 17:02:16 -0700834 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700835 if (ksm_test_exit(vma->vm_mm))
836 break;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700837 if (signal_pending(current))
838 err = -ERESTARTSYS;
839 else
840 err = break_ksm(vma, addr);
841 }
842 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700843}
844
Mike Rapoport88484822018-06-07 17:07:11 -0700845static inline struct stable_node *page_stable_node(struct page *page)
846{
847 return PageKsm(page) ? page_rmapping(page) : NULL;
848}
849
850static inline void set_page_stable_node(struct page *page,
851 struct stable_node *stable_node)
852{
853 page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM);
854}
855
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700856#ifdef CONFIG_SYSFS
857/*
858 * Only called through the sysfs control interface:
859 */
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800860static int remove_stable_node(struct stable_node *stable_node)
861{
862 struct page *page;
863 int err;
864
865 page = get_ksm_page(stable_node, true);
866 if (!page) {
867 /*
868 * get_ksm_page did remove_node_from_stable_tree itself.
869 */
870 return 0;
871 }
872
Andrey Ryabinine8d355b2019-11-21 17:54:01 -0800873 /*
874 * Page could be still mapped if this races with __mmput() running in
875 * between ksm_exit() and exit_mmap(). Just refuse to let
876 * merge_across_nodes/max_page_sharing be switched.
877 */
878 err = -EBUSY;
879 if (!page_mapped(page)) {
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800880 /*
881 * The stable node did not yet appear stale to get_ksm_page(),
882 * since that allows for an unmapped ksm page to be recognized
883 * right up until it is freed; but the node is safe to remove.
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800884 * This page might be in a pagevec waiting to be freed,
885 * or it might be PageSwapCache (perhaps under writeback),
886 * or it might have been removed from swapcache a moment ago.
887 */
888 set_page_stable_node(page, NULL);
889 remove_node_from_stable_tree(stable_node);
890 err = 0;
891 }
892
893 unlock_page(page);
894 put_page(page);
895 return err;
896}
897
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700898static int remove_stable_node_chain(struct stable_node *stable_node,
899 struct rb_root *root)
900{
901 struct stable_node *dup;
902 struct hlist_node *hlist_safe;
903
904 if (!is_stable_node_chain(stable_node)) {
905 VM_BUG_ON(is_stable_node_dup(stable_node));
906 if (remove_stable_node(stable_node))
907 return true;
908 else
909 return false;
910 }
911
912 hlist_for_each_entry_safe(dup, hlist_safe,
913 &stable_node->hlist, hlist_dup) {
914 VM_BUG_ON(!is_stable_node_dup(dup));
915 if (remove_stable_node(dup))
916 return true;
917 }
918 BUG_ON(!hlist_empty(&stable_node->hlist));
919 free_stable_node_chain(stable_node, root);
920 return false;
921}
922
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800923static int remove_all_stable_nodes(void)
924{
Geliang Tang03640412016-01-14 15:20:54 -0800925 struct stable_node *stable_node, *next;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800926 int nid;
927 int err = 0;
928
Hugh Dickinsef53d162013-02-22 16:36:12 -0800929 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800930 while (root_stable_tree[nid].rb_node) {
931 stable_node = rb_entry(root_stable_tree[nid].rb_node,
932 struct stable_node, node);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -0700933 if (remove_stable_node_chain(stable_node,
934 root_stable_tree + nid)) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800935 err = -EBUSY;
936 break; /* proceed to next nid */
937 }
938 cond_resched();
939 }
940 }
Geliang Tang03640412016-01-14 15:20:54 -0800941 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800942 if (remove_stable_node(stable_node))
943 err = -EBUSY;
944 cond_resched();
945 }
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800946 return err;
947}
948
Hugh Dickinsd952b792009-09-21 17:02:16 -0700949static int unmerge_and_remove_all_rmap_items(void)
Izik Eidus31dbd012009-09-21 17:02:03 -0700950{
951 struct mm_slot *mm_slot;
952 struct mm_struct *mm;
953 struct vm_area_struct *vma;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700954 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700955
Hugh Dickinsd952b792009-09-21 17:02:16 -0700956 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700957 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700958 struct mm_slot, mm_list);
959 spin_unlock(&ksm_mmlist_lock);
960
Hugh Dickins9ba69292009-09-21 17:02:20 -0700961 for (mm_slot = ksm_scan.mm_slot;
962 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700963 mm = mm_slot->mm;
964 down_read(&mm->mmap_sem);
965 for (vma = mm->mmap; vma; vma = vma->vm_next) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700966 if (ksm_test_exit(mm))
967 break;
Izik Eidus31dbd012009-09-21 17:02:03 -0700968 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
969 continue;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700970 err = unmerge_ksm_pages(vma,
971 vma->vm_start, vma->vm_end);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700972 if (err)
973 goto error;
Izik Eidus31dbd012009-09-21 17:02:03 -0700974 }
Hugh Dickins9ba69292009-09-21 17:02:20 -0700975
Hugh Dickins6514d512009-12-14 17:59:19 -0800976 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
Zhou Chengming7496fea2016-05-12 15:42:21 -0700977 up_read(&mm->mmap_sem);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700978
979 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700980 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700981 struct mm_slot, mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700982 if (ksm_test_exit(mm)) {
Sasha Levin4ca3a692013-02-22 16:32:28 -0800983 hash_del(&mm_slot->link);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700984 list_del(&mm_slot->mm_list);
985 spin_unlock(&ksm_mmlist_lock);
986
987 free_mm_slot(mm_slot);
988 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700989 mmdrop(mm);
Zhou Chengming7496fea2016-05-12 15:42:21 -0700990 } else
Hugh Dickins9ba69292009-09-21 17:02:20 -0700991 spin_unlock(&ksm_mmlist_lock);
Izik Eidus31dbd012009-09-21 17:02:03 -0700992 }
993
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800994 /* Clean up stable nodes, but don't worry if some are still busy */
995 remove_all_stable_nodes();
Hugh Dickinsd952b792009-09-21 17:02:16 -0700996 ksm_scan.seqnr = 0;
Hugh Dickins9ba69292009-09-21 17:02:20 -0700997 return 0;
998
999error:
1000 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -07001001 spin_lock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -07001002 ksm_scan.mm_slot = &ksm_mm_head;
Izik Eidus31dbd012009-09-21 17:02:03 -07001003 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -07001004 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -07001005}
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001006#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07001007
Izik Eidus31dbd012009-09-21 17:02:03 -07001008static u32 calc_checksum(struct page *page)
1009{
1010 u32 checksum;
Cong Wang9b04c5f2011-11-25 23:14:39 +08001011 void *addr = kmap_atomic(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001012 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
Cong Wang9b04c5f2011-11-25 23:14:39 +08001013 kunmap_atomic(addr);
Izik Eidus31dbd012009-09-21 17:02:03 -07001014 return checksum;
1015}
1016
1017static int memcmp_pages(struct page *page1, struct page *page2)
1018{
1019 char *addr1, *addr2;
1020 int ret;
1021
Cong Wang9b04c5f2011-11-25 23:14:39 +08001022 addr1 = kmap_atomic(page1);
1023 addr2 = kmap_atomic(page2);
Izik Eidus31dbd012009-09-21 17:02:03 -07001024 ret = memcmp(addr1, addr2, PAGE_SIZE);
Cong Wang9b04c5f2011-11-25 23:14:39 +08001025 kunmap_atomic(addr2);
1026 kunmap_atomic(addr1);
Izik Eidus31dbd012009-09-21 17:02:03 -07001027 return ret;
1028}
1029
1030static inline int pages_identical(struct page *page1, struct page *page2)
1031{
1032 return !memcmp_pages(page1, page2);
1033}
1034
1035static int write_protect_page(struct vm_area_struct *vma, struct page *page,
1036 pte_t *orig_pte)
1037{
1038 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -08001039 struct page_vma_mapped_walk pvmw = {
1040 .page = page,
1041 .vma = vma,
1042 };
Izik Eidus31dbd012009-09-21 17:02:03 -07001043 int swapped;
1044 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -07001045 unsigned long mmun_start; /* For mmu_notifiers */
1046 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -07001047
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -08001048 pvmw.address = page_address_in_vma(page, vma);
1049 if (pvmw.address == -EFAULT)
Izik Eidus31dbd012009-09-21 17:02:03 -07001050 goto out;
1051
Andrea Arcangeli29ad7682011-01-13 15:47:19 -08001052 BUG_ON(PageTransCompound(page));
Haggai Eran6bdb9132012-10-08 16:33:35 -07001053
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -08001054 mmun_start = pvmw.address;
1055 mmun_end = pvmw.address + PAGE_SIZE;
Haggai Eran6bdb9132012-10-08 16:33:35 -07001056 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1057
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -08001058 if (!page_vma_mapped_walk(&pvmw))
Haggai Eran6bdb9132012-10-08 16:33:35 -07001059 goto out_mn;
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -08001060 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
1061 goto out_unlock;
Izik Eidus31dbd012009-09-21 17:02:03 -07001062
Aneesh Kumar K.V595cd8f2017-02-24 14:59:19 -08001063 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
Minchan Kimb3a81d02017-08-10 15:24:15 -07001064 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) ||
1065 mm_tlb_flush_pending(mm)) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001066 pte_t entry;
1067
1068 swapped = PageSwapCache(page);
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -08001069 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
Izik Eidus31dbd012009-09-21 17:02:03 -07001070 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001071 * Ok this is tricky, when get_user_pages_fast() run it doesn't
Izik Eidus31dbd012009-09-21 17:02:03 -07001072 * take any lock, therefore the check that we are going to make
1073 * with the pagecount against the mapcount is racey and
1074 * O_DIRECT can happen right after the check.
1075 * So we clear the pte and flush the tlb before the check
1076 * this assure us that no O_DIRECT can happen after the check
1077 * or in the middle of the check.
Jérôme Glisse0f108512017-11-15 17:34:07 -08001078 *
1079 * No need to notify as we are downgrading page table to read
1080 * only not changing it to point to a new page.
1081 *
Mike Rapoportad56b732018-03-21 21:22:47 +02001082 * See Documentation/vm/mmu_notifier.rst
Izik Eidus31dbd012009-09-21 17:02:03 -07001083 */
Jérôme Glisse0f108512017-11-15 17:34:07 -08001084 entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte);
Izik Eidus31dbd012009-09-21 17:02:03 -07001085 /*
1086 * Check that no O_DIRECT or similar I/O is in progress on the
1087 * page
1088 */
Hugh Dickins31e855e2009-12-14 17:59:17 -08001089 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -08001090 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
Izik Eidus31dbd012009-09-21 17:02:03 -07001091 goto out_unlock;
1092 }
Hugh Dickins4e316352010-10-02 17:49:08 -07001093 if (pte_dirty(entry))
1094 set_page_dirty(page);
Aneesh Kumar K.V595cd8f2017-02-24 14:59:19 -08001095
1096 if (pte_protnone(entry))
1097 entry = pte_mkclean(pte_clear_savedwrite(entry));
1098 else
1099 entry = pte_mkclean(pte_wrprotect(entry));
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -08001100 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
Izik Eidus31dbd012009-09-21 17:02:03 -07001101 }
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -08001102 *orig_pte = *pvmw.pte;
Izik Eidus31dbd012009-09-21 17:02:03 -07001103 err = 0;
1104
1105out_unlock:
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -08001106 page_vma_mapped_walk_done(&pvmw);
Haggai Eran6bdb9132012-10-08 16:33:35 -07001107out_mn:
1108 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -07001109out:
1110 return err;
1111}
1112
1113/**
1114 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -08001115 * @vma: vma that holds the pte pointing to page
1116 * @page: the page we are replacing by kpage
1117 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -07001118 * @orig_pte: the original value of the pte
1119 *
1120 * Returns 0 on success, -EFAULT on failure.
1121 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001122static int replace_page(struct vm_area_struct *vma, struct page *page,
1123 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -07001124{
1125 struct mm_struct *mm = vma->vm_mm;
Izik Eidus31dbd012009-09-21 17:02:03 -07001126 pmd_t *pmd;
1127 pte_t *ptep;
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08001128 pte_t newpte;
Izik Eidus31dbd012009-09-21 17:02:03 -07001129 spinlock_t *ptl;
1130 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -07001131 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -07001132 unsigned long mmun_start; /* For mmu_notifiers */
1133 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -07001134
Hugh Dickins8dd35572009-12-14 17:59:18 -08001135 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -07001136 if (addr == -EFAULT)
1137 goto out;
1138
Bob Liu62190492012-12-11 16:00:37 -08001139 pmd = mm_find_pmd(mm, addr);
1140 if (!pmd)
Izik Eidus31dbd012009-09-21 17:02:03 -07001141 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -07001142
Haggai Eran6bdb9132012-10-08 16:33:35 -07001143 mmun_start = addr;
1144 mmun_end = addr + PAGE_SIZE;
1145 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1146
Izik Eidus31dbd012009-09-21 17:02:03 -07001147 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
1148 if (!pte_same(*ptep, orig_pte)) {
1149 pte_unmap_unlock(ptep, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -07001150 goto out_mn;
Izik Eidus31dbd012009-09-21 17:02:03 -07001151 }
1152
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08001153 /*
1154 * No need to check ksm_use_zero_pages here: we can only have a
1155 * zero_page here if ksm_use_zero_pages was enabled alreaady.
1156 */
1157 if (!is_zero_pfn(page_to_pfn(kpage))) {
1158 get_page(kpage);
1159 page_add_anon_rmap(kpage, vma, addr, false);
1160 newpte = mk_pte(kpage, vma->vm_page_prot);
1161 } else {
1162 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
1163 vma->vm_page_prot));
Claudio Imbrendaa38c0152018-04-10 16:29:41 -07001164 /*
1165 * We're replacing an anonymous page with a zero page, which is
1166 * not anonymous. We need to do proper accounting otherwise we
1167 * will get wrong values in /proc, and a BUG message in dmesg
1168 * when tearing down the mm.
1169 */
1170 dec_mm_counter(mm, MM_ANONPAGES);
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08001171 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001172
1173 flush_cache_page(vma, addr, pte_pfn(*ptep));
Jérôme Glisse0f108512017-11-15 17:34:07 -08001174 /*
1175 * No need to notify as we are replacing a read only page with another
1176 * read only page with the same content.
1177 *
Mike Rapoportad56b732018-03-21 21:22:47 +02001178 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08001179 */
1180 ptep_clear_flush(vma, addr, ptep);
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08001181 set_pte_at_notify(mm, addr, ptep, newpte);
Izik Eidus31dbd012009-09-21 17:02:03 -07001182
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001183 page_remove_rmap(page, false);
Hugh Dickinsae52a2a2011-01-13 15:46:28 -08001184 if (!page_mapped(page))
1185 try_to_free_swap(page);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001186 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001187
1188 pte_unmap_unlock(ptep, ptl);
1189 err = 0;
Haggai Eran6bdb9132012-10-08 16:33:35 -07001190out_mn:
1191 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -07001192out:
1193 return err;
1194}
1195
1196/*
1197 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -08001198 * @vma: the vma that holds the pte pointing to page
1199 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins80e148222009-12-14 17:59:29 -08001200 * @kpage: the PageKsm page that we want to map instead of page,
1201 * or NULL the first time when we want to use page as kpage.
Izik Eidus31dbd012009-09-21 17:02:03 -07001202 *
1203 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1204 */
1205static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -08001206 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001207{
1208 pte_t orig_pte = __pte(0);
1209 int err = -EFAULT;
1210
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001211 if (page == kpage) /* ksm page forked */
1212 return 0;
1213
Hugh Dickins8dd35572009-12-14 17:59:18 -08001214 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -07001215 goto out;
1216
Izik Eidus31dbd012009-09-21 17:02:03 -07001217 /*
1218 * We need the page lock to read a stable PageSwapCache in
1219 * write_protect_page(). We use trylock_page() instead of
1220 * lock_page() because we don't want to wait here - we
1221 * prefer to continue scanning and merging different pages,
1222 * then come back to this page when it is unlocked.
1223 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001224 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -08001225 goto out;
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001226
1227 if (PageTransCompound(page)) {
Andrea Arcangelia7306c32017-06-02 14:46:11 -07001228 if (split_huge_page(page))
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001229 goto out_unlock;
1230 }
1231
Izik Eidus31dbd012009-09-21 17:02:03 -07001232 /*
1233 * If this anonymous page is mapped only here, its pte may need
1234 * to be write-protected. If it's mapped elsewhere, all of its
1235 * ptes are necessarily already write-protected. But in either
1236 * case, we need to lock and check page_count is not raised.
1237 */
Hugh Dickins80e148222009-12-14 17:59:29 -08001238 if (write_protect_page(vma, page, &orig_pte) == 0) {
1239 if (!kpage) {
1240 /*
1241 * While we hold page lock, upgrade page from
1242 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1243 * stable_tree_insert() will update stable_node.
1244 */
1245 set_page_stable_node(page, NULL);
1246 mark_page_accessed(page);
Minchan Kim337ed7e2016-01-15 16:55:15 -08001247 /*
1248 * Page reclaim just frees a clean page with no dirty
1249 * ptes: make sure that the ksm page would be swapped.
1250 */
1251 if (!PageDirty(page))
1252 SetPageDirty(page);
Hugh Dickins80e148222009-12-14 17:59:29 -08001253 err = 0;
1254 } else if (pages_identical(page, kpage))
1255 err = replace_page(vma, page, kpage, orig_pte);
1256 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001257
Hugh Dickins80e148222009-12-14 17:59:29 -08001258 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
Hugh Dickins73848b42009-12-14 17:59:22 -08001259 munlock_vma_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001260 if (!PageMlocked(kpage)) {
1261 unlock_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001262 lock_page(kpage);
1263 mlock_vma_page(kpage);
1264 page = kpage; /* for final unlock */
1265 }
1266 }
Hugh Dickins73848b42009-12-14 17:59:22 -08001267
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001268out_unlock:
Hugh Dickins8dd35572009-12-14 17:59:18 -08001269 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001270out:
1271 return err;
1272}
1273
1274/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001275 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1276 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -08001277 *
1278 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -07001279 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001280static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1281 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -07001282{
Hugh Dickins8dd35572009-12-14 17:59:18 -08001283 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -07001284 struct vm_area_struct *vma;
1285 int err = -EFAULT;
1286
Hugh Dickins8dd35572009-12-14 17:59:18 -08001287 down_read(&mm->mmap_sem);
Andrea Arcangeli85c6e8d2015-11-05 18:49:16 -08001288 vma = find_mergeable_vma(mm, rmap_item->address);
1289 if (!vma)
Hugh Dickins9ba69292009-09-21 17:02:20 -07001290 goto out;
1291
Hugh Dickins8dd35572009-12-14 17:59:18 -08001292 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001293 if (err)
1294 goto out;
1295
Hugh Dickinsbc566202013-02-22 16:36:06 -08001296 /* Unstable nid is in union with stable anon_vma: remove first */
1297 remove_rmap_item_from_tree(rmap_item);
1298
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001299 /* Must get reference to anon_vma while still holding mmap_sem */
Peter Zijlstra9e601092011-03-22 16:32:46 -07001300 rmap_item->anon_vma = vma->anon_vma;
1301 get_anon_vma(vma->anon_vma);
Hugh Dickins81464e302009-09-21 17:02:15 -07001302out:
Hugh Dickins8dd35572009-12-14 17:59:18 -08001303 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -07001304 return err;
1305}
1306
1307/*
Izik Eidus31dbd012009-09-21 17:02:03 -07001308 * try_to_merge_two_pages - take two identical pages and prepare them
1309 * to be merged into one page.
1310 *
Hugh Dickins8dd35572009-12-14 17:59:18 -08001311 * This function returns the kpage if we successfully merged two identical
1312 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001313 *
Hugh Dickins80e148222009-12-14 17:59:29 -08001314 * Note that this function upgrades page to ksm page: if one of the pages
Izik Eidus31dbd012009-09-21 17:02:03 -07001315 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1316 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001317static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1318 struct page *page,
1319 struct rmap_item *tree_rmap_item,
1320 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001321{
Hugh Dickins80e148222009-12-14 17:59:29 -08001322 int err;
Izik Eidus31dbd012009-09-21 17:02:03 -07001323
Hugh Dickins80e148222009-12-14 17:59:29 -08001324 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
Izik Eidus31dbd012009-09-21 17:02:03 -07001325 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001326 err = try_to_merge_with_ksm_page(tree_rmap_item,
Hugh Dickins80e148222009-12-14 17:59:29 -08001327 tree_page, page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001328 /*
Hugh Dickins81464e302009-09-21 17:02:15 -07001329 * If that fails, we have a ksm page with only one pte
1330 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -07001331 */
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001332 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -08001333 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001334 }
Hugh Dickins80e148222009-12-14 17:59:29 -08001335 return err ? NULL : page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001336}
1337
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001338static __always_inline
1339bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset)
1340{
1341 VM_BUG_ON(stable_node->rmap_hlist_len < 0);
1342 /*
1343 * Check that at least one mapping still exists, otherwise
1344 * there's no much point to merge and share with this
1345 * stable_node, as the underlying tree_page of the other
1346 * sharer is going to be freed soon.
1347 */
1348 return stable_node->rmap_hlist_len &&
1349 stable_node->rmap_hlist_len + offset < ksm_max_page_sharing;
1350}
1351
1352static __always_inline
1353bool is_page_sharing_candidate(struct stable_node *stable_node)
1354{
1355 return __is_page_sharing_candidate(stable_node, 0);
1356}
1357
Colin Ian Kingc01f0b52018-04-05 16:22:01 -07001358static struct page *stable_node_dup(struct stable_node **_stable_node_dup,
1359 struct stable_node **_stable_node,
1360 struct rb_root *root,
1361 bool prune_stale_stable_nodes)
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001362{
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001363 struct stable_node *dup, *found = NULL, *stable_node = *_stable_node;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001364 struct hlist_node *hlist_safe;
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001365 struct page *_tree_page, *tree_page = NULL;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001366 int nr = 0;
1367 int found_rmap_hlist_len;
1368
1369 if (!prune_stale_stable_nodes ||
1370 time_before(jiffies, stable_node->chain_prune_time +
1371 msecs_to_jiffies(
1372 ksm_stable_node_chains_prune_millisecs)))
1373 prune_stale_stable_nodes = false;
1374 else
1375 stable_node->chain_prune_time = jiffies;
1376
1377 hlist_for_each_entry_safe(dup, hlist_safe,
1378 &stable_node->hlist, hlist_dup) {
1379 cond_resched();
1380 /*
1381 * We must walk all stable_node_dup to prune the stale
1382 * stable nodes during lookup.
1383 *
1384 * get_ksm_page can drop the nodes from the
1385 * stable_node->hlist if they point to freed pages
1386 * (that's why we do a _safe walk). The "dup"
1387 * stable_node parameter itself will be freed from
1388 * under us if it returns NULL.
1389 */
1390 _tree_page = get_ksm_page(dup, false);
1391 if (!_tree_page)
1392 continue;
1393 nr += 1;
1394 if (is_page_sharing_candidate(dup)) {
1395 if (!found ||
1396 dup->rmap_hlist_len > found_rmap_hlist_len) {
1397 if (found)
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001398 put_page(tree_page);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001399 found = dup;
1400 found_rmap_hlist_len = found->rmap_hlist_len;
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001401 tree_page = _tree_page;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001402
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001403 /* skip put_page for found dup */
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001404 if (!prune_stale_stable_nodes)
1405 break;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001406 continue;
1407 }
1408 }
1409 put_page(_tree_page);
1410 }
1411
Andrea Arcangeli80b18df2017-07-06 15:37:08 -07001412 if (found) {
1413 /*
1414 * nr is counting all dups in the chain only if
1415 * prune_stale_stable_nodes is true, otherwise we may
1416 * break the loop at nr == 1 even if there are
1417 * multiple entries.
1418 */
1419 if (prune_stale_stable_nodes && nr == 1) {
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001420 /*
1421 * If there's not just one entry it would
1422 * corrupt memory, better BUG_ON. In KSM
1423 * context with no lock held it's not even
1424 * fatal.
1425 */
1426 BUG_ON(stable_node->hlist.first->next);
1427
1428 /*
1429 * There's just one entry and it is below the
1430 * deduplication limit so drop the chain.
1431 */
1432 rb_replace_node(&stable_node->node, &found->node,
1433 root);
1434 free_stable_node(stable_node);
1435 ksm_stable_node_chains--;
1436 ksm_stable_node_dups--;
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001437 /*
Andrea Arcangeli0ba1d0f2017-07-06 15:37:02 -07001438 * NOTE: the caller depends on the stable_node
1439 * to be equal to stable_node_dup if the chain
1440 * was collapsed.
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001441 */
Andrea Arcangeli0ba1d0f2017-07-06 15:37:02 -07001442 *_stable_node = found;
1443 /*
1444 * Just for robustneess as stable_node is
1445 * otherwise left as a stable pointer, the
1446 * compiler shall optimize it away at build
1447 * time.
1448 */
1449 stable_node = NULL;
Andrea Arcangeli80b18df2017-07-06 15:37:08 -07001450 } else if (stable_node->hlist.first != &found->hlist_dup &&
1451 __is_page_sharing_candidate(found, 1)) {
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001452 /*
Andrea Arcangeli80b18df2017-07-06 15:37:08 -07001453 * If the found stable_node dup can accept one
1454 * more future merge (in addition to the one
1455 * that is underway) and is not at the head of
1456 * the chain, put it there so next search will
1457 * be quicker in the !prune_stale_stable_nodes
1458 * case.
1459 *
1460 * NOTE: it would be inaccurate to use nr > 1
1461 * instead of checking the hlist.first pointer
1462 * directly, because in the
1463 * prune_stale_stable_nodes case "nr" isn't
1464 * the position of the found dup in the chain,
1465 * but the total number of dups in the chain.
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001466 */
1467 hlist_del(&found->hlist_dup);
1468 hlist_add_head(&found->hlist_dup,
1469 &stable_node->hlist);
1470 }
1471 }
1472
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001473 *_stable_node_dup = found;
1474 return tree_page;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001475}
1476
1477static struct stable_node *stable_node_dup_any(struct stable_node *stable_node,
1478 struct rb_root *root)
1479{
1480 if (!is_stable_node_chain(stable_node))
1481 return stable_node;
1482 if (hlist_empty(&stable_node->hlist)) {
1483 free_stable_node_chain(stable_node, root);
1484 return NULL;
1485 }
1486 return hlist_entry(stable_node->hlist.first,
1487 typeof(*stable_node), hlist_dup);
1488}
1489
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001490/*
1491 * Like for get_ksm_page, this function can free the *_stable_node and
1492 * *_stable_node_dup if the returned tree_page is NULL.
1493 *
1494 * It can also free and overwrite *_stable_node with the found
1495 * stable_node_dup if the chain is collapsed (in which case
1496 * *_stable_node will be equal to *_stable_node_dup like if the chain
1497 * never existed). It's up to the caller to verify tree_page is not
1498 * NULL before dereferencing *_stable_node or *_stable_node_dup.
1499 *
1500 * *_stable_node_dup is really a second output parameter of this
1501 * function and will be overwritten in all cases, the caller doesn't
1502 * need to initialize it.
1503 */
1504static struct page *__stable_node_chain(struct stable_node **_stable_node_dup,
1505 struct stable_node **_stable_node,
1506 struct rb_root *root,
1507 bool prune_stale_stable_nodes)
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001508{
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001509 struct stable_node *stable_node = *_stable_node;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001510 if (!is_stable_node_chain(stable_node)) {
1511 if (is_page_sharing_candidate(stable_node)) {
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001512 *_stable_node_dup = stable_node;
1513 return get_ksm_page(stable_node, false);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001514 }
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001515 /*
1516 * _stable_node_dup set to NULL means the stable_node
1517 * reached the ksm_max_page_sharing limit.
1518 */
1519 *_stable_node_dup = NULL;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001520 return NULL;
1521 }
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001522 return stable_node_dup(_stable_node_dup, _stable_node, root,
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001523 prune_stale_stable_nodes);
1524}
1525
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001526static __always_inline struct page *chain_prune(struct stable_node **s_n_d,
1527 struct stable_node **s_n,
1528 struct rb_root *root)
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001529{
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001530 return __stable_node_chain(s_n_d, s_n, root, true);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001531}
1532
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001533static __always_inline struct page *chain(struct stable_node **s_n_d,
1534 struct stable_node *s_n,
1535 struct rb_root *root)
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001536{
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001537 struct stable_node *old_stable_node = s_n;
1538 struct page *tree_page;
1539
1540 tree_page = __stable_node_chain(s_n_d, &s_n, root, false);
1541 /* not pruning dups so s_n cannot have changed */
1542 VM_BUG_ON(s_n != old_stable_node);
1543 return tree_page;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001544}
1545
Izik Eidus31dbd012009-09-21 17:02:03 -07001546/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001547 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -07001548 *
1549 * This function checks if there is a page inside the stable tree
1550 * with identical content to the page that we are scanning right now.
1551 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001552 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -07001553 * NULL otherwise.
1554 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001555static struct page *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001556{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001557 int nid;
Hugh Dickinsef53d162013-02-22 16:36:12 -08001558 struct rb_root *root;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001559 struct rb_node **new;
1560 struct rb_node *parent;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001561 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001562 struct stable_node *page_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001563
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001564 page_node = page_stable_node(page);
1565 if (page_node && page_node->head != &migrate_nodes) {
1566 /* ksm page forked */
Hugh Dickins08beca42009-12-14 17:59:21 -08001567 get_page(page);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001568 return page;
Hugh Dickins08beca42009-12-14 17:59:21 -08001569 }
1570
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001571 nid = get_kpfn_nid(page_to_pfn(page));
Hugh Dickinsef53d162013-02-22 16:36:12 -08001572 root = root_stable_tree + nid;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001573again:
Hugh Dickinsef53d162013-02-22 16:36:12 -08001574 new = &root->rb_node;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001575 parent = NULL;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001576
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001577 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001578 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001579 int ret;
1580
Hugh Dickins08beca42009-12-14 17:59:21 -08001581 cond_resched();
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001582 stable_node = rb_entry(*new, struct stable_node, node);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001583 stable_node_any = NULL;
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001584 tree_page = chain_prune(&stable_node_dup, &stable_node, root);
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001585 /*
1586 * NOTE: stable_node may have been freed by
1587 * chain_prune() if the returned stable_node_dup is
1588 * not NULL. stable_node_dup may have been inserted in
1589 * the rbtree instead as a regular stable_node (in
1590 * order to collapse the stable_node chain if a single
Andrea Arcangeli0ba1d0f2017-07-06 15:37:02 -07001591 * stable_node dup was found in it). In such case the
1592 * stable_node is overwritten by the calleee to point
1593 * to the stable_node_dup that was collapsed in the
1594 * stable rbtree and stable_node will be equal to
1595 * stable_node_dup like if the chain never existed.
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001596 */
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001597 if (!stable_node_dup) {
1598 /*
1599 * Either all stable_node dups were full in
1600 * this stable_node chain, or this chain was
1601 * empty and should be rb_erased.
1602 */
1603 stable_node_any = stable_node_dup_any(stable_node,
1604 root);
1605 if (!stable_node_any) {
1606 /* rb_erase just run */
1607 goto again;
1608 }
1609 /*
1610 * Take any of the stable_node dups page of
1611 * this stable_node chain to let the tree walk
1612 * continue. All KSM pages belonging to the
1613 * stable_node dups in a stable_node chain
1614 * have the same content and they're
1615 * wrprotected at all times. Any will work
1616 * fine to continue the walk.
1617 */
1618 tree_page = get_ksm_page(stable_node_any, false);
1619 }
1620 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001621 if (!tree_page) {
1622 /*
1623 * If we walked over a stale stable_node,
1624 * get_ksm_page() will call rb_erase() and it
1625 * may rebalance the tree from under us. So
1626 * restart the search from scratch. Returning
1627 * NULL would be safe too, but we'd generate
1628 * false negative insertions just because some
1629 * stable_node was stale.
1630 */
1631 goto again;
1632 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001633
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001634 ret = memcmp_pages(page, tree_page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001635 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001636
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001637 parent = *new;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001638 if (ret < 0)
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001639 new = &parent->rb_left;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001640 else if (ret > 0)
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001641 new = &parent->rb_right;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001642 else {
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001643 if (page_node) {
1644 VM_BUG_ON(page_node->head != &migrate_nodes);
1645 /*
1646 * Test if the migrated page should be merged
1647 * into a stable node dup. If the mapcount is
1648 * 1 we can migrate it with another KSM page
1649 * without adding it to the chain.
1650 */
1651 if (page_mapcount(page) > 1)
1652 goto chain_append;
1653 }
1654
1655 if (!stable_node_dup) {
1656 /*
1657 * If the stable_node is a chain and
1658 * we got a payload match in memcmp
1659 * but we cannot merge the scanned
1660 * page in any of the existing
1661 * stable_node dups because they're
1662 * all full, we need to wait the
1663 * scanned page to find itself a match
1664 * in the unstable tree to create a
1665 * brand new KSM page to add later to
1666 * the dups of this stable_node.
1667 */
1668 return NULL;
1669 }
1670
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001671 /*
1672 * Lock and unlock the stable_node's page (which
1673 * might already have been migrated) so that page
1674 * migration is sure to notice its raised count.
1675 * It would be more elegant to return stable_node
1676 * than kpage, but that involves more changes.
1677 */
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001678 tree_page = get_ksm_page(stable_node_dup, true);
1679 if (unlikely(!tree_page))
1680 /*
1681 * The tree may have been rebalanced,
1682 * so re-evaluate parent and new.
1683 */
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001684 goto again;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001685 unlock_page(tree_page);
1686
1687 if (get_kpfn_nid(stable_node_dup->kpfn) !=
1688 NUMA(stable_node_dup->nid)) {
1689 put_page(tree_page);
1690 goto replace;
1691 }
1692 return tree_page;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001693 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001694 }
1695
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001696 if (!page_node)
1697 return NULL;
1698
1699 list_del(&page_node->list);
1700 DO_NUMA(page_node->nid = nid);
1701 rb_link_node(&page_node->node, parent, new);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001702 rb_insert_color(&page_node->node, root);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001703out:
1704 if (is_page_sharing_candidate(page_node)) {
1705 get_page(page);
1706 return page;
1707 } else
1708 return NULL;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001709
1710replace:
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001711 /*
1712 * If stable_node was a chain and chain_prune collapsed it,
Andrea Arcangeli0ba1d0f2017-07-06 15:37:02 -07001713 * stable_node has been updated to be the new regular
1714 * stable_node. A collapse of the chain is indistinguishable
1715 * from the case there was no chain in the stable
1716 * rbtree. Otherwise stable_node is the chain and
1717 * stable_node_dup is the dup to replace.
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001718 */
Andrea Arcangeli0ba1d0f2017-07-06 15:37:02 -07001719 if (stable_node_dup == stable_node) {
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001720 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1721 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001722 /* there is no chain */
1723 if (page_node) {
1724 VM_BUG_ON(page_node->head != &migrate_nodes);
1725 list_del(&page_node->list);
1726 DO_NUMA(page_node->nid = nid);
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001727 rb_replace_node(&stable_node_dup->node,
1728 &page_node->node,
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001729 root);
1730 if (is_page_sharing_candidate(page_node))
1731 get_page(page);
1732 else
1733 page = NULL;
1734 } else {
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001735 rb_erase(&stable_node_dup->node, root);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001736 page = NULL;
1737 }
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001738 } else {
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001739 VM_BUG_ON(!is_stable_node_chain(stable_node));
1740 __stable_node_dup_del(stable_node_dup);
1741 if (page_node) {
1742 VM_BUG_ON(page_node->head != &migrate_nodes);
1743 list_del(&page_node->list);
1744 DO_NUMA(page_node->nid = nid);
1745 stable_node_chain_add_dup(page_node, stable_node);
1746 if (is_page_sharing_candidate(page_node))
1747 get_page(page);
1748 else
1749 page = NULL;
1750 } else {
1751 page = NULL;
1752 }
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001753 }
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001754 stable_node_dup->head = &migrate_nodes;
1755 list_add(&stable_node_dup->list, stable_node_dup->head);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001756 return page;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001757
1758chain_append:
1759 /* stable_node_dup could be null if it reached the limit */
1760 if (!stable_node_dup)
1761 stable_node_dup = stable_node_any;
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001762 /*
1763 * If stable_node was a chain and chain_prune collapsed it,
Andrea Arcangeli0ba1d0f2017-07-06 15:37:02 -07001764 * stable_node has been updated to be the new regular
1765 * stable_node. A collapse of the chain is indistinguishable
1766 * from the case there was no chain in the stable
1767 * rbtree. Otherwise stable_node is the chain and
1768 * stable_node_dup is the dup to replace.
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001769 */
Andrea Arcangeli0ba1d0f2017-07-06 15:37:02 -07001770 if (stable_node_dup == stable_node) {
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001771 VM_BUG_ON(is_stable_node_chain(stable_node_dup));
1772 VM_BUG_ON(is_stable_node_dup(stable_node_dup));
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001773 /* chain is missing so create it */
1774 stable_node = alloc_stable_node_chain(stable_node_dup,
1775 root);
1776 if (!stable_node)
1777 return NULL;
1778 }
1779 /*
1780 * Add this stable_node dup that was
1781 * migrated to the stable_node chain
1782 * of the current nid for this page
1783 * content.
1784 */
Andrea Arcangelib4fecc62017-07-06 15:36:59 -07001785 VM_BUG_ON(!is_stable_node_chain(stable_node));
1786 VM_BUG_ON(!is_stable_node_dup(stable_node_dup));
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001787 VM_BUG_ON(page_node->head != &migrate_nodes);
1788 list_del(&page_node->list);
1789 DO_NUMA(page_node->nid = nid);
1790 stable_node_chain_add_dup(page_node, stable_node);
1791 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -07001792}
1793
1794/*
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001795 * stable_tree_insert - insert stable tree node pointing to new ksm page
Izik Eidus31dbd012009-09-21 17:02:03 -07001796 * into the stable tree.
1797 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001798 * This function returns the stable tree node just allocated on success,
1799 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001800 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001801static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001802{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001803 int nid;
1804 unsigned long kpfn;
Hugh Dickinsef53d162013-02-22 16:36:12 -08001805 struct rb_root *root;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001806 struct rb_node **new;
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001807 struct rb_node *parent;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001808 struct stable_node *stable_node, *stable_node_dup, *stable_node_any;
1809 bool need_chain = false;
Izik Eidus31dbd012009-09-21 17:02:03 -07001810
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001811 kpfn = page_to_pfn(kpage);
1812 nid = get_kpfn_nid(kpfn);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001813 root = root_stable_tree + nid;
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001814again:
1815 parent = NULL;
Hugh Dickinsef53d162013-02-22 16:36:12 -08001816 new = &root->rb_node;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001817
Izik Eidus31dbd012009-09-21 17:02:03 -07001818 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001819 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001820 int ret;
1821
Hugh Dickins08beca42009-12-14 17:59:21 -08001822 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001823 stable_node = rb_entry(*new, struct stable_node, node);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001824 stable_node_any = NULL;
Andrea Arcangeli8dc5ffc2017-07-06 15:37:05 -07001825 tree_page = chain(&stable_node_dup, stable_node, root);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001826 if (!stable_node_dup) {
1827 /*
1828 * Either all stable_node dups were full in
1829 * this stable_node chain, or this chain was
1830 * empty and should be rb_erased.
1831 */
1832 stable_node_any = stable_node_dup_any(stable_node,
1833 root);
1834 if (!stable_node_any) {
1835 /* rb_erase just run */
1836 goto again;
1837 }
1838 /*
1839 * Take any of the stable_node dups page of
1840 * this stable_node chain to let the tree walk
1841 * continue. All KSM pages belonging to the
1842 * stable_node dups in a stable_node chain
1843 * have the same content and they're
1844 * wrprotected at all times. Any will work
1845 * fine to continue the walk.
1846 */
1847 tree_page = get_ksm_page(stable_node_any, false);
1848 }
1849 VM_BUG_ON(!stable_node_dup ^ !!stable_node_any);
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001850 if (!tree_page) {
1851 /*
1852 * If we walked over a stale stable_node,
1853 * get_ksm_page() will call rb_erase() and it
1854 * may rebalance the tree from under us. So
1855 * restart the search from scratch. Returning
1856 * NULL would be safe too, but we'd generate
1857 * false negative insertions just because some
1858 * stable_node was stale.
1859 */
1860 goto again;
1861 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001862
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001863 ret = memcmp_pages(kpage, tree_page);
1864 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001865
1866 parent = *new;
1867 if (ret < 0)
1868 new = &parent->rb_left;
1869 else if (ret > 0)
1870 new = &parent->rb_right;
1871 else {
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001872 need_chain = true;
1873 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001874 }
1875 }
1876
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001877 stable_node_dup = alloc_stable_node();
1878 if (!stable_node_dup)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001879 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001880
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001881 INIT_HLIST_HEAD(&stable_node_dup->hlist);
1882 stable_node_dup->kpfn = kpfn;
1883 set_page_stable_node(kpage, stable_node_dup);
1884 stable_node_dup->rmap_hlist_len = 0;
1885 DO_NUMA(stable_node_dup->nid = nid);
1886 if (!need_chain) {
1887 rb_link_node(&stable_node_dup->node, parent, new);
1888 rb_insert_color(&stable_node_dup->node, root);
1889 } else {
1890 if (!is_stable_node_chain(stable_node)) {
1891 struct stable_node *orig = stable_node;
1892 /* chain is missing so create it */
1893 stable_node = alloc_stable_node_chain(orig, root);
1894 if (!stable_node) {
1895 free_stable_node(stable_node_dup);
1896 return NULL;
1897 }
1898 }
1899 stable_node_chain_add_dup(stable_node_dup, stable_node);
1900 }
Hugh Dickins08beca42009-12-14 17:59:21 -08001901
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001902 return stable_node_dup;
Izik Eidus31dbd012009-09-21 17:02:03 -07001903}
1904
1905/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001906 * unstable_tree_search_insert - search for identical page,
1907 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001908 *
1909 * This function searches for a page in the unstable tree identical to the
1910 * page currently being scanned; and if no identical page is found in the
1911 * tree, we insert rmap_item as a new object into the unstable tree.
1912 *
1913 * This function returns pointer to rmap_item found to be identical
1914 * to the currently scanned page, NULL otherwise.
1915 *
1916 * This function does both searching and inserting, because they share
1917 * the same walking algorithm in an rbtree.
1918 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001919static
1920struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1921 struct page *page,
1922 struct page **tree_pagep)
Izik Eidus31dbd012009-09-21 17:02:03 -07001923{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001924 struct rb_node **new;
1925 struct rb_root *root;
Izik Eidus31dbd012009-09-21 17:02:03 -07001926 struct rb_node *parent = NULL;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001927 int nid;
1928
1929 nid = get_kpfn_nid(page_to_pfn(page));
Hugh Dickinsef53d162013-02-22 16:36:12 -08001930 root = root_unstable_tree + nid;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001931 new = &root->rb_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001932
1933 while (*new) {
1934 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001935 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001936 int ret;
1937
Hugh Dickinsd178f272009-11-09 15:58:23 +00001938 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -07001939 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001940 tree_page = get_mergeable_page(tree_rmap_item);
Andrea Arcangelic8f95ed2015-11-05 18:49:19 -08001941 if (!tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001942 return NULL;
1943
1944 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001945 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -07001946 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001947 if (page == tree_page) {
1948 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001949 return NULL;
1950 }
1951
Hugh Dickins8dd35572009-12-14 17:59:18 -08001952 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001953
1954 parent = *new;
1955 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001956 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001957 new = &parent->rb_left;
1958 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001959 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001960 new = &parent->rb_right;
Hugh Dickinsb599cbd2013-02-22 16:36:05 -08001961 } else if (!ksm_merge_across_nodes &&
1962 page_to_nid(tree_page) != nid) {
1963 /*
1964 * If tree_page has been migrated to another NUMA node,
1965 * it will be flushed out and put in the right unstable
1966 * tree next time: only merge with it when across_nodes.
1967 */
1968 put_page(tree_page);
1969 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001970 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001971 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001972 return tree_rmap_item;
1973 }
1974 }
1975
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001976 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001977 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001978 DO_NUMA(rmap_item->nid = nid);
Izik Eidus31dbd012009-09-21 17:02:03 -07001979 rb_link_node(&rmap_item->node, parent, new);
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001980 rb_insert_color(&rmap_item->node, root);
Izik Eidus31dbd012009-09-21 17:02:03 -07001981
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001982 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001983 return NULL;
1984}
1985
1986/*
1987 * stable_tree_append - add another rmap_item to the linked list of
1988 * rmap_items hanging off a given node of the stable tree, all sharing
1989 * the same ksm page.
1990 */
1991static void stable_tree_append(struct rmap_item *rmap_item,
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001992 struct stable_node *stable_node,
1993 bool max_page_sharing_bypass)
Izik Eidus31dbd012009-09-21 17:02:03 -07001994{
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07001995 /*
1996 * rmap won't find this mapping if we don't insert the
1997 * rmap_item in the right stable_node
1998 * duplicate. page_migration could break later if rmap breaks,
1999 * so we can as well crash here. We really need to check for
2000 * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check
2001 * for other negative values as an undeflow if detected here
2002 * for the first time (and not when decreasing rmap_hlist_len)
2003 * would be sign of memory corruption in the stable_node.
2004 */
2005 BUG_ON(stable_node->rmap_hlist_len < 0);
2006
2007 stable_node->rmap_hlist_len++;
2008 if (!max_page_sharing_bypass)
2009 /* possibly non fatal but unexpected overflow, only warn */
2010 WARN_ON_ONCE(stable_node->rmap_hlist_len >
2011 ksm_max_page_sharing);
2012
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08002013 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07002014 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08002015 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07002016
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08002017 if (rmap_item->hlist.next)
2018 ksm_pages_sharing++;
2019 else
2020 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07002021}
2022
2023/*
Hugh Dickins81464e302009-09-21 17:02:15 -07002024 * cmp_and_merge_page - first see if page can be merged into the stable tree;
2025 * if not, compare checksum to previous and if it's the same, see if page can
2026 * be inserted into the unstable tree, or merged with a page already there and
2027 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07002028 *
2029 * @page: the page that we are searching identical page to.
2030 * @rmap_item: the reverse mapping into the virtual address of this page
2031 */
2032static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
2033{
Kirill Tkhai4b229272017-10-03 16:14:27 -07002034 struct mm_struct *mm = rmap_item->mm;
Izik Eidus31dbd012009-09-21 17:02:03 -07002035 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08002036 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08002037 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08002038 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07002039 unsigned int checksum;
2040 int err;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07002041 bool max_page_sharing_bypass = false;
Izik Eidus31dbd012009-09-21 17:02:03 -07002042
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002043 stable_node = page_stable_node(page);
2044 if (stable_node) {
2045 if (stable_node->head != &migrate_nodes &&
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07002046 get_kpfn_nid(READ_ONCE(stable_node->kpfn)) !=
2047 NUMA(stable_node->nid)) {
2048 stable_node_dup_del(stable_node);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002049 stable_node->head = &migrate_nodes;
2050 list_add(&stable_node->list, stable_node->head);
2051 }
2052 if (stable_node->head != &migrate_nodes &&
2053 rmap_item->head == stable_node)
2054 return;
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07002055 /*
2056 * If it's a KSM fork, allow it to go over the sharing limit
2057 * without warnings.
2058 */
2059 if (!is_page_sharing_candidate(stable_node))
2060 max_page_sharing_bypass = true;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002061 }
Izik Eidus31dbd012009-09-21 17:02:03 -07002062
2063 /* We first start with searching the page inside the stable tree */
Hugh Dickins62b61f62009-12-14 17:59:33 -08002064 kpage = stable_tree_search(page);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002065 if (kpage == page && rmap_item->head == stable_node) {
2066 put_page(kpage);
2067 return;
2068 }
2069
2070 remove_rmap_item_from_tree(rmap_item);
2071
Hugh Dickins62b61f62009-12-14 17:59:33 -08002072 if (kpage) {
Hugh Dickins08beca42009-12-14 17:59:21 -08002073 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07002074 if (!err) {
2075 /*
2076 * The page was successfully merged:
2077 * add its rmap_item to the stable tree.
2078 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08002079 lock_page(kpage);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07002080 stable_tree_append(rmap_item, page_stable_node(kpage),
2081 max_page_sharing_bypass);
Hugh Dickins5ad64682009-12-14 17:59:24 -08002082 unlock_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07002083 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08002084 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07002085 return;
2086 }
2087
2088 /*
Hugh Dickins4035c07a2009-12-14 17:59:27 -08002089 * If the hash value of the page has changed from the last time
2090 * we calculated it, this page is changing frequently: therefore we
2091 * don't want to insert it in the unstable tree, and we don't want
2092 * to waste our time searching for something identical to it there.
Izik Eidus31dbd012009-09-21 17:02:03 -07002093 */
2094 checksum = calc_checksum(page);
2095 if (rmap_item->oldchecksum != checksum) {
2096 rmap_item->oldchecksum = checksum;
2097 return;
2098 }
2099
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08002100 /*
2101 * Same checksum as an empty page. We attempt to merge it with the
2102 * appropriate zero page if the user enabled this via sysfs.
2103 */
2104 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
2105 struct vm_area_struct *vma;
2106
Kirill Tkhai4b229272017-10-03 16:14:27 -07002107 down_read(&mm->mmap_sem);
2108 vma = find_mergeable_vma(mm, rmap_item->address);
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08002109 err = try_to_merge_one_page(vma, page,
2110 ZERO_PAGE(rmap_item->address));
Kirill Tkhai4b229272017-10-03 16:14:27 -07002111 up_read(&mm->mmap_sem);
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08002112 /*
2113 * In case of failure, the page was not really empty, so we
2114 * need to continue. Otherwise we're done.
2115 */
2116 if (!err)
2117 return;
2118 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08002119 tree_rmap_item =
2120 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07002121 if (tree_rmap_item) {
Claudio Imbrenda77da2ba2018-04-05 16:25:41 -07002122 bool split;
2123
Hugh Dickins8dd35572009-12-14 17:59:18 -08002124 kpage = try_to_merge_two_pages(rmap_item, page,
2125 tree_rmap_item, tree_page);
Claudio Imbrenda77da2ba2018-04-05 16:25:41 -07002126 /*
2127 * If both pages we tried to merge belong to the same compound
2128 * page, then we actually ended up increasing the reference
2129 * count of the same compound page twice, and split_huge_page
2130 * failed.
2131 * Here we set a flag if that happened, and we use it later to
2132 * try split_huge_page again. Since we call put_page right
2133 * afterwards, the reference count will be correct and
2134 * split_huge_page should succeed.
2135 */
2136 split = PageTransCompound(page)
2137 && compound_head(page) == compound_head(tree_page);
Hugh Dickins8dd35572009-12-14 17:59:18 -08002138 put_page(tree_page);
Hugh Dickins8dd35572009-12-14 17:59:18 -08002139 if (kpage) {
Hugh Dickinsbc566202013-02-22 16:36:06 -08002140 /*
2141 * The pages were successfully merged: insert new
2142 * node in the stable tree and add both rmap_items.
2143 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08002144 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08002145 stable_node = stable_tree_insert(kpage);
2146 if (stable_node) {
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07002147 stable_tree_append(tree_rmap_item, stable_node,
2148 false);
2149 stable_tree_append(rmap_item, stable_node,
2150 false);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08002151 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08002152 unlock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08002153
Izik Eidus31dbd012009-09-21 17:02:03 -07002154 /*
2155 * If we fail to insert the page into the stable tree,
2156 * we will have 2 virtual addresses that are pointing
2157 * to a ksm page left outside the stable tree,
2158 * in which case we need to break_cow on both.
2159 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08002160 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08002161 break_cow(tree_rmap_item);
2162 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07002163 }
Claudio Imbrenda77da2ba2018-04-05 16:25:41 -07002164 } else if (split) {
2165 /*
2166 * We are here if we tried to merge two pages and
2167 * failed because they both belonged to the same
2168 * compound page. We will split the page now, but no
2169 * merging will take place.
2170 * We do not want to add the cost of a full lock; if
2171 * the page is locked, it is better to skip it and
2172 * perhaps try again later.
2173 */
2174 if (!trylock_page(page))
2175 return;
2176 split_huge_page(page);
2177 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07002178 }
Izik Eidus31dbd012009-09-21 17:02:03 -07002179 }
2180}
2181
2182static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08002183 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07002184 unsigned long addr)
2185{
2186 struct rmap_item *rmap_item;
2187
Hugh Dickins6514d512009-12-14 17:59:19 -08002188 while (*rmap_list) {
2189 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08002190 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07002191 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07002192 if (rmap_item->address > addr)
2193 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08002194 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07002195 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07002196 free_rmap_item(rmap_item);
2197 }
2198
2199 rmap_item = alloc_rmap_item();
2200 if (rmap_item) {
2201 /* It has already been zeroed */
2202 rmap_item->mm = mm_slot->mm;
2203 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08002204 rmap_item->rmap_list = *rmap_list;
2205 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07002206 }
2207 return rmap_item;
2208}
2209
2210static struct rmap_item *scan_get_next_rmap_item(struct page **page)
2211{
2212 struct mm_struct *mm;
2213 struct mm_slot *slot;
2214 struct vm_area_struct *vma;
2215 struct rmap_item *rmap_item;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002216 int nid;
Izik Eidus31dbd012009-09-21 17:02:03 -07002217
2218 if (list_empty(&ksm_mm_head.mm_list))
2219 return NULL;
2220
2221 slot = ksm_scan.mm_slot;
2222 if (slot == &ksm_mm_head) {
Hugh Dickins2919bfd2011-01-13 15:47:29 -08002223 /*
2224 * A number of pages can hang around indefinitely on per-cpu
2225 * pagevecs, raised page count preventing write_protect_page
2226 * from merging them. Though it doesn't really matter much,
2227 * it is puzzling to see some stuck in pages_volatile until
2228 * other activity jostles them out, and they also prevented
2229 * LTP's KSM test from succeeding deterministically; so drain
2230 * them here (here rather than on entry to ksm_do_scan(),
2231 * so we don't IPI too often when pages_to_scan is set low).
2232 */
2233 lru_add_drain_all();
2234
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002235 /*
2236 * Whereas stale stable_nodes on the stable_tree itself
2237 * get pruned in the regular course of stable_tree_search(),
2238 * those moved out to the migrate_nodes list can accumulate:
2239 * so prune them once before each full scan.
2240 */
2241 if (!ksm_merge_across_nodes) {
Geliang Tang03640412016-01-14 15:20:54 -08002242 struct stable_node *stable_node, *next;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002243 struct page *page;
2244
Geliang Tang03640412016-01-14 15:20:54 -08002245 list_for_each_entry_safe(stable_node, next,
2246 &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002247 page = get_ksm_page(stable_node, false);
2248 if (page)
2249 put_page(page);
2250 cond_resched();
2251 }
2252 }
2253
Hugh Dickinsef53d162013-02-22 16:36:12 -08002254 for (nid = 0; nid < ksm_nr_node_ids; nid++)
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002255 root_unstable_tree[nid] = RB_ROOT;
Izik Eidus31dbd012009-09-21 17:02:03 -07002256
2257 spin_lock(&ksm_mmlist_lock);
2258 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
2259 ksm_scan.mm_slot = slot;
2260 spin_unlock(&ksm_mmlist_lock);
Hugh Dickins2b472612011-06-15 15:08:58 -07002261 /*
2262 * Although we tested list_empty() above, a racing __ksm_exit
2263 * of the last mm on the list may have removed it since then.
2264 */
2265 if (slot == &ksm_mm_head)
2266 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07002267next_mm:
2268 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08002269 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07002270 }
2271
2272 mm = slot->mm;
2273 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07002274 if (ksm_test_exit(mm))
2275 vma = NULL;
2276 else
2277 vma = find_vma(mm, ksm_scan.address);
2278
2279 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07002280 if (!(vma->vm_flags & VM_MERGEABLE))
2281 continue;
2282 if (ksm_scan.address < vma->vm_start)
2283 ksm_scan.address = vma->vm_start;
2284 if (!vma->anon_vma)
2285 ksm_scan.address = vma->vm_end;
2286
2287 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07002288 if (ksm_test_exit(mm))
2289 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07002290 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08002291 if (IS_ERR_OR_NULL(*page)) {
2292 ksm_scan.address += PAGE_SIZE;
2293 cond_resched();
2294 continue;
2295 }
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08002296 if (PageAnon(*page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -07002297 flush_anon_page(vma, *page, ksm_scan.address);
2298 flush_dcache_page(*page);
2299 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08002300 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07002301 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08002302 ksm_scan.rmap_list =
2303 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07002304 ksm_scan.address += PAGE_SIZE;
2305 } else
2306 put_page(*page);
2307 up_read(&mm->mmap_sem);
2308 return rmap_item;
2309 }
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08002310 put_page(*page);
Izik Eidus31dbd012009-09-21 17:02:03 -07002311 ksm_scan.address += PAGE_SIZE;
2312 cond_resched();
2313 }
2314 }
2315
Hugh Dickins9ba69292009-09-21 17:02:20 -07002316 if (ksm_test_exit(mm)) {
2317 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08002318 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002319 }
Izik Eidus31dbd012009-09-21 17:02:03 -07002320 /*
2321 * Nuke all the rmap_items that are above this current rmap:
2322 * because there were no VM_MERGEABLE vmas with such addresses.
2323 */
Hugh Dickins6514d512009-12-14 17:59:19 -08002324 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07002325
2326 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07002327 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
2328 struct mm_slot, mm_list);
2329 if (ksm_scan.address == 0) {
2330 /*
2331 * We've completed a full scan of all vmas, holding mmap_sem
2332 * throughout, and found no VM_MERGEABLE: so do the same as
2333 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07002334 * This applies either when cleaning up after __ksm_exit
2335 * (but beware: we can reach here even before __ksm_exit),
2336 * or when all VM_MERGEABLE areas have been unmapped (and
2337 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07002338 */
Sasha Levin4ca3a692013-02-22 16:32:28 -08002339 hash_del(&slot->link);
Hugh Dickinscd551f92009-09-21 17:02:17 -07002340 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07002341 spin_unlock(&ksm_mmlist_lock);
2342
Hugh Dickinscd551f92009-09-21 17:02:17 -07002343 free_mm_slot(slot);
2344 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07002345 up_read(&mm->mmap_sem);
2346 mmdrop(mm);
2347 } else {
Hugh Dickins9ba69292009-09-21 17:02:20 -07002348 up_read(&mm->mmap_sem);
Zhou Chengming7496fea2016-05-12 15:42:21 -07002349 /*
2350 * up_read(&mm->mmap_sem) first because after
2351 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
2352 * already have been freed under us by __ksm_exit()
2353 * because the "mm_slot" is still hashed and
2354 * ksm_scan.mm_slot doesn't point to it anymore.
2355 */
2356 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07002357 }
Izik Eidus31dbd012009-09-21 17:02:03 -07002358
2359 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07002360 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07002361 if (slot != &ksm_mm_head)
2362 goto next_mm;
2363
Izik Eidus31dbd012009-09-21 17:02:03 -07002364 ksm_scan.seqnr++;
2365 return NULL;
2366}
2367
2368/**
2369 * ksm_do_scan - the ksm scanner main worker function.
Mike Rapoportb7701a52018-02-06 15:42:13 -08002370 * @scan_npages: number of pages we want to scan before we return.
Izik Eidus31dbd012009-09-21 17:02:03 -07002371 */
2372static void ksm_do_scan(unsigned int scan_npages)
2373{
2374 struct rmap_item *rmap_item;
Dan Carpenter22eccdd2010-04-23 13:18:10 -04002375 struct page *uninitialized_var(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07002376
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002377 while (scan_npages-- && likely(!freezing(current))) {
Izik Eidus31dbd012009-09-21 17:02:03 -07002378 cond_resched();
2379 rmap_item = scan_get_next_rmap_item(&page);
2380 if (!rmap_item)
2381 return;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002382 cmp_and_merge_page(page, rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07002383 put_page(page);
2384 }
2385}
2386
Hugh Dickins6e1583842009-09-21 17:02:14 -07002387static int ksmd_should_run(void)
2388{
2389 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
2390}
2391
Izik Eidus31dbd012009-09-21 17:02:03 -07002392static int ksm_scan_thread(void *nothing)
2393{
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002394 set_freezable();
Izik Eidus339aa622009-09-21 17:02:07 -07002395 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07002396
2397 while (!kthread_should_stop()) {
Hugh Dickins6e1583842009-09-21 17:02:14 -07002398 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002399 wait_while_offlining();
Hugh Dickins6e1583842009-09-21 17:02:14 -07002400 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07002401 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e1583842009-09-21 17:02:14 -07002402 mutex_unlock(&ksm_thread_mutex);
2403
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002404 try_to_freeze();
2405
Hugh Dickins6e1583842009-09-21 17:02:14 -07002406 if (ksmd_should_run()) {
Izik Eidus31dbd012009-09-21 17:02:03 -07002407 schedule_timeout_interruptible(
2408 msecs_to_jiffies(ksm_thread_sleep_millisecs));
2409 } else {
Andrea Arcangeli878aee72011-01-13 15:47:10 -08002410 wait_event_freezable(ksm_thread_wait,
Hugh Dickins6e1583842009-09-21 17:02:14 -07002411 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07002412 }
2413 }
2414 return 0;
2415}
2416
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07002417int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
2418 unsigned long end, int advice, unsigned long *vm_flags)
2419{
2420 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07002421 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07002422
2423 switch (advice) {
2424 case MADV_MERGEABLE:
2425 /*
2426 * Be somewhat over-protective for now!
2427 */
2428 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
2429 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
Kirill A. Shutemov0661a332015-02-10 14:10:04 -08002430 VM_HUGETLB | VM_MIXEDMAP))
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07002431 return 0; /* just ignore the advice */
2432
Dave Jiange1fb4a02018-08-17 15:43:40 -07002433 if (vma_is_dax(vma))
2434 return 0;
2435
Konstantin Khlebnikovcc2383e2012-10-08 16:28:37 -07002436#ifdef VM_SAO
2437 if (*vm_flags & VM_SAO)
2438 return 0;
2439#endif
Khalid Aziz74a04962018-02-23 15:46:41 -07002440#ifdef VM_SPARC_ADI
2441 if (*vm_flags & VM_SPARC_ADI)
2442 return 0;
2443#endif
Konstantin Khlebnikovcc2383e2012-10-08 16:28:37 -07002444
Hugh Dickinsd952b792009-09-21 17:02:16 -07002445 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
2446 err = __ksm_enter(mm);
2447 if (err)
2448 return err;
2449 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07002450
2451 *vm_flags |= VM_MERGEABLE;
2452 break;
2453
2454 case MADV_UNMERGEABLE:
2455 if (!(*vm_flags & VM_MERGEABLE))
2456 return 0; /* just ignore the advice */
2457
Hugh Dickinsd952b792009-09-21 17:02:16 -07002458 if (vma->anon_vma) {
2459 err = unmerge_ksm_pages(vma, start, end);
2460 if (err)
2461 return err;
2462 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07002463
2464 *vm_flags &= ~VM_MERGEABLE;
2465 break;
2466 }
2467
2468 return 0;
2469}
2470
2471int __ksm_enter(struct mm_struct *mm)
2472{
Hugh Dickins6e1583842009-09-21 17:02:14 -07002473 struct mm_slot *mm_slot;
2474 int needs_wakeup;
2475
2476 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07002477 if (!mm_slot)
2478 return -ENOMEM;
2479
Hugh Dickins6e1583842009-09-21 17:02:14 -07002480 /* Check ksm_run too? Would need tighter locking */
2481 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
2482
Izik Eidus31dbd012009-09-21 17:02:03 -07002483 spin_lock(&ksm_mmlist_lock);
2484 insert_to_mm_slots_hash(mm, mm_slot);
2485 /*
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002486 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
2487 * insert just behind the scanning cursor, to let the area settle
Izik Eidus31dbd012009-09-21 17:02:03 -07002488 * down a little; when fork is followed by immediate exec, we don't
2489 * want ksmd to waste time setting up and tearing down an rmap_list.
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002490 *
2491 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
2492 * scanning cursor, otherwise KSM pages in newly forked mms will be
2493 * missed: then we might as well insert at the end of the list.
Izik Eidus31dbd012009-09-21 17:02:03 -07002494 */
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002495 if (ksm_run & KSM_RUN_UNMERGE)
2496 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
2497 else
2498 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07002499 spin_unlock(&ksm_mmlist_lock);
2500
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07002501 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Vegard Nossumf1f10072017-02-27 14:30:07 -08002502 mmgrab(mm);
Hugh Dickins6e1583842009-09-21 17:02:14 -07002503
2504 if (needs_wakeup)
2505 wake_up_interruptible(&ksm_thread_wait);
2506
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07002507 return 0;
2508}
2509
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07002510void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07002511{
Hugh Dickinscd551f92009-09-21 17:02:17 -07002512 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002513 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07002514
Izik Eidus31dbd012009-09-21 17:02:03 -07002515 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07002516 * This process is exiting: if it's straightforward (as is the
2517 * case when ksmd was never running), free mm_slot immediately.
2518 * But if it's at the cursor or has rmap_items linked to it, use
2519 * mmap_sem to synchronize with any break_cows before pagetables
2520 * are freed, and leave the mm_slot on the list for ksmd to free.
2521 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07002522 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07002523
Hugh Dickinscd551f92009-09-21 17:02:17 -07002524 spin_lock(&ksm_mmlist_lock);
2525 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07002526 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08002527 if (!mm_slot->rmap_list) {
Sasha Levin4ca3a692013-02-22 16:32:28 -08002528 hash_del(&mm_slot->link);
Hugh Dickins9ba69292009-09-21 17:02:20 -07002529 list_del(&mm_slot->mm_list);
2530 easy_to_free = 1;
2531 } else {
2532 list_move(&mm_slot->mm_list,
2533 &ksm_scan.mm_slot->mm_list);
2534 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07002535 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07002536 spin_unlock(&ksm_mmlist_lock);
2537
Hugh Dickins9ba69292009-09-21 17:02:20 -07002538 if (easy_to_free) {
2539 free_mm_slot(mm_slot);
2540 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
2541 mmdrop(mm);
2542 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07002543 down_write(&mm->mmap_sem);
2544 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07002545 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07002546}
Izik Eidus31dbd012009-09-21 17:02:03 -07002547
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002548struct page *ksm_might_need_to_copy(struct page *page,
Hugh Dickins5ad64682009-12-14 17:59:24 -08002549 struct vm_area_struct *vma, unsigned long address)
2550{
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002551 struct anon_vma *anon_vma = page_anon_vma(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08002552 struct page *new_page;
2553
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002554 if (PageKsm(page)) {
2555 if (page_stable_node(page) &&
2556 !(ksm_run & KSM_RUN_UNMERGE))
2557 return page; /* no need to copy it */
2558 } else if (!anon_vma) {
2559 return page; /* no need to copy it */
2560 } else if (anon_vma->root == vma->anon_vma->root &&
2561 page->index == linear_page_index(vma, address)) {
2562 return page; /* still no need to copy it */
2563 }
2564 if (!PageUptodate(page))
2565 return page; /* let do_swap_page report the error */
2566
Hugh Dickins5ad64682009-12-14 17:59:24 -08002567 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2568 if (new_page) {
2569 copy_user_highpage(new_page, page, address, vma);
2570
2571 SetPageDirty(new_page);
2572 __SetPageUptodate(new_page);
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08002573 __SetPageLocked(new_page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08002574 }
2575
Hugh Dickins5ad64682009-12-14 17:59:24 -08002576 return new_page;
2577}
2578
Minchan Kim1df631a2017-05-03 14:54:23 -07002579void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002580{
2581 struct stable_node *stable_node;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002582 struct rmap_item *rmap_item;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002583 int search_new_forks = 0;
2584
Sasha Levin309381fea2014-01-23 15:52:54 -08002585 VM_BUG_ON_PAGE(!PageKsm(page), page);
Joonsoo Kim9f326242014-01-21 15:49:53 -08002586
2587 /*
2588 * Rely on the page lock to protect against concurrent modifications
2589 * to that page's node of the stable tree.
2590 */
Sasha Levin309381fea2014-01-23 15:52:54 -08002591 VM_BUG_ON_PAGE(!PageLocked(page), page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002592
2593 stable_node = page_stable_node(page);
2594 if (!stable_node)
Minchan Kim1df631a2017-05-03 14:54:23 -07002595 return;
Minchan Kim5a83f942013-05-09 16:21:27 +09002596
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002597again:
Sasha Levinb67bfe02013-02-27 17:06:00 -08002598 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002599 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08002600 struct anon_vma_chain *vmac;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002601 struct vm_area_struct *vma;
2602
Andrea Arcangeliad126952015-11-05 18:49:07 -08002603 cond_resched();
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08002604 anon_vma_lock_read(anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -07002605 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2606 0, ULONG_MAX) {
Jia He1105a2f2018-06-14 15:26:14 -07002607 unsigned long addr;
2608
Andrea Arcangeliad126952015-11-05 18:49:07 -08002609 cond_resched();
Rik van Riel5beb4932010-03-05 13:42:07 -08002610 vma = vmac->vma;
Jia He1105a2f2018-06-14 15:26:14 -07002611
2612 /* Ignore the stable/unstable/sqnr flags */
2613 addr = rmap_item->address & ~KSM_FLAG_MASK;
2614
2615 if (addr < vma->vm_start || addr >= vma->vm_end)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002616 continue;
2617 /*
2618 * Initially we examine only the vma which covers this
2619 * rmap_item; but later, if there is still work to do,
2620 * we examine covering vmas in other mms: in case they
2621 * were forked from the original since ksmd passed.
2622 */
2623 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2624 continue;
2625
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08002626 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2627 continue;
2628
Jia He1105a2f2018-06-14 15:26:14 -07002629 if (!rwc->rmap_one(page, vma, addr, rwc->arg)) {
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08002630 anon_vma_unlock_read(anon_vma);
Minchan Kim1df631a2017-05-03 14:54:23 -07002631 return;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002632 }
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08002633 if (rwc->done && rwc->done(page)) {
2634 anon_vma_unlock_read(anon_vma);
Minchan Kim1df631a2017-05-03 14:54:23 -07002635 return;
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08002636 }
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002637 }
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08002638 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002639 }
2640 if (!search_new_forks++)
2641 goto again;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002642}
2643
Joonsoo Kim52629502014-01-21 15:49:50 -08002644#ifdef CONFIG_MIGRATION
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002645void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2646{
2647 struct stable_node *stable_node;
2648
Sasha Levin309381fea2014-01-23 15:52:54 -08002649 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2650 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2651 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002652
2653 stable_node = page_stable_node(newpage);
2654 if (stable_node) {
Sasha Levin309381fea2014-01-23 15:52:54 -08002655 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002656 stable_node->kpfn = page_to_pfn(newpage);
Hugh Dickinsc8d65532013-02-22 16:35:10 -08002657 /*
2658 * newpage->mapping was set in advance; now we need smp_wmb()
2659 * to make sure that the new stable_node->kpfn is visible
2660 * to get_ksm_page() before it can see that oldpage->mapping
2661 * has gone stale (or that PageSwapCache has been cleared).
2662 */
2663 smp_wmb();
2664 set_page_stable_node(oldpage, NULL);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002665 }
2666}
2667#endif /* CONFIG_MIGRATION */
2668
Hugh Dickins62b61f62009-12-14 17:59:33 -08002669#ifdef CONFIG_MEMORY_HOTREMOVE
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002670static void wait_while_offlining(void)
2671{
2672 while (ksm_run & KSM_RUN_OFFLINE) {
2673 mutex_unlock(&ksm_thread_mutex);
2674 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
NeilBrown74316202014-07-07 15:16:04 +10002675 TASK_UNINTERRUPTIBLE);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002676 mutex_lock(&ksm_thread_mutex);
2677 }
2678}
2679
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07002680static bool stable_node_dup_remove_range(struct stable_node *stable_node,
2681 unsigned long start_pfn,
2682 unsigned long end_pfn)
2683{
2684 if (stable_node->kpfn >= start_pfn &&
2685 stable_node->kpfn < end_pfn) {
2686 /*
2687 * Don't get_ksm_page, page has already gone:
2688 * which is why we keep kpfn instead of page*
2689 */
2690 remove_node_from_stable_tree(stable_node);
2691 return true;
2692 }
2693 return false;
2694}
2695
2696static bool stable_node_chain_remove_range(struct stable_node *stable_node,
2697 unsigned long start_pfn,
2698 unsigned long end_pfn,
2699 struct rb_root *root)
2700{
2701 struct stable_node *dup;
2702 struct hlist_node *hlist_safe;
2703
2704 if (!is_stable_node_chain(stable_node)) {
2705 VM_BUG_ON(is_stable_node_dup(stable_node));
2706 return stable_node_dup_remove_range(stable_node, start_pfn,
2707 end_pfn);
2708 }
2709
2710 hlist_for_each_entry_safe(dup, hlist_safe,
2711 &stable_node->hlist, hlist_dup) {
2712 VM_BUG_ON(!is_stable_node_dup(dup));
2713 stable_node_dup_remove_range(dup, start_pfn, end_pfn);
2714 }
2715 if (hlist_empty(&stable_node->hlist)) {
2716 free_stable_node_chain(stable_node, root);
2717 return true; /* notify caller that tree was rebalanced */
2718 } else
2719 return false;
2720}
2721
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002722static void ksm_check_stable_tree(unsigned long start_pfn,
2723 unsigned long end_pfn)
Hugh Dickins62b61f62009-12-14 17:59:33 -08002724{
Geliang Tang03640412016-01-14 15:20:54 -08002725 struct stable_node *stable_node, *next;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002726 struct rb_node *node;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002727 int nid;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002728
Hugh Dickinsef53d162013-02-22 16:36:12 -08002729 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2730 node = rb_first(root_stable_tree + nid);
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002731 while (node) {
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002732 stable_node = rb_entry(node, struct stable_node, node);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07002733 if (stable_node_chain_remove_range(stable_node,
2734 start_pfn, end_pfn,
2735 root_stable_tree +
2736 nid))
Hugh Dickinsef53d162013-02-22 16:36:12 -08002737 node = rb_first(root_stable_tree + nid);
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07002738 else
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002739 node = rb_next(node);
2740 cond_resched();
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002741 }
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002742 }
Geliang Tang03640412016-01-14 15:20:54 -08002743 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002744 if (stable_node->kpfn >= start_pfn &&
2745 stable_node->kpfn < end_pfn)
2746 remove_node_from_stable_tree(stable_node);
2747 cond_resched();
2748 }
Hugh Dickins62b61f62009-12-14 17:59:33 -08002749}
2750
2751static int ksm_memory_callback(struct notifier_block *self,
2752 unsigned long action, void *arg)
2753{
2754 struct memory_notify *mn = arg;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002755
2756 switch (action) {
2757 case MEM_GOING_OFFLINE:
2758 /*
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002759 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2760 * and remove_all_stable_nodes() while memory is going offline:
2761 * it is unsafe for them to touch the stable tree at this time.
2762 * But unmerge_ksm_pages(), rmap lookups and other entry points
2763 * which do not need the ksm_thread_mutex are all safe.
Hugh Dickins62b61f62009-12-14 17:59:33 -08002764 */
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002765 mutex_lock(&ksm_thread_mutex);
2766 ksm_run |= KSM_RUN_OFFLINE;
2767 mutex_unlock(&ksm_thread_mutex);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002768 break;
2769
2770 case MEM_OFFLINE:
2771 /*
2772 * Most of the work is done by page migration; but there might
2773 * be a few stable_nodes left over, still pointing to struct
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002774 * pages which have been offlined: prune those from the tree,
2775 * otherwise get_ksm_page() might later try to access a
2776 * non-existent struct page.
Hugh Dickins62b61f62009-12-14 17:59:33 -08002777 */
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002778 ksm_check_stable_tree(mn->start_pfn,
2779 mn->start_pfn + mn->nr_pages);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002780 /* fallthrough */
2781
2782 case MEM_CANCEL_OFFLINE:
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002783 mutex_lock(&ksm_thread_mutex);
2784 ksm_run &= ~KSM_RUN_OFFLINE;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002785 mutex_unlock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002786
2787 smp_mb(); /* wake_up_bit advises this */
2788 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
Hugh Dickins62b61f62009-12-14 17:59:33 -08002789 break;
2790 }
2791 return NOTIFY_OK;
2792}
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002793#else
2794static void wait_while_offlining(void)
2795{
2796}
Hugh Dickins62b61f62009-12-14 17:59:33 -08002797#endif /* CONFIG_MEMORY_HOTREMOVE */
2798
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002799#ifdef CONFIG_SYSFS
2800/*
2801 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2802 */
2803
Izik Eidus31dbd012009-09-21 17:02:03 -07002804#define KSM_ATTR_RO(_name) \
2805 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2806#define KSM_ATTR(_name) \
2807 static struct kobj_attribute _name##_attr = \
2808 __ATTR(_name, 0644, _name##_show, _name##_store)
2809
2810static ssize_t sleep_millisecs_show(struct kobject *kobj,
2811 struct kobj_attribute *attr, char *buf)
2812{
2813 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2814}
2815
2816static ssize_t sleep_millisecs_store(struct kobject *kobj,
2817 struct kobj_attribute *attr,
2818 const char *buf, size_t count)
2819{
2820 unsigned long msecs;
2821 int err;
2822
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002823 err = kstrtoul(buf, 10, &msecs);
Izik Eidus31dbd012009-09-21 17:02:03 -07002824 if (err || msecs > UINT_MAX)
2825 return -EINVAL;
2826
2827 ksm_thread_sleep_millisecs = msecs;
2828
2829 return count;
2830}
2831KSM_ATTR(sleep_millisecs);
2832
2833static ssize_t pages_to_scan_show(struct kobject *kobj,
2834 struct kobj_attribute *attr, char *buf)
2835{
2836 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2837}
2838
2839static ssize_t pages_to_scan_store(struct kobject *kobj,
2840 struct kobj_attribute *attr,
2841 const char *buf, size_t count)
2842{
2843 int err;
2844 unsigned long nr_pages;
2845
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002846 err = kstrtoul(buf, 10, &nr_pages);
Izik Eidus31dbd012009-09-21 17:02:03 -07002847 if (err || nr_pages > UINT_MAX)
2848 return -EINVAL;
2849
2850 ksm_thread_pages_to_scan = nr_pages;
2851
2852 return count;
2853}
2854KSM_ATTR(pages_to_scan);
2855
2856static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2857 char *buf)
2858{
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002859 return sprintf(buf, "%lu\n", ksm_run);
Izik Eidus31dbd012009-09-21 17:02:03 -07002860}
2861
2862static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2863 const char *buf, size_t count)
2864{
2865 int err;
2866 unsigned long flags;
2867
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002868 err = kstrtoul(buf, 10, &flags);
Izik Eidus31dbd012009-09-21 17:02:03 -07002869 if (err || flags > UINT_MAX)
2870 return -EINVAL;
2871 if (flags > KSM_RUN_UNMERGE)
2872 return -EINVAL;
2873
2874 /*
2875 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2876 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsd0f209f2009-12-14 17:59:34 -08002877 * breaking COW to free the pages_shared (but leaves mm_slots
2878 * on the list for when ksmd may be set running again).
Izik Eidus31dbd012009-09-21 17:02:03 -07002879 */
2880
2881 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002882 wait_while_offlining();
Izik Eidus31dbd012009-09-21 17:02:03 -07002883 if (ksm_run != flags) {
2884 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07002885 if (flags & KSM_RUN_UNMERGE) {
David Rientjese1e12d22012-12-11 16:02:56 -08002886 set_current_oom_origin();
Hugh Dickinsd952b792009-09-21 17:02:16 -07002887 err = unmerge_and_remove_all_rmap_items();
David Rientjese1e12d22012-12-11 16:02:56 -08002888 clear_current_oom_origin();
Hugh Dickinsd952b792009-09-21 17:02:16 -07002889 if (err) {
2890 ksm_run = KSM_RUN_STOP;
2891 count = err;
2892 }
2893 }
Izik Eidus31dbd012009-09-21 17:02:03 -07002894 }
2895 mutex_unlock(&ksm_thread_mutex);
2896
2897 if (flags & KSM_RUN_MERGE)
2898 wake_up_interruptible(&ksm_thread_wait);
2899
2900 return count;
2901}
2902KSM_ATTR(run);
2903
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002904#ifdef CONFIG_NUMA
2905static ssize_t merge_across_nodes_show(struct kobject *kobj,
2906 struct kobj_attribute *attr, char *buf)
2907{
2908 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2909}
2910
2911static ssize_t merge_across_nodes_store(struct kobject *kobj,
2912 struct kobj_attribute *attr,
2913 const char *buf, size_t count)
2914{
2915 int err;
2916 unsigned long knob;
2917
2918 err = kstrtoul(buf, 10, &knob);
2919 if (err)
2920 return err;
2921 if (knob > 1)
2922 return -EINVAL;
2923
2924 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002925 wait_while_offlining();
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002926 if (ksm_merge_across_nodes != knob) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002927 if (ksm_pages_shared || remove_all_stable_nodes())
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002928 err = -EBUSY;
Hugh Dickinsef53d162013-02-22 16:36:12 -08002929 else if (root_stable_tree == one_stable_tree) {
2930 struct rb_root *buf;
2931 /*
2932 * This is the first time that we switch away from the
2933 * default of merging across nodes: must now allocate
2934 * a buffer to hold as many roots as may be needed.
2935 * Allocate stable and unstable together:
2936 * MAXSMP NODES_SHIFT 10 will use 16kB.
2937 */
Joe Perchesbafe1e12013-11-12 15:07:10 -08002938 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2939 GFP_KERNEL);
Hugh Dickinsef53d162013-02-22 16:36:12 -08002940 /* Let us assume that RB_ROOT is NULL is zero */
2941 if (!buf)
2942 err = -ENOMEM;
2943 else {
2944 root_stable_tree = buf;
2945 root_unstable_tree = buf + nr_node_ids;
2946 /* Stable tree is empty but not the unstable */
2947 root_unstable_tree[0] = one_unstable_tree[0];
2948 }
2949 }
2950 if (!err) {
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002951 ksm_merge_across_nodes = knob;
Hugh Dickinsef53d162013-02-22 16:36:12 -08002952 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2953 }
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002954 }
2955 mutex_unlock(&ksm_thread_mutex);
2956
2957 return err ? err : count;
2958}
2959KSM_ATTR(merge_across_nodes);
2960#endif
2961
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08002962static ssize_t use_zero_pages_show(struct kobject *kobj,
2963 struct kobj_attribute *attr, char *buf)
2964{
2965 return sprintf(buf, "%u\n", ksm_use_zero_pages);
2966}
2967static ssize_t use_zero_pages_store(struct kobject *kobj,
2968 struct kobj_attribute *attr,
2969 const char *buf, size_t count)
2970{
2971 int err;
2972 bool value;
2973
2974 err = kstrtobool(buf, &value);
2975 if (err)
2976 return -EINVAL;
2977
2978 ksm_use_zero_pages = value;
2979
2980 return count;
2981}
2982KSM_ATTR(use_zero_pages);
2983
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07002984static ssize_t max_page_sharing_show(struct kobject *kobj,
2985 struct kobj_attribute *attr, char *buf)
2986{
2987 return sprintf(buf, "%u\n", ksm_max_page_sharing);
2988}
2989
2990static ssize_t max_page_sharing_store(struct kobject *kobj,
2991 struct kobj_attribute *attr,
2992 const char *buf, size_t count)
2993{
2994 int err;
2995 int knob;
2996
2997 err = kstrtoint(buf, 10, &knob);
2998 if (err)
2999 return err;
3000 /*
3001 * When a KSM page is created it is shared by 2 mappings. This
3002 * being a signed comparison, it implicitly verifies it's not
3003 * negative.
3004 */
3005 if (knob < 2)
3006 return -EINVAL;
3007
3008 if (READ_ONCE(ksm_max_page_sharing) == knob)
3009 return count;
3010
3011 mutex_lock(&ksm_thread_mutex);
3012 wait_while_offlining();
3013 if (ksm_max_page_sharing != knob) {
3014 if (ksm_pages_shared || remove_all_stable_nodes())
3015 err = -EBUSY;
3016 else
3017 ksm_max_page_sharing = knob;
3018 }
3019 mutex_unlock(&ksm_thread_mutex);
3020
3021 return err ? err : count;
3022}
3023KSM_ATTR(max_page_sharing);
3024
Hugh Dickinsb4028262009-09-21 17:02:09 -07003025static ssize_t pages_shared_show(struct kobject *kobj,
3026 struct kobj_attribute *attr, char *buf)
3027{
3028 return sprintf(buf, "%lu\n", ksm_pages_shared);
3029}
3030KSM_ATTR_RO(pages_shared);
3031
3032static ssize_t pages_sharing_show(struct kobject *kobj,
3033 struct kobj_attribute *attr, char *buf)
3034{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07003035 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07003036}
3037KSM_ATTR_RO(pages_sharing);
3038
Hugh Dickins473b0ce2009-09-21 17:02:11 -07003039static ssize_t pages_unshared_show(struct kobject *kobj,
3040 struct kobj_attribute *attr, char *buf)
3041{
3042 return sprintf(buf, "%lu\n", ksm_pages_unshared);
3043}
3044KSM_ATTR_RO(pages_unshared);
3045
3046static ssize_t pages_volatile_show(struct kobject *kobj,
3047 struct kobj_attribute *attr, char *buf)
3048{
3049 long ksm_pages_volatile;
3050
3051 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
3052 - ksm_pages_sharing - ksm_pages_unshared;
3053 /*
3054 * It was not worth any locking to calculate that statistic,
3055 * but it might therefore sometimes be negative: conceal that.
3056 */
3057 if (ksm_pages_volatile < 0)
3058 ksm_pages_volatile = 0;
3059 return sprintf(buf, "%ld\n", ksm_pages_volatile);
3060}
3061KSM_ATTR_RO(pages_volatile);
3062
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07003063static ssize_t stable_node_dups_show(struct kobject *kobj,
3064 struct kobj_attribute *attr, char *buf)
3065{
3066 return sprintf(buf, "%lu\n", ksm_stable_node_dups);
3067}
3068KSM_ATTR_RO(stable_node_dups);
3069
3070static ssize_t stable_node_chains_show(struct kobject *kobj,
3071 struct kobj_attribute *attr, char *buf)
3072{
3073 return sprintf(buf, "%lu\n", ksm_stable_node_chains);
3074}
3075KSM_ATTR_RO(stable_node_chains);
3076
3077static ssize_t
3078stable_node_chains_prune_millisecs_show(struct kobject *kobj,
3079 struct kobj_attribute *attr,
3080 char *buf)
3081{
3082 return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs);
3083}
3084
3085static ssize_t
3086stable_node_chains_prune_millisecs_store(struct kobject *kobj,
3087 struct kobj_attribute *attr,
3088 const char *buf, size_t count)
3089{
3090 unsigned long msecs;
3091 int err;
3092
3093 err = kstrtoul(buf, 10, &msecs);
3094 if (err || msecs > UINT_MAX)
3095 return -EINVAL;
3096
3097 ksm_stable_node_chains_prune_millisecs = msecs;
3098
3099 return count;
3100}
3101KSM_ATTR(stable_node_chains_prune_millisecs);
3102
Hugh Dickins473b0ce2009-09-21 17:02:11 -07003103static ssize_t full_scans_show(struct kobject *kobj,
3104 struct kobj_attribute *attr, char *buf)
3105{
3106 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
3107}
3108KSM_ATTR_RO(full_scans);
3109
Izik Eidus31dbd012009-09-21 17:02:03 -07003110static struct attribute *ksm_attrs[] = {
3111 &sleep_millisecs_attr.attr,
3112 &pages_to_scan_attr.attr,
3113 &run_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07003114 &pages_shared_attr.attr,
3115 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07003116 &pages_unshared_attr.attr,
3117 &pages_volatile_attr.attr,
3118 &full_scans_attr.attr,
Petr Holasek90bd6fd2013-02-22 16:35:00 -08003119#ifdef CONFIG_NUMA
3120 &merge_across_nodes_attr.attr,
3121#endif
Andrea Arcangeli2c653d02017-07-06 15:36:55 -07003122 &max_page_sharing_attr.attr,
3123 &stable_node_chains_attr.attr,
3124 &stable_node_dups_attr.attr,
3125 &stable_node_chains_prune_millisecs_attr.attr,
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08003126 &use_zero_pages_attr.attr,
Izik Eidus31dbd012009-09-21 17:02:03 -07003127 NULL,
3128};
3129
Arvind Yadavf907c262017-09-06 16:21:53 -07003130static const struct attribute_group ksm_attr_group = {
Izik Eidus31dbd012009-09-21 17:02:03 -07003131 .attrs = ksm_attrs,
3132 .name = "ksm",
3133};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07003134#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07003135
3136static int __init ksm_init(void)
3137{
3138 struct task_struct *ksm_thread;
3139 int err;
3140
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08003141 /* The correct value depends on page size and endianness */
3142 zero_checksum = calc_checksum(ZERO_PAGE(0));
3143 /* Default to false for backwards compatibility */
3144 ksm_use_zero_pages = false;
3145
Izik Eidus31dbd012009-09-21 17:02:03 -07003146 err = ksm_slab_init();
3147 if (err)
3148 goto out;
3149
Izik Eidus31dbd012009-09-21 17:02:03 -07003150 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
3151 if (IS_ERR(ksm_thread)) {
Paul McQuade25acde32014-10-09 15:29:09 -07003152 pr_err("ksm: creating kthread failed\n");
Izik Eidus31dbd012009-09-21 17:02:03 -07003153 err = PTR_ERR(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07003154 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07003155 }
3156
Hugh Dickins2ffd8672009-09-21 17:02:23 -07003157#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07003158 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
3159 if (err) {
Paul McQuade25acde32014-10-09 15:29:09 -07003160 pr_err("ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07003161 kthread_stop(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07003162 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07003163 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07003164#else
3165 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
3166
Hugh Dickins2ffd8672009-09-21 17:02:23 -07003167#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07003168
Hugh Dickins62b61f62009-12-14 17:59:33 -08003169#ifdef CONFIG_MEMORY_HOTREMOVE
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08003170 /* There is no significance to this priority 100 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08003171 hotplug_memory_notifier(ksm_memory_callback, 100);
3172#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07003173 return 0;
3174
Lai Jiangshand9f89842010-08-09 17:20:02 -07003175out_free:
Izik Eidus31dbd012009-09-21 17:02:03 -07003176 ksm_slab_free();
3177out:
3178 return err;
3179}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -08003180subsys_initcall(ksm_init);