blob: d9fc0e4561283d9a351f6dd7c4cfb74ad26ab566 [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
Izik Eidus31dbd012009-09-21 17:02:03 -070054/*
55 * A few notes about the KSM scanning process,
56 * to make it easier to understand the data structures below:
57 *
58 * In order to reduce excessive scanning, KSM sorts the memory pages by their
59 * contents into a data structure that holds pointers to the pages' locations.
60 *
61 * Since the contents of the pages may change at any moment, KSM cannot just
62 * insert the pages into a normal sorted tree and expect it to find anything.
63 * Therefore KSM uses two data structures - the stable and the unstable tree.
64 *
65 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
66 * by their contents. Because each such page is write-protected, searching on
67 * this tree is fully assured to be working (except when pages are unmapped),
68 * and therefore this tree is called the stable tree.
69 *
70 * In addition to the stable tree, KSM uses a second data structure called the
71 * unstable tree: this tree holds pointers to pages which have been found to
72 * be "unchanged for a period of time". The unstable tree sorts these pages
73 * by their contents, but since they are not write-protected, KSM cannot rely
74 * upon the unstable tree to work correctly - the unstable tree is liable to
75 * be corrupted as its contents are modified, and so it is called unstable.
76 *
77 * KSM solves this problem by several techniques:
78 *
79 * 1) The unstable tree is flushed every time KSM completes scanning all
80 * memory areas, and then the tree is rebuilt again from the beginning.
81 * 2) KSM will only insert into the unstable tree, pages whose hash value
82 * has not changed since the previous scan of all memory areas.
83 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
84 * colors of the nodes and not on their contents, assuring that even when
85 * the tree gets "corrupted" it won't get out of balance, so scanning time
86 * remains the same (also, searching and inserting nodes in an rbtree uses
87 * the same algorithm, so we have no overhead when we flush and rebuild).
88 * 4) KSM never flushes the stable tree, which means that even if it were to
89 * take 10 attempts to find a page in the unstable tree, once it is found,
90 * it is secured in the stable tree. (When we scan a new page, we first
91 * compare it against the stable tree, and then against the unstable tree.)
Hugh Dickins8fdb3db2013-02-22 16:36:03 -080092 *
93 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
94 * stable trees and multiple unstable trees: one of each for each NUMA node.
Izik Eidus31dbd012009-09-21 17:02:03 -070095 */
96
97/**
98 * struct mm_slot - ksm information per mm that is being scanned
99 * @link: link to the mm_slots hash list
100 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
Hugh Dickins6514d512009-12-14 17:59:19 -0800101 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
Izik Eidus31dbd012009-09-21 17:02:03 -0700102 * @mm: the mm that this information is valid for
103 */
104struct mm_slot {
105 struct hlist_node link;
106 struct list_head mm_list;
Hugh Dickins6514d512009-12-14 17:59:19 -0800107 struct rmap_item *rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700108 struct mm_struct *mm;
109};
110
111/**
112 * struct ksm_scan - cursor for scanning
113 * @mm_slot: the current mm_slot we are scanning
114 * @address: the next address inside that to be scanned
Hugh Dickins6514d512009-12-14 17:59:19 -0800115 * @rmap_list: link to the next rmap to be scanned in the rmap_list
Izik Eidus31dbd012009-09-21 17:02:03 -0700116 * @seqnr: count of completed full scans (needed when removing unstable node)
117 *
118 * There is only the one ksm_scan instance of this cursor structure.
119 */
120struct ksm_scan {
121 struct mm_slot *mm_slot;
122 unsigned long address;
Hugh Dickins6514d512009-12-14 17:59:19 -0800123 struct rmap_item **rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700124 unsigned long seqnr;
125};
126
127/**
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800128 * struct stable_node - node of the stable rbtree
129 * @node: rb node of this ksm page in the stable tree
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800130 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
131 * @list: linked into migrate_nodes, pending placement in the proper node tree
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800132 * @hlist: hlist head of rmap_items using this ksm page
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800133 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
134 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800135 */
136struct stable_node {
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800137 union {
138 struct rb_node node; /* when node of stable tree */
139 struct { /* when listed for migration */
140 struct list_head *head;
141 struct list_head list;
142 };
143 };
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800144 struct hlist_head hlist;
Hugh Dickins62b61f62009-12-14 17:59:33 -0800145 unsigned long kpfn;
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800146#ifdef CONFIG_NUMA
147 int nid;
148#endif
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800149};
150
151/**
Izik Eidus31dbd012009-09-21 17:02:03 -0700152 * struct rmap_item - reverse mapping item for virtual addresses
Hugh Dickins6514d512009-12-14 17:59:19 -0800153 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800154 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
Hugh Dickinsbc566202013-02-22 16:36:06 -0800155 * @nid: NUMA node id of unstable tree in which linked (may not match page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700156 * @mm: the memory structure this rmap_item is pointing into
157 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
158 * @oldchecksum: previous checksum of the page at that virtual address
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800159 * @node: rb node of this rmap_item in the unstable tree
160 * @head: pointer to stable_node heading this list in the stable tree
161 * @hlist: link into hlist of rmap_items hanging off that stable_node
Izik Eidus31dbd012009-09-21 17:02:03 -0700162 */
163struct rmap_item {
Hugh Dickins6514d512009-12-14 17:59:19 -0800164 struct rmap_item *rmap_list;
Hugh Dickinsbc566202013-02-22 16:36:06 -0800165 union {
166 struct anon_vma *anon_vma; /* when stable */
167#ifdef CONFIG_NUMA
168 int nid; /* when node of unstable tree */
169#endif
170 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700171 struct mm_struct *mm;
172 unsigned long address; /* + low bits used for flags below */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800173 unsigned int oldchecksum; /* when unstable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700174 union {
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800175 struct rb_node node; /* when node of unstable tree */
176 struct { /* when listed from stable tree */
177 struct stable_node *head;
178 struct hlist_node hlist;
179 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700180 };
181};
182
183#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800184#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
185#define STABLE_FLAG 0x200 /* is listed from the stable tree */
Izik Eidus31dbd012009-09-21 17:02:03 -0700186
187/* The stable and unstable tree heads */
Hugh Dickinsef53d162013-02-22 16:36:12 -0800188static struct rb_root one_stable_tree[1] = { RB_ROOT };
189static struct rb_root one_unstable_tree[1] = { RB_ROOT };
190static struct rb_root *root_stable_tree = one_stable_tree;
191static struct rb_root *root_unstable_tree = one_unstable_tree;
Izik Eidus31dbd012009-09-21 17:02:03 -0700192
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800193/* Recently migrated nodes of stable tree, pending proper placement */
194static LIST_HEAD(migrate_nodes);
195
Sasha Levin4ca3a692013-02-22 16:32:28 -0800196#define MM_SLOTS_HASH_BITS 10
197static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
Izik Eidus31dbd012009-09-21 17:02:03 -0700198
199static struct mm_slot ksm_mm_head = {
200 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
201};
202static struct ksm_scan ksm_scan = {
203 .mm_slot = &ksm_mm_head,
204};
205
206static struct kmem_cache *rmap_item_cache;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800207static struct kmem_cache *stable_node_cache;
Izik Eidus31dbd012009-09-21 17:02:03 -0700208static struct kmem_cache *mm_slot_cache;
209
210/* The number of nodes in the stable tree */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700211static unsigned long ksm_pages_shared;
Izik Eidus31dbd012009-09-21 17:02:03 -0700212
Hugh Dickinse178dfd2009-09-21 17:02:10 -0700213/* The number of page slots additionally sharing those nodes */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700214static unsigned long ksm_pages_sharing;
Izik Eidus31dbd012009-09-21 17:02:03 -0700215
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700216/* The number of nodes in the unstable tree */
217static unsigned long ksm_pages_unshared;
218
219/* The number of rmap_items in use: to calculate pages_volatile */
220static unsigned long ksm_rmap_items;
221
Izik Eidus31dbd012009-09-21 17:02:03 -0700222/* Number of pages ksmd should scan in one batch */
Izik Eidus2c6854f2009-09-23 15:56:04 -0700223static unsigned int ksm_thread_pages_to_scan = 100;
Izik Eidus31dbd012009-09-21 17:02:03 -0700224
225/* Milliseconds ksmd should sleep between batches */
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700226static unsigned int ksm_thread_sleep_millisecs = 20;
Izik Eidus31dbd012009-09-21 17:02:03 -0700227
Claudio Imbrendae86c59b2017-02-24 14:55:39 -0800228/* Checksum of an empty (zeroed) page */
229static unsigned int zero_checksum __read_mostly;
230
231/* Whether to merge empty (zeroed) pages with actual zero pages */
232static bool ksm_use_zero_pages __read_mostly;
233
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800234#ifdef CONFIG_NUMA
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800235/* Zeroed when merging across nodes is not allowed */
236static unsigned int ksm_merge_across_nodes = 1;
Hugh Dickinsef53d162013-02-22 16:36:12 -0800237static int ksm_nr_node_ids = 1;
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800238#else
239#define ksm_merge_across_nodes 1U
Hugh Dickinsef53d162013-02-22 16:36:12 -0800240#define ksm_nr_node_ids 1
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800241#endif
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800242
Izik Eidus31dbd012009-09-21 17:02:03 -0700243#define KSM_RUN_STOP 0
244#define KSM_RUN_MERGE 1
245#define KSM_RUN_UNMERGE 2
Hugh Dickinsef4d43a2013-02-22 16:35:16 -0800246#define KSM_RUN_OFFLINE 4
247static unsigned long ksm_run = KSM_RUN_STOP;
248static void wait_while_offlining(void);
Izik Eidus31dbd012009-09-21 17:02:03 -0700249
250static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
251static DEFINE_MUTEX(ksm_thread_mutex);
252static DEFINE_SPINLOCK(ksm_mmlist_lock);
253
254#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
255 sizeof(struct __struct), __alignof__(struct __struct),\
256 (__flags), NULL)
257
258static int __init ksm_slab_init(void)
259{
260 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
261 if (!rmap_item_cache)
262 goto out;
263
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800264 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
265 if (!stable_node_cache)
266 goto out_free1;
267
Izik Eidus31dbd012009-09-21 17:02:03 -0700268 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
269 if (!mm_slot_cache)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800270 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -0700271
272 return 0;
273
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800274out_free2:
275 kmem_cache_destroy(stable_node_cache);
276out_free1:
Izik Eidus31dbd012009-09-21 17:02:03 -0700277 kmem_cache_destroy(rmap_item_cache);
278out:
279 return -ENOMEM;
280}
281
282static void __init ksm_slab_free(void)
283{
284 kmem_cache_destroy(mm_slot_cache);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800285 kmem_cache_destroy(stable_node_cache);
Izik Eidus31dbd012009-09-21 17:02:03 -0700286 kmem_cache_destroy(rmap_item_cache);
287 mm_slot_cache = NULL;
288}
289
290static inline struct rmap_item *alloc_rmap_item(void)
291{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700292 struct rmap_item *rmap_item;
293
zhong jiang5b398e42016-09-28 15:22:30 -0700294 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
295 __GFP_NORETRY | __GFP_NOWARN);
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700296 if (rmap_item)
297 ksm_rmap_items++;
298 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -0700299}
300
301static inline void free_rmap_item(struct rmap_item *rmap_item)
302{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700303 ksm_rmap_items--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700304 rmap_item->mm = NULL; /* debug safety */
305 kmem_cache_free(rmap_item_cache, rmap_item);
306}
307
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800308static inline struct stable_node *alloc_stable_node(void)
309{
zhong jiang62130552016-10-07 17:01:19 -0700310 /*
311 * The allocation can take too long with GFP_KERNEL when memory is under
312 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
313 * grants access to memory reserves, helping to avoid this problem.
314 */
315 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800316}
317
318static inline void free_stable_node(struct stable_node *stable_node)
319{
320 kmem_cache_free(stable_node_cache, stable_node);
321}
322
Izik Eidus31dbd012009-09-21 17:02:03 -0700323static inline struct mm_slot *alloc_mm_slot(void)
324{
325 if (!mm_slot_cache) /* initialization failed */
326 return NULL;
327 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
328}
329
330static inline void free_mm_slot(struct mm_slot *mm_slot)
331{
332 kmem_cache_free(mm_slot_cache, mm_slot);
333}
334
Izik Eidus31dbd012009-09-21 17:02:03 -0700335static struct mm_slot *get_mm_slot(struct mm_struct *mm)
336{
Sasha Levin4ca3a692013-02-22 16:32:28 -0800337 struct mm_slot *slot;
Izik Eidus31dbd012009-09-21 17:02:03 -0700338
Sasha Levinb67bfe02013-02-27 17:06:00 -0800339 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
Sasha Levin4ca3a692013-02-22 16:32:28 -0800340 if (slot->mm == mm)
341 return slot;
342
Izik Eidus31dbd012009-09-21 17:02:03 -0700343 return NULL;
344}
345
346static void insert_to_mm_slots_hash(struct mm_struct *mm,
347 struct mm_slot *mm_slot)
348{
Izik Eidus31dbd012009-09-21 17:02:03 -0700349 mm_slot->mm = mm;
Sasha Levin4ca3a692013-02-22 16:32:28 -0800350 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
Izik Eidus31dbd012009-09-21 17:02:03 -0700351}
352
Izik Eidus31dbd012009-09-21 17:02:03 -0700353/*
Hugh Dickinsa913e182009-09-21 17:02:26 -0700354 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
355 * page tables after it has passed through ksm_exit() - which, if necessary,
356 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
357 * a special flag: they can just back out as soon as mm_users goes to zero.
358 * ksm_test_exit() is used throughout to make this test for exit: in some
359 * places for correctness, in some places just to avoid unnecessary work.
360 */
361static inline bool ksm_test_exit(struct mm_struct *mm)
362{
363 return atomic_read(&mm->mm_users) == 0;
364}
365
366/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700367 * We use break_ksm to break COW on a ksm page: it's a stripped down
368 *
Dave Hansend4edcf02016-02-12 13:01:56 -0800369 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
Izik Eidus31dbd012009-09-21 17:02:03 -0700370 * put_page(page);
371 *
372 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
373 * in case the application has unmapped and remapped mm,addr meanwhile.
374 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
375 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
Dave Hansen1b2ee122016-02-12 13:02:21 -0800376 *
377 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
378 * of the process that owns 'vma'. We also do not want to enforce
379 * protection keys here anyway.
Izik Eidus31dbd012009-09-21 17:02:03 -0700380 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700381static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
Izik Eidus31dbd012009-09-21 17:02:03 -0700382{
383 struct page *page;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700384 int ret = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700385
386 do {
387 cond_resched();
Dave Hansen1b2ee122016-02-12 13:02:21 -0800388 page = follow_page(vma, addr,
389 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400390 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700391 break;
392 if (PageKsm(page))
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700393 ret = handle_mm_fault(vma, addr,
394 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
Izik Eidus31dbd012009-09-21 17:02:03 -0700395 else
396 ret = VM_FAULT_WRITE;
397 put_page(page);
Linus Torvalds33692f22015-01-29 10:51:32 -0800398 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
Hugh Dickinsd952b792009-09-21 17:02:16 -0700399 /*
400 * We must loop because handle_mm_fault() may back out if there's
401 * any difficulty e.g. if pte accessed bit gets updated concurrently.
402 *
403 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
404 * COW has been broken, even if the vma does not permit VM_WRITE;
405 * but note that a concurrent fault might break PageKsm for us.
406 *
407 * VM_FAULT_SIGBUS could occur if we race with truncation of the
408 * backing file, which also invalidates anonymous pages: that's
409 * okay, that truncation will have unmapped the PageKsm for us.
410 *
411 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
412 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
413 * current task has TIF_MEMDIE set, and will be OOM killed on return
414 * to user; and ksmd, having no mm, would never be chosen for that.
415 *
416 * But if the mm is in a limited mem_cgroup, then the fault may fail
417 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
418 * even ksmd can fail in this way - though it's usually breaking ksm
419 * just to undo a merge it made a moment before, so unlikely to oom.
420 *
421 * That's a pity: we might therefore have more kernel pages allocated
422 * than we're counting as nodes in the stable tree; but ksm_do_scan
423 * will retry to break_cow on each pass, so should recover the page
424 * in due course. The important thing is to not let VM_MERGEABLE
425 * be cleared while any such pages might remain in the area.
426 */
427 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700428}
429
Bob Liuef694222012-03-21 16:34:11 -0700430static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
431 unsigned long addr)
432{
433 struct vm_area_struct *vma;
434 if (ksm_test_exit(mm))
435 return NULL;
436 vma = find_vma(mm, addr);
437 if (!vma || vma->vm_start > addr)
438 return NULL;
439 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
440 return NULL;
441 return vma;
442}
443
Hugh Dickins8dd35572009-12-14 17:59:18 -0800444static void break_cow(struct rmap_item *rmap_item)
Izik Eidus31dbd012009-09-21 17:02:03 -0700445{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800446 struct mm_struct *mm = rmap_item->mm;
447 unsigned long addr = rmap_item->address;
Izik Eidus31dbd012009-09-21 17:02:03 -0700448 struct vm_area_struct *vma;
449
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800450 /*
451 * It is not an accident that whenever we want to break COW
452 * to undo, we also need to drop a reference to the anon_vma.
453 */
Peter Zijlstra9e601092011-03-22 16:32:46 -0700454 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800455
Hugh Dickins81464e302009-09-21 17:02:15 -0700456 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700457 vma = find_mergeable_vma(mm, addr);
458 if (vma)
459 break_ksm(vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700460 up_read(&mm->mmap_sem);
461}
462
463static struct page *get_mergeable_page(struct rmap_item *rmap_item)
464{
465 struct mm_struct *mm = rmap_item->mm;
466 unsigned long addr = rmap_item->address;
467 struct vm_area_struct *vma;
468 struct page *page;
469
470 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700471 vma = find_mergeable_vma(mm, addr);
472 if (!vma)
Izik Eidus31dbd012009-09-21 17:02:03 -0700473 goto out;
474
475 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400476 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700477 goto out;
Kirill A. Shutemovf765f542016-01-15 16:53:03 -0800478 if (PageAnon(page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700479 flush_anon_page(vma, page, addr);
480 flush_dcache_page(page);
481 } else {
482 put_page(page);
Andrea Arcangelic8f95ed2015-11-05 18:49:19 -0800483out:
484 page = NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700485 }
486 up_read(&mm->mmap_sem);
487 return page;
488}
489
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800490/*
491 * This helper is used for getting right index into array of tree roots.
492 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
493 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
494 * every node has its own stable and unstable tree.
495 */
496static inline int get_kpfn_nid(unsigned long kpfn)
497{
Hugh Dickinsd8fc16a2013-03-08 12:43:34 -0800498 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800499}
500
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800501static void remove_node_from_stable_tree(struct stable_node *stable_node)
502{
503 struct rmap_item *rmap_item;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800504
Sasha Levinb67bfe02013-02-27 17:06:00 -0800505 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800506 if (rmap_item->hlist.next)
507 ksm_pages_sharing--;
508 else
509 ksm_pages_shared--;
Peter Zijlstra9e601092011-03-22 16:32:46 -0700510 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800511 rmap_item->address &= PAGE_MASK;
512 cond_resched();
513 }
514
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800515 if (stable_node->head == &migrate_nodes)
516 list_del(&stable_node->list);
517 else
518 rb_erase(&stable_node->node,
Hugh Dickinsef53d162013-02-22 16:36:12 -0800519 root_stable_tree + NUMA(stable_node->nid));
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800520 free_stable_node(stable_node);
521}
522
523/*
524 * get_ksm_page: checks if the page indicated by the stable node
525 * is still its ksm page, despite having held no reference to it.
526 * In which case we can trust the content of the page, and it
527 * returns the gotten page; but if the page has now been zapped,
528 * remove the stale node from the stable tree and return NULL.
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800529 * But beware, the stable node's page might be being migrated.
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800530 *
531 * You would expect the stable_node to hold a reference to the ksm page.
532 * But if it increments the page's count, swapping out has to wait for
533 * ksmd to come around again before it can free the page, which may take
534 * seconds or even minutes: much too unresponsive. So instead we use a
535 * "keyhole reference": access to the ksm page from the stable node peeps
536 * out through its keyhole to see if that page still holds the right key,
537 * pointing back to this stable node. This relies on freeing a PageAnon
538 * page to reset its page->mapping to NULL, and relies on no other use of
539 * a page to put something that might look like our key in page->mapping.
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800540 * is on its way to being freed; but it is an anomaly to bear in mind.
541 */
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800542static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800543{
544 struct page *page;
545 void *expected_mapping;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800546 unsigned long kpfn;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800547
Minchan Kimbda807d2016-07-26 15:23:05 -0700548 expected_mapping = (void *)((unsigned long)stable_node |
549 PAGE_MAPPING_KSM);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800550again:
Jason Low4db0c3c2015-04-15 16:14:08 -0700551 kpfn = READ_ONCE(stable_node->kpfn);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800552 page = pfn_to_page(kpfn);
553
554 /*
555 * page is computed from kpfn, so on most architectures reading
556 * page->mapping is naturally ordered after reading node->kpfn,
557 * but on Alpha we need to be more careful.
558 */
559 smp_read_barrier_depends();
Jason Low4db0c3c2015-04-15 16:14:08 -0700560 if (READ_ONCE(page->mapping) != expected_mapping)
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800561 goto stale;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800562
563 /*
564 * We cannot do anything with the page while its refcount is 0.
565 * Usually 0 means free, or tail of a higher-order page: in which
566 * case this node is no longer referenced, and should be freed;
567 * however, it might mean that the page is under page_freeze_refs().
568 * The __remove_mapping() case is easy, again the node is now stale;
569 * but if page is swapcache in migrate_page_move_mapping(), it might
570 * still be our page, in which case it's essential to keep the node.
571 */
572 while (!get_page_unless_zero(page)) {
573 /*
574 * Another check for page->mapping != expected_mapping would
575 * work here too. We have chosen the !PageSwapCache test to
576 * optimize the common case, when the page is or is about to
577 * be freed: PageSwapCache is cleared (under spin_lock_irq)
578 * in the freeze_refs section of __remove_mapping(); but Anon
579 * page->mapping reset to NULL later, in free_pages_prepare().
580 */
581 if (!PageSwapCache(page))
582 goto stale;
583 cpu_relax();
584 }
585
Jason Low4db0c3c2015-04-15 16:14:08 -0700586 if (READ_ONCE(page->mapping) != expected_mapping) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800587 put_page(page);
588 goto stale;
589 }
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800590
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800591 if (lock_it) {
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800592 lock_page(page);
Jason Low4db0c3c2015-04-15 16:14:08 -0700593 if (READ_ONCE(page->mapping) != expected_mapping) {
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800594 unlock_page(page);
595 put_page(page);
596 goto stale;
597 }
598 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800599 return page;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800600
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800601stale:
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800602 /*
603 * We come here from above when page->mapping or !PageSwapCache
604 * suggests that the node is stale; but it might be under migration.
605 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
606 * before checking whether node->kpfn has been changed.
607 */
608 smp_rmb();
Jason Low4db0c3c2015-04-15 16:14:08 -0700609 if (READ_ONCE(stable_node->kpfn) != kpfn)
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800610 goto again;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800611 remove_node_from_stable_tree(stable_node);
612 return NULL;
613}
614
Izik Eidus31dbd012009-09-21 17:02:03 -0700615/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700616 * Removing rmap_item from stable or unstable tree.
617 * This function will clean the information from the stable/unstable tree.
618 */
619static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
620{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800621 if (rmap_item->address & STABLE_FLAG) {
622 struct stable_node *stable_node;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800623 struct page *page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700624
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800625 stable_node = rmap_item->head;
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800626 page = get_ksm_page(stable_node, true);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800627 if (!page)
628 goto out;
629
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800630 hlist_del(&rmap_item->hlist);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800631 unlock_page(page);
632 put_page(page);
Hugh Dickins08beca42009-12-14 17:59:21 -0800633
Andrea Arcangeli98666f8a2015-11-05 18:49:13 -0800634 if (!hlist_empty(&stable_node->hlist))
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800635 ksm_pages_sharing--;
636 else
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800637 ksm_pages_shared--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700638
Peter Zijlstra9e601092011-03-22 16:32:46 -0700639 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins93d17712009-12-14 17:59:16 -0800640 rmap_item->address &= PAGE_MASK;
Izik Eidus31dbd012009-09-21 17:02:03 -0700641
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800642 } else if (rmap_item->address & UNSTABLE_FLAG) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700643 unsigned char age;
644 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -0700645 * Usually ksmd can and must skip the rb_erase, because
Izik Eidus31dbd012009-09-21 17:02:03 -0700646 * root_unstable_tree was already reset to RB_ROOT.
Hugh Dickins9ba69292009-09-21 17:02:20 -0700647 * But be careful when an mm is exiting: do the rb_erase
648 * if this rmap_item was inserted by this scan, rather
649 * than left over from before.
Izik Eidus31dbd012009-09-21 17:02:03 -0700650 */
651 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
Hugh Dickinscd551f92009-09-21 17:02:17 -0700652 BUG_ON(age > 1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700653 if (!age)
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800654 rb_erase(&rmap_item->node,
Hugh Dickinsef53d162013-02-22 16:36:12 -0800655 root_unstable_tree + NUMA(rmap_item->nid));
Hugh Dickins93d17712009-12-14 17:59:16 -0800656 ksm_pages_unshared--;
657 rmap_item->address &= PAGE_MASK;
658 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800659out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700660 cond_resched(); /* we're called from many long loops */
661}
662
Izik Eidus31dbd012009-09-21 17:02:03 -0700663static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -0800664 struct rmap_item **rmap_list)
Izik Eidus31dbd012009-09-21 17:02:03 -0700665{
Hugh Dickins6514d512009-12-14 17:59:19 -0800666 while (*rmap_list) {
667 struct rmap_item *rmap_item = *rmap_list;
668 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700669 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700670 free_rmap_item(rmap_item);
671 }
672}
673
674/*
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800675 * Though it's very tempting to unmerge rmap_items from stable tree rather
Izik Eidus31dbd012009-09-21 17:02:03 -0700676 * than check every pte of a given vma, the locking doesn't quite work for
677 * that - an rmap_item is assigned to the stable tree after inserting ksm
678 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
679 * rmap_items from parent to child at fork time (so as not to waste time
680 * if exit comes before the next scan reaches it).
Hugh Dickins81464e302009-09-21 17:02:15 -0700681 *
682 * Similarly, although we'd like to remove rmap_items (so updating counts
683 * and freeing memory) when unmerging an area, it's easier to leave that
684 * to the next pass of ksmd - consider, for example, how ksmd might be
685 * in cmp_and_merge_page on one of the rmap_items we would be removing.
Izik Eidus31dbd012009-09-21 17:02:03 -0700686 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700687static int unmerge_ksm_pages(struct vm_area_struct *vma,
688 unsigned long start, unsigned long end)
Izik Eidus31dbd012009-09-21 17:02:03 -0700689{
690 unsigned long addr;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700691 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700692
Hugh Dickinsd952b792009-09-21 17:02:16 -0700693 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700694 if (ksm_test_exit(vma->vm_mm))
695 break;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700696 if (signal_pending(current))
697 err = -ERESTARTSYS;
698 else
699 err = break_ksm(vma, addr);
700 }
701 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700702}
703
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700704#ifdef CONFIG_SYSFS
705/*
706 * Only called through the sysfs control interface:
707 */
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800708static int remove_stable_node(struct stable_node *stable_node)
709{
710 struct page *page;
711 int err;
712
713 page = get_ksm_page(stable_node, true);
714 if (!page) {
715 /*
716 * get_ksm_page did remove_node_from_stable_tree itself.
717 */
718 return 0;
719 }
720
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800721 if (WARN_ON_ONCE(page_mapped(page))) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800722 /*
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800723 * This should not happen: but if it does, just refuse to let
724 * merge_across_nodes be switched - there is no need to panic.
725 */
726 err = -EBUSY;
727 } else {
728 /*
729 * The stable node did not yet appear stale to get_ksm_page(),
730 * since that allows for an unmapped ksm page to be recognized
731 * right up until it is freed; but the node is safe to remove.
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800732 * This page might be in a pagevec waiting to be freed,
733 * or it might be PageSwapCache (perhaps under writeback),
734 * or it might have been removed from swapcache a moment ago.
735 */
736 set_page_stable_node(page, NULL);
737 remove_node_from_stable_tree(stable_node);
738 err = 0;
739 }
740
741 unlock_page(page);
742 put_page(page);
743 return err;
744}
745
746static int remove_all_stable_nodes(void)
747{
Geliang Tang03640412016-01-14 15:20:54 -0800748 struct stable_node *stable_node, *next;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800749 int nid;
750 int err = 0;
751
Hugh Dickinsef53d162013-02-22 16:36:12 -0800752 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800753 while (root_stable_tree[nid].rb_node) {
754 stable_node = rb_entry(root_stable_tree[nid].rb_node,
755 struct stable_node, node);
756 if (remove_stable_node(stable_node)) {
757 err = -EBUSY;
758 break; /* proceed to next nid */
759 }
760 cond_resched();
761 }
762 }
Geliang Tang03640412016-01-14 15:20:54 -0800763 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800764 if (remove_stable_node(stable_node))
765 err = -EBUSY;
766 cond_resched();
767 }
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800768 return err;
769}
770
Hugh Dickinsd952b792009-09-21 17:02:16 -0700771static int unmerge_and_remove_all_rmap_items(void)
Izik Eidus31dbd012009-09-21 17:02:03 -0700772{
773 struct mm_slot *mm_slot;
774 struct mm_struct *mm;
775 struct vm_area_struct *vma;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700776 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700777
Hugh Dickinsd952b792009-09-21 17:02:16 -0700778 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700779 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700780 struct mm_slot, mm_list);
781 spin_unlock(&ksm_mmlist_lock);
782
Hugh Dickins9ba69292009-09-21 17:02:20 -0700783 for (mm_slot = ksm_scan.mm_slot;
784 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700785 mm = mm_slot->mm;
786 down_read(&mm->mmap_sem);
787 for (vma = mm->mmap; vma; vma = vma->vm_next) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700788 if (ksm_test_exit(mm))
789 break;
Izik Eidus31dbd012009-09-21 17:02:03 -0700790 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
791 continue;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700792 err = unmerge_ksm_pages(vma,
793 vma->vm_start, vma->vm_end);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700794 if (err)
795 goto error;
Izik Eidus31dbd012009-09-21 17:02:03 -0700796 }
Hugh Dickins9ba69292009-09-21 17:02:20 -0700797
Hugh Dickins6514d512009-12-14 17:59:19 -0800798 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
Zhou Chengming7496fea2016-05-12 15:42:21 -0700799 up_read(&mm->mmap_sem);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700800
801 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700802 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700803 struct mm_slot, mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700804 if (ksm_test_exit(mm)) {
Sasha Levin4ca3a692013-02-22 16:32:28 -0800805 hash_del(&mm_slot->link);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700806 list_del(&mm_slot->mm_list);
807 spin_unlock(&ksm_mmlist_lock);
808
809 free_mm_slot(mm_slot);
810 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700811 mmdrop(mm);
Zhou Chengming7496fea2016-05-12 15:42:21 -0700812 } else
Hugh Dickins9ba69292009-09-21 17:02:20 -0700813 spin_unlock(&ksm_mmlist_lock);
Izik Eidus31dbd012009-09-21 17:02:03 -0700814 }
815
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800816 /* Clean up stable nodes, but don't worry if some are still busy */
817 remove_all_stable_nodes();
Hugh Dickinsd952b792009-09-21 17:02:16 -0700818 ksm_scan.seqnr = 0;
Hugh Dickins9ba69292009-09-21 17:02:20 -0700819 return 0;
820
821error:
822 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700823 spin_lock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700824 ksm_scan.mm_slot = &ksm_mm_head;
Izik Eidus31dbd012009-09-21 17:02:03 -0700825 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700826 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700827}
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700828#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -0700829
Izik Eidus31dbd012009-09-21 17:02:03 -0700830static u32 calc_checksum(struct page *page)
831{
832 u32 checksum;
Cong Wang9b04c5f2011-11-25 23:14:39 +0800833 void *addr = kmap_atomic(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700834 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
Cong Wang9b04c5f2011-11-25 23:14:39 +0800835 kunmap_atomic(addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700836 return checksum;
837}
838
839static int memcmp_pages(struct page *page1, struct page *page2)
840{
841 char *addr1, *addr2;
842 int ret;
843
Cong Wang9b04c5f2011-11-25 23:14:39 +0800844 addr1 = kmap_atomic(page1);
845 addr2 = kmap_atomic(page2);
Izik Eidus31dbd012009-09-21 17:02:03 -0700846 ret = memcmp(addr1, addr2, PAGE_SIZE);
Cong Wang9b04c5f2011-11-25 23:14:39 +0800847 kunmap_atomic(addr2);
848 kunmap_atomic(addr1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700849 return ret;
850}
851
852static inline int pages_identical(struct page *page1, struct page *page2)
853{
854 return !memcmp_pages(page1, page2);
855}
856
857static int write_protect_page(struct vm_area_struct *vma, struct page *page,
858 pte_t *orig_pte)
859{
860 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -0800861 struct page_vma_mapped_walk pvmw = {
862 .page = page,
863 .vma = vma,
864 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700865 int swapped;
866 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700867 unsigned long mmun_start; /* For mmu_notifiers */
868 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -0700869
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -0800870 pvmw.address = page_address_in_vma(page, vma);
871 if (pvmw.address == -EFAULT)
Izik Eidus31dbd012009-09-21 17:02:03 -0700872 goto out;
873
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800874 BUG_ON(PageTransCompound(page));
Haggai Eran6bdb9132012-10-08 16:33:35 -0700875
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -0800876 mmun_start = pvmw.address;
877 mmun_end = pvmw.address + PAGE_SIZE;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700878 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
879
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -0800880 if (!page_vma_mapped_walk(&pvmw))
Haggai Eran6bdb9132012-10-08 16:33:35 -0700881 goto out_mn;
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -0800882 if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?"))
883 goto out_unlock;
Izik Eidus31dbd012009-09-21 17:02:03 -0700884
Aneesh Kumar K.V595cd8f2017-02-24 14:59:19 -0800885 if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) ||
886 (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte))) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700887 pte_t entry;
888
889 swapped = PageSwapCache(page);
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -0800890 flush_cache_page(vma, pvmw.address, page_to_pfn(page));
Izik Eidus31dbd012009-09-21 17:02:03 -0700891 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300892 * Ok this is tricky, when get_user_pages_fast() run it doesn't
Izik Eidus31dbd012009-09-21 17:02:03 -0700893 * take any lock, therefore the check that we are going to make
894 * with the pagecount against the mapcount is racey and
895 * O_DIRECT can happen right after the check.
896 * So we clear the pte and flush the tlb before the check
897 * this assure us that no O_DIRECT can happen after the check
898 * or in the middle of the check.
899 */
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -0800900 entry = ptep_clear_flush_notify(vma, pvmw.address, pvmw.pte);
Izik Eidus31dbd012009-09-21 17:02:03 -0700901 /*
902 * Check that no O_DIRECT or similar I/O is in progress on the
903 * page
904 */
Hugh Dickins31e855e2009-12-14 17:59:17 -0800905 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -0800906 set_pte_at(mm, pvmw.address, pvmw.pte, entry);
Izik Eidus31dbd012009-09-21 17:02:03 -0700907 goto out_unlock;
908 }
Hugh Dickins4e316352010-10-02 17:49:08 -0700909 if (pte_dirty(entry))
910 set_page_dirty(page);
Aneesh Kumar K.V595cd8f2017-02-24 14:59:19 -0800911
912 if (pte_protnone(entry))
913 entry = pte_mkclean(pte_clear_savedwrite(entry));
914 else
915 entry = pte_mkclean(pte_wrprotect(entry));
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -0800916 set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry);
Izik Eidus31dbd012009-09-21 17:02:03 -0700917 }
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -0800918 *orig_pte = *pvmw.pte;
Izik Eidus31dbd012009-09-21 17:02:03 -0700919 err = 0;
920
921out_unlock:
Kirill A. Shutemov36eaff32017-02-24 14:58:04 -0800922 page_vma_mapped_walk_done(&pvmw);
Haggai Eran6bdb9132012-10-08 16:33:35 -0700923out_mn:
924 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -0700925out:
926 return err;
927}
928
929/**
930 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -0800931 * @vma: vma that holds the pte pointing to page
932 * @page: the page we are replacing by kpage
933 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -0700934 * @orig_pte: the original value of the pte
935 *
936 * Returns 0 on success, -EFAULT on failure.
937 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800938static int replace_page(struct vm_area_struct *vma, struct page *page,
939 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -0700940{
941 struct mm_struct *mm = vma->vm_mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700942 pmd_t *pmd;
943 pte_t *ptep;
Claudio Imbrendae86c59b2017-02-24 14:55:39 -0800944 pte_t newpte;
Izik Eidus31dbd012009-09-21 17:02:03 -0700945 spinlock_t *ptl;
946 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -0700947 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700948 unsigned long mmun_start; /* For mmu_notifiers */
949 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -0700950
Hugh Dickins8dd35572009-12-14 17:59:18 -0800951 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -0700952 if (addr == -EFAULT)
953 goto out;
954
Bob Liu62190492012-12-11 16:00:37 -0800955 pmd = mm_find_pmd(mm, addr);
956 if (!pmd)
Izik Eidus31dbd012009-09-21 17:02:03 -0700957 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700958
Haggai Eran6bdb9132012-10-08 16:33:35 -0700959 mmun_start = addr;
960 mmun_end = addr + PAGE_SIZE;
961 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
962
Izik Eidus31dbd012009-09-21 17:02:03 -0700963 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
964 if (!pte_same(*ptep, orig_pte)) {
965 pte_unmap_unlock(ptep, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -0700966 goto out_mn;
Izik Eidus31dbd012009-09-21 17:02:03 -0700967 }
968
Claudio Imbrendae86c59b2017-02-24 14:55:39 -0800969 /*
970 * No need to check ksm_use_zero_pages here: we can only have a
971 * zero_page here if ksm_use_zero_pages was enabled alreaady.
972 */
973 if (!is_zero_pfn(page_to_pfn(kpage))) {
974 get_page(kpage);
975 page_add_anon_rmap(kpage, vma, addr, false);
976 newpte = mk_pte(kpage, vma->vm_page_prot);
977 } else {
978 newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage),
979 vma->vm_page_prot));
980 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700981
982 flush_cache_page(vma, addr, pte_pfn(*ptep));
Joerg Roedel34ee6452014-11-13 13:46:09 +1100983 ptep_clear_flush_notify(vma, addr, ptep);
Claudio Imbrendae86c59b2017-02-24 14:55:39 -0800984 set_pte_at_notify(mm, addr, ptep, newpte);
Izik Eidus31dbd012009-09-21 17:02:03 -0700985
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800986 page_remove_rmap(page, false);
Hugh Dickinsae52a2a2011-01-13 15:46:28 -0800987 if (!page_mapped(page))
988 try_to_free_swap(page);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800989 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700990
991 pte_unmap_unlock(ptep, ptl);
992 err = 0;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700993out_mn:
994 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -0700995out:
996 return err;
997}
998
999/*
1000 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -08001001 * @vma: the vma that holds the pte pointing to page
1002 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins80e148222009-12-14 17:59:29 -08001003 * @kpage: the PageKsm page that we want to map instead of page,
1004 * or NULL the first time when we want to use page as kpage.
Izik Eidus31dbd012009-09-21 17:02:03 -07001005 *
1006 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1007 */
1008static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -08001009 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001010{
1011 pte_t orig_pte = __pte(0);
1012 int err = -EFAULT;
1013
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001014 if (page == kpage) /* ksm page forked */
1015 return 0;
1016
Hugh Dickins8dd35572009-12-14 17:59:18 -08001017 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -07001018 goto out;
1019
Izik Eidus31dbd012009-09-21 17:02:03 -07001020 /*
1021 * We need the page lock to read a stable PageSwapCache in
1022 * write_protect_page(). We use trylock_page() instead of
1023 * lock_page() because we don't want to wait here - we
1024 * prefer to continue scanning and merging different pages,
1025 * then come back to this page when it is unlocked.
1026 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001027 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -08001028 goto out;
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001029
1030 if (PageTransCompound(page)) {
1031 err = split_huge_page(page);
1032 if (err)
1033 goto out_unlock;
1034 }
1035
Izik Eidus31dbd012009-09-21 17:02:03 -07001036 /*
1037 * If this anonymous page is mapped only here, its pte may need
1038 * to be write-protected. If it's mapped elsewhere, all of its
1039 * ptes are necessarily already write-protected. But in either
1040 * case, we need to lock and check page_count is not raised.
1041 */
Hugh Dickins80e148222009-12-14 17:59:29 -08001042 if (write_protect_page(vma, page, &orig_pte) == 0) {
1043 if (!kpage) {
1044 /*
1045 * While we hold page lock, upgrade page from
1046 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1047 * stable_tree_insert() will update stable_node.
1048 */
1049 set_page_stable_node(page, NULL);
1050 mark_page_accessed(page);
Minchan Kim337ed7e2016-01-15 16:55:15 -08001051 /*
1052 * Page reclaim just frees a clean page with no dirty
1053 * ptes: make sure that the ksm page would be swapped.
1054 */
1055 if (!PageDirty(page))
1056 SetPageDirty(page);
Hugh Dickins80e148222009-12-14 17:59:29 -08001057 err = 0;
1058 } else if (pages_identical(page, kpage))
1059 err = replace_page(vma, page, kpage, orig_pte);
1060 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001061
Hugh Dickins80e148222009-12-14 17:59:29 -08001062 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
Hugh Dickins73848b42009-12-14 17:59:22 -08001063 munlock_vma_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001064 if (!PageMlocked(kpage)) {
1065 unlock_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001066 lock_page(kpage);
1067 mlock_vma_page(kpage);
1068 page = kpage; /* for final unlock */
1069 }
1070 }
Hugh Dickins73848b42009-12-14 17:59:22 -08001071
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001072out_unlock:
Hugh Dickins8dd35572009-12-14 17:59:18 -08001073 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001074out:
1075 return err;
1076}
1077
1078/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001079 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1080 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -08001081 *
1082 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -07001083 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001084static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1085 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -07001086{
Hugh Dickins8dd35572009-12-14 17:59:18 -08001087 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -07001088 struct vm_area_struct *vma;
1089 int err = -EFAULT;
1090
Hugh Dickins8dd35572009-12-14 17:59:18 -08001091 down_read(&mm->mmap_sem);
Andrea Arcangeli85c6e8d2015-11-05 18:49:16 -08001092 vma = find_mergeable_vma(mm, rmap_item->address);
1093 if (!vma)
Hugh Dickins9ba69292009-09-21 17:02:20 -07001094 goto out;
1095
Hugh Dickins8dd35572009-12-14 17:59:18 -08001096 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001097 if (err)
1098 goto out;
1099
Hugh Dickinsbc566202013-02-22 16:36:06 -08001100 /* Unstable nid is in union with stable anon_vma: remove first */
1101 remove_rmap_item_from_tree(rmap_item);
1102
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001103 /* Must get reference to anon_vma while still holding mmap_sem */
Peter Zijlstra9e601092011-03-22 16:32:46 -07001104 rmap_item->anon_vma = vma->anon_vma;
1105 get_anon_vma(vma->anon_vma);
Hugh Dickins81464e302009-09-21 17:02:15 -07001106out:
Hugh Dickins8dd35572009-12-14 17:59:18 -08001107 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -07001108 return err;
1109}
1110
1111/*
Izik Eidus31dbd012009-09-21 17:02:03 -07001112 * try_to_merge_two_pages - take two identical pages and prepare them
1113 * to be merged into one page.
1114 *
Hugh Dickins8dd35572009-12-14 17:59:18 -08001115 * This function returns the kpage if we successfully merged two identical
1116 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001117 *
Hugh Dickins80e148222009-12-14 17:59:29 -08001118 * Note that this function upgrades page to ksm page: if one of the pages
Izik Eidus31dbd012009-09-21 17:02:03 -07001119 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1120 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001121static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1122 struct page *page,
1123 struct rmap_item *tree_rmap_item,
1124 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001125{
Hugh Dickins80e148222009-12-14 17:59:29 -08001126 int err;
Izik Eidus31dbd012009-09-21 17:02:03 -07001127
Hugh Dickins80e148222009-12-14 17:59:29 -08001128 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
Izik Eidus31dbd012009-09-21 17:02:03 -07001129 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001130 err = try_to_merge_with_ksm_page(tree_rmap_item,
Hugh Dickins80e148222009-12-14 17:59:29 -08001131 tree_page, page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001132 /*
Hugh Dickins81464e302009-09-21 17:02:15 -07001133 * If that fails, we have a ksm page with only one pte
1134 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -07001135 */
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001136 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -08001137 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001138 }
Hugh Dickins80e148222009-12-14 17:59:29 -08001139 return err ? NULL : page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001140}
1141
1142/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001143 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -07001144 *
1145 * This function checks if there is a page inside the stable tree
1146 * with identical content to the page that we are scanning right now.
1147 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001148 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -07001149 * NULL otherwise.
1150 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001151static struct page *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001152{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001153 int nid;
Hugh Dickinsef53d162013-02-22 16:36:12 -08001154 struct rb_root *root;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001155 struct rb_node **new;
1156 struct rb_node *parent;
1157 struct stable_node *stable_node;
1158 struct stable_node *page_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001159
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001160 page_node = page_stable_node(page);
1161 if (page_node && page_node->head != &migrate_nodes) {
1162 /* ksm page forked */
Hugh Dickins08beca42009-12-14 17:59:21 -08001163 get_page(page);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001164 return page;
Hugh Dickins08beca42009-12-14 17:59:21 -08001165 }
1166
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001167 nid = get_kpfn_nid(page_to_pfn(page));
Hugh Dickinsef53d162013-02-22 16:36:12 -08001168 root = root_stable_tree + nid;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001169again:
Hugh Dickinsef53d162013-02-22 16:36:12 -08001170 new = &root->rb_node;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001171 parent = NULL;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001172
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001173 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001174 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001175 int ret;
1176
Hugh Dickins08beca42009-12-14 17:59:21 -08001177 cond_resched();
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001178 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins8aafa6a2013-02-22 16:35:06 -08001179 tree_page = get_ksm_page(stable_node, false);
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001180 if (!tree_page) {
1181 /*
1182 * If we walked over a stale stable_node,
1183 * get_ksm_page() will call rb_erase() and it
1184 * may rebalance the tree from under us. So
1185 * restart the search from scratch. Returning
1186 * NULL would be safe too, but we'd generate
1187 * false negative insertions just because some
1188 * stable_node was stale.
1189 */
1190 goto again;
1191 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001192
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001193 ret = memcmp_pages(page, tree_page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001194 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001195
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001196 parent = *new;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001197 if (ret < 0)
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001198 new = &parent->rb_left;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001199 else if (ret > 0)
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001200 new = &parent->rb_right;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001201 else {
1202 /*
1203 * Lock and unlock the stable_node's page (which
1204 * might already have been migrated) so that page
1205 * migration is sure to notice its raised count.
1206 * It would be more elegant to return stable_node
1207 * than kpage, but that involves more changes.
1208 */
1209 tree_page = get_ksm_page(stable_node, true);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001210 if (tree_page) {
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001211 unlock_page(tree_page);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001212 if (get_kpfn_nid(stable_node->kpfn) !=
1213 NUMA(stable_node->nid)) {
1214 put_page(tree_page);
1215 goto replace;
1216 }
1217 return tree_page;
1218 }
1219 /*
1220 * There is now a place for page_node, but the tree may
1221 * have been rebalanced, so re-evaluate parent and new.
1222 */
1223 if (page_node)
1224 goto again;
1225 return NULL;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001226 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001227 }
1228
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001229 if (!page_node)
1230 return NULL;
1231
1232 list_del(&page_node->list);
1233 DO_NUMA(page_node->nid = nid);
1234 rb_link_node(&page_node->node, parent, new);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001235 rb_insert_color(&page_node->node, root);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001236 get_page(page);
1237 return page;
1238
1239replace:
1240 if (page_node) {
1241 list_del(&page_node->list);
1242 DO_NUMA(page_node->nid = nid);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001243 rb_replace_node(&stable_node->node, &page_node->node, root);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001244 get_page(page);
1245 } else {
Hugh Dickinsef53d162013-02-22 16:36:12 -08001246 rb_erase(&stable_node->node, root);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001247 page = NULL;
1248 }
1249 stable_node->head = &migrate_nodes;
1250 list_add(&stable_node->list, stable_node->head);
1251 return page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001252}
1253
1254/*
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001255 * stable_tree_insert - insert stable tree node pointing to new ksm page
Izik Eidus31dbd012009-09-21 17:02:03 -07001256 * into the stable tree.
1257 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001258 * This function returns the stable tree node just allocated on success,
1259 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001260 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001261static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001262{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001263 int nid;
1264 unsigned long kpfn;
Hugh Dickinsef53d162013-02-22 16:36:12 -08001265 struct rb_root *root;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001266 struct rb_node **new;
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001267 struct rb_node *parent;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001268 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001269
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001270 kpfn = page_to_pfn(kpage);
1271 nid = get_kpfn_nid(kpfn);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001272 root = root_stable_tree + nid;
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001273again:
1274 parent = NULL;
Hugh Dickinsef53d162013-02-22 16:36:12 -08001275 new = &root->rb_node;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001276
Izik Eidus31dbd012009-09-21 17:02:03 -07001277 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001278 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001279 int ret;
1280
Hugh Dickins08beca42009-12-14 17:59:21 -08001281 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001282 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins8aafa6a2013-02-22 16:35:06 -08001283 tree_page = get_ksm_page(stable_node, false);
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001284 if (!tree_page) {
1285 /*
1286 * If we walked over a stale stable_node,
1287 * get_ksm_page() will call rb_erase() and it
1288 * may rebalance the tree from under us. So
1289 * restart the search from scratch. Returning
1290 * NULL would be safe too, but we'd generate
1291 * false negative insertions just because some
1292 * stable_node was stale.
1293 */
1294 goto again;
1295 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001296
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001297 ret = memcmp_pages(kpage, tree_page);
1298 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001299
1300 parent = *new;
1301 if (ret < 0)
1302 new = &parent->rb_left;
1303 else if (ret > 0)
1304 new = &parent->rb_right;
1305 else {
1306 /*
1307 * It is not a bug that stable_tree_search() didn't
1308 * find this node: because at that time our page was
1309 * not yet write-protected, so may have changed since.
1310 */
1311 return NULL;
1312 }
1313 }
1314
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001315 stable_node = alloc_stable_node();
1316 if (!stable_node)
1317 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001318
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001319 INIT_HLIST_HEAD(&stable_node->hlist);
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001320 stable_node->kpfn = kpfn;
Hugh Dickins08beca42009-12-14 17:59:21 -08001321 set_page_stable_node(kpage, stable_node);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001322 DO_NUMA(stable_node->nid = nid);
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001323 rb_link_node(&stable_node->node, parent, new);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001324 rb_insert_color(&stable_node->node, root);
Hugh Dickins08beca42009-12-14 17:59:21 -08001325
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001326 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001327}
1328
1329/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001330 * unstable_tree_search_insert - search for identical page,
1331 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001332 *
1333 * This function searches for a page in the unstable tree identical to the
1334 * page currently being scanned; and if no identical page is found in the
1335 * tree, we insert rmap_item as a new object into the unstable tree.
1336 *
1337 * This function returns pointer to rmap_item found to be identical
1338 * to the currently scanned page, NULL otherwise.
1339 *
1340 * This function does both searching and inserting, because they share
1341 * the same walking algorithm in an rbtree.
1342 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001343static
1344struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1345 struct page *page,
1346 struct page **tree_pagep)
Izik Eidus31dbd012009-09-21 17:02:03 -07001347{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001348 struct rb_node **new;
1349 struct rb_root *root;
Izik Eidus31dbd012009-09-21 17:02:03 -07001350 struct rb_node *parent = NULL;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001351 int nid;
1352
1353 nid = get_kpfn_nid(page_to_pfn(page));
Hugh Dickinsef53d162013-02-22 16:36:12 -08001354 root = root_unstable_tree + nid;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001355 new = &root->rb_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001356
1357 while (*new) {
1358 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001359 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001360 int ret;
1361
Hugh Dickinsd178f272009-11-09 15:58:23 +00001362 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -07001363 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001364 tree_page = get_mergeable_page(tree_rmap_item);
Andrea Arcangelic8f95ed2015-11-05 18:49:19 -08001365 if (!tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001366 return NULL;
1367
1368 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001369 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -07001370 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001371 if (page == tree_page) {
1372 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001373 return NULL;
1374 }
1375
Hugh Dickins8dd35572009-12-14 17:59:18 -08001376 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001377
1378 parent = *new;
1379 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001380 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001381 new = &parent->rb_left;
1382 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001383 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001384 new = &parent->rb_right;
Hugh Dickinsb599cbd2013-02-22 16:36:05 -08001385 } else if (!ksm_merge_across_nodes &&
1386 page_to_nid(tree_page) != nid) {
1387 /*
1388 * If tree_page has been migrated to another NUMA node,
1389 * it will be flushed out and put in the right unstable
1390 * tree next time: only merge with it when across_nodes.
1391 */
1392 put_page(tree_page);
1393 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001394 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001395 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001396 return tree_rmap_item;
1397 }
1398 }
1399
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001400 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001401 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001402 DO_NUMA(rmap_item->nid = nid);
Izik Eidus31dbd012009-09-21 17:02:03 -07001403 rb_link_node(&rmap_item->node, parent, new);
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001404 rb_insert_color(&rmap_item->node, root);
Izik Eidus31dbd012009-09-21 17:02:03 -07001405
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001406 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001407 return NULL;
1408}
1409
1410/*
1411 * stable_tree_append - add another rmap_item to the linked list of
1412 * rmap_items hanging off a given node of the stable tree, all sharing
1413 * the same ksm page.
1414 */
1415static void stable_tree_append(struct rmap_item *rmap_item,
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001416 struct stable_node *stable_node)
Izik Eidus31dbd012009-09-21 17:02:03 -07001417{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001418 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001419 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001420 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001421
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001422 if (rmap_item->hlist.next)
1423 ksm_pages_sharing++;
1424 else
1425 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001426}
1427
1428/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001429 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1430 * if not, compare checksum to previous and if it's the same, see if page can
1431 * be inserted into the unstable tree, or merged with a page already there and
1432 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001433 *
1434 * @page: the page that we are searching identical page to.
1435 * @rmap_item: the reverse mapping into the virtual address of this page
1436 */
1437static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1438{
Izik Eidus31dbd012009-09-21 17:02:03 -07001439 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001440 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001441 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001442 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07001443 unsigned int checksum;
1444 int err;
1445
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001446 stable_node = page_stable_node(page);
1447 if (stable_node) {
1448 if (stable_node->head != &migrate_nodes &&
1449 get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1450 rb_erase(&stable_node->node,
Hugh Dickinsef53d162013-02-22 16:36:12 -08001451 root_stable_tree + NUMA(stable_node->nid));
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001452 stable_node->head = &migrate_nodes;
1453 list_add(&stable_node->list, stable_node->head);
1454 }
1455 if (stable_node->head != &migrate_nodes &&
1456 rmap_item->head == stable_node)
1457 return;
1458 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001459
1460 /* We first start with searching the page inside the stable tree */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001461 kpage = stable_tree_search(page);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001462 if (kpage == page && rmap_item->head == stable_node) {
1463 put_page(kpage);
1464 return;
1465 }
1466
1467 remove_rmap_item_from_tree(rmap_item);
1468
Hugh Dickins62b61f62009-12-14 17:59:33 -08001469 if (kpage) {
Hugh Dickins08beca42009-12-14 17:59:21 -08001470 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001471 if (!err) {
1472 /*
1473 * The page was successfully merged:
1474 * add its rmap_item to the stable tree.
1475 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001476 lock_page(kpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001477 stable_tree_append(rmap_item, page_stable_node(kpage));
Hugh Dickins5ad64682009-12-14 17:59:24 -08001478 unlock_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001479 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001480 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001481 return;
1482 }
1483
1484 /*
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001485 * If the hash value of the page has changed from the last time
1486 * we calculated it, this page is changing frequently: therefore we
1487 * don't want to insert it in the unstable tree, and we don't want
1488 * to waste our time searching for something identical to it there.
Izik Eidus31dbd012009-09-21 17:02:03 -07001489 */
1490 checksum = calc_checksum(page);
1491 if (rmap_item->oldchecksum != checksum) {
1492 rmap_item->oldchecksum = checksum;
1493 return;
1494 }
1495
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08001496 /*
1497 * Same checksum as an empty page. We attempt to merge it with the
1498 * appropriate zero page if the user enabled this via sysfs.
1499 */
1500 if (ksm_use_zero_pages && (checksum == zero_checksum)) {
1501 struct vm_area_struct *vma;
1502
1503 vma = find_mergeable_vma(rmap_item->mm, rmap_item->address);
1504 err = try_to_merge_one_page(vma, page,
1505 ZERO_PAGE(rmap_item->address));
1506 /*
1507 * In case of failure, the page was not really empty, so we
1508 * need to continue. Otherwise we're done.
1509 */
1510 if (!err)
1511 return;
1512 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001513 tree_rmap_item =
1514 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001515 if (tree_rmap_item) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001516 kpage = try_to_merge_two_pages(rmap_item, page,
1517 tree_rmap_item, tree_page);
1518 put_page(tree_page);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001519 if (kpage) {
Hugh Dickinsbc566202013-02-22 16:36:06 -08001520 /*
1521 * The pages were successfully merged: insert new
1522 * node in the stable tree and add both rmap_items.
1523 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001524 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001525 stable_node = stable_tree_insert(kpage);
1526 if (stable_node) {
1527 stable_tree_append(tree_rmap_item, stable_node);
1528 stable_tree_append(rmap_item, stable_node);
1529 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001530 unlock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001531
Izik Eidus31dbd012009-09-21 17:02:03 -07001532 /*
1533 * If we fail to insert the page into the stable tree,
1534 * we will have 2 virtual addresses that are pointing
1535 * to a ksm page left outside the stable tree,
1536 * in which case we need to break_cow on both.
1537 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001538 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001539 break_cow(tree_rmap_item);
1540 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001541 }
1542 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001543 }
1544}
1545
1546static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001547 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07001548 unsigned long addr)
1549{
1550 struct rmap_item *rmap_item;
1551
Hugh Dickins6514d512009-12-14 17:59:19 -08001552 while (*rmap_list) {
1553 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08001554 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07001555 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001556 if (rmap_item->address > addr)
1557 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08001558 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001559 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001560 free_rmap_item(rmap_item);
1561 }
1562
1563 rmap_item = alloc_rmap_item();
1564 if (rmap_item) {
1565 /* It has already been zeroed */
1566 rmap_item->mm = mm_slot->mm;
1567 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08001568 rmap_item->rmap_list = *rmap_list;
1569 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001570 }
1571 return rmap_item;
1572}
1573
1574static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1575{
1576 struct mm_struct *mm;
1577 struct mm_slot *slot;
1578 struct vm_area_struct *vma;
1579 struct rmap_item *rmap_item;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001580 int nid;
Izik Eidus31dbd012009-09-21 17:02:03 -07001581
1582 if (list_empty(&ksm_mm_head.mm_list))
1583 return NULL;
1584
1585 slot = ksm_scan.mm_slot;
1586 if (slot == &ksm_mm_head) {
Hugh Dickins2919bfd2011-01-13 15:47:29 -08001587 /*
1588 * A number of pages can hang around indefinitely on per-cpu
1589 * pagevecs, raised page count preventing write_protect_page
1590 * from merging them. Though it doesn't really matter much,
1591 * it is puzzling to see some stuck in pages_volatile until
1592 * other activity jostles them out, and they also prevented
1593 * LTP's KSM test from succeeding deterministically; so drain
1594 * them here (here rather than on entry to ksm_do_scan(),
1595 * so we don't IPI too often when pages_to_scan is set low).
1596 */
1597 lru_add_drain_all();
1598
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001599 /*
1600 * Whereas stale stable_nodes on the stable_tree itself
1601 * get pruned in the regular course of stable_tree_search(),
1602 * those moved out to the migrate_nodes list can accumulate:
1603 * so prune them once before each full scan.
1604 */
1605 if (!ksm_merge_across_nodes) {
Geliang Tang03640412016-01-14 15:20:54 -08001606 struct stable_node *stable_node, *next;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001607 struct page *page;
1608
Geliang Tang03640412016-01-14 15:20:54 -08001609 list_for_each_entry_safe(stable_node, next,
1610 &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001611 page = get_ksm_page(stable_node, false);
1612 if (page)
1613 put_page(page);
1614 cond_resched();
1615 }
1616 }
1617
Hugh Dickinsef53d162013-02-22 16:36:12 -08001618 for (nid = 0; nid < ksm_nr_node_ids; nid++)
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001619 root_unstable_tree[nid] = RB_ROOT;
Izik Eidus31dbd012009-09-21 17:02:03 -07001620
1621 spin_lock(&ksm_mmlist_lock);
1622 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1623 ksm_scan.mm_slot = slot;
1624 spin_unlock(&ksm_mmlist_lock);
Hugh Dickins2b472612011-06-15 15:08:58 -07001625 /*
1626 * Although we tested list_empty() above, a racing __ksm_exit
1627 * of the last mm on the list may have removed it since then.
1628 */
1629 if (slot == &ksm_mm_head)
1630 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001631next_mm:
1632 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001633 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001634 }
1635
1636 mm = slot->mm;
1637 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001638 if (ksm_test_exit(mm))
1639 vma = NULL;
1640 else
1641 vma = find_vma(mm, ksm_scan.address);
1642
1643 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001644 if (!(vma->vm_flags & VM_MERGEABLE))
1645 continue;
1646 if (ksm_scan.address < vma->vm_start)
1647 ksm_scan.address = vma->vm_start;
1648 if (!vma->anon_vma)
1649 ksm_scan.address = vma->vm_end;
1650
1651 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001652 if (ksm_test_exit(mm))
1653 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001654 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001655 if (IS_ERR_OR_NULL(*page)) {
1656 ksm_scan.address += PAGE_SIZE;
1657 cond_resched();
1658 continue;
1659 }
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001660 if (PageAnon(*page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001661 flush_anon_page(vma, *page, ksm_scan.address);
1662 flush_dcache_page(*page);
1663 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001664 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07001665 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001666 ksm_scan.rmap_list =
1667 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001668 ksm_scan.address += PAGE_SIZE;
1669 } else
1670 put_page(*page);
1671 up_read(&mm->mmap_sem);
1672 return rmap_item;
1673 }
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001674 put_page(*page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001675 ksm_scan.address += PAGE_SIZE;
1676 cond_resched();
1677 }
1678 }
1679
Hugh Dickins9ba69292009-09-21 17:02:20 -07001680 if (ksm_test_exit(mm)) {
1681 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001682 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001683 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001684 /*
1685 * Nuke all the rmap_items that are above this current rmap:
1686 * because there were no VM_MERGEABLE vmas with such addresses.
1687 */
Hugh Dickins6514d512009-12-14 17:59:19 -08001688 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001689
1690 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001691 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1692 struct mm_slot, mm_list);
1693 if (ksm_scan.address == 0) {
1694 /*
1695 * We've completed a full scan of all vmas, holding mmap_sem
1696 * throughout, and found no VM_MERGEABLE: so do the same as
1697 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07001698 * This applies either when cleaning up after __ksm_exit
1699 * (but beware: we can reach here even before __ksm_exit),
1700 * or when all VM_MERGEABLE areas have been unmapped (and
1701 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07001702 */
Sasha Levin4ca3a692013-02-22 16:32:28 -08001703 hash_del(&slot->link);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001704 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001705 spin_unlock(&ksm_mmlist_lock);
1706
Hugh Dickinscd551f92009-09-21 17:02:17 -07001707 free_mm_slot(slot);
1708 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001709 up_read(&mm->mmap_sem);
1710 mmdrop(mm);
1711 } else {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001712 up_read(&mm->mmap_sem);
Zhou Chengming7496fea2016-05-12 15:42:21 -07001713 /*
1714 * up_read(&mm->mmap_sem) first because after
1715 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
1716 * already have been freed under us by __ksm_exit()
1717 * because the "mm_slot" is still hashed and
1718 * ksm_scan.mm_slot doesn't point to it anymore.
1719 */
1720 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001721 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001722
1723 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07001724 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07001725 if (slot != &ksm_mm_head)
1726 goto next_mm;
1727
Izik Eidus31dbd012009-09-21 17:02:03 -07001728 ksm_scan.seqnr++;
1729 return NULL;
1730}
1731
1732/**
1733 * ksm_do_scan - the ksm scanner main worker function.
1734 * @scan_npages - number of pages we want to scan before we return.
1735 */
1736static void ksm_do_scan(unsigned int scan_npages)
1737{
1738 struct rmap_item *rmap_item;
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001739 struct page *uninitialized_var(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001740
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001741 while (scan_npages-- && likely(!freezing(current))) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001742 cond_resched();
1743 rmap_item = scan_get_next_rmap_item(&page);
1744 if (!rmap_item)
1745 return;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001746 cmp_and_merge_page(page, rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001747 put_page(page);
1748 }
1749}
1750
Hugh Dickins6e1583842009-09-21 17:02:14 -07001751static int ksmd_should_run(void)
1752{
1753 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1754}
1755
Izik Eidus31dbd012009-09-21 17:02:03 -07001756static int ksm_scan_thread(void *nothing)
1757{
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001758 set_freezable();
Izik Eidus339aa622009-09-21 17:02:07 -07001759 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07001760
1761 while (!kthread_should_stop()) {
Hugh Dickins6e1583842009-09-21 17:02:14 -07001762 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08001763 wait_while_offlining();
Hugh Dickins6e1583842009-09-21 17:02:14 -07001764 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07001765 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001766 mutex_unlock(&ksm_thread_mutex);
1767
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001768 try_to_freeze();
1769
Hugh Dickins6e1583842009-09-21 17:02:14 -07001770 if (ksmd_should_run()) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001771 schedule_timeout_interruptible(
1772 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1773 } else {
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001774 wait_event_freezable(ksm_thread_wait,
Hugh Dickins6e1583842009-09-21 17:02:14 -07001775 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07001776 }
1777 }
1778 return 0;
1779}
1780
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001781int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1782 unsigned long end, int advice, unsigned long *vm_flags)
1783{
1784 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001785 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001786
1787 switch (advice) {
1788 case MADV_MERGEABLE:
1789 /*
1790 * Be somewhat over-protective for now!
1791 */
1792 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1793 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
Kirill A. Shutemov0661a332015-02-10 14:10:04 -08001794 VM_HUGETLB | VM_MIXEDMAP))
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001795 return 0; /* just ignore the advice */
1796
Konstantin Khlebnikovcc2383e2012-10-08 16:28:37 -07001797#ifdef VM_SAO
1798 if (*vm_flags & VM_SAO)
1799 return 0;
1800#endif
1801
Hugh Dickinsd952b792009-09-21 17:02:16 -07001802 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1803 err = __ksm_enter(mm);
1804 if (err)
1805 return err;
1806 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001807
1808 *vm_flags |= VM_MERGEABLE;
1809 break;
1810
1811 case MADV_UNMERGEABLE:
1812 if (!(*vm_flags & VM_MERGEABLE))
1813 return 0; /* just ignore the advice */
1814
Hugh Dickinsd952b792009-09-21 17:02:16 -07001815 if (vma->anon_vma) {
1816 err = unmerge_ksm_pages(vma, start, end);
1817 if (err)
1818 return err;
1819 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001820
1821 *vm_flags &= ~VM_MERGEABLE;
1822 break;
1823 }
1824
1825 return 0;
1826}
1827
1828int __ksm_enter(struct mm_struct *mm)
1829{
Hugh Dickins6e1583842009-09-21 17:02:14 -07001830 struct mm_slot *mm_slot;
1831 int needs_wakeup;
1832
1833 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07001834 if (!mm_slot)
1835 return -ENOMEM;
1836
Hugh Dickins6e1583842009-09-21 17:02:14 -07001837 /* Check ksm_run too? Would need tighter locking */
1838 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1839
Izik Eidus31dbd012009-09-21 17:02:03 -07001840 spin_lock(&ksm_mmlist_lock);
1841 insert_to_mm_slots_hash(mm, mm_slot);
1842 /*
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001843 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1844 * insert just behind the scanning cursor, to let the area settle
Izik Eidus31dbd012009-09-21 17:02:03 -07001845 * down a little; when fork is followed by immediate exec, we don't
1846 * want ksmd to waste time setting up and tearing down an rmap_list.
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001847 *
1848 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1849 * scanning cursor, otherwise KSM pages in newly forked mms will be
1850 * missed: then we might as well insert at the end of the list.
Izik Eidus31dbd012009-09-21 17:02:03 -07001851 */
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001852 if (ksm_run & KSM_RUN_UNMERGE)
1853 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1854 else
1855 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001856 spin_unlock(&ksm_mmlist_lock);
1857
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001858 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Vegard Nossumf1f10072017-02-27 14:30:07 -08001859 mmgrab(mm);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001860
1861 if (needs_wakeup)
1862 wake_up_interruptible(&ksm_thread_wait);
1863
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001864 return 0;
1865}
1866
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07001867void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001868{
Hugh Dickinscd551f92009-09-21 17:02:17 -07001869 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001870 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07001871
Izik Eidus31dbd012009-09-21 17:02:03 -07001872 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07001873 * This process is exiting: if it's straightforward (as is the
1874 * case when ksmd was never running), free mm_slot immediately.
1875 * But if it's at the cursor or has rmap_items linked to it, use
1876 * mmap_sem to synchronize with any break_cows before pagetables
1877 * are freed, and leave the mm_slot on the list for ksmd to free.
1878 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07001879 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07001880
Hugh Dickinscd551f92009-09-21 17:02:17 -07001881 spin_lock(&ksm_mmlist_lock);
1882 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001883 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001884 if (!mm_slot->rmap_list) {
Sasha Levin4ca3a692013-02-22 16:32:28 -08001885 hash_del(&mm_slot->link);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001886 list_del(&mm_slot->mm_list);
1887 easy_to_free = 1;
1888 } else {
1889 list_move(&mm_slot->mm_list,
1890 &ksm_scan.mm_slot->mm_list);
1891 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001892 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001893 spin_unlock(&ksm_mmlist_lock);
1894
Hugh Dickins9ba69292009-09-21 17:02:20 -07001895 if (easy_to_free) {
1896 free_mm_slot(mm_slot);
1897 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1898 mmdrop(mm);
1899 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001900 down_write(&mm->mmap_sem);
1901 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001902 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001903}
Izik Eidus31dbd012009-09-21 17:02:03 -07001904
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001905struct page *ksm_might_need_to_copy(struct page *page,
Hugh Dickins5ad64682009-12-14 17:59:24 -08001906 struct vm_area_struct *vma, unsigned long address)
1907{
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001908 struct anon_vma *anon_vma = page_anon_vma(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001909 struct page *new_page;
1910
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001911 if (PageKsm(page)) {
1912 if (page_stable_node(page) &&
1913 !(ksm_run & KSM_RUN_UNMERGE))
1914 return page; /* no need to copy it */
1915 } else if (!anon_vma) {
1916 return page; /* no need to copy it */
1917 } else if (anon_vma->root == vma->anon_vma->root &&
1918 page->index == linear_page_index(vma, address)) {
1919 return page; /* still no need to copy it */
1920 }
1921 if (!PageUptodate(page))
1922 return page; /* let do_swap_page report the error */
1923
Hugh Dickins5ad64682009-12-14 17:59:24 -08001924 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1925 if (new_page) {
1926 copy_user_highpage(new_page, page, address, vma);
1927
1928 SetPageDirty(new_page);
1929 __SetPageUptodate(new_page);
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001930 __SetPageLocked(new_page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001931 }
1932
Hugh Dickins5ad64682009-12-14 17:59:24 -08001933 return new_page;
1934}
1935
Minchan Kim1df631a2017-05-03 14:54:23 -07001936void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001937{
1938 struct stable_node *stable_node;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001939 struct rmap_item *rmap_item;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001940 int search_new_forks = 0;
1941
Sasha Levin309381fea2014-01-23 15:52:54 -08001942 VM_BUG_ON_PAGE(!PageKsm(page), page);
Joonsoo Kim9f326242014-01-21 15:49:53 -08001943
1944 /*
1945 * Rely on the page lock to protect against concurrent modifications
1946 * to that page's node of the stable tree.
1947 */
Sasha Levin309381fea2014-01-23 15:52:54 -08001948 VM_BUG_ON_PAGE(!PageLocked(page), page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001949
1950 stable_node = page_stable_node(page);
1951 if (!stable_node)
Minchan Kim1df631a2017-05-03 14:54:23 -07001952 return;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001953again:
Sasha Levinb67bfe02013-02-27 17:06:00 -08001954 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001955 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001956 struct anon_vma_chain *vmac;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001957 struct vm_area_struct *vma;
1958
Andrea Arcangeliad126952015-11-05 18:49:07 -08001959 cond_resched();
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001960 anon_vma_lock_read(anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -07001961 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1962 0, ULONG_MAX) {
Andrea Arcangeliad126952015-11-05 18:49:07 -08001963 cond_resched();
Rik van Riel5beb4932010-03-05 13:42:07 -08001964 vma = vmac->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001965 if (rmap_item->address < vma->vm_start ||
1966 rmap_item->address >= vma->vm_end)
1967 continue;
1968 /*
1969 * Initially we examine only the vma which covers this
1970 * rmap_item; but later, if there is still work to do,
1971 * we examine covering vmas in other mms: in case they
1972 * were forked from the original since ksmd passed.
1973 */
1974 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1975 continue;
1976
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001977 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1978 continue;
1979
Minchan Kime4b82222017-05-03 14:54:27 -07001980 if (!rwc->rmap_one(page, vma,
Minchan Kim1df631a2017-05-03 14:54:23 -07001981 rmap_item->address, rwc->arg)) {
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001982 anon_vma_unlock_read(anon_vma);
Minchan Kim1df631a2017-05-03 14:54:23 -07001983 return;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001984 }
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001985 if (rwc->done && rwc->done(page)) {
1986 anon_vma_unlock_read(anon_vma);
Minchan Kim1df631a2017-05-03 14:54:23 -07001987 return;
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001988 }
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001989 }
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001990 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001991 }
1992 if (!search_new_forks++)
1993 goto again;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001994}
1995
Joonsoo Kim52629502014-01-21 15:49:50 -08001996#ifdef CONFIG_MIGRATION
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001997void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1998{
1999 struct stable_node *stable_node;
2000
Sasha Levin309381fea2014-01-23 15:52:54 -08002001 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
2002 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
2003 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002004
2005 stable_node = page_stable_node(newpage);
2006 if (stable_node) {
Sasha Levin309381fea2014-01-23 15:52:54 -08002007 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002008 stable_node->kpfn = page_to_pfn(newpage);
Hugh Dickinsc8d65532013-02-22 16:35:10 -08002009 /*
2010 * newpage->mapping was set in advance; now we need smp_wmb()
2011 * to make sure that the new stable_node->kpfn is visible
2012 * to get_ksm_page() before it can see that oldpage->mapping
2013 * has gone stale (or that PageSwapCache has been cleared).
2014 */
2015 smp_wmb();
2016 set_page_stable_node(oldpage, NULL);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002017 }
2018}
2019#endif /* CONFIG_MIGRATION */
2020
Hugh Dickins62b61f62009-12-14 17:59:33 -08002021#ifdef CONFIG_MEMORY_HOTREMOVE
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002022static void wait_while_offlining(void)
2023{
2024 while (ksm_run & KSM_RUN_OFFLINE) {
2025 mutex_unlock(&ksm_thread_mutex);
2026 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
NeilBrown74316202014-07-07 15:16:04 +10002027 TASK_UNINTERRUPTIBLE);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002028 mutex_lock(&ksm_thread_mutex);
2029 }
2030}
2031
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002032static void ksm_check_stable_tree(unsigned long start_pfn,
2033 unsigned long end_pfn)
Hugh Dickins62b61f62009-12-14 17:59:33 -08002034{
Geliang Tang03640412016-01-14 15:20:54 -08002035 struct stable_node *stable_node, *next;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002036 struct rb_node *node;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002037 int nid;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002038
Hugh Dickinsef53d162013-02-22 16:36:12 -08002039 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2040 node = rb_first(root_stable_tree + nid);
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002041 while (node) {
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002042 stable_node = rb_entry(node, struct stable_node, node);
2043 if (stable_node->kpfn >= start_pfn &&
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002044 stable_node->kpfn < end_pfn) {
2045 /*
2046 * Don't get_ksm_page, page has already gone:
2047 * which is why we keep kpfn instead of page*
2048 */
2049 remove_node_from_stable_tree(stable_node);
Hugh Dickinsef53d162013-02-22 16:36:12 -08002050 node = rb_first(root_stable_tree + nid);
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002051 } else
2052 node = rb_next(node);
2053 cond_resched();
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002054 }
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002055 }
Geliang Tang03640412016-01-14 15:20:54 -08002056 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002057 if (stable_node->kpfn >= start_pfn &&
2058 stable_node->kpfn < end_pfn)
2059 remove_node_from_stable_tree(stable_node);
2060 cond_resched();
2061 }
Hugh Dickins62b61f62009-12-14 17:59:33 -08002062}
2063
2064static int ksm_memory_callback(struct notifier_block *self,
2065 unsigned long action, void *arg)
2066{
2067 struct memory_notify *mn = arg;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002068
2069 switch (action) {
2070 case MEM_GOING_OFFLINE:
2071 /*
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002072 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2073 * and remove_all_stable_nodes() while memory is going offline:
2074 * it is unsafe for them to touch the stable tree at this time.
2075 * But unmerge_ksm_pages(), rmap lookups and other entry points
2076 * which do not need the ksm_thread_mutex are all safe.
Hugh Dickins62b61f62009-12-14 17:59:33 -08002077 */
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002078 mutex_lock(&ksm_thread_mutex);
2079 ksm_run |= KSM_RUN_OFFLINE;
2080 mutex_unlock(&ksm_thread_mutex);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002081 break;
2082
2083 case MEM_OFFLINE:
2084 /*
2085 * Most of the work is done by page migration; but there might
2086 * be a few stable_nodes left over, still pointing to struct
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002087 * pages which have been offlined: prune those from the tree,
2088 * otherwise get_ksm_page() might later try to access a
2089 * non-existent struct page.
Hugh Dickins62b61f62009-12-14 17:59:33 -08002090 */
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002091 ksm_check_stable_tree(mn->start_pfn,
2092 mn->start_pfn + mn->nr_pages);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002093 /* fallthrough */
2094
2095 case MEM_CANCEL_OFFLINE:
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002096 mutex_lock(&ksm_thread_mutex);
2097 ksm_run &= ~KSM_RUN_OFFLINE;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002098 mutex_unlock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002099
2100 smp_mb(); /* wake_up_bit advises this */
2101 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
Hugh Dickins62b61f62009-12-14 17:59:33 -08002102 break;
2103 }
2104 return NOTIFY_OK;
2105}
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002106#else
2107static void wait_while_offlining(void)
2108{
2109}
Hugh Dickins62b61f62009-12-14 17:59:33 -08002110#endif /* CONFIG_MEMORY_HOTREMOVE */
2111
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002112#ifdef CONFIG_SYSFS
2113/*
2114 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2115 */
2116
Izik Eidus31dbd012009-09-21 17:02:03 -07002117#define KSM_ATTR_RO(_name) \
2118 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2119#define KSM_ATTR(_name) \
2120 static struct kobj_attribute _name##_attr = \
2121 __ATTR(_name, 0644, _name##_show, _name##_store)
2122
2123static ssize_t sleep_millisecs_show(struct kobject *kobj,
2124 struct kobj_attribute *attr, char *buf)
2125{
2126 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2127}
2128
2129static ssize_t sleep_millisecs_store(struct kobject *kobj,
2130 struct kobj_attribute *attr,
2131 const char *buf, size_t count)
2132{
2133 unsigned long msecs;
2134 int err;
2135
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002136 err = kstrtoul(buf, 10, &msecs);
Izik Eidus31dbd012009-09-21 17:02:03 -07002137 if (err || msecs > UINT_MAX)
2138 return -EINVAL;
2139
2140 ksm_thread_sleep_millisecs = msecs;
2141
2142 return count;
2143}
2144KSM_ATTR(sleep_millisecs);
2145
2146static ssize_t pages_to_scan_show(struct kobject *kobj,
2147 struct kobj_attribute *attr, char *buf)
2148{
2149 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2150}
2151
2152static ssize_t pages_to_scan_store(struct kobject *kobj,
2153 struct kobj_attribute *attr,
2154 const char *buf, size_t count)
2155{
2156 int err;
2157 unsigned long nr_pages;
2158
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002159 err = kstrtoul(buf, 10, &nr_pages);
Izik Eidus31dbd012009-09-21 17:02:03 -07002160 if (err || nr_pages > UINT_MAX)
2161 return -EINVAL;
2162
2163 ksm_thread_pages_to_scan = nr_pages;
2164
2165 return count;
2166}
2167KSM_ATTR(pages_to_scan);
2168
2169static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2170 char *buf)
2171{
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002172 return sprintf(buf, "%lu\n", ksm_run);
Izik Eidus31dbd012009-09-21 17:02:03 -07002173}
2174
2175static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2176 const char *buf, size_t count)
2177{
2178 int err;
2179 unsigned long flags;
2180
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002181 err = kstrtoul(buf, 10, &flags);
Izik Eidus31dbd012009-09-21 17:02:03 -07002182 if (err || flags > UINT_MAX)
2183 return -EINVAL;
2184 if (flags > KSM_RUN_UNMERGE)
2185 return -EINVAL;
2186
2187 /*
2188 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2189 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsd0f209f2009-12-14 17:59:34 -08002190 * breaking COW to free the pages_shared (but leaves mm_slots
2191 * on the list for when ksmd may be set running again).
Izik Eidus31dbd012009-09-21 17:02:03 -07002192 */
2193
2194 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002195 wait_while_offlining();
Izik Eidus31dbd012009-09-21 17:02:03 -07002196 if (ksm_run != flags) {
2197 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07002198 if (flags & KSM_RUN_UNMERGE) {
David Rientjese1e12d22012-12-11 16:02:56 -08002199 set_current_oom_origin();
Hugh Dickinsd952b792009-09-21 17:02:16 -07002200 err = unmerge_and_remove_all_rmap_items();
David Rientjese1e12d22012-12-11 16:02:56 -08002201 clear_current_oom_origin();
Hugh Dickinsd952b792009-09-21 17:02:16 -07002202 if (err) {
2203 ksm_run = KSM_RUN_STOP;
2204 count = err;
2205 }
2206 }
Izik Eidus31dbd012009-09-21 17:02:03 -07002207 }
2208 mutex_unlock(&ksm_thread_mutex);
2209
2210 if (flags & KSM_RUN_MERGE)
2211 wake_up_interruptible(&ksm_thread_wait);
2212
2213 return count;
2214}
2215KSM_ATTR(run);
2216
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002217#ifdef CONFIG_NUMA
2218static ssize_t merge_across_nodes_show(struct kobject *kobj,
2219 struct kobj_attribute *attr, char *buf)
2220{
2221 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2222}
2223
2224static ssize_t merge_across_nodes_store(struct kobject *kobj,
2225 struct kobj_attribute *attr,
2226 const char *buf, size_t count)
2227{
2228 int err;
2229 unsigned long knob;
2230
2231 err = kstrtoul(buf, 10, &knob);
2232 if (err)
2233 return err;
2234 if (knob > 1)
2235 return -EINVAL;
2236
2237 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002238 wait_while_offlining();
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002239 if (ksm_merge_across_nodes != knob) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002240 if (ksm_pages_shared || remove_all_stable_nodes())
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002241 err = -EBUSY;
Hugh Dickinsef53d162013-02-22 16:36:12 -08002242 else if (root_stable_tree == one_stable_tree) {
2243 struct rb_root *buf;
2244 /*
2245 * This is the first time that we switch away from the
2246 * default of merging across nodes: must now allocate
2247 * a buffer to hold as many roots as may be needed.
2248 * Allocate stable and unstable together:
2249 * MAXSMP NODES_SHIFT 10 will use 16kB.
2250 */
Joe Perchesbafe1e12013-11-12 15:07:10 -08002251 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2252 GFP_KERNEL);
Hugh Dickinsef53d162013-02-22 16:36:12 -08002253 /* Let us assume that RB_ROOT is NULL is zero */
2254 if (!buf)
2255 err = -ENOMEM;
2256 else {
2257 root_stable_tree = buf;
2258 root_unstable_tree = buf + nr_node_ids;
2259 /* Stable tree is empty but not the unstable */
2260 root_unstable_tree[0] = one_unstable_tree[0];
2261 }
2262 }
2263 if (!err) {
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002264 ksm_merge_across_nodes = knob;
Hugh Dickinsef53d162013-02-22 16:36:12 -08002265 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2266 }
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002267 }
2268 mutex_unlock(&ksm_thread_mutex);
2269
2270 return err ? err : count;
2271}
2272KSM_ATTR(merge_across_nodes);
2273#endif
2274
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08002275static ssize_t use_zero_pages_show(struct kobject *kobj,
2276 struct kobj_attribute *attr, char *buf)
2277{
2278 return sprintf(buf, "%u\n", ksm_use_zero_pages);
2279}
2280static ssize_t use_zero_pages_store(struct kobject *kobj,
2281 struct kobj_attribute *attr,
2282 const char *buf, size_t count)
2283{
2284 int err;
2285 bool value;
2286
2287 err = kstrtobool(buf, &value);
2288 if (err)
2289 return -EINVAL;
2290
2291 ksm_use_zero_pages = value;
2292
2293 return count;
2294}
2295KSM_ATTR(use_zero_pages);
2296
Hugh Dickinsb4028262009-09-21 17:02:09 -07002297static ssize_t pages_shared_show(struct kobject *kobj,
2298 struct kobj_attribute *attr, char *buf)
2299{
2300 return sprintf(buf, "%lu\n", ksm_pages_shared);
2301}
2302KSM_ATTR_RO(pages_shared);
2303
2304static ssize_t pages_sharing_show(struct kobject *kobj,
2305 struct kobj_attribute *attr, char *buf)
2306{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07002307 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07002308}
2309KSM_ATTR_RO(pages_sharing);
2310
Hugh Dickins473b0ce2009-09-21 17:02:11 -07002311static ssize_t pages_unshared_show(struct kobject *kobj,
2312 struct kobj_attribute *attr, char *buf)
2313{
2314 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2315}
2316KSM_ATTR_RO(pages_unshared);
2317
2318static ssize_t pages_volatile_show(struct kobject *kobj,
2319 struct kobj_attribute *attr, char *buf)
2320{
2321 long ksm_pages_volatile;
2322
2323 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2324 - ksm_pages_sharing - ksm_pages_unshared;
2325 /*
2326 * It was not worth any locking to calculate that statistic,
2327 * but it might therefore sometimes be negative: conceal that.
2328 */
2329 if (ksm_pages_volatile < 0)
2330 ksm_pages_volatile = 0;
2331 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2332}
2333KSM_ATTR_RO(pages_volatile);
2334
2335static ssize_t full_scans_show(struct kobject *kobj,
2336 struct kobj_attribute *attr, char *buf)
2337{
2338 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2339}
2340KSM_ATTR_RO(full_scans);
2341
Izik Eidus31dbd012009-09-21 17:02:03 -07002342static struct attribute *ksm_attrs[] = {
2343 &sleep_millisecs_attr.attr,
2344 &pages_to_scan_attr.attr,
2345 &run_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07002346 &pages_shared_attr.attr,
2347 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07002348 &pages_unshared_attr.attr,
2349 &pages_volatile_attr.attr,
2350 &full_scans_attr.attr,
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002351#ifdef CONFIG_NUMA
2352 &merge_across_nodes_attr.attr,
2353#endif
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08002354 &use_zero_pages_attr.attr,
Izik Eidus31dbd012009-09-21 17:02:03 -07002355 NULL,
2356};
2357
2358static struct attribute_group ksm_attr_group = {
2359 .attrs = ksm_attrs,
2360 .name = "ksm",
2361};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002362#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002363
2364static int __init ksm_init(void)
2365{
2366 struct task_struct *ksm_thread;
2367 int err;
2368
Claudio Imbrendae86c59b2017-02-24 14:55:39 -08002369 /* The correct value depends on page size and endianness */
2370 zero_checksum = calc_checksum(ZERO_PAGE(0));
2371 /* Default to false for backwards compatibility */
2372 ksm_use_zero_pages = false;
2373
Izik Eidus31dbd012009-09-21 17:02:03 -07002374 err = ksm_slab_init();
2375 if (err)
2376 goto out;
2377
Izik Eidus31dbd012009-09-21 17:02:03 -07002378 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2379 if (IS_ERR(ksm_thread)) {
Paul McQuade25acde32014-10-09 15:29:09 -07002380 pr_err("ksm: creating kthread failed\n");
Izik Eidus31dbd012009-09-21 17:02:03 -07002381 err = PTR_ERR(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002382 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002383 }
2384
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002385#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07002386 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2387 if (err) {
Paul McQuade25acde32014-10-09 15:29:09 -07002388 pr_err("ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002389 kthread_stop(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002390 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002391 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07002392#else
2393 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2394
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002395#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002396
Hugh Dickins62b61f62009-12-14 17:59:33 -08002397#ifdef CONFIG_MEMORY_HOTREMOVE
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002398 /* There is no significance to this priority 100 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08002399 hotplug_memory_notifier(ksm_memory_callback, 100);
2400#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07002401 return 0;
2402
Lai Jiangshand9f89842010-08-09 17:20:02 -07002403out_free:
Izik Eidus31dbd012009-09-21 17:02:03 -07002404 ksm_slab_free();
2405out:
2406 return err;
2407}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -08002408subsys_initcall(ksm_init);