blob: 9ae6011a41f895d56942175814d1fe0158a591b7 [file] [log] [blame]
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001/*
Izik Eidus31dbd012009-09-21 17:02:03 -07002 * Memory merging support.
3 *
4 * This code enables dynamic sharing of identical pages found in different
5 * memory areas, even if they are not shared by fork()
6 *
Izik Eidus36b25282009-09-21 17:02:06 -07007 * Copyright (C) 2008-2009 Red Hat, Inc.
Izik Eidus31dbd012009-09-21 17:02:03 -07008 * Authors:
9 * Izik Eidus
10 * Andrea Arcangeli
11 * Chris Wright
Izik Eidus36b25282009-09-21 17:02:06 -070012 * Hugh Dickins
Izik Eidus31dbd012009-09-21 17:02:03 -070013 *
14 * This work is licensed under the terms of the GNU GPL, version 2.
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070015 */
16
17#include <linux/errno.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070018#include <linux/mm.h>
19#include <linux/fs.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070020#include <linux/mman.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070021#include <linux/sched.h>
22#include <linux/rwsem.h>
23#include <linux/pagemap.h>
24#include <linux/rmap.h>
25#include <linux/spinlock.h>
26#include <linux/jhash.h>
27#include <linux/delay.h>
28#include <linux/kthread.h>
29#include <linux/wait.h>
30#include <linux/slab.h>
31#include <linux/rbtree.h>
Hugh Dickins62b61f62009-12-14 17:59:33 -080032#include <linux/memory.h>
Izik Eidus31dbd012009-09-21 17:02:03 -070033#include <linux/mmu_notifier.h>
Izik Eidus2c6854f2009-09-23 15:56:04 -070034#include <linux/swap.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070035#include <linux/ksm.h>
Sasha Levin4ca3a692013-02-22 16:32:28 -080036#include <linux/hashtable.h>
Andrea Arcangeli878aee72011-01-13 15:47:10 -080037#include <linux/freezer.h>
David Rientjes72788c32011-05-24 17:11:40 -070038#include <linux/oom.h>
Petr Holasek90bd6fd2013-02-22 16:35:00 -080039#include <linux/numa.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070040
Izik Eidus31dbd012009-09-21 17:02:03 -070041#include <asm/tlbflush.h>
Hugh Dickins73848b42009-12-14 17:59:22 -080042#include "internal.h"
Izik Eidus31dbd012009-09-21 17:02:03 -070043
Hugh Dickinse850dcf2013-02-22 16:35:03 -080044#ifdef CONFIG_NUMA
45#define NUMA(x) (x)
46#define DO_NUMA(x) do { (x); } while (0)
47#else
48#define NUMA(x) (0)
49#define DO_NUMA(x) do { } while (0)
50#endif
51
Izik Eidus31dbd012009-09-21 17:02:03 -070052/*
53 * A few notes about the KSM scanning process,
54 * to make it easier to understand the data structures below:
55 *
56 * In order to reduce excessive scanning, KSM sorts the memory pages by their
57 * contents into a data structure that holds pointers to the pages' locations.
58 *
59 * Since the contents of the pages may change at any moment, KSM cannot just
60 * insert the pages into a normal sorted tree and expect it to find anything.
61 * Therefore KSM uses two data structures - the stable and the unstable tree.
62 *
63 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
64 * by their contents. Because each such page is write-protected, searching on
65 * this tree is fully assured to be working (except when pages are unmapped),
66 * and therefore this tree is called the stable tree.
67 *
68 * In addition to the stable tree, KSM uses a second data structure called the
69 * unstable tree: this tree holds pointers to pages which have been found to
70 * be "unchanged for a period of time". The unstable tree sorts these pages
71 * by their contents, but since they are not write-protected, KSM cannot rely
72 * upon the unstable tree to work correctly - the unstable tree is liable to
73 * be corrupted as its contents are modified, and so it is called unstable.
74 *
75 * KSM solves this problem by several techniques:
76 *
77 * 1) The unstable tree is flushed every time KSM completes scanning all
78 * memory areas, and then the tree is rebuilt again from the beginning.
79 * 2) KSM will only insert into the unstable tree, pages whose hash value
80 * has not changed since the previous scan of all memory areas.
81 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
82 * colors of the nodes and not on their contents, assuring that even when
83 * the tree gets "corrupted" it won't get out of balance, so scanning time
84 * remains the same (also, searching and inserting nodes in an rbtree uses
85 * the same algorithm, so we have no overhead when we flush and rebuild).
86 * 4) KSM never flushes the stable tree, which means that even if it were to
87 * take 10 attempts to find a page in the unstable tree, once it is found,
88 * it is secured in the stable tree. (When we scan a new page, we first
89 * compare it against the stable tree, and then against the unstable tree.)
Hugh Dickins8fdb3db2013-02-22 16:36:03 -080090 *
91 * If the merge_across_nodes tunable is unset, then KSM maintains multiple
92 * stable trees and multiple unstable trees: one of each for each NUMA node.
Izik Eidus31dbd012009-09-21 17:02:03 -070093 */
94
95/**
96 * struct mm_slot - ksm information per mm that is being scanned
97 * @link: link to the mm_slots hash list
98 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
Hugh Dickins6514d512009-12-14 17:59:19 -080099 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
Izik Eidus31dbd012009-09-21 17:02:03 -0700100 * @mm: the mm that this information is valid for
101 */
102struct mm_slot {
103 struct hlist_node link;
104 struct list_head mm_list;
Hugh Dickins6514d512009-12-14 17:59:19 -0800105 struct rmap_item *rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700106 struct mm_struct *mm;
107};
108
109/**
110 * struct ksm_scan - cursor for scanning
111 * @mm_slot: the current mm_slot we are scanning
112 * @address: the next address inside that to be scanned
Hugh Dickins6514d512009-12-14 17:59:19 -0800113 * @rmap_list: link to the next rmap to be scanned in the rmap_list
Izik Eidus31dbd012009-09-21 17:02:03 -0700114 * @seqnr: count of completed full scans (needed when removing unstable node)
115 *
116 * There is only the one ksm_scan instance of this cursor structure.
117 */
118struct ksm_scan {
119 struct mm_slot *mm_slot;
120 unsigned long address;
Hugh Dickins6514d512009-12-14 17:59:19 -0800121 struct rmap_item **rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700122 unsigned long seqnr;
123};
124
125/**
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800126 * struct stable_node - node of the stable rbtree
127 * @node: rb node of this ksm page in the stable tree
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800128 * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list
129 * @list: linked into migrate_nodes, pending placement in the proper node tree
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800130 * @hlist: hlist head of rmap_items using this ksm page
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800131 * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid)
132 * @nid: NUMA node id of stable tree in which linked (may not match kpfn)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800133 */
134struct stable_node {
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800135 union {
136 struct rb_node node; /* when node of stable tree */
137 struct { /* when listed for migration */
138 struct list_head *head;
139 struct list_head list;
140 };
141 };
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800142 struct hlist_head hlist;
Hugh Dickins62b61f62009-12-14 17:59:33 -0800143 unsigned long kpfn;
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800144#ifdef CONFIG_NUMA
145 int nid;
146#endif
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800147};
148
149/**
Izik Eidus31dbd012009-09-21 17:02:03 -0700150 * struct rmap_item - reverse mapping item for virtual addresses
Hugh Dickins6514d512009-12-14 17:59:19 -0800151 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800152 * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree
Hugh Dickinsbc566202013-02-22 16:36:06 -0800153 * @nid: NUMA node id of unstable tree in which linked (may not match page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700154 * @mm: the memory structure this rmap_item is pointing into
155 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
156 * @oldchecksum: previous checksum of the page at that virtual address
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800157 * @node: rb node of this rmap_item in the unstable tree
158 * @head: pointer to stable_node heading this list in the stable tree
159 * @hlist: link into hlist of rmap_items hanging off that stable_node
Izik Eidus31dbd012009-09-21 17:02:03 -0700160 */
161struct rmap_item {
Hugh Dickins6514d512009-12-14 17:59:19 -0800162 struct rmap_item *rmap_list;
Hugh Dickinsbc566202013-02-22 16:36:06 -0800163 union {
164 struct anon_vma *anon_vma; /* when stable */
165#ifdef CONFIG_NUMA
166 int nid; /* when node of unstable tree */
167#endif
168 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700169 struct mm_struct *mm;
170 unsigned long address; /* + low bits used for flags below */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800171 unsigned int oldchecksum; /* when unstable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700172 union {
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800173 struct rb_node node; /* when node of unstable tree */
174 struct { /* when listed from stable tree */
175 struct stable_node *head;
176 struct hlist_node hlist;
177 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700178 };
179};
180
181#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800182#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
183#define STABLE_FLAG 0x200 /* is listed from the stable tree */
Izik Eidus31dbd012009-09-21 17:02:03 -0700184
185/* The stable and unstable tree heads */
Hugh Dickinsef53d162013-02-22 16:36:12 -0800186static struct rb_root one_stable_tree[1] = { RB_ROOT };
187static struct rb_root one_unstable_tree[1] = { RB_ROOT };
188static struct rb_root *root_stable_tree = one_stable_tree;
189static struct rb_root *root_unstable_tree = one_unstable_tree;
Izik Eidus31dbd012009-09-21 17:02:03 -0700190
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800191/* Recently migrated nodes of stable tree, pending proper placement */
192static LIST_HEAD(migrate_nodes);
193
Sasha Levin4ca3a692013-02-22 16:32:28 -0800194#define MM_SLOTS_HASH_BITS 10
195static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
Izik Eidus31dbd012009-09-21 17:02:03 -0700196
197static struct mm_slot ksm_mm_head = {
198 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
199};
200static struct ksm_scan ksm_scan = {
201 .mm_slot = &ksm_mm_head,
202};
203
204static struct kmem_cache *rmap_item_cache;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800205static struct kmem_cache *stable_node_cache;
Izik Eidus31dbd012009-09-21 17:02:03 -0700206static struct kmem_cache *mm_slot_cache;
207
208/* The number of nodes in the stable tree */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700209static unsigned long ksm_pages_shared;
Izik Eidus31dbd012009-09-21 17:02:03 -0700210
Hugh Dickinse178dfd2009-09-21 17:02:10 -0700211/* The number of page slots additionally sharing those nodes */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700212static unsigned long ksm_pages_sharing;
Izik Eidus31dbd012009-09-21 17:02:03 -0700213
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700214/* The number of nodes in the unstable tree */
215static unsigned long ksm_pages_unshared;
216
217/* The number of rmap_items in use: to calculate pages_volatile */
218static unsigned long ksm_rmap_items;
219
Izik Eidus31dbd012009-09-21 17:02:03 -0700220/* Number of pages ksmd should scan in one batch */
Izik Eidus2c6854f2009-09-23 15:56:04 -0700221static unsigned int ksm_thread_pages_to_scan = 100;
Izik Eidus31dbd012009-09-21 17:02:03 -0700222
223/* Milliseconds ksmd should sleep between batches */
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700224static unsigned int ksm_thread_sleep_millisecs = 20;
Izik Eidus31dbd012009-09-21 17:02:03 -0700225
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800226#ifdef CONFIG_NUMA
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800227/* Zeroed when merging across nodes is not allowed */
228static unsigned int ksm_merge_across_nodes = 1;
Hugh Dickinsef53d162013-02-22 16:36:12 -0800229static int ksm_nr_node_ids = 1;
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800230#else
231#define ksm_merge_across_nodes 1U
Hugh Dickinsef53d162013-02-22 16:36:12 -0800232#define ksm_nr_node_ids 1
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800233#endif
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800234
Izik Eidus31dbd012009-09-21 17:02:03 -0700235#define KSM_RUN_STOP 0
236#define KSM_RUN_MERGE 1
237#define KSM_RUN_UNMERGE 2
Hugh Dickinsef4d43a2013-02-22 16:35:16 -0800238#define KSM_RUN_OFFLINE 4
239static unsigned long ksm_run = KSM_RUN_STOP;
240static void wait_while_offlining(void);
Izik Eidus31dbd012009-09-21 17:02:03 -0700241
242static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
243static DEFINE_MUTEX(ksm_thread_mutex);
244static DEFINE_SPINLOCK(ksm_mmlist_lock);
245
246#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
247 sizeof(struct __struct), __alignof__(struct __struct),\
248 (__flags), NULL)
249
250static int __init ksm_slab_init(void)
251{
252 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
253 if (!rmap_item_cache)
254 goto out;
255
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800256 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
257 if (!stable_node_cache)
258 goto out_free1;
259
Izik Eidus31dbd012009-09-21 17:02:03 -0700260 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
261 if (!mm_slot_cache)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800262 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -0700263
264 return 0;
265
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800266out_free2:
267 kmem_cache_destroy(stable_node_cache);
268out_free1:
Izik Eidus31dbd012009-09-21 17:02:03 -0700269 kmem_cache_destroy(rmap_item_cache);
270out:
271 return -ENOMEM;
272}
273
274static void __init ksm_slab_free(void)
275{
276 kmem_cache_destroy(mm_slot_cache);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800277 kmem_cache_destroy(stable_node_cache);
Izik Eidus31dbd012009-09-21 17:02:03 -0700278 kmem_cache_destroy(rmap_item_cache);
279 mm_slot_cache = NULL;
280}
281
282static inline struct rmap_item *alloc_rmap_item(void)
283{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700284 struct rmap_item *rmap_item;
285
zhong jiang5b398e42016-09-28 15:22:30 -0700286 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL |
287 __GFP_NORETRY | __GFP_NOWARN);
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700288 if (rmap_item)
289 ksm_rmap_items++;
290 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -0700291}
292
293static inline void free_rmap_item(struct rmap_item *rmap_item)
294{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700295 ksm_rmap_items--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700296 rmap_item->mm = NULL; /* debug safety */
297 kmem_cache_free(rmap_item_cache, rmap_item);
298}
299
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800300static inline struct stable_node *alloc_stable_node(void)
301{
zhong jiang62130552016-10-07 17:01:19 -0700302 /*
303 * The allocation can take too long with GFP_KERNEL when memory is under
304 * pressure, which may lead to hung task warnings. Adding __GFP_HIGH
305 * grants access to memory reserves, helping to avoid this problem.
306 */
307 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800308}
309
310static inline void free_stable_node(struct stable_node *stable_node)
311{
312 kmem_cache_free(stable_node_cache, stable_node);
313}
314
Izik Eidus31dbd012009-09-21 17:02:03 -0700315static inline struct mm_slot *alloc_mm_slot(void)
316{
317 if (!mm_slot_cache) /* initialization failed */
318 return NULL;
319 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
320}
321
322static inline void free_mm_slot(struct mm_slot *mm_slot)
323{
324 kmem_cache_free(mm_slot_cache, mm_slot);
325}
326
Izik Eidus31dbd012009-09-21 17:02:03 -0700327static struct mm_slot *get_mm_slot(struct mm_struct *mm)
328{
Sasha Levin4ca3a692013-02-22 16:32:28 -0800329 struct mm_slot *slot;
Izik Eidus31dbd012009-09-21 17:02:03 -0700330
Sasha Levinb67bfe02013-02-27 17:06:00 -0800331 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
Sasha Levin4ca3a692013-02-22 16:32:28 -0800332 if (slot->mm == mm)
333 return slot;
334
Izik Eidus31dbd012009-09-21 17:02:03 -0700335 return NULL;
336}
337
338static void insert_to_mm_slots_hash(struct mm_struct *mm,
339 struct mm_slot *mm_slot)
340{
Izik Eidus31dbd012009-09-21 17:02:03 -0700341 mm_slot->mm = mm;
Sasha Levin4ca3a692013-02-22 16:32:28 -0800342 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
Izik Eidus31dbd012009-09-21 17:02:03 -0700343}
344
Izik Eidus31dbd012009-09-21 17:02:03 -0700345/*
Hugh Dickinsa913e182009-09-21 17:02:26 -0700346 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
347 * page tables after it has passed through ksm_exit() - which, if necessary,
348 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
349 * a special flag: they can just back out as soon as mm_users goes to zero.
350 * ksm_test_exit() is used throughout to make this test for exit: in some
351 * places for correctness, in some places just to avoid unnecessary work.
352 */
353static inline bool ksm_test_exit(struct mm_struct *mm)
354{
355 return atomic_read(&mm->mm_users) == 0;
356}
357
358/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700359 * We use break_ksm to break COW on a ksm page: it's a stripped down
360 *
Dave Hansend4edcf02016-02-12 13:01:56 -0800361 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
Izik Eidus31dbd012009-09-21 17:02:03 -0700362 * put_page(page);
363 *
364 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
365 * in case the application has unmapped and remapped mm,addr meanwhile.
366 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
367 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
Dave Hansen1b2ee122016-02-12 13:02:21 -0800368 *
369 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
370 * of the process that owns 'vma'. We also do not want to enforce
371 * protection keys here anyway.
Izik Eidus31dbd012009-09-21 17:02:03 -0700372 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700373static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
Izik Eidus31dbd012009-09-21 17:02:03 -0700374{
375 struct page *page;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700376 int ret = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700377
378 do {
379 cond_resched();
Dave Hansen1b2ee122016-02-12 13:02:21 -0800380 page = follow_page(vma, addr,
381 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400382 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700383 break;
384 if (PageKsm(page))
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700385 ret = handle_mm_fault(vma, addr,
386 FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE);
Izik Eidus31dbd012009-09-21 17:02:03 -0700387 else
388 ret = VM_FAULT_WRITE;
389 put_page(page);
Linus Torvalds33692f22015-01-29 10:51:32 -0800390 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
Hugh Dickinsd952b792009-09-21 17:02:16 -0700391 /*
392 * We must loop because handle_mm_fault() may back out if there's
393 * any difficulty e.g. if pte accessed bit gets updated concurrently.
394 *
395 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
396 * COW has been broken, even if the vma does not permit VM_WRITE;
397 * but note that a concurrent fault might break PageKsm for us.
398 *
399 * VM_FAULT_SIGBUS could occur if we race with truncation of the
400 * backing file, which also invalidates anonymous pages: that's
401 * okay, that truncation will have unmapped the PageKsm for us.
402 *
403 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
404 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
405 * current task has TIF_MEMDIE set, and will be OOM killed on return
406 * to user; and ksmd, having no mm, would never be chosen for that.
407 *
408 * But if the mm is in a limited mem_cgroup, then the fault may fail
409 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
410 * even ksmd can fail in this way - though it's usually breaking ksm
411 * just to undo a merge it made a moment before, so unlikely to oom.
412 *
413 * That's a pity: we might therefore have more kernel pages allocated
414 * than we're counting as nodes in the stable tree; but ksm_do_scan
415 * will retry to break_cow on each pass, so should recover the page
416 * in due course. The important thing is to not let VM_MERGEABLE
417 * be cleared while any such pages might remain in the area.
418 */
419 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700420}
421
Bob Liuef694222012-03-21 16:34:11 -0700422static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
423 unsigned long addr)
424{
425 struct vm_area_struct *vma;
426 if (ksm_test_exit(mm))
427 return NULL;
428 vma = find_vma(mm, addr);
429 if (!vma || vma->vm_start > addr)
430 return NULL;
431 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
432 return NULL;
433 return vma;
434}
435
Hugh Dickins8dd35572009-12-14 17:59:18 -0800436static void break_cow(struct rmap_item *rmap_item)
Izik Eidus31dbd012009-09-21 17:02:03 -0700437{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800438 struct mm_struct *mm = rmap_item->mm;
439 unsigned long addr = rmap_item->address;
Izik Eidus31dbd012009-09-21 17:02:03 -0700440 struct vm_area_struct *vma;
441
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800442 /*
443 * It is not an accident that whenever we want to break COW
444 * to undo, we also need to drop a reference to the anon_vma.
445 */
Peter Zijlstra9e601092011-03-22 16:32:46 -0700446 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800447
Hugh Dickins81464e302009-09-21 17:02:15 -0700448 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700449 vma = find_mergeable_vma(mm, addr);
450 if (vma)
451 break_ksm(vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700452 up_read(&mm->mmap_sem);
453}
454
455static struct page *get_mergeable_page(struct rmap_item *rmap_item)
456{
457 struct mm_struct *mm = rmap_item->mm;
458 unsigned long addr = rmap_item->address;
459 struct vm_area_struct *vma;
460 struct page *page;
461
462 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700463 vma = find_mergeable_vma(mm, addr);
464 if (!vma)
Izik Eidus31dbd012009-09-21 17:02:03 -0700465 goto out;
466
467 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400468 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700469 goto out;
Kirill A. Shutemovf765f542016-01-15 16:53:03 -0800470 if (PageAnon(page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700471 flush_anon_page(vma, page, addr);
472 flush_dcache_page(page);
473 } else {
474 put_page(page);
Andrea Arcangelic8f95ed2015-11-05 18:49:19 -0800475out:
476 page = NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700477 }
478 up_read(&mm->mmap_sem);
479 return page;
480}
481
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800482/*
483 * This helper is used for getting right index into array of tree roots.
484 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
485 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
486 * every node has its own stable and unstable tree.
487 */
488static inline int get_kpfn_nid(unsigned long kpfn)
489{
Hugh Dickinsd8fc16a2013-03-08 12:43:34 -0800490 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800491}
492
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800493static void remove_node_from_stable_tree(struct stable_node *stable_node)
494{
495 struct rmap_item *rmap_item;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800496
Sasha Levinb67bfe02013-02-27 17:06:00 -0800497 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800498 if (rmap_item->hlist.next)
499 ksm_pages_sharing--;
500 else
501 ksm_pages_shared--;
Peter Zijlstra9e601092011-03-22 16:32:46 -0700502 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800503 rmap_item->address &= PAGE_MASK;
504 cond_resched();
505 }
506
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800507 if (stable_node->head == &migrate_nodes)
508 list_del(&stable_node->list);
509 else
510 rb_erase(&stable_node->node,
Hugh Dickinsef53d162013-02-22 16:36:12 -0800511 root_stable_tree + NUMA(stable_node->nid));
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800512 free_stable_node(stable_node);
513}
514
515/*
516 * get_ksm_page: checks if the page indicated by the stable node
517 * is still its ksm page, despite having held no reference to it.
518 * In which case we can trust the content of the page, and it
519 * returns the gotten page; but if the page has now been zapped,
520 * remove the stale node from the stable tree and return NULL.
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800521 * But beware, the stable node's page might be being migrated.
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800522 *
523 * You would expect the stable_node to hold a reference to the ksm page.
524 * But if it increments the page's count, swapping out has to wait for
525 * ksmd to come around again before it can free the page, which may take
526 * seconds or even minutes: much too unresponsive. So instead we use a
527 * "keyhole reference": access to the ksm page from the stable node peeps
528 * out through its keyhole to see if that page still holds the right key,
529 * pointing back to this stable node. This relies on freeing a PageAnon
530 * page to reset its page->mapping to NULL, and relies on no other use of
531 * a page to put something that might look like our key in page->mapping.
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800532 * is on its way to being freed; but it is an anomaly to bear in mind.
533 */
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800534static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800535{
536 struct page *page;
537 void *expected_mapping;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800538 unsigned long kpfn;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800539
Minchan Kimbda807d2016-07-26 15:23:05 -0700540 expected_mapping = (void *)((unsigned long)stable_node |
541 PAGE_MAPPING_KSM);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800542again:
Jason Low4db0c3c2015-04-15 16:14:08 -0700543 kpfn = READ_ONCE(stable_node->kpfn);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800544 page = pfn_to_page(kpfn);
545
546 /*
547 * page is computed from kpfn, so on most architectures reading
548 * page->mapping is naturally ordered after reading node->kpfn,
549 * but on Alpha we need to be more careful.
550 */
551 smp_read_barrier_depends();
Jason Low4db0c3c2015-04-15 16:14:08 -0700552 if (READ_ONCE(page->mapping) != expected_mapping)
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800553 goto stale;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800554
555 /*
556 * We cannot do anything with the page while its refcount is 0.
557 * Usually 0 means free, or tail of a higher-order page: in which
558 * case this node is no longer referenced, and should be freed;
559 * however, it might mean that the page is under page_freeze_refs().
560 * The __remove_mapping() case is easy, again the node is now stale;
561 * but if page is swapcache in migrate_page_move_mapping(), it might
562 * still be our page, in which case it's essential to keep the node.
563 */
564 while (!get_page_unless_zero(page)) {
565 /*
566 * Another check for page->mapping != expected_mapping would
567 * work here too. We have chosen the !PageSwapCache test to
568 * optimize the common case, when the page is or is about to
569 * be freed: PageSwapCache is cleared (under spin_lock_irq)
570 * in the freeze_refs section of __remove_mapping(); but Anon
571 * page->mapping reset to NULL later, in free_pages_prepare().
572 */
573 if (!PageSwapCache(page))
574 goto stale;
575 cpu_relax();
576 }
577
Jason Low4db0c3c2015-04-15 16:14:08 -0700578 if (READ_ONCE(page->mapping) != expected_mapping) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800579 put_page(page);
580 goto stale;
581 }
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800582
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800583 if (lock_it) {
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800584 lock_page(page);
Jason Low4db0c3c2015-04-15 16:14:08 -0700585 if (READ_ONCE(page->mapping) != expected_mapping) {
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800586 unlock_page(page);
587 put_page(page);
588 goto stale;
589 }
590 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800591 return page;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800592
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800593stale:
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800594 /*
595 * We come here from above when page->mapping or !PageSwapCache
596 * suggests that the node is stale; but it might be under migration.
597 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
598 * before checking whether node->kpfn has been changed.
599 */
600 smp_rmb();
Jason Low4db0c3c2015-04-15 16:14:08 -0700601 if (READ_ONCE(stable_node->kpfn) != kpfn)
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800602 goto again;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800603 remove_node_from_stable_tree(stable_node);
604 return NULL;
605}
606
Izik Eidus31dbd012009-09-21 17:02:03 -0700607/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700608 * Removing rmap_item from stable or unstable tree.
609 * This function will clean the information from the stable/unstable tree.
610 */
611static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
612{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800613 if (rmap_item->address & STABLE_FLAG) {
614 struct stable_node *stable_node;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800615 struct page *page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700616
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800617 stable_node = rmap_item->head;
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800618 page = get_ksm_page(stable_node, true);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800619 if (!page)
620 goto out;
621
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800622 hlist_del(&rmap_item->hlist);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800623 unlock_page(page);
624 put_page(page);
Hugh Dickins08beca42009-12-14 17:59:21 -0800625
Andrea Arcangeli98666f8a2015-11-05 18:49:13 -0800626 if (!hlist_empty(&stable_node->hlist))
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800627 ksm_pages_sharing--;
628 else
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800629 ksm_pages_shared--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700630
Peter Zijlstra9e601092011-03-22 16:32:46 -0700631 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins93d17712009-12-14 17:59:16 -0800632 rmap_item->address &= PAGE_MASK;
Izik Eidus31dbd012009-09-21 17:02:03 -0700633
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800634 } else if (rmap_item->address & UNSTABLE_FLAG) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700635 unsigned char age;
636 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -0700637 * Usually ksmd can and must skip the rb_erase, because
Izik Eidus31dbd012009-09-21 17:02:03 -0700638 * root_unstable_tree was already reset to RB_ROOT.
Hugh Dickins9ba69292009-09-21 17:02:20 -0700639 * But be careful when an mm is exiting: do the rb_erase
640 * if this rmap_item was inserted by this scan, rather
641 * than left over from before.
Izik Eidus31dbd012009-09-21 17:02:03 -0700642 */
643 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
Hugh Dickinscd551f92009-09-21 17:02:17 -0700644 BUG_ON(age > 1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700645 if (!age)
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800646 rb_erase(&rmap_item->node,
Hugh Dickinsef53d162013-02-22 16:36:12 -0800647 root_unstable_tree + NUMA(rmap_item->nid));
Hugh Dickins93d17712009-12-14 17:59:16 -0800648 ksm_pages_unshared--;
649 rmap_item->address &= PAGE_MASK;
650 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800651out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700652 cond_resched(); /* we're called from many long loops */
653}
654
Izik Eidus31dbd012009-09-21 17:02:03 -0700655static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -0800656 struct rmap_item **rmap_list)
Izik Eidus31dbd012009-09-21 17:02:03 -0700657{
Hugh Dickins6514d512009-12-14 17:59:19 -0800658 while (*rmap_list) {
659 struct rmap_item *rmap_item = *rmap_list;
660 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700661 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700662 free_rmap_item(rmap_item);
663 }
664}
665
666/*
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800667 * Though it's very tempting to unmerge rmap_items from stable tree rather
Izik Eidus31dbd012009-09-21 17:02:03 -0700668 * than check every pte of a given vma, the locking doesn't quite work for
669 * that - an rmap_item is assigned to the stable tree after inserting ksm
670 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
671 * rmap_items from parent to child at fork time (so as not to waste time
672 * if exit comes before the next scan reaches it).
Hugh Dickins81464e302009-09-21 17:02:15 -0700673 *
674 * Similarly, although we'd like to remove rmap_items (so updating counts
675 * and freeing memory) when unmerging an area, it's easier to leave that
676 * to the next pass of ksmd - consider, for example, how ksmd might be
677 * in cmp_and_merge_page on one of the rmap_items we would be removing.
Izik Eidus31dbd012009-09-21 17:02:03 -0700678 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700679static int unmerge_ksm_pages(struct vm_area_struct *vma,
680 unsigned long start, unsigned long end)
Izik Eidus31dbd012009-09-21 17:02:03 -0700681{
682 unsigned long addr;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700683 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700684
Hugh Dickinsd952b792009-09-21 17:02:16 -0700685 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700686 if (ksm_test_exit(vma->vm_mm))
687 break;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700688 if (signal_pending(current))
689 err = -ERESTARTSYS;
690 else
691 err = break_ksm(vma, addr);
692 }
693 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700694}
695
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700696#ifdef CONFIG_SYSFS
697/*
698 * Only called through the sysfs control interface:
699 */
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800700static int remove_stable_node(struct stable_node *stable_node)
701{
702 struct page *page;
703 int err;
704
705 page = get_ksm_page(stable_node, true);
706 if (!page) {
707 /*
708 * get_ksm_page did remove_node_from_stable_tree itself.
709 */
710 return 0;
711 }
712
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800713 if (WARN_ON_ONCE(page_mapped(page))) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800714 /*
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800715 * This should not happen: but if it does, just refuse to let
716 * merge_across_nodes be switched - there is no need to panic.
717 */
718 err = -EBUSY;
719 } else {
720 /*
721 * The stable node did not yet appear stale to get_ksm_page(),
722 * since that allows for an unmapped ksm page to be recognized
723 * right up until it is freed; but the node is safe to remove.
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800724 * This page might be in a pagevec waiting to be freed,
725 * or it might be PageSwapCache (perhaps under writeback),
726 * or it might have been removed from swapcache a moment ago.
727 */
728 set_page_stable_node(page, NULL);
729 remove_node_from_stable_tree(stable_node);
730 err = 0;
731 }
732
733 unlock_page(page);
734 put_page(page);
735 return err;
736}
737
738static int remove_all_stable_nodes(void)
739{
Geliang Tang03640412016-01-14 15:20:54 -0800740 struct stable_node *stable_node, *next;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800741 int nid;
742 int err = 0;
743
Hugh Dickinsef53d162013-02-22 16:36:12 -0800744 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800745 while (root_stable_tree[nid].rb_node) {
746 stable_node = rb_entry(root_stable_tree[nid].rb_node,
747 struct stable_node, node);
748 if (remove_stable_node(stable_node)) {
749 err = -EBUSY;
750 break; /* proceed to next nid */
751 }
752 cond_resched();
753 }
754 }
Geliang Tang03640412016-01-14 15:20:54 -0800755 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800756 if (remove_stable_node(stable_node))
757 err = -EBUSY;
758 cond_resched();
759 }
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800760 return err;
761}
762
Hugh Dickinsd952b792009-09-21 17:02:16 -0700763static int unmerge_and_remove_all_rmap_items(void)
Izik Eidus31dbd012009-09-21 17:02:03 -0700764{
765 struct mm_slot *mm_slot;
766 struct mm_struct *mm;
767 struct vm_area_struct *vma;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700768 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700769
Hugh Dickinsd952b792009-09-21 17:02:16 -0700770 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700771 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700772 struct mm_slot, mm_list);
773 spin_unlock(&ksm_mmlist_lock);
774
Hugh Dickins9ba69292009-09-21 17:02:20 -0700775 for (mm_slot = ksm_scan.mm_slot;
776 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700777 mm = mm_slot->mm;
778 down_read(&mm->mmap_sem);
779 for (vma = mm->mmap; vma; vma = vma->vm_next) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700780 if (ksm_test_exit(mm))
781 break;
Izik Eidus31dbd012009-09-21 17:02:03 -0700782 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
783 continue;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700784 err = unmerge_ksm_pages(vma,
785 vma->vm_start, vma->vm_end);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700786 if (err)
787 goto error;
Izik Eidus31dbd012009-09-21 17:02:03 -0700788 }
Hugh Dickins9ba69292009-09-21 17:02:20 -0700789
Hugh Dickins6514d512009-12-14 17:59:19 -0800790 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
Zhou Chengming7496fea2016-05-12 15:42:21 -0700791 up_read(&mm->mmap_sem);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700792
793 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700794 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700795 struct mm_slot, mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700796 if (ksm_test_exit(mm)) {
Sasha Levin4ca3a692013-02-22 16:32:28 -0800797 hash_del(&mm_slot->link);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700798 list_del(&mm_slot->mm_list);
799 spin_unlock(&ksm_mmlist_lock);
800
801 free_mm_slot(mm_slot);
802 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700803 mmdrop(mm);
Zhou Chengming7496fea2016-05-12 15:42:21 -0700804 } else
Hugh Dickins9ba69292009-09-21 17:02:20 -0700805 spin_unlock(&ksm_mmlist_lock);
Izik Eidus31dbd012009-09-21 17:02:03 -0700806 }
807
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800808 /* Clean up stable nodes, but don't worry if some are still busy */
809 remove_all_stable_nodes();
Hugh Dickinsd952b792009-09-21 17:02:16 -0700810 ksm_scan.seqnr = 0;
Hugh Dickins9ba69292009-09-21 17:02:20 -0700811 return 0;
812
813error:
814 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700815 spin_lock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700816 ksm_scan.mm_slot = &ksm_mm_head;
Izik Eidus31dbd012009-09-21 17:02:03 -0700817 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700818 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700819}
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700820#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -0700821
Izik Eidus31dbd012009-09-21 17:02:03 -0700822static u32 calc_checksum(struct page *page)
823{
824 u32 checksum;
Cong Wang9b04c5f2011-11-25 23:14:39 +0800825 void *addr = kmap_atomic(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700826 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
Cong Wang9b04c5f2011-11-25 23:14:39 +0800827 kunmap_atomic(addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700828 return checksum;
829}
830
831static int memcmp_pages(struct page *page1, struct page *page2)
832{
833 char *addr1, *addr2;
834 int ret;
835
Cong Wang9b04c5f2011-11-25 23:14:39 +0800836 addr1 = kmap_atomic(page1);
837 addr2 = kmap_atomic(page2);
Izik Eidus31dbd012009-09-21 17:02:03 -0700838 ret = memcmp(addr1, addr2, PAGE_SIZE);
Cong Wang9b04c5f2011-11-25 23:14:39 +0800839 kunmap_atomic(addr2);
840 kunmap_atomic(addr1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700841 return ret;
842}
843
844static inline int pages_identical(struct page *page1, struct page *page2)
845{
846 return !memcmp_pages(page1, page2);
847}
848
849static int write_protect_page(struct vm_area_struct *vma, struct page *page,
850 pte_t *orig_pte)
851{
852 struct mm_struct *mm = vma->vm_mm;
853 unsigned long addr;
854 pte_t *ptep;
855 spinlock_t *ptl;
856 int swapped;
857 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700858 unsigned long mmun_start; /* For mmu_notifiers */
859 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -0700860
861 addr = page_address_in_vma(page, vma);
862 if (addr == -EFAULT)
863 goto out;
864
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800865 BUG_ON(PageTransCompound(page));
Haggai Eran6bdb9132012-10-08 16:33:35 -0700866
867 mmun_start = addr;
868 mmun_end = addr + PAGE_SIZE;
869 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
870
Izik Eidus31dbd012009-09-21 17:02:03 -0700871 ptep = page_check_address(page, mm, addr, &ptl, 0);
872 if (!ptep)
Haggai Eran6bdb9132012-10-08 16:33:35 -0700873 goto out_mn;
Izik Eidus31dbd012009-09-21 17:02:03 -0700874
Hugh Dickins4e316352010-10-02 17:49:08 -0700875 if (pte_write(*ptep) || pte_dirty(*ptep)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700876 pte_t entry;
877
878 swapped = PageSwapCache(page);
879 flush_cache_page(vma, addr, page_to_pfn(page));
880 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300881 * Ok this is tricky, when get_user_pages_fast() run it doesn't
Izik Eidus31dbd012009-09-21 17:02:03 -0700882 * take any lock, therefore the check that we are going to make
883 * with the pagecount against the mapcount is racey and
884 * O_DIRECT can happen right after the check.
885 * So we clear the pte and flush the tlb before the check
886 * this assure us that no O_DIRECT can happen after the check
887 * or in the middle of the check.
888 */
Joerg Roedel34ee6452014-11-13 13:46:09 +1100889 entry = ptep_clear_flush_notify(vma, addr, ptep);
Izik Eidus31dbd012009-09-21 17:02:03 -0700890 /*
891 * Check that no O_DIRECT or similar I/O is in progress on the
892 * page
893 */
Hugh Dickins31e855e2009-12-14 17:59:17 -0800894 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
Robin Holtcb532372010-03-23 13:35:26 -0700895 set_pte_at(mm, addr, ptep, entry);
Izik Eidus31dbd012009-09-21 17:02:03 -0700896 goto out_unlock;
897 }
Hugh Dickins4e316352010-10-02 17:49:08 -0700898 if (pte_dirty(entry))
899 set_page_dirty(page);
900 entry = pte_mkclean(pte_wrprotect(entry));
Izik Eidus31dbd012009-09-21 17:02:03 -0700901 set_pte_at_notify(mm, addr, ptep, entry);
902 }
903 *orig_pte = *ptep;
904 err = 0;
905
906out_unlock:
907 pte_unmap_unlock(ptep, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -0700908out_mn:
909 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -0700910out:
911 return err;
912}
913
914/**
915 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -0800916 * @vma: vma that holds the pte pointing to page
917 * @page: the page we are replacing by kpage
918 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -0700919 * @orig_pte: the original value of the pte
920 *
921 * Returns 0 on success, -EFAULT on failure.
922 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800923static int replace_page(struct vm_area_struct *vma, struct page *page,
924 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -0700925{
926 struct mm_struct *mm = vma->vm_mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700927 pmd_t *pmd;
928 pte_t *ptep;
929 spinlock_t *ptl;
930 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -0700931 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700932 unsigned long mmun_start; /* For mmu_notifiers */
933 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -0700934
Hugh Dickins8dd35572009-12-14 17:59:18 -0800935 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -0700936 if (addr == -EFAULT)
937 goto out;
938
Bob Liu62190492012-12-11 16:00:37 -0800939 pmd = mm_find_pmd(mm, addr);
940 if (!pmd)
Izik Eidus31dbd012009-09-21 17:02:03 -0700941 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700942
Haggai Eran6bdb9132012-10-08 16:33:35 -0700943 mmun_start = addr;
944 mmun_end = addr + PAGE_SIZE;
945 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
946
Izik Eidus31dbd012009-09-21 17:02:03 -0700947 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
948 if (!pte_same(*ptep, orig_pte)) {
949 pte_unmap_unlock(ptep, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -0700950 goto out_mn;
Izik Eidus31dbd012009-09-21 17:02:03 -0700951 }
952
Hugh Dickins8dd35572009-12-14 17:59:18 -0800953 get_page(kpage);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800954 page_add_anon_rmap(kpage, vma, addr, false);
Izik Eidus31dbd012009-09-21 17:02:03 -0700955
956 flush_cache_page(vma, addr, pte_pfn(*ptep));
Joerg Roedel34ee6452014-11-13 13:46:09 +1100957 ptep_clear_flush_notify(vma, addr, ptep);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800958 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
Izik Eidus31dbd012009-09-21 17:02:03 -0700959
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800960 page_remove_rmap(page, false);
Hugh Dickinsae52a2a2011-01-13 15:46:28 -0800961 if (!page_mapped(page))
962 try_to_free_swap(page);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800963 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700964
965 pte_unmap_unlock(ptep, ptl);
966 err = 0;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700967out_mn:
968 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -0700969out:
970 return err;
971}
972
973/*
974 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -0800975 * @vma: the vma that holds the pte pointing to page
976 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins80e148222009-12-14 17:59:29 -0800977 * @kpage: the PageKsm page that we want to map instead of page,
978 * or NULL the first time when we want to use page as kpage.
Izik Eidus31dbd012009-09-21 17:02:03 -0700979 *
980 * This function returns 0 if the pages were merged, -EFAULT otherwise.
981 */
982static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -0800983 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700984{
985 pte_t orig_pte = __pte(0);
986 int err = -EFAULT;
987
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800988 if (page == kpage) /* ksm page forked */
989 return 0;
990
Hugh Dickins8dd35572009-12-14 17:59:18 -0800991 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700992 goto out;
993
Izik Eidus31dbd012009-09-21 17:02:03 -0700994 /*
995 * We need the page lock to read a stable PageSwapCache in
996 * write_protect_page(). We use trylock_page() instead of
997 * lock_page() because we don't want to wait here - we
998 * prefer to continue scanning and merging different pages,
999 * then come back to this page when it is unlocked.
1000 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001001 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -08001002 goto out;
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001003
1004 if (PageTransCompound(page)) {
1005 err = split_huge_page(page);
1006 if (err)
1007 goto out_unlock;
1008 }
1009
Izik Eidus31dbd012009-09-21 17:02:03 -07001010 /*
1011 * If this anonymous page is mapped only here, its pte may need
1012 * to be write-protected. If it's mapped elsewhere, all of its
1013 * ptes are necessarily already write-protected. But in either
1014 * case, we need to lock and check page_count is not raised.
1015 */
Hugh Dickins80e148222009-12-14 17:59:29 -08001016 if (write_protect_page(vma, page, &orig_pte) == 0) {
1017 if (!kpage) {
1018 /*
1019 * While we hold page lock, upgrade page from
1020 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1021 * stable_tree_insert() will update stable_node.
1022 */
1023 set_page_stable_node(page, NULL);
1024 mark_page_accessed(page);
Minchan Kim337ed7e2016-01-15 16:55:15 -08001025 /*
1026 * Page reclaim just frees a clean page with no dirty
1027 * ptes: make sure that the ksm page would be swapped.
1028 */
1029 if (!PageDirty(page))
1030 SetPageDirty(page);
Hugh Dickins80e148222009-12-14 17:59:29 -08001031 err = 0;
1032 } else if (pages_identical(page, kpage))
1033 err = replace_page(vma, page, kpage, orig_pte);
1034 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001035
Hugh Dickins80e148222009-12-14 17:59:29 -08001036 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
Hugh Dickins73848b42009-12-14 17:59:22 -08001037 munlock_vma_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001038 if (!PageMlocked(kpage)) {
1039 unlock_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001040 lock_page(kpage);
1041 mlock_vma_page(kpage);
1042 page = kpage; /* for final unlock */
1043 }
1044 }
Hugh Dickins73848b42009-12-14 17:59:22 -08001045
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001046out_unlock:
Hugh Dickins8dd35572009-12-14 17:59:18 -08001047 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001048out:
1049 return err;
1050}
1051
1052/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001053 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1054 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -08001055 *
1056 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -07001057 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001058static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1059 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -07001060{
Hugh Dickins8dd35572009-12-14 17:59:18 -08001061 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -07001062 struct vm_area_struct *vma;
1063 int err = -EFAULT;
1064
Hugh Dickins8dd35572009-12-14 17:59:18 -08001065 down_read(&mm->mmap_sem);
Andrea Arcangeli85c6e8d2015-11-05 18:49:16 -08001066 vma = find_mergeable_vma(mm, rmap_item->address);
1067 if (!vma)
Hugh Dickins9ba69292009-09-21 17:02:20 -07001068 goto out;
1069
Hugh Dickins8dd35572009-12-14 17:59:18 -08001070 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001071 if (err)
1072 goto out;
1073
Hugh Dickinsbc566202013-02-22 16:36:06 -08001074 /* Unstable nid is in union with stable anon_vma: remove first */
1075 remove_rmap_item_from_tree(rmap_item);
1076
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001077 /* Must get reference to anon_vma while still holding mmap_sem */
Peter Zijlstra9e601092011-03-22 16:32:46 -07001078 rmap_item->anon_vma = vma->anon_vma;
1079 get_anon_vma(vma->anon_vma);
Hugh Dickins81464e302009-09-21 17:02:15 -07001080out:
Hugh Dickins8dd35572009-12-14 17:59:18 -08001081 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -07001082 return err;
1083}
1084
1085/*
Izik Eidus31dbd012009-09-21 17:02:03 -07001086 * try_to_merge_two_pages - take two identical pages and prepare them
1087 * to be merged into one page.
1088 *
Hugh Dickins8dd35572009-12-14 17:59:18 -08001089 * This function returns the kpage if we successfully merged two identical
1090 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001091 *
Hugh Dickins80e148222009-12-14 17:59:29 -08001092 * Note that this function upgrades page to ksm page: if one of the pages
Izik Eidus31dbd012009-09-21 17:02:03 -07001093 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1094 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001095static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1096 struct page *page,
1097 struct rmap_item *tree_rmap_item,
1098 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001099{
Hugh Dickins80e148222009-12-14 17:59:29 -08001100 int err;
Izik Eidus31dbd012009-09-21 17:02:03 -07001101
Hugh Dickins80e148222009-12-14 17:59:29 -08001102 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
Izik Eidus31dbd012009-09-21 17:02:03 -07001103 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001104 err = try_to_merge_with_ksm_page(tree_rmap_item,
Hugh Dickins80e148222009-12-14 17:59:29 -08001105 tree_page, page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001106 /*
Hugh Dickins81464e302009-09-21 17:02:15 -07001107 * If that fails, we have a ksm page with only one pte
1108 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -07001109 */
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001110 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -08001111 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001112 }
Hugh Dickins80e148222009-12-14 17:59:29 -08001113 return err ? NULL : page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001114}
1115
1116/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001117 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -07001118 *
1119 * This function checks if there is a page inside the stable tree
1120 * with identical content to the page that we are scanning right now.
1121 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001122 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -07001123 * NULL otherwise.
1124 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001125static struct page *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001126{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001127 int nid;
Hugh Dickinsef53d162013-02-22 16:36:12 -08001128 struct rb_root *root;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001129 struct rb_node **new;
1130 struct rb_node *parent;
1131 struct stable_node *stable_node;
1132 struct stable_node *page_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001133
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001134 page_node = page_stable_node(page);
1135 if (page_node && page_node->head != &migrate_nodes) {
1136 /* ksm page forked */
Hugh Dickins08beca42009-12-14 17:59:21 -08001137 get_page(page);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001138 return page;
Hugh Dickins08beca42009-12-14 17:59:21 -08001139 }
1140
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001141 nid = get_kpfn_nid(page_to_pfn(page));
Hugh Dickinsef53d162013-02-22 16:36:12 -08001142 root = root_stable_tree + nid;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001143again:
Hugh Dickinsef53d162013-02-22 16:36:12 -08001144 new = &root->rb_node;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001145 parent = NULL;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001146
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001147 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001148 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001149 int ret;
1150
Hugh Dickins08beca42009-12-14 17:59:21 -08001151 cond_resched();
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001152 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins8aafa6a2013-02-22 16:35:06 -08001153 tree_page = get_ksm_page(stable_node, false);
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001154 if (!tree_page) {
1155 /*
1156 * If we walked over a stale stable_node,
1157 * get_ksm_page() will call rb_erase() and it
1158 * may rebalance the tree from under us. So
1159 * restart the search from scratch. Returning
1160 * NULL would be safe too, but we'd generate
1161 * false negative insertions just because some
1162 * stable_node was stale.
1163 */
1164 goto again;
1165 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001166
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001167 ret = memcmp_pages(page, tree_page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001168 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001169
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001170 parent = *new;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001171 if (ret < 0)
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001172 new = &parent->rb_left;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001173 else if (ret > 0)
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001174 new = &parent->rb_right;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001175 else {
1176 /*
1177 * Lock and unlock the stable_node's page (which
1178 * might already have been migrated) so that page
1179 * migration is sure to notice its raised count.
1180 * It would be more elegant to return stable_node
1181 * than kpage, but that involves more changes.
1182 */
1183 tree_page = get_ksm_page(stable_node, true);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001184 if (tree_page) {
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001185 unlock_page(tree_page);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001186 if (get_kpfn_nid(stable_node->kpfn) !=
1187 NUMA(stable_node->nid)) {
1188 put_page(tree_page);
1189 goto replace;
1190 }
1191 return tree_page;
1192 }
1193 /*
1194 * There is now a place for page_node, but the tree may
1195 * have been rebalanced, so re-evaluate parent and new.
1196 */
1197 if (page_node)
1198 goto again;
1199 return NULL;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001200 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001201 }
1202
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001203 if (!page_node)
1204 return NULL;
1205
1206 list_del(&page_node->list);
1207 DO_NUMA(page_node->nid = nid);
1208 rb_link_node(&page_node->node, parent, new);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001209 rb_insert_color(&page_node->node, root);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001210 get_page(page);
1211 return page;
1212
1213replace:
1214 if (page_node) {
1215 list_del(&page_node->list);
1216 DO_NUMA(page_node->nid = nid);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001217 rb_replace_node(&stable_node->node, &page_node->node, root);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001218 get_page(page);
1219 } else {
Hugh Dickinsef53d162013-02-22 16:36:12 -08001220 rb_erase(&stable_node->node, root);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001221 page = NULL;
1222 }
1223 stable_node->head = &migrate_nodes;
1224 list_add(&stable_node->list, stable_node->head);
1225 return page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001226}
1227
1228/*
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001229 * stable_tree_insert - insert stable tree node pointing to new ksm page
Izik Eidus31dbd012009-09-21 17:02:03 -07001230 * into the stable tree.
1231 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001232 * This function returns the stable tree node just allocated on success,
1233 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001234 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001235static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001236{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001237 int nid;
1238 unsigned long kpfn;
Hugh Dickinsef53d162013-02-22 16:36:12 -08001239 struct rb_root *root;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001240 struct rb_node **new;
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001241 struct rb_node *parent;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001242 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001243
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001244 kpfn = page_to_pfn(kpage);
1245 nid = get_kpfn_nid(kpfn);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001246 root = root_stable_tree + nid;
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001247again:
1248 parent = NULL;
Hugh Dickinsef53d162013-02-22 16:36:12 -08001249 new = &root->rb_node;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001250
Izik Eidus31dbd012009-09-21 17:02:03 -07001251 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001252 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001253 int ret;
1254
Hugh Dickins08beca42009-12-14 17:59:21 -08001255 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001256 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins8aafa6a2013-02-22 16:35:06 -08001257 tree_page = get_ksm_page(stable_node, false);
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001258 if (!tree_page) {
1259 /*
1260 * If we walked over a stale stable_node,
1261 * get_ksm_page() will call rb_erase() and it
1262 * may rebalance the tree from under us. So
1263 * restart the search from scratch. Returning
1264 * NULL would be safe too, but we'd generate
1265 * false negative insertions just because some
1266 * stable_node was stale.
1267 */
1268 goto again;
1269 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001270
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001271 ret = memcmp_pages(kpage, tree_page);
1272 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001273
1274 parent = *new;
1275 if (ret < 0)
1276 new = &parent->rb_left;
1277 else if (ret > 0)
1278 new = &parent->rb_right;
1279 else {
1280 /*
1281 * It is not a bug that stable_tree_search() didn't
1282 * find this node: because at that time our page was
1283 * not yet write-protected, so may have changed since.
1284 */
1285 return NULL;
1286 }
1287 }
1288
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001289 stable_node = alloc_stable_node();
1290 if (!stable_node)
1291 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001292
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001293 INIT_HLIST_HEAD(&stable_node->hlist);
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001294 stable_node->kpfn = kpfn;
Hugh Dickins08beca42009-12-14 17:59:21 -08001295 set_page_stable_node(kpage, stable_node);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001296 DO_NUMA(stable_node->nid = nid);
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001297 rb_link_node(&stable_node->node, parent, new);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001298 rb_insert_color(&stable_node->node, root);
Hugh Dickins08beca42009-12-14 17:59:21 -08001299
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001300 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001301}
1302
1303/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001304 * unstable_tree_search_insert - search for identical page,
1305 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001306 *
1307 * This function searches for a page in the unstable tree identical to the
1308 * page currently being scanned; and if no identical page is found in the
1309 * tree, we insert rmap_item as a new object into the unstable tree.
1310 *
1311 * This function returns pointer to rmap_item found to be identical
1312 * to the currently scanned page, NULL otherwise.
1313 *
1314 * This function does both searching and inserting, because they share
1315 * the same walking algorithm in an rbtree.
1316 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001317static
1318struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1319 struct page *page,
1320 struct page **tree_pagep)
Izik Eidus31dbd012009-09-21 17:02:03 -07001321{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001322 struct rb_node **new;
1323 struct rb_root *root;
Izik Eidus31dbd012009-09-21 17:02:03 -07001324 struct rb_node *parent = NULL;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001325 int nid;
1326
1327 nid = get_kpfn_nid(page_to_pfn(page));
Hugh Dickinsef53d162013-02-22 16:36:12 -08001328 root = root_unstable_tree + nid;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001329 new = &root->rb_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001330
1331 while (*new) {
1332 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001333 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001334 int ret;
1335
Hugh Dickinsd178f272009-11-09 15:58:23 +00001336 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -07001337 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001338 tree_page = get_mergeable_page(tree_rmap_item);
Andrea Arcangelic8f95ed2015-11-05 18:49:19 -08001339 if (!tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001340 return NULL;
1341
1342 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001343 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -07001344 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001345 if (page == tree_page) {
1346 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001347 return NULL;
1348 }
1349
Hugh Dickins8dd35572009-12-14 17:59:18 -08001350 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001351
1352 parent = *new;
1353 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001354 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001355 new = &parent->rb_left;
1356 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001357 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001358 new = &parent->rb_right;
Hugh Dickinsb599cbd2013-02-22 16:36:05 -08001359 } else if (!ksm_merge_across_nodes &&
1360 page_to_nid(tree_page) != nid) {
1361 /*
1362 * If tree_page has been migrated to another NUMA node,
1363 * it will be flushed out and put in the right unstable
1364 * tree next time: only merge with it when across_nodes.
1365 */
1366 put_page(tree_page);
1367 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001368 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001369 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001370 return tree_rmap_item;
1371 }
1372 }
1373
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001374 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001375 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001376 DO_NUMA(rmap_item->nid = nid);
Izik Eidus31dbd012009-09-21 17:02:03 -07001377 rb_link_node(&rmap_item->node, parent, new);
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001378 rb_insert_color(&rmap_item->node, root);
Izik Eidus31dbd012009-09-21 17:02:03 -07001379
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001380 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001381 return NULL;
1382}
1383
1384/*
1385 * stable_tree_append - add another rmap_item to the linked list of
1386 * rmap_items hanging off a given node of the stable tree, all sharing
1387 * the same ksm page.
1388 */
1389static void stable_tree_append(struct rmap_item *rmap_item,
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001390 struct stable_node *stable_node)
Izik Eidus31dbd012009-09-21 17:02:03 -07001391{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001392 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001393 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001394 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001395
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001396 if (rmap_item->hlist.next)
1397 ksm_pages_sharing++;
1398 else
1399 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001400}
1401
1402/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001403 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1404 * if not, compare checksum to previous and if it's the same, see if page can
1405 * be inserted into the unstable tree, or merged with a page already there and
1406 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001407 *
1408 * @page: the page that we are searching identical page to.
1409 * @rmap_item: the reverse mapping into the virtual address of this page
1410 */
1411static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1412{
Izik Eidus31dbd012009-09-21 17:02:03 -07001413 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001414 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001415 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001416 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07001417 unsigned int checksum;
1418 int err;
1419
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001420 stable_node = page_stable_node(page);
1421 if (stable_node) {
1422 if (stable_node->head != &migrate_nodes &&
1423 get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1424 rb_erase(&stable_node->node,
Hugh Dickinsef53d162013-02-22 16:36:12 -08001425 root_stable_tree + NUMA(stable_node->nid));
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001426 stable_node->head = &migrate_nodes;
1427 list_add(&stable_node->list, stable_node->head);
1428 }
1429 if (stable_node->head != &migrate_nodes &&
1430 rmap_item->head == stable_node)
1431 return;
1432 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001433
1434 /* We first start with searching the page inside the stable tree */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001435 kpage = stable_tree_search(page);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001436 if (kpage == page && rmap_item->head == stable_node) {
1437 put_page(kpage);
1438 return;
1439 }
1440
1441 remove_rmap_item_from_tree(rmap_item);
1442
Hugh Dickins62b61f62009-12-14 17:59:33 -08001443 if (kpage) {
Hugh Dickins08beca42009-12-14 17:59:21 -08001444 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001445 if (!err) {
1446 /*
1447 * The page was successfully merged:
1448 * add its rmap_item to the stable tree.
1449 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001450 lock_page(kpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001451 stable_tree_append(rmap_item, page_stable_node(kpage));
Hugh Dickins5ad64682009-12-14 17:59:24 -08001452 unlock_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001453 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001454 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001455 return;
1456 }
1457
1458 /*
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001459 * If the hash value of the page has changed from the last time
1460 * we calculated it, this page is changing frequently: therefore we
1461 * don't want to insert it in the unstable tree, and we don't want
1462 * to waste our time searching for something identical to it there.
Izik Eidus31dbd012009-09-21 17:02:03 -07001463 */
1464 checksum = calc_checksum(page);
1465 if (rmap_item->oldchecksum != checksum) {
1466 rmap_item->oldchecksum = checksum;
1467 return;
1468 }
1469
Hugh Dickins8dd35572009-12-14 17:59:18 -08001470 tree_rmap_item =
1471 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001472 if (tree_rmap_item) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001473 kpage = try_to_merge_two_pages(rmap_item, page,
1474 tree_rmap_item, tree_page);
1475 put_page(tree_page);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001476 if (kpage) {
Hugh Dickinsbc566202013-02-22 16:36:06 -08001477 /*
1478 * The pages were successfully merged: insert new
1479 * node in the stable tree and add both rmap_items.
1480 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001481 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001482 stable_node = stable_tree_insert(kpage);
1483 if (stable_node) {
1484 stable_tree_append(tree_rmap_item, stable_node);
1485 stable_tree_append(rmap_item, stable_node);
1486 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001487 unlock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001488
Izik Eidus31dbd012009-09-21 17:02:03 -07001489 /*
1490 * If we fail to insert the page into the stable tree,
1491 * we will have 2 virtual addresses that are pointing
1492 * to a ksm page left outside the stable tree,
1493 * in which case we need to break_cow on both.
1494 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001495 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001496 break_cow(tree_rmap_item);
1497 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001498 }
1499 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001500 }
1501}
1502
1503static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001504 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07001505 unsigned long addr)
1506{
1507 struct rmap_item *rmap_item;
1508
Hugh Dickins6514d512009-12-14 17:59:19 -08001509 while (*rmap_list) {
1510 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08001511 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07001512 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001513 if (rmap_item->address > addr)
1514 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08001515 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001516 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001517 free_rmap_item(rmap_item);
1518 }
1519
1520 rmap_item = alloc_rmap_item();
1521 if (rmap_item) {
1522 /* It has already been zeroed */
1523 rmap_item->mm = mm_slot->mm;
1524 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08001525 rmap_item->rmap_list = *rmap_list;
1526 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001527 }
1528 return rmap_item;
1529}
1530
1531static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1532{
1533 struct mm_struct *mm;
1534 struct mm_slot *slot;
1535 struct vm_area_struct *vma;
1536 struct rmap_item *rmap_item;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001537 int nid;
Izik Eidus31dbd012009-09-21 17:02:03 -07001538
1539 if (list_empty(&ksm_mm_head.mm_list))
1540 return NULL;
1541
1542 slot = ksm_scan.mm_slot;
1543 if (slot == &ksm_mm_head) {
Hugh Dickins2919bfd2011-01-13 15:47:29 -08001544 /*
1545 * A number of pages can hang around indefinitely on per-cpu
1546 * pagevecs, raised page count preventing write_protect_page
1547 * from merging them. Though it doesn't really matter much,
1548 * it is puzzling to see some stuck in pages_volatile until
1549 * other activity jostles them out, and they also prevented
1550 * LTP's KSM test from succeeding deterministically; so drain
1551 * them here (here rather than on entry to ksm_do_scan(),
1552 * so we don't IPI too often when pages_to_scan is set low).
1553 */
1554 lru_add_drain_all();
1555
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001556 /*
1557 * Whereas stale stable_nodes on the stable_tree itself
1558 * get pruned in the regular course of stable_tree_search(),
1559 * those moved out to the migrate_nodes list can accumulate:
1560 * so prune them once before each full scan.
1561 */
1562 if (!ksm_merge_across_nodes) {
Geliang Tang03640412016-01-14 15:20:54 -08001563 struct stable_node *stable_node, *next;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001564 struct page *page;
1565
Geliang Tang03640412016-01-14 15:20:54 -08001566 list_for_each_entry_safe(stable_node, next,
1567 &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001568 page = get_ksm_page(stable_node, false);
1569 if (page)
1570 put_page(page);
1571 cond_resched();
1572 }
1573 }
1574
Hugh Dickinsef53d162013-02-22 16:36:12 -08001575 for (nid = 0; nid < ksm_nr_node_ids; nid++)
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001576 root_unstable_tree[nid] = RB_ROOT;
Izik Eidus31dbd012009-09-21 17:02:03 -07001577
1578 spin_lock(&ksm_mmlist_lock);
1579 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1580 ksm_scan.mm_slot = slot;
1581 spin_unlock(&ksm_mmlist_lock);
Hugh Dickins2b472612011-06-15 15:08:58 -07001582 /*
1583 * Although we tested list_empty() above, a racing __ksm_exit
1584 * of the last mm on the list may have removed it since then.
1585 */
1586 if (slot == &ksm_mm_head)
1587 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001588next_mm:
1589 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001590 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001591 }
1592
1593 mm = slot->mm;
1594 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001595 if (ksm_test_exit(mm))
1596 vma = NULL;
1597 else
1598 vma = find_vma(mm, ksm_scan.address);
1599
1600 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001601 if (!(vma->vm_flags & VM_MERGEABLE))
1602 continue;
1603 if (ksm_scan.address < vma->vm_start)
1604 ksm_scan.address = vma->vm_start;
1605 if (!vma->anon_vma)
1606 ksm_scan.address = vma->vm_end;
1607
1608 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001609 if (ksm_test_exit(mm))
1610 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001611 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001612 if (IS_ERR_OR_NULL(*page)) {
1613 ksm_scan.address += PAGE_SIZE;
1614 cond_resched();
1615 continue;
1616 }
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001617 if (PageAnon(*page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001618 flush_anon_page(vma, *page, ksm_scan.address);
1619 flush_dcache_page(*page);
1620 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001621 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07001622 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001623 ksm_scan.rmap_list =
1624 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001625 ksm_scan.address += PAGE_SIZE;
1626 } else
1627 put_page(*page);
1628 up_read(&mm->mmap_sem);
1629 return rmap_item;
1630 }
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001631 put_page(*page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001632 ksm_scan.address += PAGE_SIZE;
1633 cond_resched();
1634 }
1635 }
1636
Hugh Dickins9ba69292009-09-21 17:02:20 -07001637 if (ksm_test_exit(mm)) {
1638 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001639 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001640 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001641 /*
1642 * Nuke all the rmap_items that are above this current rmap:
1643 * because there were no VM_MERGEABLE vmas with such addresses.
1644 */
Hugh Dickins6514d512009-12-14 17:59:19 -08001645 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001646
1647 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001648 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1649 struct mm_slot, mm_list);
1650 if (ksm_scan.address == 0) {
1651 /*
1652 * We've completed a full scan of all vmas, holding mmap_sem
1653 * throughout, and found no VM_MERGEABLE: so do the same as
1654 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07001655 * This applies either when cleaning up after __ksm_exit
1656 * (but beware: we can reach here even before __ksm_exit),
1657 * or when all VM_MERGEABLE areas have been unmapped (and
1658 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07001659 */
Sasha Levin4ca3a692013-02-22 16:32:28 -08001660 hash_del(&slot->link);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001661 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001662 spin_unlock(&ksm_mmlist_lock);
1663
Hugh Dickinscd551f92009-09-21 17:02:17 -07001664 free_mm_slot(slot);
1665 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001666 up_read(&mm->mmap_sem);
1667 mmdrop(mm);
1668 } else {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001669 up_read(&mm->mmap_sem);
Zhou Chengming7496fea2016-05-12 15:42:21 -07001670 /*
1671 * up_read(&mm->mmap_sem) first because after
1672 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
1673 * already have been freed under us by __ksm_exit()
1674 * because the "mm_slot" is still hashed and
1675 * ksm_scan.mm_slot doesn't point to it anymore.
1676 */
1677 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001678 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001679
1680 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07001681 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07001682 if (slot != &ksm_mm_head)
1683 goto next_mm;
1684
Izik Eidus31dbd012009-09-21 17:02:03 -07001685 ksm_scan.seqnr++;
1686 return NULL;
1687}
1688
1689/**
1690 * ksm_do_scan - the ksm scanner main worker function.
1691 * @scan_npages - number of pages we want to scan before we return.
1692 */
1693static void ksm_do_scan(unsigned int scan_npages)
1694{
1695 struct rmap_item *rmap_item;
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001696 struct page *uninitialized_var(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001697
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001698 while (scan_npages-- && likely(!freezing(current))) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001699 cond_resched();
1700 rmap_item = scan_get_next_rmap_item(&page);
1701 if (!rmap_item)
1702 return;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001703 cmp_and_merge_page(page, rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001704 put_page(page);
1705 }
1706}
1707
Hugh Dickins6e1583842009-09-21 17:02:14 -07001708static int ksmd_should_run(void)
1709{
1710 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1711}
1712
Izik Eidus31dbd012009-09-21 17:02:03 -07001713static int ksm_scan_thread(void *nothing)
1714{
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001715 set_freezable();
Izik Eidus339aa622009-09-21 17:02:07 -07001716 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07001717
1718 while (!kthread_should_stop()) {
Hugh Dickins6e1583842009-09-21 17:02:14 -07001719 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08001720 wait_while_offlining();
Hugh Dickins6e1583842009-09-21 17:02:14 -07001721 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07001722 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001723 mutex_unlock(&ksm_thread_mutex);
1724
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001725 try_to_freeze();
1726
Hugh Dickins6e1583842009-09-21 17:02:14 -07001727 if (ksmd_should_run()) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001728 schedule_timeout_interruptible(
1729 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1730 } else {
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001731 wait_event_freezable(ksm_thread_wait,
Hugh Dickins6e1583842009-09-21 17:02:14 -07001732 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07001733 }
1734 }
1735 return 0;
1736}
1737
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001738int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1739 unsigned long end, int advice, unsigned long *vm_flags)
1740{
1741 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001742 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001743
1744 switch (advice) {
1745 case MADV_MERGEABLE:
1746 /*
1747 * Be somewhat over-protective for now!
1748 */
1749 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1750 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
Kirill A. Shutemov0661a332015-02-10 14:10:04 -08001751 VM_HUGETLB | VM_MIXEDMAP))
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001752 return 0; /* just ignore the advice */
1753
Konstantin Khlebnikovcc2383e2012-10-08 16:28:37 -07001754#ifdef VM_SAO
1755 if (*vm_flags & VM_SAO)
1756 return 0;
1757#endif
1758
Hugh Dickinsd952b792009-09-21 17:02:16 -07001759 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1760 err = __ksm_enter(mm);
1761 if (err)
1762 return err;
1763 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001764
1765 *vm_flags |= VM_MERGEABLE;
1766 break;
1767
1768 case MADV_UNMERGEABLE:
1769 if (!(*vm_flags & VM_MERGEABLE))
1770 return 0; /* just ignore the advice */
1771
Hugh Dickinsd952b792009-09-21 17:02:16 -07001772 if (vma->anon_vma) {
1773 err = unmerge_ksm_pages(vma, start, end);
1774 if (err)
1775 return err;
1776 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001777
1778 *vm_flags &= ~VM_MERGEABLE;
1779 break;
1780 }
1781
1782 return 0;
1783}
1784
1785int __ksm_enter(struct mm_struct *mm)
1786{
Hugh Dickins6e1583842009-09-21 17:02:14 -07001787 struct mm_slot *mm_slot;
1788 int needs_wakeup;
1789
1790 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07001791 if (!mm_slot)
1792 return -ENOMEM;
1793
Hugh Dickins6e1583842009-09-21 17:02:14 -07001794 /* Check ksm_run too? Would need tighter locking */
1795 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1796
Izik Eidus31dbd012009-09-21 17:02:03 -07001797 spin_lock(&ksm_mmlist_lock);
1798 insert_to_mm_slots_hash(mm, mm_slot);
1799 /*
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001800 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1801 * insert just behind the scanning cursor, to let the area settle
Izik Eidus31dbd012009-09-21 17:02:03 -07001802 * down a little; when fork is followed by immediate exec, we don't
1803 * want ksmd to waste time setting up and tearing down an rmap_list.
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001804 *
1805 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1806 * scanning cursor, otherwise KSM pages in newly forked mms will be
1807 * missed: then we might as well insert at the end of the list.
Izik Eidus31dbd012009-09-21 17:02:03 -07001808 */
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001809 if (ksm_run & KSM_RUN_UNMERGE)
1810 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1811 else
1812 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001813 spin_unlock(&ksm_mmlist_lock);
1814
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001815 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001816 atomic_inc(&mm->mm_count);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001817
1818 if (needs_wakeup)
1819 wake_up_interruptible(&ksm_thread_wait);
1820
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001821 return 0;
1822}
1823
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07001824void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001825{
Hugh Dickinscd551f92009-09-21 17:02:17 -07001826 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001827 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07001828
Izik Eidus31dbd012009-09-21 17:02:03 -07001829 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07001830 * This process is exiting: if it's straightforward (as is the
1831 * case when ksmd was never running), free mm_slot immediately.
1832 * But if it's at the cursor or has rmap_items linked to it, use
1833 * mmap_sem to synchronize with any break_cows before pagetables
1834 * are freed, and leave the mm_slot on the list for ksmd to free.
1835 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07001836 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07001837
Hugh Dickinscd551f92009-09-21 17:02:17 -07001838 spin_lock(&ksm_mmlist_lock);
1839 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001840 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001841 if (!mm_slot->rmap_list) {
Sasha Levin4ca3a692013-02-22 16:32:28 -08001842 hash_del(&mm_slot->link);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001843 list_del(&mm_slot->mm_list);
1844 easy_to_free = 1;
1845 } else {
1846 list_move(&mm_slot->mm_list,
1847 &ksm_scan.mm_slot->mm_list);
1848 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001849 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001850 spin_unlock(&ksm_mmlist_lock);
1851
Hugh Dickins9ba69292009-09-21 17:02:20 -07001852 if (easy_to_free) {
1853 free_mm_slot(mm_slot);
1854 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1855 mmdrop(mm);
1856 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001857 down_write(&mm->mmap_sem);
1858 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001859 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001860}
Izik Eidus31dbd012009-09-21 17:02:03 -07001861
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001862struct page *ksm_might_need_to_copy(struct page *page,
Hugh Dickins5ad64682009-12-14 17:59:24 -08001863 struct vm_area_struct *vma, unsigned long address)
1864{
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001865 struct anon_vma *anon_vma = page_anon_vma(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001866 struct page *new_page;
1867
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001868 if (PageKsm(page)) {
1869 if (page_stable_node(page) &&
1870 !(ksm_run & KSM_RUN_UNMERGE))
1871 return page; /* no need to copy it */
1872 } else if (!anon_vma) {
1873 return page; /* no need to copy it */
1874 } else if (anon_vma->root == vma->anon_vma->root &&
1875 page->index == linear_page_index(vma, address)) {
1876 return page; /* still no need to copy it */
1877 }
1878 if (!PageUptodate(page))
1879 return page; /* let do_swap_page report the error */
1880
Hugh Dickins5ad64682009-12-14 17:59:24 -08001881 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1882 if (new_page) {
1883 copy_user_highpage(new_page, page, address, vma);
1884
1885 SetPageDirty(new_page);
1886 __SetPageUptodate(new_page);
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001887 __SetPageLocked(new_page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001888 }
1889
Hugh Dickins5ad64682009-12-14 17:59:24 -08001890 return new_page;
1891}
1892
Joonsoo Kim051ac832014-01-21 15:49:48 -08001893int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001894{
1895 struct stable_node *stable_node;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001896 struct rmap_item *rmap_item;
1897 int ret = SWAP_AGAIN;
1898 int search_new_forks = 0;
1899
Sasha Levin309381fea2014-01-23 15:52:54 -08001900 VM_BUG_ON_PAGE(!PageKsm(page), page);
Joonsoo Kim9f326242014-01-21 15:49:53 -08001901
1902 /*
1903 * Rely on the page lock to protect against concurrent modifications
1904 * to that page's node of the stable tree.
1905 */
Sasha Levin309381fea2014-01-23 15:52:54 -08001906 VM_BUG_ON_PAGE(!PageLocked(page), page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001907
1908 stable_node = page_stable_node(page);
1909 if (!stable_node)
1910 return ret;
1911again:
Sasha Levinb67bfe02013-02-27 17:06:00 -08001912 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001913 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001914 struct anon_vma_chain *vmac;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001915 struct vm_area_struct *vma;
1916
Andrea Arcangeliad126952015-11-05 18:49:07 -08001917 cond_resched();
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001918 anon_vma_lock_read(anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -07001919 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1920 0, ULONG_MAX) {
Andrea Arcangeliad126952015-11-05 18:49:07 -08001921 cond_resched();
Rik van Riel5beb4932010-03-05 13:42:07 -08001922 vma = vmac->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001923 if (rmap_item->address < vma->vm_start ||
1924 rmap_item->address >= vma->vm_end)
1925 continue;
1926 /*
1927 * Initially we examine only the vma which covers this
1928 * rmap_item; but later, if there is still work to do,
1929 * we examine covering vmas in other mms: in case they
1930 * were forked from the original since ksmd passed.
1931 */
1932 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1933 continue;
1934
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001935 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1936 continue;
1937
Joonsoo Kim051ac832014-01-21 15:49:48 -08001938 ret = rwc->rmap_one(page, vma,
1939 rmap_item->address, rwc->arg);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001940 if (ret != SWAP_AGAIN) {
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001941 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001942 goto out;
1943 }
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001944 if (rwc->done && rwc->done(page)) {
1945 anon_vma_unlock_read(anon_vma);
1946 goto out;
1947 }
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001948 }
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001949 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001950 }
1951 if (!search_new_forks++)
1952 goto again;
1953out:
1954 return ret;
1955}
1956
Joonsoo Kim52629502014-01-21 15:49:50 -08001957#ifdef CONFIG_MIGRATION
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001958void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1959{
1960 struct stable_node *stable_node;
1961
Sasha Levin309381fea2014-01-23 15:52:54 -08001962 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
1963 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1964 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001965
1966 stable_node = page_stable_node(newpage);
1967 if (stable_node) {
Sasha Levin309381fea2014-01-23 15:52:54 -08001968 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001969 stable_node->kpfn = page_to_pfn(newpage);
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001970 /*
1971 * newpage->mapping was set in advance; now we need smp_wmb()
1972 * to make sure that the new stable_node->kpfn is visible
1973 * to get_ksm_page() before it can see that oldpage->mapping
1974 * has gone stale (or that PageSwapCache has been cleared).
1975 */
1976 smp_wmb();
1977 set_page_stable_node(oldpage, NULL);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001978 }
1979}
1980#endif /* CONFIG_MIGRATION */
1981
Hugh Dickins62b61f62009-12-14 17:59:33 -08001982#ifdef CONFIG_MEMORY_HOTREMOVE
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08001983static void wait_while_offlining(void)
1984{
1985 while (ksm_run & KSM_RUN_OFFLINE) {
1986 mutex_unlock(&ksm_thread_mutex);
1987 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
NeilBrown74316202014-07-07 15:16:04 +10001988 TASK_UNINTERRUPTIBLE);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08001989 mutex_lock(&ksm_thread_mutex);
1990 }
1991}
1992
Hugh Dickinsee0ea592013-02-22 16:35:05 -08001993static void ksm_check_stable_tree(unsigned long start_pfn,
1994 unsigned long end_pfn)
Hugh Dickins62b61f62009-12-14 17:59:33 -08001995{
Geliang Tang03640412016-01-14 15:20:54 -08001996 struct stable_node *stable_node, *next;
Hugh Dickins62b61f62009-12-14 17:59:33 -08001997 struct rb_node *node;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001998 int nid;
Hugh Dickins62b61f62009-12-14 17:59:33 -08001999
Hugh Dickinsef53d162013-02-22 16:36:12 -08002000 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
2001 node = rb_first(root_stable_tree + nid);
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002002 while (node) {
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002003 stable_node = rb_entry(node, struct stable_node, node);
2004 if (stable_node->kpfn >= start_pfn &&
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002005 stable_node->kpfn < end_pfn) {
2006 /*
2007 * Don't get_ksm_page, page has already gone:
2008 * which is why we keep kpfn instead of page*
2009 */
2010 remove_node_from_stable_tree(stable_node);
Hugh Dickinsef53d162013-02-22 16:36:12 -08002011 node = rb_first(root_stable_tree + nid);
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002012 } else
2013 node = rb_next(node);
2014 cond_resched();
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002015 }
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002016 }
Geliang Tang03640412016-01-14 15:20:54 -08002017 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002018 if (stable_node->kpfn >= start_pfn &&
2019 stable_node->kpfn < end_pfn)
2020 remove_node_from_stable_tree(stable_node);
2021 cond_resched();
2022 }
Hugh Dickins62b61f62009-12-14 17:59:33 -08002023}
2024
2025static int ksm_memory_callback(struct notifier_block *self,
2026 unsigned long action, void *arg)
2027{
2028 struct memory_notify *mn = arg;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002029
2030 switch (action) {
2031 case MEM_GOING_OFFLINE:
2032 /*
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002033 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2034 * and remove_all_stable_nodes() while memory is going offline:
2035 * it is unsafe for them to touch the stable tree at this time.
2036 * But unmerge_ksm_pages(), rmap lookups and other entry points
2037 * which do not need the ksm_thread_mutex are all safe.
Hugh Dickins62b61f62009-12-14 17:59:33 -08002038 */
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002039 mutex_lock(&ksm_thread_mutex);
2040 ksm_run |= KSM_RUN_OFFLINE;
2041 mutex_unlock(&ksm_thread_mutex);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002042 break;
2043
2044 case MEM_OFFLINE:
2045 /*
2046 * Most of the work is done by page migration; but there might
2047 * be a few stable_nodes left over, still pointing to struct
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002048 * pages which have been offlined: prune those from the tree,
2049 * otherwise get_ksm_page() might later try to access a
2050 * non-existent struct page.
Hugh Dickins62b61f62009-12-14 17:59:33 -08002051 */
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002052 ksm_check_stable_tree(mn->start_pfn,
2053 mn->start_pfn + mn->nr_pages);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002054 /* fallthrough */
2055
2056 case MEM_CANCEL_OFFLINE:
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002057 mutex_lock(&ksm_thread_mutex);
2058 ksm_run &= ~KSM_RUN_OFFLINE;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002059 mutex_unlock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002060
2061 smp_mb(); /* wake_up_bit advises this */
2062 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
Hugh Dickins62b61f62009-12-14 17:59:33 -08002063 break;
2064 }
2065 return NOTIFY_OK;
2066}
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002067#else
2068static void wait_while_offlining(void)
2069{
2070}
Hugh Dickins62b61f62009-12-14 17:59:33 -08002071#endif /* CONFIG_MEMORY_HOTREMOVE */
2072
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002073#ifdef CONFIG_SYSFS
2074/*
2075 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2076 */
2077
Izik Eidus31dbd012009-09-21 17:02:03 -07002078#define KSM_ATTR_RO(_name) \
2079 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2080#define KSM_ATTR(_name) \
2081 static struct kobj_attribute _name##_attr = \
2082 __ATTR(_name, 0644, _name##_show, _name##_store)
2083
2084static ssize_t sleep_millisecs_show(struct kobject *kobj,
2085 struct kobj_attribute *attr, char *buf)
2086{
2087 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2088}
2089
2090static ssize_t sleep_millisecs_store(struct kobject *kobj,
2091 struct kobj_attribute *attr,
2092 const char *buf, size_t count)
2093{
2094 unsigned long msecs;
2095 int err;
2096
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002097 err = kstrtoul(buf, 10, &msecs);
Izik Eidus31dbd012009-09-21 17:02:03 -07002098 if (err || msecs > UINT_MAX)
2099 return -EINVAL;
2100
2101 ksm_thread_sleep_millisecs = msecs;
2102
2103 return count;
2104}
2105KSM_ATTR(sleep_millisecs);
2106
2107static ssize_t pages_to_scan_show(struct kobject *kobj,
2108 struct kobj_attribute *attr, char *buf)
2109{
2110 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2111}
2112
2113static ssize_t pages_to_scan_store(struct kobject *kobj,
2114 struct kobj_attribute *attr,
2115 const char *buf, size_t count)
2116{
2117 int err;
2118 unsigned long nr_pages;
2119
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002120 err = kstrtoul(buf, 10, &nr_pages);
Izik Eidus31dbd012009-09-21 17:02:03 -07002121 if (err || nr_pages > UINT_MAX)
2122 return -EINVAL;
2123
2124 ksm_thread_pages_to_scan = nr_pages;
2125
2126 return count;
2127}
2128KSM_ATTR(pages_to_scan);
2129
2130static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2131 char *buf)
2132{
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002133 return sprintf(buf, "%lu\n", ksm_run);
Izik Eidus31dbd012009-09-21 17:02:03 -07002134}
2135
2136static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2137 const char *buf, size_t count)
2138{
2139 int err;
2140 unsigned long flags;
2141
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002142 err = kstrtoul(buf, 10, &flags);
Izik Eidus31dbd012009-09-21 17:02:03 -07002143 if (err || flags > UINT_MAX)
2144 return -EINVAL;
2145 if (flags > KSM_RUN_UNMERGE)
2146 return -EINVAL;
2147
2148 /*
2149 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2150 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsd0f209f2009-12-14 17:59:34 -08002151 * breaking COW to free the pages_shared (but leaves mm_slots
2152 * on the list for when ksmd may be set running again).
Izik Eidus31dbd012009-09-21 17:02:03 -07002153 */
2154
2155 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002156 wait_while_offlining();
Izik Eidus31dbd012009-09-21 17:02:03 -07002157 if (ksm_run != flags) {
2158 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07002159 if (flags & KSM_RUN_UNMERGE) {
David Rientjese1e12d22012-12-11 16:02:56 -08002160 set_current_oom_origin();
Hugh Dickinsd952b792009-09-21 17:02:16 -07002161 err = unmerge_and_remove_all_rmap_items();
David Rientjese1e12d22012-12-11 16:02:56 -08002162 clear_current_oom_origin();
Hugh Dickinsd952b792009-09-21 17:02:16 -07002163 if (err) {
2164 ksm_run = KSM_RUN_STOP;
2165 count = err;
2166 }
2167 }
Izik Eidus31dbd012009-09-21 17:02:03 -07002168 }
2169 mutex_unlock(&ksm_thread_mutex);
2170
2171 if (flags & KSM_RUN_MERGE)
2172 wake_up_interruptible(&ksm_thread_wait);
2173
2174 return count;
2175}
2176KSM_ATTR(run);
2177
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002178#ifdef CONFIG_NUMA
2179static ssize_t merge_across_nodes_show(struct kobject *kobj,
2180 struct kobj_attribute *attr, char *buf)
2181{
2182 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2183}
2184
2185static ssize_t merge_across_nodes_store(struct kobject *kobj,
2186 struct kobj_attribute *attr,
2187 const char *buf, size_t count)
2188{
2189 int err;
2190 unsigned long knob;
2191
2192 err = kstrtoul(buf, 10, &knob);
2193 if (err)
2194 return err;
2195 if (knob > 1)
2196 return -EINVAL;
2197
2198 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002199 wait_while_offlining();
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002200 if (ksm_merge_across_nodes != knob) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002201 if (ksm_pages_shared || remove_all_stable_nodes())
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002202 err = -EBUSY;
Hugh Dickinsef53d162013-02-22 16:36:12 -08002203 else if (root_stable_tree == one_stable_tree) {
2204 struct rb_root *buf;
2205 /*
2206 * This is the first time that we switch away from the
2207 * default of merging across nodes: must now allocate
2208 * a buffer to hold as many roots as may be needed.
2209 * Allocate stable and unstable together:
2210 * MAXSMP NODES_SHIFT 10 will use 16kB.
2211 */
Joe Perchesbafe1e12013-11-12 15:07:10 -08002212 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2213 GFP_KERNEL);
Hugh Dickinsef53d162013-02-22 16:36:12 -08002214 /* Let us assume that RB_ROOT is NULL is zero */
2215 if (!buf)
2216 err = -ENOMEM;
2217 else {
2218 root_stable_tree = buf;
2219 root_unstable_tree = buf + nr_node_ids;
2220 /* Stable tree is empty but not the unstable */
2221 root_unstable_tree[0] = one_unstable_tree[0];
2222 }
2223 }
2224 if (!err) {
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002225 ksm_merge_across_nodes = knob;
Hugh Dickinsef53d162013-02-22 16:36:12 -08002226 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2227 }
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002228 }
2229 mutex_unlock(&ksm_thread_mutex);
2230
2231 return err ? err : count;
2232}
2233KSM_ATTR(merge_across_nodes);
2234#endif
2235
Hugh Dickinsb4028262009-09-21 17:02:09 -07002236static ssize_t pages_shared_show(struct kobject *kobj,
2237 struct kobj_attribute *attr, char *buf)
2238{
2239 return sprintf(buf, "%lu\n", ksm_pages_shared);
2240}
2241KSM_ATTR_RO(pages_shared);
2242
2243static ssize_t pages_sharing_show(struct kobject *kobj,
2244 struct kobj_attribute *attr, char *buf)
2245{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07002246 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07002247}
2248KSM_ATTR_RO(pages_sharing);
2249
Hugh Dickins473b0ce2009-09-21 17:02:11 -07002250static ssize_t pages_unshared_show(struct kobject *kobj,
2251 struct kobj_attribute *attr, char *buf)
2252{
2253 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2254}
2255KSM_ATTR_RO(pages_unshared);
2256
2257static ssize_t pages_volatile_show(struct kobject *kobj,
2258 struct kobj_attribute *attr, char *buf)
2259{
2260 long ksm_pages_volatile;
2261
2262 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2263 - ksm_pages_sharing - ksm_pages_unshared;
2264 /*
2265 * It was not worth any locking to calculate that statistic,
2266 * but it might therefore sometimes be negative: conceal that.
2267 */
2268 if (ksm_pages_volatile < 0)
2269 ksm_pages_volatile = 0;
2270 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2271}
2272KSM_ATTR_RO(pages_volatile);
2273
2274static ssize_t full_scans_show(struct kobject *kobj,
2275 struct kobj_attribute *attr, char *buf)
2276{
2277 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2278}
2279KSM_ATTR_RO(full_scans);
2280
Izik Eidus31dbd012009-09-21 17:02:03 -07002281static struct attribute *ksm_attrs[] = {
2282 &sleep_millisecs_attr.attr,
2283 &pages_to_scan_attr.attr,
2284 &run_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07002285 &pages_shared_attr.attr,
2286 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07002287 &pages_unshared_attr.attr,
2288 &pages_volatile_attr.attr,
2289 &full_scans_attr.attr,
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002290#ifdef CONFIG_NUMA
2291 &merge_across_nodes_attr.attr,
2292#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07002293 NULL,
2294};
2295
2296static struct attribute_group ksm_attr_group = {
2297 .attrs = ksm_attrs,
2298 .name = "ksm",
2299};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002300#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002301
2302static int __init ksm_init(void)
2303{
2304 struct task_struct *ksm_thread;
2305 int err;
2306
2307 err = ksm_slab_init();
2308 if (err)
2309 goto out;
2310
Izik Eidus31dbd012009-09-21 17:02:03 -07002311 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2312 if (IS_ERR(ksm_thread)) {
Paul McQuade25acde32014-10-09 15:29:09 -07002313 pr_err("ksm: creating kthread failed\n");
Izik Eidus31dbd012009-09-21 17:02:03 -07002314 err = PTR_ERR(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002315 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002316 }
2317
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002318#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07002319 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2320 if (err) {
Paul McQuade25acde32014-10-09 15:29:09 -07002321 pr_err("ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002322 kthread_stop(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002323 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002324 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07002325#else
2326 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2327
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002328#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002329
Hugh Dickins62b61f62009-12-14 17:59:33 -08002330#ifdef CONFIG_MEMORY_HOTREMOVE
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002331 /* There is no significance to this priority 100 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08002332 hotplug_memory_notifier(ksm_memory_callback, 100);
2333#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07002334 return 0;
2335
Lai Jiangshand9f89842010-08-09 17:20:02 -07002336out_free:
Izik Eidus31dbd012009-09-21 17:02:03 -07002337 ksm_slab_free();
2338out:
2339 return err;
2340}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -08002341subsys_initcall(ksm_init);