blob: 4786b4150f62108438f5aabaf77d339280eb51dc [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
286 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
287 if (rmap_item)
288 ksm_rmap_items++;
289 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -0700290}
291
292static inline void free_rmap_item(struct rmap_item *rmap_item)
293{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700294 ksm_rmap_items--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700295 rmap_item->mm = NULL; /* debug safety */
296 kmem_cache_free(rmap_item_cache, rmap_item);
297}
298
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800299static inline struct stable_node *alloc_stable_node(void)
300{
301 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
302}
303
304static inline void free_stable_node(struct stable_node *stable_node)
305{
306 kmem_cache_free(stable_node_cache, stable_node);
307}
308
Izik Eidus31dbd012009-09-21 17:02:03 -0700309static inline struct mm_slot *alloc_mm_slot(void)
310{
311 if (!mm_slot_cache) /* initialization failed */
312 return NULL;
313 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
314}
315
316static inline void free_mm_slot(struct mm_slot *mm_slot)
317{
318 kmem_cache_free(mm_slot_cache, mm_slot);
319}
320
Izik Eidus31dbd012009-09-21 17:02:03 -0700321static struct mm_slot *get_mm_slot(struct mm_struct *mm)
322{
Sasha Levin4ca3a692013-02-22 16:32:28 -0800323 struct mm_slot *slot;
Izik Eidus31dbd012009-09-21 17:02:03 -0700324
Sasha Levinb67bfe02013-02-27 17:06:00 -0800325 hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm)
Sasha Levin4ca3a692013-02-22 16:32:28 -0800326 if (slot->mm == mm)
327 return slot;
328
Izik Eidus31dbd012009-09-21 17:02:03 -0700329 return NULL;
330}
331
332static void insert_to_mm_slots_hash(struct mm_struct *mm,
333 struct mm_slot *mm_slot)
334{
Izik Eidus31dbd012009-09-21 17:02:03 -0700335 mm_slot->mm = mm;
Sasha Levin4ca3a692013-02-22 16:32:28 -0800336 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
Izik Eidus31dbd012009-09-21 17:02:03 -0700337}
338
Izik Eidus31dbd012009-09-21 17:02:03 -0700339/*
Hugh Dickinsa913e182009-09-21 17:02:26 -0700340 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
341 * page tables after it has passed through ksm_exit() - which, if necessary,
342 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
343 * a special flag: they can just back out as soon as mm_users goes to zero.
344 * ksm_test_exit() is used throughout to make this test for exit: in some
345 * places for correctness, in some places just to avoid unnecessary work.
346 */
347static inline bool ksm_test_exit(struct mm_struct *mm)
348{
349 return atomic_read(&mm->mm_users) == 0;
350}
351
352/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700353 * We use break_ksm to break COW on a ksm page: it's a stripped down
354 *
Dave Hansend4edcf02016-02-12 13:01:56 -0800355 * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1)
Izik Eidus31dbd012009-09-21 17:02:03 -0700356 * put_page(page);
357 *
358 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
359 * in case the application has unmapped and remapped mm,addr meanwhile.
360 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
361 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
Dave Hansen1b2ee122016-02-12 13:02:21 -0800362 *
363 * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context
364 * of the process that owns 'vma'. We also do not want to enforce
365 * protection keys here anyway.
Izik Eidus31dbd012009-09-21 17:02:03 -0700366 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700367static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
Izik Eidus31dbd012009-09-21 17:02:03 -0700368{
369 struct page *page;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700370 int ret = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700371
372 do {
373 cond_resched();
Dave Hansen1b2ee122016-02-12 13:02:21 -0800374 page = follow_page(vma, addr,
375 FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400376 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700377 break;
378 if (PageKsm(page))
379 ret = handle_mm_fault(vma->vm_mm, vma, addr,
Dave Hansen1b2ee122016-02-12 13:02:21 -0800380 FAULT_FLAG_WRITE |
381 FAULT_FLAG_REMOTE);
Izik Eidus31dbd012009-09-21 17:02:03 -0700382 else
383 ret = VM_FAULT_WRITE;
384 put_page(page);
Linus Torvalds33692f22015-01-29 10:51:32 -0800385 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM)));
Hugh Dickinsd952b792009-09-21 17:02:16 -0700386 /*
387 * We must loop because handle_mm_fault() may back out if there's
388 * any difficulty e.g. if pte accessed bit gets updated concurrently.
389 *
390 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
391 * COW has been broken, even if the vma does not permit VM_WRITE;
392 * but note that a concurrent fault might break PageKsm for us.
393 *
394 * VM_FAULT_SIGBUS could occur if we race with truncation of the
395 * backing file, which also invalidates anonymous pages: that's
396 * okay, that truncation will have unmapped the PageKsm for us.
397 *
398 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
399 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
400 * current task has TIF_MEMDIE set, and will be OOM killed on return
401 * to user; and ksmd, having no mm, would never be chosen for that.
402 *
403 * But if the mm is in a limited mem_cgroup, then the fault may fail
404 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
405 * even ksmd can fail in this way - though it's usually breaking ksm
406 * just to undo a merge it made a moment before, so unlikely to oom.
407 *
408 * That's a pity: we might therefore have more kernel pages allocated
409 * than we're counting as nodes in the stable tree; but ksm_do_scan
410 * will retry to break_cow on each pass, so should recover the page
411 * in due course. The important thing is to not let VM_MERGEABLE
412 * be cleared while any such pages might remain in the area.
413 */
414 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700415}
416
Bob Liuef694222012-03-21 16:34:11 -0700417static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
418 unsigned long addr)
419{
420 struct vm_area_struct *vma;
421 if (ksm_test_exit(mm))
422 return NULL;
423 vma = find_vma(mm, addr);
424 if (!vma || vma->vm_start > addr)
425 return NULL;
426 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
427 return NULL;
428 return vma;
429}
430
Hugh Dickins8dd35572009-12-14 17:59:18 -0800431static void break_cow(struct rmap_item *rmap_item)
Izik Eidus31dbd012009-09-21 17:02:03 -0700432{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800433 struct mm_struct *mm = rmap_item->mm;
434 unsigned long addr = rmap_item->address;
Izik Eidus31dbd012009-09-21 17:02:03 -0700435 struct vm_area_struct *vma;
436
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800437 /*
438 * It is not an accident that whenever we want to break COW
439 * to undo, we also need to drop a reference to the anon_vma.
440 */
Peter Zijlstra9e601092011-03-22 16:32:46 -0700441 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800442
Hugh Dickins81464e302009-09-21 17:02:15 -0700443 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700444 vma = find_mergeable_vma(mm, addr);
445 if (vma)
446 break_ksm(vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700447 up_read(&mm->mmap_sem);
448}
449
450static struct page *get_mergeable_page(struct rmap_item *rmap_item)
451{
452 struct mm_struct *mm = rmap_item->mm;
453 unsigned long addr = rmap_item->address;
454 struct vm_area_struct *vma;
455 struct page *page;
456
457 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700458 vma = find_mergeable_vma(mm, addr);
459 if (!vma)
Izik Eidus31dbd012009-09-21 17:02:03 -0700460 goto out;
461
462 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400463 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700464 goto out;
Kirill A. Shutemovf765f542016-01-15 16:53:03 -0800465 if (PageAnon(page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700466 flush_anon_page(vma, page, addr);
467 flush_dcache_page(page);
468 } else {
469 put_page(page);
Andrea Arcangelic8f95ed2015-11-05 18:49:19 -0800470out:
471 page = NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700472 }
473 up_read(&mm->mmap_sem);
474 return page;
475}
476
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800477/*
478 * This helper is used for getting right index into array of tree roots.
479 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
480 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
481 * every node has its own stable and unstable tree.
482 */
483static inline int get_kpfn_nid(unsigned long kpfn)
484{
Hugh Dickinsd8fc16a2013-03-08 12:43:34 -0800485 return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn));
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800486}
487
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800488static void remove_node_from_stable_tree(struct stable_node *stable_node)
489{
490 struct rmap_item *rmap_item;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800491
Sasha Levinb67bfe02013-02-27 17:06:00 -0800492 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800493 if (rmap_item->hlist.next)
494 ksm_pages_sharing--;
495 else
496 ksm_pages_shared--;
Peter Zijlstra9e601092011-03-22 16:32:46 -0700497 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800498 rmap_item->address &= PAGE_MASK;
499 cond_resched();
500 }
501
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800502 if (stable_node->head == &migrate_nodes)
503 list_del(&stable_node->list);
504 else
505 rb_erase(&stable_node->node,
Hugh Dickinsef53d162013-02-22 16:36:12 -0800506 root_stable_tree + NUMA(stable_node->nid));
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800507 free_stable_node(stable_node);
508}
509
510/*
511 * get_ksm_page: checks if the page indicated by the stable node
512 * is still its ksm page, despite having held no reference to it.
513 * In which case we can trust the content of the page, and it
514 * returns the gotten page; but if the page has now been zapped,
515 * remove the stale node from the stable tree and return NULL.
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800516 * But beware, the stable node's page might be being migrated.
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800517 *
518 * You would expect the stable_node to hold a reference to the ksm page.
519 * But if it increments the page's count, swapping out has to wait for
520 * ksmd to come around again before it can free the page, which may take
521 * seconds or even minutes: much too unresponsive. So instead we use a
522 * "keyhole reference": access to the ksm page from the stable node peeps
523 * out through its keyhole to see if that page still holds the right key,
524 * pointing back to this stable node. This relies on freeing a PageAnon
525 * page to reset its page->mapping to NULL, and relies on no other use of
526 * a page to put something that might look like our key in page->mapping.
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800527 * is on its way to being freed; but it is an anomaly to bear in mind.
528 */
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800529static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800530{
531 struct page *page;
532 void *expected_mapping;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800533 unsigned long kpfn;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800534
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800535 expected_mapping = (void *)stable_node +
536 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800537again:
Jason Low4db0c3c2015-04-15 16:14:08 -0700538 kpfn = READ_ONCE(stable_node->kpfn);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800539 page = pfn_to_page(kpfn);
540
541 /*
542 * page is computed from kpfn, so on most architectures reading
543 * page->mapping is naturally ordered after reading node->kpfn,
544 * but on Alpha we need to be more careful.
545 */
546 smp_read_barrier_depends();
Jason Low4db0c3c2015-04-15 16:14:08 -0700547 if (READ_ONCE(page->mapping) != expected_mapping)
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800548 goto stale;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800549
550 /*
551 * We cannot do anything with the page while its refcount is 0.
552 * Usually 0 means free, or tail of a higher-order page: in which
553 * case this node is no longer referenced, and should be freed;
554 * however, it might mean that the page is under page_freeze_refs().
555 * The __remove_mapping() case is easy, again the node is now stale;
556 * but if page is swapcache in migrate_page_move_mapping(), it might
557 * still be our page, in which case it's essential to keep the node.
558 */
559 while (!get_page_unless_zero(page)) {
560 /*
561 * Another check for page->mapping != expected_mapping would
562 * work here too. We have chosen the !PageSwapCache test to
563 * optimize the common case, when the page is or is about to
564 * be freed: PageSwapCache is cleared (under spin_lock_irq)
565 * in the freeze_refs section of __remove_mapping(); but Anon
566 * page->mapping reset to NULL later, in free_pages_prepare().
567 */
568 if (!PageSwapCache(page))
569 goto stale;
570 cpu_relax();
571 }
572
Jason Low4db0c3c2015-04-15 16:14:08 -0700573 if (READ_ONCE(page->mapping) != expected_mapping) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800574 put_page(page);
575 goto stale;
576 }
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800577
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800578 if (lock_it) {
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800579 lock_page(page);
Jason Low4db0c3c2015-04-15 16:14:08 -0700580 if (READ_ONCE(page->mapping) != expected_mapping) {
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800581 unlock_page(page);
582 put_page(page);
583 goto stale;
584 }
585 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800586 return page;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800587
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800588stale:
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800589 /*
590 * We come here from above when page->mapping or !PageSwapCache
591 * suggests that the node is stale; but it might be under migration.
592 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
593 * before checking whether node->kpfn has been changed.
594 */
595 smp_rmb();
Jason Low4db0c3c2015-04-15 16:14:08 -0700596 if (READ_ONCE(stable_node->kpfn) != kpfn)
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800597 goto again;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800598 remove_node_from_stable_tree(stable_node);
599 return NULL;
600}
601
Izik Eidus31dbd012009-09-21 17:02:03 -0700602/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700603 * Removing rmap_item from stable or unstable tree.
604 * This function will clean the information from the stable/unstable tree.
605 */
606static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
607{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800608 if (rmap_item->address & STABLE_FLAG) {
609 struct stable_node *stable_node;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800610 struct page *page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700611
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800612 stable_node = rmap_item->head;
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800613 page = get_ksm_page(stable_node, true);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800614 if (!page)
615 goto out;
616
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800617 hlist_del(&rmap_item->hlist);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800618 unlock_page(page);
619 put_page(page);
Hugh Dickins08beca42009-12-14 17:59:21 -0800620
Andrea Arcangeli98666f8a2015-11-05 18:49:13 -0800621 if (!hlist_empty(&stable_node->hlist))
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800622 ksm_pages_sharing--;
623 else
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800624 ksm_pages_shared--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700625
Peter Zijlstra9e601092011-03-22 16:32:46 -0700626 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins93d17712009-12-14 17:59:16 -0800627 rmap_item->address &= PAGE_MASK;
Izik Eidus31dbd012009-09-21 17:02:03 -0700628
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800629 } else if (rmap_item->address & UNSTABLE_FLAG) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700630 unsigned char age;
631 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -0700632 * Usually ksmd can and must skip the rb_erase, because
Izik Eidus31dbd012009-09-21 17:02:03 -0700633 * root_unstable_tree was already reset to RB_ROOT.
Hugh Dickins9ba69292009-09-21 17:02:20 -0700634 * But be careful when an mm is exiting: do the rb_erase
635 * if this rmap_item was inserted by this scan, rather
636 * than left over from before.
Izik Eidus31dbd012009-09-21 17:02:03 -0700637 */
638 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
Hugh Dickinscd551f92009-09-21 17:02:17 -0700639 BUG_ON(age > 1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700640 if (!age)
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800641 rb_erase(&rmap_item->node,
Hugh Dickinsef53d162013-02-22 16:36:12 -0800642 root_unstable_tree + NUMA(rmap_item->nid));
Hugh Dickins93d17712009-12-14 17:59:16 -0800643 ksm_pages_unshared--;
644 rmap_item->address &= PAGE_MASK;
645 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800646out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700647 cond_resched(); /* we're called from many long loops */
648}
649
Izik Eidus31dbd012009-09-21 17:02:03 -0700650static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -0800651 struct rmap_item **rmap_list)
Izik Eidus31dbd012009-09-21 17:02:03 -0700652{
Hugh Dickins6514d512009-12-14 17:59:19 -0800653 while (*rmap_list) {
654 struct rmap_item *rmap_item = *rmap_list;
655 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700656 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700657 free_rmap_item(rmap_item);
658 }
659}
660
661/*
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800662 * Though it's very tempting to unmerge rmap_items from stable tree rather
Izik Eidus31dbd012009-09-21 17:02:03 -0700663 * than check every pte of a given vma, the locking doesn't quite work for
664 * that - an rmap_item is assigned to the stable tree after inserting ksm
665 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
666 * rmap_items from parent to child at fork time (so as not to waste time
667 * if exit comes before the next scan reaches it).
Hugh Dickins81464e302009-09-21 17:02:15 -0700668 *
669 * Similarly, although we'd like to remove rmap_items (so updating counts
670 * and freeing memory) when unmerging an area, it's easier to leave that
671 * to the next pass of ksmd - consider, for example, how ksmd might be
672 * in cmp_and_merge_page on one of the rmap_items we would be removing.
Izik Eidus31dbd012009-09-21 17:02:03 -0700673 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700674static int unmerge_ksm_pages(struct vm_area_struct *vma,
675 unsigned long start, unsigned long end)
Izik Eidus31dbd012009-09-21 17:02:03 -0700676{
677 unsigned long addr;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700678 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700679
Hugh Dickinsd952b792009-09-21 17:02:16 -0700680 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700681 if (ksm_test_exit(vma->vm_mm))
682 break;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700683 if (signal_pending(current))
684 err = -ERESTARTSYS;
685 else
686 err = break_ksm(vma, addr);
687 }
688 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700689}
690
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700691#ifdef CONFIG_SYSFS
692/*
693 * Only called through the sysfs control interface:
694 */
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800695static int remove_stable_node(struct stable_node *stable_node)
696{
697 struct page *page;
698 int err;
699
700 page = get_ksm_page(stable_node, true);
701 if (!page) {
702 /*
703 * get_ksm_page did remove_node_from_stable_tree itself.
704 */
705 return 0;
706 }
707
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800708 if (WARN_ON_ONCE(page_mapped(page))) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800709 /*
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800710 * This should not happen: but if it does, just refuse to let
711 * merge_across_nodes be switched - there is no need to panic.
712 */
713 err = -EBUSY;
714 } else {
715 /*
716 * The stable node did not yet appear stale to get_ksm_page(),
717 * since that allows for an unmapped ksm page to be recognized
718 * right up until it is freed; but the node is safe to remove.
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800719 * This page might be in a pagevec waiting to be freed,
720 * or it might be PageSwapCache (perhaps under writeback),
721 * or it might have been removed from swapcache a moment ago.
722 */
723 set_page_stable_node(page, NULL);
724 remove_node_from_stable_tree(stable_node);
725 err = 0;
726 }
727
728 unlock_page(page);
729 put_page(page);
730 return err;
731}
732
733static int remove_all_stable_nodes(void)
734{
Geliang Tang03640412016-01-14 15:20:54 -0800735 struct stable_node *stable_node, *next;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800736 int nid;
737 int err = 0;
738
Hugh Dickinsef53d162013-02-22 16:36:12 -0800739 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800740 while (root_stable_tree[nid].rb_node) {
741 stable_node = rb_entry(root_stable_tree[nid].rb_node,
742 struct stable_node, node);
743 if (remove_stable_node(stable_node)) {
744 err = -EBUSY;
745 break; /* proceed to next nid */
746 }
747 cond_resched();
748 }
749 }
Geliang Tang03640412016-01-14 15:20:54 -0800750 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800751 if (remove_stable_node(stable_node))
752 err = -EBUSY;
753 cond_resched();
754 }
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800755 return err;
756}
757
Hugh Dickinsd952b792009-09-21 17:02:16 -0700758static int unmerge_and_remove_all_rmap_items(void)
Izik Eidus31dbd012009-09-21 17:02:03 -0700759{
760 struct mm_slot *mm_slot;
761 struct mm_struct *mm;
762 struct vm_area_struct *vma;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700763 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700764
Hugh Dickinsd952b792009-09-21 17:02:16 -0700765 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700766 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700767 struct mm_slot, mm_list);
768 spin_unlock(&ksm_mmlist_lock);
769
Hugh Dickins9ba69292009-09-21 17:02:20 -0700770 for (mm_slot = ksm_scan.mm_slot;
771 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700772 mm = mm_slot->mm;
773 down_read(&mm->mmap_sem);
774 for (vma = mm->mmap; vma; vma = vma->vm_next) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700775 if (ksm_test_exit(mm))
776 break;
Izik Eidus31dbd012009-09-21 17:02:03 -0700777 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
778 continue;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700779 err = unmerge_ksm_pages(vma,
780 vma->vm_start, vma->vm_end);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700781 if (err)
782 goto error;
Izik Eidus31dbd012009-09-21 17:02:03 -0700783 }
Hugh Dickins9ba69292009-09-21 17:02:20 -0700784
Hugh Dickins6514d512009-12-14 17:59:19 -0800785 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
Zhou Chengming7496fea2016-05-12 15:42:21 -0700786 up_read(&mm->mmap_sem);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700787
788 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700789 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700790 struct mm_slot, mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700791 if (ksm_test_exit(mm)) {
Sasha Levin4ca3a692013-02-22 16:32:28 -0800792 hash_del(&mm_slot->link);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700793 list_del(&mm_slot->mm_list);
794 spin_unlock(&ksm_mmlist_lock);
795
796 free_mm_slot(mm_slot);
797 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700798 mmdrop(mm);
Zhou Chengming7496fea2016-05-12 15:42:21 -0700799 } else
Hugh Dickins9ba69292009-09-21 17:02:20 -0700800 spin_unlock(&ksm_mmlist_lock);
Izik Eidus31dbd012009-09-21 17:02:03 -0700801 }
802
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800803 /* Clean up stable nodes, but don't worry if some are still busy */
804 remove_all_stable_nodes();
Hugh Dickinsd952b792009-09-21 17:02:16 -0700805 ksm_scan.seqnr = 0;
Hugh Dickins9ba69292009-09-21 17:02:20 -0700806 return 0;
807
808error:
809 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700810 spin_lock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700811 ksm_scan.mm_slot = &ksm_mm_head;
Izik Eidus31dbd012009-09-21 17:02:03 -0700812 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700813 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700814}
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700815#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -0700816
Izik Eidus31dbd012009-09-21 17:02:03 -0700817static u32 calc_checksum(struct page *page)
818{
819 u32 checksum;
Cong Wang9b04c5f2011-11-25 23:14:39 +0800820 void *addr = kmap_atomic(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700821 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
Cong Wang9b04c5f2011-11-25 23:14:39 +0800822 kunmap_atomic(addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700823 return checksum;
824}
825
826static int memcmp_pages(struct page *page1, struct page *page2)
827{
828 char *addr1, *addr2;
829 int ret;
830
Cong Wang9b04c5f2011-11-25 23:14:39 +0800831 addr1 = kmap_atomic(page1);
832 addr2 = kmap_atomic(page2);
Izik Eidus31dbd012009-09-21 17:02:03 -0700833 ret = memcmp(addr1, addr2, PAGE_SIZE);
Cong Wang9b04c5f2011-11-25 23:14:39 +0800834 kunmap_atomic(addr2);
835 kunmap_atomic(addr1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700836 return ret;
837}
838
839static inline int pages_identical(struct page *page1, struct page *page2)
840{
841 return !memcmp_pages(page1, page2);
842}
843
844static int write_protect_page(struct vm_area_struct *vma, struct page *page,
845 pte_t *orig_pte)
846{
847 struct mm_struct *mm = vma->vm_mm;
848 unsigned long addr;
849 pte_t *ptep;
850 spinlock_t *ptl;
851 int swapped;
852 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700853 unsigned long mmun_start; /* For mmu_notifiers */
854 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -0700855
856 addr = page_address_in_vma(page, vma);
857 if (addr == -EFAULT)
858 goto out;
859
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800860 BUG_ON(PageTransCompound(page));
Haggai Eran6bdb9132012-10-08 16:33:35 -0700861
862 mmun_start = addr;
863 mmun_end = addr + PAGE_SIZE;
864 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
865
Izik Eidus31dbd012009-09-21 17:02:03 -0700866 ptep = page_check_address(page, mm, addr, &ptl, 0);
867 if (!ptep)
Haggai Eran6bdb9132012-10-08 16:33:35 -0700868 goto out_mn;
Izik Eidus31dbd012009-09-21 17:02:03 -0700869
Hugh Dickins4e316352010-10-02 17:49:08 -0700870 if (pte_write(*ptep) || pte_dirty(*ptep)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700871 pte_t entry;
872
873 swapped = PageSwapCache(page);
874 flush_cache_page(vma, addr, page_to_pfn(page));
875 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300876 * Ok this is tricky, when get_user_pages_fast() run it doesn't
Izik Eidus31dbd012009-09-21 17:02:03 -0700877 * take any lock, therefore the check that we are going to make
878 * with the pagecount against the mapcount is racey and
879 * O_DIRECT can happen right after the check.
880 * So we clear the pte and flush the tlb before the check
881 * this assure us that no O_DIRECT can happen after the check
882 * or in the middle of the check.
883 */
Joerg Roedel34ee6452014-11-13 13:46:09 +1100884 entry = ptep_clear_flush_notify(vma, addr, ptep);
Izik Eidus31dbd012009-09-21 17:02:03 -0700885 /*
886 * Check that no O_DIRECT or similar I/O is in progress on the
887 * page
888 */
Hugh Dickins31e855e2009-12-14 17:59:17 -0800889 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
Robin Holtcb532372010-03-23 13:35:26 -0700890 set_pte_at(mm, addr, ptep, entry);
Izik Eidus31dbd012009-09-21 17:02:03 -0700891 goto out_unlock;
892 }
Hugh Dickins4e316352010-10-02 17:49:08 -0700893 if (pte_dirty(entry))
894 set_page_dirty(page);
895 entry = pte_mkclean(pte_wrprotect(entry));
Izik Eidus31dbd012009-09-21 17:02:03 -0700896 set_pte_at_notify(mm, addr, ptep, entry);
897 }
898 *orig_pte = *ptep;
899 err = 0;
900
901out_unlock:
902 pte_unmap_unlock(ptep, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -0700903out_mn:
904 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -0700905out:
906 return err;
907}
908
909/**
910 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -0800911 * @vma: vma that holds the pte pointing to page
912 * @page: the page we are replacing by kpage
913 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -0700914 * @orig_pte: the original value of the pte
915 *
916 * Returns 0 on success, -EFAULT on failure.
917 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800918static int replace_page(struct vm_area_struct *vma, struct page *page,
919 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -0700920{
921 struct mm_struct *mm = vma->vm_mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700922 pmd_t *pmd;
923 pte_t *ptep;
924 spinlock_t *ptl;
925 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -0700926 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700927 unsigned long mmun_start; /* For mmu_notifiers */
928 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -0700929
Hugh Dickins8dd35572009-12-14 17:59:18 -0800930 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -0700931 if (addr == -EFAULT)
932 goto out;
933
Bob Liu62190492012-12-11 16:00:37 -0800934 pmd = mm_find_pmd(mm, addr);
935 if (!pmd)
Izik Eidus31dbd012009-09-21 17:02:03 -0700936 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700937
Haggai Eran6bdb9132012-10-08 16:33:35 -0700938 mmun_start = addr;
939 mmun_end = addr + PAGE_SIZE;
940 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
941
Izik Eidus31dbd012009-09-21 17:02:03 -0700942 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
943 if (!pte_same(*ptep, orig_pte)) {
944 pte_unmap_unlock(ptep, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -0700945 goto out_mn;
Izik Eidus31dbd012009-09-21 17:02:03 -0700946 }
947
Hugh Dickins8dd35572009-12-14 17:59:18 -0800948 get_page(kpage);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800949 page_add_anon_rmap(kpage, vma, addr, false);
Izik Eidus31dbd012009-09-21 17:02:03 -0700950
951 flush_cache_page(vma, addr, pte_pfn(*ptep));
Joerg Roedel34ee6452014-11-13 13:46:09 +1100952 ptep_clear_flush_notify(vma, addr, ptep);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800953 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
Izik Eidus31dbd012009-09-21 17:02:03 -0700954
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -0800955 page_remove_rmap(page, false);
Hugh Dickinsae52a2a2011-01-13 15:46:28 -0800956 if (!page_mapped(page))
957 try_to_free_swap(page);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800958 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700959
960 pte_unmap_unlock(ptep, ptl);
961 err = 0;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700962out_mn:
963 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -0700964out:
965 return err;
966}
967
968/*
969 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -0800970 * @vma: the vma that holds the pte pointing to page
971 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins80e148222009-12-14 17:59:29 -0800972 * @kpage: the PageKsm page that we want to map instead of page,
973 * or NULL the first time when we want to use page as kpage.
Izik Eidus31dbd012009-09-21 17:02:03 -0700974 *
975 * This function returns 0 if the pages were merged, -EFAULT otherwise.
976 */
977static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -0800978 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700979{
980 pte_t orig_pte = __pte(0);
981 int err = -EFAULT;
982
Hugh Dickinsdb114b82009-12-14 17:59:25 -0800983 if (page == kpage) /* ksm page forked */
984 return 0;
985
Hugh Dickins8dd35572009-12-14 17:59:18 -0800986 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700987 goto out;
988
Izik Eidus31dbd012009-09-21 17:02:03 -0700989 /*
990 * We need the page lock to read a stable PageSwapCache in
991 * write_protect_page(). We use trylock_page() instead of
992 * lock_page() because we don't want to wait here - we
993 * prefer to continue scanning and merging different pages,
994 * then come back to this page when it is unlocked.
995 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800996 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -0800997 goto out;
Kirill A. Shutemovf765f542016-01-15 16:53:03 -0800998
999 if (PageTransCompound(page)) {
1000 err = split_huge_page(page);
1001 if (err)
1002 goto out_unlock;
1003 }
1004
Izik Eidus31dbd012009-09-21 17:02:03 -07001005 /*
1006 * If this anonymous page is mapped only here, its pte may need
1007 * to be write-protected. If it's mapped elsewhere, all of its
1008 * ptes are necessarily already write-protected. But in either
1009 * case, we need to lock and check page_count is not raised.
1010 */
Hugh Dickins80e148222009-12-14 17:59:29 -08001011 if (write_protect_page(vma, page, &orig_pte) == 0) {
1012 if (!kpage) {
1013 /*
1014 * While we hold page lock, upgrade page from
1015 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1016 * stable_tree_insert() will update stable_node.
1017 */
1018 set_page_stable_node(page, NULL);
1019 mark_page_accessed(page);
Minchan Kim337ed7e2016-01-15 16:55:15 -08001020 /*
1021 * Page reclaim just frees a clean page with no dirty
1022 * ptes: make sure that the ksm page would be swapped.
1023 */
1024 if (!PageDirty(page))
1025 SetPageDirty(page);
Hugh Dickins80e148222009-12-14 17:59:29 -08001026 err = 0;
1027 } else if (pages_identical(page, kpage))
1028 err = replace_page(vma, page, kpage, orig_pte);
1029 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001030
Hugh Dickins80e148222009-12-14 17:59:29 -08001031 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
Hugh Dickins73848b42009-12-14 17:59:22 -08001032 munlock_vma_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001033 if (!PageMlocked(kpage)) {
1034 unlock_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001035 lock_page(kpage);
1036 mlock_vma_page(kpage);
1037 page = kpage; /* for final unlock */
1038 }
1039 }
Hugh Dickins73848b42009-12-14 17:59:22 -08001040
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001041out_unlock:
Hugh Dickins8dd35572009-12-14 17:59:18 -08001042 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001043out:
1044 return err;
1045}
1046
1047/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001048 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1049 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -08001050 *
1051 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -07001052 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001053static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1054 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -07001055{
Hugh Dickins8dd35572009-12-14 17:59:18 -08001056 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -07001057 struct vm_area_struct *vma;
1058 int err = -EFAULT;
1059
Hugh Dickins8dd35572009-12-14 17:59:18 -08001060 down_read(&mm->mmap_sem);
Andrea Arcangeli85c6e8d2015-11-05 18:49:16 -08001061 vma = find_mergeable_vma(mm, rmap_item->address);
1062 if (!vma)
Hugh Dickins9ba69292009-09-21 17:02:20 -07001063 goto out;
1064
Hugh Dickins8dd35572009-12-14 17:59:18 -08001065 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001066 if (err)
1067 goto out;
1068
Hugh Dickinsbc566202013-02-22 16:36:06 -08001069 /* Unstable nid is in union with stable anon_vma: remove first */
1070 remove_rmap_item_from_tree(rmap_item);
1071
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001072 /* Must get reference to anon_vma while still holding mmap_sem */
Peter Zijlstra9e601092011-03-22 16:32:46 -07001073 rmap_item->anon_vma = vma->anon_vma;
1074 get_anon_vma(vma->anon_vma);
Hugh Dickins81464e302009-09-21 17:02:15 -07001075out:
Hugh Dickins8dd35572009-12-14 17:59:18 -08001076 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -07001077 return err;
1078}
1079
1080/*
Izik Eidus31dbd012009-09-21 17:02:03 -07001081 * try_to_merge_two_pages - take two identical pages and prepare them
1082 * to be merged into one page.
1083 *
Hugh Dickins8dd35572009-12-14 17:59:18 -08001084 * This function returns the kpage if we successfully merged two identical
1085 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001086 *
Hugh Dickins80e148222009-12-14 17:59:29 -08001087 * Note that this function upgrades page to ksm page: if one of the pages
Izik Eidus31dbd012009-09-21 17:02:03 -07001088 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1089 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001090static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1091 struct page *page,
1092 struct rmap_item *tree_rmap_item,
1093 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001094{
Hugh Dickins80e148222009-12-14 17:59:29 -08001095 int err;
Izik Eidus31dbd012009-09-21 17:02:03 -07001096
Hugh Dickins80e148222009-12-14 17:59:29 -08001097 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
Izik Eidus31dbd012009-09-21 17:02:03 -07001098 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001099 err = try_to_merge_with_ksm_page(tree_rmap_item,
Hugh Dickins80e148222009-12-14 17:59:29 -08001100 tree_page, page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001101 /*
Hugh Dickins81464e302009-09-21 17:02:15 -07001102 * If that fails, we have a ksm page with only one pte
1103 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -07001104 */
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001105 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -08001106 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001107 }
Hugh Dickins80e148222009-12-14 17:59:29 -08001108 return err ? NULL : page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001109}
1110
1111/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001112 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -07001113 *
1114 * This function checks if there is a page inside the stable tree
1115 * with identical content to the page that we are scanning right now.
1116 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001117 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -07001118 * NULL otherwise.
1119 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001120static struct page *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001121{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001122 int nid;
Hugh Dickinsef53d162013-02-22 16:36:12 -08001123 struct rb_root *root;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001124 struct rb_node **new;
1125 struct rb_node *parent;
1126 struct stable_node *stable_node;
1127 struct stable_node *page_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001128
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001129 page_node = page_stable_node(page);
1130 if (page_node && page_node->head != &migrate_nodes) {
1131 /* ksm page forked */
Hugh Dickins08beca42009-12-14 17:59:21 -08001132 get_page(page);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001133 return page;
Hugh Dickins08beca42009-12-14 17:59:21 -08001134 }
1135
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001136 nid = get_kpfn_nid(page_to_pfn(page));
Hugh Dickinsef53d162013-02-22 16:36:12 -08001137 root = root_stable_tree + nid;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001138again:
Hugh Dickinsef53d162013-02-22 16:36:12 -08001139 new = &root->rb_node;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001140 parent = NULL;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001141
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001142 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001143 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001144 int ret;
1145
Hugh Dickins08beca42009-12-14 17:59:21 -08001146 cond_resched();
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001147 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins8aafa6a2013-02-22 16:35:06 -08001148 tree_page = get_ksm_page(stable_node, false);
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001149 if (!tree_page) {
1150 /*
1151 * If we walked over a stale stable_node,
1152 * get_ksm_page() will call rb_erase() and it
1153 * may rebalance the tree from under us. So
1154 * restart the search from scratch. Returning
1155 * NULL would be safe too, but we'd generate
1156 * false negative insertions just because some
1157 * stable_node was stale.
1158 */
1159 goto again;
1160 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001161
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001162 ret = memcmp_pages(page, tree_page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001163 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001164
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001165 parent = *new;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001166 if (ret < 0)
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001167 new = &parent->rb_left;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001168 else if (ret > 0)
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001169 new = &parent->rb_right;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001170 else {
1171 /*
1172 * Lock and unlock the stable_node's page (which
1173 * might already have been migrated) so that page
1174 * migration is sure to notice its raised count.
1175 * It would be more elegant to return stable_node
1176 * than kpage, but that involves more changes.
1177 */
1178 tree_page = get_ksm_page(stable_node, true);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001179 if (tree_page) {
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001180 unlock_page(tree_page);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001181 if (get_kpfn_nid(stable_node->kpfn) !=
1182 NUMA(stable_node->nid)) {
1183 put_page(tree_page);
1184 goto replace;
1185 }
1186 return tree_page;
1187 }
1188 /*
1189 * There is now a place for page_node, but the tree may
1190 * have been rebalanced, so re-evaluate parent and new.
1191 */
1192 if (page_node)
1193 goto again;
1194 return NULL;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001195 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001196 }
1197
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001198 if (!page_node)
1199 return NULL;
1200
1201 list_del(&page_node->list);
1202 DO_NUMA(page_node->nid = nid);
1203 rb_link_node(&page_node->node, parent, new);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001204 rb_insert_color(&page_node->node, root);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001205 get_page(page);
1206 return page;
1207
1208replace:
1209 if (page_node) {
1210 list_del(&page_node->list);
1211 DO_NUMA(page_node->nid = nid);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001212 rb_replace_node(&stable_node->node, &page_node->node, root);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001213 get_page(page);
1214 } else {
Hugh Dickinsef53d162013-02-22 16:36:12 -08001215 rb_erase(&stable_node->node, root);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001216 page = NULL;
1217 }
1218 stable_node->head = &migrate_nodes;
1219 list_add(&stable_node->list, stable_node->head);
1220 return page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001221}
1222
1223/*
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001224 * stable_tree_insert - insert stable tree node pointing to new ksm page
Izik Eidus31dbd012009-09-21 17:02:03 -07001225 * into the stable tree.
1226 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001227 * This function returns the stable tree node just allocated on success,
1228 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001229 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001230static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001231{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001232 int nid;
1233 unsigned long kpfn;
Hugh Dickinsef53d162013-02-22 16:36:12 -08001234 struct rb_root *root;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001235 struct rb_node **new;
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001236 struct rb_node *parent;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001237 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001238
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001239 kpfn = page_to_pfn(kpage);
1240 nid = get_kpfn_nid(kpfn);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001241 root = root_stable_tree + nid;
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001242again:
1243 parent = NULL;
Hugh Dickinsef53d162013-02-22 16:36:12 -08001244 new = &root->rb_node;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001245
Izik Eidus31dbd012009-09-21 17:02:03 -07001246 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001247 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001248 int ret;
1249
Hugh Dickins08beca42009-12-14 17:59:21 -08001250 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001251 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins8aafa6a2013-02-22 16:35:06 -08001252 tree_page = get_ksm_page(stable_node, false);
Andrea Arcangelif2e5ff82015-11-05 18:49:10 -08001253 if (!tree_page) {
1254 /*
1255 * If we walked over a stale stable_node,
1256 * get_ksm_page() will call rb_erase() and it
1257 * may rebalance the tree from under us. So
1258 * restart the search from scratch. Returning
1259 * NULL would be safe too, but we'd generate
1260 * false negative insertions just because some
1261 * stable_node was stale.
1262 */
1263 goto again;
1264 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001265
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001266 ret = memcmp_pages(kpage, tree_page);
1267 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001268
1269 parent = *new;
1270 if (ret < 0)
1271 new = &parent->rb_left;
1272 else if (ret > 0)
1273 new = &parent->rb_right;
1274 else {
1275 /*
1276 * It is not a bug that stable_tree_search() didn't
1277 * find this node: because at that time our page was
1278 * not yet write-protected, so may have changed since.
1279 */
1280 return NULL;
1281 }
1282 }
1283
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001284 stable_node = alloc_stable_node();
1285 if (!stable_node)
1286 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001287
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001288 INIT_HLIST_HEAD(&stable_node->hlist);
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001289 stable_node->kpfn = kpfn;
Hugh Dickins08beca42009-12-14 17:59:21 -08001290 set_page_stable_node(kpage, stable_node);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001291 DO_NUMA(stable_node->nid = nid);
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001292 rb_link_node(&stable_node->node, parent, new);
Hugh Dickinsef53d162013-02-22 16:36:12 -08001293 rb_insert_color(&stable_node->node, root);
Hugh Dickins08beca42009-12-14 17:59:21 -08001294
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001295 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001296}
1297
1298/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001299 * unstable_tree_search_insert - search for identical page,
1300 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001301 *
1302 * This function searches for a page in the unstable tree identical to the
1303 * page currently being scanned; and if no identical page is found in the
1304 * tree, we insert rmap_item as a new object into the unstable tree.
1305 *
1306 * This function returns pointer to rmap_item found to be identical
1307 * to the currently scanned page, NULL otherwise.
1308 *
1309 * This function does both searching and inserting, because they share
1310 * the same walking algorithm in an rbtree.
1311 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001312static
1313struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1314 struct page *page,
1315 struct page **tree_pagep)
Izik Eidus31dbd012009-09-21 17:02:03 -07001316{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001317 struct rb_node **new;
1318 struct rb_root *root;
Izik Eidus31dbd012009-09-21 17:02:03 -07001319 struct rb_node *parent = NULL;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001320 int nid;
1321
1322 nid = get_kpfn_nid(page_to_pfn(page));
Hugh Dickinsef53d162013-02-22 16:36:12 -08001323 root = root_unstable_tree + nid;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001324 new = &root->rb_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001325
1326 while (*new) {
1327 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001328 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001329 int ret;
1330
Hugh Dickinsd178f272009-11-09 15:58:23 +00001331 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -07001332 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001333 tree_page = get_mergeable_page(tree_rmap_item);
Andrea Arcangelic8f95ed2015-11-05 18:49:19 -08001334 if (!tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001335 return NULL;
1336
1337 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001338 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -07001339 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001340 if (page == tree_page) {
1341 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001342 return NULL;
1343 }
1344
Hugh Dickins8dd35572009-12-14 17:59:18 -08001345 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001346
1347 parent = *new;
1348 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001349 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001350 new = &parent->rb_left;
1351 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001352 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001353 new = &parent->rb_right;
Hugh Dickinsb599cbd2013-02-22 16:36:05 -08001354 } else if (!ksm_merge_across_nodes &&
1355 page_to_nid(tree_page) != nid) {
1356 /*
1357 * If tree_page has been migrated to another NUMA node,
1358 * it will be flushed out and put in the right unstable
1359 * tree next time: only merge with it when across_nodes.
1360 */
1361 put_page(tree_page);
1362 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001363 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001364 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001365 return tree_rmap_item;
1366 }
1367 }
1368
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001369 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001370 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001371 DO_NUMA(rmap_item->nid = nid);
Izik Eidus31dbd012009-09-21 17:02:03 -07001372 rb_link_node(&rmap_item->node, parent, new);
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001373 rb_insert_color(&rmap_item->node, root);
Izik Eidus31dbd012009-09-21 17:02:03 -07001374
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001375 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001376 return NULL;
1377}
1378
1379/*
1380 * stable_tree_append - add another rmap_item to the linked list of
1381 * rmap_items hanging off a given node of the stable tree, all sharing
1382 * the same ksm page.
1383 */
1384static void stable_tree_append(struct rmap_item *rmap_item,
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001385 struct stable_node *stable_node)
Izik Eidus31dbd012009-09-21 17:02:03 -07001386{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001387 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001388 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001389 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001390
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001391 if (rmap_item->hlist.next)
1392 ksm_pages_sharing++;
1393 else
1394 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001395}
1396
1397/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001398 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1399 * if not, compare checksum to previous and if it's the same, see if page can
1400 * be inserted into the unstable tree, or merged with a page already there and
1401 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001402 *
1403 * @page: the page that we are searching identical page to.
1404 * @rmap_item: the reverse mapping into the virtual address of this page
1405 */
1406static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1407{
Izik Eidus31dbd012009-09-21 17:02:03 -07001408 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001409 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001410 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001411 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07001412 unsigned int checksum;
1413 int err;
1414
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001415 stable_node = page_stable_node(page);
1416 if (stable_node) {
1417 if (stable_node->head != &migrate_nodes &&
1418 get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1419 rb_erase(&stable_node->node,
Hugh Dickinsef53d162013-02-22 16:36:12 -08001420 root_stable_tree + NUMA(stable_node->nid));
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001421 stable_node->head = &migrate_nodes;
1422 list_add(&stable_node->list, stable_node->head);
1423 }
1424 if (stable_node->head != &migrate_nodes &&
1425 rmap_item->head == stable_node)
1426 return;
1427 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001428
1429 /* We first start with searching the page inside the stable tree */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001430 kpage = stable_tree_search(page);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001431 if (kpage == page && rmap_item->head == stable_node) {
1432 put_page(kpage);
1433 return;
1434 }
1435
1436 remove_rmap_item_from_tree(rmap_item);
1437
Hugh Dickins62b61f62009-12-14 17:59:33 -08001438 if (kpage) {
Hugh Dickins08beca42009-12-14 17:59:21 -08001439 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001440 if (!err) {
1441 /*
1442 * The page was successfully merged:
1443 * add its rmap_item to the stable tree.
1444 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001445 lock_page(kpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001446 stable_tree_append(rmap_item, page_stable_node(kpage));
Hugh Dickins5ad64682009-12-14 17:59:24 -08001447 unlock_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001448 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001449 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001450 return;
1451 }
1452
1453 /*
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001454 * If the hash value of the page has changed from the last time
1455 * we calculated it, this page is changing frequently: therefore we
1456 * don't want to insert it in the unstable tree, and we don't want
1457 * to waste our time searching for something identical to it there.
Izik Eidus31dbd012009-09-21 17:02:03 -07001458 */
1459 checksum = calc_checksum(page);
1460 if (rmap_item->oldchecksum != checksum) {
1461 rmap_item->oldchecksum = checksum;
1462 return;
1463 }
1464
Hugh Dickins8dd35572009-12-14 17:59:18 -08001465 tree_rmap_item =
1466 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001467 if (tree_rmap_item) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001468 kpage = try_to_merge_two_pages(rmap_item, page,
1469 tree_rmap_item, tree_page);
1470 put_page(tree_page);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001471 if (kpage) {
Hugh Dickinsbc566202013-02-22 16:36:06 -08001472 /*
1473 * The pages were successfully merged: insert new
1474 * node in the stable tree and add both rmap_items.
1475 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001476 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001477 stable_node = stable_tree_insert(kpage);
1478 if (stable_node) {
1479 stable_tree_append(tree_rmap_item, stable_node);
1480 stable_tree_append(rmap_item, stable_node);
1481 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001482 unlock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001483
Izik Eidus31dbd012009-09-21 17:02:03 -07001484 /*
1485 * If we fail to insert the page into the stable tree,
1486 * we will have 2 virtual addresses that are pointing
1487 * to a ksm page left outside the stable tree,
1488 * in which case we need to break_cow on both.
1489 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001490 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001491 break_cow(tree_rmap_item);
1492 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001493 }
1494 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001495 }
1496}
1497
1498static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001499 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07001500 unsigned long addr)
1501{
1502 struct rmap_item *rmap_item;
1503
Hugh Dickins6514d512009-12-14 17:59:19 -08001504 while (*rmap_list) {
1505 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08001506 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07001507 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001508 if (rmap_item->address > addr)
1509 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08001510 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001511 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001512 free_rmap_item(rmap_item);
1513 }
1514
1515 rmap_item = alloc_rmap_item();
1516 if (rmap_item) {
1517 /* It has already been zeroed */
1518 rmap_item->mm = mm_slot->mm;
1519 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08001520 rmap_item->rmap_list = *rmap_list;
1521 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001522 }
1523 return rmap_item;
1524}
1525
1526static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1527{
1528 struct mm_struct *mm;
1529 struct mm_slot *slot;
1530 struct vm_area_struct *vma;
1531 struct rmap_item *rmap_item;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001532 int nid;
Izik Eidus31dbd012009-09-21 17:02:03 -07001533
1534 if (list_empty(&ksm_mm_head.mm_list))
1535 return NULL;
1536
1537 slot = ksm_scan.mm_slot;
1538 if (slot == &ksm_mm_head) {
Hugh Dickins2919bfd2011-01-13 15:47:29 -08001539 /*
1540 * A number of pages can hang around indefinitely on per-cpu
1541 * pagevecs, raised page count preventing write_protect_page
1542 * from merging them. Though it doesn't really matter much,
1543 * it is puzzling to see some stuck in pages_volatile until
1544 * other activity jostles them out, and they also prevented
1545 * LTP's KSM test from succeeding deterministically; so drain
1546 * them here (here rather than on entry to ksm_do_scan(),
1547 * so we don't IPI too often when pages_to_scan is set low).
1548 */
1549 lru_add_drain_all();
1550
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001551 /*
1552 * Whereas stale stable_nodes on the stable_tree itself
1553 * get pruned in the regular course of stable_tree_search(),
1554 * those moved out to the migrate_nodes list can accumulate:
1555 * so prune them once before each full scan.
1556 */
1557 if (!ksm_merge_across_nodes) {
Geliang Tang03640412016-01-14 15:20:54 -08001558 struct stable_node *stable_node, *next;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001559 struct page *page;
1560
Geliang Tang03640412016-01-14 15:20:54 -08001561 list_for_each_entry_safe(stable_node, next,
1562 &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001563 page = get_ksm_page(stable_node, false);
1564 if (page)
1565 put_page(page);
1566 cond_resched();
1567 }
1568 }
1569
Hugh Dickinsef53d162013-02-22 16:36:12 -08001570 for (nid = 0; nid < ksm_nr_node_ids; nid++)
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001571 root_unstable_tree[nid] = RB_ROOT;
Izik Eidus31dbd012009-09-21 17:02:03 -07001572
1573 spin_lock(&ksm_mmlist_lock);
1574 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1575 ksm_scan.mm_slot = slot;
1576 spin_unlock(&ksm_mmlist_lock);
Hugh Dickins2b472612011-06-15 15:08:58 -07001577 /*
1578 * Although we tested list_empty() above, a racing __ksm_exit
1579 * of the last mm on the list may have removed it since then.
1580 */
1581 if (slot == &ksm_mm_head)
1582 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001583next_mm:
1584 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001585 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001586 }
1587
1588 mm = slot->mm;
1589 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001590 if (ksm_test_exit(mm))
1591 vma = NULL;
1592 else
1593 vma = find_vma(mm, ksm_scan.address);
1594
1595 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001596 if (!(vma->vm_flags & VM_MERGEABLE))
1597 continue;
1598 if (ksm_scan.address < vma->vm_start)
1599 ksm_scan.address = vma->vm_start;
1600 if (!vma->anon_vma)
1601 ksm_scan.address = vma->vm_end;
1602
1603 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001604 if (ksm_test_exit(mm))
1605 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001606 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001607 if (IS_ERR_OR_NULL(*page)) {
1608 ksm_scan.address += PAGE_SIZE;
1609 cond_resched();
1610 continue;
1611 }
Kirill A. Shutemovf765f542016-01-15 16:53:03 -08001612 if (PageAnon(*page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001613 flush_anon_page(vma, *page, ksm_scan.address);
1614 flush_dcache_page(*page);
1615 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001616 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07001617 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001618 ksm_scan.rmap_list =
1619 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001620 ksm_scan.address += PAGE_SIZE;
1621 } else
1622 put_page(*page);
1623 up_read(&mm->mmap_sem);
1624 return rmap_item;
1625 }
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001626 put_page(*page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001627 ksm_scan.address += PAGE_SIZE;
1628 cond_resched();
1629 }
1630 }
1631
Hugh Dickins9ba69292009-09-21 17:02:20 -07001632 if (ksm_test_exit(mm)) {
1633 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001634 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001635 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001636 /*
1637 * Nuke all the rmap_items that are above this current rmap:
1638 * because there were no VM_MERGEABLE vmas with such addresses.
1639 */
Hugh Dickins6514d512009-12-14 17:59:19 -08001640 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001641
1642 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001643 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1644 struct mm_slot, mm_list);
1645 if (ksm_scan.address == 0) {
1646 /*
1647 * We've completed a full scan of all vmas, holding mmap_sem
1648 * throughout, and found no VM_MERGEABLE: so do the same as
1649 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07001650 * This applies either when cleaning up after __ksm_exit
1651 * (but beware: we can reach here even before __ksm_exit),
1652 * or when all VM_MERGEABLE areas have been unmapped (and
1653 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07001654 */
Sasha Levin4ca3a692013-02-22 16:32:28 -08001655 hash_del(&slot->link);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001656 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001657 spin_unlock(&ksm_mmlist_lock);
1658
Hugh Dickinscd551f92009-09-21 17:02:17 -07001659 free_mm_slot(slot);
1660 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001661 up_read(&mm->mmap_sem);
1662 mmdrop(mm);
1663 } else {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001664 up_read(&mm->mmap_sem);
Zhou Chengming7496fea2016-05-12 15:42:21 -07001665 /*
1666 * up_read(&mm->mmap_sem) first because after
1667 * spin_unlock(&ksm_mmlist_lock) run, the "mm" may
1668 * already have been freed under us by __ksm_exit()
1669 * because the "mm_slot" is still hashed and
1670 * ksm_scan.mm_slot doesn't point to it anymore.
1671 */
1672 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001673 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001674
1675 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07001676 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07001677 if (slot != &ksm_mm_head)
1678 goto next_mm;
1679
Izik Eidus31dbd012009-09-21 17:02:03 -07001680 ksm_scan.seqnr++;
1681 return NULL;
1682}
1683
1684/**
1685 * ksm_do_scan - the ksm scanner main worker function.
1686 * @scan_npages - number of pages we want to scan before we return.
1687 */
1688static void ksm_do_scan(unsigned int scan_npages)
1689{
1690 struct rmap_item *rmap_item;
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001691 struct page *uninitialized_var(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001692
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001693 while (scan_npages-- && likely(!freezing(current))) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001694 cond_resched();
1695 rmap_item = scan_get_next_rmap_item(&page);
1696 if (!rmap_item)
1697 return;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001698 cmp_and_merge_page(page, rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001699 put_page(page);
1700 }
1701}
1702
Hugh Dickins6e1583842009-09-21 17:02:14 -07001703static int ksmd_should_run(void)
1704{
1705 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1706}
1707
Izik Eidus31dbd012009-09-21 17:02:03 -07001708static int ksm_scan_thread(void *nothing)
1709{
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001710 set_freezable();
Izik Eidus339aa622009-09-21 17:02:07 -07001711 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07001712
1713 while (!kthread_should_stop()) {
Hugh Dickins6e1583842009-09-21 17:02:14 -07001714 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08001715 wait_while_offlining();
Hugh Dickins6e1583842009-09-21 17:02:14 -07001716 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07001717 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001718 mutex_unlock(&ksm_thread_mutex);
1719
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001720 try_to_freeze();
1721
Hugh Dickins6e1583842009-09-21 17:02:14 -07001722 if (ksmd_should_run()) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001723 schedule_timeout_interruptible(
1724 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1725 } else {
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001726 wait_event_freezable(ksm_thread_wait,
Hugh Dickins6e1583842009-09-21 17:02:14 -07001727 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07001728 }
1729 }
1730 return 0;
1731}
1732
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001733int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1734 unsigned long end, int advice, unsigned long *vm_flags)
1735{
1736 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001737 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001738
1739 switch (advice) {
1740 case MADV_MERGEABLE:
1741 /*
1742 * Be somewhat over-protective for now!
1743 */
1744 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1745 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
Kirill A. Shutemov0661a332015-02-10 14:10:04 -08001746 VM_HUGETLB | VM_MIXEDMAP))
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001747 return 0; /* just ignore the advice */
1748
Konstantin Khlebnikovcc2383e2012-10-08 16:28:37 -07001749#ifdef VM_SAO
1750 if (*vm_flags & VM_SAO)
1751 return 0;
1752#endif
1753
Hugh Dickinsd952b792009-09-21 17:02:16 -07001754 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1755 err = __ksm_enter(mm);
1756 if (err)
1757 return err;
1758 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001759
1760 *vm_flags |= VM_MERGEABLE;
1761 break;
1762
1763 case MADV_UNMERGEABLE:
1764 if (!(*vm_flags & VM_MERGEABLE))
1765 return 0; /* just ignore the advice */
1766
Hugh Dickinsd952b792009-09-21 17:02:16 -07001767 if (vma->anon_vma) {
1768 err = unmerge_ksm_pages(vma, start, end);
1769 if (err)
1770 return err;
1771 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001772
1773 *vm_flags &= ~VM_MERGEABLE;
1774 break;
1775 }
1776
1777 return 0;
1778}
1779
1780int __ksm_enter(struct mm_struct *mm)
1781{
Hugh Dickins6e1583842009-09-21 17:02:14 -07001782 struct mm_slot *mm_slot;
1783 int needs_wakeup;
1784
1785 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07001786 if (!mm_slot)
1787 return -ENOMEM;
1788
Hugh Dickins6e1583842009-09-21 17:02:14 -07001789 /* Check ksm_run too? Would need tighter locking */
1790 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1791
Izik Eidus31dbd012009-09-21 17:02:03 -07001792 spin_lock(&ksm_mmlist_lock);
1793 insert_to_mm_slots_hash(mm, mm_slot);
1794 /*
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001795 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1796 * insert just behind the scanning cursor, to let the area settle
Izik Eidus31dbd012009-09-21 17:02:03 -07001797 * down a little; when fork is followed by immediate exec, we don't
1798 * want ksmd to waste time setting up and tearing down an rmap_list.
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001799 *
1800 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1801 * scanning cursor, otherwise KSM pages in newly forked mms will be
1802 * missed: then we might as well insert at the end of the list.
Izik Eidus31dbd012009-09-21 17:02:03 -07001803 */
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001804 if (ksm_run & KSM_RUN_UNMERGE)
1805 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1806 else
1807 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001808 spin_unlock(&ksm_mmlist_lock);
1809
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001810 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001811 atomic_inc(&mm->mm_count);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001812
1813 if (needs_wakeup)
1814 wake_up_interruptible(&ksm_thread_wait);
1815
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001816 return 0;
1817}
1818
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07001819void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001820{
Hugh Dickinscd551f92009-09-21 17:02:17 -07001821 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001822 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07001823
Izik Eidus31dbd012009-09-21 17:02:03 -07001824 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07001825 * This process is exiting: if it's straightforward (as is the
1826 * case when ksmd was never running), free mm_slot immediately.
1827 * But if it's at the cursor or has rmap_items linked to it, use
1828 * mmap_sem to synchronize with any break_cows before pagetables
1829 * are freed, and leave the mm_slot on the list for ksmd to free.
1830 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07001831 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07001832
Hugh Dickinscd551f92009-09-21 17:02:17 -07001833 spin_lock(&ksm_mmlist_lock);
1834 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001835 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001836 if (!mm_slot->rmap_list) {
Sasha Levin4ca3a692013-02-22 16:32:28 -08001837 hash_del(&mm_slot->link);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001838 list_del(&mm_slot->mm_list);
1839 easy_to_free = 1;
1840 } else {
1841 list_move(&mm_slot->mm_list,
1842 &ksm_scan.mm_slot->mm_list);
1843 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001844 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001845 spin_unlock(&ksm_mmlist_lock);
1846
Hugh Dickins9ba69292009-09-21 17:02:20 -07001847 if (easy_to_free) {
1848 free_mm_slot(mm_slot);
1849 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1850 mmdrop(mm);
1851 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001852 down_write(&mm->mmap_sem);
1853 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001854 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001855}
Izik Eidus31dbd012009-09-21 17:02:03 -07001856
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001857struct page *ksm_might_need_to_copy(struct page *page,
Hugh Dickins5ad64682009-12-14 17:59:24 -08001858 struct vm_area_struct *vma, unsigned long address)
1859{
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001860 struct anon_vma *anon_vma = page_anon_vma(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001861 struct page *new_page;
1862
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001863 if (PageKsm(page)) {
1864 if (page_stable_node(page) &&
1865 !(ksm_run & KSM_RUN_UNMERGE))
1866 return page; /* no need to copy it */
1867 } else if (!anon_vma) {
1868 return page; /* no need to copy it */
1869 } else if (anon_vma->root == vma->anon_vma->root &&
1870 page->index == linear_page_index(vma, address)) {
1871 return page; /* still no need to copy it */
1872 }
1873 if (!PageUptodate(page))
1874 return page; /* let do_swap_page report the error */
1875
Hugh Dickins5ad64682009-12-14 17:59:24 -08001876 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1877 if (new_page) {
1878 copy_user_highpage(new_page, page, address, vma);
1879
1880 SetPageDirty(new_page);
1881 __SetPageUptodate(new_page);
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001882 __SetPageLocked(new_page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001883 }
1884
Hugh Dickins5ad64682009-12-14 17:59:24 -08001885 return new_page;
1886}
1887
Joonsoo Kim051ac832014-01-21 15:49:48 -08001888int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc)
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001889{
1890 struct stable_node *stable_node;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001891 struct rmap_item *rmap_item;
1892 int ret = SWAP_AGAIN;
1893 int search_new_forks = 0;
1894
Sasha Levin309381fea2014-01-23 15:52:54 -08001895 VM_BUG_ON_PAGE(!PageKsm(page), page);
Joonsoo Kim9f326242014-01-21 15:49:53 -08001896
1897 /*
1898 * Rely on the page lock to protect against concurrent modifications
1899 * to that page's node of the stable tree.
1900 */
Sasha Levin309381fea2014-01-23 15:52:54 -08001901 VM_BUG_ON_PAGE(!PageLocked(page), page);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001902
1903 stable_node = page_stable_node(page);
1904 if (!stable_node)
1905 return ret;
1906again:
Sasha Levinb67bfe02013-02-27 17:06:00 -08001907 hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) {
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001908 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001909 struct anon_vma_chain *vmac;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001910 struct vm_area_struct *vma;
1911
Andrea Arcangeliad126952015-11-05 18:49:07 -08001912 cond_resched();
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001913 anon_vma_lock_read(anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -07001914 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1915 0, ULONG_MAX) {
Andrea Arcangeliad126952015-11-05 18:49:07 -08001916 cond_resched();
Rik van Riel5beb4932010-03-05 13:42:07 -08001917 vma = vmac->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001918 if (rmap_item->address < vma->vm_start ||
1919 rmap_item->address >= vma->vm_end)
1920 continue;
1921 /*
1922 * Initially we examine only the vma which covers this
1923 * rmap_item; but later, if there is still work to do,
1924 * we examine covering vmas in other mms: in case they
1925 * were forked from the original since ksmd passed.
1926 */
1927 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1928 continue;
1929
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001930 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1931 continue;
1932
Joonsoo Kim051ac832014-01-21 15:49:48 -08001933 ret = rwc->rmap_one(page, vma,
1934 rmap_item->address, rwc->arg);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001935 if (ret != SWAP_AGAIN) {
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001936 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001937 goto out;
1938 }
Joonsoo Kim0dd1c7b2014-01-21 15:49:49 -08001939 if (rwc->done && rwc->done(page)) {
1940 anon_vma_unlock_read(anon_vma);
1941 goto out;
1942 }
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001943 }
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001944 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001945 }
1946 if (!search_new_forks++)
1947 goto again;
1948out:
1949 return ret;
1950}
1951
Joonsoo Kim52629502014-01-21 15:49:50 -08001952#ifdef CONFIG_MIGRATION
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001953void ksm_migrate_page(struct page *newpage, struct page *oldpage)
1954{
1955 struct stable_node *stable_node;
1956
Sasha Levin309381fea2014-01-23 15:52:54 -08001957 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
1958 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1959 VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001960
1961 stable_node = page_stable_node(newpage);
1962 if (stable_node) {
Sasha Levin309381fea2014-01-23 15:52:54 -08001963 VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001964 stable_node->kpfn = page_to_pfn(newpage);
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001965 /*
1966 * newpage->mapping was set in advance; now we need smp_wmb()
1967 * to make sure that the new stable_node->kpfn is visible
1968 * to get_ksm_page() before it can see that oldpage->mapping
1969 * has gone stale (or that PageSwapCache has been cleared).
1970 */
1971 smp_wmb();
1972 set_page_stable_node(oldpage, NULL);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001973 }
1974}
1975#endif /* CONFIG_MIGRATION */
1976
Hugh Dickins62b61f62009-12-14 17:59:33 -08001977#ifdef CONFIG_MEMORY_HOTREMOVE
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08001978static void wait_while_offlining(void)
1979{
1980 while (ksm_run & KSM_RUN_OFFLINE) {
1981 mutex_unlock(&ksm_thread_mutex);
1982 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
NeilBrown74316202014-07-07 15:16:04 +10001983 TASK_UNINTERRUPTIBLE);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08001984 mutex_lock(&ksm_thread_mutex);
1985 }
1986}
1987
Hugh Dickinsee0ea592013-02-22 16:35:05 -08001988static void ksm_check_stable_tree(unsigned long start_pfn,
1989 unsigned long end_pfn)
Hugh Dickins62b61f62009-12-14 17:59:33 -08001990{
Geliang Tang03640412016-01-14 15:20:54 -08001991 struct stable_node *stable_node, *next;
Hugh Dickins62b61f62009-12-14 17:59:33 -08001992 struct rb_node *node;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001993 int nid;
Hugh Dickins62b61f62009-12-14 17:59:33 -08001994
Hugh Dickinsef53d162013-02-22 16:36:12 -08001995 for (nid = 0; nid < ksm_nr_node_ids; nid++) {
1996 node = rb_first(root_stable_tree + nid);
Hugh Dickinsee0ea592013-02-22 16:35:05 -08001997 while (node) {
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001998 stable_node = rb_entry(node, struct stable_node, node);
1999 if (stable_node->kpfn >= start_pfn &&
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002000 stable_node->kpfn < end_pfn) {
2001 /*
2002 * Don't get_ksm_page, page has already gone:
2003 * which is why we keep kpfn instead of page*
2004 */
2005 remove_node_from_stable_tree(stable_node);
Hugh Dickinsef53d162013-02-22 16:36:12 -08002006 node = rb_first(root_stable_tree + nid);
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002007 } else
2008 node = rb_next(node);
2009 cond_resched();
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002010 }
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002011 }
Geliang Tang03640412016-01-14 15:20:54 -08002012 list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) {
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002013 if (stable_node->kpfn >= start_pfn &&
2014 stable_node->kpfn < end_pfn)
2015 remove_node_from_stable_tree(stable_node);
2016 cond_resched();
2017 }
Hugh Dickins62b61f62009-12-14 17:59:33 -08002018}
2019
2020static int ksm_memory_callback(struct notifier_block *self,
2021 unsigned long action, void *arg)
2022{
2023 struct memory_notify *mn = arg;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002024
2025 switch (action) {
2026 case MEM_GOING_OFFLINE:
2027 /*
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002028 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2029 * and remove_all_stable_nodes() while memory is going offline:
2030 * it is unsafe for them to touch the stable tree at this time.
2031 * But unmerge_ksm_pages(), rmap lookups and other entry points
2032 * which do not need the ksm_thread_mutex are all safe.
Hugh Dickins62b61f62009-12-14 17:59:33 -08002033 */
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002034 mutex_lock(&ksm_thread_mutex);
2035 ksm_run |= KSM_RUN_OFFLINE;
2036 mutex_unlock(&ksm_thread_mutex);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002037 break;
2038
2039 case MEM_OFFLINE:
2040 /*
2041 * Most of the work is done by page migration; but there might
2042 * be a few stable_nodes left over, still pointing to struct
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002043 * pages which have been offlined: prune those from the tree,
2044 * otherwise get_ksm_page() might later try to access a
2045 * non-existent struct page.
Hugh Dickins62b61f62009-12-14 17:59:33 -08002046 */
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002047 ksm_check_stable_tree(mn->start_pfn,
2048 mn->start_pfn + mn->nr_pages);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002049 /* fallthrough */
2050
2051 case MEM_CANCEL_OFFLINE:
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002052 mutex_lock(&ksm_thread_mutex);
2053 ksm_run &= ~KSM_RUN_OFFLINE;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002054 mutex_unlock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002055
2056 smp_mb(); /* wake_up_bit advises this */
2057 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
Hugh Dickins62b61f62009-12-14 17:59:33 -08002058 break;
2059 }
2060 return NOTIFY_OK;
2061}
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002062#else
2063static void wait_while_offlining(void)
2064{
2065}
Hugh Dickins62b61f62009-12-14 17:59:33 -08002066#endif /* CONFIG_MEMORY_HOTREMOVE */
2067
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002068#ifdef CONFIG_SYSFS
2069/*
2070 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2071 */
2072
Izik Eidus31dbd012009-09-21 17:02:03 -07002073#define KSM_ATTR_RO(_name) \
2074 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2075#define KSM_ATTR(_name) \
2076 static struct kobj_attribute _name##_attr = \
2077 __ATTR(_name, 0644, _name##_show, _name##_store)
2078
2079static ssize_t sleep_millisecs_show(struct kobject *kobj,
2080 struct kobj_attribute *attr, char *buf)
2081{
2082 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2083}
2084
2085static ssize_t sleep_millisecs_store(struct kobject *kobj,
2086 struct kobj_attribute *attr,
2087 const char *buf, size_t count)
2088{
2089 unsigned long msecs;
2090 int err;
2091
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002092 err = kstrtoul(buf, 10, &msecs);
Izik Eidus31dbd012009-09-21 17:02:03 -07002093 if (err || msecs > UINT_MAX)
2094 return -EINVAL;
2095
2096 ksm_thread_sleep_millisecs = msecs;
2097
2098 return count;
2099}
2100KSM_ATTR(sleep_millisecs);
2101
2102static ssize_t pages_to_scan_show(struct kobject *kobj,
2103 struct kobj_attribute *attr, char *buf)
2104{
2105 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2106}
2107
2108static ssize_t pages_to_scan_store(struct kobject *kobj,
2109 struct kobj_attribute *attr,
2110 const char *buf, size_t count)
2111{
2112 int err;
2113 unsigned long nr_pages;
2114
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002115 err = kstrtoul(buf, 10, &nr_pages);
Izik Eidus31dbd012009-09-21 17:02:03 -07002116 if (err || nr_pages > UINT_MAX)
2117 return -EINVAL;
2118
2119 ksm_thread_pages_to_scan = nr_pages;
2120
2121 return count;
2122}
2123KSM_ATTR(pages_to_scan);
2124
2125static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2126 char *buf)
2127{
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002128 return sprintf(buf, "%lu\n", ksm_run);
Izik Eidus31dbd012009-09-21 17:02:03 -07002129}
2130
2131static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2132 const char *buf, size_t count)
2133{
2134 int err;
2135 unsigned long flags;
2136
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002137 err = kstrtoul(buf, 10, &flags);
Izik Eidus31dbd012009-09-21 17:02:03 -07002138 if (err || flags > UINT_MAX)
2139 return -EINVAL;
2140 if (flags > KSM_RUN_UNMERGE)
2141 return -EINVAL;
2142
2143 /*
2144 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2145 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsd0f209f2009-12-14 17:59:34 -08002146 * breaking COW to free the pages_shared (but leaves mm_slots
2147 * on the list for when ksmd may be set running again).
Izik Eidus31dbd012009-09-21 17:02:03 -07002148 */
2149
2150 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002151 wait_while_offlining();
Izik Eidus31dbd012009-09-21 17:02:03 -07002152 if (ksm_run != flags) {
2153 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07002154 if (flags & KSM_RUN_UNMERGE) {
David Rientjese1e12d22012-12-11 16:02:56 -08002155 set_current_oom_origin();
Hugh Dickinsd952b792009-09-21 17:02:16 -07002156 err = unmerge_and_remove_all_rmap_items();
David Rientjese1e12d22012-12-11 16:02:56 -08002157 clear_current_oom_origin();
Hugh Dickinsd952b792009-09-21 17:02:16 -07002158 if (err) {
2159 ksm_run = KSM_RUN_STOP;
2160 count = err;
2161 }
2162 }
Izik Eidus31dbd012009-09-21 17:02:03 -07002163 }
2164 mutex_unlock(&ksm_thread_mutex);
2165
2166 if (flags & KSM_RUN_MERGE)
2167 wake_up_interruptible(&ksm_thread_wait);
2168
2169 return count;
2170}
2171KSM_ATTR(run);
2172
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002173#ifdef CONFIG_NUMA
2174static ssize_t merge_across_nodes_show(struct kobject *kobj,
2175 struct kobj_attribute *attr, char *buf)
2176{
2177 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2178}
2179
2180static ssize_t merge_across_nodes_store(struct kobject *kobj,
2181 struct kobj_attribute *attr,
2182 const char *buf, size_t count)
2183{
2184 int err;
2185 unsigned long knob;
2186
2187 err = kstrtoul(buf, 10, &knob);
2188 if (err)
2189 return err;
2190 if (knob > 1)
2191 return -EINVAL;
2192
2193 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002194 wait_while_offlining();
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002195 if (ksm_merge_across_nodes != knob) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002196 if (ksm_pages_shared || remove_all_stable_nodes())
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002197 err = -EBUSY;
Hugh Dickinsef53d162013-02-22 16:36:12 -08002198 else if (root_stable_tree == one_stable_tree) {
2199 struct rb_root *buf;
2200 /*
2201 * This is the first time that we switch away from the
2202 * default of merging across nodes: must now allocate
2203 * a buffer to hold as many roots as may be needed.
2204 * Allocate stable and unstable together:
2205 * MAXSMP NODES_SHIFT 10 will use 16kB.
2206 */
Joe Perchesbafe1e12013-11-12 15:07:10 -08002207 buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf),
2208 GFP_KERNEL);
Hugh Dickinsef53d162013-02-22 16:36:12 -08002209 /* Let us assume that RB_ROOT is NULL is zero */
2210 if (!buf)
2211 err = -ENOMEM;
2212 else {
2213 root_stable_tree = buf;
2214 root_unstable_tree = buf + nr_node_ids;
2215 /* Stable tree is empty but not the unstable */
2216 root_unstable_tree[0] = one_unstable_tree[0];
2217 }
2218 }
2219 if (!err) {
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002220 ksm_merge_across_nodes = knob;
Hugh Dickinsef53d162013-02-22 16:36:12 -08002221 ksm_nr_node_ids = knob ? 1 : nr_node_ids;
2222 }
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002223 }
2224 mutex_unlock(&ksm_thread_mutex);
2225
2226 return err ? err : count;
2227}
2228KSM_ATTR(merge_across_nodes);
2229#endif
2230
Hugh Dickinsb4028262009-09-21 17:02:09 -07002231static ssize_t pages_shared_show(struct kobject *kobj,
2232 struct kobj_attribute *attr, char *buf)
2233{
2234 return sprintf(buf, "%lu\n", ksm_pages_shared);
2235}
2236KSM_ATTR_RO(pages_shared);
2237
2238static ssize_t pages_sharing_show(struct kobject *kobj,
2239 struct kobj_attribute *attr, char *buf)
2240{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07002241 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07002242}
2243KSM_ATTR_RO(pages_sharing);
2244
Hugh Dickins473b0ce2009-09-21 17:02:11 -07002245static ssize_t pages_unshared_show(struct kobject *kobj,
2246 struct kobj_attribute *attr, char *buf)
2247{
2248 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2249}
2250KSM_ATTR_RO(pages_unshared);
2251
2252static ssize_t pages_volatile_show(struct kobject *kobj,
2253 struct kobj_attribute *attr, char *buf)
2254{
2255 long ksm_pages_volatile;
2256
2257 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2258 - ksm_pages_sharing - ksm_pages_unshared;
2259 /*
2260 * It was not worth any locking to calculate that statistic,
2261 * but it might therefore sometimes be negative: conceal that.
2262 */
2263 if (ksm_pages_volatile < 0)
2264 ksm_pages_volatile = 0;
2265 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2266}
2267KSM_ATTR_RO(pages_volatile);
2268
2269static ssize_t full_scans_show(struct kobject *kobj,
2270 struct kobj_attribute *attr, char *buf)
2271{
2272 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2273}
2274KSM_ATTR_RO(full_scans);
2275
Izik Eidus31dbd012009-09-21 17:02:03 -07002276static struct attribute *ksm_attrs[] = {
2277 &sleep_millisecs_attr.attr,
2278 &pages_to_scan_attr.attr,
2279 &run_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07002280 &pages_shared_attr.attr,
2281 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07002282 &pages_unshared_attr.attr,
2283 &pages_volatile_attr.attr,
2284 &full_scans_attr.attr,
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002285#ifdef CONFIG_NUMA
2286 &merge_across_nodes_attr.attr,
2287#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07002288 NULL,
2289};
2290
2291static struct attribute_group ksm_attr_group = {
2292 .attrs = ksm_attrs,
2293 .name = "ksm",
2294};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002295#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002296
2297static int __init ksm_init(void)
2298{
2299 struct task_struct *ksm_thread;
2300 int err;
2301
2302 err = ksm_slab_init();
2303 if (err)
2304 goto out;
2305
Izik Eidus31dbd012009-09-21 17:02:03 -07002306 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2307 if (IS_ERR(ksm_thread)) {
Paul McQuade25acde32014-10-09 15:29:09 -07002308 pr_err("ksm: creating kthread failed\n");
Izik Eidus31dbd012009-09-21 17:02:03 -07002309 err = PTR_ERR(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002310 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002311 }
2312
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002313#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07002314 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2315 if (err) {
Paul McQuade25acde32014-10-09 15:29:09 -07002316 pr_err("ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002317 kthread_stop(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002318 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002319 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07002320#else
2321 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2322
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002323#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002324
Hugh Dickins62b61f62009-12-14 17:59:33 -08002325#ifdef CONFIG_MEMORY_HOTREMOVE
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002326 /* There is no significance to this priority 100 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08002327 hotplug_memory_notifier(ksm_memory_callback, 100);
2328#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07002329 return 0;
2330
Lai Jiangshand9f89842010-08-09 17:20:02 -07002331out_free:
Izik Eidus31dbd012009-09-21 17:02:03 -07002332 ksm_slab_free();
2333out:
2334 return err;
2335}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -08002336subsys_initcall(ksm_init);