blob: d61cba6fa1dc4380677dd03f6e092b12e208c881 [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 */
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800186static struct rb_root root_unstable_tree[MAX_NUMNODES];
187static struct rb_root root_stable_tree[MAX_NUMNODES];
Izik Eidus31dbd012009-09-21 17:02:03 -0700188
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800189/* Recently migrated nodes of stable tree, pending proper placement */
190static LIST_HEAD(migrate_nodes);
191
Sasha Levin4ca3a692013-02-22 16:32:28 -0800192#define MM_SLOTS_HASH_BITS 10
193static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
Izik Eidus31dbd012009-09-21 17:02:03 -0700194
195static struct mm_slot ksm_mm_head = {
196 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
197};
198static struct ksm_scan ksm_scan = {
199 .mm_slot = &ksm_mm_head,
200};
201
202static struct kmem_cache *rmap_item_cache;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800203static struct kmem_cache *stable_node_cache;
Izik Eidus31dbd012009-09-21 17:02:03 -0700204static struct kmem_cache *mm_slot_cache;
205
206/* The number of nodes in the stable tree */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700207static unsigned long ksm_pages_shared;
Izik Eidus31dbd012009-09-21 17:02:03 -0700208
Hugh Dickinse178dfd2009-09-21 17:02:10 -0700209/* The number of page slots additionally sharing those nodes */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700210static unsigned long ksm_pages_sharing;
Izik Eidus31dbd012009-09-21 17:02:03 -0700211
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700212/* The number of nodes in the unstable tree */
213static unsigned long ksm_pages_unshared;
214
215/* The number of rmap_items in use: to calculate pages_volatile */
216static unsigned long ksm_rmap_items;
217
Izik Eidus31dbd012009-09-21 17:02:03 -0700218/* Number of pages ksmd should scan in one batch */
Izik Eidus2c6854f2009-09-23 15:56:04 -0700219static unsigned int ksm_thread_pages_to_scan = 100;
Izik Eidus31dbd012009-09-21 17:02:03 -0700220
221/* Milliseconds ksmd should sleep between batches */
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700222static unsigned int ksm_thread_sleep_millisecs = 20;
Izik Eidus31dbd012009-09-21 17:02:03 -0700223
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800224#ifdef CONFIG_NUMA
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800225/* Zeroed when merging across nodes is not allowed */
226static unsigned int ksm_merge_across_nodes = 1;
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800227#else
228#define ksm_merge_across_nodes 1U
229#endif
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800230
Izik Eidus31dbd012009-09-21 17:02:03 -0700231#define KSM_RUN_STOP 0
232#define KSM_RUN_MERGE 1
233#define KSM_RUN_UNMERGE 2
Hugh Dickinsef4d43a2013-02-22 16:35:16 -0800234#define KSM_RUN_OFFLINE 4
235static unsigned long ksm_run = KSM_RUN_STOP;
236static void wait_while_offlining(void);
Izik Eidus31dbd012009-09-21 17:02:03 -0700237
238static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
239static DEFINE_MUTEX(ksm_thread_mutex);
240static DEFINE_SPINLOCK(ksm_mmlist_lock);
241
242#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
243 sizeof(struct __struct), __alignof__(struct __struct),\
244 (__flags), NULL)
245
246static int __init ksm_slab_init(void)
247{
248 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
249 if (!rmap_item_cache)
250 goto out;
251
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800252 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
253 if (!stable_node_cache)
254 goto out_free1;
255
Izik Eidus31dbd012009-09-21 17:02:03 -0700256 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
257 if (!mm_slot_cache)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800258 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -0700259
260 return 0;
261
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800262out_free2:
263 kmem_cache_destroy(stable_node_cache);
264out_free1:
Izik Eidus31dbd012009-09-21 17:02:03 -0700265 kmem_cache_destroy(rmap_item_cache);
266out:
267 return -ENOMEM;
268}
269
270static void __init ksm_slab_free(void)
271{
272 kmem_cache_destroy(mm_slot_cache);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800273 kmem_cache_destroy(stable_node_cache);
Izik Eidus31dbd012009-09-21 17:02:03 -0700274 kmem_cache_destroy(rmap_item_cache);
275 mm_slot_cache = NULL;
276}
277
278static inline struct rmap_item *alloc_rmap_item(void)
279{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700280 struct rmap_item *rmap_item;
281
282 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
283 if (rmap_item)
284 ksm_rmap_items++;
285 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -0700286}
287
288static inline void free_rmap_item(struct rmap_item *rmap_item)
289{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700290 ksm_rmap_items--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700291 rmap_item->mm = NULL; /* debug safety */
292 kmem_cache_free(rmap_item_cache, rmap_item);
293}
294
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800295static inline struct stable_node *alloc_stable_node(void)
296{
297 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
298}
299
300static inline void free_stable_node(struct stable_node *stable_node)
301{
302 kmem_cache_free(stable_node_cache, stable_node);
303}
304
Izik Eidus31dbd012009-09-21 17:02:03 -0700305static inline struct mm_slot *alloc_mm_slot(void)
306{
307 if (!mm_slot_cache) /* initialization failed */
308 return NULL;
309 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
310}
311
312static inline void free_mm_slot(struct mm_slot *mm_slot)
313{
314 kmem_cache_free(mm_slot_cache, mm_slot);
315}
316
Izik Eidus31dbd012009-09-21 17:02:03 -0700317static struct mm_slot *get_mm_slot(struct mm_struct *mm)
318{
Izik Eidus31dbd012009-09-21 17:02:03 -0700319 struct hlist_node *node;
Sasha Levin4ca3a692013-02-22 16:32:28 -0800320 struct mm_slot *slot;
Izik Eidus31dbd012009-09-21 17:02:03 -0700321
Sasha Levin4ca3a692013-02-22 16:32:28 -0800322 hash_for_each_possible(mm_slots_hash, slot, node, link, (unsigned long)mm)
323 if (slot->mm == mm)
324 return slot;
325
Izik Eidus31dbd012009-09-21 17:02:03 -0700326 return NULL;
327}
328
329static void insert_to_mm_slots_hash(struct mm_struct *mm,
330 struct mm_slot *mm_slot)
331{
Izik Eidus31dbd012009-09-21 17:02:03 -0700332 mm_slot->mm = mm;
Sasha Levin4ca3a692013-02-22 16:32:28 -0800333 hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm);
Izik Eidus31dbd012009-09-21 17:02:03 -0700334}
335
Izik Eidus31dbd012009-09-21 17:02:03 -0700336/*
Hugh Dickinsa913e182009-09-21 17:02:26 -0700337 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
338 * page tables after it has passed through ksm_exit() - which, if necessary,
339 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
340 * a special flag: they can just back out as soon as mm_users goes to zero.
341 * ksm_test_exit() is used throughout to make this test for exit: in some
342 * places for correctness, in some places just to avoid unnecessary work.
343 */
344static inline bool ksm_test_exit(struct mm_struct *mm)
345{
346 return atomic_read(&mm->mm_users) == 0;
347}
348
349/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700350 * We use break_ksm to break COW on a ksm page: it's a stripped down
351 *
352 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
353 * put_page(page);
354 *
355 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
356 * in case the application has unmapped and remapped mm,addr meanwhile.
357 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
358 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
359 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700360static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
Izik Eidus31dbd012009-09-21 17:02:03 -0700361{
362 struct page *page;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700363 int ret = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700364
365 do {
366 cond_resched();
Hugh Dickins5117b3b2013-02-22 16:36:07 -0800367 page = follow_page(vma, addr, FOLL_GET | FOLL_MIGRATION);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400368 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700369 break;
370 if (PageKsm(page))
371 ret = handle_mm_fault(vma->vm_mm, vma, addr,
372 FAULT_FLAG_WRITE);
373 else
374 ret = VM_FAULT_WRITE;
375 put_page(page);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700376 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
377 /*
378 * We must loop because handle_mm_fault() may back out if there's
379 * any difficulty e.g. if pte accessed bit gets updated concurrently.
380 *
381 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
382 * COW has been broken, even if the vma does not permit VM_WRITE;
383 * but note that a concurrent fault might break PageKsm for us.
384 *
385 * VM_FAULT_SIGBUS could occur if we race with truncation of the
386 * backing file, which also invalidates anonymous pages: that's
387 * okay, that truncation will have unmapped the PageKsm for us.
388 *
389 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
390 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
391 * current task has TIF_MEMDIE set, and will be OOM killed on return
392 * to user; and ksmd, having no mm, would never be chosen for that.
393 *
394 * But if the mm is in a limited mem_cgroup, then the fault may fail
395 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
396 * even ksmd can fail in this way - though it's usually breaking ksm
397 * just to undo a merge it made a moment before, so unlikely to oom.
398 *
399 * That's a pity: we might therefore have more kernel pages allocated
400 * than we're counting as nodes in the stable tree; but ksm_do_scan
401 * will retry to break_cow on each pass, so should recover the page
402 * in due course. The important thing is to not let VM_MERGEABLE
403 * be cleared while any such pages might remain in the area.
404 */
405 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700406}
407
Bob Liuef694222012-03-21 16:34:11 -0700408static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm,
409 unsigned long addr)
410{
411 struct vm_area_struct *vma;
412 if (ksm_test_exit(mm))
413 return NULL;
414 vma = find_vma(mm, addr);
415 if (!vma || vma->vm_start > addr)
416 return NULL;
417 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
418 return NULL;
419 return vma;
420}
421
Hugh Dickins8dd35572009-12-14 17:59:18 -0800422static void break_cow(struct rmap_item *rmap_item)
Izik Eidus31dbd012009-09-21 17:02:03 -0700423{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800424 struct mm_struct *mm = rmap_item->mm;
425 unsigned long addr = rmap_item->address;
Izik Eidus31dbd012009-09-21 17:02:03 -0700426 struct vm_area_struct *vma;
427
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800428 /*
429 * It is not an accident that whenever we want to break COW
430 * to undo, we also need to drop a reference to the anon_vma.
431 */
Peter Zijlstra9e601092011-03-22 16:32:46 -0700432 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800433
Hugh Dickins81464e302009-09-21 17:02:15 -0700434 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700435 vma = find_mergeable_vma(mm, addr);
436 if (vma)
437 break_ksm(vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700438 up_read(&mm->mmap_sem);
439}
440
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800441static struct page *page_trans_compound_anon(struct page *page)
442{
443 if (PageTransCompound(page)) {
Andrea Arcangeli22e5c472011-01-13 15:47:20 -0800444 struct page *head = compound_trans_head(page);
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800445 /*
Andrea Arcangeli22e5c472011-01-13 15:47:20 -0800446 * head may actually be splitted and freed from under
447 * us but it's ok here.
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800448 */
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800449 if (PageAnon(head))
450 return head;
451 }
452 return NULL;
453}
454
Izik Eidus31dbd012009-09-21 17:02:03 -0700455static struct page *get_mergeable_page(struct rmap_item *rmap_item)
456{
457 struct mm_struct *mm = rmap_item->mm;
458 unsigned long addr = rmap_item->address;
459 struct vm_area_struct *vma;
460 struct page *page;
461
462 down_read(&mm->mmap_sem);
Bob Liuef694222012-03-21 16:34:11 -0700463 vma = find_mergeable_vma(mm, addr);
464 if (!vma)
Izik Eidus31dbd012009-09-21 17:02:03 -0700465 goto out;
466
467 page = follow_page(vma, addr, FOLL_GET);
Dan Carpenter22eccdd2010-04-23 13:18:10 -0400468 if (IS_ERR_OR_NULL(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700469 goto out;
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800470 if (PageAnon(page) || page_trans_compound_anon(page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700471 flush_anon_page(vma, page, addr);
472 flush_dcache_page(page);
473 } else {
474 put_page(page);
475out: page = NULL;
476 }
477 up_read(&mm->mmap_sem);
478 return page;
479}
480
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800481/*
482 * This helper is used for getting right index into array of tree roots.
483 * When merge_across_nodes knob is set to 1, there are only two rb-trees for
484 * stable and unstable pages from all nodes with roots in index 0. Otherwise,
485 * every node has its own stable and unstable tree.
486 */
487static inline int get_kpfn_nid(unsigned long kpfn)
488{
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800489 return ksm_merge_across_nodes ? 0 : pfn_to_nid(kpfn);
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800490}
491
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800492static void remove_node_from_stable_tree(struct stable_node *stable_node)
493{
494 struct rmap_item *rmap_item;
495 struct hlist_node *hlist;
496
497 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
498 if (rmap_item->hlist.next)
499 ksm_pages_sharing--;
500 else
501 ksm_pages_shared--;
Peter Zijlstra9e601092011-03-22 16:32:46 -0700502 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800503 rmap_item->address &= PAGE_MASK;
504 cond_resched();
505 }
506
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800507 if (stable_node->head == &migrate_nodes)
508 list_del(&stable_node->list);
509 else
510 rb_erase(&stable_node->node,
511 &root_stable_tree[NUMA(stable_node->nid)]);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800512 free_stable_node(stable_node);
513}
514
515/*
516 * get_ksm_page: checks if the page indicated by the stable node
517 * is still its ksm page, despite having held no reference to it.
518 * In which case we can trust the content of the page, and it
519 * returns the gotten page; but if the page has now been zapped,
520 * remove the stale node from the stable tree and return NULL.
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800521 * But beware, the stable node's page might be being migrated.
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800522 *
523 * You would expect the stable_node to hold a reference to the ksm page.
524 * But if it increments the page's count, swapping out has to wait for
525 * ksmd to come around again before it can free the page, which may take
526 * seconds or even minutes: much too unresponsive. So instead we use a
527 * "keyhole reference": access to the ksm page from the stable node peeps
528 * out through its keyhole to see if that page still holds the right key,
529 * pointing back to this stable node. This relies on freeing a PageAnon
530 * page to reset its page->mapping to NULL, and relies on no other use of
531 * a page to put something that might look like our key in page->mapping.
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800532 * is on its way to being freed; but it is an anomaly to bear in mind.
533 */
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800534static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it)
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800535{
536 struct page *page;
537 void *expected_mapping;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800538 unsigned long kpfn;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800539
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800540 expected_mapping = (void *)stable_node +
541 (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800542again:
543 kpfn = ACCESS_ONCE(stable_node->kpfn);
544 page = pfn_to_page(kpfn);
545
546 /*
547 * page is computed from kpfn, so on most architectures reading
548 * page->mapping is naturally ordered after reading node->kpfn,
549 * but on Alpha we need to be more careful.
550 */
551 smp_read_barrier_depends();
552 if (ACCESS_ONCE(page->mapping) != expected_mapping)
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800553 goto stale;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800554
555 /*
556 * We cannot do anything with the page while its refcount is 0.
557 * Usually 0 means free, or tail of a higher-order page: in which
558 * case this node is no longer referenced, and should be freed;
559 * however, it might mean that the page is under page_freeze_refs().
560 * The __remove_mapping() case is easy, again the node is now stale;
561 * but if page is swapcache in migrate_page_move_mapping(), it might
562 * still be our page, in which case it's essential to keep the node.
563 */
564 while (!get_page_unless_zero(page)) {
565 /*
566 * Another check for page->mapping != expected_mapping would
567 * work here too. We have chosen the !PageSwapCache test to
568 * optimize the common case, when the page is or is about to
569 * be freed: PageSwapCache is cleared (under spin_lock_irq)
570 * in the freeze_refs section of __remove_mapping(); but Anon
571 * page->mapping reset to NULL later, in free_pages_prepare().
572 */
573 if (!PageSwapCache(page))
574 goto stale;
575 cpu_relax();
576 }
577
578 if (ACCESS_ONCE(page->mapping) != expected_mapping) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800579 put_page(page);
580 goto stale;
581 }
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800582
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800583 if (lock_it) {
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800584 lock_page(page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800585 if (ACCESS_ONCE(page->mapping) != expected_mapping) {
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800586 unlock_page(page);
587 put_page(page);
588 goto stale;
589 }
590 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800591 return page;
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800592
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800593stale:
Hugh Dickinsc8d65532013-02-22 16:35:10 -0800594 /*
595 * We come here from above when page->mapping or !PageSwapCache
596 * suggests that the node is stale; but it might be under migration.
597 * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(),
598 * before checking whether node->kpfn has been changed.
599 */
600 smp_rmb();
601 if (ACCESS_ONCE(stable_node->kpfn) != kpfn)
602 goto again;
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800603 remove_node_from_stable_tree(stable_node);
604 return NULL;
605}
606
Izik Eidus31dbd012009-09-21 17:02:03 -0700607/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700608 * Removing rmap_item from stable or unstable tree.
609 * This function will clean the information from the stable/unstable tree.
610 */
611static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
612{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800613 if (rmap_item->address & STABLE_FLAG) {
614 struct stable_node *stable_node;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800615 struct page *page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700616
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800617 stable_node = rmap_item->head;
Hugh Dickins8aafa6a2013-02-22 16:35:06 -0800618 page = get_ksm_page(stable_node, true);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800619 if (!page)
620 goto out;
621
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800622 hlist_del(&rmap_item->hlist);
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800623 unlock_page(page);
624 put_page(page);
Hugh Dickins08beca42009-12-14 17:59:21 -0800625
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800626 if (stable_node->hlist.first)
627 ksm_pages_sharing--;
628 else
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800629 ksm_pages_shared--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700630
Peter Zijlstra9e601092011-03-22 16:32:46 -0700631 put_anon_vma(rmap_item->anon_vma);
Hugh Dickins93d17712009-12-14 17:59:16 -0800632 rmap_item->address &= PAGE_MASK;
Izik Eidus31dbd012009-09-21 17:02:03 -0700633
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800634 } else if (rmap_item->address & UNSTABLE_FLAG) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700635 unsigned char age;
636 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -0700637 * Usually ksmd can and must skip the rb_erase, because
Izik Eidus31dbd012009-09-21 17:02:03 -0700638 * root_unstable_tree was already reset to RB_ROOT.
Hugh Dickins9ba69292009-09-21 17:02:20 -0700639 * But be careful when an mm is exiting: do the rb_erase
640 * if this rmap_item was inserted by this scan, rather
641 * than left over from before.
Izik Eidus31dbd012009-09-21 17:02:03 -0700642 */
643 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
Hugh Dickinscd551f92009-09-21 17:02:17 -0700644 BUG_ON(age > 1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700645 if (!age)
Petr Holasek90bd6fd2013-02-22 16:35:00 -0800646 rb_erase(&rmap_item->node,
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800647 &root_unstable_tree[NUMA(rmap_item->nid)]);
Hugh Dickins93d17712009-12-14 17:59:16 -0800648 ksm_pages_unshared--;
649 rmap_item->address &= PAGE_MASK;
650 }
Hugh Dickins4035c07a2009-12-14 17:59:27 -0800651out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700652 cond_resched(); /* we're called from many long loops */
653}
654
Izik Eidus31dbd012009-09-21 17:02:03 -0700655static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -0800656 struct rmap_item **rmap_list)
Izik Eidus31dbd012009-09-21 17:02:03 -0700657{
Hugh Dickins6514d512009-12-14 17:59:19 -0800658 while (*rmap_list) {
659 struct rmap_item *rmap_item = *rmap_list;
660 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700661 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700662 free_rmap_item(rmap_item);
663 }
664}
665
666/*
Hugh Dickinse850dcf2013-02-22 16:35:03 -0800667 * Though it's very tempting to unmerge rmap_items from stable tree rather
Izik Eidus31dbd012009-09-21 17:02:03 -0700668 * than check every pte of a given vma, the locking doesn't quite work for
669 * that - an rmap_item is assigned to the stable tree after inserting ksm
670 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
671 * rmap_items from parent to child at fork time (so as not to waste time
672 * if exit comes before the next scan reaches it).
Hugh Dickins81464e302009-09-21 17:02:15 -0700673 *
674 * Similarly, although we'd like to remove rmap_items (so updating counts
675 * and freeing memory) when unmerging an area, it's easier to leave that
676 * to the next pass of ksmd - consider, for example, how ksmd might be
677 * in cmp_and_merge_page on one of the rmap_items we would be removing.
Izik Eidus31dbd012009-09-21 17:02:03 -0700678 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700679static int unmerge_ksm_pages(struct vm_area_struct *vma,
680 unsigned long start, unsigned long end)
Izik Eidus31dbd012009-09-21 17:02:03 -0700681{
682 unsigned long addr;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700683 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700684
Hugh Dickinsd952b792009-09-21 17:02:16 -0700685 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700686 if (ksm_test_exit(vma->vm_mm))
687 break;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700688 if (signal_pending(current))
689 err = -ERESTARTSYS;
690 else
691 err = break_ksm(vma, addr);
692 }
693 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700694}
695
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700696#ifdef CONFIG_SYSFS
697/*
698 * Only called through the sysfs control interface:
699 */
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800700static int remove_stable_node(struct stable_node *stable_node)
701{
702 struct page *page;
703 int err;
704
705 page = get_ksm_page(stable_node, true);
706 if (!page) {
707 /*
708 * get_ksm_page did remove_node_from_stable_tree itself.
709 */
710 return 0;
711 }
712
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800713 if (WARN_ON_ONCE(page_mapped(page))) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800714 /*
Hugh Dickins8fdb3db2013-02-22 16:36:03 -0800715 * This should not happen: but if it does, just refuse to let
716 * merge_across_nodes be switched - there is no need to panic.
717 */
718 err = -EBUSY;
719 } else {
720 /*
721 * The stable node did not yet appear stale to get_ksm_page(),
722 * since that allows for an unmapped ksm page to be recognized
723 * right up until it is freed; but the node is safe to remove.
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800724 * This page might be in a pagevec waiting to be freed,
725 * or it might be PageSwapCache (perhaps under writeback),
726 * or it might have been removed from swapcache a moment ago.
727 */
728 set_page_stable_node(page, NULL);
729 remove_node_from_stable_tree(stable_node);
730 err = 0;
731 }
732
733 unlock_page(page);
734 put_page(page);
735 return err;
736}
737
738static int remove_all_stable_nodes(void)
739{
740 struct stable_node *stable_node;
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800741 struct list_head *this, *next;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800742 int nid;
743 int err = 0;
744
745 for (nid = 0; nid < nr_node_ids; nid++) {
746 while (root_stable_tree[nid].rb_node) {
747 stable_node = rb_entry(root_stable_tree[nid].rb_node,
748 struct stable_node, node);
749 if (remove_stable_node(stable_node)) {
750 err = -EBUSY;
751 break; /* proceed to next nid */
752 }
753 cond_resched();
754 }
755 }
Hugh Dickins4146d2d2013-02-22 16:35:11 -0800756 list_for_each_safe(this, next, &migrate_nodes) {
757 stable_node = list_entry(this, struct stable_node, list);
758 if (remove_stable_node(stable_node))
759 err = -EBUSY;
760 cond_resched();
761 }
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800762 return err;
763}
764
Hugh Dickinsd952b792009-09-21 17:02:16 -0700765static int unmerge_and_remove_all_rmap_items(void)
Izik Eidus31dbd012009-09-21 17:02:03 -0700766{
767 struct mm_slot *mm_slot;
768 struct mm_struct *mm;
769 struct vm_area_struct *vma;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700770 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700771
Hugh Dickinsd952b792009-09-21 17:02:16 -0700772 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700773 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700774 struct mm_slot, mm_list);
775 spin_unlock(&ksm_mmlist_lock);
776
Hugh Dickins9ba69292009-09-21 17:02:20 -0700777 for (mm_slot = ksm_scan.mm_slot;
778 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700779 mm = mm_slot->mm;
780 down_read(&mm->mmap_sem);
781 for (vma = mm->mmap; vma; vma = vma->vm_next) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700782 if (ksm_test_exit(mm))
783 break;
Izik Eidus31dbd012009-09-21 17:02:03 -0700784 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
785 continue;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700786 err = unmerge_ksm_pages(vma,
787 vma->vm_start, vma->vm_end);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700788 if (err)
789 goto error;
Izik Eidus31dbd012009-09-21 17:02:03 -0700790 }
Hugh Dickins9ba69292009-09-21 17:02:20 -0700791
Hugh Dickins6514d512009-12-14 17:59:19 -0800792 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700793
794 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700795 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700796 struct mm_slot, mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700797 if (ksm_test_exit(mm)) {
Sasha Levin4ca3a692013-02-22 16:32:28 -0800798 hash_del(&mm_slot->link);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700799 list_del(&mm_slot->mm_list);
800 spin_unlock(&ksm_mmlist_lock);
801
802 free_mm_slot(mm_slot);
803 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
804 up_read(&mm->mmap_sem);
805 mmdrop(mm);
806 } else {
807 spin_unlock(&ksm_mmlist_lock);
808 up_read(&mm->mmap_sem);
809 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700810 }
811
Hugh Dickinscbf86cf2013-02-22 16:35:08 -0800812 /* Clean up stable nodes, but don't worry if some are still busy */
813 remove_all_stable_nodes();
Hugh Dickinsd952b792009-09-21 17:02:16 -0700814 ksm_scan.seqnr = 0;
Hugh Dickins9ba69292009-09-21 17:02:20 -0700815 return 0;
816
817error:
818 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700819 spin_lock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700820 ksm_scan.mm_slot = &ksm_mm_head;
Izik Eidus31dbd012009-09-21 17:02:03 -0700821 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700822 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700823}
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700824#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -0700825
Izik Eidus31dbd012009-09-21 17:02:03 -0700826static u32 calc_checksum(struct page *page)
827{
828 u32 checksum;
Cong Wang9b04c5f2011-11-25 23:14:39 +0800829 void *addr = kmap_atomic(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700830 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
Cong Wang9b04c5f2011-11-25 23:14:39 +0800831 kunmap_atomic(addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700832 return checksum;
833}
834
835static int memcmp_pages(struct page *page1, struct page *page2)
836{
837 char *addr1, *addr2;
838 int ret;
839
Cong Wang9b04c5f2011-11-25 23:14:39 +0800840 addr1 = kmap_atomic(page1);
841 addr2 = kmap_atomic(page2);
Izik Eidus31dbd012009-09-21 17:02:03 -0700842 ret = memcmp(addr1, addr2, PAGE_SIZE);
Cong Wang9b04c5f2011-11-25 23:14:39 +0800843 kunmap_atomic(addr2);
844 kunmap_atomic(addr1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700845 return ret;
846}
847
848static inline int pages_identical(struct page *page1, struct page *page2)
849{
850 return !memcmp_pages(page1, page2);
851}
852
853static int write_protect_page(struct vm_area_struct *vma, struct page *page,
854 pte_t *orig_pte)
855{
856 struct mm_struct *mm = vma->vm_mm;
857 unsigned long addr;
858 pte_t *ptep;
859 spinlock_t *ptl;
860 int swapped;
861 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700862 unsigned long mmun_start; /* For mmu_notifiers */
863 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -0700864
865 addr = page_address_in_vma(page, vma);
866 if (addr == -EFAULT)
867 goto out;
868
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800869 BUG_ON(PageTransCompound(page));
Haggai Eran6bdb9132012-10-08 16:33:35 -0700870
871 mmun_start = addr;
872 mmun_end = addr + PAGE_SIZE;
873 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
874
Izik Eidus31dbd012009-09-21 17:02:03 -0700875 ptep = page_check_address(page, mm, addr, &ptl, 0);
876 if (!ptep)
Haggai Eran6bdb9132012-10-08 16:33:35 -0700877 goto out_mn;
Izik Eidus31dbd012009-09-21 17:02:03 -0700878
Hugh Dickins4e316352010-10-02 17:49:08 -0700879 if (pte_write(*ptep) || pte_dirty(*ptep)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700880 pte_t entry;
881
882 swapped = PageSwapCache(page);
883 flush_cache_page(vma, addr, page_to_pfn(page));
884 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300885 * Ok this is tricky, when get_user_pages_fast() run it doesn't
Izik Eidus31dbd012009-09-21 17:02:03 -0700886 * take any lock, therefore the check that we are going to make
887 * with the pagecount against the mapcount is racey and
888 * O_DIRECT can happen right after the check.
889 * So we clear the pte and flush the tlb before the check
890 * this assure us that no O_DIRECT can happen after the check
891 * or in the middle of the check.
892 */
893 entry = ptep_clear_flush(vma, addr, ptep);
894 /*
895 * Check that no O_DIRECT or similar I/O is in progress on the
896 * page
897 */
Hugh Dickins31e855e2009-12-14 17:59:17 -0800898 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
Robin Holtcb532372010-03-23 13:35:26 -0700899 set_pte_at(mm, addr, ptep, entry);
Izik Eidus31dbd012009-09-21 17:02:03 -0700900 goto out_unlock;
901 }
Hugh Dickins4e316352010-10-02 17:49:08 -0700902 if (pte_dirty(entry))
903 set_page_dirty(page);
904 entry = pte_mkclean(pte_wrprotect(entry));
Izik Eidus31dbd012009-09-21 17:02:03 -0700905 set_pte_at_notify(mm, addr, ptep, entry);
906 }
907 *orig_pte = *ptep;
908 err = 0;
909
910out_unlock:
911 pte_unmap_unlock(ptep, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -0700912out_mn:
913 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -0700914out:
915 return err;
916}
917
918/**
919 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -0800920 * @vma: vma that holds the pte pointing to page
921 * @page: the page we are replacing by kpage
922 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -0700923 * @orig_pte: the original value of the pte
924 *
925 * Returns 0 on success, -EFAULT on failure.
926 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800927static int replace_page(struct vm_area_struct *vma, struct page *page,
928 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -0700929{
930 struct mm_struct *mm = vma->vm_mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700931 pmd_t *pmd;
932 pte_t *ptep;
933 spinlock_t *ptl;
934 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -0700935 int err = -EFAULT;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700936 unsigned long mmun_start; /* For mmu_notifiers */
937 unsigned long mmun_end; /* For mmu_notifiers */
Izik Eidus31dbd012009-09-21 17:02:03 -0700938
Hugh Dickins8dd35572009-12-14 17:59:18 -0800939 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -0700940 if (addr == -EFAULT)
941 goto out;
942
Bob Liu62190492012-12-11 16:00:37 -0800943 pmd = mm_find_pmd(mm, addr);
944 if (!pmd)
Izik Eidus31dbd012009-09-21 17:02:03 -0700945 goto out;
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800946 BUG_ON(pmd_trans_huge(*pmd));
Izik Eidus31dbd012009-09-21 17:02:03 -0700947
Haggai Eran6bdb9132012-10-08 16:33:35 -0700948 mmun_start = addr;
949 mmun_end = addr + PAGE_SIZE;
950 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
951
Izik Eidus31dbd012009-09-21 17:02:03 -0700952 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
953 if (!pte_same(*ptep, orig_pte)) {
954 pte_unmap_unlock(ptep, ptl);
Haggai Eran6bdb9132012-10-08 16:33:35 -0700955 goto out_mn;
Izik Eidus31dbd012009-09-21 17:02:03 -0700956 }
957
Hugh Dickins8dd35572009-12-14 17:59:18 -0800958 get_page(kpage);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800959 page_add_anon_rmap(kpage, vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700960
961 flush_cache_page(vma, addr, pte_pfn(*ptep));
962 ptep_clear_flush(vma, addr, ptep);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800963 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
Izik Eidus31dbd012009-09-21 17:02:03 -0700964
Hugh Dickins8dd35572009-12-14 17:59:18 -0800965 page_remove_rmap(page);
Hugh Dickinsae52a2a2011-01-13 15:46:28 -0800966 if (!page_mapped(page))
967 try_to_free_swap(page);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800968 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700969
970 pte_unmap_unlock(ptep, ptl);
971 err = 0;
Haggai Eran6bdb9132012-10-08 16:33:35 -0700972out_mn:
973 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Izik Eidus31dbd012009-09-21 17:02:03 -0700974out:
975 return err;
976}
977
Andrea Arcangeli29ad7682011-01-13 15:47:19 -0800978static int page_trans_compound_anon_split(struct page *page)
979{
980 int ret = 0;
981 struct page *transhuge_head = page_trans_compound_anon(page);
982 if (transhuge_head) {
983 /* Get the reference on the head to split it. */
984 if (get_page_unless_zero(transhuge_head)) {
985 /*
986 * Recheck we got the reference while the head
987 * was still anonymous.
988 */
989 if (PageAnon(transhuge_head))
990 ret = split_huge_page(transhuge_head);
991 else
992 /*
993 * Retry later if split_huge_page run
994 * from under us.
995 */
996 ret = 1;
997 put_page(transhuge_head);
998 } else
999 /* Retry later if split_huge_page run from under us. */
1000 ret = 1;
1001 }
1002 return ret;
1003}
1004
Izik Eidus31dbd012009-09-21 17:02:03 -07001005/*
1006 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -08001007 * @vma: the vma that holds the pte pointing to page
1008 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins80e148222009-12-14 17:59:29 -08001009 * @kpage: the PageKsm page that we want to map instead of page,
1010 * or NULL the first time when we want to use page as kpage.
Izik Eidus31dbd012009-09-21 17:02:03 -07001011 *
1012 * This function returns 0 if the pages were merged, -EFAULT otherwise.
1013 */
1014static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -08001015 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001016{
1017 pte_t orig_pte = __pte(0);
1018 int err = -EFAULT;
1019
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001020 if (page == kpage) /* ksm page forked */
1021 return 0;
1022
Izik Eidus31dbd012009-09-21 17:02:03 -07001023 if (!(vma->vm_flags & VM_MERGEABLE))
1024 goto out;
Andrea Arcangeli29ad7682011-01-13 15:47:19 -08001025 if (PageTransCompound(page) && page_trans_compound_anon_split(page))
1026 goto out;
1027 BUG_ON(PageTransCompound(page));
Hugh Dickins8dd35572009-12-14 17:59:18 -08001028 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -07001029 goto out;
1030
Izik Eidus31dbd012009-09-21 17:02:03 -07001031 /*
1032 * We need the page lock to read a stable PageSwapCache in
1033 * write_protect_page(). We use trylock_page() instead of
1034 * lock_page() because we don't want to wait here - we
1035 * prefer to continue scanning and merging different pages,
1036 * then come back to this page when it is unlocked.
1037 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001038 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -08001039 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -07001040 /*
1041 * If this anonymous page is mapped only here, its pte may need
1042 * to be write-protected. If it's mapped elsewhere, all of its
1043 * ptes are necessarily already write-protected. But in either
1044 * case, we need to lock and check page_count is not raised.
1045 */
Hugh Dickins80e148222009-12-14 17:59:29 -08001046 if (write_protect_page(vma, page, &orig_pte) == 0) {
1047 if (!kpage) {
1048 /*
1049 * While we hold page lock, upgrade page from
1050 * PageAnon+anon_vma to PageKsm+NULL stable_node:
1051 * stable_tree_insert() will update stable_node.
1052 */
1053 set_page_stable_node(page, NULL);
1054 mark_page_accessed(page);
1055 err = 0;
1056 } else if (pages_identical(page, kpage))
1057 err = replace_page(vma, page, kpage, orig_pte);
1058 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001059
Hugh Dickins80e148222009-12-14 17:59:29 -08001060 if ((vma->vm_flags & VM_LOCKED) && kpage && !err) {
Hugh Dickins73848b42009-12-14 17:59:22 -08001061 munlock_vma_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001062 if (!PageMlocked(kpage)) {
1063 unlock_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001064 lock_page(kpage);
1065 mlock_vma_page(kpage);
1066 page = kpage; /* for final unlock */
1067 }
1068 }
Hugh Dickins73848b42009-12-14 17:59:22 -08001069
Hugh Dickins8dd35572009-12-14 17:59:18 -08001070 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001071out:
1072 return err;
1073}
1074
1075/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001076 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
1077 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -08001078 *
1079 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -07001080 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001081static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
1082 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -07001083{
Hugh Dickins8dd35572009-12-14 17:59:18 -08001084 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -07001085 struct vm_area_struct *vma;
1086 int err = -EFAULT;
1087
Hugh Dickins8dd35572009-12-14 17:59:18 -08001088 down_read(&mm->mmap_sem);
1089 if (ksm_test_exit(mm))
1090 goto out;
1091 vma = find_vma(mm, rmap_item->address);
1092 if (!vma || vma->vm_start > rmap_item->address)
Hugh Dickins9ba69292009-09-21 17:02:20 -07001093 goto out;
1094
Hugh Dickins8dd35572009-12-14 17:59:18 -08001095 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001096 if (err)
1097 goto out;
1098
Hugh Dickinsbc566202013-02-22 16:36:06 -08001099 /* Unstable nid is in union with stable anon_vma: remove first */
1100 remove_rmap_item_from_tree(rmap_item);
1101
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001102 /* Must get reference to anon_vma while still holding mmap_sem */
Peter Zijlstra9e601092011-03-22 16:32:46 -07001103 rmap_item->anon_vma = vma->anon_vma;
1104 get_anon_vma(vma->anon_vma);
Hugh Dickins81464e302009-09-21 17:02:15 -07001105out:
Hugh Dickins8dd35572009-12-14 17:59:18 -08001106 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -07001107 return err;
1108}
1109
1110/*
Izik Eidus31dbd012009-09-21 17:02:03 -07001111 * try_to_merge_two_pages - take two identical pages and prepare them
1112 * to be merged into one page.
1113 *
Hugh Dickins8dd35572009-12-14 17:59:18 -08001114 * This function returns the kpage if we successfully merged two identical
1115 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001116 *
Hugh Dickins80e148222009-12-14 17:59:29 -08001117 * Note that this function upgrades page to ksm page: if one of the pages
Izik Eidus31dbd012009-09-21 17:02:03 -07001118 * is already a ksm page, try_to_merge_with_ksm_page should be used.
1119 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001120static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
1121 struct page *page,
1122 struct rmap_item *tree_rmap_item,
1123 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001124{
Hugh Dickins80e148222009-12-14 17:59:29 -08001125 int err;
Izik Eidus31dbd012009-09-21 17:02:03 -07001126
Hugh Dickins80e148222009-12-14 17:59:29 -08001127 err = try_to_merge_with_ksm_page(rmap_item, page, NULL);
Izik Eidus31dbd012009-09-21 17:02:03 -07001128 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001129 err = try_to_merge_with_ksm_page(tree_rmap_item,
Hugh Dickins80e148222009-12-14 17:59:29 -08001130 tree_page, page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001131 /*
Hugh Dickins81464e302009-09-21 17:02:15 -07001132 * If that fails, we have a ksm page with only one pte
1133 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -07001134 */
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001135 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -08001136 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001137 }
Hugh Dickins80e148222009-12-14 17:59:29 -08001138 return err ? NULL : page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001139}
1140
1141/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001142 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -07001143 *
1144 * This function checks if there is a page inside the stable tree
1145 * with identical content to the page that we are scanning right now.
1146 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001147 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -07001148 * NULL otherwise.
1149 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001150static struct page *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001151{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001152 int nid;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001153 struct rb_node **new;
1154 struct rb_node *parent;
1155 struct stable_node *stable_node;
1156 struct stable_node *page_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001157
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001158 page_node = page_stable_node(page);
1159 if (page_node && page_node->head != &migrate_nodes) {
1160 /* ksm page forked */
Hugh Dickins08beca42009-12-14 17:59:21 -08001161 get_page(page);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001162 return page;
Hugh Dickins08beca42009-12-14 17:59:21 -08001163 }
1164
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001165 nid = get_kpfn_nid(page_to_pfn(page));
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001166again:
1167 new = &root_stable_tree[nid].rb_node;
1168 parent = NULL;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001169
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001170 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001171 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001172 int ret;
1173
Hugh Dickins08beca42009-12-14 17:59:21 -08001174 cond_resched();
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001175 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins8aafa6a2013-02-22 16:35:06 -08001176 tree_page = get_ksm_page(stable_node, false);
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001177 if (!tree_page)
1178 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001179
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001180 ret = memcmp_pages(page, tree_page);
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001181 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001182
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001183 parent = *new;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001184 if (ret < 0)
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001185 new = &parent->rb_left;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001186 else if (ret > 0)
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001187 new = &parent->rb_right;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001188 else {
1189 /*
1190 * Lock and unlock the stable_node's page (which
1191 * might already have been migrated) so that page
1192 * migration is sure to notice its raised count.
1193 * It would be more elegant to return stable_node
1194 * than kpage, but that involves more changes.
1195 */
1196 tree_page = get_ksm_page(stable_node, true);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001197 if (tree_page) {
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001198 unlock_page(tree_page);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001199 if (get_kpfn_nid(stable_node->kpfn) !=
1200 NUMA(stable_node->nid)) {
1201 put_page(tree_page);
1202 goto replace;
1203 }
1204 return tree_page;
1205 }
1206 /*
1207 * There is now a place for page_node, but the tree may
1208 * have been rebalanced, so re-evaluate parent and new.
1209 */
1210 if (page_node)
1211 goto again;
1212 return NULL;
Hugh Dickinsc8d65532013-02-22 16:35:10 -08001213 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001214 }
1215
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001216 if (!page_node)
1217 return NULL;
1218
1219 list_del(&page_node->list);
1220 DO_NUMA(page_node->nid = nid);
1221 rb_link_node(&page_node->node, parent, new);
1222 rb_insert_color(&page_node->node, &root_stable_tree[nid]);
1223 get_page(page);
1224 return page;
1225
1226replace:
1227 if (page_node) {
1228 list_del(&page_node->list);
1229 DO_NUMA(page_node->nid = nid);
1230 rb_replace_node(&stable_node->node,
1231 &page_node->node, &root_stable_tree[nid]);
1232 get_page(page);
1233 } else {
1234 rb_erase(&stable_node->node, &root_stable_tree[nid]);
1235 page = NULL;
1236 }
1237 stable_node->head = &migrate_nodes;
1238 list_add(&stable_node->list, stable_node->head);
1239 return page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001240}
1241
1242/*
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001243 * stable_tree_insert - insert stable tree node pointing to new ksm page
Izik Eidus31dbd012009-09-21 17:02:03 -07001244 * into the stable tree.
1245 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001246 * This function returns the stable tree node just allocated on success,
1247 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -07001248 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001249static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -07001250{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001251 int nid;
1252 unsigned long kpfn;
1253 struct rb_node **new;
Izik Eidus31dbd012009-09-21 17:02:03 -07001254 struct rb_node *parent = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001255 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001256
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001257 kpfn = page_to_pfn(kpage);
1258 nid = get_kpfn_nid(kpfn);
1259 new = &root_stable_tree[nid].rb_node;
1260
Izik Eidus31dbd012009-09-21 17:02:03 -07001261 while (*new) {
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001262 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001263 int ret;
1264
Hugh Dickins08beca42009-12-14 17:59:21 -08001265 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001266 stable_node = rb_entry(*new, struct stable_node, node);
Hugh Dickins8aafa6a2013-02-22 16:35:06 -08001267 tree_page = get_ksm_page(stable_node, false);
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001268 if (!tree_page)
1269 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001270
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001271 ret = memcmp_pages(kpage, tree_page);
1272 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001273
1274 parent = *new;
1275 if (ret < 0)
1276 new = &parent->rb_left;
1277 else if (ret > 0)
1278 new = &parent->rb_right;
1279 else {
1280 /*
1281 * It is not a bug that stable_tree_search() didn't
1282 * find this node: because at that time our page was
1283 * not yet write-protected, so may have changed since.
1284 */
1285 return NULL;
1286 }
1287 }
1288
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001289 stable_node = alloc_stable_node();
1290 if (!stable_node)
1291 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001292
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001293 INIT_HLIST_HEAD(&stable_node->hlist);
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001294 stable_node->kpfn = kpfn;
Hugh Dickins08beca42009-12-14 17:59:21 -08001295 set_page_stable_node(kpage, stable_node);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001296 DO_NUMA(stable_node->nid = nid);
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001297 rb_link_node(&stable_node->node, parent, new);
1298 rb_insert_color(&stable_node->node, &root_stable_tree[nid]);
Hugh Dickins08beca42009-12-14 17:59:21 -08001299
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001300 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001301}
1302
1303/*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001304 * unstable_tree_search_insert - search for identical page,
1305 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001306 *
1307 * This function searches for a page in the unstable tree identical to the
1308 * page currently being scanned; and if no identical page is found in the
1309 * tree, we insert rmap_item as a new object into the unstable tree.
1310 *
1311 * This function returns pointer to rmap_item found to be identical
1312 * to the currently scanned page, NULL otherwise.
1313 *
1314 * This function does both searching and inserting, because they share
1315 * the same walking algorithm in an rbtree.
1316 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001317static
1318struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1319 struct page *page,
1320 struct page **tree_pagep)
Izik Eidus31dbd012009-09-21 17:02:03 -07001321{
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001322 struct rb_node **new;
1323 struct rb_root *root;
Izik Eidus31dbd012009-09-21 17:02:03 -07001324 struct rb_node *parent = NULL;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001325 int nid;
1326
1327 nid = get_kpfn_nid(page_to_pfn(page));
1328 root = &root_unstable_tree[nid];
1329 new = &root->rb_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001330
1331 while (*new) {
1332 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001333 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001334 int ret;
1335
Hugh Dickinsd178f272009-11-09 15:58:23 +00001336 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -07001337 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001338 tree_page = get_mergeable_page(tree_rmap_item);
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001339 if (IS_ERR_OR_NULL(tree_page))
Izik Eidus31dbd012009-09-21 17:02:03 -07001340 return NULL;
1341
1342 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001343 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -07001344 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001345 if (page == tree_page) {
1346 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001347 return NULL;
1348 }
1349
Hugh Dickins8dd35572009-12-14 17:59:18 -08001350 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001351
1352 parent = *new;
1353 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001354 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001355 new = &parent->rb_left;
1356 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001357 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001358 new = &parent->rb_right;
Hugh Dickinsb599cbd2013-02-22 16:36:05 -08001359 } else if (!ksm_merge_across_nodes &&
1360 page_to_nid(tree_page) != nid) {
1361 /*
1362 * If tree_page has been migrated to another NUMA node,
1363 * it will be flushed out and put in the right unstable
1364 * tree next time: only merge with it when across_nodes.
1365 */
1366 put_page(tree_page);
1367 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001368 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001369 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001370 return tree_rmap_item;
1371 }
1372 }
1373
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001374 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001375 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
Hugh Dickinse850dcf2013-02-22 16:35:03 -08001376 DO_NUMA(rmap_item->nid = nid);
Izik Eidus31dbd012009-09-21 17:02:03 -07001377 rb_link_node(&rmap_item->node, parent, new);
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001378 rb_insert_color(&rmap_item->node, root);
Izik Eidus31dbd012009-09-21 17:02:03 -07001379
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001380 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001381 return NULL;
1382}
1383
1384/*
1385 * stable_tree_append - add another rmap_item to the linked list of
1386 * rmap_items hanging off a given node of the stable tree, all sharing
1387 * the same ksm page.
1388 */
1389static void stable_tree_append(struct rmap_item *rmap_item,
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001390 struct stable_node *stable_node)
Izik Eidus31dbd012009-09-21 17:02:03 -07001391{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001392 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001393 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001394 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001395
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001396 if (rmap_item->hlist.next)
1397 ksm_pages_sharing++;
1398 else
1399 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001400}
1401
1402/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001403 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1404 * if not, compare checksum to previous and if it's the same, see if page can
1405 * be inserted into the unstable tree, or merged with a page already there and
1406 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001407 *
1408 * @page: the page that we are searching identical page to.
1409 * @rmap_item: the reverse mapping into the virtual address of this page
1410 */
1411static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1412{
Izik Eidus31dbd012009-09-21 17:02:03 -07001413 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001414 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001415 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001416 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07001417 unsigned int checksum;
1418 int err;
1419
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001420 stable_node = page_stable_node(page);
1421 if (stable_node) {
1422 if (stable_node->head != &migrate_nodes &&
1423 get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) {
1424 rb_erase(&stable_node->node,
1425 &root_stable_tree[NUMA(stable_node->nid)]);
1426 stable_node->head = &migrate_nodes;
1427 list_add(&stable_node->list, stable_node->head);
1428 }
1429 if (stable_node->head != &migrate_nodes &&
1430 rmap_item->head == stable_node)
1431 return;
1432 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001433
1434 /* We first start with searching the page inside the stable tree */
Hugh Dickins62b61f62009-12-14 17:59:33 -08001435 kpage = stable_tree_search(page);
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001436 if (kpage == page && rmap_item->head == stable_node) {
1437 put_page(kpage);
1438 return;
1439 }
1440
1441 remove_rmap_item_from_tree(rmap_item);
1442
Hugh Dickins62b61f62009-12-14 17:59:33 -08001443 if (kpage) {
Hugh Dickins08beca42009-12-14 17:59:21 -08001444 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001445 if (!err) {
1446 /*
1447 * The page was successfully merged:
1448 * add its rmap_item to the stable tree.
1449 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001450 lock_page(kpage);
Hugh Dickins62b61f62009-12-14 17:59:33 -08001451 stable_tree_append(rmap_item, page_stable_node(kpage));
Hugh Dickins5ad64682009-12-14 17:59:24 -08001452 unlock_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001453 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001454 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001455 return;
1456 }
1457
1458 /*
Hugh Dickins4035c07a2009-12-14 17:59:27 -08001459 * If the hash value of the page has changed from the last time
1460 * we calculated it, this page is changing frequently: therefore we
1461 * don't want to insert it in the unstable tree, and we don't want
1462 * to waste our time searching for something identical to it there.
Izik Eidus31dbd012009-09-21 17:02:03 -07001463 */
1464 checksum = calc_checksum(page);
1465 if (rmap_item->oldchecksum != checksum) {
1466 rmap_item->oldchecksum = checksum;
1467 return;
1468 }
1469
Hugh Dickins8dd35572009-12-14 17:59:18 -08001470 tree_rmap_item =
1471 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001472 if (tree_rmap_item) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001473 kpage = try_to_merge_two_pages(rmap_item, page,
1474 tree_rmap_item, tree_page);
1475 put_page(tree_page);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001476 if (kpage) {
Hugh Dickinsbc566202013-02-22 16:36:06 -08001477 /*
1478 * The pages were successfully merged: insert new
1479 * node in the stable tree and add both rmap_items.
1480 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001481 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001482 stable_node = stable_tree_insert(kpage);
1483 if (stable_node) {
1484 stable_tree_append(tree_rmap_item, stable_node);
1485 stable_tree_append(rmap_item, stable_node);
1486 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001487 unlock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001488
Izik Eidus31dbd012009-09-21 17:02:03 -07001489 /*
1490 * If we fail to insert the page into the stable tree,
1491 * we will have 2 virtual addresses that are pointing
1492 * to a ksm page left outside the stable tree,
1493 * in which case we need to break_cow on both.
1494 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001495 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001496 break_cow(tree_rmap_item);
1497 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001498 }
1499 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001500 }
1501}
1502
1503static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001504 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07001505 unsigned long addr)
1506{
1507 struct rmap_item *rmap_item;
1508
Hugh Dickins6514d512009-12-14 17:59:19 -08001509 while (*rmap_list) {
1510 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08001511 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07001512 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001513 if (rmap_item->address > addr)
1514 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08001515 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001516 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001517 free_rmap_item(rmap_item);
1518 }
1519
1520 rmap_item = alloc_rmap_item();
1521 if (rmap_item) {
1522 /* It has already been zeroed */
1523 rmap_item->mm = mm_slot->mm;
1524 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08001525 rmap_item->rmap_list = *rmap_list;
1526 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001527 }
1528 return rmap_item;
1529}
1530
1531static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1532{
1533 struct mm_struct *mm;
1534 struct mm_slot *slot;
1535 struct vm_area_struct *vma;
1536 struct rmap_item *rmap_item;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001537 int nid;
Izik Eidus31dbd012009-09-21 17:02:03 -07001538
1539 if (list_empty(&ksm_mm_head.mm_list))
1540 return NULL;
1541
1542 slot = ksm_scan.mm_slot;
1543 if (slot == &ksm_mm_head) {
Hugh Dickins2919bfd2011-01-13 15:47:29 -08001544 /*
1545 * A number of pages can hang around indefinitely on per-cpu
1546 * pagevecs, raised page count preventing write_protect_page
1547 * from merging them. Though it doesn't really matter much,
1548 * it is puzzling to see some stuck in pages_volatile until
1549 * other activity jostles them out, and they also prevented
1550 * LTP's KSM test from succeeding deterministically; so drain
1551 * them here (here rather than on entry to ksm_do_scan(),
1552 * so we don't IPI too often when pages_to_scan is set low).
1553 */
1554 lru_add_drain_all();
1555
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001556 /*
1557 * Whereas stale stable_nodes on the stable_tree itself
1558 * get pruned in the regular course of stable_tree_search(),
1559 * those moved out to the migrate_nodes list can accumulate:
1560 * so prune them once before each full scan.
1561 */
1562 if (!ksm_merge_across_nodes) {
1563 struct stable_node *stable_node;
1564 struct list_head *this, *next;
1565 struct page *page;
1566
1567 list_for_each_safe(this, next, &migrate_nodes) {
1568 stable_node = list_entry(this,
1569 struct stable_node, list);
1570 page = get_ksm_page(stable_node, false);
1571 if (page)
1572 put_page(page);
1573 cond_resched();
1574 }
1575 }
1576
Petr Holasek90bd6fd2013-02-22 16:35:00 -08001577 for (nid = 0; nid < nr_node_ids; nid++)
1578 root_unstable_tree[nid] = RB_ROOT;
Izik Eidus31dbd012009-09-21 17:02:03 -07001579
1580 spin_lock(&ksm_mmlist_lock);
1581 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1582 ksm_scan.mm_slot = slot;
1583 spin_unlock(&ksm_mmlist_lock);
Hugh Dickins2b472612011-06-15 15:08:58 -07001584 /*
1585 * Although we tested list_empty() above, a racing __ksm_exit
1586 * of the last mm on the list may have removed it since then.
1587 */
1588 if (slot == &ksm_mm_head)
1589 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -07001590next_mm:
1591 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001592 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001593 }
1594
1595 mm = slot->mm;
1596 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001597 if (ksm_test_exit(mm))
1598 vma = NULL;
1599 else
1600 vma = find_vma(mm, ksm_scan.address);
1601
1602 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001603 if (!(vma->vm_flags & VM_MERGEABLE))
1604 continue;
1605 if (ksm_scan.address < vma->vm_start)
1606 ksm_scan.address = vma->vm_start;
1607 if (!vma->anon_vma)
1608 ksm_scan.address = vma->vm_end;
1609
1610 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001611 if (ksm_test_exit(mm))
1612 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001613 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001614 if (IS_ERR_OR_NULL(*page)) {
1615 ksm_scan.address += PAGE_SIZE;
1616 cond_resched();
1617 continue;
1618 }
Andrea Arcangeli29ad7682011-01-13 15:47:19 -08001619 if (PageAnon(*page) ||
1620 page_trans_compound_anon(*page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001621 flush_anon_page(vma, *page, ksm_scan.address);
1622 flush_dcache_page(*page);
1623 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001624 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07001625 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001626 ksm_scan.rmap_list =
1627 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001628 ksm_scan.address += PAGE_SIZE;
1629 } else
1630 put_page(*page);
1631 up_read(&mm->mmap_sem);
1632 return rmap_item;
1633 }
Andrea Arcangeli21ae5b02011-01-13 15:47:00 -08001634 put_page(*page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001635 ksm_scan.address += PAGE_SIZE;
1636 cond_resched();
1637 }
1638 }
1639
Hugh Dickins9ba69292009-09-21 17:02:20 -07001640 if (ksm_test_exit(mm)) {
1641 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001642 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001643 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001644 /*
1645 * Nuke all the rmap_items that are above this current rmap:
1646 * because there were no VM_MERGEABLE vmas with such addresses.
1647 */
Hugh Dickins6514d512009-12-14 17:59:19 -08001648 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001649
1650 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001651 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1652 struct mm_slot, mm_list);
1653 if (ksm_scan.address == 0) {
1654 /*
1655 * We've completed a full scan of all vmas, holding mmap_sem
1656 * throughout, and found no VM_MERGEABLE: so do the same as
1657 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07001658 * This applies either when cleaning up after __ksm_exit
1659 * (but beware: we can reach here even before __ksm_exit),
1660 * or when all VM_MERGEABLE areas have been unmapped (and
1661 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07001662 */
Sasha Levin4ca3a692013-02-22 16:32:28 -08001663 hash_del(&slot->link);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001664 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001665 spin_unlock(&ksm_mmlist_lock);
1666
Hugh Dickinscd551f92009-09-21 17:02:17 -07001667 free_mm_slot(slot);
1668 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001669 up_read(&mm->mmap_sem);
1670 mmdrop(mm);
1671 } else {
1672 spin_unlock(&ksm_mmlist_lock);
1673 up_read(&mm->mmap_sem);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001674 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001675
1676 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07001677 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07001678 if (slot != &ksm_mm_head)
1679 goto next_mm;
1680
Izik Eidus31dbd012009-09-21 17:02:03 -07001681 ksm_scan.seqnr++;
1682 return NULL;
1683}
1684
1685/**
1686 * ksm_do_scan - the ksm scanner main worker function.
1687 * @scan_npages - number of pages we want to scan before we return.
1688 */
1689static void ksm_do_scan(unsigned int scan_npages)
1690{
1691 struct rmap_item *rmap_item;
Dan Carpenter22eccdd2010-04-23 13:18:10 -04001692 struct page *uninitialized_var(page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001693
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001694 while (scan_npages-- && likely(!freezing(current))) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001695 cond_resched();
1696 rmap_item = scan_get_next_rmap_item(&page);
1697 if (!rmap_item)
1698 return;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08001699 cmp_and_merge_page(page, rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001700 put_page(page);
1701 }
1702}
1703
Hugh Dickins6e1583842009-09-21 17:02:14 -07001704static int ksmd_should_run(void)
1705{
1706 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1707}
1708
Izik Eidus31dbd012009-09-21 17:02:03 -07001709static int ksm_scan_thread(void *nothing)
1710{
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001711 set_freezable();
Izik Eidus339aa622009-09-21 17:02:07 -07001712 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07001713
1714 while (!kthread_should_stop()) {
Hugh Dickins6e1583842009-09-21 17:02:14 -07001715 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08001716 wait_while_offlining();
Hugh Dickins6e1583842009-09-21 17:02:14 -07001717 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07001718 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001719 mutex_unlock(&ksm_thread_mutex);
1720
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001721 try_to_freeze();
1722
Hugh Dickins6e1583842009-09-21 17:02:14 -07001723 if (ksmd_should_run()) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001724 schedule_timeout_interruptible(
1725 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1726 } else {
Andrea Arcangeli878aee72011-01-13 15:47:10 -08001727 wait_event_freezable(ksm_thread_wait,
Hugh Dickins6e1583842009-09-21 17:02:14 -07001728 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07001729 }
1730 }
1731 return 0;
1732}
1733
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001734int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1735 unsigned long end, int advice, unsigned long *vm_flags)
1736{
1737 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001738 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001739
1740 switch (advice) {
1741 case MADV_MERGEABLE:
1742 /*
1743 * Be somewhat over-protective for now!
1744 */
1745 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1746 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001747 VM_HUGETLB | VM_NONLINEAR | VM_MIXEDMAP))
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001748 return 0; /* just ignore the advice */
1749
Konstantin Khlebnikovcc2383e2012-10-08 16:28:37 -07001750#ifdef VM_SAO
1751 if (*vm_flags & VM_SAO)
1752 return 0;
1753#endif
1754
Hugh Dickinsd952b792009-09-21 17:02:16 -07001755 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1756 err = __ksm_enter(mm);
1757 if (err)
1758 return err;
1759 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001760
1761 *vm_flags |= VM_MERGEABLE;
1762 break;
1763
1764 case MADV_UNMERGEABLE:
1765 if (!(*vm_flags & VM_MERGEABLE))
1766 return 0; /* just ignore the advice */
1767
Hugh Dickinsd952b792009-09-21 17:02:16 -07001768 if (vma->anon_vma) {
1769 err = unmerge_ksm_pages(vma, start, end);
1770 if (err)
1771 return err;
1772 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001773
1774 *vm_flags &= ~VM_MERGEABLE;
1775 break;
1776 }
1777
1778 return 0;
1779}
1780
1781int __ksm_enter(struct mm_struct *mm)
1782{
Hugh Dickins6e1583842009-09-21 17:02:14 -07001783 struct mm_slot *mm_slot;
1784 int needs_wakeup;
1785
1786 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07001787 if (!mm_slot)
1788 return -ENOMEM;
1789
Hugh Dickins6e1583842009-09-21 17:02:14 -07001790 /* Check ksm_run too? Would need tighter locking */
1791 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1792
Izik Eidus31dbd012009-09-21 17:02:03 -07001793 spin_lock(&ksm_mmlist_lock);
1794 insert_to_mm_slots_hash(mm, mm_slot);
1795 /*
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001796 * When KSM_RUN_MERGE (or KSM_RUN_STOP),
1797 * insert just behind the scanning cursor, to let the area settle
Izik Eidus31dbd012009-09-21 17:02:03 -07001798 * down a little; when fork is followed by immediate exec, we don't
1799 * want ksmd to waste time setting up and tearing down an rmap_list.
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001800 *
1801 * But when KSM_RUN_UNMERGE, it's important to insert ahead of its
1802 * scanning cursor, otherwise KSM pages in newly forked mms will be
1803 * missed: then we might as well insert at the end of the list.
Izik Eidus31dbd012009-09-21 17:02:03 -07001804 */
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001805 if (ksm_run & KSM_RUN_UNMERGE)
1806 list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list);
1807 else
1808 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001809 spin_unlock(&ksm_mmlist_lock);
1810
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001811 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001812 atomic_inc(&mm->mm_count);
Hugh Dickins6e1583842009-09-21 17:02:14 -07001813
1814 if (needs_wakeup)
1815 wake_up_interruptible(&ksm_thread_wait);
1816
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001817 return 0;
1818}
1819
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07001820void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001821{
Hugh Dickinscd551f92009-09-21 17:02:17 -07001822 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001823 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07001824
Izik Eidus31dbd012009-09-21 17:02:03 -07001825 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07001826 * This process is exiting: if it's straightforward (as is the
1827 * case when ksmd was never running), free mm_slot immediately.
1828 * But if it's at the cursor or has rmap_items linked to it, use
1829 * mmap_sem to synchronize with any break_cows before pagetables
1830 * are freed, and leave the mm_slot on the list for ksmd to free.
1831 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07001832 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07001833
Hugh Dickinscd551f92009-09-21 17:02:17 -07001834 spin_lock(&ksm_mmlist_lock);
1835 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001836 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001837 if (!mm_slot->rmap_list) {
Sasha Levin4ca3a692013-02-22 16:32:28 -08001838 hash_del(&mm_slot->link);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001839 list_del(&mm_slot->mm_list);
1840 easy_to_free = 1;
1841 } else {
1842 list_move(&mm_slot->mm_list,
1843 &ksm_scan.mm_slot->mm_list);
1844 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001845 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001846 spin_unlock(&ksm_mmlist_lock);
1847
Hugh Dickins9ba69292009-09-21 17:02:20 -07001848 if (easy_to_free) {
1849 free_mm_slot(mm_slot);
1850 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1851 mmdrop(mm);
1852 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001853 down_write(&mm->mmap_sem);
1854 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001855 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001856}
Izik Eidus31dbd012009-09-21 17:02:03 -07001857
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001858struct page *ksm_might_need_to_copy(struct page *page,
Hugh Dickins5ad64682009-12-14 17:59:24 -08001859 struct vm_area_struct *vma, unsigned long address)
1860{
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001861 struct anon_vma *anon_vma = page_anon_vma(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001862 struct page *new_page;
1863
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08001864 if (PageKsm(page)) {
1865 if (page_stable_node(page) &&
1866 !(ksm_run & KSM_RUN_UNMERGE))
1867 return page; /* no need to copy it */
1868 } else if (!anon_vma) {
1869 return page; /* no need to copy it */
1870 } else if (anon_vma->root == vma->anon_vma->root &&
1871 page->index == linear_page_index(vma, address)) {
1872 return page; /* still no need to copy it */
1873 }
1874 if (!PageUptodate(page))
1875 return page; /* let do_swap_page report the error */
1876
Hugh Dickins5ad64682009-12-14 17:59:24 -08001877 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1878 if (new_page) {
1879 copy_user_highpage(new_page, page, address, vma);
1880
1881 SetPageDirty(new_page);
1882 __SetPageUptodate(new_page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001883 __set_page_locked(new_page);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001884 }
1885
Hugh Dickins5ad64682009-12-14 17:59:24 -08001886 return new_page;
1887}
1888
1889int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1890 unsigned long *vm_flags)
1891{
1892 struct stable_node *stable_node;
1893 struct rmap_item *rmap_item;
1894 struct hlist_node *hlist;
1895 unsigned int mapcount = page_mapcount(page);
1896 int referenced = 0;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001897 int search_new_forks = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001898
1899 VM_BUG_ON(!PageKsm(page));
1900 VM_BUG_ON(!PageLocked(page));
1901
1902 stable_node = page_stable_node(page);
1903 if (!stable_node)
1904 return 0;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001905again:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001906 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001907 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001908 struct anon_vma_chain *vmac;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001909 struct vm_area_struct *vma;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001910
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001911 anon_vma_lock_read(anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -07001912 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1913 0, ULONG_MAX) {
Rik van Riel5beb4932010-03-05 13:42:07 -08001914 vma = vmac->vma;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001915 if (rmap_item->address < vma->vm_start ||
1916 rmap_item->address >= vma->vm_end)
1917 continue;
1918 /*
1919 * Initially we examine only the vma which covers this
1920 * rmap_item; but later, if there is still work to do,
1921 * we examine covering vmas in other mms: in case they
1922 * were forked from the original since ksmd passed.
1923 */
1924 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1925 continue;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001926
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001927 if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
1928 continue;
1929
1930 referenced += page_referenced_one(page, vma,
Hugh Dickins5ad64682009-12-14 17:59:24 -08001931 rmap_item->address, &mapcount, vm_flags);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001932 if (!search_new_forks || !mapcount)
1933 break;
1934 }
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001935 anon_vma_unlock_read(anon_vma);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001936 if (!mapcount)
1937 goto out;
1938 }
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001939 if (!search_new_forks++)
1940 goto again;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001941out:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001942 return referenced;
1943}
1944
1945int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1946{
1947 struct stable_node *stable_node;
1948 struct hlist_node *hlist;
1949 struct rmap_item *rmap_item;
1950 int ret = SWAP_AGAIN;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001951 int search_new_forks = 0;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001952
1953 VM_BUG_ON(!PageKsm(page));
1954 VM_BUG_ON(!PageLocked(page));
1955
1956 stable_node = page_stable_node(page);
1957 if (!stable_node)
1958 return SWAP_FAIL;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001959again:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001960 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001961 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08001962 struct anon_vma_chain *vmac;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001963 struct vm_area_struct *vma;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001964
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001965 anon_vma_lock_read(anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -07001966 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
1967 0, ULONG_MAX) {
Rik van Riel5beb4932010-03-05 13:42:07 -08001968 vma = vmac->vma;
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001969 if (rmap_item->address < vma->vm_start ||
1970 rmap_item->address >= vma->vm_end)
1971 continue;
1972 /*
1973 * Initially we examine only the vma which covers this
1974 * rmap_item; but later, if there is still work to do,
1975 * we examine covering vmas in other mms: in case they
1976 * were forked from the original since ksmd passed.
1977 */
1978 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
1979 continue;
1980
1981 ret = try_to_unmap_one(page, vma,
1982 rmap_item->address, flags);
1983 if (ret != SWAP_AGAIN || !page_mapped(page)) {
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001984 anon_vma_unlock_read(anon_vma);
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001985 goto out;
1986 }
1987 }
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08001988 anon_vma_unlock_read(anon_vma);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001989 }
Hugh Dickinsdb114b82009-12-14 17:59:25 -08001990 if (!search_new_forks++)
1991 goto again;
Hugh Dickins5ad64682009-12-14 17:59:24 -08001992out:
Hugh Dickins5ad64682009-12-14 17:59:24 -08001993 return ret;
1994}
1995
Hugh Dickinse9995ef2009-12-14 17:59:31 -08001996#ifdef CONFIG_MIGRATION
1997int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *,
1998 struct vm_area_struct *, unsigned long, void *), void *arg)
1999{
2000 struct stable_node *stable_node;
2001 struct hlist_node *hlist;
2002 struct rmap_item *rmap_item;
2003 int ret = SWAP_AGAIN;
2004 int search_new_forks = 0;
2005
2006 VM_BUG_ON(!PageKsm(page));
2007 VM_BUG_ON(!PageLocked(page));
2008
2009 stable_node = page_stable_node(page);
2010 if (!stable_node)
2011 return ret;
2012again:
2013 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
2014 struct anon_vma *anon_vma = rmap_item->anon_vma;
Rik van Riel5beb4932010-03-05 13:42:07 -08002015 struct anon_vma_chain *vmac;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002016 struct vm_area_struct *vma;
2017
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08002018 anon_vma_lock_read(anon_vma);
Michel Lespinassebf181b92012-10-08 16:31:39 -07002019 anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root,
2020 0, ULONG_MAX) {
Rik van Riel5beb4932010-03-05 13:42:07 -08002021 vma = vmac->vma;
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002022 if (rmap_item->address < vma->vm_start ||
2023 rmap_item->address >= vma->vm_end)
2024 continue;
2025 /*
2026 * Initially we examine only the vma which covers this
2027 * rmap_item; but later, if there is still work to do,
2028 * we examine covering vmas in other mms: in case they
2029 * were forked from the original since ksmd passed.
2030 */
2031 if ((rmap_item->mm == vma->vm_mm) == search_new_forks)
2032 continue;
2033
2034 ret = rmap_one(page, vma, rmap_item->address, arg);
2035 if (ret != SWAP_AGAIN) {
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08002036 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002037 goto out;
2038 }
2039 }
Hugh Dickinsb6b19f22012-12-19 17:44:29 -08002040 anon_vma_unlock_read(anon_vma);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002041 }
2042 if (!search_new_forks++)
2043 goto again;
2044out:
2045 return ret;
2046}
2047
2048void ksm_migrate_page(struct page *newpage, struct page *oldpage)
2049{
2050 struct stable_node *stable_node;
2051
2052 VM_BUG_ON(!PageLocked(oldpage));
2053 VM_BUG_ON(!PageLocked(newpage));
2054 VM_BUG_ON(newpage->mapping != oldpage->mapping);
2055
2056 stable_node = page_stable_node(newpage);
2057 if (stable_node) {
Hugh Dickins62b61f62009-12-14 17:59:33 -08002058 VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage));
2059 stable_node->kpfn = page_to_pfn(newpage);
Hugh Dickinsc8d65532013-02-22 16:35:10 -08002060 /*
2061 * newpage->mapping was set in advance; now we need smp_wmb()
2062 * to make sure that the new stable_node->kpfn is visible
2063 * to get_ksm_page() before it can see that oldpage->mapping
2064 * has gone stale (or that PageSwapCache has been cleared).
2065 */
2066 smp_wmb();
2067 set_page_stable_node(oldpage, NULL);
Hugh Dickinse9995ef2009-12-14 17:59:31 -08002068 }
2069}
2070#endif /* CONFIG_MIGRATION */
2071
Hugh Dickins62b61f62009-12-14 17:59:33 -08002072#ifdef CONFIG_MEMORY_HOTREMOVE
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002073static int just_wait(void *word)
2074{
2075 schedule();
2076 return 0;
2077}
2078
2079static void wait_while_offlining(void)
2080{
2081 while (ksm_run & KSM_RUN_OFFLINE) {
2082 mutex_unlock(&ksm_thread_mutex);
2083 wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE),
2084 just_wait, TASK_UNINTERRUPTIBLE);
2085 mutex_lock(&ksm_thread_mutex);
2086 }
2087}
2088
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002089static void ksm_check_stable_tree(unsigned long start_pfn,
2090 unsigned long end_pfn)
Hugh Dickins62b61f62009-12-14 17:59:33 -08002091{
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002092 struct stable_node *stable_node;
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002093 struct list_head *this, *next;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002094 struct rb_node *node;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002095 int nid;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002096
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002097 for (nid = 0; nid < nr_node_ids; nid++) {
2098 node = rb_first(&root_stable_tree[nid]);
2099 while (node) {
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002100 stable_node = rb_entry(node, struct stable_node, node);
2101 if (stable_node->kpfn >= start_pfn &&
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002102 stable_node->kpfn < end_pfn) {
2103 /*
2104 * Don't get_ksm_page, page has already gone:
2105 * which is why we keep kpfn instead of page*
2106 */
2107 remove_node_from_stable_tree(stable_node);
2108 node = rb_first(&root_stable_tree[nid]);
2109 } else
2110 node = rb_next(node);
2111 cond_resched();
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002112 }
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002113 }
Hugh Dickins4146d2d2013-02-22 16:35:11 -08002114 list_for_each_safe(this, next, &migrate_nodes) {
2115 stable_node = list_entry(this, struct stable_node, list);
2116 if (stable_node->kpfn >= start_pfn &&
2117 stable_node->kpfn < end_pfn)
2118 remove_node_from_stable_tree(stable_node);
2119 cond_resched();
2120 }
Hugh Dickins62b61f62009-12-14 17:59:33 -08002121}
2122
2123static int ksm_memory_callback(struct notifier_block *self,
2124 unsigned long action, void *arg)
2125{
2126 struct memory_notify *mn = arg;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002127
2128 switch (action) {
2129 case MEM_GOING_OFFLINE:
2130 /*
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002131 * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items()
2132 * and remove_all_stable_nodes() while memory is going offline:
2133 * it is unsafe for them to touch the stable tree at this time.
2134 * But unmerge_ksm_pages(), rmap lookups and other entry points
2135 * which do not need the ksm_thread_mutex are all safe.
Hugh Dickins62b61f62009-12-14 17:59:33 -08002136 */
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002137 mutex_lock(&ksm_thread_mutex);
2138 ksm_run |= KSM_RUN_OFFLINE;
2139 mutex_unlock(&ksm_thread_mutex);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002140 break;
2141
2142 case MEM_OFFLINE:
2143 /*
2144 * Most of the work is done by page migration; but there might
2145 * be a few stable_nodes left over, still pointing to struct
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002146 * pages which have been offlined: prune those from the tree,
2147 * otherwise get_ksm_page() might later try to access a
2148 * non-existent struct page.
Hugh Dickins62b61f62009-12-14 17:59:33 -08002149 */
Hugh Dickinsee0ea592013-02-22 16:35:05 -08002150 ksm_check_stable_tree(mn->start_pfn,
2151 mn->start_pfn + mn->nr_pages);
Hugh Dickins62b61f62009-12-14 17:59:33 -08002152 /* fallthrough */
2153
2154 case MEM_CANCEL_OFFLINE:
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002155 mutex_lock(&ksm_thread_mutex);
2156 ksm_run &= ~KSM_RUN_OFFLINE;
Hugh Dickins62b61f62009-12-14 17:59:33 -08002157 mutex_unlock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002158
2159 smp_mb(); /* wake_up_bit advises this */
2160 wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE));
Hugh Dickins62b61f62009-12-14 17:59:33 -08002161 break;
2162 }
2163 return NOTIFY_OK;
2164}
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002165#else
2166static void wait_while_offlining(void)
2167{
2168}
Hugh Dickins62b61f62009-12-14 17:59:33 -08002169#endif /* CONFIG_MEMORY_HOTREMOVE */
2170
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002171#ifdef CONFIG_SYSFS
2172/*
2173 * This all compiles without CONFIG_SYSFS, but is a waste of space.
2174 */
2175
Izik Eidus31dbd012009-09-21 17:02:03 -07002176#define KSM_ATTR_RO(_name) \
2177 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2178#define KSM_ATTR(_name) \
2179 static struct kobj_attribute _name##_attr = \
2180 __ATTR(_name, 0644, _name##_show, _name##_store)
2181
2182static ssize_t sleep_millisecs_show(struct kobject *kobj,
2183 struct kobj_attribute *attr, char *buf)
2184{
2185 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
2186}
2187
2188static ssize_t sleep_millisecs_store(struct kobject *kobj,
2189 struct kobj_attribute *attr,
2190 const char *buf, size_t count)
2191{
2192 unsigned long msecs;
2193 int err;
2194
2195 err = strict_strtoul(buf, 10, &msecs);
2196 if (err || msecs > UINT_MAX)
2197 return -EINVAL;
2198
2199 ksm_thread_sleep_millisecs = msecs;
2200
2201 return count;
2202}
2203KSM_ATTR(sleep_millisecs);
2204
2205static ssize_t pages_to_scan_show(struct kobject *kobj,
2206 struct kobj_attribute *attr, char *buf)
2207{
2208 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
2209}
2210
2211static ssize_t pages_to_scan_store(struct kobject *kobj,
2212 struct kobj_attribute *attr,
2213 const char *buf, size_t count)
2214{
2215 int err;
2216 unsigned long nr_pages;
2217
2218 err = strict_strtoul(buf, 10, &nr_pages);
2219 if (err || nr_pages > UINT_MAX)
2220 return -EINVAL;
2221
2222 ksm_thread_pages_to_scan = nr_pages;
2223
2224 return count;
2225}
2226KSM_ATTR(pages_to_scan);
2227
2228static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
2229 char *buf)
2230{
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002231 return sprintf(buf, "%lu\n", ksm_run);
Izik Eidus31dbd012009-09-21 17:02:03 -07002232}
2233
2234static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
2235 const char *buf, size_t count)
2236{
2237 int err;
2238 unsigned long flags;
2239
2240 err = strict_strtoul(buf, 10, &flags);
2241 if (err || flags > UINT_MAX)
2242 return -EINVAL;
2243 if (flags > KSM_RUN_UNMERGE)
2244 return -EINVAL;
2245
2246 /*
2247 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
2248 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsd0f209f2009-12-14 17:59:34 -08002249 * breaking COW to free the pages_shared (but leaves mm_slots
2250 * on the list for when ksmd may be set running again).
Izik Eidus31dbd012009-09-21 17:02:03 -07002251 */
2252
2253 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002254 wait_while_offlining();
Izik Eidus31dbd012009-09-21 17:02:03 -07002255 if (ksm_run != flags) {
2256 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07002257 if (flags & KSM_RUN_UNMERGE) {
David Rientjese1e12d22012-12-11 16:02:56 -08002258 set_current_oom_origin();
Hugh Dickinsd952b792009-09-21 17:02:16 -07002259 err = unmerge_and_remove_all_rmap_items();
David Rientjese1e12d22012-12-11 16:02:56 -08002260 clear_current_oom_origin();
Hugh Dickinsd952b792009-09-21 17:02:16 -07002261 if (err) {
2262 ksm_run = KSM_RUN_STOP;
2263 count = err;
2264 }
2265 }
Izik Eidus31dbd012009-09-21 17:02:03 -07002266 }
2267 mutex_unlock(&ksm_thread_mutex);
2268
2269 if (flags & KSM_RUN_MERGE)
2270 wake_up_interruptible(&ksm_thread_wait);
2271
2272 return count;
2273}
2274KSM_ATTR(run);
2275
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002276#ifdef CONFIG_NUMA
2277static ssize_t merge_across_nodes_show(struct kobject *kobj,
2278 struct kobj_attribute *attr, char *buf)
2279{
2280 return sprintf(buf, "%u\n", ksm_merge_across_nodes);
2281}
2282
2283static ssize_t merge_across_nodes_store(struct kobject *kobj,
2284 struct kobj_attribute *attr,
2285 const char *buf, size_t count)
2286{
2287 int err;
2288 unsigned long knob;
2289
2290 err = kstrtoul(buf, 10, &knob);
2291 if (err)
2292 return err;
2293 if (knob > 1)
2294 return -EINVAL;
2295
2296 mutex_lock(&ksm_thread_mutex);
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002297 wait_while_offlining();
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002298 if (ksm_merge_across_nodes != knob) {
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002299 if (ksm_pages_shared || remove_all_stable_nodes())
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002300 err = -EBUSY;
2301 else
2302 ksm_merge_across_nodes = knob;
2303 }
2304 mutex_unlock(&ksm_thread_mutex);
2305
2306 return err ? err : count;
2307}
2308KSM_ATTR(merge_across_nodes);
2309#endif
2310
Hugh Dickinsb4028262009-09-21 17:02:09 -07002311static ssize_t pages_shared_show(struct kobject *kobj,
2312 struct kobj_attribute *attr, char *buf)
2313{
2314 return sprintf(buf, "%lu\n", ksm_pages_shared);
2315}
2316KSM_ATTR_RO(pages_shared);
2317
2318static ssize_t pages_sharing_show(struct kobject *kobj,
2319 struct kobj_attribute *attr, char *buf)
2320{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07002321 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07002322}
2323KSM_ATTR_RO(pages_sharing);
2324
Hugh Dickins473b0ce2009-09-21 17:02:11 -07002325static ssize_t pages_unshared_show(struct kobject *kobj,
2326 struct kobj_attribute *attr, char *buf)
2327{
2328 return sprintf(buf, "%lu\n", ksm_pages_unshared);
2329}
2330KSM_ATTR_RO(pages_unshared);
2331
2332static ssize_t pages_volatile_show(struct kobject *kobj,
2333 struct kobj_attribute *attr, char *buf)
2334{
2335 long ksm_pages_volatile;
2336
2337 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
2338 - ksm_pages_sharing - ksm_pages_unshared;
2339 /*
2340 * It was not worth any locking to calculate that statistic,
2341 * but it might therefore sometimes be negative: conceal that.
2342 */
2343 if (ksm_pages_volatile < 0)
2344 ksm_pages_volatile = 0;
2345 return sprintf(buf, "%ld\n", ksm_pages_volatile);
2346}
2347KSM_ATTR_RO(pages_volatile);
2348
2349static ssize_t full_scans_show(struct kobject *kobj,
2350 struct kobj_attribute *attr, char *buf)
2351{
2352 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
2353}
2354KSM_ATTR_RO(full_scans);
2355
Izik Eidus31dbd012009-09-21 17:02:03 -07002356static struct attribute *ksm_attrs[] = {
2357 &sleep_millisecs_attr.attr,
2358 &pages_to_scan_attr.attr,
2359 &run_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07002360 &pages_shared_attr.attr,
2361 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07002362 &pages_unshared_attr.attr,
2363 &pages_volatile_attr.attr,
2364 &full_scans_attr.attr,
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002365#ifdef CONFIG_NUMA
2366 &merge_across_nodes_attr.attr,
2367#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07002368 NULL,
2369};
2370
2371static struct attribute_group ksm_attr_group = {
2372 .attrs = ksm_attrs,
2373 .name = "ksm",
2374};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002375#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002376
2377static int __init ksm_init(void)
2378{
2379 struct task_struct *ksm_thread;
2380 int err;
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002381 int nid;
Izik Eidus31dbd012009-09-21 17:02:03 -07002382
2383 err = ksm_slab_init();
2384 if (err)
2385 goto out;
2386
Petr Holasek90bd6fd2013-02-22 16:35:00 -08002387 for (nid = 0; nid < nr_node_ids; nid++)
2388 root_stable_tree[nid] = RB_ROOT;
2389
Izik Eidus31dbd012009-09-21 17:02:03 -07002390 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
2391 if (IS_ERR(ksm_thread)) {
2392 printk(KERN_ERR "ksm: creating kthread failed\n");
2393 err = PTR_ERR(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002394 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002395 }
2396
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002397#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07002398 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
2399 if (err) {
2400 printk(KERN_ERR "ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002401 kthread_stop(ksm_thread);
Lai Jiangshand9f89842010-08-09 17:20:02 -07002402 goto out_free;
Izik Eidus31dbd012009-09-21 17:02:03 -07002403 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07002404#else
2405 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
2406
Hugh Dickins2ffd8672009-09-21 17:02:23 -07002407#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07002408
Hugh Dickins62b61f62009-12-14 17:59:33 -08002409#ifdef CONFIG_MEMORY_HOTREMOVE
Hugh Dickinsef4d43a2013-02-22 16:35:16 -08002410 /* There is no significance to this priority 100 */
Hugh Dickins62b61f62009-12-14 17:59:33 -08002411 hotplug_memory_notifier(ksm_memory_callback, 100);
2412#endif
Izik Eidus31dbd012009-09-21 17:02:03 -07002413 return 0;
2414
Lai Jiangshand9f89842010-08-09 17:20:02 -07002415out_free:
Izik Eidus31dbd012009-09-21 17:02:03 -07002416 ksm_slab_free();
2417out:
2418 return err;
2419}
2420module_init(ksm_init)