blob: 2f58ceebfe8f54b89f2ddc48b676045a68f7bb58 [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>
32#include <linux/mmu_notifier.h>
Izik Eidus2c6854f2009-09-23 15:56:04 -070033#include <linux/swap.h>
Hugh Dickinsf8af4da2009-09-21 17:01:57 -070034#include <linux/ksm.h>
35
Izik Eidus31dbd012009-09-21 17:02:03 -070036#include <asm/tlbflush.h>
Hugh Dickins73848b42009-12-14 17:59:22 -080037#include "internal.h"
Izik Eidus31dbd012009-09-21 17:02:03 -070038
39/*
40 * A few notes about the KSM scanning process,
41 * to make it easier to understand the data structures below:
42 *
43 * In order to reduce excessive scanning, KSM sorts the memory pages by their
44 * contents into a data structure that holds pointers to the pages' locations.
45 *
46 * Since the contents of the pages may change at any moment, KSM cannot just
47 * insert the pages into a normal sorted tree and expect it to find anything.
48 * Therefore KSM uses two data structures - the stable and the unstable tree.
49 *
50 * The stable tree holds pointers to all the merged pages (ksm pages), sorted
51 * by their contents. Because each such page is write-protected, searching on
52 * this tree is fully assured to be working (except when pages are unmapped),
53 * and therefore this tree is called the stable tree.
54 *
55 * In addition to the stable tree, KSM uses a second data structure called the
56 * unstable tree: this tree holds pointers to pages which have been found to
57 * be "unchanged for a period of time". The unstable tree sorts these pages
58 * by their contents, but since they are not write-protected, KSM cannot rely
59 * upon the unstable tree to work correctly - the unstable tree is liable to
60 * be corrupted as its contents are modified, and so it is called unstable.
61 *
62 * KSM solves this problem by several techniques:
63 *
64 * 1) The unstable tree is flushed every time KSM completes scanning all
65 * memory areas, and then the tree is rebuilt again from the beginning.
66 * 2) KSM will only insert into the unstable tree, pages whose hash value
67 * has not changed since the previous scan of all memory areas.
68 * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the
69 * colors of the nodes and not on their contents, assuring that even when
70 * the tree gets "corrupted" it won't get out of balance, so scanning time
71 * remains the same (also, searching and inserting nodes in an rbtree uses
72 * the same algorithm, so we have no overhead when we flush and rebuild).
73 * 4) KSM never flushes the stable tree, which means that even if it were to
74 * take 10 attempts to find a page in the unstable tree, once it is found,
75 * it is secured in the stable tree. (When we scan a new page, we first
76 * compare it against the stable tree, and then against the unstable tree.)
77 */
78
79/**
80 * struct mm_slot - ksm information per mm that is being scanned
81 * @link: link to the mm_slots hash list
82 * @mm_list: link into the mm_slots list, rooted in ksm_mm_head
Hugh Dickins6514d512009-12-14 17:59:19 -080083 * @rmap_list: head for this mm_slot's singly-linked list of rmap_items
Izik Eidus31dbd012009-09-21 17:02:03 -070084 * @mm: the mm that this information is valid for
85 */
86struct mm_slot {
87 struct hlist_node link;
88 struct list_head mm_list;
Hugh Dickins6514d512009-12-14 17:59:19 -080089 struct rmap_item *rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -070090 struct mm_struct *mm;
91};
92
93/**
94 * struct ksm_scan - cursor for scanning
95 * @mm_slot: the current mm_slot we are scanning
96 * @address: the next address inside that to be scanned
Hugh Dickins6514d512009-12-14 17:59:19 -080097 * @rmap_list: link to the next rmap to be scanned in the rmap_list
Izik Eidus31dbd012009-09-21 17:02:03 -070098 * @seqnr: count of completed full scans (needed when removing unstable node)
99 *
100 * There is only the one ksm_scan instance of this cursor structure.
101 */
102struct ksm_scan {
103 struct mm_slot *mm_slot;
104 unsigned long address;
Hugh Dickins6514d512009-12-14 17:59:19 -0800105 struct rmap_item **rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700106 unsigned long seqnr;
107};
108
109/**
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800110 * struct stable_node - node of the stable rbtree
Hugh Dickins08beca42009-12-14 17:59:21 -0800111 * @page: pointer to struct page of the ksm page
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800112 * @node: rb node of this ksm page in the stable tree
113 * @hlist: hlist head of rmap_items using this ksm page
114 */
115struct stable_node {
Hugh Dickins08beca42009-12-14 17:59:21 -0800116 struct page *page;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800117 struct rb_node node;
118 struct hlist_head hlist;
119};
120
121/**
Izik Eidus31dbd012009-09-21 17:02:03 -0700122 * struct rmap_item - reverse mapping item for virtual addresses
Hugh Dickins6514d512009-12-14 17:59:19 -0800123 * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list
124 * @filler: unused space we're making available in this patch
Izik Eidus31dbd012009-09-21 17:02:03 -0700125 * @mm: the memory structure this rmap_item is pointing into
126 * @address: the virtual address this rmap_item tracks (+ flags in low bits)
127 * @oldchecksum: previous checksum of the page at that virtual address
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800128 * @node: rb node of this rmap_item in the unstable tree
129 * @head: pointer to stable_node heading this list in the stable tree
130 * @hlist: link into hlist of rmap_items hanging off that stable_node
Izik Eidus31dbd012009-09-21 17:02:03 -0700131 */
132struct rmap_item {
Hugh Dickins6514d512009-12-14 17:59:19 -0800133 struct rmap_item *rmap_list;
134 unsigned long filler;
Izik Eidus31dbd012009-09-21 17:02:03 -0700135 struct mm_struct *mm;
136 unsigned long address; /* + low bits used for flags below */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800137 unsigned int oldchecksum; /* when unstable */
Izik Eidus31dbd012009-09-21 17:02:03 -0700138 union {
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800139 struct rb_node node; /* when node of unstable tree */
140 struct { /* when listed from stable tree */
141 struct stable_node *head;
142 struct hlist_node hlist;
143 };
Izik Eidus31dbd012009-09-21 17:02:03 -0700144 };
145};
146
147#define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800148#define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */
149#define STABLE_FLAG 0x200 /* is listed from the stable tree */
Izik Eidus31dbd012009-09-21 17:02:03 -0700150
151/* The stable and unstable tree heads */
152static struct rb_root root_stable_tree = RB_ROOT;
153static struct rb_root root_unstable_tree = RB_ROOT;
154
155#define MM_SLOTS_HASH_HEADS 1024
156static struct hlist_head *mm_slots_hash;
157
158static struct mm_slot ksm_mm_head = {
159 .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list),
160};
161static struct ksm_scan ksm_scan = {
162 .mm_slot = &ksm_mm_head,
163};
164
165static struct kmem_cache *rmap_item_cache;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800166static struct kmem_cache *stable_node_cache;
Izik Eidus31dbd012009-09-21 17:02:03 -0700167static struct kmem_cache *mm_slot_cache;
168
169/* The number of nodes in the stable tree */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700170static unsigned long ksm_pages_shared;
Izik Eidus31dbd012009-09-21 17:02:03 -0700171
Hugh Dickinse178dfd2009-09-21 17:02:10 -0700172/* The number of page slots additionally sharing those nodes */
Hugh Dickinsb4028262009-09-21 17:02:09 -0700173static unsigned long ksm_pages_sharing;
Izik Eidus31dbd012009-09-21 17:02:03 -0700174
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700175/* The number of nodes in the unstable tree */
176static unsigned long ksm_pages_unshared;
177
178/* The number of rmap_items in use: to calculate pages_volatile */
179static unsigned long ksm_rmap_items;
180
Izik Eidus31dbd012009-09-21 17:02:03 -0700181/* Limit on the number of unswappable pages used */
Izik Eidus2c6854f2009-09-23 15:56:04 -0700182static unsigned long ksm_max_kernel_pages;
Izik Eidus31dbd012009-09-21 17:02:03 -0700183
184/* Number of pages ksmd should scan in one batch */
Izik Eidus2c6854f2009-09-23 15:56:04 -0700185static unsigned int ksm_thread_pages_to_scan = 100;
Izik Eidus31dbd012009-09-21 17:02:03 -0700186
187/* Milliseconds ksmd should sleep between batches */
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700188static unsigned int ksm_thread_sleep_millisecs = 20;
Izik Eidus31dbd012009-09-21 17:02:03 -0700189
190#define KSM_RUN_STOP 0
191#define KSM_RUN_MERGE 1
192#define KSM_RUN_UNMERGE 2
Izik Eidus2c6854f2009-09-23 15:56:04 -0700193static unsigned int ksm_run = KSM_RUN_STOP;
Izik Eidus31dbd012009-09-21 17:02:03 -0700194
195static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait);
196static DEFINE_MUTEX(ksm_thread_mutex);
197static DEFINE_SPINLOCK(ksm_mmlist_lock);
198
Hugh Dickins5ad64682009-12-14 17:59:24 -0800199/*
200 * Temporary hack for page_referenced_ksm() and try_to_unmap_ksm(),
201 * later we rework things a little to get the right vma to them.
202 */
203static DEFINE_SPINLOCK(ksm_fallback_vma_lock);
204static struct vm_area_struct ksm_fallback_vma;
205
Izik Eidus31dbd012009-09-21 17:02:03 -0700206#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
207 sizeof(struct __struct), __alignof__(struct __struct),\
208 (__flags), NULL)
209
210static int __init ksm_slab_init(void)
211{
212 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
213 if (!rmap_item_cache)
214 goto out;
215
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800216 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
217 if (!stable_node_cache)
218 goto out_free1;
219
Izik Eidus31dbd012009-09-21 17:02:03 -0700220 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
221 if (!mm_slot_cache)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800222 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -0700223
224 return 0;
225
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800226out_free2:
227 kmem_cache_destroy(stable_node_cache);
228out_free1:
Izik Eidus31dbd012009-09-21 17:02:03 -0700229 kmem_cache_destroy(rmap_item_cache);
230out:
231 return -ENOMEM;
232}
233
234static void __init ksm_slab_free(void)
235{
236 kmem_cache_destroy(mm_slot_cache);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800237 kmem_cache_destroy(stable_node_cache);
Izik Eidus31dbd012009-09-21 17:02:03 -0700238 kmem_cache_destroy(rmap_item_cache);
239 mm_slot_cache = NULL;
240}
241
242static inline struct rmap_item *alloc_rmap_item(void)
243{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700244 struct rmap_item *rmap_item;
245
246 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
247 if (rmap_item)
248 ksm_rmap_items++;
249 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -0700250}
251
252static inline void free_rmap_item(struct rmap_item *rmap_item)
253{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700254 ksm_rmap_items--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700255 rmap_item->mm = NULL; /* debug safety */
256 kmem_cache_free(rmap_item_cache, rmap_item);
257}
258
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800259static inline struct stable_node *alloc_stable_node(void)
260{
261 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
262}
263
264static inline void free_stable_node(struct stable_node *stable_node)
265{
266 kmem_cache_free(stable_node_cache, stable_node);
267}
268
Izik Eidus31dbd012009-09-21 17:02:03 -0700269static inline struct mm_slot *alloc_mm_slot(void)
270{
271 if (!mm_slot_cache) /* initialization failed */
272 return NULL;
273 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
274}
275
276static inline void free_mm_slot(struct mm_slot *mm_slot)
277{
278 kmem_cache_free(mm_slot_cache, mm_slot);
279}
280
281static int __init mm_slots_hash_init(void)
282{
283 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
284 GFP_KERNEL);
285 if (!mm_slots_hash)
286 return -ENOMEM;
287 return 0;
288}
289
290static void __init mm_slots_hash_free(void)
291{
292 kfree(mm_slots_hash);
293}
294
295static struct mm_slot *get_mm_slot(struct mm_struct *mm)
296{
297 struct mm_slot *mm_slot;
298 struct hlist_head *bucket;
299 struct hlist_node *node;
300
301 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
302 % MM_SLOTS_HASH_HEADS];
303 hlist_for_each_entry(mm_slot, node, bucket, link) {
304 if (mm == mm_slot->mm)
305 return mm_slot;
306 }
307 return NULL;
308}
309
310static void insert_to_mm_slots_hash(struct mm_struct *mm,
311 struct mm_slot *mm_slot)
312{
313 struct hlist_head *bucket;
314
315 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
316 % MM_SLOTS_HASH_HEADS];
317 mm_slot->mm = mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700318 hlist_add_head(&mm_slot->link, bucket);
319}
320
321static inline int in_stable_tree(struct rmap_item *rmap_item)
322{
323 return rmap_item->address & STABLE_FLAG;
324}
325
326/*
Hugh Dickinsa913e182009-09-21 17:02:26 -0700327 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
328 * page tables after it has passed through ksm_exit() - which, if necessary,
329 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
330 * a special flag: they can just back out as soon as mm_users goes to zero.
331 * ksm_test_exit() is used throughout to make this test for exit: in some
332 * places for correctness, in some places just to avoid unnecessary work.
333 */
334static inline bool ksm_test_exit(struct mm_struct *mm)
335{
336 return atomic_read(&mm->mm_users) == 0;
337}
338
339/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700340 * We use break_ksm to break COW on a ksm page: it's a stripped down
341 *
342 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
343 * put_page(page);
344 *
345 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
346 * in case the application has unmapped and remapped mm,addr meanwhile.
347 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
348 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
349 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700350static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
Izik Eidus31dbd012009-09-21 17:02:03 -0700351{
352 struct page *page;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700353 int ret = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700354
355 do {
356 cond_resched();
357 page = follow_page(vma, addr, FOLL_GET);
358 if (!page)
359 break;
360 if (PageKsm(page))
361 ret = handle_mm_fault(vma->vm_mm, vma, addr,
362 FAULT_FLAG_WRITE);
363 else
364 ret = VM_FAULT_WRITE;
365 put_page(page);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700366 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
367 /*
368 * We must loop because handle_mm_fault() may back out if there's
369 * any difficulty e.g. if pte accessed bit gets updated concurrently.
370 *
371 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
372 * COW has been broken, even if the vma does not permit VM_WRITE;
373 * but note that a concurrent fault might break PageKsm for us.
374 *
375 * VM_FAULT_SIGBUS could occur if we race with truncation of the
376 * backing file, which also invalidates anonymous pages: that's
377 * okay, that truncation will have unmapped the PageKsm for us.
378 *
379 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
380 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
381 * current task has TIF_MEMDIE set, and will be OOM killed on return
382 * to user; and ksmd, having no mm, would never be chosen for that.
383 *
384 * But if the mm is in a limited mem_cgroup, then the fault may fail
385 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
386 * even ksmd can fail in this way - though it's usually breaking ksm
387 * just to undo a merge it made a moment before, so unlikely to oom.
388 *
389 * That's a pity: we might therefore have more kernel pages allocated
390 * than we're counting as nodes in the stable tree; but ksm_do_scan
391 * will retry to break_cow on each pass, so should recover the page
392 * in due course. The important thing is to not let VM_MERGEABLE
393 * be cleared while any such pages might remain in the area.
394 */
395 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700396}
397
Hugh Dickins8dd35572009-12-14 17:59:18 -0800398static void break_cow(struct rmap_item *rmap_item)
Izik Eidus31dbd012009-09-21 17:02:03 -0700399{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800400 struct mm_struct *mm = rmap_item->mm;
401 unsigned long addr = rmap_item->address;
Izik Eidus31dbd012009-09-21 17:02:03 -0700402 struct vm_area_struct *vma;
403
Hugh Dickins81464e302009-09-21 17:02:15 -0700404 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700405 if (ksm_test_exit(mm))
406 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700407 vma = find_vma(mm, addr);
408 if (!vma || vma->vm_start > addr)
Hugh Dickins81464e302009-09-21 17:02:15 -0700409 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700410 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
Hugh Dickins81464e302009-09-21 17:02:15 -0700411 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700412 break_ksm(vma, addr);
Hugh Dickins81464e302009-09-21 17:02:15 -0700413out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700414 up_read(&mm->mmap_sem);
415}
416
417static struct page *get_mergeable_page(struct rmap_item *rmap_item)
418{
419 struct mm_struct *mm = rmap_item->mm;
420 unsigned long addr = rmap_item->address;
421 struct vm_area_struct *vma;
422 struct page *page;
423
424 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700425 if (ksm_test_exit(mm))
426 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700427 vma = find_vma(mm, addr);
428 if (!vma || vma->vm_start > addr)
429 goto out;
430 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
431 goto out;
432
433 page = follow_page(vma, addr, FOLL_GET);
434 if (!page)
435 goto out;
436 if (PageAnon(page)) {
437 flush_anon_page(vma, page, addr);
438 flush_dcache_page(page);
439 } else {
440 put_page(page);
441out: page = NULL;
442 }
443 up_read(&mm->mmap_sem);
444 return page;
445}
446
447/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700448 * Removing rmap_item from stable or unstable tree.
449 * This function will clean the information from the stable/unstable tree.
450 */
451static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
452{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800453 if (rmap_item->address & STABLE_FLAG) {
454 struct stable_node *stable_node;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800455 struct page *page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700456
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800457 stable_node = rmap_item->head;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800458 page = stable_node->page;
459 lock_page(page);
460
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800461 hlist_del(&rmap_item->hlist);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800462 if (stable_node->hlist.first) {
463 unlock_page(page);
Hugh Dickinse178dfd2009-09-21 17:02:10 -0700464 ksm_pages_sharing--;
Hugh Dickins5ad64682009-12-14 17:59:24 -0800465 } else {
466 set_page_stable_node(page, NULL);
467 unlock_page(page);
468 put_page(page);
Hugh Dickins08beca42009-12-14 17:59:21 -0800469
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800470 rb_erase(&stable_node->node, &root_stable_tree);
471 free_stable_node(stable_node);
472 ksm_pages_shared--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700473 }
474
Hugh Dickins93d17712009-12-14 17:59:16 -0800475 rmap_item->address &= PAGE_MASK;
Izik Eidus31dbd012009-09-21 17:02:03 -0700476
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800477 } else if (rmap_item->address & UNSTABLE_FLAG) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700478 unsigned char age;
479 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -0700480 * Usually ksmd can and must skip the rb_erase, because
Izik Eidus31dbd012009-09-21 17:02:03 -0700481 * root_unstable_tree was already reset to RB_ROOT.
Hugh Dickins9ba69292009-09-21 17:02:20 -0700482 * But be careful when an mm is exiting: do the rb_erase
483 * if this rmap_item was inserted by this scan, rather
484 * than left over from before.
Izik Eidus31dbd012009-09-21 17:02:03 -0700485 */
486 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
Hugh Dickinscd551f92009-09-21 17:02:17 -0700487 BUG_ON(age > 1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700488 if (!age)
489 rb_erase(&rmap_item->node, &root_unstable_tree);
Izik Eidus31dbd012009-09-21 17:02:03 -0700490
Hugh Dickins93d17712009-12-14 17:59:16 -0800491 ksm_pages_unshared--;
492 rmap_item->address &= PAGE_MASK;
493 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700494
495 cond_resched(); /* we're called from many long loops */
496}
497
Izik Eidus31dbd012009-09-21 17:02:03 -0700498static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -0800499 struct rmap_item **rmap_list)
Izik Eidus31dbd012009-09-21 17:02:03 -0700500{
Hugh Dickins6514d512009-12-14 17:59:19 -0800501 while (*rmap_list) {
502 struct rmap_item *rmap_item = *rmap_list;
503 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700504 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700505 free_rmap_item(rmap_item);
506 }
507}
508
509/*
510 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
511 * than check every pte of a given vma, the locking doesn't quite work for
512 * that - an rmap_item is assigned to the stable tree after inserting ksm
513 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
514 * rmap_items from parent to child at fork time (so as not to waste time
515 * if exit comes before the next scan reaches it).
Hugh Dickins81464e302009-09-21 17:02:15 -0700516 *
517 * Similarly, although we'd like to remove rmap_items (so updating counts
518 * and freeing memory) when unmerging an area, it's easier to leave that
519 * to the next pass of ksmd - consider, for example, how ksmd might be
520 * in cmp_and_merge_page on one of the rmap_items we would be removing.
Izik Eidus31dbd012009-09-21 17:02:03 -0700521 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700522static int unmerge_ksm_pages(struct vm_area_struct *vma,
523 unsigned long start, unsigned long end)
Izik Eidus31dbd012009-09-21 17:02:03 -0700524{
525 unsigned long addr;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700526 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700527
Hugh Dickinsd952b792009-09-21 17:02:16 -0700528 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700529 if (ksm_test_exit(vma->vm_mm))
530 break;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700531 if (signal_pending(current))
532 err = -ERESTARTSYS;
533 else
534 err = break_ksm(vma, addr);
535 }
536 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700537}
538
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700539#ifdef CONFIG_SYSFS
540/*
541 * Only called through the sysfs control interface:
542 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700543static int unmerge_and_remove_all_rmap_items(void)
Izik Eidus31dbd012009-09-21 17:02:03 -0700544{
545 struct mm_slot *mm_slot;
546 struct mm_struct *mm;
547 struct vm_area_struct *vma;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700548 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700549
Hugh Dickinsd952b792009-09-21 17:02:16 -0700550 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700551 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700552 struct mm_slot, mm_list);
553 spin_unlock(&ksm_mmlist_lock);
554
Hugh Dickins9ba69292009-09-21 17:02:20 -0700555 for (mm_slot = ksm_scan.mm_slot;
556 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700557 mm = mm_slot->mm;
558 down_read(&mm->mmap_sem);
559 for (vma = mm->mmap; vma; vma = vma->vm_next) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700560 if (ksm_test_exit(mm))
561 break;
Izik Eidus31dbd012009-09-21 17:02:03 -0700562 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
563 continue;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700564 err = unmerge_ksm_pages(vma,
565 vma->vm_start, vma->vm_end);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700566 if (err)
567 goto error;
Izik Eidus31dbd012009-09-21 17:02:03 -0700568 }
Hugh Dickins9ba69292009-09-21 17:02:20 -0700569
Hugh Dickins6514d512009-12-14 17:59:19 -0800570 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700571
572 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700573 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700574 struct mm_slot, mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700575 if (ksm_test_exit(mm)) {
576 hlist_del(&mm_slot->link);
577 list_del(&mm_slot->mm_list);
578 spin_unlock(&ksm_mmlist_lock);
579
580 free_mm_slot(mm_slot);
581 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
582 up_read(&mm->mmap_sem);
583 mmdrop(mm);
584 } else {
585 spin_unlock(&ksm_mmlist_lock);
586 up_read(&mm->mmap_sem);
587 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700588 }
589
Hugh Dickinsd952b792009-09-21 17:02:16 -0700590 ksm_scan.seqnr = 0;
Hugh Dickins9ba69292009-09-21 17:02:20 -0700591 return 0;
592
593error:
594 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700595 spin_lock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700596 ksm_scan.mm_slot = &ksm_mm_head;
Izik Eidus31dbd012009-09-21 17:02:03 -0700597 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700598 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700599}
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700600#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -0700601
Izik Eidus31dbd012009-09-21 17:02:03 -0700602static u32 calc_checksum(struct page *page)
603{
604 u32 checksum;
605 void *addr = kmap_atomic(page, KM_USER0);
606 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
607 kunmap_atomic(addr, KM_USER0);
608 return checksum;
609}
610
611static int memcmp_pages(struct page *page1, struct page *page2)
612{
613 char *addr1, *addr2;
614 int ret;
615
616 addr1 = kmap_atomic(page1, KM_USER0);
617 addr2 = kmap_atomic(page2, KM_USER1);
618 ret = memcmp(addr1, addr2, PAGE_SIZE);
619 kunmap_atomic(addr2, KM_USER1);
620 kunmap_atomic(addr1, KM_USER0);
621 return ret;
622}
623
624static inline int pages_identical(struct page *page1, struct page *page2)
625{
626 return !memcmp_pages(page1, page2);
627}
628
629static int write_protect_page(struct vm_area_struct *vma, struct page *page,
630 pte_t *orig_pte)
631{
632 struct mm_struct *mm = vma->vm_mm;
633 unsigned long addr;
634 pte_t *ptep;
635 spinlock_t *ptl;
636 int swapped;
637 int err = -EFAULT;
638
639 addr = page_address_in_vma(page, vma);
640 if (addr == -EFAULT)
641 goto out;
642
643 ptep = page_check_address(page, mm, addr, &ptl, 0);
644 if (!ptep)
645 goto out;
646
647 if (pte_write(*ptep)) {
648 pte_t entry;
649
650 swapped = PageSwapCache(page);
651 flush_cache_page(vma, addr, page_to_pfn(page));
652 /*
653 * Ok this is tricky, when get_user_pages_fast() run it doesnt
654 * take any lock, therefore the check that we are going to make
655 * with the pagecount against the mapcount is racey and
656 * O_DIRECT can happen right after the check.
657 * So we clear the pte and flush the tlb before the check
658 * this assure us that no O_DIRECT can happen after the check
659 * or in the middle of the check.
660 */
661 entry = ptep_clear_flush(vma, addr, ptep);
662 /*
663 * Check that no O_DIRECT or similar I/O is in progress on the
664 * page
665 */
Hugh Dickins31e855e2009-12-14 17:59:17 -0800666 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700667 set_pte_at_notify(mm, addr, ptep, entry);
668 goto out_unlock;
669 }
670 entry = pte_wrprotect(entry);
671 set_pte_at_notify(mm, addr, ptep, entry);
672 }
673 *orig_pte = *ptep;
674 err = 0;
675
676out_unlock:
677 pte_unmap_unlock(ptep, ptl);
678out:
679 return err;
680}
681
682/**
683 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -0800684 * @vma: vma that holds the pte pointing to page
685 * @page: the page we are replacing by kpage
686 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -0700687 * @orig_pte: the original value of the pte
688 *
689 * Returns 0 on success, -EFAULT on failure.
690 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800691static int replace_page(struct vm_area_struct *vma, struct page *page,
692 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -0700693{
694 struct mm_struct *mm = vma->vm_mm;
695 pgd_t *pgd;
696 pud_t *pud;
697 pmd_t *pmd;
698 pte_t *ptep;
699 spinlock_t *ptl;
700 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -0700701 int err = -EFAULT;
702
Hugh Dickins8dd35572009-12-14 17:59:18 -0800703 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -0700704 if (addr == -EFAULT)
705 goto out;
706
707 pgd = pgd_offset(mm, addr);
708 if (!pgd_present(*pgd))
709 goto out;
710
711 pud = pud_offset(pgd, addr);
712 if (!pud_present(*pud))
713 goto out;
714
715 pmd = pmd_offset(pud, addr);
716 if (!pmd_present(*pmd))
717 goto out;
718
719 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
720 if (!pte_same(*ptep, orig_pte)) {
721 pte_unmap_unlock(ptep, ptl);
722 goto out;
723 }
724
Hugh Dickins8dd35572009-12-14 17:59:18 -0800725 get_page(kpage);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800726 page_add_anon_rmap(kpage, vma, addr);
Izik Eidus31dbd012009-09-21 17:02:03 -0700727
728 flush_cache_page(vma, addr, pte_pfn(*ptep));
729 ptep_clear_flush(vma, addr, ptep);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800730 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
Izik Eidus31dbd012009-09-21 17:02:03 -0700731
Hugh Dickins8dd35572009-12-14 17:59:18 -0800732 page_remove_rmap(page);
733 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700734
735 pte_unmap_unlock(ptep, ptl);
736 err = 0;
737out:
738 return err;
739}
740
741/*
742 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -0800743 * @vma: the vma that holds the pte pointing to page
744 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins08beca42009-12-14 17:59:21 -0800745 * @kpage: the PageKsm page that we want to map instead of page
Izik Eidus31dbd012009-09-21 17:02:03 -0700746 *
747 * This function returns 0 if the pages were merged, -EFAULT otherwise.
748 */
749static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -0800750 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700751{
752 pte_t orig_pte = __pte(0);
753 int err = -EFAULT;
754
755 if (!(vma->vm_flags & VM_MERGEABLE))
756 goto out;
Hugh Dickins8dd35572009-12-14 17:59:18 -0800757 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700758 goto out;
759
Izik Eidus31dbd012009-09-21 17:02:03 -0700760 /*
761 * We need the page lock to read a stable PageSwapCache in
762 * write_protect_page(). We use trylock_page() instead of
763 * lock_page() because we don't want to wait here - we
764 * prefer to continue scanning and merging different pages,
765 * then come back to this page when it is unlocked.
766 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800767 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -0800768 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700769 /*
770 * If this anonymous page is mapped only here, its pte may need
771 * to be write-protected. If it's mapped elsewhere, all of its
772 * ptes are necessarily already write-protected. But in either
773 * case, we need to lock and check page_count is not raised.
774 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800775 if (write_protect_page(vma, page, &orig_pte) == 0 &&
776 pages_identical(page, kpage))
777 err = replace_page(vma, page, kpage, orig_pte);
Izik Eidus31dbd012009-09-21 17:02:03 -0700778
Hugh Dickins5ad64682009-12-14 17:59:24 -0800779 if ((vma->vm_flags & VM_LOCKED) && !err) {
Hugh Dickins73848b42009-12-14 17:59:22 -0800780 munlock_vma_page(page);
Hugh Dickins5ad64682009-12-14 17:59:24 -0800781 if (!PageMlocked(kpage)) {
782 unlock_page(page);
783 lru_add_drain();
784 lock_page(kpage);
785 mlock_vma_page(kpage);
786 page = kpage; /* for final unlock */
787 }
788 }
Hugh Dickins73848b42009-12-14 17:59:22 -0800789
Hugh Dickins8dd35572009-12-14 17:59:18 -0800790 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700791out:
792 return err;
793}
794
795/*
Hugh Dickins81464e302009-09-21 17:02:15 -0700796 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
797 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -0800798 *
799 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -0700800 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800801static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
802 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -0700803{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800804 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -0700805 struct vm_area_struct *vma;
806 int err = -EFAULT;
807
Hugh Dickins08beca42009-12-14 17:59:21 -0800808 if (page == kpage) /* ksm page forked */
809 return 0;
810
Hugh Dickins8dd35572009-12-14 17:59:18 -0800811 down_read(&mm->mmap_sem);
812 if (ksm_test_exit(mm))
813 goto out;
814 vma = find_vma(mm, rmap_item->address);
815 if (!vma || vma->vm_start > rmap_item->address)
Hugh Dickins9ba69292009-09-21 17:02:20 -0700816 goto out;
817
Hugh Dickins8dd35572009-12-14 17:59:18 -0800818 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickins81464e302009-09-21 17:02:15 -0700819out:
Hugh Dickins8dd35572009-12-14 17:59:18 -0800820 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -0700821 return err;
822}
823
824/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700825 * try_to_merge_two_pages - take two identical pages and prepare them
826 * to be merged into one page.
827 *
Hugh Dickins8dd35572009-12-14 17:59:18 -0800828 * This function returns the kpage if we successfully merged two identical
829 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -0700830 *
831 * Note that this function allocates a new kernel page: if one of the pages
832 * is already a ksm page, try_to_merge_with_ksm_page should be used.
833 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800834static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
835 struct page *page,
836 struct rmap_item *tree_rmap_item,
837 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700838{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800839 struct mm_struct *mm = rmap_item->mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700840 struct vm_area_struct *vma;
841 struct page *kpage;
842 int err = -EFAULT;
843
844 /*
845 * The number of nodes in the stable tree
846 * is the number of kernel pages that we hold.
847 */
848 if (ksm_max_kernel_pages &&
Hugh Dickinsb4028262009-09-21 17:02:09 -0700849 ksm_max_kernel_pages <= ksm_pages_shared)
Hugh Dickins8dd35572009-12-14 17:59:18 -0800850 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700851
852 kpage = alloc_page(GFP_HIGHUSER);
853 if (!kpage)
Hugh Dickins8dd35572009-12-14 17:59:18 -0800854 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700855
Hugh Dickins8dd35572009-12-14 17:59:18 -0800856 down_read(&mm->mmap_sem);
857 if (ksm_test_exit(mm))
858 goto up;
859 vma = find_vma(mm, rmap_item->address);
860 if (!vma || vma->vm_start > rmap_item->address)
861 goto up;
Izik Eidus31dbd012009-09-21 17:02:03 -0700862
Hugh Dickins8dd35572009-12-14 17:59:18 -0800863 copy_user_highpage(kpage, page, rmap_item->address, vma);
Hugh Dickins08beca42009-12-14 17:59:21 -0800864
Hugh Dickins5ad64682009-12-14 17:59:24 -0800865 SetPageDirty(kpage);
866 __SetPageUptodate(kpage);
867 SetPageSwapBacked(kpage);
Hugh Dickins08beca42009-12-14 17:59:21 -0800868 set_page_stable_node(kpage, NULL); /* mark it PageKsm */
Hugh Dickins5ad64682009-12-14 17:59:24 -0800869 lru_cache_add_lru(kpage, LRU_ACTIVE_ANON);
Hugh Dickins08beca42009-12-14 17:59:21 -0800870
Hugh Dickins8dd35572009-12-14 17:59:18 -0800871 err = try_to_merge_one_page(vma, page, kpage);
872up:
873 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700874
875 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -0800876 err = try_to_merge_with_ksm_page(tree_rmap_item,
877 tree_page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -0700878 /*
Hugh Dickins81464e302009-09-21 17:02:15 -0700879 * If that fails, we have a ksm page with only one pte
880 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -0700881 */
882 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -0800883 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700884 }
Hugh Dickins8dd35572009-12-14 17:59:18 -0800885 if (err) {
886 put_page(kpage);
887 kpage = NULL;
888 }
889 return kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -0700890}
891
892/*
Hugh Dickins8dd35572009-12-14 17:59:18 -0800893 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -0700894 *
895 * This function checks if there is a page inside the stable tree
896 * with identical content to the page that we are scanning right now.
897 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800898 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -0700899 * NULL otherwise.
900 */
Hugh Dickins08beca42009-12-14 17:59:21 -0800901static struct stable_node *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700902{
903 struct rb_node *node = root_stable_tree.rb_node;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800904 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700905
Hugh Dickins08beca42009-12-14 17:59:21 -0800906 stable_node = page_stable_node(page);
907 if (stable_node) { /* ksm page forked */
908 get_page(page);
909 return stable_node;
910 }
911
Izik Eidus31dbd012009-09-21 17:02:03 -0700912 while (node) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700913 int ret;
914
Hugh Dickins08beca42009-12-14 17:59:21 -0800915 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800916 stable_node = rb_entry(node, struct stable_node, node);
Izik Eidus31dbd012009-09-21 17:02:03 -0700917
Hugh Dickins08beca42009-12-14 17:59:21 -0800918 ret = memcmp_pages(page, stable_node->page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700919
Hugh Dickins08beca42009-12-14 17:59:21 -0800920 if (ret < 0)
Izik Eidus31dbd012009-09-21 17:02:03 -0700921 node = node->rb_left;
Hugh Dickins08beca42009-12-14 17:59:21 -0800922 else if (ret > 0)
Izik Eidus31dbd012009-09-21 17:02:03 -0700923 node = node->rb_right;
Hugh Dickins08beca42009-12-14 17:59:21 -0800924 else {
925 get_page(stable_node->page);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800926 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700927 }
928 }
929
930 return NULL;
931}
932
933/*
934 * stable_tree_insert - insert rmap_item pointing to new ksm page
935 * into the stable tree.
936 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800937 * This function returns the stable tree node just allocated on success,
938 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -0700939 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800940static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700941{
942 struct rb_node **new = &root_stable_tree.rb_node;
943 struct rb_node *parent = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800944 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700945
946 while (*new) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700947 int ret;
948
Hugh Dickins08beca42009-12-14 17:59:21 -0800949 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800950 stable_node = rb_entry(*new, struct stable_node, node);
Izik Eidus31dbd012009-09-21 17:02:03 -0700951
Hugh Dickins08beca42009-12-14 17:59:21 -0800952 ret = memcmp_pages(kpage, stable_node->page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700953
954 parent = *new;
955 if (ret < 0)
956 new = &parent->rb_left;
957 else if (ret > 0)
958 new = &parent->rb_right;
959 else {
960 /*
961 * It is not a bug that stable_tree_search() didn't
962 * find this node: because at that time our page was
963 * not yet write-protected, so may have changed since.
964 */
965 return NULL;
966 }
967 }
968
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800969 stable_node = alloc_stable_node();
970 if (!stable_node)
971 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700972
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800973 rb_link_node(&stable_node->node, parent, new);
974 rb_insert_color(&stable_node->node, &root_stable_tree);
975
976 INIT_HLIST_HEAD(&stable_node->hlist);
977
Hugh Dickins08beca42009-12-14 17:59:21 -0800978 get_page(kpage);
979 stable_node->page = kpage;
980 set_page_stable_node(kpage, stable_node);
981
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800982 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700983}
984
985/*
Hugh Dickins8dd35572009-12-14 17:59:18 -0800986 * unstable_tree_search_insert - search for identical page,
987 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -0700988 *
989 * This function searches for a page in the unstable tree identical to the
990 * page currently being scanned; and if no identical page is found in the
991 * tree, we insert rmap_item as a new object into the unstable tree.
992 *
993 * This function returns pointer to rmap_item found to be identical
994 * to the currently scanned page, NULL otherwise.
995 *
996 * This function does both searching and inserting, because they share
997 * the same walking algorithm in an rbtree.
998 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800999static
1000struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
1001 struct page *page,
1002 struct page **tree_pagep)
1003
Izik Eidus31dbd012009-09-21 17:02:03 -07001004{
1005 struct rb_node **new = &root_unstable_tree.rb_node;
1006 struct rb_node *parent = NULL;
1007
1008 while (*new) {
1009 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001010 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001011 int ret;
1012
Hugh Dickinsd178f272009-11-09 15:58:23 +00001013 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -07001014 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -08001015 tree_page = get_mergeable_page(tree_rmap_item);
1016 if (!tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -07001017 return NULL;
1018
1019 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -08001020 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -07001021 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001022 if (page == tree_page) {
1023 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001024 return NULL;
1025 }
1026
Hugh Dickins8dd35572009-12-14 17:59:18 -08001027 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001028
1029 parent = *new;
1030 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001031 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001032 new = &parent->rb_left;
1033 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001034 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001035 new = &parent->rb_right;
1036 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001037 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001038 return tree_rmap_item;
1039 }
1040 }
1041
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001042 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001043 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1044 rb_link_node(&rmap_item->node, parent, new);
1045 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1046
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001047 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001048 return NULL;
1049}
1050
1051/*
1052 * stable_tree_append - add another rmap_item to the linked list of
1053 * rmap_items hanging off a given node of the stable tree, all sharing
1054 * the same ksm page.
1055 */
1056static void stable_tree_append(struct rmap_item *rmap_item,
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001057 struct stable_node *stable_node)
Izik Eidus31dbd012009-09-21 17:02:03 -07001058{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001059 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001060 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001061 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001062
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001063 if (rmap_item->hlist.next)
1064 ksm_pages_sharing++;
1065 else
1066 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001067}
1068
1069/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001070 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1071 * if not, compare checksum to previous and if it's the same, see if page can
1072 * be inserted into the unstable tree, or merged with a page already there and
1073 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001074 *
1075 * @page: the page that we are searching identical page to.
1076 * @rmap_item: the reverse mapping into the virtual address of this page
1077 */
1078static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1079{
Izik Eidus31dbd012009-09-21 17:02:03 -07001080 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001081 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001082 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001083 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07001084 unsigned int checksum;
1085 int err;
1086
Hugh Dickins93d17712009-12-14 17:59:16 -08001087 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001088
1089 /* We first start with searching the page inside the stable tree */
Hugh Dickins08beca42009-12-14 17:59:21 -08001090 stable_node = stable_tree_search(page);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001091 if (stable_node) {
Hugh Dickins08beca42009-12-14 17:59:21 -08001092 kpage = stable_node->page;
1093 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001094 if (!err) {
1095 /*
1096 * The page was successfully merged:
1097 * add its rmap_item to the stable tree.
1098 */
Hugh Dickins5ad64682009-12-14 17:59:24 -08001099 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001100 stable_tree_append(rmap_item, stable_node);
Hugh Dickins5ad64682009-12-14 17:59:24 -08001101 unlock_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001102 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001103 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001104 return;
1105 }
1106
1107 /*
1108 * A ksm page might have got here by fork, but its other
1109 * references have already been removed from the stable tree.
Hugh Dickinsd952b792009-09-21 17:02:16 -07001110 * Or it might be left over from a break_ksm which failed
1111 * when the mem_cgroup had reached its limit: try again now.
Izik Eidus31dbd012009-09-21 17:02:03 -07001112 */
1113 if (PageKsm(page))
Hugh Dickins8dd35572009-12-14 17:59:18 -08001114 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001115
1116 /*
1117 * In case the hash value of the page was changed from the last time we
1118 * have calculated it, this page to be changed frequely, therefore we
1119 * don't want to insert it to the unstable tree, and we don't want to
1120 * waste our time to search if there is something identical to it there.
1121 */
1122 checksum = calc_checksum(page);
1123 if (rmap_item->oldchecksum != checksum) {
1124 rmap_item->oldchecksum = checksum;
1125 return;
1126 }
1127
Hugh Dickins8dd35572009-12-14 17:59:18 -08001128 tree_rmap_item =
1129 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001130 if (tree_rmap_item) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001131 kpage = try_to_merge_two_pages(rmap_item, page,
1132 tree_rmap_item, tree_page);
1133 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001134 /*
1135 * As soon as we merge this page, we want to remove the
1136 * rmap_item of the page we have merged with from the unstable
1137 * tree, and insert it instead as new node in the stable tree.
1138 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001139 if (kpage) {
Hugh Dickins93d17712009-12-14 17:59:16 -08001140 remove_rmap_item_from_tree(tree_rmap_item);
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001141
Hugh Dickins5ad64682009-12-14 17:59:24 -08001142 lock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001143 stable_node = stable_tree_insert(kpage);
1144 if (stable_node) {
1145 stable_tree_append(tree_rmap_item, stable_node);
1146 stable_tree_append(rmap_item, stable_node);
1147 }
Hugh Dickins5ad64682009-12-14 17:59:24 -08001148 unlock_page(kpage);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001149 put_page(kpage);
1150
Izik Eidus31dbd012009-09-21 17:02:03 -07001151 /*
1152 * If we fail to insert the page into the stable tree,
1153 * we will have 2 virtual addresses that are pointing
1154 * to a ksm page left outside the stable tree,
1155 * in which case we need to break_cow on both.
1156 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001157 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001158 break_cow(tree_rmap_item);
1159 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001160 }
1161 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001162 }
1163}
1164
1165static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001166 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07001167 unsigned long addr)
1168{
1169 struct rmap_item *rmap_item;
1170
Hugh Dickins6514d512009-12-14 17:59:19 -08001171 while (*rmap_list) {
1172 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08001173 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07001174 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001175 if (rmap_item->address > addr)
1176 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08001177 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001178 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001179 free_rmap_item(rmap_item);
1180 }
1181
1182 rmap_item = alloc_rmap_item();
1183 if (rmap_item) {
1184 /* It has already been zeroed */
1185 rmap_item->mm = mm_slot->mm;
1186 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08001187 rmap_item->rmap_list = *rmap_list;
1188 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001189 }
1190 return rmap_item;
1191}
1192
1193static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1194{
1195 struct mm_struct *mm;
1196 struct mm_slot *slot;
1197 struct vm_area_struct *vma;
1198 struct rmap_item *rmap_item;
1199
1200 if (list_empty(&ksm_mm_head.mm_list))
1201 return NULL;
1202
1203 slot = ksm_scan.mm_slot;
1204 if (slot == &ksm_mm_head) {
1205 root_unstable_tree = RB_ROOT;
1206
1207 spin_lock(&ksm_mmlist_lock);
1208 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1209 ksm_scan.mm_slot = slot;
1210 spin_unlock(&ksm_mmlist_lock);
1211next_mm:
1212 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001213 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001214 }
1215
1216 mm = slot->mm;
1217 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001218 if (ksm_test_exit(mm))
1219 vma = NULL;
1220 else
1221 vma = find_vma(mm, ksm_scan.address);
1222
1223 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001224 if (!(vma->vm_flags & VM_MERGEABLE))
1225 continue;
1226 if (ksm_scan.address < vma->vm_start)
1227 ksm_scan.address = vma->vm_start;
1228 if (!vma->anon_vma)
1229 ksm_scan.address = vma->vm_end;
1230
1231 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001232 if (ksm_test_exit(mm))
1233 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001234 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1235 if (*page && PageAnon(*page)) {
1236 flush_anon_page(vma, *page, ksm_scan.address);
1237 flush_dcache_page(*page);
1238 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001239 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07001240 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001241 ksm_scan.rmap_list =
1242 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001243 ksm_scan.address += PAGE_SIZE;
1244 } else
1245 put_page(*page);
1246 up_read(&mm->mmap_sem);
1247 return rmap_item;
1248 }
1249 if (*page)
1250 put_page(*page);
1251 ksm_scan.address += PAGE_SIZE;
1252 cond_resched();
1253 }
1254 }
1255
Hugh Dickins9ba69292009-09-21 17:02:20 -07001256 if (ksm_test_exit(mm)) {
1257 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001258 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001259 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001260 /*
1261 * Nuke all the rmap_items that are above this current rmap:
1262 * because there were no VM_MERGEABLE vmas with such addresses.
1263 */
Hugh Dickins6514d512009-12-14 17:59:19 -08001264 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001265
1266 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001267 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1268 struct mm_slot, mm_list);
1269 if (ksm_scan.address == 0) {
1270 /*
1271 * We've completed a full scan of all vmas, holding mmap_sem
1272 * throughout, and found no VM_MERGEABLE: so do the same as
1273 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07001274 * This applies either when cleaning up after __ksm_exit
1275 * (but beware: we can reach here even before __ksm_exit),
1276 * or when all VM_MERGEABLE areas have been unmapped (and
1277 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07001278 */
1279 hlist_del(&slot->link);
1280 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001281 spin_unlock(&ksm_mmlist_lock);
1282
Hugh Dickinscd551f92009-09-21 17:02:17 -07001283 free_mm_slot(slot);
1284 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001285 up_read(&mm->mmap_sem);
1286 mmdrop(mm);
1287 } else {
1288 spin_unlock(&ksm_mmlist_lock);
1289 up_read(&mm->mmap_sem);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001290 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001291
1292 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07001293 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07001294 if (slot != &ksm_mm_head)
1295 goto next_mm;
1296
Izik Eidus31dbd012009-09-21 17:02:03 -07001297 ksm_scan.seqnr++;
1298 return NULL;
1299}
1300
1301/**
1302 * ksm_do_scan - the ksm scanner main worker function.
1303 * @scan_npages - number of pages we want to scan before we return.
1304 */
1305static void ksm_do_scan(unsigned int scan_npages)
1306{
1307 struct rmap_item *rmap_item;
1308 struct page *page;
1309
1310 while (scan_npages--) {
1311 cond_resched();
1312 rmap_item = scan_get_next_rmap_item(&page);
1313 if (!rmap_item)
1314 return;
1315 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1316 cmp_and_merge_page(page, rmap_item);
1317 put_page(page);
1318 }
1319}
1320
Hugh Dickins6e158382009-09-21 17:02:14 -07001321static int ksmd_should_run(void)
1322{
1323 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1324}
1325
Izik Eidus31dbd012009-09-21 17:02:03 -07001326static int ksm_scan_thread(void *nothing)
1327{
Izik Eidus339aa622009-09-21 17:02:07 -07001328 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07001329
1330 while (!kthread_should_stop()) {
Hugh Dickins6e158382009-09-21 17:02:14 -07001331 mutex_lock(&ksm_thread_mutex);
1332 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07001333 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e158382009-09-21 17:02:14 -07001334 mutex_unlock(&ksm_thread_mutex);
1335
1336 if (ksmd_should_run()) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001337 schedule_timeout_interruptible(
1338 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1339 } else {
1340 wait_event_interruptible(ksm_thread_wait,
Hugh Dickins6e158382009-09-21 17:02:14 -07001341 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07001342 }
1343 }
1344 return 0;
1345}
1346
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001347int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1348 unsigned long end, int advice, unsigned long *vm_flags)
1349{
1350 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001351 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001352
1353 switch (advice) {
1354 case MADV_MERGEABLE:
1355 /*
1356 * Be somewhat over-protective for now!
1357 */
1358 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1359 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1360 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
Hugh Dickins5ad64682009-12-14 17:59:24 -08001361 VM_NONLINEAR | VM_MIXEDMAP | VM_SAO))
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001362 return 0; /* just ignore the advice */
1363
Hugh Dickinsd952b792009-09-21 17:02:16 -07001364 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1365 err = __ksm_enter(mm);
1366 if (err)
1367 return err;
1368 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001369
1370 *vm_flags |= VM_MERGEABLE;
1371 break;
1372
1373 case MADV_UNMERGEABLE:
1374 if (!(*vm_flags & VM_MERGEABLE))
1375 return 0; /* just ignore the advice */
1376
Hugh Dickinsd952b792009-09-21 17:02:16 -07001377 if (vma->anon_vma) {
1378 err = unmerge_ksm_pages(vma, start, end);
1379 if (err)
1380 return err;
1381 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001382
1383 *vm_flags &= ~VM_MERGEABLE;
1384 break;
1385 }
1386
1387 return 0;
1388}
1389
1390int __ksm_enter(struct mm_struct *mm)
1391{
Hugh Dickins6e158382009-09-21 17:02:14 -07001392 struct mm_slot *mm_slot;
1393 int needs_wakeup;
1394
1395 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07001396 if (!mm_slot)
1397 return -ENOMEM;
1398
Hugh Dickins6e158382009-09-21 17:02:14 -07001399 /* Check ksm_run too? Would need tighter locking */
1400 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1401
Izik Eidus31dbd012009-09-21 17:02:03 -07001402 spin_lock(&ksm_mmlist_lock);
1403 insert_to_mm_slots_hash(mm, mm_slot);
1404 /*
1405 * Insert just behind the scanning cursor, to let the area settle
1406 * down a little; when fork is followed by immediate exec, we don't
1407 * want ksmd to waste time setting up and tearing down an rmap_list.
1408 */
1409 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1410 spin_unlock(&ksm_mmlist_lock);
1411
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001412 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001413 atomic_inc(&mm->mm_count);
Hugh Dickins6e158382009-09-21 17:02:14 -07001414
1415 if (needs_wakeup)
1416 wake_up_interruptible(&ksm_thread_wait);
1417
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001418 return 0;
1419}
1420
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07001421void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001422{
Hugh Dickinscd551f92009-09-21 17:02:17 -07001423 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001424 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07001425
Izik Eidus31dbd012009-09-21 17:02:03 -07001426 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07001427 * This process is exiting: if it's straightforward (as is the
1428 * case when ksmd was never running), free mm_slot immediately.
1429 * But if it's at the cursor or has rmap_items linked to it, use
1430 * mmap_sem to synchronize with any break_cows before pagetables
1431 * are freed, and leave the mm_slot on the list for ksmd to free.
1432 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07001433 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07001434
Hugh Dickinscd551f92009-09-21 17:02:17 -07001435 spin_lock(&ksm_mmlist_lock);
1436 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001437 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001438 if (!mm_slot->rmap_list) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001439 hlist_del(&mm_slot->link);
1440 list_del(&mm_slot->mm_list);
1441 easy_to_free = 1;
1442 } else {
1443 list_move(&mm_slot->mm_list,
1444 &ksm_scan.mm_slot->mm_list);
1445 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001446 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001447 spin_unlock(&ksm_mmlist_lock);
1448
Hugh Dickins9ba69292009-09-21 17:02:20 -07001449 if (easy_to_free) {
1450 free_mm_slot(mm_slot);
1451 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1452 mmdrop(mm);
1453 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001454 down_write(&mm->mmap_sem);
1455 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001456 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001457}
Izik Eidus31dbd012009-09-21 17:02:03 -07001458
Hugh Dickins5ad64682009-12-14 17:59:24 -08001459struct page *ksm_does_need_to_copy(struct page *page,
1460 struct vm_area_struct *vma, unsigned long address)
1461{
1462 struct page *new_page;
1463
1464 unlock_page(page); /* any racers will COW it, not modify it */
1465
1466 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1467 if (new_page) {
1468 copy_user_highpage(new_page, page, address, vma);
1469
1470 SetPageDirty(new_page);
1471 __SetPageUptodate(new_page);
1472 SetPageSwapBacked(new_page);
1473 __set_page_locked(new_page);
1474
1475 if (page_evictable(new_page, vma))
1476 lru_cache_add_lru(new_page, LRU_ACTIVE_ANON);
1477 else
1478 add_page_to_unevictable_list(new_page);
1479 }
1480
1481 page_cache_release(page);
1482 return new_page;
1483}
1484
1485int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg,
1486 unsigned long *vm_flags)
1487{
1488 struct stable_node *stable_node;
1489 struct rmap_item *rmap_item;
1490 struct hlist_node *hlist;
1491 unsigned int mapcount = page_mapcount(page);
1492 int referenced = 0;
1493 struct vm_area_struct *vma;
1494
1495 VM_BUG_ON(!PageKsm(page));
1496 VM_BUG_ON(!PageLocked(page));
1497
1498 stable_node = page_stable_node(page);
1499 if (!stable_node)
1500 return 0;
1501
1502 /*
1503 * Temporary hack: really we need anon_vma in rmap_item, to
1504 * provide the correct vma, and to find recently forked instances.
1505 * Use zalloc to avoid weirdness if any other fields are involved.
1506 */
1507 vma = kmem_cache_zalloc(vm_area_cachep, GFP_ATOMIC);
1508 if (!vma) {
1509 spin_lock(&ksm_fallback_vma_lock);
1510 vma = &ksm_fallback_vma;
1511 }
1512
1513 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1514 if (memcg && !mm_match_cgroup(rmap_item->mm, memcg))
1515 continue;
1516
1517 vma->vm_mm = rmap_item->mm;
1518 vma->vm_start = rmap_item->address;
1519 vma->vm_end = vma->vm_start + PAGE_SIZE;
1520
1521 referenced += page_referenced_one(page, vma,
1522 rmap_item->address, &mapcount, vm_flags);
1523 if (!mapcount)
1524 goto out;
1525 }
1526out:
1527 if (vma == &ksm_fallback_vma)
1528 spin_unlock(&ksm_fallback_vma_lock);
1529 else
1530 kmem_cache_free(vm_area_cachep, vma);
1531 return referenced;
1532}
1533
1534int try_to_unmap_ksm(struct page *page, enum ttu_flags flags)
1535{
1536 struct stable_node *stable_node;
1537 struct hlist_node *hlist;
1538 struct rmap_item *rmap_item;
1539 int ret = SWAP_AGAIN;
1540 struct vm_area_struct *vma;
1541
1542 VM_BUG_ON(!PageKsm(page));
1543 VM_BUG_ON(!PageLocked(page));
1544
1545 stable_node = page_stable_node(page);
1546 if (!stable_node)
1547 return SWAP_FAIL;
1548
1549 /*
1550 * Temporary hack: really we need anon_vma in rmap_item, to
1551 * provide the correct vma, and to find recently forked instances.
1552 * Use zalloc to avoid weirdness if any other fields are involved.
1553 */
1554 if (TTU_ACTION(flags) != TTU_UNMAP)
1555 return SWAP_FAIL;
1556
1557 vma = kmem_cache_zalloc(vm_area_cachep, GFP_ATOMIC);
1558 if (!vma) {
1559 spin_lock(&ksm_fallback_vma_lock);
1560 vma = &ksm_fallback_vma;
1561 }
1562
1563 hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) {
1564 vma->vm_mm = rmap_item->mm;
1565 vma->vm_start = rmap_item->address;
1566 vma->vm_end = vma->vm_start + PAGE_SIZE;
1567
1568 ret = try_to_unmap_one(page, vma, rmap_item->address, flags);
1569 if (ret != SWAP_AGAIN || !page_mapped(page))
1570 goto out;
1571 }
1572out:
1573 if (vma == &ksm_fallback_vma)
1574 spin_unlock(&ksm_fallback_vma_lock);
1575 else
1576 kmem_cache_free(vm_area_cachep, vma);
1577 return ret;
1578}
1579
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001580#ifdef CONFIG_SYSFS
1581/*
1582 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1583 */
1584
Izik Eidus31dbd012009-09-21 17:02:03 -07001585#define KSM_ATTR_RO(_name) \
1586 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1587#define KSM_ATTR(_name) \
1588 static struct kobj_attribute _name##_attr = \
1589 __ATTR(_name, 0644, _name##_show, _name##_store)
1590
1591static ssize_t sleep_millisecs_show(struct kobject *kobj,
1592 struct kobj_attribute *attr, char *buf)
1593{
1594 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1595}
1596
1597static ssize_t sleep_millisecs_store(struct kobject *kobj,
1598 struct kobj_attribute *attr,
1599 const char *buf, size_t count)
1600{
1601 unsigned long msecs;
1602 int err;
1603
1604 err = strict_strtoul(buf, 10, &msecs);
1605 if (err || msecs > UINT_MAX)
1606 return -EINVAL;
1607
1608 ksm_thread_sleep_millisecs = msecs;
1609
1610 return count;
1611}
1612KSM_ATTR(sleep_millisecs);
1613
1614static ssize_t pages_to_scan_show(struct kobject *kobj,
1615 struct kobj_attribute *attr, char *buf)
1616{
1617 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1618}
1619
1620static ssize_t pages_to_scan_store(struct kobject *kobj,
1621 struct kobj_attribute *attr,
1622 const char *buf, size_t count)
1623{
1624 int err;
1625 unsigned long nr_pages;
1626
1627 err = strict_strtoul(buf, 10, &nr_pages);
1628 if (err || nr_pages > UINT_MAX)
1629 return -EINVAL;
1630
1631 ksm_thread_pages_to_scan = nr_pages;
1632
1633 return count;
1634}
1635KSM_ATTR(pages_to_scan);
1636
1637static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1638 char *buf)
1639{
1640 return sprintf(buf, "%u\n", ksm_run);
1641}
1642
1643static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1644 const char *buf, size_t count)
1645{
1646 int err;
1647 unsigned long flags;
1648
1649 err = strict_strtoul(buf, 10, &flags);
1650 if (err || flags > UINT_MAX)
1651 return -EINVAL;
1652 if (flags > KSM_RUN_UNMERGE)
1653 return -EINVAL;
1654
1655 /*
1656 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1657 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsb4028262009-09-21 17:02:09 -07001658 * breaking COW to free the unswappable pages_shared (but leaves
Izik Eidus31dbd012009-09-21 17:02:03 -07001659 * mm_slots on the list for when ksmd may be set running again).
1660 */
1661
1662 mutex_lock(&ksm_thread_mutex);
1663 if (ksm_run != flags) {
1664 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001665 if (flags & KSM_RUN_UNMERGE) {
Hugh Dickins35451be2009-09-21 17:02:27 -07001666 current->flags |= PF_OOM_ORIGIN;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001667 err = unmerge_and_remove_all_rmap_items();
Hugh Dickins35451be2009-09-21 17:02:27 -07001668 current->flags &= ~PF_OOM_ORIGIN;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001669 if (err) {
1670 ksm_run = KSM_RUN_STOP;
1671 count = err;
1672 }
1673 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001674 }
1675 mutex_unlock(&ksm_thread_mutex);
1676
1677 if (flags & KSM_RUN_MERGE)
1678 wake_up_interruptible(&ksm_thread_wait);
1679
1680 return count;
1681}
1682KSM_ATTR(run);
1683
Izik Eidus31dbd012009-09-21 17:02:03 -07001684static ssize_t max_kernel_pages_store(struct kobject *kobj,
1685 struct kobj_attribute *attr,
1686 const char *buf, size_t count)
1687{
1688 int err;
1689 unsigned long nr_pages;
1690
1691 err = strict_strtoul(buf, 10, &nr_pages);
1692 if (err)
1693 return -EINVAL;
1694
1695 ksm_max_kernel_pages = nr_pages;
1696
1697 return count;
1698}
1699
1700static ssize_t max_kernel_pages_show(struct kobject *kobj,
1701 struct kobj_attribute *attr, char *buf)
1702{
1703 return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1704}
1705KSM_ATTR(max_kernel_pages);
1706
Hugh Dickinsb4028262009-09-21 17:02:09 -07001707static ssize_t pages_shared_show(struct kobject *kobj,
1708 struct kobj_attribute *attr, char *buf)
1709{
1710 return sprintf(buf, "%lu\n", ksm_pages_shared);
1711}
1712KSM_ATTR_RO(pages_shared);
1713
1714static ssize_t pages_sharing_show(struct kobject *kobj,
1715 struct kobj_attribute *attr, char *buf)
1716{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001717 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07001718}
1719KSM_ATTR_RO(pages_sharing);
1720
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001721static ssize_t pages_unshared_show(struct kobject *kobj,
1722 struct kobj_attribute *attr, char *buf)
1723{
1724 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1725}
1726KSM_ATTR_RO(pages_unshared);
1727
1728static ssize_t pages_volatile_show(struct kobject *kobj,
1729 struct kobj_attribute *attr, char *buf)
1730{
1731 long ksm_pages_volatile;
1732
1733 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1734 - ksm_pages_sharing - ksm_pages_unshared;
1735 /*
1736 * It was not worth any locking to calculate that statistic,
1737 * but it might therefore sometimes be negative: conceal that.
1738 */
1739 if (ksm_pages_volatile < 0)
1740 ksm_pages_volatile = 0;
1741 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1742}
1743KSM_ATTR_RO(pages_volatile);
1744
1745static ssize_t full_scans_show(struct kobject *kobj,
1746 struct kobj_attribute *attr, char *buf)
1747{
1748 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1749}
1750KSM_ATTR_RO(full_scans);
1751
Izik Eidus31dbd012009-09-21 17:02:03 -07001752static struct attribute *ksm_attrs[] = {
1753 &sleep_millisecs_attr.attr,
1754 &pages_to_scan_attr.attr,
1755 &run_attr.attr,
Izik Eidus31dbd012009-09-21 17:02:03 -07001756 &max_kernel_pages_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07001757 &pages_shared_attr.attr,
1758 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001759 &pages_unshared_attr.attr,
1760 &pages_volatile_attr.attr,
1761 &full_scans_attr.attr,
Izik Eidus31dbd012009-09-21 17:02:03 -07001762 NULL,
1763};
1764
1765static struct attribute_group ksm_attr_group = {
1766 .attrs = ksm_attrs,
1767 .name = "ksm",
1768};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001769#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07001770
1771static int __init ksm_init(void)
1772{
1773 struct task_struct *ksm_thread;
1774 int err;
1775
Hugh Dickinsc73602a2009-10-07 16:32:22 -07001776 ksm_max_kernel_pages = totalram_pages / 4;
Izik Eidus2c6854f2009-09-23 15:56:04 -07001777
Izik Eidus31dbd012009-09-21 17:02:03 -07001778 err = ksm_slab_init();
1779 if (err)
1780 goto out;
1781
1782 err = mm_slots_hash_init();
1783 if (err)
1784 goto out_free1;
1785
1786 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1787 if (IS_ERR(ksm_thread)) {
1788 printk(KERN_ERR "ksm: creating kthread failed\n");
1789 err = PTR_ERR(ksm_thread);
1790 goto out_free2;
1791 }
1792
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001793#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07001794 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1795 if (err) {
1796 printk(KERN_ERR "ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001797 kthread_stop(ksm_thread);
1798 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -07001799 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07001800#else
1801 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
1802
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001803#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07001804
1805 return 0;
1806
Izik Eidus31dbd012009-09-21 17:02:03 -07001807out_free2:
1808 mm_slots_hash_free();
1809out_free1:
1810 ksm_slab_free();
1811out:
1812 return err;
1813}
1814module_init(ksm_init)