blob: af5f571185d5e645a6a9b028d125e2c41145a60b [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
199#define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\
200 sizeof(struct __struct), __alignof__(struct __struct),\
201 (__flags), NULL)
202
203static int __init ksm_slab_init(void)
204{
205 rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0);
206 if (!rmap_item_cache)
207 goto out;
208
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800209 stable_node_cache = KSM_KMEM_CACHE(stable_node, 0);
210 if (!stable_node_cache)
211 goto out_free1;
212
Izik Eidus31dbd012009-09-21 17:02:03 -0700213 mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0);
214 if (!mm_slot_cache)
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800215 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -0700216
217 return 0;
218
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800219out_free2:
220 kmem_cache_destroy(stable_node_cache);
221out_free1:
Izik Eidus31dbd012009-09-21 17:02:03 -0700222 kmem_cache_destroy(rmap_item_cache);
223out:
224 return -ENOMEM;
225}
226
227static void __init ksm_slab_free(void)
228{
229 kmem_cache_destroy(mm_slot_cache);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800230 kmem_cache_destroy(stable_node_cache);
Izik Eidus31dbd012009-09-21 17:02:03 -0700231 kmem_cache_destroy(rmap_item_cache);
232 mm_slot_cache = NULL;
233}
234
235static inline struct rmap_item *alloc_rmap_item(void)
236{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700237 struct rmap_item *rmap_item;
238
239 rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL);
240 if (rmap_item)
241 ksm_rmap_items++;
242 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -0700243}
244
245static inline void free_rmap_item(struct rmap_item *rmap_item)
246{
Hugh Dickins473b0ce2009-09-21 17:02:11 -0700247 ksm_rmap_items--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700248 rmap_item->mm = NULL; /* debug safety */
249 kmem_cache_free(rmap_item_cache, rmap_item);
250}
251
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800252static inline struct stable_node *alloc_stable_node(void)
253{
254 return kmem_cache_alloc(stable_node_cache, GFP_KERNEL);
255}
256
257static inline void free_stable_node(struct stable_node *stable_node)
258{
259 kmem_cache_free(stable_node_cache, stable_node);
260}
261
Izik Eidus31dbd012009-09-21 17:02:03 -0700262static inline struct mm_slot *alloc_mm_slot(void)
263{
264 if (!mm_slot_cache) /* initialization failed */
265 return NULL;
266 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
267}
268
269static inline void free_mm_slot(struct mm_slot *mm_slot)
270{
271 kmem_cache_free(mm_slot_cache, mm_slot);
272}
273
274static int __init mm_slots_hash_init(void)
275{
276 mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
277 GFP_KERNEL);
278 if (!mm_slots_hash)
279 return -ENOMEM;
280 return 0;
281}
282
283static void __init mm_slots_hash_free(void)
284{
285 kfree(mm_slots_hash);
286}
287
288static struct mm_slot *get_mm_slot(struct mm_struct *mm)
289{
290 struct mm_slot *mm_slot;
291 struct hlist_head *bucket;
292 struct hlist_node *node;
293
294 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
295 % MM_SLOTS_HASH_HEADS];
296 hlist_for_each_entry(mm_slot, node, bucket, link) {
297 if (mm == mm_slot->mm)
298 return mm_slot;
299 }
300 return NULL;
301}
302
303static void insert_to_mm_slots_hash(struct mm_struct *mm,
304 struct mm_slot *mm_slot)
305{
306 struct hlist_head *bucket;
307
308 bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
309 % MM_SLOTS_HASH_HEADS];
310 mm_slot->mm = mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700311 hlist_add_head(&mm_slot->link, bucket);
312}
313
314static inline int in_stable_tree(struct rmap_item *rmap_item)
315{
316 return rmap_item->address & STABLE_FLAG;
317}
318
319/*
Hugh Dickinsa913e182009-09-21 17:02:26 -0700320 * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's
321 * page tables after it has passed through ksm_exit() - which, if necessary,
322 * takes mmap_sem briefly to serialize against them. ksm_exit() does not set
323 * a special flag: they can just back out as soon as mm_users goes to zero.
324 * ksm_test_exit() is used throughout to make this test for exit: in some
325 * places for correctness, in some places just to avoid unnecessary work.
326 */
327static inline bool ksm_test_exit(struct mm_struct *mm)
328{
329 return atomic_read(&mm->mm_users) == 0;
330}
331
332/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700333 * We use break_ksm to break COW on a ksm page: it's a stripped down
334 *
335 * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1)
336 * put_page(page);
337 *
338 * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma,
339 * in case the application has unmapped and remapped mm,addr meanwhile.
340 * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP
341 * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
342 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700343static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
Izik Eidus31dbd012009-09-21 17:02:03 -0700344{
345 struct page *page;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700346 int ret = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700347
348 do {
349 cond_resched();
350 page = follow_page(vma, addr, FOLL_GET);
351 if (!page)
352 break;
353 if (PageKsm(page))
354 ret = handle_mm_fault(vma->vm_mm, vma, addr,
355 FAULT_FLAG_WRITE);
356 else
357 ret = VM_FAULT_WRITE;
358 put_page(page);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700359 } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
360 /*
361 * We must loop because handle_mm_fault() may back out if there's
362 * any difficulty e.g. if pte accessed bit gets updated concurrently.
363 *
364 * VM_FAULT_WRITE is what we have been hoping for: it indicates that
365 * COW has been broken, even if the vma does not permit VM_WRITE;
366 * but note that a concurrent fault might break PageKsm for us.
367 *
368 * VM_FAULT_SIGBUS could occur if we race with truncation of the
369 * backing file, which also invalidates anonymous pages: that's
370 * okay, that truncation will have unmapped the PageKsm for us.
371 *
372 * VM_FAULT_OOM: at the time of writing (late July 2009), setting
373 * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
374 * current task has TIF_MEMDIE set, and will be OOM killed on return
375 * to user; and ksmd, having no mm, would never be chosen for that.
376 *
377 * But if the mm is in a limited mem_cgroup, then the fault may fail
378 * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
379 * even ksmd can fail in this way - though it's usually breaking ksm
380 * just to undo a merge it made a moment before, so unlikely to oom.
381 *
382 * That's a pity: we might therefore have more kernel pages allocated
383 * than we're counting as nodes in the stable tree; but ksm_do_scan
384 * will retry to break_cow on each pass, so should recover the page
385 * in due course. The important thing is to not let VM_MERGEABLE
386 * be cleared while any such pages might remain in the area.
387 */
388 return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700389}
390
Hugh Dickins8dd35572009-12-14 17:59:18 -0800391static void break_cow(struct rmap_item *rmap_item)
Izik Eidus31dbd012009-09-21 17:02:03 -0700392{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800393 struct mm_struct *mm = rmap_item->mm;
394 unsigned long addr = rmap_item->address;
Izik Eidus31dbd012009-09-21 17:02:03 -0700395 struct vm_area_struct *vma;
396
Hugh Dickins81464e302009-09-21 17:02:15 -0700397 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700398 if (ksm_test_exit(mm))
399 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700400 vma = find_vma(mm, addr);
401 if (!vma || vma->vm_start > addr)
Hugh Dickins81464e302009-09-21 17:02:15 -0700402 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700403 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
Hugh Dickins81464e302009-09-21 17:02:15 -0700404 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700405 break_ksm(vma, addr);
Hugh Dickins81464e302009-09-21 17:02:15 -0700406out:
Izik Eidus31dbd012009-09-21 17:02:03 -0700407 up_read(&mm->mmap_sem);
408}
409
410static struct page *get_mergeable_page(struct rmap_item *rmap_item)
411{
412 struct mm_struct *mm = rmap_item->mm;
413 unsigned long addr = rmap_item->address;
414 struct vm_area_struct *vma;
415 struct page *page;
416
417 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700418 if (ksm_test_exit(mm))
419 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700420 vma = find_vma(mm, addr);
421 if (!vma || vma->vm_start > addr)
422 goto out;
423 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
424 goto out;
425
426 page = follow_page(vma, addr, FOLL_GET);
427 if (!page)
428 goto out;
429 if (PageAnon(page)) {
430 flush_anon_page(vma, page, addr);
431 flush_dcache_page(page);
432 } else {
433 put_page(page);
434out: page = NULL;
435 }
436 up_read(&mm->mmap_sem);
437 return page;
438}
439
440/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700441 * Removing rmap_item from stable or unstable tree.
442 * This function will clean the information from the stable/unstable tree.
443 */
444static void remove_rmap_item_from_tree(struct rmap_item *rmap_item)
445{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800446 if (rmap_item->address & STABLE_FLAG) {
447 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700448
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800449 stable_node = rmap_item->head;
450 hlist_del(&rmap_item->hlist);
451 if (stable_node->hlist.first)
Hugh Dickinse178dfd2009-09-21 17:02:10 -0700452 ksm_pages_sharing--;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800453 else {
Hugh Dickins08beca42009-12-14 17:59:21 -0800454 set_page_stable_node(stable_node->page, NULL);
455 put_page(stable_node->page);
456
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800457 rb_erase(&stable_node->node, &root_stable_tree);
458 free_stable_node(stable_node);
459 ksm_pages_shared--;
Izik Eidus31dbd012009-09-21 17:02:03 -0700460 }
461
Hugh Dickins93d17712009-12-14 17:59:16 -0800462 rmap_item->address &= PAGE_MASK;
Izik Eidus31dbd012009-09-21 17:02:03 -0700463
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800464 } else if (rmap_item->address & UNSTABLE_FLAG) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700465 unsigned char age;
466 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -0700467 * Usually ksmd can and must skip the rb_erase, because
Izik Eidus31dbd012009-09-21 17:02:03 -0700468 * root_unstable_tree was already reset to RB_ROOT.
Hugh Dickins9ba69292009-09-21 17:02:20 -0700469 * But be careful when an mm is exiting: do the rb_erase
470 * if this rmap_item was inserted by this scan, rather
471 * than left over from before.
Izik Eidus31dbd012009-09-21 17:02:03 -0700472 */
473 age = (unsigned char)(ksm_scan.seqnr - rmap_item->address);
Hugh Dickinscd551f92009-09-21 17:02:17 -0700474 BUG_ON(age > 1);
Izik Eidus31dbd012009-09-21 17:02:03 -0700475 if (!age)
476 rb_erase(&rmap_item->node, &root_unstable_tree);
Izik Eidus31dbd012009-09-21 17:02:03 -0700477
Hugh Dickins93d17712009-12-14 17:59:16 -0800478 ksm_pages_unshared--;
479 rmap_item->address &= PAGE_MASK;
480 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700481
482 cond_resched(); /* we're called from many long loops */
483}
484
Izik Eidus31dbd012009-09-21 17:02:03 -0700485static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -0800486 struct rmap_item **rmap_list)
Izik Eidus31dbd012009-09-21 17:02:03 -0700487{
Hugh Dickins6514d512009-12-14 17:59:19 -0800488 while (*rmap_list) {
489 struct rmap_item *rmap_item = *rmap_list;
490 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -0700491 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700492 free_rmap_item(rmap_item);
493 }
494}
495
496/*
497 * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather
498 * than check every pte of a given vma, the locking doesn't quite work for
499 * that - an rmap_item is assigned to the stable tree after inserting ksm
500 * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing
501 * rmap_items from parent to child at fork time (so as not to waste time
502 * if exit comes before the next scan reaches it).
Hugh Dickins81464e302009-09-21 17:02:15 -0700503 *
504 * Similarly, although we'd like to remove rmap_items (so updating counts
505 * and freeing memory) when unmerging an area, it's easier to leave that
506 * to the next pass of ksmd - consider, for example, how ksmd might be
507 * in cmp_and_merge_page on one of the rmap_items we would be removing.
Izik Eidus31dbd012009-09-21 17:02:03 -0700508 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700509static int unmerge_ksm_pages(struct vm_area_struct *vma,
510 unsigned long start, unsigned long end)
Izik Eidus31dbd012009-09-21 17:02:03 -0700511{
512 unsigned long addr;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700513 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700514
Hugh Dickinsd952b792009-09-21 17:02:16 -0700515 for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700516 if (ksm_test_exit(vma->vm_mm))
517 break;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700518 if (signal_pending(current))
519 err = -ERESTARTSYS;
520 else
521 err = break_ksm(vma, addr);
522 }
523 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700524}
525
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700526#ifdef CONFIG_SYSFS
527/*
528 * Only called through the sysfs control interface:
529 */
Hugh Dickinsd952b792009-09-21 17:02:16 -0700530static int unmerge_and_remove_all_rmap_items(void)
Izik Eidus31dbd012009-09-21 17:02:03 -0700531{
532 struct mm_slot *mm_slot;
533 struct mm_struct *mm;
534 struct vm_area_struct *vma;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700535 int err = 0;
Izik Eidus31dbd012009-09-21 17:02:03 -0700536
Hugh Dickinsd952b792009-09-21 17:02:16 -0700537 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700538 ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700539 struct mm_slot, mm_list);
540 spin_unlock(&ksm_mmlist_lock);
541
Hugh Dickins9ba69292009-09-21 17:02:20 -0700542 for (mm_slot = ksm_scan.mm_slot;
543 mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700544 mm = mm_slot->mm;
545 down_read(&mm->mmap_sem);
546 for (vma = mm->mmap; vma; vma = vma->vm_next) {
Hugh Dickins9ba69292009-09-21 17:02:20 -0700547 if (ksm_test_exit(mm))
548 break;
Izik Eidus31dbd012009-09-21 17:02:03 -0700549 if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
550 continue;
Hugh Dickinsd952b792009-09-21 17:02:16 -0700551 err = unmerge_ksm_pages(vma,
552 vma->vm_start, vma->vm_end);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700553 if (err)
554 goto error;
Izik Eidus31dbd012009-09-21 17:02:03 -0700555 }
Hugh Dickins9ba69292009-09-21 17:02:20 -0700556
Hugh Dickins6514d512009-12-14 17:59:19 -0800557 remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700558
559 spin_lock(&ksm_mmlist_lock);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700560 ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next,
Hugh Dickinsd952b792009-09-21 17:02:16 -0700561 struct mm_slot, mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -0700562 if (ksm_test_exit(mm)) {
563 hlist_del(&mm_slot->link);
564 list_del(&mm_slot->mm_list);
565 spin_unlock(&ksm_mmlist_lock);
566
567 free_mm_slot(mm_slot);
568 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
569 up_read(&mm->mmap_sem);
570 mmdrop(mm);
571 } else {
572 spin_unlock(&ksm_mmlist_lock);
573 up_read(&mm->mmap_sem);
574 }
Izik Eidus31dbd012009-09-21 17:02:03 -0700575 }
576
Hugh Dickinsd952b792009-09-21 17:02:16 -0700577 ksm_scan.seqnr = 0;
Hugh Dickins9ba69292009-09-21 17:02:20 -0700578 return 0;
579
580error:
581 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700582 spin_lock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700583 ksm_scan.mm_slot = &ksm_mm_head;
Izik Eidus31dbd012009-09-21 17:02:03 -0700584 spin_unlock(&ksm_mmlist_lock);
Hugh Dickinsd952b792009-09-21 17:02:16 -0700585 return err;
Izik Eidus31dbd012009-09-21 17:02:03 -0700586}
Hugh Dickins2ffd8672009-09-21 17:02:23 -0700587#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -0700588
Izik Eidus31dbd012009-09-21 17:02:03 -0700589static u32 calc_checksum(struct page *page)
590{
591 u32 checksum;
592 void *addr = kmap_atomic(page, KM_USER0);
593 checksum = jhash2(addr, PAGE_SIZE / 4, 17);
594 kunmap_atomic(addr, KM_USER0);
595 return checksum;
596}
597
598static int memcmp_pages(struct page *page1, struct page *page2)
599{
600 char *addr1, *addr2;
601 int ret;
602
603 addr1 = kmap_atomic(page1, KM_USER0);
604 addr2 = kmap_atomic(page2, KM_USER1);
605 ret = memcmp(addr1, addr2, PAGE_SIZE);
606 kunmap_atomic(addr2, KM_USER1);
607 kunmap_atomic(addr1, KM_USER0);
608 return ret;
609}
610
611static inline int pages_identical(struct page *page1, struct page *page2)
612{
613 return !memcmp_pages(page1, page2);
614}
615
616static int write_protect_page(struct vm_area_struct *vma, struct page *page,
617 pte_t *orig_pte)
618{
619 struct mm_struct *mm = vma->vm_mm;
620 unsigned long addr;
621 pte_t *ptep;
622 spinlock_t *ptl;
623 int swapped;
624 int err = -EFAULT;
625
626 addr = page_address_in_vma(page, vma);
627 if (addr == -EFAULT)
628 goto out;
629
630 ptep = page_check_address(page, mm, addr, &ptl, 0);
631 if (!ptep)
632 goto out;
633
634 if (pte_write(*ptep)) {
635 pte_t entry;
636
637 swapped = PageSwapCache(page);
638 flush_cache_page(vma, addr, page_to_pfn(page));
639 /*
640 * Ok this is tricky, when get_user_pages_fast() run it doesnt
641 * take any lock, therefore the check that we are going to make
642 * with the pagecount against the mapcount is racey and
643 * O_DIRECT can happen right after the check.
644 * So we clear the pte and flush the tlb before the check
645 * this assure us that no O_DIRECT can happen after the check
646 * or in the middle of the check.
647 */
648 entry = ptep_clear_flush(vma, addr, ptep);
649 /*
650 * Check that no O_DIRECT or similar I/O is in progress on the
651 * page
652 */
Hugh Dickins31e855e2009-12-14 17:59:17 -0800653 if (page_mapcount(page) + 1 + swapped != page_count(page)) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700654 set_pte_at_notify(mm, addr, ptep, entry);
655 goto out_unlock;
656 }
657 entry = pte_wrprotect(entry);
658 set_pte_at_notify(mm, addr, ptep, entry);
659 }
660 *orig_pte = *ptep;
661 err = 0;
662
663out_unlock:
664 pte_unmap_unlock(ptep, ptl);
665out:
666 return err;
667}
668
669/**
670 * replace_page - replace page in vma by new ksm page
Hugh Dickins8dd35572009-12-14 17:59:18 -0800671 * @vma: vma that holds the pte pointing to page
672 * @page: the page we are replacing by kpage
673 * @kpage: the ksm page we replace page by
Izik Eidus31dbd012009-09-21 17:02:03 -0700674 * @orig_pte: the original value of the pte
675 *
676 * Returns 0 on success, -EFAULT on failure.
677 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800678static int replace_page(struct vm_area_struct *vma, struct page *page,
679 struct page *kpage, pte_t orig_pte)
Izik Eidus31dbd012009-09-21 17:02:03 -0700680{
681 struct mm_struct *mm = vma->vm_mm;
682 pgd_t *pgd;
683 pud_t *pud;
684 pmd_t *pmd;
685 pte_t *ptep;
686 spinlock_t *ptl;
687 unsigned long addr;
Izik Eidus31dbd012009-09-21 17:02:03 -0700688 int err = -EFAULT;
689
Hugh Dickins8dd35572009-12-14 17:59:18 -0800690 addr = page_address_in_vma(page, vma);
Izik Eidus31dbd012009-09-21 17:02:03 -0700691 if (addr == -EFAULT)
692 goto out;
693
694 pgd = pgd_offset(mm, addr);
695 if (!pgd_present(*pgd))
696 goto out;
697
698 pud = pud_offset(pgd, addr);
699 if (!pud_present(*pud))
700 goto out;
701
702 pmd = pmd_offset(pud, addr);
703 if (!pmd_present(*pmd))
704 goto out;
705
706 ptep = pte_offset_map_lock(mm, pmd, addr, &ptl);
707 if (!pte_same(*ptep, orig_pte)) {
708 pte_unmap_unlock(ptep, ptl);
709 goto out;
710 }
711
Hugh Dickins8dd35572009-12-14 17:59:18 -0800712 get_page(kpage);
713 page_add_ksm_rmap(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -0700714
715 flush_cache_page(vma, addr, pte_pfn(*ptep));
716 ptep_clear_flush(vma, addr, ptep);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800717 set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
Izik Eidus31dbd012009-09-21 17:02:03 -0700718
Hugh Dickins8dd35572009-12-14 17:59:18 -0800719 page_remove_rmap(page);
720 put_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700721
722 pte_unmap_unlock(ptep, ptl);
723 err = 0;
724out:
725 return err;
726}
727
728/*
729 * try_to_merge_one_page - take two pages and merge them into one
Hugh Dickins8dd35572009-12-14 17:59:18 -0800730 * @vma: the vma that holds the pte pointing to page
731 * @page: the PageAnon page that we want to replace with kpage
Hugh Dickins08beca42009-12-14 17:59:21 -0800732 * @kpage: the PageKsm page that we want to map instead of page
Izik Eidus31dbd012009-09-21 17:02:03 -0700733 *
734 * This function returns 0 if the pages were merged, -EFAULT otherwise.
735 */
736static int try_to_merge_one_page(struct vm_area_struct *vma,
Hugh Dickins8dd35572009-12-14 17:59:18 -0800737 struct page *page, struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700738{
739 pte_t orig_pte = __pte(0);
740 int err = -EFAULT;
741
742 if (!(vma->vm_flags & VM_MERGEABLE))
743 goto out;
Hugh Dickins8dd35572009-12-14 17:59:18 -0800744 if (!PageAnon(page))
Izik Eidus31dbd012009-09-21 17:02:03 -0700745 goto out;
746
Izik Eidus31dbd012009-09-21 17:02:03 -0700747 /*
748 * We need the page lock to read a stable PageSwapCache in
749 * write_protect_page(). We use trylock_page() instead of
750 * lock_page() because we don't want to wait here - we
751 * prefer to continue scanning and merging different pages,
752 * then come back to this page when it is unlocked.
753 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800754 if (!trylock_page(page))
Hugh Dickins31e855e2009-12-14 17:59:17 -0800755 goto out;
Izik Eidus31dbd012009-09-21 17:02:03 -0700756 /*
757 * If this anonymous page is mapped only here, its pte may need
758 * to be write-protected. If it's mapped elsewhere, all of its
759 * ptes are necessarily already write-protected. But in either
760 * case, we need to lock and check page_count is not raised.
761 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800762 if (write_protect_page(vma, page, &orig_pte) == 0 &&
763 pages_identical(page, kpage))
764 err = replace_page(vma, page, kpage, orig_pte);
Izik Eidus31dbd012009-09-21 17:02:03 -0700765
Hugh Dickins73848b42009-12-14 17:59:22 -0800766 if ((vma->vm_flags & VM_LOCKED) && !err)
767 munlock_vma_page(page);
768
Hugh Dickins8dd35572009-12-14 17:59:18 -0800769 unlock_page(page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700770out:
771 return err;
772}
773
774/*
Hugh Dickins81464e302009-09-21 17:02:15 -0700775 * try_to_merge_with_ksm_page - like try_to_merge_two_pages,
776 * but no new kernel page is allocated: kpage must already be a ksm page.
Hugh Dickins8dd35572009-12-14 17:59:18 -0800777 *
778 * This function returns 0 if the pages were merged, -EFAULT otherwise.
Hugh Dickins81464e302009-09-21 17:02:15 -0700779 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800780static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item,
781 struct page *page, struct page *kpage)
Hugh Dickins81464e302009-09-21 17:02:15 -0700782{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800783 struct mm_struct *mm = rmap_item->mm;
Hugh Dickins81464e302009-09-21 17:02:15 -0700784 struct vm_area_struct *vma;
785 int err = -EFAULT;
786
Hugh Dickins08beca42009-12-14 17:59:21 -0800787 if (page == kpage) /* ksm page forked */
788 return 0;
789
Hugh Dickins8dd35572009-12-14 17:59:18 -0800790 down_read(&mm->mmap_sem);
791 if (ksm_test_exit(mm))
792 goto out;
793 vma = find_vma(mm, rmap_item->address);
794 if (!vma || vma->vm_start > rmap_item->address)
Hugh Dickins9ba69292009-09-21 17:02:20 -0700795 goto out;
796
Hugh Dickins8dd35572009-12-14 17:59:18 -0800797 err = try_to_merge_one_page(vma, page, kpage);
Hugh Dickins81464e302009-09-21 17:02:15 -0700798out:
Hugh Dickins8dd35572009-12-14 17:59:18 -0800799 up_read(&mm->mmap_sem);
Hugh Dickins81464e302009-09-21 17:02:15 -0700800 return err;
801}
802
803/*
Izik Eidus31dbd012009-09-21 17:02:03 -0700804 * try_to_merge_two_pages - take two identical pages and prepare them
805 * to be merged into one page.
806 *
Hugh Dickins8dd35572009-12-14 17:59:18 -0800807 * This function returns the kpage if we successfully merged two identical
808 * pages into one ksm page, NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -0700809 *
810 * Note that this function allocates a new kernel page: if one of the pages
811 * is already a ksm page, try_to_merge_with_ksm_page should be used.
812 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800813static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item,
814 struct page *page,
815 struct rmap_item *tree_rmap_item,
816 struct page *tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700817{
Hugh Dickins8dd35572009-12-14 17:59:18 -0800818 struct mm_struct *mm = rmap_item->mm;
Izik Eidus31dbd012009-09-21 17:02:03 -0700819 struct vm_area_struct *vma;
820 struct page *kpage;
821 int err = -EFAULT;
822
823 /*
824 * The number of nodes in the stable tree
825 * is the number of kernel pages that we hold.
826 */
827 if (ksm_max_kernel_pages &&
Hugh Dickinsb4028262009-09-21 17:02:09 -0700828 ksm_max_kernel_pages <= ksm_pages_shared)
Hugh Dickins8dd35572009-12-14 17:59:18 -0800829 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700830
831 kpage = alloc_page(GFP_HIGHUSER);
832 if (!kpage)
Hugh Dickins8dd35572009-12-14 17:59:18 -0800833 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700834
Hugh Dickins8dd35572009-12-14 17:59:18 -0800835 down_read(&mm->mmap_sem);
836 if (ksm_test_exit(mm))
837 goto up;
838 vma = find_vma(mm, rmap_item->address);
839 if (!vma || vma->vm_start > rmap_item->address)
840 goto up;
Izik Eidus31dbd012009-09-21 17:02:03 -0700841
Hugh Dickins8dd35572009-12-14 17:59:18 -0800842 copy_user_highpage(kpage, page, rmap_item->address, vma);
Hugh Dickins08beca42009-12-14 17:59:21 -0800843
844 set_page_stable_node(kpage, NULL); /* mark it PageKsm */
845
Hugh Dickins8dd35572009-12-14 17:59:18 -0800846 err = try_to_merge_one_page(vma, page, kpage);
847up:
848 up_read(&mm->mmap_sem);
Izik Eidus31dbd012009-09-21 17:02:03 -0700849
850 if (!err) {
Hugh Dickins8dd35572009-12-14 17:59:18 -0800851 err = try_to_merge_with_ksm_page(tree_rmap_item,
852 tree_page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -0700853 /*
Hugh Dickins81464e302009-09-21 17:02:15 -0700854 * If that fails, we have a ksm page with only one pte
855 * pointing to it: so break it.
Izik Eidus31dbd012009-09-21 17:02:03 -0700856 */
857 if (err)
Hugh Dickins8dd35572009-12-14 17:59:18 -0800858 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -0700859 }
Hugh Dickins8dd35572009-12-14 17:59:18 -0800860 if (err) {
861 put_page(kpage);
862 kpage = NULL;
863 }
864 return kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -0700865}
866
867/*
Hugh Dickins8dd35572009-12-14 17:59:18 -0800868 * stable_tree_search - search for page inside the stable tree
Izik Eidus31dbd012009-09-21 17:02:03 -0700869 *
870 * This function checks if there is a page inside the stable tree
871 * with identical content to the page that we are scanning right now.
872 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800873 * This function returns the stable tree node of identical content if found,
Izik Eidus31dbd012009-09-21 17:02:03 -0700874 * NULL otherwise.
875 */
Hugh Dickins08beca42009-12-14 17:59:21 -0800876static struct stable_node *stable_tree_search(struct page *page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700877{
878 struct rb_node *node = root_stable_tree.rb_node;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800879 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700880
Hugh Dickins08beca42009-12-14 17:59:21 -0800881 stable_node = page_stable_node(page);
882 if (stable_node) { /* ksm page forked */
883 get_page(page);
884 return stable_node;
885 }
886
Izik Eidus31dbd012009-09-21 17:02:03 -0700887 while (node) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700888 int ret;
889
Hugh Dickins08beca42009-12-14 17:59:21 -0800890 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800891 stable_node = rb_entry(node, struct stable_node, node);
Izik Eidus31dbd012009-09-21 17:02:03 -0700892
Hugh Dickins08beca42009-12-14 17:59:21 -0800893 ret = memcmp_pages(page, stable_node->page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700894
Hugh Dickins08beca42009-12-14 17:59:21 -0800895 if (ret < 0)
Izik Eidus31dbd012009-09-21 17:02:03 -0700896 node = node->rb_left;
Hugh Dickins08beca42009-12-14 17:59:21 -0800897 else if (ret > 0)
Izik Eidus31dbd012009-09-21 17:02:03 -0700898 node = node->rb_right;
Hugh Dickins08beca42009-12-14 17:59:21 -0800899 else {
900 get_page(stable_node->page);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800901 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700902 }
903 }
904
905 return NULL;
906}
907
908/*
909 * stable_tree_insert - insert rmap_item pointing to new ksm page
910 * into the stable tree.
911 *
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800912 * This function returns the stable tree node just allocated on success,
913 * NULL otherwise.
Izik Eidus31dbd012009-09-21 17:02:03 -0700914 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800915static struct stable_node *stable_tree_insert(struct page *kpage)
Izik Eidus31dbd012009-09-21 17:02:03 -0700916{
917 struct rb_node **new = &root_stable_tree.rb_node;
918 struct rb_node *parent = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800919 struct stable_node *stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700920
921 while (*new) {
Izik Eidus31dbd012009-09-21 17:02:03 -0700922 int ret;
923
Hugh Dickins08beca42009-12-14 17:59:21 -0800924 cond_resched();
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800925 stable_node = rb_entry(*new, struct stable_node, node);
Izik Eidus31dbd012009-09-21 17:02:03 -0700926
Hugh Dickins08beca42009-12-14 17:59:21 -0800927 ret = memcmp_pages(kpage, stable_node->page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700928
929 parent = *new;
930 if (ret < 0)
931 new = &parent->rb_left;
932 else if (ret > 0)
933 new = &parent->rb_right;
934 else {
935 /*
936 * It is not a bug that stable_tree_search() didn't
937 * find this node: because at that time our page was
938 * not yet write-protected, so may have changed since.
939 */
940 return NULL;
941 }
942 }
943
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800944 stable_node = alloc_stable_node();
945 if (!stable_node)
946 return NULL;
Izik Eidus31dbd012009-09-21 17:02:03 -0700947
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800948 rb_link_node(&stable_node->node, parent, new);
949 rb_insert_color(&stable_node->node, &root_stable_tree);
950
951 INIT_HLIST_HEAD(&stable_node->hlist);
952
Hugh Dickins08beca42009-12-14 17:59:21 -0800953 get_page(kpage);
954 stable_node->page = kpage;
955 set_page_stable_node(kpage, stable_node);
956
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -0800957 return stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -0700958}
959
960/*
Hugh Dickins8dd35572009-12-14 17:59:18 -0800961 * unstable_tree_search_insert - search for identical page,
962 * else insert rmap_item into the unstable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -0700963 *
964 * This function searches for a page in the unstable tree identical to the
965 * page currently being scanned; and if no identical page is found in the
966 * tree, we insert rmap_item as a new object into the unstable tree.
967 *
968 * This function returns pointer to rmap_item found to be identical
969 * to the currently scanned page, NULL otherwise.
970 *
971 * This function does both searching and inserting, because they share
972 * the same walking algorithm in an rbtree.
973 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800974static
975struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item,
976 struct page *page,
977 struct page **tree_pagep)
978
Izik Eidus31dbd012009-09-21 17:02:03 -0700979{
980 struct rb_node **new = &root_unstable_tree.rb_node;
981 struct rb_node *parent = NULL;
982
983 while (*new) {
984 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -0800985 struct page *tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -0700986 int ret;
987
Hugh Dickinsd178f272009-11-09 15:58:23 +0000988 cond_resched();
Izik Eidus31dbd012009-09-21 17:02:03 -0700989 tree_rmap_item = rb_entry(*new, struct rmap_item, node);
Hugh Dickins8dd35572009-12-14 17:59:18 -0800990 tree_page = get_mergeable_page(tree_rmap_item);
991 if (!tree_page)
Izik Eidus31dbd012009-09-21 17:02:03 -0700992 return NULL;
993
994 /*
Hugh Dickins8dd35572009-12-14 17:59:18 -0800995 * Don't substitute a ksm page for a forked page.
Izik Eidus31dbd012009-09-21 17:02:03 -0700996 */
Hugh Dickins8dd35572009-12-14 17:59:18 -0800997 if (page == tree_page) {
998 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -0700999 return NULL;
1000 }
1001
Hugh Dickins8dd35572009-12-14 17:59:18 -08001002 ret = memcmp_pages(page, tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001003
1004 parent = *new;
1005 if (ret < 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001006 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001007 new = &parent->rb_left;
1008 } else if (ret > 0) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001009 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001010 new = &parent->rb_right;
1011 } else {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001012 *tree_pagep = tree_page;
Izik Eidus31dbd012009-09-21 17:02:03 -07001013 return tree_rmap_item;
1014 }
1015 }
1016
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001017 rmap_item->address |= UNSTABLE_FLAG;
Izik Eidus31dbd012009-09-21 17:02:03 -07001018 rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK);
1019 rb_link_node(&rmap_item->node, parent, new);
1020 rb_insert_color(&rmap_item->node, &root_unstable_tree);
1021
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001022 ksm_pages_unshared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001023 return NULL;
1024}
1025
1026/*
1027 * stable_tree_append - add another rmap_item to the linked list of
1028 * rmap_items hanging off a given node of the stable tree, all sharing
1029 * the same ksm page.
1030 */
1031static void stable_tree_append(struct rmap_item *rmap_item,
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001032 struct stable_node *stable_node)
Izik Eidus31dbd012009-09-21 17:02:03 -07001033{
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001034 rmap_item->head = stable_node;
Izik Eidus31dbd012009-09-21 17:02:03 -07001035 rmap_item->address |= STABLE_FLAG;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001036 hlist_add_head(&rmap_item->hlist, &stable_node->hlist);
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001037
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001038 if (rmap_item->hlist.next)
1039 ksm_pages_sharing++;
1040 else
1041 ksm_pages_shared++;
Izik Eidus31dbd012009-09-21 17:02:03 -07001042}
1043
1044/*
Hugh Dickins81464e302009-09-21 17:02:15 -07001045 * cmp_and_merge_page - first see if page can be merged into the stable tree;
1046 * if not, compare checksum to previous and if it's the same, see if page can
1047 * be inserted into the unstable tree, or merged with a page already there and
1048 * both transferred to the stable tree.
Izik Eidus31dbd012009-09-21 17:02:03 -07001049 *
1050 * @page: the page that we are searching identical page to.
1051 * @rmap_item: the reverse mapping into the virtual address of this page
1052 */
1053static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
1054{
Izik Eidus31dbd012009-09-21 17:02:03 -07001055 struct rmap_item *tree_rmap_item;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001056 struct page *tree_page = NULL;
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001057 struct stable_node *stable_node;
Hugh Dickins8dd35572009-12-14 17:59:18 -08001058 struct page *kpage;
Izik Eidus31dbd012009-09-21 17:02:03 -07001059 unsigned int checksum;
1060 int err;
1061
Hugh Dickins93d17712009-12-14 17:59:16 -08001062 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001063
1064 /* We first start with searching the page inside the stable tree */
Hugh Dickins08beca42009-12-14 17:59:21 -08001065 stable_node = stable_tree_search(page);
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001066 if (stable_node) {
Hugh Dickins08beca42009-12-14 17:59:21 -08001067 kpage = stable_node->page;
1068 err = try_to_merge_with_ksm_page(rmap_item, page, kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001069 if (!err) {
1070 /*
1071 * The page was successfully merged:
1072 * add its rmap_item to the stable tree.
1073 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001074 stable_tree_append(rmap_item, stable_node);
Izik Eidus31dbd012009-09-21 17:02:03 -07001075 }
Hugh Dickins8dd35572009-12-14 17:59:18 -08001076 put_page(kpage);
Izik Eidus31dbd012009-09-21 17:02:03 -07001077 return;
1078 }
1079
1080 /*
1081 * A ksm page might have got here by fork, but its other
1082 * references have already been removed from the stable tree.
Hugh Dickinsd952b792009-09-21 17:02:16 -07001083 * Or it might be left over from a break_ksm which failed
1084 * when the mem_cgroup had reached its limit: try again now.
Izik Eidus31dbd012009-09-21 17:02:03 -07001085 */
1086 if (PageKsm(page))
Hugh Dickins8dd35572009-12-14 17:59:18 -08001087 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001088
1089 /*
1090 * In case the hash value of the page was changed from the last time we
1091 * have calculated it, this page to be changed frequely, therefore we
1092 * don't want to insert it to the unstable tree, and we don't want to
1093 * waste our time to search if there is something identical to it there.
1094 */
1095 checksum = calc_checksum(page);
1096 if (rmap_item->oldchecksum != checksum) {
1097 rmap_item->oldchecksum = checksum;
1098 return;
1099 }
1100
Hugh Dickins8dd35572009-12-14 17:59:18 -08001101 tree_rmap_item =
1102 unstable_tree_search_insert(rmap_item, page, &tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001103 if (tree_rmap_item) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001104 kpage = try_to_merge_two_pages(rmap_item, page,
1105 tree_rmap_item, tree_page);
1106 put_page(tree_page);
Izik Eidus31dbd012009-09-21 17:02:03 -07001107 /*
1108 * As soon as we merge this page, we want to remove the
1109 * rmap_item of the page we have merged with from the unstable
1110 * tree, and insert it instead as new node in the stable tree.
1111 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001112 if (kpage) {
Hugh Dickins93d17712009-12-14 17:59:16 -08001113 remove_rmap_item_from_tree(tree_rmap_item);
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001114
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001115 stable_node = stable_tree_insert(kpage);
1116 if (stable_node) {
1117 stable_tree_append(tree_rmap_item, stable_node);
1118 stable_tree_append(rmap_item, stable_node);
1119 }
1120 put_page(kpage);
1121
Izik Eidus31dbd012009-09-21 17:02:03 -07001122 /*
1123 * If we fail to insert the page into the stable tree,
1124 * we will have 2 virtual addresses that are pointing
1125 * to a ksm page left outside the stable tree,
1126 * in which case we need to break_cow on both.
1127 */
Hugh Dickins7b6ba2c2009-12-14 17:59:20 -08001128 if (!stable_node) {
Hugh Dickins8dd35572009-12-14 17:59:18 -08001129 break_cow(tree_rmap_item);
1130 break_cow(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001131 }
1132 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001133 }
1134}
1135
1136static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001137 struct rmap_item **rmap_list,
Izik Eidus31dbd012009-09-21 17:02:03 -07001138 unsigned long addr)
1139{
1140 struct rmap_item *rmap_item;
1141
Hugh Dickins6514d512009-12-14 17:59:19 -08001142 while (*rmap_list) {
1143 rmap_item = *rmap_list;
Hugh Dickins93d17712009-12-14 17:59:16 -08001144 if ((rmap_item->address & PAGE_MASK) == addr)
Izik Eidus31dbd012009-09-21 17:02:03 -07001145 return rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001146 if (rmap_item->address > addr)
1147 break;
Hugh Dickins6514d512009-12-14 17:59:19 -08001148 *rmap_list = rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001149 remove_rmap_item_from_tree(rmap_item);
Izik Eidus31dbd012009-09-21 17:02:03 -07001150 free_rmap_item(rmap_item);
1151 }
1152
1153 rmap_item = alloc_rmap_item();
1154 if (rmap_item) {
1155 /* It has already been zeroed */
1156 rmap_item->mm = mm_slot->mm;
1157 rmap_item->address = addr;
Hugh Dickins6514d512009-12-14 17:59:19 -08001158 rmap_item->rmap_list = *rmap_list;
1159 *rmap_list = rmap_item;
Izik Eidus31dbd012009-09-21 17:02:03 -07001160 }
1161 return rmap_item;
1162}
1163
1164static struct rmap_item *scan_get_next_rmap_item(struct page **page)
1165{
1166 struct mm_struct *mm;
1167 struct mm_slot *slot;
1168 struct vm_area_struct *vma;
1169 struct rmap_item *rmap_item;
1170
1171 if (list_empty(&ksm_mm_head.mm_list))
1172 return NULL;
1173
1174 slot = ksm_scan.mm_slot;
1175 if (slot == &ksm_mm_head) {
1176 root_unstable_tree = RB_ROOT;
1177
1178 spin_lock(&ksm_mmlist_lock);
1179 slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list);
1180 ksm_scan.mm_slot = slot;
1181 spin_unlock(&ksm_mmlist_lock);
1182next_mm:
1183 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001184 ksm_scan.rmap_list = &slot->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001185 }
1186
1187 mm = slot->mm;
1188 down_read(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001189 if (ksm_test_exit(mm))
1190 vma = NULL;
1191 else
1192 vma = find_vma(mm, ksm_scan.address);
1193
1194 for (; vma; vma = vma->vm_next) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001195 if (!(vma->vm_flags & VM_MERGEABLE))
1196 continue;
1197 if (ksm_scan.address < vma->vm_start)
1198 ksm_scan.address = vma->vm_start;
1199 if (!vma->anon_vma)
1200 ksm_scan.address = vma->vm_end;
1201
1202 while (ksm_scan.address < vma->vm_end) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001203 if (ksm_test_exit(mm))
1204 break;
Izik Eidus31dbd012009-09-21 17:02:03 -07001205 *page = follow_page(vma, ksm_scan.address, FOLL_GET);
1206 if (*page && PageAnon(*page)) {
1207 flush_anon_page(vma, *page, ksm_scan.address);
1208 flush_dcache_page(*page);
1209 rmap_item = get_next_rmap_item(slot,
Hugh Dickins6514d512009-12-14 17:59:19 -08001210 ksm_scan.rmap_list, ksm_scan.address);
Izik Eidus31dbd012009-09-21 17:02:03 -07001211 if (rmap_item) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001212 ksm_scan.rmap_list =
1213 &rmap_item->rmap_list;
Izik Eidus31dbd012009-09-21 17:02:03 -07001214 ksm_scan.address += PAGE_SIZE;
1215 } else
1216 put_page(*page);
1217 up_read(&mm->mmap_sem);
1218 return rmap_item;
1219 }
1220 if (*page)
1221 put_page(*page);
1222 ksm_scan.address += PAGE_SIZE;
1223 cond_resched();
1224 }
1225 }
1226
Hugh Dickins9ba69292009-09-21 17:02:20 -07001227 if (ksm_test_exit(mm)) {
1228 ksm_scan.address = 0;
Hugh Dickins6514d512009-12-14 17:59:19 -08001229 ksm_scan.rmap_list = &slot->rmap_list;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001230 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001231 /*
1232 * Nuke all the rmap_items that are above this current rmap:
1233 * because there were no VM_MERGEABLE vmas with such addresses.
1234 */
Hugh Dickins6514d512009-12-14 17:59:19 -08001235 remove_trailing_rmap_items(slot, ksm_scan.rmap_list);
Izik Eidus31dbd012009-09-21 17:02:03 -07001236
1237 spin_lock(&ksm_mmlist_lock);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001238 ksm_scan.mm_slot = list_entry(slot->mm_list.next,
1239 struct mm_slot, mm_list);
1240 if (ksm_scan.address == 0) {
1241 /*
1242 * We've completed a full scan of all vmas, holding mmap_sem
1243 * throughout, and found no VM_MERGEABLE: so do the same as
1244 * __ksm_exit does to remove this mm from all our lists now.
Hugh Dickins9ba69292009-09-21 17:02:20 -07001245 * This applies either when cleaning up after __ksm_exit
1246 * (but beware: we can reach here even before __ksm_exit),
1247 * or when all VM_MERGEABLE areas have been unmapped (and
1248 * mmap_sem then protects against race with MADV_MERGEABLE).
Hugh Dickinscd551f92009-09-21 17:02:17 -07001249 */
1250 hlist_del(&slot->link);
1251 list_del(&slot->mm_list);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001252 spin_unlock(&ksm_mmlist_lock);
1253
Hugh Dickinscd551f92009-09-21 17:02:17 -07001254 free_mm_slot(slot);
1255 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001256 up_read(&mm->mmap_sem);
1257 mmdrop(mm);
1258 } else {
1259 spin_unlock(&ksm_mmlist_lock);
1260 up_read(&mm->mmap_sem);
Hugh Dickinscd551f92009-09-21 17:02:17 -07001261 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001262
1263 /* Repeat until we've completed scanning the whole list */
Hugh Dickinscd551f92009-09-21 17:02:17 -07001264 slot = ksm_scan.mm_slot;
Izik Eidus31dbd012009-09-21 17:02:03 -07001265 if (slot != &ksm_mm_head)
1266 goto next_mm;
1267
Izik Eidus31dbd012009-09-21 17:02:03 -07001268 ksm_scan.seqnr++;
1269 return NULL;
1270}
1271
1272/**
1273 * ksm_do_scan - the ksm scanner main worker function.
1274 * @scan_npages - number of pages we want to scan before we return.
1275 */
1276static void ksm_do_scan(unsigned int scan_npages)
1277{
1278 struct rmap_item *rmap_item;
1279 struct page *page;
1280
1281 while (scan_npages--) {
1282 cond_resched();
1283 rmap_item = scan_get_next_rmap_item(&page);
1284 if (!rmap_item)
1285 return;
1286 if (!PageKsm(page) || !in_stable_tree(rmap_item))
1287 cmp_and_merge_page(page, rmap_item);
Hugh Dickins26465d32009-09-21 17:02:12 -07001288 else if (page_mapcount(page) == 1) {
1289 /*
1290 * Replace now-unshared ksm page by ordinary page.
1291 */
Hugh Dickins8dd35572009-12-14 17:59:18 -08001292 break_cow(rmap_item);
Hugh Dickins26465d32009-09-21 17:02:12 -07001293 remove_rmap_item_from_tree(rmap_item);
1294 rmap_item->oldchecksum = calc_checksum(page);
1295 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001296 put_page(page);
1297 }
1298}
1299
Hugh Dickins6e158382009-09-21 17:02:14 -07001300static int ksmd_should_run(void)
1301{
1302 return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list);
1303}
1304
Izik Eidus31dbd012009-09-21 17:02:03 -07001305static int ksm_scan_thread(void *nothing)
1306{
Izik Eidus339aa622009-09-21 17:02:07 -07001307 set_user_nice(current, 5);
Izik Eidus31dbd012009-09-21 17:02:03 -07001308
1309 while (!kthread_should_stop()) {
Hugh Dickins6e158382009-09-21 17:02:14 -07001310 mutex_lock(&ksm_thread_mutex);
1311 if (ksmd_should_run())
Izik Eidus31dbd012009-09-21 17:02:03 -07001312 ksm_do_scan(ksm_thread_pages_to_scan);
Hugh Dickins6e158382009-09-21 17:02:14 -07001313 mutex_unlock(&ksm_thread_mutex);
1314
1315 if (ksmd_should_run()) {
Izik Eidus31dbd012009-09-21 17:02:03 -07001316 schedule_timeout_interruptible(
1317 msecs_to_jiffies(ksm_thread_sleep_millisecs));
1318 } else {
1319 wait_event_interruptible(ksm_thread_wait,
Hugh Dickins6e158382009-09-21 17:02:14 -07001320 ksmd_should_run() || kthread_should_stop());
Izik Eidus31dbd012009-09-21 17:02:03 -07001321 }
1322 }
1323 return 0;
1324}
1325
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001326int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
1327 unsigned long end, int advice, unsigned long *vm_flags)
1328{
1329 struct mm_struct *mm = vma->vm_mm;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001330 int err;
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001331
1332 switch (advice) {
1333 case MADV_MERGEABLE:
1334 /*
1335 * Be somewhat over-protective for now!
1336 */
1337 if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE |
1338 VM_PFNMAP | VM_IO | VM_DONTEXPAND |
1339 VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE |
1340 VM_MIXEDMAP | VM_SAO))
1341 return 0; /* just ignore the advice */
1342
Hugh Dickinsd952b792009-09-21 17:02:16 -07001343 if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
1344 err = __ksm_enter(mm);
1345 if (err)
1346 return err;
1347 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001348
1349 *vm_flags |= VM_MERGEABLE;
1350 break;
1351
1352 case MADV_UNMERGEABLE:
1353 if (!(*vm_flags & VM_MERGEABLE))
1354 return 0; /* just ignore the advice */
1355
Hugh Dickinsd952b792009-09-21 17:02:16 -07001356 if (vma->anon_vma) {
1357 err = unmerge_ksm_pages(vma, start, end);
1358 if (err)
1359 return err;
1360 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001361
1362 *vm_flags &= ~VM_MERGEABLE;
1363 break;
1364 }
1365
1366 return 0;
1367}
1368
1369int __ksm_enter(struct mm_struct *mm)
1370{
Hugh Dickins6e158382009-09-21 17:02:14 -07001371 struct mm_slot *mm_slot;
1372 int needs_wakeup;
1373
1374 mm_slot = alloc_mm_slot();
Izik Eidus31dbd012009-09-21 17:02:03 -07001375 if (!mm_slot)
1376 return -ENOMEM;
1377
Hugh Dickins6e158382009-09-21 17:02:14 -07001378 /* Check ksm_run too? Would need tighter locking */
1379 needs_wakeup = list_empty(&ksm_mm_head.mm_list);
1380
Izik Eidus31dbd012009-09-21 17:02:03 -07001381 spin_lock(&ksm_mmlist_lock);
1382 insert_to_mm_slots_hash(mm, mm_slot);
1383 /*
1384 * Insert just behind the scanning cursor, to let the area settle
1385 * down a little; when fork is followed by immediate exec, we don't
1386 * want ksmd to waste time setting up and tearing down an rmap_list.
1387 */
1388 list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list);
1389 spin_unlock(&ksm_mmlist_lock);
1390
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001391 set_bit(MMF_VM_MERGEABLE, &mm->flags);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001392 atomic_inc(&mm->mm_count);
Hugh Dickins6e158382009-09-21 17:02:14 -07001393
1394 if (needs_wakeup)
1395 wake_up_interruptible(&ksm_thread_wait);
1396
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001397 return 0;
1398}
1399
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07001400void __ksm_exit(struct mm_struct *mm)
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001401{
Hugh Dickinscd551f92009-09-21 17:02:17 -07001402 struct mm_slot *mm_slot;
Hugh Dickins9ba69292009-09-21 17:02:20 -07001403 int easy_to_free = 0;
Hugh Dickinscd551f92009-09-21 17:02:17 -07001404
Izik Eidus31dbd012009-09-21 17:02:03 -07001405 /*
Hugh Dickins9ba69292009-09-21 17:02:20 -07001406 * This process is exiting: if it's straightforward (as is the
1407 * case when ksmd was never running), free mm_slot immediately.
1408 * But if it's at the cursor or has rmap_items linked to it, use
1409 * mmap_sem to synchronize with any break_cows before pagetables
1410 * are freed, and leave the mm_slot on the list for ksmd to free.
1411 * Beware: ksm may already have noticed it exiting and freed the slot.
Izik Eidus31dbd012009-09-21 17:02:03 -07001412 */
Hugh Dickins9ba69292009-09-21 17:02:20 -07001413
Hugh Dickinscd551f92009-09-21 17:02:17 -07001414 spin_lock(&ksm_mmlist_lock);
1415 mm_slot = get_mm_slot(mm);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001416 if (mm_slot && ksm_scan.mm_slot != mm_slot) {
Hugh Dickins6514d512009-12-14 17:59:19 -08001417 if (!mm_slot->rmap_list) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001418 hlist_del(&mm_slot->link);
1419 list_del(&mm_slot->mm_list);
1420 easy_to_free = 1;
1421 } else {
1422 list_move(&mm_slot->mm_list,
1423 &ksm_scan.mm_slot->mm_list);
1424 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001425 }
Hugh Dickinscd551f92009-09-21 17:02:17 -07001426 spin_unlock(&ksm_mmlist_lock);
1427
Hugh Dickins9ba69292009-09-21 17:02:20 -07001428 if (easy_to_free) {
1429 free_mm_slot(mm_slot);
1430 clear_bit(MMF_VM_MERGEABLE, &mm->flags);
1431 mmdrop(mm);
1432 } else if (mm_slot) {
Hugh Dickins9ba69292009-09-21 17:02:20 -07001433 down_write(&mm->mmap_sem);
1434 up_write(&mm->mmap_sem);
Hugh Dickins9ba69292009-09-21 17:02:20 -07001435 }
Hugh Dickinsf8af4da2009-09-21 17:01:57 -07001436}
Izik Eidus31dbd012009-09-21 17:02:03 -07001437
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001438#ifdef CONFIG_SYSFS
1439/*
1440 * This all compiles without CONFIG_SYSFS, but is a waste of space.
1441 */
1442
Izik Eidus31dbd012009-09-21 17:02:03 -07001443#define KSM_ATTR_RO(_name) \
1444 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1445#define KSM_ATTR(_name) \
1446 static struct kobj_attribute _name##_attr = \
1447 __ATTR(_name, 0644, _name##_show, _name##_store)
1448
1449static ssize_t sleep_millisecs_show(struct kobject *kobj,
1450 struct kobj_attribute *attr, char *buf)
1451{
1452 return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs);
1453}
1454
1455static ssize_t sleep_millisecs_store(struct kobject *kobj,
1456 struct kobj_attribute *attr,
1457 const char *buf, size_t count)
1458{
1459 unsigned long msecs;
1460 int err;
1461
1462 err = strict_strtoul(buf, 10, &msecs);
1463 if (err || msecs > UINT_MAX)
1464 return -EINVAL;
1465
1466 ksm_thread_sleep_millisecs = msecs;
1467
1468 return count;
1469}
1470KSM_ATTR(sleep_millisecs);
1471
1472static ssize_t pages_to_scan_show(struct kobject *kobj,
1473 struct kobj_attribute *attr, char *buf)
1474{
1475 return sprintf(buf, "%u\n", ksm_thread_pages_to_scan);
1476}
1477
1478static ssize_t pages_to_scan_store(struct kobject *kobj,
1479 struct kobj_attribute *attr,
1480 const char *buf, size_t count)
1481{
1482 int err;
1483 unsigned long nr_pages;
1484
1485 err = strict_strtoul(buf, 10, &nr_pages);
1486 if (err || nr_pages > UINT_MAX)
1487 return -EINVAL;
1488
1489 ksm_thread_pages_to_scan = nr_pages;
1490
1491 return count;
1492}
1493KSM_ATTR(pages_to_scan);
1494
1495static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr,
1496 char *buf)
1497{
1498 return sprintf(buf, "%u\n", ksm_run);
1499}
1500
1501static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
1502 const char *buf, size_t count)
1503{
1504 int err;
1505 unsigned long flags;
1506
1507 err = strict_strtoul(buf, 10, &flags);
1508 if (err || flags > UINT_MAX)
1509 return -EINVAL;
1510 if (flags > KSM_RUN_UNMERGE)
1511 return -EINVAL;
1512
1513 /*
1514 * KSM_RUN_MERGE sets ksmd running, and 0 stops it running.
1515 * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items,
Hugh Dickinsb4028262009-09-21 17:02:09 -07001516 * breaking COW to free the unswappable pages_shared (but leaves
Izik Eidus31dbd012009-09-21 17:02:03 -07001517 * mm_slots on the list for when ksmd may be set running again).
1518 */
1519
1520 mutex_lock(&ksm_thread_mutex);
1521 if (ksm_run != flags) {
1522 ksm_run = flags;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001523 if (flags & KSM_RUN_UNMERGE) {
Hugh Dickins35451be2009-09-21 17:02:27 -07001524 current->flags |= PF_OOM_ORIGIN;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001525 err = unmerge_and_remove_all_rmap_items();
Hugh Dickins35451be2009-09-21 17:02:27 -07001526 current->flags &= ~PF_OOM_ORIGIN;
Hugh Dickinsd952b792009-09-21 17:02:16 -07001527 if (err) {
1528 ksm_run = KSM_RUN_STOP;
1529 count = err;
1530 }
1531 }
Izik Eidus31dbd012009-09-21 17:02:03 -07001532 }
1533 mutex_unlock(&ksm_thread_mutex);
1534
1535 if (flags & KSM_RUN_MERGE)
1536 wake_up_interruptible(&ksm_thread_wait);
1537
1538 return count;
1539}
1540KSM_ATTR(run);
1541
Izik Eidus31dbd012009-09-21 17:02:03 -07001542static ssize_t max_kernel_pages_store(struct kobject *kobj,
1543 struct kobj_attribute *attr,
1544 const char *buf, size_t count)
1545{
1546 int err;
1547 unsigned long nr_pages;
1548
1549 err = strict_strtoul(buf, 10, &nr_pages);
1550 if (err)
1551 return -EINVAL;
1552
1553 ksm_max_kernel_pages = nr_pages;
1554
1555 return count;
1556}
1557
1558static ssize_t max_kernel_pages_show(struct kobject *kobj,
1559 struct kobj_attribute *attr, char *buf)
1560{
1561 return sprintf(buf, "%lu\n", ksm_max_kernel_pages);
1562}
1563KSM_ATTR(max_kernel_pages);
1564
Hugh Dickinsb4028262009-09-21 17:02:09 -07001565static ssize_t pages_shared_show(struct kobject *kobj,
1566 struct kobj_attribute *attr, char *buf)
1567{
1568 return sprintf(buf, "%lu\n", ksm_pages_shared);
1569}
1570KSM_ATTR_RO(pages_shared);
1571
1572static ssize_t pages_sharing_show(struct kobject *kobj,
1573 struct kobj_attribute *attr, char *buf)
1574{
Hugh Dickinse178dfd2009-09-21 17:02:10 -07001575 return sprintf(buf, "%lu\n", ksm_pages_sharing);
Hugh Dickinsb4028262009-09-21 17:02:09 -07001576}
1577KSM_ATTR_RO(pages_sharing);
1578
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001579static ssize_t pages_unshared_show(struct kobject *kobj,
1580 struct kobj_attribute *attr, char *buf)
1581{
1582 return sprintf(buf, "%lu\n", ksm_pages_unshared);
1583}
1584KSM_ATTR_RO(pages_unshared);
1585
1586static ssize_t pages_volatile_show(struct kobject *kobj,
1587 struct kobj_attribute *attr, char *buf)
1588{
1589 long ksm_pages_volatile;
1590
1591 ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared
1592 - ksm_pages_sharing - ksm_pages_unshared;
1593 /*
1594 * It was not worth any locking to calculate that statistic,
1595 * but it might therefore sometimes be negative: conceal that.
1596 */
1597 if (ksm_pages_volatile < 0)
1598 ksm_pages_volatile = 0;
1599 return sprintf(buf, "%ld\n", ksm_pages_volatile);
1600}
1601KSM_ATTR_RO(pages_volatile);
1602
1603static ssize_t full_scans_show(struct kobject *kobj,
1604 struct kobj_attribute *attr, char *buf)
1605{
1606 return sprintf(buf, "%lu\n", ksm_scan.seqnr);
1607}
1608KSM_ATTR_RO(full_scans);
1609
Izik Eidus31dbd012009-09-21 17:02:03 -07001610static struct attribute *ksm_attrs[] = {
1611 &sleep_millisecs_attr.attr,
1612 &pages_to_scan_attr.attr,
1613 &run_attr.attr,
Izik Eidus31dbd012009-09-21 17:02:03 -07001614 &max_kernel_pages_attr.attr,
Hugh Dickinsb4028262009-09-21 17:02:09 -07001615 &pages_shared_attr.attr,
1616 &pages_sharing_attr.attr,
Hugh Dickins473b0ce2009-09-21 17:02:11 -07001617 &pages_unshared_attr.attr,
1618 &pages_volatile_attr.attr,
1619 &full_scans_attr.attr,
Izik Eidus31dbd012009-09-21 17:02:03 -07001620 NULL,
1621};
1622
1623static struct attribute_group ksm_attr_group = {
1624 .attrs = ksm_attrs,
1625 .name = "ksm",
1626};
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001627#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07001628
1629static int __init ksm_init(void)
1630{
1631 struct task_struct *ksm_thread;
1632 int err;
1633
Hugh Dickinsc73602a2009-10-07 16:32:22 -07001634 ksm_max_kernel_pages = totalram_pages / 4;
Izik Eidus2c6854f2009-09-23 15:56:04 -07001635
Izik Eidus31dbd012009-09-21 17:02:03 -07001636 err = ksm_slab_init();
1637 if (err)
1638 goto out;
1639
1640 err = mm_slots_hash_init();
1641 if (err)
1642 goto out_free1;
1643
1644 ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd");
1645 if (IS_ERR(ksm_thread)) {
1646 printk(KERN_ERR "ksm: creating kthread failed\n");
1647 err = PTR_ERR(ksm_thread);
1648 goto out_free2;
1649 }
1650
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001651#ifdef CONFIG_SYSFS
Izik Eidus31dbd012009-09-21 17:02:03 -07001652 err = sysfs_create_group(mm_kobj, &ksm_attr_group);
1653 if (err) {
1654 printk(KERN_ERR "ksm: register sysfs failed\n");
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001655 kthread_stop(ksm_thread);
1656 goto out_free2;
Izik Eidus31dbd012009-09-21 17:02:03 -07001657 }
Hugh Dickinsc73602a2009-10-07 16:32:22 -07001658#else
1659 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
1660
Hugh Dickins2ffd8672009-09-21 17:02:23 -07001661#endif /* CONFIG_SYSFS */
Izik Eidus31dbd012009-09-21 17:02:03 -07001662
1663 return 0;
1664
Izik Eidus31dbd012009-09-21 17:02:03 -07001665out_free2:
1666 mm_slots_hash_free();
1667out_free1:
1668 ksm_slab_free();
1669out:
1670 return err;
1671}
1672module_init(ksm_init)