]>
Commit | Line | Data |
---|---|---|
f8af4da3 | 1 | /* |
31dbd01f IE |
2 | * 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 | * | |
36b2528d | 7 | * Copyright (C) 2008-2009 Red Hat, Inc. |
31dbd01f IE |
8 | * Authors: |
9 | * Izik Eidus | |
10 | * Andrea Arcangeli | |
11 | * Chris Wright | |
36b2528d | 12 | * Hugh Dickins |
31dbd01f IE |
13 | * |
14 | * This work is licensed under the terms of the GNU GPL, version 2. | |
f8af4da3 HD |
15 | */ |
16 | ||
17 | #include <linux/errno.h> | |
31dbd01f IE |
18 | #include <linux/mm.h> |
19 | #include <linux/fs.h> | |
f8af4da3 | 20 | #include <linux/mman.h> |
31dbd01f IE |
21 | #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> | |
62b61f61 | 32 | #include <linux/memory.h> |
31dbd01f | 33 | #include <linux/mmu_notifier.h> |
2c6854fd | 34 | #include <linux/swap.h> |
f8af4da3 | 35 | #include <linux/ksm.h> |
d9f8984c | 36 | #include <linux/hash.h> |
f8af4da3 | 37 | |
31dbd01f | 38 | #include <asm/tlbflush.h> |
73848b46 | 39 | #include "internal.h" |
31dbd01f IE |
40 | |
41 | /* | |
42 | * A few notes about the KSM scanning process, | |
43 | * to make it easier to understand the data structures below: | |
44 | * | |
45 | * In order to reduce excessive scanning, KSM sorts the memory pages by their | |
46 | * contents into a data structure that holds pointers to the pages' locations. | |
47 | * | |
48 | * Since the contents of the pages may change at any moment, KSM cannot just | |
49 | * insert the pages into a normal sorted tree and expect it to find anything. | |
50 | * Therefore KSM uses two data structures - the stable and the unstable tree. | |
51 | * | |
52 | * The stable tree holds pointers to all the merged pages (ksm pages), sorted | |
53 | * by their contents. Because each such page is write-protected, searching on | |
54 | * this tree is fully assured to be working (except when pages are unmapped), | |
55 | * and therefore this tree is called the stable tree. | |
56 | * | |
57 | * In addition to the stable tree, KSM uses a second data structure called the | |
58 | * unstable tree: this tree holds pointers to pages which have been found to | |
59 | * be "unchanged for a period of time". The unstable tree sorts these pages | |
60 | * by their contents, but since they are not write-protected, KSM cannot rely | |
61 | * upon the unstable tree to work correctly - the unstable tree is liable to | |
62 | * be corrupted as its contents are modified, and so it is called unstable. | |
63 | * | |
64 | * KSM solves this problem by several techniques: | |
65 | * | |
66 | * 1) The unstable tree is flushed every time KSM completes scanning all | |
67 | * memory areas, and then the tree is rebuilt again from the beginning. | |
68 | * 2) KSM will only insert into the unstable tree, pages whose hash value | |
69 | * has not changed since the previous scan of all memory areas. | |
70 | * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the | |
71 | * colors of the nodes and not on their contents, assuring that even when | |
72 | * the tree gets "corrupted" it won't get out of balance, so scanning time | |
73 | * remains the same (also, searching and inserting nodes in an rbtree uses | |
74 | * the same algorithm, so we have no overhead when we flush and rebuild). | |
75 | * 4) KSM never flushes the stable tree, which means that even if it were to | |
76 | * take 10 attempts to find a page in the unstable tree, once it is found, | |
77 | * it is secured in the stable tree. (When we scan a new page, we first | |
78 | * compare it against the stable tree, and then against the unstable tree.) | |
79 | */ | |
80 | ||
81 | /** | |
82 | * struct mm_slot - ksm information per mm that is being scanned | |
83 | * @link: link to the mm_slots hash list | |
84 | * @mm_list: link into the mm_slots list, rooted in ksm_mm_head | |
6514d511 | 85 | * @rmap_list: head for this mm_slot's singly-linked list of rmap_items |
31dbd01f IE |
86 | * @mm: the mm that this information is valid for |
87 | */ | |
88 | struct mm_slot { | |
89 | struct hlist_node link; | |
90 | struct list_head mm_list; | |
6514d511 | 91 | struct rmap_item *rmap_list; |
31dbd01f IE |
92 | struct mm_struct *mm; |
93 | }; | |
94 | ||
95 | /** | |
96 | * struct ksm_scan - cursor for scanning | |
97 | * @mm_slot: the current mm_slot we are scanning | |
98 | * @address: the next address inside that to be scanned | |
6514d511 | 99 | * @rmap_list: link to the next rmap to be scanned in the rmap_list |
31dbd01f IE |
100 | * @seqnr: count of completed full scans (needed when removing unstable node) |
101 | * | |
102 | * There is only the one ksm_scan instance of this cursor structure. | |
103 | */ | |
104 | struct ksm_scan { | |
105 | struct mm_slot *mm_slot; | |
106 | unsigned long address; | |
6514d511 | 107 | struct rmap_item **rmap_list; |
31dbd01f IE |
108 | unsigned long seqnr; |
109 | }; | |
110 | ||
7b6ba2c7 HD |
111 | /** |
112 | * struct stable_node - node of the stable rbtree | |
113 | * @node: rb node of this ksm page in the stable tree | |
114 | * @hlist: hlist head of rmap_items using this ksm page | |
62b61f61 | 115 | * @kpfn: page frame number of this ksm page |
7b6ba2c7 HD |
116 | */ |
117 | struct stable_node { | |
118 | struct rb_node node; | |
119 | struct hlist_head hlist; | |
62b61f61 | 120 | unsigned long kpfn; |
7b6ba2c7 HD |
121 | }; |
122 | ||
31dbd01f IE |
123 | /** |
124 | * struct rmap_item - reverse mapping item for virtual addresses | |
6514d511 | 125 | * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list |
db114b83 | 126 | * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree |
31dbd01f IE |
127 | * @mm: the memory structure this rmap_item is pointing into |
128 | * @address: the virtual address this rmap_item tracks (+ flags in low bits) | |
129 | * @oldchecksum: previous checksum of the page at that virtual address | |
7b6ba2c7 HD |
130 | * @node: rb node of this rmap_item in the unstable tree |
131 | * @head: pointer to stable_node heading this list in the stable tree | |
132 | * @hlist: link into hlist of rmap_items hanging off that stable_node | |
31dbd01f IE |
133 | */ |
134 | struct rmap_item { | |
6514d511 | 135 | struct rmap_item *rmap_list; |
db114b83 | 136 | struct anon_vma *anon_vma; /* when stable */ |
31dbd01f IE |
137 | struct mm_struct *mm; |
138 | unsigned long address; /* + low bits used for flags below */ | |
7b6ba2c7 | 139 | unsigned int oldchecksum; /* when unstable */ |
31dbd01f | 140 | union { |
7b6ba2c7 HD |
141 | struct rb_node node; /* when node of unstable tree */ |
142 | struct { /* when listed from stable tree */ | |
143 | struct stable_node *head; | |
144 | struct hlist_node hlist; | |
145 | }; | |
31dbd01f IE |
146 | }; |
147 | }; | |
148 | ||
149 | #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ | |
7b6ba2c7 HD |
150 | #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ |
151 | #define STABLE_FLAG 0x200 /* is listed from the stable tree */ | |
31dbd01f IE |
152 | |
153 | /* The stable and unstable tree heads */ | |
154 | static struct rb_root root_stable_tree = RB_ROOT; | |
155 | static struct rb_root root_unstable_tree = RB_ROOT; | |
156 | ||
d9f8984c LJ |
157 | #define MM_SLOTS_HASH_SHIFT 10 |
158 | #define MM_SLOTS_HASH_HEADS (1 << MM_SLOTS_HASH_SHIFT) | |
159 | static struct hlist_head mm_slots_hash[MM_SLOTS_HASH_HEADS]; | |
31dbd01f IE |
160 | |
161 | static struct mm_slot ksm_mm_head = { | |
162 | .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list), | |
163 | }; | |
164 | static struct ksm_scan ksm_scan = { | |
165 | .mm_slot = &ksm_mm_head, | |
166 | }; | |
167 | ||
168 | static struct kmem_cache *rmap_item_cache; | |
7b6ba2c7 | 169 | static struct kmem_cache *stable_node_cache; |
31dbd01f IE |
170 | static struct kmem_cache *mm_slot_cache; |
171 | ||
172 | /* The number of nodes in the stable tree */ | |
b4028260 | 173 | static unsigned long ksm_pages_shared; |
31dbd01f | 174 | |
e178dfde | 175 | /* The number of page slots additionally sharing those nodes */ |
b4028260 | 176 | static unsigned long ksm_pages_sharing; |
31dbd01f | 177 | |
473b0ce4 HD |
178 | /* The number of nodes in the unstable tree */ |
179 | static unsigned long ksm_pages_unshared; | |
180 | ||
181 | /* The number of rmap_items in use: to calculate pages_volatile */ | |
182 | static unsigned long ksm_rmap_items; | |
183 | ||
31dbd01f | 184 | /* Number of pages ksmd should scan in one batch */ |
2c6854fd | 185 | static unsigned int ksm_thread_pages_to_scan = 100; |
31dbd01f IE |
186 | |
187 | /* Milliseconds ksmd should sleep between batches */ | |
2ffd8679 | 188 | static unsigned int ksm_thread_sleep_millisecs = 20; |
31dbd01f IE |
189 | |
190 | #define KSM_RUN_STOP 0 | |
191 | #define KSM_RUN_MERGE 1 | |
192 | #define KSM_RUN_UNMERGE 2 | |
2c6854fd | 193 | static unsigned int ksm_run = KSM_RUN_STOP; |
31dbd01f IE |
194 | |
195 | static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); | |
196 | static DEFINE_MUTEX(ksm_thread_mutex); | |
197 | static 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 | ||
203 | static 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 | ||
7b6ba2c7 HD |
209 | stable_node_cache = KSM_KMEM_CACHE(stable_node, 0); |
210 | if (!stable_node_cache) | |
211 | goto out_free1; | |
212 | ||
31dbd01f IE |
213 | mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0); |
214 | if (!mm_slot_cache) | |
7b6ba2c7 | 215 | goto out_free2; |
31dbd01f IE |
216 | |
217 | return 0; | |
218 | ||
7b6ba2c7 HD |
219 | out_free2: |
220 | kmem_cache_destroy(stable_node_cache); | |
221 | out_free1: | |
31dbd01f IE |
222 | kmem_cache_destroy(rmap_item_cache); |
223 | out: | |
224 | return -ENOMEM; | |
225 | } | |
226 | ||
227 | static void __init ksm_slab_free(void) | |
228 | { | |
229 | kmem_cache_destroy(mm_slot_cache); | |
7b6ba2c7 | 230 | kmem_cache_destroy(stable_node_cache); |
31dbd01f IE |
231 | kmem_cache_destroy(rmap_item_cache); |
232 | mm_slot_cache = NULL; | |
233 | } | |
234 | ||
235 | static inline struct rmap_item *alloc_rmap_item(void) | |
236 | { | |
473b0ce4 HD |
237 | 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; | |
31dbd01f IE |
243 | } |
244 | ||
245 | static inline void free_rmap_item(struct rmap_item *rmap_item) | |
246 | { | |
473b0ce4 | 247 | ksm_rmap_items--; |
31dbd01f IE |
248 | rmap_item->mm = NULL; /* debug safety */ |
249 | kmem_cache_free(rmap_item_cache, rmap_item); | |
250 | } | |
251 | ||
7b6ba2c7 HD |
252 | static inline struct stable_node *alloc_stable_node(void) |
253 | { | |
254 | return kmem_cache_alloc(stable_node_cache, GFP_KERNEL); | |
255 | } | |
256 | ||
257 | static inline void free_stable_node(struct stable_node *stable_node) | |
258 | { | |
259 | kmem_cache_free(stable_node_cache, stable_node); | |
260 | } | |
261 | ||
31dbd01f IE |
262 | static 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 | ||
269 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
270 | { | |
271 | kmem_cache_free(mm_slot_cache, mm_slot); | |
272 | } | |
273 | ||
31dbd01f IE |
274 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) |
275 | { | |
276 | struct mm_slot *mm_slot; | |
277 | struct hlist_head *bucket; | |
278 | struct hlist_node *node; | |
279 | ||
d9f8984c | 280 | bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)]; |
31dbd01f IE |
281 | hlist_for_each_entry(mm_slot, node, bucket, link) { |
282 | if (mm == mm_slot->mm) | |
283 | return mm_slot; | |
284 | } | |
285 | return NULL; | |
286 | } | |
287 | ||
288 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
289 | struct mm_slot *mm_slot) | |
290 | { | |
291 | struct hlist_head *bucket; | |
292 | ||
d9f8984c | 293 | bucket = &mm_slots_hash[hash_ptr(mm, MM_SLOTS_HASH_SHIFT)]; |
31dbd01f | 294 | mm_slot->mm = mm; |
31dbd01f IE |
295 | hlist_add_head(&mm_slot->link, bucket); |
296 | } | |
297 | ||
298 | static inline int in_stable_tree(struct rmap_item *rmap_item) | |
299 | { | |
300 | return rmap_item->address & STABLE_FLAG; | |
301 | } | |
302 | ||
db114b83 HD |
303 | static void hold_anon_vma(struct rmap_item *rmap_item, |
304 | struct anon_vma *anon_vma) | |
305 | { | |
306 | rmap_item->anon_vma = anon_vma; | |
76545066 | 307 | get_anon_vma(anon_vma); |
db114b83 HD |
308 | } |
309 | ||
76545066 | 310 | static void ksm_drop_anon_vma(struct rmap_item *rmap_item) |
db114b83 HD |
311 | { |
312 | struct anon_vma *anon_vma = rmap_item->anon_vma; | |
313 | ||
76545066 | 314 | drop_anon_vma(anon_vma); |
db114b83 HD |
315 | } |
316 | ||
a913e182 HD |
317 | /* |
318 | * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's | |
319 | * page tables after it has passed through ksm_exit() - which, if necessary, | |
320 | * takes mmap_sem briefly to serialize against them. ksm_exit() does not set | |
321 | * a special flag: they can just back out as soon as mm_users goes to zero. | |
322 | * ksm_test_exit() is used throughout to make this test for exit: in some | |
323 | * places for correctness, in some places just to avoid unnecessary work. | |
324 | */ | |
325 | static inline bool ksm_test_exit(struct mm_struct *mm) | |
326 | { | |
327 | return atomic_read(&mm->mm_users) == 0; | |
328 | } | |
329 | ||
31dbd01f IE |
330 | /* |
331 | * We use break_ksm to break COW on a ksm page: it's a stripped down | |
332 | * | |
333 | * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1) | |
334 | * put_page(page); | |
335 | * | |
336 | * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma, | |
337 | * in case the application has unmapped and remapped mm,addr meanwhile. | |
338 | * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP | |
339 | * mmap of /dev/mem or /dev/kmem, where we would not want to touch it. | |
340 | */ | |
d952b791 | 341 | static int break_ksm(struct vm_area_struct *vma, unsigned long addr) |
31dbd01f IE |
342 | { |
343 | struct page *page; | |
d952b791 | 344 | int ret = 0; |
31dbd01f IE |
345 | |
346 | do { | |
347 | cond_resched(); | |
348 | page = follow_page(vma, addr, FOLL_GET); | |
22eccdd7 | 349 | if (IS_ERR_OR_NULL(page)) |
31dbd01f IE |
350 | break; |
351 | if (PageKsm(page)) | |
352 | ret = handle_mm_fault(vma->vm_mm, vma, addr, | |
353 | FAULT_FLAG_WRITE); | |
354 | else | |
355 | ret = VM_FAULT_WRITE; | |
356 | put_page(page); | |
d952b791 HD |
357 | } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM))); |
358 | /* | |
359 | * We must loop because handle_mm_fault() may back out if there's | |
360 | * any difficulty e.g. if pte accessed bit gets updated concurrently. | |
361 | * | |
362 | * VM_FAULT_WRITE is what we have been hoping for: it indicates that | |
363 | * COW has been broken, even if the vma does not permit VM_WRITE; | |
364 | * but note that a concurrent fault might break PageKsm for us. | |
365 | * | |
366 | * VM_FAULT_SIGBUS could occur if we race with truncation of the | |
367 | * backing file, which also invalidates anonymous pages: that's | |
368 | * okay, that truncation will have unmapped the PageKsm for us. | |
369 | * | |
370 | * VM_FAULT_OOM: at the time of writing (late July 2009), setting | |
371 | * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the | |
372 | * current task has TIF_MEMDIE set, and will be OOM killed on return | |
373 | * to user; and ksmd, having no mm, would never be chosen for that. | |
374 | * | |
375 | * But if the mm is in a limited mem_cgroup, then the fault may fail | |
376 | * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and | |
377 | * even ksmd can fail in this way - though it's usually breaking ksm | |
378 | * just to undo a merge it made a moment before, so unlikely to oom. | |
379 | * | |
380 | * That's a pity: we might therefore have more kernel pages allocated | |
381 | * than we're counting as nodes in the stable tree; but ksm_do_scan | |
382 | * will retry to break_cow on each pass, so should recover the page | |
383 | * in due course. The important thing is to not let VM_MERGEABLE | |
384 | * be cleared while any such pages might remain in the area. | |
385 | */ | |
386 | return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; | |
31dbd01f IE |
387 | } |
388 | ||
8dd3557a | 389 | static void break_cow(struct rmap_item *rmap_item) |
31dbd01f | 390 | { |
8dd3557a HD |
391 | struct mm_struct *mm = rmap_item->mm; |
392 | unsigned long addr = rmap_item->address; | |
31dbd01f IE |
393 | struct vm_area_struct *vma; |
394 | ||
4035c07a HD |
395 | /* |
396 | * It is not an accident that whenever we want to break COW | |
397 | * to undo, we also need to drop a reference to the anon_vma. | |
398 | */ | |
76545066 | 399 | ksm_drop_anon_vma(rmap_item); |
4035c07a | 400 | |
81464e30 | 401 | down_read(&mm->mmap_sem); |
9ba69294 HD |
402 | if (ksm_test_exit(mm)) |
403 | goto out; | |
31dbd01f IE |
404 | vma = find_vma(mm, addr); |
405 | if (!vma || vma->vm_start > addr) | |
81464e30 | 406 | goto out; |
31dbd01f | 407 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
81464e30 | 408 | goto out; |
31dbd01f | 409 | break_ksm(vma, addr); |
81464e30 | 410 | out: |
31dbd01f IE |
411 | up_read(&mm->mmap_sem); |
412 | } | |
413 | ||
414 | static struct page *get_mergeable_page(struct rmap_item *rmap_item) | |
415 | { | |
416 | struct mm_struct *mm = rmap_item->mm; | |
417 | unsigned long addr = rmap_item->address; | |
418 | struct vm_area_struct *vma; | |
419 | struct page *page; | |
420 | ||
421 | down_read(&mm->mmap_sem); | |
9ba69294 HD |
422 | if (ksm_test_exit(mm)) |
423 | goto out; | |
31dbd01f IE |
424 | vma = find_vma(mm, addr); |
425 | if (!vma || vma->vm_start > addr) | |
426 | goto out; | |
427 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) | |
428 | goto out; | |
429 | ||
430 | page = follow_page(vma, addr, FOLL_GET); | |
22eccdd7 | 431 | if (IS_ERR_OR_NULL(page)) |
31dbd01f IE |
432 | goto out; |
433 | if (PageAnon(page)) { | |
434 | flush_anon_page(vma, page, addr); | |
435 | flush_dcache_page(page); | |
436 | } else { | |
437 | put_page(page); | |
438 | out: page = NULL; | |
439 | } | |
440 | up_read(&mm->mmap_sem); | |
441 | return page; | |
442 | } | |
443 | ||
4035c07a HD |
444 | static void remove_node_from_stable_tree(struct stable_node *stable_node) |
445 | { | |
446 | struct rmap_item *rmap_item; | |
447 | struct hlist_node *hlist; | |
448 | ||
449 | hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) { | |
450 | if (rmap_item->hlist.next) | |
451 | ksm_pages_sharing--; | |
452 | else | |
453 | ksm_pages_shared--; | |
76545066 | 454 | ksm_drop_anon_vma(rmap_item); |
4035c07a HD |
455 | rmap_item->address &= PAGE_MASK; |
456 | cond_resched(); | |
457 | } | |
458 | ||
459 | rb_erase(&stable_node->node, &root_stable_tree); | |
460 | free_stable_node(stable_node); | |
461 | } | |
462 | ||
463 | /* | |
464 | * get_ksm_page: checks if the page indicated by the stable node | |
465 | * is still its ksm page, despite having held no reference to it. | |
466 | * In which case we can trust the content of the page, and it | |
467 | * returns the gotten page; but if the page has now been zapped, | |
468 | * remove the stale node from the stable tree and return NULL. | |
469 | * | |
470 | * You would expect the stable_node to hold a reference to the ksm page. | |
471 | * But if it increments the page's count, swapping out has to wait for | |
472 | * ksmd to come around again before it can free the page, which may take | |
473 | * seconds or even minutes: much too unresponsive. So instead we use a | |
474 | * "keyhole reference": access to the ksm page from the stable node peeps | |
475 | * out through its keyhole to see if that page still holds the right key, | |
476 | * pointing back to this stable node. This relies on freeing a PageAnon | |
477 | * page to reset its page->mapping to NULL, and relies on no other use of | |
478 | * a page to put something that might look like our key in page->mapping. | |
479 | * | |
480 | * include/linux/pagemap.h page_cache_get_speculative() is a good reference, | |
481 | * but this is different - made simpler by ksm_thread_mutex being held, but | |
482 | * interesting for assuming that no other use of the struct page could ever | |
483 | * put our expected_mapping into page->mapping (or a field of the union which | |
484 | * coincides with page->mapping). The RCU calls are not for KSM at all, but | |
485 | * to keep the page_count protocol described with page_cache_get_speculative. | |
486 | * | |
487 | * Note: it is possible that get_ksm_page() will return NULL one moment, | |
488 | * then page the next, if the page is in between page_freeze_refs() and | |
489 | * page_unfreeze_refs(): this shouldn't be a problem anywhere, the page | |
490 | * is on its way to being freed; but it is an anomaly to bear in mind. | |
491 | */ | |
492 | static struct page *get_ksm_page(struct stable_node *stable_node) | |
493 | { | |
494 | struct page *page; | |
495 | void *expected_mapping; | |
496 | ||
62b61f61 | 497 | page = pfn_to_page(stable_node->kpfn); |
4035c07a HD |
498 | expected_mapping = (void *)stable_node + |
499 | (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM); | |
500 | rcu_read_lock(); | |
501 | if (page->mapping != expected_mapping) | |
502 | goto stale; | |
503 | if (!get_page_unless_zero(page)) | |
504 | goto stale; | |
505 | if (page->mapping != expected_mapping) { | |
506 | put_page(page); | |
507 | goto stale; | |
508 | } | |
509 | rcu_read_unlock(); | |
510 | return page; | |
511 | stale: | |
512 | rcu_read_unlock(); | |
513 | remove_node_from_stable_tree(stable_node); | |
514 | return NULL; | |
515 | } | |
516 | ||
31dbd01f IE |
517 | /* |
518 | * Removing rmap_item from stable or unstable tree. | |
519 | * This function will clean the information from the stable/unstable tree. | |
520 | */ | |
521 | static void remove_rmap_item_from_tree(struct rmap_item *rmap_item) | |
522 | { | |
7b6ba2c7 HD |
523 | if (rmap_item->address & STABLE_FLAG) { |
524 | struct stable_node *stable_node; | |
5ad64688 | 525 | struct page *page; |
31dbd01f | 526 | |
7b6ba2c7 | 527 | stable_node = rmap_item->head; |
4035c07a HD |
528 | page = get_ksm_page(stable_node); |
529 | if (!page) | |
530 | goto out; | |
5ad64688 | 531 | |
4035c07a | 532 | lock_page(page); |
7b6ba2c7 | 533 | hlist_del(&rmap_item->hlist); |
4035c07a HD |
534 | unlock_page(page); |
535 | put_page(page); | |
08beca44 | 536 | |
4035c07a HD |
537 | if (stable_node->hlist.first) |
538 | ksm_pages_sharing--; | |
539 | else | |
7b6ba2c7 | 540 | ksm_pages_shared--; |
31dbd01f | 541 | |
76545066 | 542 | ksm_drop_anon_vma(rmap_item); |
93d17715 | 543 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 544 | |
7b6ba2c7 | 545 | } else if (rmap_item->address & UNSTABLE_FLAG) { |
31dbd01f IE |
546 | unsigned char age; |
547 | /* | |
9ba69294 | 548 | * Usually ksmd can and must skip the rb_erase, because |
31dbd01f | 549 | * root_unstable_tree was already reset to RB_ROOT. |
9ba69294 HD |
550 | * But be careful when an mm is exiting: do the rb_erase |
551 | * if this rmap_item was inserted by this scan, rather | |
552 | * than left over from before. | |
31dbd01f IE |
553 | */ |
554 | age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); | |
cd551f97 | 555 | BUG_ON(age > 1); |
31dbd01f IE |
556 | if (!age) |
557 | rb_erase(&rmap_item->node, &root_unstable_tree); | |
93d17715 | 558 | |
473b0ce4 | 559 | ksm_pages_unshared--; |
93d17715 | 560 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 561 | } |
4035c07a | 562 | out: |
31dbd01f IE |
563 | cond_resched(); /* we're called from many long loops */ |
564 | } | |
565 | ||
31dbd01f | 566 | static void remove_trailing_rmap_items(struct mm_slot *mm_slot, |
6514d511 | 567 | struct rmap_item **rmap_list) |
31dbd01f | 568 | { |
6514d511 HD |
569 | while (*rmap_list) { |
570 | struct rmap_item *rmap_item = *rmap_list; | |
571 | *rmap_list = rmap_item->rmap_list; | |
31dbd01f | 572 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
573 | free_rmap_item(rmap_item); |
574 | } | |
575 | } | |
576 | ||
577 | /* | |
578 | * Though it's very tempting to unmerge in_stable_tree(rmap_item)s rather | |
579 | * than check every pte of a given vma, the locking doesn't quite work for | |
580 | * that - an rmap_item is assigned to the stable tree after inserting ksm | |
581 | * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing | |
582 | * rmap_items from parent to child at fork time (so as not to waste time | |
583 | * if exit comes before the next scan reaches it). | |
81464e30 HD |
584 | * |
585 | * Similarly, although we'd like to remove rmap_items (so updating counts | |
586 | * and freeing memory) when unmerging an area, it's easier to leave that | |
587 | * to the next pass of ksmd - consider, for example, how ksmd might be | |
588 | * in cmp_and_merge_page on one of the rmap_items we would be removing. | |
31dbd01f | 589 | */ |
d952b791 HD |
590 | static int unmerge_ksm_pages(struct vm_area_struct *vma, |
591 | unsigned long start, unsigned long end) | |
31dbd01f IE |
592 | { |
593 | unsigned long addr; | |
d952b791 | 594 | int err = 0; |
31dbd01f | 595 | |
d952b791 | 596 | for (addr = start; addr < end && !err; addr += PAGE_SIZE) { |
9ba69294 HD |
597 | if (ksm_test_exit(vma->vm_mm)) |
598 | break; | |
d952b791 HD |
599 | if (signal_pending(current)) |
600 | err = -ERESTARTSYS; | |
601 | else | |
602 | err = break_ksm(vma, addr); | |
603 | } | |
604 | return err; | |
31dbd01f IE |
605 | } |
606 | ||
2ffd8679 HD |
607 | #ifdef CONFIG_SYSFS |
608 | /* | |
609 | * Only called through the sysfs control interface: | |
610 | */ | |
d952b791 | 611 | static int unmerge_and_remove_all_rmap_items(void) |
31dbd01f IE |
612 | { |
613 | struct mm_slot *mm_slot; | |
614 | struct mm_struct *mm; | |
615 | struct vm_area_struct *vma; | |
d952b791 HD |
616 | int err = 0; |
617 | ||
618 | spin_lock(&ksm_mmlist_lock); | |
9ba69294 | 619 | ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next, |
d952b791 HD |
620 | struct mm_slot, mm_list); |
621 | spin_unlock(&ksm_mmlist_lock); | |
31dbd01f | 622 | |
9ba69294 HD |
623 | for (mm_slot = ksm_scan.mm_slot; |
624 | mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) { | |
31dbd01f IE |
625 | mm = mm_slot->mm; |
626 | down_read(&mm->mmap_sem); | |
627 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
9ba69294 HD |
628 | if (ksm_test_exit(mm)) |
629 | break; | |
31dbd01f IE |
630 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
631 | continue; | |
d952b791 HD |
632 | err = unmerge_ksm_pages(vma, |
633 | vma->vm_start, vma->vm_end); | |
9ba69294 HD |
634 | if (err) |
635 | goto error; | |
31dbd01f | 636 | } |
9ba69294 | 637 | |
6514d511 | 638 | remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list); |
d952b791 HD |
639 | |
640 | spin_lock(&ksm_mmlist_lock); | |
9ba69294 | 641 | ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next, |
d952b791 | 642 | struct mm_slot, mm_list); |
9ba69294 HD |
643 | if (ksm_test_exit(mm)) { |
644 | hlist_del(&mm_slot->link); | |
645 | list_del(&mm_slot->mm_list); | |
646 | spin_unlock(&ksm_mmlist_lock); | |
647 | ||
648 | free_mm_slot(mm_slot); | |
649 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
650 | up_read(&mm->mmap_sem); | |
651 | mmdrop(mm); | |
652 | } else { | |
653 | spin_unlock(&ksm_mmlist_lock); | |
654 | up_read(&mm->mmap_sem); | |
655 | } | |
31dbd01f IE |
656 | } |
657 | ||
d952b791 | 658 | ksm_scan.seqnr = 0; |
9ba69294 HD |
659 | return 0; |
660 | ||
661 | error: | |
662 | up_read(&mm->mmap_sem); | |
31dbd01f | 663 | spin_lock(&ksm_mmlist_lock); |
d952b791 | 664 | ksm_scan.mm_slot = &ksm_mm_head; |
31dbd01f | 665 | spin_unlock(&ksm_mmlist_lock); |
d952b791 | 666 | return err; |
31dbd01f | 667 | } |
2ffd8679 | 668 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 669 | |
31dbd01f IE |
670 | static u32 calc_checksum(struct page *page) |
671 | { | |
672 | u32 checksum; | |
673 | void *addr = kmap_atomic(page, KM_USER0); | |
674 | checksum = jhash2(addr, PAGE_SIZE / 4, 17); | |
675 | kunmap_atomic(addr, KM_USER0); | |
676 | return checksum; | |
677 | } | |
678 | ||
679 | static int memcmp_pages(struct page *page1, struct page *page2) | |
680 | { | |
681 | char *addr1, *addr2; | |
682 | int ret; | |
683 | ||
684 | addr1 = kmap_atomic(page1, KM_USER0); | |
685 | addr2 = kmap_atomic(page2, KM_USER1); | |
686 | ret = memcmp(addr1, addr2, PAGE_SIZE); | |
687 | kunmap_atomic(addr2, KM_USER1); | |
688 | kunmap_atomic(addr1, KM_USER0); | |
689 | return ret; | |
690 | } | |
691 | ||
692 | static inline int pages_identical(struct page *page1, struct page *page2) | |
693 | { | |
694 | return !memcmp_pages(page1, page2); | |
695 | } | |
696 | ||
697 | static int write_protect_page(struct vm_area_struct *vma, struct page *page, | |
698 | pte_t *orig_pte) | |
699 | { | |
700 | struct mm_struct *mm = vma->vm_mm; | |
701 | unsigned long addr; | |
702 | pte_t *ptep; | |
703 | spinlock_t *ptl; | |
704 | int swapped; | |
705 | int err = -EFAULT; | |
706 | ||
707 | addr = page_address_in_vma(page, vma); | |
708 | if (addr == -EFAULT) | |
709 | goto out; | |
710 | ||
711 | ptep = page_check_address(page, mm, addr, &ptl, 0); | |
712 | if (!ptep) | |
713 | goto out; | |
714 | ||
4e31635c | 715 | if (pte_write(*ptep) || pte_dirty(*ptep)) { |
31dbd01f IE |
716 | pte_t entry; |
717 | ||
718 | swapped = PageSwapCache(page); | |
719 | flush_cache_page(vma, addr, page_to_pfn(page)); | |
720 | /* | |
721 | * Ok this is tricky, when get_user_pages_fast() run it doesnt | |
722 | * take any lock, therefore the check that we are going to make | |
723 | * with the pagecount against the mapcount is racey and | |
724 | * O_DIRECT can happen right after the check. | |
725 | * So we clear the pte and flush the tlb before the check | |
726 | * this assure us that no O_DIRECT can happen after the check | |
727 | * or in the middle of the check. | |
728 | */ | |
729 | entry = ptep_clear_flush(vma, addr, ptep); | |
730 | /* | |
731 | * Check that no O_DIRECT or similar I/O is in progress on the | |
732 | * page | |
733 | */ | |
31e855ea | 734 | if (page_mapcount(page) + 1 + swapped != page_count(page)) { |
cb532375 | 735 | set_pte_at(mm, addr, ptep, entry); |
31dbd01f IE |
736 | goto out_unlock; |
737 | } | |
4e31635c HD |
738 | if (pte_dirty(entry)) |
739 | set_page_dirty(page); | |
740 | entry = pte_mkclean(pte_wrprotect(entry)); | |
31dbd01f IE |
741 | set_pte_at_notify(mm, addr, ptep, entry); |
742 | } | |
743 | *orig_pte = *ptep; | |
744 | err = 0; | |
745 | ||
746 | out_unlock: | |
747 | pte_unmap_unlock(ptep, ptl); | |
748 | out: | |
749 | return err; | |
750 | } | |
751 | ||
752 | /** | |
753 | * replace_page - replace page in vma by new ksm page | |
8dd3557a HD |
754 | * @vma: vma that holds the pte pointing to page |
755 | * @page: the page we are replacing by kpage | |
756 | * @kpage: the ksm page we replace page by | |
31dbd01f IE |
757 | * @orig_pte: the original value of the pte |
758 | * | |
759 | * Returns 0 on success, -EFAULT on failure. | |
760 | */ | |
8dd3557a HD |
761 | static int replace_page(struct vm_area_struct *vma, struct page *page, |
762 | struct page *kpage, pte_t orig_pte) | |
31dbd01f IE |
763 | { |
764 | struct mm_struct *mm = vma->vm_mm; | |
765 | pgd_t *pgd; | |
766 | pud_t *pud; | |
767 | pmd_t *pmd; | |
768 | pte_t *ptep; | |
769 | spinlock_t *ptl; | |
770 | unsigned long addr; | |
31dbd01f IE |
771 | int err = -EFAULT; |
772 | ||
8dd3557a | 773 | addr = page_address_in_vma(page, vma); |
31dbd01f IE |
774 | if (addr == -EFAULT) |
775 | goto out; | |
776 | ||
777 | pgd = pgd_offset(mm, addr); | |
778 | if (!pgd_present(*pgd)) | |
779 | goto out; | |
780 | ||
781 | pud = pud_offset(pgd, addr); | |
782 | if (!pud_present(*pud)) | |
783 | goto out; | |
784 | ||
785 | pmd = pmd_offset(pud, addr); | |
786 | if (!pmd_present(*pmd)) | |
787 | goto out; | |
788 | ||
789 | ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); | |
790 | if (!pte_same(*ptep, orig_pte)) { | |
791 | pte_unmap_unlock(ptep, ptl); | |
792 | goto out; | |
793 | } | |
794 | ||
8dd3557a | 795 | get_page(kpage); |
5ad64688 | 796 | page_add_anon_rmap(kpage, vma, addr); |
31dbd01f IE |
797 | |
798 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
799 | ptep_clear_flush(vma, addr, ptep); | |
8dd3557a | 800 | set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot)); |
31dbd01f | 801 | |
8dd3557a | 802 | page_remove_rmap(page); |
ae52a2ad HD |
803 | if (!page_mapped(page)) |
804 | try_to_free_swap(page); | |
8dd3557a | 805 | put_page(page); |
31dbd01f IE |
806 | |
807 | pte_unmap_unlock(ptep, ptl); | |
808 | err = 0; | |
809 | out: | |
810 | return err; | |
811 | } | |
812 | ||
813 | /* | |
814 | * try_to_merge_one_page - take two pages and merge them into one | |
8dd3557a HD |
815 | * @vma: the vma that holds the pte pointing to page |
816 | * @page: the PageAnon page that we want to replace with kpage | |
80e14822 HD |
817 | * @kpage: the PageKsm page that we want to map instead of page, |
818 | * or NULL the first time when we want to use page as kpage. | |
31dbd01f IE |
819 | * |
820 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
821 | */ | |
822 | static int try_to_merge_one_page(struct vm_area_struct *vma, | |
8dd3557a | 823 | struct page *page, struct page *kpage) |
31dbd01f IE |
824 | { |
825 | pte_t orig_pte = __pte(0); | |
826 | int err = -EFAULT; | |
827 | ||
db114b83 HD |
828 | if (page == kpage) /* ksm page forked */ |
829 | return 0; | |
830 | ||
31dbd01f IE |
831 | if (!(vma->vm_flags & VM_MERGEABLE)) |
832 | goto out; | |
8dd3557a | 833 | if (!PageAnon(page)) |
31dbd01f IE |
834 | goto out; |
835 | ||
31dbd01f IE |
836 | /* |
837 | * We need the page lock to read a stable PageSwapCache in | |
838 | * write_protect_page(). We use trylock_page() instead of | |
839 | * lock_page() because we don't want to wait here - we | |
840 | * prefer to continue scanning and merging different pages, | |
841 | * then come back to this page when it is unlocked. | |
842 | */ | |
8dd3557a | 843 | if (!trylock_page(page)) |
31e855ea | 844 | goto out; |
31dbd01f IE |
845 | /* |
846 | * If this anonymous page is mapped only here, its pte may need | |
847 | * to be write-protected. If it's mapped elsewhere, all of its | |
848 | * ptes are necessarily already write-protected. But in either | |
849 | * case, we need to lock and check page_count is not raised. | |
850 | */ | |
80e14822 HD |
851 | if (write_protect_page(vma, page, &orig_pte) == 0) { |
852 | if (!kpage) { | |
853 | /* | |
854 | * While we hold page lock, upgrade page from | |
855 | * PageAnon+anon_vma to PageKsm+NULL stable_node: | |
856 | * stable_tree_insert() will update stable_node. | |
857 | */ | |
858 | set_page_stable_node(page, NULL); | |
859 | mark_page_accessed(page); | |
860 | err = 0; | |
861 | } else if (pages_identical(page, kpage)) | |
862 | err = replace_page(vma, page, kpage, orig_pte); | |
863 | } | |
31dbd01f | 864 | |
80e14822 | 865 | if ((vma->vm_flags & VM_LOCKED) && kpage && !err) { |
73848b46 | 866 | munlock_vma_page(page); |
5ad64688 HD |
867 | if (!PageMlocked(kpage)) { |
868 | unlock_page(page); | |
5ad64688 HD |
869 | lock_page(kpage); |
870 | mlock_vma_page(kpage); | |
871 | page = kpage; /* for final unlock */ | |
872 | } | |
873 | } | |
73848b46 | 874 | |
8dd3557a | 875 | unlock_page(page); |
31dbd01f IE |
876 | out: |
877 | return err; | |
878 | } | |
879 | ||
81464e30 HD |
880 | /* |
881 | * try_to_merge_with_ksm_page - like try_to_merge_two_pages, | |
882 | * but no new kernel page is allocated: kpage must already be a ksm page. | |
8dd3557a HD |
883 | * |
884 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
81464e30 | 885 | */ |
8dd3557a HD |
886 | static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item, |
887 | struct page *page, struct page *kpage) | |
81464e30 | 888 | { |
8dd3557a | 889 | struct mm_struct *mm = rmap_item->mm; |
81464e30 HD |
890 | struct vm_area_struct *vma; |
891 | int err = -EFAULT; | |
892 | ||
8dd3557a HD |
893 | down_read(&mm->mmap_sem); |
894 | if (ksm_test_exit(mm)) | |
9ba69294 | 895 | goto out; |
8dd3557a HD |
896 | vma = find_vma(mm, rmap_item->address); |
897 | if (!vma || vma->vm_start > rmap_item->address) | |
81464e30 HD |
898 | goto out; |
899 | ||
8dd3557a | 900 | err = try_to_merge_one_page(vma, page, kpage); |
db114b83 HD |
901 | if (err) |
902 | goto out; | |
903 | ||
904 | /* Must get reference to anon_vma while still holding mmap_sem */ | |
905 | hold_anon_vma(rmap_item, vma->anon_vma); | |
81464e30 | 906 | out: |
8dd3557a | 907 | up_read(&mm->mmap_sem); |
81464e30 HD |
908 | return err; |
909 | } | |
910 | ||
31dbd01f IE |
911 | /* |
912 | * try_to_merge_two_pages - take two identical pages and prepare them | |
913 | * to be merged into one page. | |
914 | * | |
8dd3557a HD |
915 | * This function returns the kpage if we successfully merged two identical |
916 | * pages into one ksm page, NULL otherwise. | |
31dbd01f | 917 | * |
80e14822 | 918 | * Note that this function upgrades page to ksm page: if one of the pages |
31dbd01f IE |
919 | * is already a ksm page, try_to_merge_with_ksm_page should be used. |
920 | */ | |
8dd3557a HD |
921 | static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item, |
922 | struct page *page, | |
923 | struct rmap_item *tree_rmap_item, | |
924 | struct page *tree_page) | |
31dbd01f | 925 | { |
80e14822 | 926 | int err; |
31dbd01f | 927 | |
80e14822 | 928 | err = try_to_merge_with_ksm_page(rmap_item, page, NULL); |
31dbd01f | 929 | if (!err) { |
8dd3557a | 930 | err = try_to_merge_with_ksm_page(tree_rmap_item, |
80e14822 | 931 | tree_page, page); |
31dbd01f | 932 | /* |
81464e30 HD |
933 | * If that fails, we have a ksm page with only one pte |
934 | * pointing to it: so break it. | |
31dbd01f | 935 | */ |
4035c07a | 936 | if (err) |
8dd3557a | 937 | break_cow(rmap_item); |
31dbd01f | 938 | } |
80e14822 | 939 | return err ? NULL : page; |
31dbd01f IE |
940 | } |
941 | ||
31dbd01f | 942 | /* |
8dd3557a | 943 | * stable_tree_search - search for page inside the stable tree |
31dbd01f IE |
944 | * |
945 | * This function checks if there is a page inside the stable tree | |
946 | * with identical content to the page that we are scanning right now. | |
947 | * | |
7b6ba2c7 | 948 | * This function returns the stable tree node of identical content if found, |
31dbd01f IE |
949 | * NULL otherwise. |
950 | */ | |
62b61f61 | 951 | static struct page *stable_tree_search(struct page *page) |
31dbd01f IE |
952 | { |
953 | struct rb_node *node = root_stable_tree.rb_node; | |
7b6ba2c7 | 954 | struct stable_node *stable_node; |
31dbd01f | 955 | |
08beca44 HD |
956 | stable_node = page_stable_node(page); |
957 | if (stable_node) { /* ksm page forked */ | |
958 | get_page(page); | |
62b61f61 | 959 | return page; |
08beca44 HD |
960 | } |
961 | ||
31dbd01f | 962 | while (node) { |
4035c07a | 963 | struct page *tree_page; |
31dbd01f IE |
964 | int ret; |
965 | ||
08beca44 | 966 | cond_resched(); |
7b6ba2c7 | 967 | stable_node = rb_entry(node, struct stable_node, node); |
4035c07a HD |
968 | tree_page = get_ksm_page(stable_node); |
969 | if (!tree_page) | |
970 | return NULL; | |
31dbd01f | 971 | |
4035c07a | 972 | ret = memcmp_pages(page, tree_page); |
31dbd01f | 973 | |
4035c07a HD |
974 | if (ret < 0) { |
975 | put_page(tree_page); | |
31dbd01f | 976 | node = node->rb_left; |
4035c07a HD |
977 | } else if (ret > 0) { |
978 | put_page(tree_page); | |
31dbd01f | 979 | node = node->rb_right; |
4035c07a | 980 | } else |
62b61f61 | 981 | return tree_page; |
31dbd01f IE |
982 | } |
983 | ||
984 | return NULL; | |
985 | } | |
986 | ||
987 | /* | |
988 | * stable_tree_insert - insert rmap_item pointing to new ksm page | |
989 | * into the stable tree. | |
990 | * | |
7b6ba2c7 HD |
991 | * This function returns the stable tree node just allocated on success, |
992 | * NULL otherwise. | |
31dbd01f | 993 | */ |
7b6ba2c7 | 994 | static struct stable_node *stable_tree_insert(struct page *kpage) |
31dbd01f IE |
995 | { |
996 | struct rb_node **new = &root_stable_tree.rb_node; | |
997 | struct rb_node *parent = NULL; | |
7b6ba2c7 | 998 | struct stable_node *stable_node; |
31dbd01f IE |
999 | |
1000 | while (*new) { | |
4035c07a | 1001 | struct page *tree_page; |
31dbd01f IE |
1002 | int ret; |
1003 | ||
08beca44 | 1004 | cond_resched(); |
7b6ba2c7 | 1005 | stable_node = rb_entry(*new, struct stable_node, node); |
4035c07a HD |
1006 | tree_page = get_ksm_page(stable_node); |
1007 | if (!tree_page) | |
1008 | return NULL; | |
31dbd01f | 1009 | |
4035c07a HD |
1010 | ret = memcmp_pages(kpage, tree_page); |
1011 | put_page(tree_page); | |
31dbd01f IE |
1012 | |
1013 | parent = *new; | |
1014 | if (ret < 0) | |
1015 | new = &parent->rb_left; | |
1016 | else if (ret > 0) | |
1017 | new = &parent->rb_right; | |
1018 | else { | |
1019 | /* | |
1020 | * It is not a bug that stable_tree_search() didn't | |
1021 | * find this node: because at that time our page was | |
1022 | * not yet write-protected, so may have changed since. | |
1023 | */ | |
1024 | return NULL; | |
1025 | } | |
1026 | } | |
1027 | ||
7b6ba2c7 HD |
1028 | stable_node = alloc_stable_node(); |
1029 | if (!stable_node) | |
1030 | return NULL; | |
31dbd01f | 1031 | |
7b6ba2c7 HD |
1032 | rb_link_node(&stable_node->node, parent, new); |
1033 | rb_insert_color(&stable_node->node, &root_stable_tree); | |
1034 | ||
1035 | INIT_HLIST_HEAD(&stable_node->hlist); | |
1036 | ||
62b61f61 | 1037 | stable_node->kpfn = page_to_pfn(kpage); |
08beca44 HD |
1038 | set_page_stable_node(kpage, stable_node); |
1039 | ||
7b6ba2c7 | 1040 | return stable_node; |
31dbd01f IE |
1041 | } |
1042 | ||
1043 | /* | |
8dd3557a HD |
1044 | * unstable_tree_search_insert - search for identical page, |
1045 | * else insert rmap_item into the unstable tree. | |
31dbd01f IE |
1046 | * |
1047 | * This function searches for a page in the unstable tree identical to the | |
1048 | * page currently being scanned; and if no identical page is found in the | |
1049 | * tree, we insert rmap_item as a new object into the unstable tree. | |
1050 | * | |
1051 | * This function returns pointer to rmap_item found to be identical | |
1052 | * to the currently scanned page, NULL otherwise. | |
1053 | * | |
1054 | * This function does both searching and inserting, because they share | |
1055 | * the same walking algorithm in an rbtree. | |
1056 | */ | |
8dd3557a HD |
1057 | static |
1058 | struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item, | |
1059 | struct page *page, | |
1060 | struct page **tree_pagep) | |
1061 | ||
31dbd01f IE |
1062 | { |
1063 | struct rb_node **new = &root_unstable_tree.rb_node; | |
1064 | struct rb_node *parent = NULL; | |
1065 | ||
1066 | while (*new) { | |
1067 | struct rmap_item *tree_rmap_item; | |
8dd3557a | 1068 | struct page *tree_page; |
31dbd01f IE |
1069 | int ret; |
1070 | ||
d178f27f | 1071 | cond_resched(); |
31dbd01f | 1072 | tree_rmap_item = rb_entry(*new, struct rmap_item, node); |
8dd3557a | 1073 | tree_page = get_mergeable_page(tree_rmap_item); |
22eccdd7 | 1074 | if (IS_ERR_OR_NULL(tree_page)) |
31dbd01f IE |
1075 | return NULL; |
1076 | ||
1077 | /* | |
8dd3557a | 1078 | * Don't substitute a ksm page for a forked page. |
31dbd01f | 1079 | */ |
8dd3557a HD |
1080 | if (page == tree_page) { |
1081 | put_page(tree_page); | |
31dbd01f IE |
1082 | return NULL; |
1083 | } | |
1084 | ||
8dd3557a | 1085 | ret = memcmp_pages(page, tree_page); |
31dbd01f IE |
1086 | |
1087 | parent = *new; | |
1088 | if (ret < 0) { | |
8dd3557a | 1089 | put_page(tree_page); |
31dbd01f IE |
1090 | new = &parent->rb_left; |
1091 | } else if (ret > 0) { | |
8dd3557a | 1092 | put_page(tree_page); |
31dbd01f IE |
1093 | new = &parent->rb_right; |
1094 | } else { | |
8dd3557a | 1095 | *tree_pagep = tree_page; |
31dbd01f IE |
1096 | return tree_rmap_item; |
1097 | } | |
1098 | } | |
1099 | ||
7b6ba2c7 | 1100 | rmap_item->address |= UNSTABLE_FLAG; |
31dbd01f IE |
1101 | rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); |
1102 | rb_link_node(&rmap_item->node, parent, new); | |
1103 | rb_insert_color(&rmap_item->node, &root_unstable_tree); | |
1104 | ||
473b0ce4 | 1105 | ksm_pages_unshared++; |
31dbd01f IE |
1106 | return NULL; |
1107 | } | |
1108 | ||
1109 | /* | |
1110 | * stable_tree_append - add another rmap_item to the linked list of | |
1111 | * rmap_items hanging off a given node of the stable tree, all sharing | |
1112 | * the same ksm page. | |
1113 | */ | |
1114 | static void stable_tree_append(struct rmap_item *rmap_item, | |
7b6ba2c7 | 1115 | struct stable_node *stable_node) |
31dbd01f | 1116 | { |
7b6ba2c7 | 1117 | rmap_item->head = stable_node; |
31dbd01f | 1118 | rmap_item->address |= STABLE_FLAG; |
7b6ba2c7 | 1119 | hlist_add_head(&rmap_item->hlist, &stable_node->hlist); |
e178dfde | 1120 | |
7b6ba2c7 HD |
1121 | if (rmap_item->hlist.next) |
1122 | ksm_pages_sharing++; | |
1123 | else | |
1124 | ksm_pages_shared++; | |
31dbd01f IE |
1125 | } |
1126 | ||
1127 | /* | |
81464e30 HD |
1128 | * cmp_and_merge_page - first see if page can be merged into the stable tree; |
1129 | * if not, compare checksum to previous and if it's the same, see if page can | |
1130 | * be inserted into the unstable tree, or merged with a page already there and | |
1131 | * both transferred to the stable tree. | |
31dbd01f IE |
1132 | * |
1133 | * @page: the page that we are searching identical page to. | |
1134 | * @rmap_item: the reverse mapping into the virtual address of this page | |
1135 | */ | |
1136 | static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item) | |
1137 | { | |
31dbd01f | 1138 | struct rmap_item *tree_rmap_item; |
8dd3557a | 1139 | struct page *tree_page = NULL; |
7b6ba2c7 | 1140 | struct stable_node *stable_node; |
8dd3557a | 1141 | struct page *kpage; |
31dbd01f IE |
1142 | unsigned int checksum; |
1143 | int err; | |
1144 | ||
93d17715 | 1145 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
1146 | |
1147 | /* We first start with searching the page inside the stable tree */ | |
62b61f61 HD |
1148 | kpage = stable_tree_search(page); |
1149 | if (kpage) { | |
08beca44 | 1150 | err = try_to_merge_with_ksm_page(rmap_item, page, kpage); |
31dbd01f IE |
1151 | if (!err) { |
1152 | /* | |
1153 | * The page was successfully merged: | |
1154 | * add its rmap_item to the stable tree. | |
1155 | */ | |
5ad64688 | 1156 | lock_page(kpage); |
62b61f61 | 1157 | stable_tree_append(rmap_item, page_stable_node(kpage)); |
5ad64688 | 1158 | unlock_page(kpage); |
31dbd01f | 1159 | } |
8dd3557a | 1160 | put_page(kpage); |
31dbd01f IE |
1161 | return; |
1162 | } | |
1163 | ||
1164 | /* | |
4035c07a HD |
1165 | * If the hash value of the page has changed from the last time |
1166 | * we calculated it, this page is changing frequently: therefore we | |
1167 | * don't want to insert it in the unstable tree, and we don't want | |
1168 | * to waste our time searching for something identical to it there. | |
31dbd01f IE |
1169 | */ |
1170 | checksum = calc_checksum(page); | |
1171 | if (rmap_item->oldchecksum != checksum) { | |
1172 | rmap_item->oldchecksum = checksum; | |
1173 | return; | |
1174 | } | |
1175 | ||
8dd3557a HD |
1176 | tree_rmap_item = |
1177 | unstable_tree_search_insert(rmap_item, page, &tree_page); | |
31dbd01f | 1178 | if (tree_rmap_item) { |
8dd3557a HD |
1179 | kpage = try_to_merge_two_pages(rmap_item, page, |
1180 | tree_rmap_item, tree_page); | |
1181 | put_page(tree_page); | |
31dbd01f IE |
1182 | /* |
1183 | * As soon as we merge this page, we want to remove the | |
1184 | * rmap_item of the page we have merged with from the unstable | |
1185 | * tree, and insert it instead as new node in the stable tree. | |
1186 | */ | |
8dd3557a | 1187 | if (kpage) { |
93d17715 | 1188 | remove_rmap_item_from_tree(tree_rmap_item); |
473b0ce4 | 1189 | |
5ad64688 | 1190 | lock_page(kpage); |
7b6ba2c7 HD |
1191 | stable_node = stable_tree_insert(kpage); |
1192 | if (stable_node) { | |
1193 | stable_tree_append(tree_rmap_item, stable_node); | |
1194 | stable_tree_append(rmap_item, stable_node); | |
1195 | } | |
5ad64688 | 1196 | unlock_page(kpage); |
7b6ba2c7 | 1197 | |
31dbd01f IE |
1198 | /* |
1199 | * If we fail to insert the page into the stable tree, | |
1200 | * we will have 2 virtual addresses that are pointing | |
1201 | * to a ksm page left outside the stable tree, | |
1202 | * in which case we need to break_cow on both. | |
1203 | */ | |
7b6ba2c7 | 1204 | if (!stable_node) { |
8dd3557a HD |
1205 | break_cow(tree_rmap_item); |
1206 | break_cow(rmap_item); | |
31dbd01f IE |
1207 | } |
1208 | } | |
31dbd01f IE |
1209 | } |
1210 | } | |
1211 | ||
1212 | static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot, | |
6514d511 | 1213 | struct rmap_item **rmap_list, |
31dbd01f IE |
1214 | unsigned long addr) |
1215 | { | |
1216 | struct rmap_item *rmap_item; | |
1217 | ||
6514d511 HD |
1218 | while (*rmap_list) { |
1219 | rmap_item = *rmap_list; | |
93d17715 | 1220 | if ((rmap_item->address & PAGE_MASK) == addr) |
31dbd01f | 1221 | return rmap_item; |
31dbd01f IE |
1222 | if (rmap_item->address > addr) |
1223 | break; | |
6514d511 | 1224 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 1225 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
1226 | free_rmap_item(rmap_item); |
1227 | } | |
1228 | ||
1229 | rmap_item = alloc_rmap_item(); | |
1230 | if (rmap_item) { | |
1231 | /* It has already been zeroed */ | |
1232 | rmap_item->mm = mm_slot->mm; | |
1233 | rmap_item->address = addr; | |
6514d511 HD |
1234 | rmap_item->rmap_list = *rmap_list; |
1235 | *rmap_list = rmap_item; | |
31dbd01f IE |
1236 | } |
1237 | return rmap_item; | |
1238 | } | |
1239 | ||
1240 | static struct rmap_item *scan_get_next_rmap_item(struct page **page) | |
1241 | { | |
1242 | struct mm_struct *mm; | |
1243 | struct mm_slot *slot; | |
1244 | struct vm_area_struct *vma; | |
1245 | struct rmap_item *rmap_item; | |
1246 | ||
1247 | if (list_empty(&ksm_mm_head.mm_list)) | |
1248 | return NULL; | |
1249 | ||
1250 | slot = ksm_scan.mm_slot; | |
1251 | if (slot == &ksm_mm_head) { | |
1252 | root_unstable_tree = RB_ROOT; | |
1253 | ||
1254 | spin_lock(&ksm_mmlist_lock); | |
1255 | slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list); | |
1256 | ksm_scan.mm_slot = slot; | |
1257 | spin_unlock(&ksm_mmlist_lock); | |
1258 | next_mm: | |
1259 | ksm_scan.address = 0; | |
6514d511 | 1260 | ksm_scan.rmap_list = &slot->rmap_list; |
31dbd01f IE |
1261 | } |
1262 | ||
1263 | mm = slot->mm; | |
1264 | down_read(&mm->mmap_sem); | |
9ba69294 HD |
1265 | if (ksm_test_exit(mm)) |
1266 | vma = NULL; | |
1267 | else | |
1268 | vma = find_vma(mm, ksm_scan.address); | |
1269 | ||
1270 | for (; vma; vma = vma->vm_next) { | |
31dbd01f IE |
1271 | if (!(vma->vm_flags & VM_MERGEABLE)) |
1272 | continue; | |
1273 | if (ksm_scan.address < vma->vm_start) | |
1274 | ksm_scan.address = vma->vm_start; | |
1275 | if (!vma->anon_vma) | |
1276 | ksm_scan.address = vma->vm_end; | |
1277 | ||
1278 | while (ksm_scan.address < vma->vm_end) { | |
9ba69294 HD |
1279 | if (ksm_test_exit(mm)) |
1280 | break; | |
31dbd01f | 1281 | *page = follow_page(vma, ksm_scan.address, FOLL_GET); |
22eccdd7 | 1282 | if (!IS_ERR_OR_NULL(*page) && PageAnon(*page)) { |
31dbd01f IE |
1283 | flush_anon_page(vma, *page, ksm_scan.address); |
1284 | flush_dcache_page(*page); | |
1285 | rmap_item = get_next_rmap_item(slot, | |
6514d511 | 1286 | ksm_scan.rmap_list, ksm_scan.address); |
31dbd01f | 1287 | if (rmap_item) { |
6514d511 HD |
1288 | ksm_scan.rmap_list = |
1289 | &rmap_item->rmap_list; | |
31dbd01f IE |
1290 | ksm_scan.address += PAGE_SIZE; |
1291 | } else | |
1292 | put_page(*page); | |
1293 | up_read(&mm->mmap_sem); | |
1294 | return rmap_item; | |
1295 | } | |
22eccdd7 | 1296 | if (!IS_ERR_OR_NULL(*page)) |
31dbd01f IE |
1297 | put_page(*page); |
1298 | ksm_scan.address += PAGE_SIZE; | |
1299 | cond_resched(); | |
1300 | } | |
1301 | } | |
1302 | ||
9ba69294 HD |
1303 | if (ksm_test_exit(mm)) { |
1304 | ksm_scan.address = 0; | |
6514d511 | 1305 | ksm_scan.rmap_list = &slot->rmap_list; |
9ba69294 | 1306 | } |
31dbd01f IE |
1307 | /* |
1308 | * Nuke all the rmap_items that are above this current rmap: | |
1309 | * because there were no VM_MERGEABLE vmas with such addresses. | |
1310 | */ | |
6514d511 | 1311 | remove_trailing_rmap_items(slot, ksm_scan.rmap_list); |
31dbd01f IE |
1312 | |
1313 | spin_lock(&ksm_mmlist_lock); | |
cd551f97 HD |
1314 | ksm_scan.mm_slot = list_entry(slot->mm_list.next, |
1315 | struct mm_slot, mm_list); | |
1316 | if (ksm_scan.address == 0) { | |
1317 | /* | |
1318 | * We've completed a full scan of all vmas, holding mmap_sem | |
1319 | * throughout, and found no VM_MERGEABLE: so do the same as | |
1320 | * __ksm_exit does to remove this mm from all our lists now. | |
9ba69294 HD |
1321 | * This applies either when cleaning up after __ksm_exit |
1322 | * (but beware: we can reach here even before __ksm_exit), | |
1323 | * or when all VM_MERGEABLE areas have been unmapped (and | |
1324 | * mmap_sem then protects against race with MADV_MERGEABLE). | |
cd551f97 HD |
1325 | */ |
1326 | hlist_del(&slot->link); | |
1327 | list_del(&slot->mm_list); | |
9ba69294 HD |
1328 | spin_unlock(&ksm_mmlist_lock); |
1329 | ||
cd551f97 HD |
1330 | free_mm_slot(slot); |
1331 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
9ba69294 HD |
1332 | up_read(&mm->mmap_sem); |
1333 | mmdrop(mm); | |
1334 | } else { | |
1335 | spin_unlock(&ksm_mmlist_lock); | |
1336 | up_read(&mm->mmap_sem); | |
cd551f97 | 1337 | } |
31dbd01f IE |
1338 | |
1339 | /* Repeat until we've completed scanning the whole list */ | |
cd551f97 | 1340 | slot = ksm_scan.mm_slot; |
31dbd01f IE |
1341 | if (slot != &ksm_mm_head) |
1342 | goto next_mm; | |
1343 | ||
31dbd01f IE |
1344 | ksm_scan.seqnr++; |
1345 | return NULL; | |
1346 | } | |
1347 | ||
1348 | /** | |
1349 | * ksm_do_scan - the ksm scanner main worker function. | |
1350 | * @scan_npages - number of pages we want to scan before we return. | |
1351 | */ | |
1352 | static void ksm_do_scan(unsigned int scan_npages) | |
1353 | { | |
1354 | struct rmap_item *rmap_item; | |
22eccdd7 | 1355 | struct page *uninitialized_var(page); |
31dbd01f IE |
1356 | |
1357 | while (scan_npages--) { | |
1358 | cond_resched(); | |
1359 | rmap_item = scan_get_next_rmap_item(&page); | |
1360 | if (!rmap_item) | |
1361 | return; | |
1362 | if (!PageKsm(page) || !in_stable_tree(rmap_item)) | |
1363 | cmp_and_merge_page(page, rmap_item); | |
1364 | put_page(page); | |
1365 | } | |
1366 | } | |
1367 | ||
6e158384 HD |
1368 | static int ksmd_should_run(void) |
1369 | { | |
1370 | return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list); | |
1371 | } | |
1372 | ||
31dbd01f IE |
1373 | static int ksm_scan_thread(void *nothing) |
1374 | { | |
339aa624 | 1375 | set_user_nice(current, 5); |
31dbd01f IE |
1376 | |
1377 | while (!kthread_should_stop()) { | |
6e158384 HD |
1378 | mutex_lock(&ksm_thread_mutex); |
1379 | if (ksmd_should_run()) | |
31dbd01f | 1380 | ksm_do_scan(ksm_thread_pages_to_scan); |
6e158384 HD |
1381 | mutex_unlock(&ksm_thread_mutex); |
1382 | ||
1383 | if (ksmd_should_run()) { | |
31dbd01f IE |
1384 | schedule_timeout_interruptible( |
1385 | msecs_to_jiffies(ksm_thread_sleep_millisecs)); | |
1386 | } else { | |
1387 | wait_event_interruptible(ksm_thread_wait, | |
6e158384 | 1388 | ksmd_should_run() || kthread_should_stop()); |
31dbd01f IE |
1389 | } |
1390 | } | |
1391 | return 0; | |
1392 | } | |
1393 | ||
f8af4da3 HD |
1394 | int ksm_madvise(struct vm_area_struct *vma, unsigned long start, |
1395 | unsigned long end, int advice, unsigned long *vm_flags) | |
1396 | { | |
1397 | struct mm_struct *mm = vma->vm_mm; | |
d952b791 | 1398 | int err; |
f8af4da3 HD |
1399 | |
1400 | switch (advice) { | |
1401 | case MADV_MERGEABLE: | |
1402 | /* | |
1403 | * Be somewhat over-protective for now! | |
1404 | */ | |
1405 | if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE | | |
1406 | VM_PFNMAP | VM_IO | VM_DONTEXPAND | | |
1407 | VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE | | |
5ad64688 | 1408 | VM_NONLINEAR | VM_MIXEDMAP | VM_SAO)) |
f8af4da3 HD |
1409 | return 0; /* just ignore the advice */ |
1410 | ||
d952b791 HD |
1411 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { |
1412 | err = __ksm_enter(mm); | |
1413 | if (err) | |
1414 | return err; | |
1415 | } | |
f8af4da3 HD |
1416 | |
1417 | *vm_flags |= VM_MERGEABLE; | |
1418 | break; | |
1419 | ||
1420 | case MADV_UNMERGEABLE: | |
1421 | if (!(*vm_flags & VM_MERGEABLE)) | |
1422 | return 0; /* just ignore the advice */ | |
1423 | ||
d952b791 HD |
1424 | if (vma->anon_vma) { |
1425 | err = unmerge_ksm_pages(vma, start, end); | |
1426 | if (err) | |
1427 | return err; | |
1428 | } | |
f8af4da3 HD |
1429 | |
1430 | *vm_flags &= ~VM_MERGEABLE; | |
1431 | break; | |
1432 | } | |
1433 | ||
1434 | return 0; | |
1435 | } | |
1436 | ||
1437 | int __ksm_enter(struct mm_struct *mm) | |
1438 | { | |
6e158384 HD |
1439 | struct mm_slot *mm_slot; |
1440 | int needs_wakeup; | |
1441 | ||
1442 | mm_slot = alloc_mm_slot(); | |
31dbd01f IE |
1443 | if (!mm_slot) |
1444 | return -ENOMEM; | |
1445 | ||
6e158384 HD |
1446 | /* Check ksm_run too? Would need tighter locking */ |
1447 | needs_wakeup = list_empty(&ksm_mm_head.mm_list); | |
1448 | ||
31dbd01f IE |
1449 | spin_lock(&ksm_mmlist_lock); |
1450 | insert_to_mm_slots_hash(mm, mm_slot); | |
1451 | /* | |
1452 | * Insert just behind the scanning cursor, to let the area settle | |
1453 | * down a little; when fork is followed by immediate exec, we don't | |
1454 | * want ksmd to waste time setting up and tearing down an rmap_list. | |
1455 | */ | |
1456 | list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list); | |
1457 | spin_unlock(&ksm_mmlist_lock); | |
1458 | ||
f8af4da3 | 1459 | set_bit(MMF_VM_MERGEABLE, &mm->flags); |
9ba69294 | 1460 | atomic_inc(&mm->mm_count); |
6e158384 HD |
1461 | |
1462 | if (needs_wakeup) | |
1463 | wake_up_interruptible(&ksm_thread_wait); | |
1464 | ||
f8af4da3 HD |
1465 | return 0; |
1466 | } | |
1467 | ||
1c2fb7a4 | 1468 | void __ksm_exit(struct mm_struct *mm) |
f8af4da3 | 1469 | { |
cd551f97 | 1470 | struct mm_slot *mm_slot; |
9ba69294 | 1471 | int easy_to_free = 0; |
cd551f97 | 1472 | |
31dbd01f | 1473 | /* |
9ba69294 HD |
1474 | * This process is exiting: if it's straightforward (as is the |
1475 | * case when ksmd was never running), free mm_slot immediately. | |
1476 | * But if it's at the cursor or has rmap_items linked to it, use | |
1477 | * mmap_sem to synchronize with any break_cows before pagetables | |
1478 | * are freed, and leave the mm_slot on the list for ksmd to free. | |
1479 | * Beware: ksm may already have noticed it exiting and freed the slot. | |
31dbd01f | 1480 | */ |
9ba69294 | 1481 | |
cd551f97 HD |
1482 | spin_lock(&ksm_mmlist_lock); |
1483 | mm_slot = get_mm_slot(mm); | |
9ba69294 | 1484 | if (mm_slot && ksm_scan.mm_slot != mm_slot) { |
6514d511 | 1485 | if (!mm_slot->rmap_list) { |
9ba69294 HD |
1486 | hlist_del(&mm_slot->link); |
1487 | list_del(&mm_slot->mm_list); | |
1488 | easy_to_free = 1; | |
1489 | } else { | |
1490 | list_move(&mm_slot->mm_list, | |
1491 | &ksm_scan.mm_slot->mm_list); | |
1492 | } | |
cd551f97 | 1493 | } |
cd551f97 HD |
1494 | spin_unlock(&ksm_mmlist_lock); |
1495 | ||
9ba69294 HD |
1496 | if (easy_to_free) { |
1497 | free_mm_slot(mm_slot); | |
1498 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
1499 | mmdrop(mm); | |
1500 | } else if (mm_slot) { | |
9ba69294 HD |
1501 | down_write(&mm->mmap_sem); |
1502 | up_write(&mm->mmap_sem); | |
9ba69294 | 1503 | } |
31dbd01f IE |
1504 | } |
1505 | ||
5ad64688 HD |
1506 | struct page *ksm_does_need_to_copy(struct page *page, |
1507 | struct vm_area_struct *vma, unsigned long address) | |
1508 | { | |
1509 | struct page *new_page; | |
1510 | ||
5ad64688 HD |
1511 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
1512 | if (new_page) { | |
1513 | copy_user_highpage(new_page, page, address, vma); | |
1514 | ||
1515 | SetPageDirty(new_page); | |
1516 | __SetPageUptodate(new_page); | |
1517 | SetPageSwapBacked(new_page); | |
1518 | __set_page_locked(new_page); | |
1519 | ||
1520 | if (page_evictable(new_page, vma)) | |
1521 | lru_cache_add_lru(new_page, LRU_ACTIVE_ANON); | |
1522 | else | |
1523 | add_page_to_unevictable_list(new_page); | |
1524 | } | |
1525 | ||
5ad64688 HD |
1526 | return new_page; |
1527 | } | |
1528 | ||
1529 | int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg, | |
1530 | unsigned long *vm_flags) | |
1531 | { | |
1532 | struct stable_node *stable_node; | |
1533 | struct rmap_item *rmap_item; | |
1534 | struct hlist_node *hlist; | |
1535 | unsigned int mapcount = page_mapcount(page); | |
1536 | int referenced = 0; | |
db114b83 | 1537 | int search_new_forks = 0; |
5ad64688 HD |
1538 | |
1539 | VM_BUG_ON(!PageKsm(page)); | |
1540 | VM_BUG_ON(!PageLocked(page)); | |
1541 | ||
1542 | stable_node = page_stable_node(page); | |
1543 | if (!stable_node) | |
1544 | return 0; | |
db114b83 | 1545 | again: |
5ad64688 | 1546 | hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) { |
db114b83 | 1547 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
5beb4930 | 1548 | struct anon_vma_chain *vmac; |
db114b83 | 1549 | struct vm_area_struct *vma; |
5ad64688 | 1550 | |
cba48b98 | 1551 | anon_vma_lock(anon_vma); |
5beb4930 RR |
1552 | list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) { |
1553 | vma = vmac->vma; | |
db114b83 HD |
1554 | if (rmap_item->address < vma->vm_start || |
1555 | rmap_item->address >= vma->vm_end) | |
1556 | continue; | |
1557 | /* | |
1558 | * Initially we examine only the vma which covers this | |
1559 | * rmap_item; but later, if there is still work to do, | |
1560 | * we examine covering vmas in other mms: in case they | |
1561 | * were forked from the original since ksmd passed. | |
1562 | */ | |
1563 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
1564 | continue; | |
1565 | ||
1566 | if (memcg && !mm_match_cgroup(vma->vm_mm, memcg)) | |
1567 | continue; | |
5ad64688 | 1568 | |
db114b83 | 1569 | referenced += page_referenced_one(page, vma, |
5ad64688 | 1570 | rmap_item->address, &mapcount, vm_flags); |
db114b83 HD |
1571 | if (!search_new_forks || !mapcount) |
1572 | break; | |
1573 | } | |
cba48b98 | 1574 | anon_vma_unlock(anon_vma); |
5ad64688 HD |
1575 | if (!mapcount) |
1576 | goto out; | |
1577 | } | |
db114b83 HD |
1578 | if (!search_new_forks++) |
1579 | goto again; | |
5ad64688 | 1580 | out: |
5ad64688 HD |
1581 | return referenced; |
1582 | } | |
1583 | ||
1584 | int try_to_unmap_ksm(struct page *page, enum ttu_flags flags) | |
1585 | { | |
1586 | struct stable_node *stable_node; | |
1587 | struct hlist_node *hlist; | |
1588 | struct rmap_item *rmap_item; | |
1589 | int ret = SWAP_AGAIN; | |
db114b83 | 1590 | int search_new_forks = 0; |
5ad64688 HD |
1591 | |
1592 | VM_BUG_ON(!PageKsm(page)); | |
1593 | VM_BUG_ON(!PageLocked(page)); | |
1594 | ||
1595 | stable_node = page_stable_node(page); | |
1596 | if (!stable_node) | |
1597 | return SWAP_FAIL; | |
db114b83 | 1598 | again: |
5ad64688 | 1599 | hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) { |
db114b83 | 1600 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
5beb4930 | 1601 | struct anon_vma_chain *vmac; |
db114b83 | 1602 | struct vm_area_struct *vma; |
5ad64688 | 1603 | |
cba48b98 | 1604 | anon_vma_lock(anon_vma); |
5beb4930 RR |
1605 | list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) { |
1606 | vma = vmac->vma; | |
db114b83 HD |
1607 | if (rmap_item->address < vma->vm_start || |
1608 | rmap_item->address >= vma->vm_end) | |
1609 | continue; | |
1610 | /* | |
1611 | * Initially we examine only the vma which covers this | |
1612 | * rmap_item; but later, if there is still work to do, | |
1613 | * we examine covering vmas in other mms: in case they | |
1614 | * were forked from the original since ksmd passed. | |
1615 | */ | |
1616 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
1617 | continue; | |
1618 | ||
1619 | ret = try_to_unmap_one(page, vma, | |
1620 | rmap_item->address, flags); | |
1621 | if (ret != SWAP_AGAIN || !page_mapped(page)) { | |
cba48b98 | 1622 | anon_vma_unlock(anon_vma); |
db114b83 HD |
1623 | goto out; |
1624 | } | |
1625 | } | |
cba48b98 | 1626 | anon_vma_unlock(anon_vma); |
5ad64688 | 1627 | } |
db114b83 HD |
1628 | if (!search_new_forks++) |
1629 | goto again; | |
5ad64688 | 1630 | out: |
5ad64688 HD |
1631 | return ret; |
1632 | } | |
1633 | ||
e9995ef9 HD |
1634 | #ifdef CONFIG_MIGRATION |
1635 | int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *, | |
1636 | struct vm_area_struct *, unsigned long, void *), void *arg) | |
1637 | { | |
1638 | struct stable_node *stable_node; | |
1639 | struct hlist_node *hlist; | |
1640 | struct rmap_item *rmap_item; | |
1641 | int ret = SWAP_AGAIN; | |
1642 | int search_new_forks = 0; | |
1643 | ||
1644 | VM_BUG_ON(!PageKsm(page)); | |
1645 | VM_BUG_ON(!PageLocked(page)); | |
1646 | ||
1647 | stable_node = page_stable_node(page); | |
1648 | if (!stable_node) | |
1649 | return ret; | |
1650 | again: | |
1651 | hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) { | |
1652 | struct anon_vma *anon_vma = rmap_item->anon_vma; | |
5beb4930 | 1653 | struct anon_vma_chain *vmac; |
e9995ef9 HD |
1654 | struct vm_area_struct *vma; |
1655 | ||
cba48b98 | 1656 | anon_vma_lock(anon_vma); |
5beb4930 RR |
1657 | list_for_each_entry(vmac, &anon_vma->head, same_anon_vma) { |
1658 | vma = vmac->vma; | |
e9995ef9 HD |
1659 | if (rmap_item->address < vma->vm_start || |
1660 | rmap_item->address >= vma->vm_end) | |
1661 | continue; | |
1662 | /* | |
1663 | * Initially we examine only the vma which covers this | |
1664 | * rmap_item; but later, if there is still work to do, | |
1665 | * we examine covering vmas in other mms: in case they | |
1666 | * were forked from the original since ksmd passed. | |
1667 | */ | |
1668 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
1669 | continue; | |
1670 | ||
1671 | ret = rmap_one(page, vma, rmap_item->address, arg); | |
1672 | if (ret != SWAP_AGAIN) { | |
cba48b98 | 1673 | anon_vma_unlock(anon_vma); |
e9995ef9 HD |
1674 | goto out; |
1675 | } | |
1676 | } | |
cba48b98 | 1677 | anon_vma_unlock(anon_vma); |
e9995ef9 HD |
1678 | } |
1679 | if (!search_new_forks++) | |
1680 | goto again; | |
1681 | out: | |
1682 | return ret; | |
1683 | } | |
1684 | ||
1685 | void ksm_migrate_page(struct page *newpage, struct page *oldpage) | |
1686 | { | |
1687 | struct stable_node *stable_node; | |
1688 | ||
1689 | VM_BUG_ON(!PageLocked(oldpage)); | |
1690 | VM_BUG_ON(!PageLocked(newpage)); | |
1691 | VM_BUG_ON(newpage->mapping != oldpage->mapping); | |
1692 | ||
1693 | stable_node = page_stable_node(newpage); | |
1694 | if (stable_node) { | |
62b61f61 HD |
1695 | VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage)); |
1696 | stable_node->kpfn = page_to_pfn(newpage); | |
e9995ef9 HD |
1697 | } |
1698 | } | |
1699 | #endif /* CONFIG_MIGRATION */ | |
1700 | ||
62b61f61 HD |
1701 | #ifdef CONFIG_MEMORY_HOTREMOVE |
1702 | static struct stable_node *ksm_check_stable_tree(unsigned long start_pfn, | |
1703 | unsigned long end_pfn) | |
1704 | { | |
1705 | struct rb_node *node; | |
1706 | ||
1707 | for (node = rb_first(&root_stable_tree); node; node = rb_next(node)) { | |
1708 | struct stable_node *stable_node; | |
1709 | ||
1710 | stable_node = rb_entry(node, struct stable_node, node); | |
1711 | if (stable_node->kpfn >= start_pfn && | |
1712 | stable_node->kpfn < end_pfn) | |
1713 | return stable_node; | |
1714 | } | |
1715 | return NULL; | |
1716 | } | |
1717 | ||
1718 | static int ksm_memory_callback(struct notifier_block *self, | |
1719 | unsigned long action, void *arg) | |
1720 | { | |
1721 | struct memory_notify *mn = arg; | |
1722 | struct stable_node *stable_node; | |
1723 | ||
1724 | switch (action) { | |
1725 | case MEM_GOING_OFFLINE: | |
1726 | /* | |
1727 | * Keep it very simple for now: just lock out ksmd and | |
1728 | * MADV_UNMERGEABLE while any memory is going offline. | |
a0b0f58c KM |
1729 | * mutex_lock_nested() is necessary because lockdep was alarmed |
1730 | * that here we take ksm_thread_mutex inside notifier chain | |
1731 | * mutex, and later take notifier chain mutex inside | |
1732 | * ksm_thread_mutex to unlock it. But that's safe because both | |
1733 | * are inside mem_hotplug_mutex. | |
62b61f61 | 1734 | */ |
a0b0f58c | 1735 | mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING); |
62b61f61 HD |
1736 | break; |
1737 | ||
1738 | case MEM_OFFLINE: | |
1739 | /* | |
1740 | * Most of the work is done by page migration; but there might | |
1741 | * be a few stable_nodes left over, still pointing to struct | |
1742 | * pages which have been offlined: prune those from the tree. | |
1743 | */ | |
1744 | while ((stable_node = ksm_check_stable_tree(mn->start_pfn, | |
1745 | mn->start_pfn + mn->nr_pages)) != NULL) | |
1746 | remove_node_from_stable_tree(stable_node); | |
1747 | /* fallthrough */ | |
1748 | ||
1749 | case MEM_CANCEL_OFFLINE: | |
1750 | mutex_unlock(&ksm_thread_mutex); | |
1751 | break; | |
1752 | } | |
1753 | return NOTIFY_OK; | |
1754 | } | |
1755 | #endif /* CONFIG_MEMORY_HOTREMOVE */ | |
1756 | ||
2ffd8679 HD |
1757 | #ifdef CONFIG_SYSFS |
1758 | /* | |
1759 | * This all compiles without CONFIG_SYSFS, but is a waste of space. | |
1760 | */ | |
1761 | ||
31dbd01f IE |
1762 | #define KSM_ATTR_RO(_name) \ |
1763 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1764 | #define KSM_ATTR(_name) \ | |
1765 | static struct kobj_attribute _name##_attr = \ | |
1766 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1767 | ||
1768 | static ssize_t sleep_millisecs_show(struct kobject *kobj, | |
1769 | struct kobj_attribute *attr, char *buf) | |
1770 | { | |
1771 | return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs); | |
1772 | } | |
1773 | ||
1774 | static ssize_t sleep_millisecs_store(struct kobject *kobj, | |
1775 | struct kobj_attribute *attr, | |
1776 | const char *buf, size_t count) | |
1777 | { | |
1778 | unsigned long msecs; | |
1779 | int err; | |
1780 | ||
1781 | err = strict_strtoul(buf, 10, &msecs); | |
1782 | if (err || msecs > UINT_MAX) | |
1783 | return -EINVAL; | |
1784 | ||
1785 | ksm_thread_sleep_millisecs = msecs; | |
1786 | ||
1787 | return count; | |
1788 | } | |
1789 | KSM_ATTR(sleep_millisecs); | |
1790 | ||
1791 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
1792 | struct kobj_attribute *attr, char *buf) | |
1793 | { | |
1794 | return sprintf(buf, "%u\n", ksm_thread_pages_to_scan); | |
1795 | } | |
1796 | ||
1797 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
1798 | struct kobj_attribute *attr, | |
1799 | const char *buf, size_t count) | |
1800 | { | |
1801 | int err; | |
1802 | unsigned long nr_pages; | |
1803 | ||
1804 | err = strict_strtoul(buf, 10, &nr_pages); | |
1805 | if (err || nr_pages > UINT_MAX) | |
1806 | return -EINVAL; | |
1807 | ||
1808 | ksm_thread_pages_to_scan = nr_pages; | |
1809 | ||
1810 | return count; | |
1811 | } | |
1812 | KSM_ATTR(pages_to_scan); | |
1813 | ||
1814 | static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, | |
1815 | char *buf) | |
1816 | { | |
1817 | return sprintf(buf, "%u\n", ksm_run); | |
1818 | } | |
1819 | ||
1820 | static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, | |
1821 | const char *buf, size_t count) | |
1822 | { | |
1823 | int err; | |
1824 | unsigned long flags; | |
1825 | ||
1826 | err = strict_strtoul(buf, 10, &flags); | |
1827 | if (err || flags > UINT_MAX) | |
1828 | return -EINVAL; | |
1829 | if (flags > KSM_RUN_UNMERGE) | |
1830 | return -EINVAL; | |
1831 | ||
1832 | /* | |
1833 | * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. | |
1834 | * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, | |
d0f209f6 HD |
1835 | * breaking COW to free the pages_shared (but leaves mm_slots |
1836 | * on the list for when ksmd may be set running again). | |
31dbd01f IE |
1837 | */ |
1838 | ||
1839 | mutex_lock(&ksm_thread_mutex); | |
1840 | if (ksm_run != flags) { | |
1841 | ksm_run = flags; | |
d952b791 | 1842 | if (flags & KSM_RUN_UNMERGE) { |
35451bee | 1843 | current->flags |= PF_OOM_ORIGIN; |
d952b791 | 1844 | err = unmerge_and_remove_all_rmap_items(); |
35451bee | 1845 | current->flags &= ~PF_OOM_ORIGIN; |
d952b791 HD |
1846 | if (err) { |
1847 | ksm_run = KSM_RUN_STOP; | |
1848 | count = err; | |
1849 | } | |
1850 | } | |
31dbd01f IE |
1851 | } |
1852 | mutex_unlock(&ksm_thread_mutex); | |
1853 | ||
1854 | if (flags & KSM_RUN_MERGE) | |
1855 | wake_up_interruptible(&ksm_thread_wait); | |
1856 | ||
1857 | return count; | |
1858 | } | |
1859 | KSM_ATTR(run); | |
1860 | ||
b4028260 HD |
1861 | static ssize_t pages_shared_show(struct kobject *kobj, |
1862 | struct kobj_attribute *attr, char *buf) | |
1863 | { | |
1864 | return sprintf(buf, "%lu\n", ksm_pages_shared); | |
1865 | } | |
1866 | KSM_ATTR_RO(pages_shared); | |
1867 | ||
1868 | static ssize_t pages_sharing_show(struct kobject *kobj, | |
1869 | struct kobj_attribute *attr, char *buf) | |
1870 | { | |
e178dfde | 1871 | return sprintf(buf, "%lu\n", ksm_pages_sharing); |
b4028260 HD |
1872 | } |
1873 | KSM_ATTR_RO(pages_sharing); | |
1874 | ||
473b0ce4 HD |
1875 | static ssize_t pages_unshared_show(struct kobject *kobj, |
1876 | struct kobj_attribute *attr, char *buf) | |
1877 | { | |
1878 | return sprintf(buf, "%lu\n", ksm_pages_unshared); | |
1879 | } | |
1880 | KSM_ATTR_RO(pages_unshared); | |
1881 | ||
1882 | static ssize_t pages_volatile_show(struct kobject *kobj, | |
1883 | struct kobj_attribute *attr, char *buf) | |
1884 | { | |
1885 | long ksm_pages_volatile; | |
1886 | ||
1887 | ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared | |
1888 | - ksm_pages_sharing - ksm_pages_unshared; | |
1889 | /* | |
1890 | * It was not worth any locking to calculate that statistic, | |
1891 | * but it might therefore sometimes be negative: conceal that. | |
1892 | */ | |
1893 | if (ksm_pages_volatile < 0) | |
1894 | ksm_pages_volatile = 0; | |
1895 | return sprintf(buf, "%ld\n", ksm_pages_volatile); | |
1896 | } | |
1897 | KSM_ATTR_RO(pages_volatile); | |
1898 | ||
1899 | static ssize_t full_scans_show(struct kobject *kobj, | |
1900 | struct kobj_attribute *attr, char *buf) | |
1901 | { | |
1902 | return sprintf(buf, "%lu\n", ksm_scan.seqnr); | |
1903 | } | |
1904 | KSM_ATTR_RO(full_scans); | |
1905 | ||
31dbd01f IE |
1906 | static struct attribute *ksm_attrs[] = { |
1907 | &sleep_millisecs_attr.attr, | |
1908 | &pages_to_scan_attr.attr, | |
1909 | &run_attr.attr, | |
b4028260 HD |
1910 | &pages_shared_attr.attr, |
1911 | &pages_sharing_attr.attr, | |
473b0ce4 HD |
1912 | &pages_unshared_attr.attr, |
1913 | &pages_volatile_attr.attr, | |
1914 | &full_scans_attr.attr, | |
31dbd01f IE |
1915 | NULL, |
1916 | }; | |
1917 | ||
1918 | static struct attribute_group ksm_attr_group = { | |
1919 | .attrs = ksm_attrs, | |
1920 | .name = "ksm", | |
1921 | }; | |
2ffd8679 | 1922 | #endif /* CONFIG_SYSFS */ |
31dbd01f IE |
1923 | |
1924 | static int __init ksm_init(void) | |
1925 | { | |
1926 | struct task_struct *ksm_thread; | |
1927 | int err; | |
1928 | ||
1929 | err = ksm_slab_init(); | |
1930 | if (err) | |
1931 | goto out; | |
1932 | ||
31dbd01f IE |
1933 | ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); |
1934 | if (IS_ERR(ksm_thread)) { | |
1935 | printk(KERN_ERR "ksm: creating kthread failed\n"); | |
1936 | err = PTR_ERR(ksm_thread); | |
d9f8984c | 1937 | goto out_free; |
31dbd01f IE |
1938 | } |
1939 | ||
2ffd8679 | 1940 | #ifdef CONFIG_SYSFS |
31dbd01f IE |
1941 | err = sysfs_create_group(mm_kobj, &ksm_attr_group); |
1942 | if (err) { | |
1943 | printk(KERN_ERR "ksm: register sysfs failed\n"); | |
2ffd8679 | 1944 | kthread_stop(ksm_thread); |
d9f8984c | 1945 | goto out_free; |
31dbd01f | 1946 | } |
c73602ad HD |
1947 | #else |
1948 | ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ | |
1949 | ||
2ffd8679 | 1950 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 1951 | |
62b61f61 HD |
1952 | #ifdef CONFIG_MEMORY_HOTREMOVE |
1953 | /* | |
1954 | * Choose a high priority since the callback takes ksm_thread_mutex: | |
1955 | * later callbacks could only be taking locks which nest within that. | |
1956 | */ | |
1957 | hotplug_memory_notifier(ksm_memory_callback, 100); | |
1958 | #endif | |
31dbd01f IE |
1959 | return 0; |
1960 | ||
d9f8984c | 1961 | out_free: |
31dbd01f IE |
1962 | ksm_slab_free(); |
1963 | out: | |
1964 | return err; | |
f8af4da3 | 1965 | } |
31dbd01f | 1966 | module_init(ksm_init) |