]>
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> |
4ca3a69b | 36 | #include <linux/hashtable.h> |
878aee7d | 37 | #include <linux/freezer.h> |
72788c38 | 38 | #include <linux/oom.h> |
90bd6fd3 | 39 | #include <linux/numa.h> |
f8af4da3 | 40 | |
31dbd01f | 41 | #include <asm/tlbflush.h> |
73848b46 | 42 | #include "internal.h" |
31dbd01f | 43 | |
e850dcf5 HD |
44 | #ifdef CONFIG_NUMA |
45 | #define NUMA(x) (x) | |
46 | #define DO_NUMA(x) do { (x); } while (0) | |
47 | #else | |
48 | #define NUMA(x) (0) | |
49 | #define DO_NUMA(x) do { } while (0) | |
50 | #endif | |
51 | ||
31dbd01f IE |
52 | /* |
53 | * A few notes about the KSM scanning process, | |
54 | * to make it easier to understand the data structures below: | |
55 | * | |
56 | * In order to reduce excessive scanning, KSM sorts the memory pages by their | |
57 | * contents into a data structure that holds pointers to the pages' locations. | |
58 | * | |
59 | * Since the contents of the pages may change at any moment, KSM cannot just | |
60 | * insert the pages into a normal sorted tree and expect it to find anything. | |
61 | * Therefore KSM uses two data structures - the stable and the unstable tree. | |
62 | * | |
63 | * The stable tree holds pointers to all the merged pages (ksm pages), sorted | |
64 | * by their contents. Because each such page is write-protected, searching on | |
65 | * this tree is fully assured to be working (except when pages are unmapped), | |
66 | * and therefore this tree is called the stable tree. | |
67 | * | |
68 | * In addition to the stable tree, KSM uses a second data structure called the | |
69 | * unstable tree: this tree holds pointers to pages which have been found to | |
70 | * be "unchanged for a period of time". The unstable tree sorts these pages | |
71 | * by their contents, but since they are not write-protected, KSM cannot rely | |
72 | * upon the unstable tree to work correctly - the unstable tree is liable to | |
73 | * be corrupted as its contents are modified, and so it is called unstable. | |
74 | * | |
75 | * KSM solves this problem by several techniques: | |
76 | * | |
77 | * 1) The unstable tree is flushed every time KSM completes scanning all | |
78 | * memory areas, and then the tree is rebuilt again from the beginning. | |
79 | * 2) KSM will only insert into the unstable tree, pages whose hash value | |
80 | * has not changed since the previous scan of all memory areas. | |
81 | * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the | |
82 | * colors of the nodes and not on their contents, assuring that even when | |
83 | * the tree gets "corrupted" it won't get out of balance, so scanning time | |
84 | * remains the same (also, searching and inserting nodes in an rbtree uses | |
85 | * the same algorithm, so we have no overhead when we flush and rebuild). | |
86 | * 4) KSM never flushes the stable tree, which means that even if it were to | |
87 | * take 10 attempts to find a page in the unstable tree, once it is found, | |
88 | * it is secured in the stable tree. (When we scan a new page, we first | |
89 | * compare it against the stable tree, and then against the unstable tree.) | |
90 | */ | |
91 | ||
92 | /** | |
93 | * struct mm_slot - ksm information per mm that is being scanned | |
94 | * @link: link to the mm_slots hash list | |
95 | * @mm_list: link into the mm_slots list, rooted in ksm_mm_head | |
6514d511 | 96 | * @rmap_list: head for this mm_slot's singly-linked list of rmap_items |
31dbd01f IE |
97 | * @mm: the mm that this information is valid for |
98 | */ | |
99 | struct mm_slot { | |
100 | struct hlist_node link; | |
101 | struct list_head mm_list; | |
6514d511 | 102 | struct rmap_item *rmap_list; |
31dbd01f IE |
103 | struct mm_struct *mm; |
104 | }; | |
105 | ||
106 | /** | |
107 | * struct ksm_scan - cursor for scanning | |
108 | * @mm_slot: the current mm_slot we are scanning | |
109 | * @address: the next address inside that to be scanned | |
6514d511 | 110 | * @rmap_list: link to the next rmap to be scanned in the rmap_list |
31dbd01f IE |
111 | * @seqnr: count of completed full scans (needed when removing unstable node) |
112 | * | |
113 | * There is only the one ksm_scan instance of this cursor structure. | |
114 | */ | |
115 | struct ksm_scan { | |
116 | struct mm_slot *mm_slot; | |
117 | unsigned long address; | |
6514d511 | 118 | struct rmap_item **rmap_list; |
31dbd01f IE |
119 | unsigned long seqnr; |
120 | }; | |
121 | ||
7b6ba2c7 HD |
122 | /** |
123 | * struct stable_node - node of the stable rbtree | |
124 | * @node: rb node of this ksm page in the stable tree | |
4146d2d6 HD |
125 | * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list |
126 | * @list: linked into migrate_nodes, pending placement in the proper node tree | |
7b6ba2c7 | 127 | * @hlist: hlist head of rmap_items using this ksm page |
4146d2d6 HD |
128 | * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) |
129 | * @nid: NUMA node id of stable tree in which linked (may not match kpfn) | |
7b6ba2c7 HD |
130 | */ |
131 | struct stable_node { | |
4146d2d6 HD |
132 | union { |
133 | struct rb_node node; /* when node of stable tree */ | |
134 | struct { /* when listed for migration */ | |
135 | struct list_head *head; | |
136 | struct list_head list; | |
137 | }; | |
138 | }; | |
7b6ba2c7 | 139 | struct hlist_head hlist; |
62b61f61 | 140 | unsigned long kpfn; |
4146d2d6 HD |
141 | #ifdef CONFIG_NUMA |
142 | int nid; | |
143 | #endif | |
7b6ba2c7 HD |
144 | }; |
145 | ||
31dbd01f IE |
146 | /** |
147 | * struct rmap_item - reverse mapping item for virtual addresses | |
6514d511 | 148 | * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list |
db114b83 | 149 | * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree |
31dbd01f IE |
150 | * @mm: the memory structure this rmap_item is pointing into |
151 | * @address: the virtual address this rmap_item tracks (+ flags in low bits) | |
152 | * @oldchecksum: previous checksum of the page at that virtual address | |
e850dcf5 | 153 | * @nid: NUMA node id of unstable tree in which linked (may not match page) |
7b6ba2c7 HD |
154 | * @node: rb node of this rmap_item in the unstable tree |
155 | * @head: pointer to stable_node heading this list in the stable tree | |
156 | * @hlist: link into hlist of rmap_items hanging off that stable_node | |
31dbd01f IE |
157 | */ |
158 | struct rmap_item { | |
6514d511 | 159 | struct rmap_item *rmap_list; |
db114b83 | 160 | struct anon_vma *anon_vma; /* when stable */ |
31dbd01f IE |
161 | struct mm_struct *mm; |
162 | unsigned long address; /* + low bits used for flags below */ | |
7b6ba2c7 | 163 | unsigned int oldchecksum; /* when unstable */ |
90bd6fd3 | 164 | #ifdef CONFIG_NUMA |
e850dcf5 | 165 | int nid; |
90bd6fd3 | 166 | #endif |
31dbd01f | 167 | union { |
7b6ba2c7 HD |
168 | struct rb_node node; /* when node of unstable tree */ |
169 | struct { /* when listed from stable tree */ | |
170 | struct stable_node *head; | |
171 | struct hlist_node hlist; | |
172 | }; | |
31dbd01f IE |
173 | }; |
174 | }; | |
175 | ||
176 | #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ | |
7b6ba2c7 HD |
177 | #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ |
178 | #define STABLE_FLAG 0x200 /* is listed from the stable tree */ | |
31dbd01f IE |
179 | |
180 | /* The stable and unstable tree heads */ | |
90bd6fd3 PH |
181 | static struct rb_root root_unstable_tree[MAX_NUMNODES]; |
182 | static struct rb_root root_stable_tree[MAX_NUMNODES]; | |
31dbd01f | 183 | |
4146d2d6 HD |
184 | /* Recently migrated nodes of stable tree, pending proper placement */ |
185 | static LIST_HEAD(migrate_nodes); | |
186 | ||
4ca3a69b SL |
187 | #define MM_SLOTS_HASH_BITS 10 |
188 | static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
31dbd01f IE |
189 | |
190 | static struct mm_slot ksm_mm_head = { | |
191 | .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list), | |
192 | }; | |
193 | static struct ksm_scan ksm_scan = { | |
194 | .mm_slot = &ksm_mm_head, | |
195 | }; | |
196 | ||
197 | static struct kmem_cache *rmap_item_cache; | |
7b6ba2c7 | 198 | static struct kmem_cache *stable_node_cache; |
31dbd01f IE |
199 | static struct kmem_cache *mm_slot_cache; |
200 | ||
201 | /* The number of nodes in the stable tree */ | |
b4028260 | 202 | static unsigned long ksm_pages_shared; |
31dbd01f | 203 | |
e178dfde | 204 | /* The number of page slots additionally sharing those nodes */ |
b4028260 | 205 | static unsigned long ksm_pages_sharing; |
31dbd01f | 206 | |
473b0ce4 HD |
207 | /* The number of nodes in the unstable tree */ |
208 | static unsigned long ksm_pages_unshared; | |
209 | ||
210 | /* The number of rmap_items in use: to calculate pages_volatile */ | |
211 | static unsigned long ksm_rmap_items; | |
212 | ||
31dbd01f | 213 | /* Number of pages ksmd should scan in one batch */ |
2c6854fd | 214 | static unsigned int ksm_thread_pages_to_scan = 100; |
31dbd01f IE |
215 | |
216 | /* Milliseconds ksmd should sleep between batches */ | |
2ffd8679 | 217 | static unsigned int ksm_thread_sleep_millisecs = 20; |
31dbd01f | 218 | |
e850dcf5 | 219 | #ifdef CONFIG_NUMA |
90bd6fd3 PH |
220 | /* Zeroed when merging across nodes is not allowed */ |
221 | static unsigned int ksm_merge_across_nodes = 1; | |
e850dcf5 HD |
222 | #else |
223 | #define ksm_merge_across_nodes 1U | |
224 | #endif | |
90bd6fd3 | 225 | |
31dbd01f IE |
226 | #define KSM_RUN_STOP 0 |
227 | #define KSM_RUN_MERGE 1 | |
228 | #define KSM_RUN_UNMERGE 2 | |
2c6854fd | 229 | static unsigned int ksm_run = KSM_RUN_STOP; |
31dbd01f IE |
230 | |
231 | static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); | |
232 | static DEFINE_MUTEX(ksm_thread_mutex); | |
233 | static DEFINE_SPINLOCK(ksm_mmlist_lock); | |
234 | ||
235 | #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\ | |
236 | sizeof(struct __struct), __alignof__(struct __struct),\ | |
237 | (__flags), NULL) | |
238 | ||
239 | static int __init ksm_slab_init(void) | |
240 | { | |
241 | rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0); | |
242 | if (!rmap_item_cache) | |
243 | goto out; | |
244 | ||
7b6ba2c7 HD |
245 | stable_node_cache = KSM_KMEM_CACHE(stable_node, 0); |
246 | if (!stable_node_cache) | |
247 | goto out_free1; | |
248 | ||
31dbd01f IE |
249 | mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0); |
250 | if (!mm_slot_cache) | |
7b6ba2c7 | 251 | goto out_free2; |
31dbd01f IE |
252 | |
253 | return 0; | |
254 | ||
7b6ba2c7 HD |
255 | out_free2: |
256 | kmem_cache_destroy(stable_node_cache); | |
257 | out_free1: | |
31dbd01f IE |
258 | kmem_cache_destroy(rmap_item_cache); |
259 | out: | |
260 | return -ENOMEM; | |
261 | } | |
262 | ||
263 | static void __init ksm_slab_free(void) | |
264 | { | |
265 | kmem_cache_destroy(mm_slot_cache); | |
7b6ba2c7 | 266 | kmem_cache_destroy(stable_node_cache); |
31dbd01f IE |
267 | kmem_cache_destroy(rmap_item_cache); |
268 | mm_slot_cache = NULL; | |
269 | } | |
270 | ||
271 | static inline struct rmap_item *alloc_rmap_item(void) | |
272 | { | |
473b0ce4 HD |
273 | struct rmap_item *rmap_item; |
274 | ||
275 | rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL); | |
276 | if (rmap_item) | |
277 | ksm_rmap_items++; | |
278 | return rmap_item; | |
31dbd01f IE |
279 | } |
280 | ||
281 | static inline void free_rmap_item(struct rmap_item *rmap_item) | |
282 | { | |
473b0ce4 | 283 | ksm_rmap_items--; |
31dbd01f IE |
284 | rmap_item->mm = NULL; /* debug safety */ |
285 | kmem_cache_free(rmap_item_cache, rmap_item); | |
286 | } | |
287 | ||
7b6ba2c7 HD |
288 | static inline struct stable_node *alloc_stable_node(void) |
289 | { | |
290 | return kmem_cache_alloc(stable_node_cache, GFP_KERNEL); | |
291 | } | |
292 | ||
293 | static inline void free_stable_node(struct stable_node *stable_node) | |
294 | { | |
295 | kmem_cache_free(stable_node_cache, stable_node); | |
296 | } | |
297 | ||
31dbd01f IE |
298 | static inline struct mm_slot *alloc_mm_slot(void) |
299 | { | |
300 | if (!mm_slot_cache) /* initialization failed */ | |
301 | return NULL; | |
302 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
303 | } | |
304 | ||
305 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
306 | { | |
307 | kmem_cache_free(mm_slot_cache, mm_slot); | |
308 | } | |
309 | ||
31dbd01f IE |
310 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) |
311 | { | |
31dbd01f | 312 | struct hlist_node *node; |
4ca3a69b SL |
313 | struct mm_slot *slot; |
314 | ||
315 | hash_for_each_possible(mm_slots_hash, slot, node, link, (unsigned long)mm) | |
316 | if (slot->mm == mm) | |
317 | return slot; | |
31dbd01f | 318 | |
31dbd01f IE |
319 | return NULL; |
320 | } | |
321 | ||
322 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
323 | struct mm_slot *mm_slot) | |
324 | { | |
31dbd01f | 325 | mm_slot->mm = mm; |
4ca3a69b | 326 | hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm); |
31dbd01f IE |
327 | } |
328 | ||
a913e182 HD |
329 | /* |
330 | * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's | |
331 | * page tables after it has passed through ksm_exit() - which, if necessary, | |
332 | * takes mmap_sem briefly to serialize against them. ksm_exit() does not set | |
333 | * a special flag: they can just back out as soon as mm_users goes to zero. | |
334 | * ksm_test_exit() is used throughout to make this test for exit: in some | |
335 | * places for correctness, in some places just to avoid unnecessary work. | |
336 | */ | |
337 | static inline bool ksm_test_exit(struct mm_struct *mm) | |
338 | { | |
339 | return atomic_read(&mm->mm_users) == 0; | |
340 | } | |
341 | ||
31dbd01f IE |
342 | /* |
343 | * We use break_ksm to break COW on a ksm page: it's a stripped down | |
344 | * | |
345 | * if (get_user_pages(current, mm, addr, 1, 1, 1, &page, NULL) == 1) | |
346 | * put_page(page); | |
347 | * | |
348 | * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma, | |
349 | * in case the application has unmapped and remapped mm,addr meanwhile. | |
350 | * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP | |
351 | * mmap of /dev/mem or /dev/kmem, where we would not want to touch it. | |
352 | */ | |
d952b791 | 353 | static int break_ksm(struct vm_area_struct *vma, unsigned long addr) |
31dbd01f IE |
354 | { |
355 | struct page *page; | |
d952b791 | 356 | int ret = 0; |
31dbd01f IE |
357 | |
358 | do { | |
359 | cond_resched(); | |
360 | page = follow_page(vma, addr, FOLL_GET); | |
22eccdd7 | 361 | if (IS_ERR_OR_NULL(page)) |
31dbd01f IE |
362 | break; |
363 | if (PageKsm(page)) | |
364 | ret = handle_mm_fault(vma->vm_mm, vma, addr, | |
365 | FAULT_FLAG_WRITE); | |
366 | else | |
367 | ret = VM_FAULT_WRITE; | |
368 | put_page(page); | |
d952b791 HD |
369 | } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM))); |
370 | /* | |
371 | * We must loop because handle_mm_fault() may back out if there's | |
372 | * any difficulty e.g. if pte accessed bit gets updated concurrently. | |
373 | * | |
374 | * VM_FAULT_WRITE is what we have been hoping for: it indicates that | |
375 | * COW has been broken, even if the vma does not permit VM_WRITE; | |
376 | * but note that a concurrent fault might break PageKsm for us. | |
377 | * | |
378 | * VM_FAULT_SIGBUS could occur if we race with truncation of the | |
379 | * backing file, which also invalidates anonymous pages: that's | |
380 | * okay, that truncation will have unmapped the PageKsm for us. | |
381 | * | |
382 | * VM_FAULT_OOM: at the time of writing (late July 2009), setting | |
383 | * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the | |
384 | * current task has TIF_MEMDIE set, and will be OOM killed on return | |
385 | * to user; and ksmd, having no mm, would never be chosen for that. | |
386 | * | |
387 | * But if the mm is in a limited mem_cgroup, then the fault may fail | |
388 | * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and | |
389 | * even ksmd can fail in this way - though it's usually breaking ksm | |
390 | * just to undo a merge it made a moment before, so unlikely to oom. | |
391 | * | |
392 | * That's a pity: we might therefore have more kernel pages allocated | |
393 | * than we're counting as nodes in the stable tree; but ksm_do_scan | |
394 | * will retry to break_cow on each pass, so should recover the page | |
395 | * in due course. The important thing is to not let VM_MERGEABLE | |
396 | * be cleared while any such pages might remain in the area. | |
397 | */ | |
398 | return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; | |
31dbd01f IE |
399 | } |
400 | ||
ef694222 BL |
401 | static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, |
402 | unsigned long addr) | |
403 | { | |
404 | struct vm_area_struct *vma; | |
405 | if (ksm_test_exit(mm)) | |
406 | return NULL; | |
407 | vma = find_vma(mm, addr); | |
408 | if (!vma || vma->vm_start > addr) | |
409 | return NULL; | |
410 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) | |
411 | return NULL; | |
412 | return vma; | |
413 | } | |
414 | ||
8dd3557a | 415 | static void break_cow(struct rmap_item *rmap_item) |
31dbd01f | 416 | { |
8dd3557a HD |
417 | struct mm_struct *mm = rmap_item->mm; |
418 | unsigned long addr = rmap_item->address; | |
31dbd01f IE |
419 | struct vm_area_struct *vma; |
420 | ||
4035c07a HD |
421 | /* |
422 | * It is not an accident that whenever we want to break COW | |
423 | * to undo, we also need to drop a reference to the anon_vma. | |
424 | */ | |
9e60109f | 425 | put_anon_vma(rmap_item->anon_vma); |
4035c07a | 426 | |
81464e30 | 427 | down_read(&mm->mmap_sem); |
ef694222 BL |
428 | vma = find_mergeable_vma(mm, addr); |
429 | if (vma) | |
430 | break_ksm(vma, addr); | |
31dbd01f IE |
431 | up_read(&mm->mmap_sem); |
432 | } | |
433 | ||
29ad768c AA |
434 | static struct page *page_trans_compound_anon(struct page *page) |
435 | { | |
436 | if (PageTransCompound(page)) { | |
22e5c47e | 437 | struct page *head = compound_trans_head(page); |
29ad768c | 438 | /* |
22e5c47e AA |
439 | * head may actually be splitted and freed from under |
440 | * us but it's ok here. | |
29ad768c | 441 | */ |
29ad768c AA |
442 | if (PageAnon(head)) |
443 | return head; | |
444 | } | |
445 | return NULL; | |
446 | } | |
447 | ||
31dbd01f IE |
448 | static struct page *get_mergeable_page(struct rmap_item *rmap_item) |
449 | { | |
450 | struct mm_struct *mm = rmap_item->mm; | |
451 | unsigned long addr = rmap_item->address; | |
452 | struct vm_area_struct *vma; | |
453 | struct page *page; | |
454 | ||
455 | down_read(&mm->mmap_sem); | |
ef694222 BL |
456 | vma = find_mergeable_vma(mm, addr); |
457 | if (!vma) | |
31dbd01f IE |
458 | goto out; |
459 | ||
460 | page = follow_page(vma, addr, FOLL_GET); | |
22eccdd7 | 461 | if (IS_ERR_OR_NULL(page)) |
31dbd01f | 462 | goto out; |
29ad768c | 463 | if (PageAnon(page) || page_trans_compound_anon(page)) { |
31dbd01f IE |
464 | flush_anon_page(vma, page, addr); |
465 | flush_dcache_page(page); | |
466 | } else { | |
467 | put_page(page); | |
468 | out: page = NULL; | |
469 | } | |
470 | up_read(&mm->mmap_sem); | |
471 | return page; | |
472 | } | |
473 | ||
90bd6fd3 PH |
474 | /* |
475 | * This helper is used for getting right index into array of tree roots. | |
476 | * When merge_across_nodes knob is set to 1, there are only two rb-trees for | |
477 | * stable and unstable pages from all nodes with roots in index 0. Otherwise, | |
478 | * every node has its own stable and unstable tree. | |
479 | */ | |
480 | static inline int get_kpfn_nid(unsigned long kpfn) | |
481 | { | |
e850dcf5 | 482 | return ksm_merge_across_nodes ? 0 : pfn_to_nid(kpfn); |
90bd6fd3 PH |
483 | } |
484 | ||
4035c07a HD |
485 | static void remove_node_from_stable_tree(struct stable_node *stable_node) |
486 | { | |
487 | struct rmap_item *rmap_item; | |
488 | struct hlist_node *hlist; | |
489 | ||
490 | hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) { | |
491 | if (rmap_item->hlist.next) | |
492 | ksm_pages_sharing--; | |
493 | else | |
494 | ksm_pages_shared--; | |
9e60109f | 495 | put_anon_vma(rmap_item->anon_vma); |
4035c07a HD |
496 | rmap_item->address &= PAGE_MASK; |
497 | cond_resched(); | |
498 | } | |
499 | ||
4146d2d6 HD |
500 | if (stable_node->head == &migrate_nodes) |
501 | list_del(&stable_node->list); | |
502 | else | |
503 | rb_erase(&stable_node->node, | |
504 | &root_stable_tree[NUMA(stable_node->nid)]); | |
4035c07a HD |
505 | free_stable_node(stable_node); |
506 | } | |
507 | ||
508 | /* | |
509 | * get_ksm_page: checks if the page indicated by the stable node | |
510 | * is still its ksm page, despite having held no reference to it. | |
511 | * In which case we can trust the content of the page, and it | |
512 | * returns the gotten page; but if the page has now been zapped, | |
513 | * remove the stale node from the stable tree and return NULL. | |
c8d6553b | 514 | * But beware, the stable node's page might be being migrated. |
4035c07a HD |
515 | * |
516 | * You would expect the stable_node to hold a reference to the ksm page. | |
517 | * But if it increments the page's count, swapping out has to wait for | |
518 | * ksmd to come around again before it can free the page, which may take | |
519 | * seconds or even minutes: much too unresponsive. So instead we use a | |
520 | * "keyhole reference": access to the ksm page from the stable node peeps | |
521 | * out through its keyhole to see if that page still holds the right key, | |
522 | * pointing back to this stable node. This relies on freeing a PageAnon | |
523 | * page to reset its page->mapping to NULL, and relies on no other use of | |
524 | * a page to put something that might look like our key in page->mapping. | |
4035c07a HD |
525 | * is on its way to being freed; but it is an anomaly to bear in mind. |
526 | */ | |
8aafa6a4 | 527 | static struct page *get_ksm_page(struct stable_node *stable_node, bool locked) |
4035c07a HD |
528 | { |
529 | struct page *page; | |
530 | void *expected_mapping; | |
c8d6553b | 531 | unsigned long kpfn; |
4035c07a | 532 | |
4035c07a HD |
533 | expected_mapping = (void *)stable_node + |
534 | (PAGE_MAPPING_ANON | PAGE_MAPPING_KSM); | |
c8d6553b HD |
535 | again: |
536 | kpfn = ACCESS_ONCE(stable_node->kpfn); | |
537 | page = pfn_to_page(kpfn); | |
538 | ||
539 | /* | |
540 | * page is computed from kpfn, so on most architectures reading | |
541 | * page->mapping is naturally ordered after reading node->kpfn, | |
542 | * but on Alpha we need to be more careful. | |
543 | */ | |
544 | smp_read_barrier_depends(); | |
545 | if (ACCESS_ONCE(page->mapping) != expected_mapping) | |
4035c07a | 546 | goto stale; |
c8d6553b HD |
547 | |
548 | /* | |
549 | * We cannot do anything with the page while its refcount is 0. | |
550 | * Usually 0 means free, or tail of a higher-order page: in which | |
551 | * case this node is no longer referenced, and should be freed; | |
552 | * however, it might mean that the page is under page_freeze_refs(). | |
553 | * The __remove_mapping() case is easy, again the node is now stale; | |
554 | * but if page is swapcache in migrate_page_move_mapping(), it might | |
555 | * still be our page, in which case it's essential to keep the node. | |
556 | */ | |
557 | while (!get_page_unless_zero(page)) { | |
558 | /* | |
559 | * Another check for page->mapping != expected_mapping would | |
560 | * work here too. We have chosen the !PageSwapCache test to | |
561 | * optimize the common case, when the page is or is about to | |
562 | * be freed: PageSwapCache is cleared (under spin_lock_irq) | |
563 | * in the freeze_refs section of __remove_mapping(); but Anon | |
564 | * page->mapping reset to NULL later, in free_pages_prepare(). | |
565 | */ | |
566 | if (!PageSwapCache(page)) | |
567 | goto stale; | |
568 | cpu_relax(); | |
569 | } | |
570 | ||
571 | if (ACCESS_ONCE(page->mapping) != expected_mapping) { | |
4035c07a HD |
572 | put_page(page); |
573 | goto stale; | |
574 | } | |
c8d6553b | 575 | |
8aafa6a4 HD |
576 | if (locked) { |
577 | lock_page(page); | |
c8d6553b | 578 | if (ACCESS_ONCE(page->mapping) != expected_mapping) { |
8aafa6a4 HD |
579 | unlock_page(page); |
580 | put_page(page); | |
581 | goto stale; | |
582 | } | |
583 | } | |
4035c07a | 584 | return page; |
c8d6553b | 585 | |
4035c07a | 586 | stale: |
c8d6553b HD |
587 | /* |
588 | * We come here from above when page->mapping or !PageSwapCache | |
589 | * suggests that the node is stale; but it might be under migration. | |
590 | * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(), | |
591 | * before checking whether node->kpfn has been changed. | |
592 | */ | |
593 | smp_rmb(); | |
594 | if (ACCESS_ONCE(stable_node->kpfn) != kpfn) | |
595 | goto again; | |
4035c07a HD |
596 | remove_node_from_stable_tree(stable_node); |
597 | return NULL; | |
598 | } | |
599 | ||
31dbd01f IE |
600 | /* |
601 | * Removing rmap_item from stable or unstable tree. | |
602 | * This function will clean the information from the stable/unstable tree. | |
603 | */ | |
604 | static void remove_rmap_item_from_tree(struct rmap_item *rmap_item) | |
605 | { | |
7b6ba2c7 HD |
606 | if (rmap_item->address & STABLE_FLAG) { |
607 | struct stable_node *stable_node; | |
5ad64688 | 608 | struct page *page; |
31dbd01f | 609 | |
7b6ba2c7 | 610 | stable_node = rmap_item->head; |
8aafa6a4 | 611 | page = get_ksm_page(stable_node, true); |
4035c07a HD |
612 | if (!page) |
613 | goto out; | |
5ad64688 | 614 | |
7b6ba2c7 | 615 | hlist_del(&rmap_item->hlist); |
4035c07a HD |
616 | unlock_page(page); |
617 | put_page(page); | |
08beca44 | 618 | |
4035c07a HD |
619 | if (stable_node->hlist.first) |
620 | ksm_pages_sharing--; | |
621 | else | |
7b6ba2c7 | 622 | ksm_pages_shared--; |
31dbd01f | 623 | |
9e60109f | 624 | put_anon_vma(rmap_item->anon_vma); |
93d17715 | 625 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 626 | |
7b6ba2c7 | 627 | } else if (rmap_item->address & UNSTABLE_FLAG) { |
31dbd01f IE |
628 | unsigned char age; |
629 | /* | |
9ba69294 | 630 | * Usually ksmd can and must skip the rb_erase, because |
31dbd01f | 631 | * root_unstable_tree was already reset to RB_ROOT. |
9ba69294 HD |
632 | * But be careful when an mm is exiting: do the rb_erase |
633 | * if this rmap_item was inserted by this scan, rather | |
634 | * than left over from before. | |
31dbd01f IE |
635 | */ |
636 | age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); | |
cd551f97 | 637 | BUG_ON(age > 1); |
31dbd01f | 638 | if (!age) |
90bd6fd3 | 639 | rb_erase(&rmap_item->node, |
e850dcf5 | 640 | &root_unstable_tree[NUMA(rmap_item->nid)]); |
473b0ce4 | 641 | ksm_pages_unshared--; |
93d17715 | 642 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 643 | } |
4035c07a | 644 | out: |
31dbd01f IE |
645 | cond_resched(); /* we're called from many long loops */ |
646 | } | |
647 | ||
31dbd01f | 648 | static void remove_trailing_rmap_items(struct mm_slot *mm_slot, |
6514d511 | 649 | struct rmap_item **rmap_list) |
31dbd01f | 650 | { |
6514d511 HD |
651 | while (*rmap_list) { |
652 | struct rmap_item *rmap_item = *rmap_list; | |
653 | *rmap_list = rmap_item->rmap_list; | |
31dbd01f | 654 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
655 | free_rmap_item(rmap_item); |
656 | } | |
657 | } | |
658 | ||
659 | /* | |
e850dcf5 | 660 | * Though it's very tempting to unmerge rmap_items from stable tree rather |
31dbd01f IE |
661 | * than check every pte of a given vma, the locking doesn't quite work for |
662 | * that - an rmap_item is assigned to the stable tree after inserting ksm | |
663 | * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing | |
664 | * rmap_items from parent to child at fork time (so as not to waste time | |
665 | * if exit comes before the next scan reaches it). | |
81464e30 HD |
666 | * |
667 | * Similarly, although we'd like to remove rmap_items (so updating counts | |
668 | * and freeing memory) when unmerging an area, it's easier to leave that | |
669 | * to the next pass of ksmd - consider, for example, how ksmd might be | |
670 | * in cmp_and_merge_page on one of the rmap_items we would be removing. | |
31dbd01f | 671 | */ |
d952b791 HD |
672 | static int unmerge_ksm_pages(struct vm_area_struct *vma, |
673 | unsigned long start, unsigned long end) | |
31dbd01f IE |
674 | { |
675 | unsigned long addr; | |
d952b791 | 676 | int err = 0; |
31dbd01f | 677 | |
d952b791 | 678 | for (addr = start; addr < end && !err; addr += PAGE_SIZE) { |
9ba69294 HD |
679 | if (ksm_test_exit(vma->vm_mm)) |
680 | break; | |
d952b791 HD |
681 | if (signal_pending(current)) |
682 | err = -ERESTARTSYS; | |
683 | else | |
684 | err = break_ksm(vma, addr); | |
685 | } | |
686 | return err; | |
31dbd01f IE |
687 | } |
688 | ||
2ffd8679 HD |
689 | #ifdef CONFIG_SYSFS |
690 | /* | |
691 | * Only called through the sysfs control interface: | |
692 | */ | |
cbf86cfe HD |
693 | static int remove_stable_node(struct stable_node *stable_node) |
694 | { | |
695 | struct page *page; | |
696 | int err; | |
697 | ||
698 | page = get_ksm_page(stable_node, true); | |
699 | if (!page) { | |
700 | /* | |
701 | * get_ksm_page did remove_node_from_stable_tree itself. | |
702 | */ | |
703 | return 0; | |
704 | } | |
705 | ||
706 | if (WARN_ON_ONCE(page_mapped(page))) | |
707 | err = -EBUSY; | |
708 | else { | |
709 | /* | |
710 | * This page might be in a pagevec waiting to be freed, | |
711 | * or it might be PageSwapCache (perhaps under writeback), | |
712 | * or it might have been removed from swapcache a moment ago. | |
713 | */ | |
714 | set_page_stable_node(page, NULL); | |
715 | remove_node_from_stable_tree(stable_node); | |
716 | err = 0; | |
717 | } | |
718 | ||
719 | unlock_page(page); | |
720 | put_page(page); | |
721 | return err; | |
722 | } | |
723 | ||
724 | static int remove_all_stable_nodes(void) | |
725 | { | |
726 | struct stable_node *stable_node; | |
4146d2d6 | 727 | struct list_head *this, *next; |
cbf86cfe HD |
728 | int nid; |
729 | int err = 0; | |
730 | ||
731 | for (nid = 0; nid < nr_node_ids; nid++) { | |
732 | while (root_stable_tree[nid].rb_node) { | |
733 | stable_node = rb_entry(root_stable_tree[nid].rb_node, | |
734 | struct stable_node, node); | |
735 | if (remove_stable_node(stable_node)) { | |
736 | err = -EBUSY; | |
737 | break; /* proceed to next nid */ | |
738 | } | |
739 | cond_resched(); | |
740 | } | |
741 | } | |
4146d2d6 HD |
742 | list_for_each_safe(this, next, &migrate_nodes) { |
743 | stable_node = list_entry(this, struct stable_node, list); | |
744 | if (remove_stable_node(stable_node)) | |
745 | err = -EBUSY; | |
746 | cond_resched(); | |
747 | } | |
cbf86cfe HD |
748 | return err; |
749 | } | |
750 | ||
d952b791 | 751 | static int unmerge_and_remove_all_rmap_items(void) |
31dbd01f IE |
752 | { |
753 | struct mm_slot *mm_slot; | |
754 | struct mm_struct *mm; | |
755 | struct vm_area_struct *vma; | |
d952b791 HD |
756 | int err = 0; |
757 | ||
758 | spin_lock(&ksm_mmlist_lock); | |
9ba69294 | 759 | ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next, |
d952b791 HD |
760 | struct mm_slot, mm_list); |
761 | spin_unlock(&ksm_mmlist_lock); | |
31dbd01f | 762 | |
9ba69294 HD |
763 | for (mm_slot = ksm_scan.mm_slot; |
764 | mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) { | |
31dbd01f IE |
765 | mm = mm_slot->mm; |
766 | down_read(&mm->mmap_sem); | |
767 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
9ba69294 HD |
768 | if (ksm_test_exit(mm)) |
769 | break; | |
31dbd01f IE |
770 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
771 | continue; | |
d952b791 HD |
772 | err = unmerge_ksm_pages(vma, |
773 | vma->vm_start, vma->vm_end); | |
9ba69294 HD |
774 | if (err) |
775 | goto error; | |
31dbd01f | 776 | } |
9ba69294 | 777 | |
6514d511 | 778 | remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list); |
d952b791 HD |
779 | |
780 | spin_lock(&ksm_mmlist_lock); | |
9ba69294 | 781 | ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next, |
d952b791 | 782 | struct mm_slot, mm_list); |
9ba69294 | 783 | if (ksm_test_exit(mm)) { |
4ca3a69b | 784 | hash_del(&mm_slot->link); |
9ba69294 HD |
785 | list_del(&mm_slot->mm_list); |
786 | spin_unlock(&ksm_mmlist_lock); | |
787 | ||
788 | free_mm_slot(mm_slot); | |
789 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
790 | up_read(&mm->mmap_sem); | |
791 | mmdrop(mm); | |
792 | } else { | |
793 | spin_unlock(&ksm_mmlist_lock); | |
794 | up_read(&mm->mmap_sem); | |
795 | } | |
31dbd01f IE |
796 | } |
797 | ||
cbf86cfe HD |
798 | /* Clean up stable nodes, but don't worry if some are still busy */ |
799 | remove_all_stable_nodes(); | |
d952b791 | 800 | ksm_scan.seqnr = 0; |
9ba69294 HD |
801 | return 0; |
802 | ||
803 | error: | |
804 | up_read(&mm->mmap_sem); | |
31dbd01f | 805 | spin_lock(&ksm_mmlist_lock); |
d952b791 | 806 | ksm_scan.mm_slot = &ksm_mm_head; |
31dbd01f | 807 | spin_unlock(&ksm_mmlist_lock); |
d952b791 | 808 | return err; |
31dbd01f | 809 | } |
2ffd8679 | 810 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 811 | |
31dbd01f IE |
812 | static u32 calc_checksum(struct page *page) |
813 | { | |
814 | u32 checksum; | |
9b04c5fe | 815 | void *addr = kmap_atomic(page); |
31dbd01f | 816 | checksum = jhash2(addr, PAGE_SIZE / 4, 17); |
9b04c5fe | 817 | kunmap_atomic(addr); |
31dbd01f IE |
818 | return checksum; |
819 | } | |
820 | ||
821 | static int memcmp_pages(struct page *page1, struct page *page2) | |
822 | { | |
823 | char *addr1, *addr2; | |
824 | int ret; | |
825 | ||
9b04c5fe CW |
826 | addr1 = kmap_atomic(page1); |
827 | addr2 = kmap_atomic(page2); | |
31dbd01f | 828 | ret = memcmp(addr1, addr2, PAGE_SIZE); |
9b04c5fe CW |
829 | kunmap_atomic(addr2); |
830 | kunmap_atomic(addr1); | |
31dbd01f IE |
831 | return ret; |
832 | } | |
833 | ||
834 | static inline int pages_identical(struct page *page1, struct page *page2) | |
835 | { | |
836 | return !memcmp_pages(page1, page2); | |
837 | } | |
838 | ||
839 | static int write_protect_page(struct vm_area_struct *vma, struct page *page, | |
840 | pte_t *orig_pte) | |
841 | { | |
842 | struct mm_struct *mm = vma->vm_mm; | |
843 | unsigned long addr; | |
844 | pte_t *ptep; | |
845 | spinlock_t *ptl; | |
846 | int swapped; | |
847 | int err = -EFAULT; | |
6bdb913f HE |
848 | unsigned long mmun_start; /* For mmu_notifiers */ |
849 | unsigned long mmun_end; /* For mmu_notifiers */ | |
31dbd01f IE |
850 | |
851 | addr = page_address_in_vma(page, vma); | |
852 | if (addr == -EFAULT) | |
853 | goto out; | |
854 | ||
29ad768c | 855 | BUG_ON(PageTransCompound(page)); |
6bdb913f HE |
856 | |
857 | mmun_start = addr; | |
858 | mmun_end = addr + PAGE_SIZE; | |
859 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
860 | ||
31dbd01f IE |
861 | ptep = page_check_address(page, mm, addr, &ptl, 0); |
862 | if (!ptep) | |
6bdb913f | 863 | goto out_mn; |
31dbd01f | 864 | |
4e31635c | 865 | if (pte_write(*ptep) || pte_dirty(*ptep)) { |
31dbd01f IE |
866 | pte_t entry; |
867 | ||
868 | swapped = PageSwapCache(page); | |
869 | flush_cache_page(vma, addr, page_to_pfn(page)); | |
870 | /* | |
25985edc | 871 | * Ok this is tricky, when get_user_pages_fast() run it doesn't |
31dbd01f IE |
872 | * take any lock, therefore the check that we are going to make |
873 | * with the pagecount against the mapcount is racey and | |
874 | * O_DIRECT can happen right after the check. | |
875 | * So we clear the pte and flush the tlb before the check | |
876 | * this assure us that no O_DIRECT can happen after the check | |
877 | * or in the middle of the check. | |
878 | */ | |
879 | entry = ptep_clear_flush(vma, addr, ptep); | |
880 | /* | |
881 | * Check that no O_DIRECT or similar I/O is in progress on the | |
882 | * page | |
883 | */ | |
31e855ea | 884 | if (page_mapcount(page) + 1 + swapped != page_count(page)) { |
cb532375 | 885 | set_pte_at(mm, addr, ptep, entry); |
31dbd01f IE |
886 | goto out_unlock; |
887 | } | |
4e31635c HD |
888 | if (pte_dirty(entry)) |
889 | set_page_dirty(page); | |
890 | entry = pte_mkclean(pte_wrprotect(entry)); | |
31dbd01f IE |
891 | set_pte_at_notify(mm, addr, ptep, entry); |
892 | } | |
893 | *orig_pte = *ptep; | |
894 | err = 0; | |
895 | ||
896 | out_unlock: | |
897 | pte_unmap_unlock(ptep, ptl); | |
6bdb913f HE |
898 | out_mn: |
899 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
31dbd01f IE |
900 | out: |
901 | return err; | |
902 | } | |
903 | ||
904 | /** | |
905 | * replace_page - replace page in vma by new ksm page | |
8dd3557a HD |
906 | * @vma: vma that holds the pte pointing to page |
907 | * @page: the page we are replacing by kpage | |
908 | * @kpage: the ksm page we replace page by | |
31dbd01f IE |
909 | * @orig_pte: the original value of the pte |
910 | * | |
911 | * Returns 0 on success, -EFAULT on failure. | |
912 | */ | |
8dd3557a HD |
913 | static int replace_page(struct vm_area_struct *vma, struct page *page, |
914 | struct page *kpage, pte_t orig_pte) | |
31dbd01f IE |
915 | { |
916 | struct mm_struct *mm = vma->vm_mm; | |
31dbd01f IE |
917 | pmd_t *pmd; |
918 | pte_t *ptep; | |
919 | spinlock_t *ptl; | |
920 | unsigned long addr; | |
31dbd01f | 921 | int err = -EFAULT; |
6bdb913f HE |
922 | unsigned long mmun_start; /* For mmu_notifiers */ |
923 | unsigned long mmun_end; /* For mmu_notifiers */ | |
31dbd01f | 924 | |
8dd3557a | 925 | addr = page_address_in_vma(page, vma); |
31dbd01f IE |
926 | if (addr == -EFAULT) |
927 | goto out; | |
928 | ||
6219049a BL |
929 | pmd = mm_find_pmd(mm, addr); |
930 | if (!pmd) | |
31dbd01f | 931 | goto out; |
29ad768c | 932 | BUG_ON(pmd_trans_huge(*pmd)); |
31dbd01f | 933 | |
6bdb913f HE |
934 | mmun_start = addr; |
935 | mmun_end = addr + PAGE_SIZE; | |
936 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
937 | ||
31dbd01f IE |
938 | ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); |
939 | if (!pte_same(*ptep, orig_pte)) { | |
940 | pte_unmap_unlock(ptep, ptl); | |
6bdb913f | 941 | goto out_mn; |
31dbd01f IE |
942 | } |
943 | ||
8dd3557a | 944 | get_page(kpage); |
5ad64688 | 945 | page_add_anon_rmap(kpage, vma, addr); |
31dbd01f IE |
946 | |
947 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
948 | ptep_clear_flush(vma, addr, ptep); | |
8dd3557a | 949 | set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot)); |
31dbd01f | 950 | |
8dd3557a | 951 | page_remove_rmap(page); |
ae52a2ad HD |
952 | if (!page_mapped(page)) |
953 | try_to_free_swap(page); | |
8dd3557a | 954 | put_page(page); |
31dbd01f IE |
955 | |
956 | pte_unmap_unlock(ptep, ptl); | |
957 | err = 0; | |
6bdb913f HE |
958 | out_mn: |
959 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
31dbd01f IE |
960 | out: |
961 | return err; | |
962 | } | |
963 | ||
29ad768c AA |
964 | static int page_trans_compound_anon_split(struct page *page) |
965 | { | |
966 | int ret = 0; | |
967 | struct page *transhuge_head = page_trans_compound_anon(page); | |
968 | if (transhuge_head) { | |
969 | /* Get the reference on the head to split it. */ | |
970 | if (get_page_unless_zero(transhuge_head)) { | |
971 | /* | |
972 | * Recheck we got the reference while the head | |
973 | * was still anonymous. | |
974 | */ | |
975 | if (PageAnon(transhuge_head)) | |
976 | ret = split_huge_page(transhuge_head); | |
977 | else | |
978 | /* | |
979 | * Retry later if split_huge_page run | |
980 | * from under us. | |
981 | */ | |
982 | ret = 1; | |
983 | put_page(transhuge_head); | |
984 | } else | |
985 | /* Retry later if split_huge_page run from under us. */ | |
986 | ret = 1; | |
987 | } | |
988 | return ret; | |
989 | } | |
990 | ||
31dbd01f IE |
991 | /* |
992 | * try_to_merge_one_page - take two pages and merge them into one | |
8dd3557a HD |
993 | * @vma: the vma that holds the pte pointing to page |
994 | * @page: the PageAnon page that we want to replace with kpage | |
80e14822 HD |
995 | * @kpage: the PageKsm page that we want to map instead of page, |
996 | * or NULL the first time when we want to use page as kpage. | |
31dbd01f IE |
997 | * |
998 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
999 | */ | |
1000 | static int try_to_merge_one_page(struct vm_area_struct *vma, | |
8dd3557a | 1001 | struct page *page, struct page *kpage) |
31dbd01f IE |
1002 | { |
1003 | pte_t orig_pte = __pte(0); | |
1004 | int err = -EFAULT; | |
1005 | ||
db114b83 HD |
1006 | if (page == kpage) /* ksm page forked */ |
1007 | return 0; | |
1008 | ||
31dbd01f IE |
1009 | if (!(vma->vm_flags & VM_MERGEABLE)) |
1010 | goto out; | |
29ad768c AA |
1011 | if (PageTransCompound(page) && page_trans_compound_anon_split(page)) |
1012 | goto out; | |
1013 | BUG_ON(PageTransCompound(page)); | |
8dd3557a | 1014 | if (!PageAnon(page)) |
31dbd01f IE |
1015 | goto out; |
1016 | ||
31dbd01f IE |
1017 | /* |
1018 | * We need the page lock to read a stable PageSwapCache in | |
1019 | * write_protect_page(). We use trylock_page() instead of | |
1020 | * lock_page() because we don't want to wait here - we | |
1021 | * prefer to continue scanning and merging different pages, | |
1022 | * then come back to this page when it is unlocked. | |
1023 | */ | |
8dd3557a | 1024 | if (!trylock_page(page)) |
31e855ea | 1025 | goto out; |
31dbd01f IE |
1026 | /* |
1027 | * If this anonymous page is mapped only here, its pte may need | |
1028 | * to be write-protected. If it's mapped elsewhere, all of its | |
1029 | * ptes are necessarily already write-protected. But in either | |
1030 | * case, we need to lock and check page_count is not raised. | |
1031 | */ | |
80e14822 HD |
1032 | if (write_protect_page(vma, page, &orig_pte) == 0) { |
1033 | if (!kpage) { | |
1034 | /* | |
1035 | * While we hold page lock, upgrade page from | |
1036 | * PageAnon+anon_vma to PageKsm+NULL stable_node: | |
1037 | * stable_tree_insert() will update stable_node. | |
1038 | */ | |
1039 | set_page_stable_node(page, NULL); | |
1040 | mark_page_accessed(page); | |
1041 | err = 0; | |
1042 | } else if (pages_identical(page, kpage)) | |
1043 | err = replace_page(vma, page, kpage, orig_pte); | |
1044 | } | |
31dbd01f | 1045 | |
80e14822 | 1046 | if ((vma->vm_flags & VM_LOCKED) && kpage && !err) { |
73848b46 | 1047 | munlock_vma_page(page); |
5ad64688 HD |
1048 | if (!PageMlocked(kpage)) { |
1049 | unlock_page(page); | |
5ad64688 HD |
1050 | lock_page(kpage); |
1051 | mlock_vma_page(kpage); | |
1052 | page = kpage; /* for final unlock */ | |
1053 | } | |
1054 | } | |
73848b46 | 1055 | |
8dd3557a | 1056 | unlock_page(page); |
31dbd01f IE |
1057 | out: |
1058 | return err; | |
1059 | } | |
1060 | ||
81464e30 HD |
1061 | /* |
1062 | * try_to_merge_with_ksm_page - like try_to_merge_two_pages, | |
1063 | * but no new kernel page is allocated: kpage must already be a ksm page. | |
8dd3557a HD |
1064 | * |
1065 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
81464e30 | 1066 | */ |
8dd3557a HD |
1067 | static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item, |
1068 | struct page *page, struct page *kpage) | |
81464e30 | 1069 | { |
8dd3557a | 1070 | struct mm_struct *mm = rmap_item->mm; |
81464e30 HD |
1071 | struct vm_area_struct *vma; |
1072 | int err = -EFAULT; | |
1073 | ||
8dd3557a HD |
1074 | down_read(&mm->mmap_sem); |
1075 | if (ksm_test_exit(mm)) | |
9ba69294 | 1076 | goto out; |
8dd3557a HD |
1077 | vma = find_vma(mm, rmap_item->address); |
1078 | if (!vma || vma->vm_start > rmap_item->address) | |
81464e30 HD |
1079 | goto out; |
1080 | ||
8dd3557a | 1081 | err = try_to_merge_one_page(vma, page, kpage); |
db114b83 HD |
1082 | if (err) |
1083 | goto out; | |
1084 | ||
1085 | /* Must get reference to anon_vma while still holding mmap_sem */ | |
9e60109f PZ |
1086 | rmap_item->anon_vma = vma->anon_vma; |
1087 | get_anon_vma(vma->anon_vma); | |
81464e30 | 1088 | out: |
8dd3557a | 1089 | up_read(&mm->mmap_sem); |
81464e30 HD |
1090 | return err; |
1091 | } | |
1092 | ||
31dbd01f IE |
1093 | /* |
1094 | * try_to_merge_two_pages - take two identical pages and prepare them | |
1095 | * to be merged into one page. | |
1096 | * | |
8dd3557a HD |
1097 | * This function returns the kpage if we successfully merged two identical |
1098 | * pages into one ksm page, NULL otherwise. | |
31dbd01f | 1099 | * |
80e14822 | 1100 | * Note that this function upgrades page to ksm page: if one of the pages |
31dbd01f IE |
1101 | * is already a ksm page, try_to_merge_with_ksm_page should be used. |
1102 | */ | |
8dd3557a HD |
1103 | static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item, |
1104 | struct page *page, | |
1105 | struct rmap_item *tree_rmap_item, | |
1106 | struct page *tree_page) | |
31dbd01f | 1107 | { |
80e14822 | 1108 | int err; |
31dbd01f | 1109 | |
80e14822 | 1110 | err = try_to_merge_with_ksm_page(rmap_item, page, NULL); |
31dbd01f | 1111 | if (!err) { |
8dd3557a | 1112 | err = try_to_merge_with_ksm_page(tree_rmap_item, |
80e14822 | 1113 | tree_page, page); |
31dbd01f | 1114 | /* |
81464e30 HD |
1115 | * If that fails, we have a ksm page with only one pte |
1116 | * pointing to it: so break it. | |
31dbd01f | 1117 | */ |
4035c07a | 1118 | if (err) |
8dd3557a | 1119 | break_cow(rmap_item); |
31dbd01f | 1120 | } |
80e14822 | 1121 | return err ? NULL : page; |
31dbd01f IE |
1122 | } |
1123 | ||
31dbd01f | 1124 | /* |
8dd3557a | 1125 | * stable_tree_search - search for page inside the stable tree |
31dbd01f IE |
1126 | * |
1127 | * This function checks if there is a page inside the stable tree | |
1128 | * with identical content to the page that we are scanning right now. | |
1129 | * | |
7b6ba2c7 | 1130 | * This function returns the stable tree node of identical content if found, |
31dbd01f IE |
1131 | * NULL otherwise. |
1132 | */ | |
62b61f61 | 1133 | static struct page *stable_tree_search(struct page *page) |
31dbd01f | 1134 | { |
90bd6fd3 | 1135 | int nid; |
4146d2d6 HD |
1136 | struct rb_node **new; |
1137 | struct rb_node *parent; | |
1138 | struct stable_node *stable_node; | |
1139 | struct stable_node *page_node; | |
31dbd01f | 1140 | |
4146d2d6 HD |
1141 | page_node = page_stable_node(page); |
1142 | if (page_node && page_node->head != &migrate_nodes) { | |
1143 | /* ksm page forked */ | |
08beca44 | 1144 | get_page(page); |
62b61f61 | 1145 | return page; |
08beca44 HD |
1146 | } |
1147 | ||
90bd6fd3 | 1148 | nid = get_kpfn_nid(page_to_pfn(page)); |
4146d2d6 HD |
1149 | again: |
1150 | new = &root_stable_tree[nid].rb_node; | |
1151 | parent = NULL; | |
90bd6fd3 | 1152 | |
4146d2d6 | 1153 | while (*new) { |
4035c07a | 1154 | struct page *tree_page; |
31dbd01f IE |
1155 | int ret; |
1156 | ||
08beca44 | 1157 | cond_resched(); |
4146d2d6 | 1158 | stable_node = rb_entry(*new, struct stable_node, node); |
8aafa6a4 | 1159 | tree_page = get_ksm_page(stable_node, false); |
4035c07a HD |
1160 | if (!tree_page) |
1161 | return NULL; | |
31dbd01f | 1162 | |
4035c07a | 1163 | ret = memcmp_pages(page, tree_page); |
c8d6553b | 1164 | put_page(tree_page); |
31dbd01f | 1165 | |
4146d2d6 | 1166 | parent = *new; |
c8d6553b | 1167 | if (ret < 0) |
4146d2d6 | 1168 | new = &parent->rb_left; |
c8d6553b | 1169 | else if (ret > 0) |
4146d2d6 | 1170 | new = &parent->rb_right; |
c8d6553b HD |
1171 | else { |
1172 | /* | |
1173 | * Lock and unlock the stable_node's page (which | |
1174 | * might already have been migrated) so that page | |
1175 | * migration is sure to notice its raised count. | |
1176 | * It would be more elegant to return stable_node | |
1177 | * than kpage, but that involves more changes. | |
1178 | */ | |
1179 | tree_page = get_ksm_page(stable_node, true); | |
4146d2d6 | 1180 | if (tree_page) { |
c8d6553b | 1181 | unlock_page(tree_page); |
4146d2d6 HD |
1182 | if (get_kpfn_nid(stable_node->kpfn) != |
1183 | NUMA(stable_node->nid)) { | |
1184 | put_page(tree_page); | |
1185 | goto replace; | |
1186 | } | |
1187 | return tree_page; | |
1188 | } | |
1189 | /* | |
1190 | * There is now a place for page_node, but the tree may | |
1191 | * have been rebalanced, so re-evaluate parent and new. | |
1192 | */ | |
1193 | if (page_node) | |
1194 | goto again; | |
1195 | return NULL; | |
c8d6553b | 1196 | } |
31dbd01f IE |
1197 | } |
1198 | ||
4146d2d6 HD |
1199 | if (!page_node) |
1200 | return NULL; | |
1201 | ||
1202 | list_del(&page_node->list); | |
1203 | DO_NUMA(page_node->nid = nid); | |
1204 | rb_link_node(&page_node->node, parent, new); | |
1205 | rb_insert_color(&page_node->node, &root_stable_tree[nid]); | |
1206 | get_page(page); | |
1207 | return page; | |
1208 | ||
1209 | replace: | |
1210 | if (page_node) { | |
1211 | list_del(&page_node->list); | |
1212 | DO_NUMA(page_node->nid = nid); | |
1213 | rb_replace_node(&stable_node->node, | |
1214 | &page_node->node, &root_stable_tree[nid]); | |
1215 | get_page(page); | |
1216 | } else { | |
1217 | rb_erase(&stable_node->node, &root_stable_tree[nid]); | |
1218 | page = NULL; | |
1219 | } | |
1220 | stable_node->head = &migrate_nodes; | |
1221 | list_add(&stable_node->list, stable_node->head); | |
1222 | return page; | |
31dbd01f IE |
1223 | } |
1224 | ||
1225 | /* | |
e850dcf5 | 1226 | * stable_tree_insert - insert stable tree node pointing to new ksm page |
31dbd01f IE |
1227 | * into the stable tree. |
1228 | * | |
7b6ba2c7 HD |
1229 | * This function returns the stable tree node just allocated on success, |
1230 | * NULL otherwise. | |
31dbd01f | 1231 | */ |
7b6ba2c7 | 1232 | static struct stable_node *stable_tree_insert(struct page *kpage) |
31dbd01f | 1233 | { |
90bd6fd3 PH |
1234 | int nid; |
1235 | unsigned long kpfn; | |
1236 | struct rb_node **new; | |
31dbd01f | 1237 | struct rb_node *parent = NULL; |
7b6ba2c7 | 1238 | struct stable_node *stable_node; |
31dbd01f | 1239 | |
90bd6fd3 PH |
1240 | kpfn = page_to_pfn(kpage); |
1241 | nid = get_kpfn_nid(kpfn); | |
1242 | new = &root_stable_tree[nid].rb_node; | |
1243 | ||
31dbd01f | 1244 | while (*new) { |
4035c07a | 1245 | struct page *tree_page; |
31dbd01f IE |
1246 | int ret; |
1247 | ||
08beca44 | 1248 | cond_resched(); |
7b6ba2c7 | 1249 | stable_node = rb_entry(*new, struct stable_node, node); |
8aafa6a4 | 1250 | tree_page = get_ksm_page(stable_node, false); |
4035c07a HD |
1251 | if (!tree_page) |
1252 | return NULL; | |
31dbd01f | 1253 | |
4035c07a HD |
1254 | ret = memcmp_pages(kpage, tree_page); |
1255 | put_page(tree_page); | |
31dbd01f IE |
1256 | |
1257 | parent = *new; | |
1258 | if (ret < 0) | |
1259 | new = &parent->rb_left; | |
1260 | else if (ret > 0) | |
1261 | new = &parent->rb_right; | |
1262 | else { | |
1263 | /* | |
1264 | * It is not a bug that stable_tree_search() didn't | |
1265 | * find this node: because at that time our page was | |
1266 | * not yet write-protected, so may have changed since. | |
1267 | */ | |
1268 | return NULL; | |
1269 | } | |
1270 | } | |
1271 | ||
7b6ba2c7 HD |
1272 | stable_node = alloc_stable_node(); |
1273 | if (!stable_node) | |
1274 | return NULL; | |
31dbd01f | 1275 | |
7b6ba2c7 | 1276 | INIT_HLIST_HEAD(&stable_node->hlist); |
90bd6fd3 | 1277 | stable_node->kpfn = kpfn; |
08beca44 | 1278 | set_page_stable_node(kpage, stable_node); |
4146d2d6 | 1279 | DO_NUMA(stable_node->nid = nid); |
e850dcf5 HD |
1280 | rb_link_node(&stable_node->node, parent, new); |
1281 | rb_insert_color(&stable_node->node, &root_stable_tree[nid]); | |
08beca44 | 1282 | |
7b6ba2c7 | 1283 | return stable_node; |
31dbd01f IE |
1284 | } |
1285 | ||
1286 | /* | |
8dd3557a HD |
1287 | * unstable_tree_search_insert - search for identical page, |
1288 | * else insert rmap_item into the unstable tree. | |
31dbd01f IE |
1289 | * |
1290 | * This function searches for a page in the unstable tree identical to the | |
1291 | * page currently being scanned; and if no identical page is found in the | |
1292 | * tree, we insert rmap_item as a new object into the unstable tree. | |
1293 | * | |
1294 | * This function returns pointer to rmap_item found to be identical | |
1295 | * to the currently scanned page, NULL otherwise. | |
1296 | * | |
1297 | * This function does both searching and inserting, because they share | |
1298 | * the same walking algorithm in an rbtree. | |
1299 | */ | |
8dd3557a HD |
1300 | static |
1301 | struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item, | |
1302 | struct page *page, | |
1303 | struct page **tree_pagep) | |
31dbd01f | 1304 | { |
90bd6fd3 PH |
1305 | struct rb_node **new; |
1306 | struct rb_root *root; | |
31dbd01f | 1307 | struct rb_node *parent = NULL; |
90bd6fd3 PH |
1308 | int nid; |
1309 | ||
1310 | nid = get_kpfn_nid(page_to_pfn(page)); | |
1311 | root = &root_unstable_tree[nid]; | |
1312 | new = &root->rb_node; | |
31dbd01f IE |
1313 | |
1314 | while (*new) { | |
1315 | struct rmap_item *tree_rmap_item; | |
8dd3557a | 1316 | struct page *tree_page; |
31dbd01f IE |
1317 | int ret; |
1318 | ||
d178f27f | 1319 | cond_resched(); |
31dbd01f | 1320 | tree_rmap_item = rb_entry(*new, struct rmap_item, node); |
8dd3557a | 1321 | tree_page = get_mergeable_page(tree_rmap_item); |
22eccdd7 | 1322 | if (IS_ERR_OR_NULL(tree_page)) |
31dbd01f IE |
1323 | return NULL; |
1324 | ||
1325 | /* | |
8dd3557a | 1326 | * Don't substitute a ksm page for a forked page. |
31dbd01f | 1327 | */ |
8dd3557a HD |
1328 | if (page == tree_page) { |
1329 | put_page(tree_page); | |
31dbd01f IE |
1330 | return NULL; |
1331 | } | |
1332 | ||
90bd6fd3 PH |
1333 | /* |
1334 | * If tree_page has been migrated to another NUMA node, it | |
1335 | * will be flushed out and put into the right unstable tree | |
1336 | * next time: only merge with it if merge_across_nodes. | |
90bd6fd3 PH |
1337 | */ |
1338 | if (!ksm_merge_across_nodes && page_to_nid(tree_page) != nid) { | |
1339 | put_page(tree_page); | |
1340 | return NULL; | |
1341 | } | |
1342 | ||
8dd3557a | 1343 | ret = memcmp_pages(page, tree_page); |
31dbd01f IE |
1344 | |
1345 | parent = *new; | |
1346 | if (ret < 0) { | |
8dd3557a | 1347 | put_page(tree_page); |
31dbd01f IE |
1348 | new = &parent->rb_left; |
1349 | } else if (ret > 0) { | |
8dd3557a | 1350 | put_page(tree_page); |
31dbd01f IE |
1351 | new = &parent->rb_right; |
1352 | } else { | |
8dd3557a | 1353 | *tree_pagep = tree_page; |
31dbd01f IE |
1354 | return tree_rmap_item; |
1355 | } | |
1356 | } | |
1357 | ||
7b6ba2c7 | 1358 | rmap_item->address |= UNSTABLE_FLAG; |
31dbd01f | 1359 | rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); |
e850dcf5 | 1360 | DO_NUMA(rmap_item->nid = nid); |
31dbd01f | 1361 | rb_link_node(&rmap_item->node, parent, new); |
90bd6fd3 | 1362 | rb_insert_color(&rmap_item->node, root); |
31dbd01f | 1363 | |
473b0ce4 | 1364 | ksm_pages_unshared++; |
31dbd01f IE |
1365 | return NULL; |
1366 | } | |
1367 | ||
1368 | /* | |
1369 | * stable_tree_append - add another rmap_item to the linked list of | |
1370 | * rmap_items hanging off a given node of the stable tree, all sharing | |
1371 | * the same ksm page. | |
1372 | */ | |
1373 | static void stable_tree_append(struct rmap_item *rmap_item, | |
7b6ba2c7 | 1374 | struct stable_node *stable_node) |
31dbd01f | 1375 | { |
7b6ba2c7 | 1376 | rmap_item->head = stable_node; |
31dbd01f | 1377 | rmap_item->address |= STABLE_FLAG; |
7b6ba2c7 | 1378 | hlist_add_head(&rmap_item->hlist, &stable_node->hlist); |
e178dfde | 1379 | |
7b6ba2c7 HD |
1380 | if (rmap_item->hlist.next) |
1381 | ksm_pages_sharing++; | |
1382 | else | |
1383 | ksm_pages_shared++; | |
31dbd01f IE |
1384 | } |
1385 | ||
1386 | /* | |
81464e30 HD |
1387 | * cmp_and_merge_page - first see if page can be merged into the stable tree; |
1388 | * if not, compare checksum to previous and if it's the same, see if page can | |
1389 | * be inserted into the unstable tree, or merged with a page already there and | |
1390 | * both transferred to the stable tree. | |
31dbd01f IE |
1391 | * |
1392 | * @page: the page that we are searching identical page to. | |
1393 | * @rmap_item: the reverse mapping into the virtual address of this page | |
1394 | */ | |
1395 | static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item) | |
1396 | { | |
31dbd01f | 1397 | struct rmap_item *tree_rmap_item; |
8dd3557a | 1398 | struct page *tree_page = NULL; |
7b6ba2c7 | 1399 | struct stable_node *stable_node; |
8dd3557a | 1400 | struct page *kpage; |
31dbd01f IE |
1401 | unsigned int checksum; |
1402 | int err; | |
1403 | ||
4146d2d6 HD |
1404 | stable_node = page_stable_node(page); |
1405 | if (stable_node) { | |
1406 | if (stable_node->head != &migrate_nodes && | |
1407 | get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) { | |
1408 | rb_erase(&stable_node->node, | |
1409 | &root_stable_tree[NUMA(stable_node->nid)]); | |
1410 | stable_node->head = &migrate_nodes; | |
1411 | list_add(&stable_node->list, stable_node->head); | |
1412 | } | |
1413 | if (stable_node->head != &migrate_nodes && | |
1414 | rmap_item->head == stable_node) | |
1415 | return; | |
1416 | } | |
31dbd01f IE |
1417 | |
1418 | /* We first start with searching the page inside the stable tree */ | |
62b61f61 | 1419 | kpage = stable_tree_search(page); |
4146d2d6 HD |
1420 | if (kpage == page && rmap_item->head == stable_node) { |
1421 | put_page(kpage); | |
1422 | return; | |
1423 | } | |
1424 | ||
1425 | remove_rmap_item_from_tree(rmap_item); | |
1426 | ||
62b61f61 | 1427 | if (kpage) { |
08beca44 | 1428 | err = try_to_merge_with_ksm_page(rmap_item, page, kpage); |
31dbd01f IE |
1429 | if (!err) { |
1430 | /* | |
1431 | * The page was successfully merged: | |
1432 | * add its rmap_item to the stable tree. | |
1433 | */ | |
5ad64688 | 1434 | lock_page(kpage); |
62b61f61 | 1435 | stable_tree_append(rmap_item, page_stable_node(kpage)); |
5ad64688 | 1436 | unlock_page(kpage); |
31dbd01f | 1437 | } |
8dd3557a | 1438 | put_page(kpage); |
31dbd01f IE |
1439 | return; |
1440 | } | |
1441 | ||
1442 | /* | |
4035c07a HD |
1443 | * If the hash value of the page has changed from the last time |
1444 | * we calculated it, this page is changing frequently: therefore we | |
1445 | * don't want to insert it in the unstable tree, and we don't want | |
1446 | * to waste our time searching for something identical to it there. | |
31dbd01f IE |
1447 | */ |
1448 | checksum = calc_checksum(page); | |
1449 | if (rmap_item->oldchecksum != checksum) { | |
1450 | rmap_item->oldchecksum = checksum; | |
1451 | return; | |
1452 | } | |
1453 | ||
8dd3557a HD |
1454 | tree_rmap_item = |
1455 | unstable_tree_search_insert(rmap_item, page, &tree_page); | |
31dbd01f | 1456 | if (tree_rmap_item) { |
8dd3557a HD |
1457 | kpage = try_to_merge_two_pages(rmap_item, page, |
1458 | tree_rmap_item, tree_page); | |
1459 | put_page(tree_page); | |
31dbd01f IE |
1460 | /* |
1461 | * As soon as we merge this page, we want to remove the | |
1462 | * rmap_item of the page we have merged with from the unstable | |
1463 | * tree, and insert it instead as new node in the stable tree. | |
1464 | */ | |
8dd3557a | 1465 | if (kpage) { |
93d17715 | 1466 | remove_rmap_item_from_tree(tree_rmap_item); |
473b0ce4 | 1467 | |
5ad64688 | 1468 | lock_page(kpage); |
7b6ba2c7 HD |
1469 | stable_node = stable_tree_insert(kpage); |
1470 | if (stable_node) { | |
1471 | stable_tree_append(tree_rmap_item, stable_node); | |
1472 | stable_tree_append(rmap_item, stable_node); | |
1473 | } | |
5ad64688 | 1474 | unlock_page(kpage); |
7b6ba2c7 | 1475 | |
31dbd01f IE |
1476 | /* |
1477 | * If we fail to insert the page into the stable tree, | |
1478 | * we will have 2 virtual addresses that are pointing | |
1479 | * to a ksm page left outside the stable tree, | |
1480 | * in which case we need to break_cow on both. | |
1481 | */ | |
7b6ba2c7 | 1482 | if (!stable_node) { |
8dd3557a HD |
1483 | break_cow(tree_rmap_item); |
1484 | break_cow(rmap_item); | |
31dbd01f IE |
1485 | } |
1486 | } | |
31dbd01f IE |
1487 | } |
1488 | } | |
1489 | ||
1490 | static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot, | |
6514d511 | 1491 | struct rmap_item **rmap_list, |
31dbd01f IE |
1492 | unsigned long addr) |
1493 | { | |
1494 | struct rmap_item *rmap_item; | |
1495 | ||
6514d511 HD |
1496 | while (*rmap_list) { |
1497 | rmap_item = *rmap_list; | |
93d17715 | 1498 | if ((rmap_item->address & PAGE_MASK) == addr) |
31dbd01f | 1499 | return rmap_item; |
31dbd01f IE |
1500 | if (rmap_item->address > addr) |
1501 | break; | |
6514d511 | 1502 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 1503 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
1504 | free_rmap_item(rmap_item); |
1505 | } | |
1506 | ||
1507 | rmap_item = alloc_rmap_item(); | |
1508 | if (rmap_item) { | |
1509 | /* It has already been zeroed */ | |
1510 | rmap_item->mm = mm_slot->mm; | |
1511 | rmap_item->address = addr; | |
6514d511 HD |
1512 | rmap_item->rmap_list = *rmap_list; |
1513 | *rmap_list = rmap_item; | |
31dbd01f IE |
1514 | } |
1515 | return rmap_item; | |
1516 | } | |
1517 | ||
1518 | static struct rmap_item *scan_get_next_rmap_item(struct page **page) | |
1519 | { | |
1520 | struct mm_struct *mm; | |
1521 | struct mm_slot *slot; | |
1522 | struct vm_area_struct *vma; | |
1523 | struct rmap_item *rmap_item; | |
90bd6fd3 | 1524 | int nid; |
31dbd01f IE |
1525 | |
1526 | if (list_empty(&ksm_mm_head.mm_list)) | |
1527 | return NULL; | |
1528 | ||
1529 | slot = ksm_scan.mm_slot; | |
1530 | if (slot == &ksm_mm_head) { | |
2919bfd0 HD |
1531 | /* |
1532 | * A number of pages can hang around indefinitely on per-cpu | |
1533 | * pagevecs, raised page count preventing write_protect_page | |
1534 | * from merging them. Though it doesn't really matter much, | |
1535 | * it is puzzling to see some stuck in pages_volatile until | |
1536 | * other activity jostles them out, and they also prevented | |
1537 | * LTP's KSM test from succeeding deterministically; so drain | |
1538 | * them here (here rather than on entry to ksm_do_scan(), | |
1539 | * so we don't IPI too often when pages_to_scan is set low). | |
1540 | */ | |
1541 | lru_add_drain_all(); | |
1542 | ||
4146d2d6 HD |
1543 | /* |
1544 | * Whereas stale stable_nodes on the stable_tree itself | |
1545 | * get pruned in the regular course of stable_tree_search(), | |
1546 | * those moved out to the migrate_nodes list can accumulate: | |
1547 | * so prune them once before each full scan. | |
1548 | */ | |
1549 | if (!ksm_merge_across_nodes) { | |
1550 | struct stable_node *stable_node; | |
1551 | struct list_head *this, *next; | |
1552 | struct page *page; | |
1553 | ||
1554 | list_for_each_safe(this, next, &migrate_nodes) { | |
1555 | stable_node = list_entry(this, | |
1556 | struct stable_node, list); | |
1557 | page = get_ksm_page(stable_node, false); | |
1558 | if (page) | |
1559 | put_page(page); | |
1560 | cond_resched(); | |
1561 | } | |
1562 | } | |
1563 | ||
90bd6fd3 PH |
1564 | for (nid = 0; nid < nr_node_ids; nid++) |
1565 | root_unstable_tree[nid] = RB_ROOT; | |
31dbd01f IE |
1566 | |
1567 | spin_lock(&ksm_mmlist_lock); | |
1568 | slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list); | |
1569 | ksm_scan.mm_slot = slot; | |
1570 | spin_unlock(&ksm_mmlist_lock); | |
2b472611 HD |
1571 | /* |
1572 | * Although we tested list_empty() above, a racing __ksm_exit | |
1573 | * of the last mm on the list may have removed it since then. | |
1574 | */ | |
1575 | if (slot == &ksm_mm_head) | |
1576 | return NULL; | |
31dbd01f IE |
1577 | next_mm: |
1578 | ksm_scan.address = 0; | |
6514d511 | 1579 | ksm_scan.rmap_list = &slot->rmap_list; |
31dbd01f IE |
1580 | } |
1581 | ||
1582 | mm = slot->mm; | |
1583 | down_read(&mm->mmap_sem); | |
9ba69294 HD |
1584 | if (ksm_test_exit(mm)) |
1585 | vma = NULL; | |
1586 | else | |
1587 | vma = find_vma(mm, ksm_scan.address); | |
1588 | ||
1589 | for (; vma; vma = vma->vm_next) { | |
31dbd01f IE |
1590 | if (!(vma->vm_flags & VM_MERGEABLE)) |
1591 | continue; | |
1592 | if (ksm_scan.address < vma->vm_start) | |
1593 | ksm_scan.address = vma->vm_start; | |
1594 | if (!vma->anon_vma) | |
1595 | ksm_scan.address = vma->vm_end; | |
1596 | ||
1597 | while (ksm_scan.address < vma->vm_end) { | |
9ba69294 HD |
1598 | if (ksm_test_exit(mm)) |
1599 | break; | |
31dbd01f | 1600 | *page = follow_page(vma, ksm_scan.address, FOLL_GET); |
21ae5b01 AA |
1601 | if (IS_ERR_OR_NULL(*page)) { |
1602 | ksm_scan.address += PAGE_SIZE; | |
1603 | cond_resched(); | |
1604 | continue; | |
1605 | } | |
29ad768c AA |
1606 | if (PageAnon(*page) || |
1607 | page_trans_compound_anon(*page)) { | |
31dbd01f IE |
1608 | flush_anon_page(vma, *page, ksm_scan.address); |
1609 | flush_dcache_page(*page); | |
1610 | rmap_item = get_next_rmap_item(slot, | |
6514d511 | 1611 | ksm_scan.rmap_list, ksm_scan.address); |
31dbd01f | 1612 | if (rmap_item) { |
6514d511 HD |
1613 | ksm_scan.rmap_list = |
1614 | &rmap_item->rmap_list; | |
31dbd01f IE |
1615 | ksm_scan.address += PAGE_SIZE; |
1616 | } else | |
1617 | put_page(*page); | |
1618 | up_read(&mm->mmap_sem); | |
1619 | return rmap_item; | |
1620 | } | |
21ae5b01 | 1621 | put_page(*page); |
31dbd01f IE |
1622 | ksm_scan.address += PAGE_SIZE; |
1623 | cond_resched(); | |
1624 | } | |
1625 | } | |
1626 | ||
9ba69294 HD |
1627 | if (ksm_test_exit(mm)) { |
1628 | ksm_scan.address = 0; | |
6514d511 | 1629 | ksm_scan.rmap_list = &slot->rmap_list; |
9ba69294 | 1630 | } |
31dbd01f IE |
1631 | /* |
1632 | * Nuke all the rmap_items that are above this current rmap: | |
1633 | * because there were no VM_MERGEABLE vmas with such addresses. | |
1634 | */ | |
6514d511 | 1635 | remove_trailing_rmap_items(slot, ksm_scan.rmap_list); |
31dbd01f IE |
1636 | |
1637 | spin_lock(&ksm_mmlist_lock); | |
cd551f97 HD |
1638 | ksm_scan.mm_slot = list_entry(slot->mm_list.next, |
1639 | struct mm_slot, mm_list); | |
1640 | if (ksm_scan.address == 0) { | |
1641 | /* | |
1642 | * We've completed a full scan of all vmas, holding mmap_sem | |
1643 | * throughout, and found no VM_MERGEABLE: so do the same as | |
1644 | * __ksm_exit does to remove this mm from all our lists now. | |
9ba69294 HD |
1645 | * This applies either when cleaning up after __ksm_exit |
1646 | * (but beware: we can reach here even before __ksm_exit), | |
1647 | * or when all VM_MERGEABLE areas have been unmapped (and | |
1648 | * mmap_sem then protects against race with MADV_MERGEABLE). | |
cd551f97 | 1649 | */ |
4ca3a69b | 1650 | hash_del(&slot->link); |
cd551f97 | 1651 | list_del(&slot->mm_list); |
9ba69294 HD |
1652 | spin_unlock(&ksm_mmlist_lock); |
1653 | ||
cd551f97 HD |
1654 | free_mm_slot(slot); |
1655 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
9ba69294 HD |
1656 | up_read(&mm->mmap_sem); |
1657 | mmdrop(mm); | |
1658 | } else { | |
1659 | spin_unlock(&ksm_mmlist_lock); | |
1660 | up_read(&mm->mmap_sem); | |
cd551f97 | 1661 | } |
31dbd01f IE |
1662 | |
1663 | /* Repeat until we've completed scanning the whole list */ | |
cd551f97 | 1664 | slot = ksm_scan.mm_slot; |
31dbd01f IE |
1665 | if (slot != &ksm_mm_head) |
1666 | goto next_mm; | |
1667 | ||
31dbd01f IE |
1668 | ksm_scan.seqnr++; |
1669 | return NULL; | |
1670 | } | |
1671 | ||
1672 | /** | |
1673 | * ksm_do_scan - the ksm scanner main worker function. | |
1674 | * @scan_npages - number of pages we want to scan before we return. | |
1675 | */ | |
1676 | static void ksm_do_scan(unsigned int scan_npages) | |
1677 | { | |
1678 | struct rmap_item *rmap_item; | |
22eccdd7 | 1679 | struct page *uninitialized_var(page); |
31dbd01f | 1680 | |
878aee7d | 1681 | while (scan_npages-- && likely(!freezing(current))) { |
31dbd01f IE |
1682 | cond_resched(); |
1683 | rmap_item = scan_get_next_rmap_item(&page); | |
1684 | if (!rmap_item) | |
1685 | return; | |
4146d2d6 | 1686 | cmp_and_merge_page(page, rmap_item); |
31dbd01f IE |
1687 | put_page(page); |
1688 | } | |
1689 | } | |
1690 | ||
6e158384 HD |
1691 | static int ksmd_should_run(void) |
1692 | { | |
1693 | return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list); | |
1694 | } | |
1695 | ||
31dbd01f IE |
1696 | static int ksm_scan_thread(void *nothing) |
1697 | { | |
878aee7d | 1698 | set_freezable(); |
339aa624 | 1699 | set_user_nice(current, 5); |
31dbd01f IE |
1700 | |
1701 | while (!kthread_should_stop()) { | |
6e158384 HD |
1702 | mutex_lock(&ksm_thread_mutex); |
1703 | if (ksmd_should_run()) | |
31dbd01f | 1704 | ksm_do_scan(ksm_thread_pages_to_scan); |
6e158384 HD |
1705 | mutex_unlock(&ksm_thread_mutex); |
1706 | ||
878aee7d AA |
1707 | try_to_freeze(); |
1708 | ||
6e158384 | 1709 | if (ksmd_should_run()) { |
31dbd01f IE |
1710 | schedule_timeout_interruptible( |
1711 | msecs_to_jiffies(ksm_thread_sleep_millisecs)); | |
1712 | } else { | |
878aee7d | 1713 | wait_event_freezable(ksm_thread_wait, |
6e158384 | 1714 | ksmd_should_run() || kthread_should_stop()); |
31dbd01f IE |
1715 | } |
1716 | } | |
1717 | return 0; | |
1718 | } | |
1719 | ||
f8af4da3 HD |
1720 | int ksm_madvise(struct vm_area_struct *vma, unsigned long start, |
1721 | unsigned long end, int advice, unsigned long *vm_flags) | |
1722 | { | |
1723 | struct mm_struct *mm = vma->vm_mm; | |
d952b791 | 1724 | int err; |
f8af4da3 HD |
1725 | |
1726 | switch (advice) { | |
1727 | case MADV_MERGEABLE: | |
1728 | /* | |
1729 | * Be somewhat over-protective for now! | |
1730 | */ | |
1731 | if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE | | |
1732 | VM_PFNMAP | VM_IO | VM_DONTEXPAND | | |
314e51b9 | 1733 | VM_HUGETLB | VM_NONLINEAR | VM_MIXEDMAP)) |
f8af4da3 HD |
1734 | return 0; /* just ignore the advice */ |
1735 | ||
cc2383ec KK |
1736 | #ifdef VM_SAO |
1737 | if (*vm_flags & VM_SAO) | |
1738 | return 0; | |
1739 | #endif | |
1740 | ||
d952b791 HD |
1741 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { |
1742 | err = __ksm_enter(mm); | |
1743 | if (err) | |
1744 | return err; | |
1745 | } | |
f8af4da3 HD |
1746 | |
1747 | *vm_flags |= VM_MERGEABLE; | |
1748 | break; | |
1749 | ||
1750 | case MADV_UNMERGEABLE: | |
1751 | if (!(*vm_flags & VM_MERGEABLE)) | |
1752 | return 0; /* just ignore the advice */ | |
1753 | ||
d952b791 HD |
1754 | if (vma->anon_vma) { |
1755 | err = unmerge_ksm_pages(vma, start, end); | |
1756 | if (err) | |
1757 | return err; | |
1758 | } | |
f8af4da3 HD |
1759 | |
1760 | *vm_flags &= ~VM_MERGEABLE; | |
1761 | break; | |
1762 | } | |
1763 | ||
1764 | return 0; | |
1765 | } | |
1766 | ||
1767 | int __ksm_enter(struct mm_struct *mm) | |
1768 | { | |
6e158384 HD |
1769 | struct mm_slot *mm_slot; |
1770 | int needs_wakeup; | |
1771 | ||
1772 | mm_slot = alloc_mm_slot(); | |
31dbd01f IE |
1773 | if (!mm_slot) |
1774 | return -ENOMEM; | |
1775 | ||
6e158384 HD |
1776 | /* Check ksm_run too? Would need tighter locking */ |
1777 | needs_wakeup = list_empty(&ksm_mm_head.mm_list); | |
1778 | ||
31dbd01f IE |
1779 | spin_lock(&ksm_mmlist_lock); |
1780 | insert_to_mm_slots_hash(mm, mm_slot); | |
1781 | /* | |
cbf86cfe HD |
1782 | * When KSM_RUN_MERGE (or KSM_RUN_STOP), |
1783 | * insert just behind the scanning cursor, to let the area settle | |
31dbd01f IE |
1784 | * down a little; when fork is followed by immediate exec, we don't |
1785 | * want ksmd to waste time setting up and tearing down an rmap_list. | |
cbf86cfe HD |
1786 | * |
1787 | * But when KSM_RUN_UNMERGE, it's important to insert ahead of its | |
1788 | * scanning cursor, otherwise KSM pages in newly forked mms will be | |
1789 | * missed: then we might as well insert at the end of the list. | |
31dbd01f | 1790 | */ |
cbf86cfe HD |
1791 | if (ksm_run & KSM_RUN_UNMERGE) |
1792 | list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list); | |
1793 | else | |
1794 | list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list); | |
31dbd01f IE |
1795 | spin_unlock(&ksm_mmlist_lock); |
1796 | ||
f8af4da3 | 1797 | set_bit(MMF_VM_MERGEABLE, &mm->flags); |
9ba69294 | 1798 | atomic_inc(&mm->mm_count); |
6e158384 HD |
1799 | |
1800 | if (needs_wakeup) | |
1801 | wake_up_interruptible(&ksm_thread_wait); | |
1802 | ||
f8af4da3 HD |
1803 | return 0; |
1804 | } | |
1805 | ||
1c2fb7a4 | 1806 | void __ksm_exit(struct mm_struct *mm) |
f8af4da3 | 1807 | { |
cd551f97 | 1808 | struct mm_slot *mm_slot; |
9ba69294 | 1809 | int easy_to_free = 0; |
cd551f97 | 1810 | |
31dbd01f | 1811 | /* |
9ba69294 HD |
1812 | * This process is exiting: if it's straightforward (as is the |
1813 | * case when ksmd was never running), free mm_slot immediately. | |
1814 | * But if it's at the cursor or has rmap_items linked to it, use | |
1815 | * mmap_sem to synchronize with any break_cows before pagetables | |
1816 | * are freed, and leave the mm_slot on the list for ksmd to free. | |
1817 | * Beware: ksm may already have noticed it exiting and freed the slot. | |
31dbd01f | 1818 | */ |
9ba69294 | 1819 | |
cd551f97 HD |
1820 | spin_lock(&ksm_mmlist_lock); |
1821 | mm_slot = get_mm_slot(mm); | |
9ba69294 | 1822 | if (mm_slot && ksm_scan.mm_slot != mm_slot) { |
6514d511 | 1823 | if (!mm_slot->rmap_list) { |
4ca3a69b | 1824 | hash_del(&mm_slot->link); |
9ba69294 HD |
1825 | list_del(&mm_slot->mm_list); |
1826 | easy_to_free = 1; | |
1827 | } else { | |
1828 | list_move(&mm_slot->mm_list, | |
1829 | &ksm_scan.mm_slot->mm_list); | |
1830 | } | |
cd551f97 | 1831 | } |
cd551f97 HD |
1832 | spin_unlock(&ksm_mmlist_lock); |
1833 | ||
9ba69294 HD |
1834 | if (easy_to_free) { |
1835 | free_mm_slot(mm_slot); | |
1836 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
1837 | mmdrop(mm); | |
1838 | } else if (mm_slot) { | |
9ba69294 HD |
1839 | down_write(&mm->mmap_sem); |
1840 | up_write(&mm->mmap_sem); | |
9ba69294 | 1841 | } |
31dbd01f IE |
1842 | } |
1843 | ||
cbf86cfe | 1844 | struct page *ksm_might_need_to_copy(struct page *page, |
5ad64688 HD |
1845 | struct vm_area_struct *vma, unsigned long address) |
1846 | { | |
cbf86cfe | 1847 | struct anon_vma *anon_vma = page_anon_vma(page); |
5ad64688 HD |
1848 | struct page *new_page; |
1849 | ||
cbf86cfe HD |
1850 | if (PageKsm(page)) { |
1851 | if (page_stable_node(page) && | |
1852 | !(ksm_run & KSM_RUN_UNMERGE)) | |
1853 | return page; /* no need to copy it */ | |
1854 | } else if (!anon_vma) { | |
1855 | return page; /* no need to copy it */ | |
1856 | } else if (anon_vma->root == vma->anon_vma->root && | |
1857 | page->index == linear_page_index(vma, address)) { | |
1858 | return page; /* still no need to copy it */ | |
1859 | } | |
1860 | if (!PageUptodate(page)) | |
1861 | return page; /* let do_swap_page report the error */ | |
1862 | ||
5ad64688 HD |
1863 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
1864 | if (new_page) { | |
1865 | copy_user_highpage(new_page, page, address, vma); | |
1866 | ||
1867 | SetPageDirty(new_page); | |
1868 | __SetPageUptodate(new_page); | |
5ad64688 | 1869 | __set_page_locked(new_page); |
5ad64688 HD |
1870 | } |
1871 | ||
5ad64688 HD |
1872 | return new_page; |
1873 | } | |
1874 | ||
1875 | int page_referenced_ksm(struct page *page, struct mem_cgroup *memcg, | |
1876 | unsigned long *vm_flags) | |
1877 | { | |
1878 | struct stable_node *stable_node; | |
1879 | struct rmap_item *rmap_item; | |
1880 | struct hlist_node *hlist; | |
1881 | unsigned int mapcount = page_mapcount(page); | |
1882 | int referenced = 0; | |
db114b83 | 1883 | int search_new_forks = 0; |
5ad64688 HD |
1884 | |
1885 | VM_BUG_ON(!PageKsm(page)); | |
1886 | VM_BUG_ON(!PageLocked(page)); | |
1887 | ||
1888 | stable_node = page_stable_node(page); | |
1889 | if (!stable_node) | |
1890 | return 0; | |
db114b83 | 1891 | again: |
5ad64688 | 1892 | hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) { |
db114b83 | 1893 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
5beb4930 | 1894 | struct anon_vma_chain *vmac; |
db114b83 | 1895 | struct vm_area_struct *vma; |
5ad64688 | 1896 | |
b6b19f25 | 1897 | anon_vma_lock_read(anon_vma); |
bf181b9f ML |
1898 | anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, |
1899 | 0, ULONG_MAX) { | |
5beb4930 | 1900 | vma = vmac->vma; |
db114b83 HD |
1901 | if (rmap_item->address < vma->vm_start || |
1902 | rmap_item->address >= vma->vm_end) | |
1903 | continue; | |
1904 | /* | |
1905 | * Initially we examine only the vma which covers this | |
1906 | * rmap_item; but later, if there is still work to do, | |
1907 | * we examine covering vmas in other mms: in case they | |
1908 | * were forked from the original since ksmd passed. | |
1909 | */ | |
1910 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
1911 | continue; | |
1912 | ||
1913 | if (memcg && !mm_match_cgroup(vma->vm_mm, memcg)) | |
1914 | continue; | |
5ad64688 | 1915 | |
db114b83 | 1916 | referenced += page_referenced_one(page, vma, |
5ad64688 | 1917 | rmap_item->address, &mapcount, vm_flags); |
db114b83 HD |
1918 | if (!search_new_forks || !mapcount) |
1919 | break; | |
1920 | } | |
b6b19f25 | 1921 | anon_vma_unlock_read(anon_vma); |
5ad64688 HD |
1922 | if (!mapcount) |
1923 | goto out; | |
1924 | } | |
db114b83 HD |
1925 | if (!search_new_forks++) |
1926 | goto again; | |
5ad64688 | 1927 | out: |
5ad64688 HD |
1928 | return referenced; |
1929 | } | |
1930 | ||
1931 | int try_to_unmap_ksm(struct page *page, enum ttu_flags flags) | |
1932 | { | |
1933 | struct stable_node *stable_node; | |
1934 | struct hlist_node *hlist; | |
1935 | struct rmap_item *rmap_item; | |
1936 | int ret = SWAP_AGAIN; | |
db114b83 | 1937 | int search_new_forks = 0; |
5ad64688 HD |
1938 | |
1939 | VM_BUG_ON(!PageKsm(page)); | |
1940 | VM_BUG_ON(!PageLocked(page)); | |
1941 | ||
1942 | stable_node = page_stable_node(page); | |
1943 | if (!stable_node) | |
1944 | return SWAP_FAIL; | |
db114b83 | 1945 | again: |
5ad64688 | 1946 | hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) { |
db114b83 | 1947 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
5beb4930 | 1948 | struct anon_vma_chain *vmac; |
db114b83 | 1949 | struct vm_area_struct *vma; |
5ad64688 | 1950 | |
b6b19f25 | 1951 | anon_vma_lock_read(anon_vma); |
bf181b9f ML |
1952 | anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, |
1953 | 0, ULONG_MAX) { | |
5beb4930 | 1954 | vma = vmac->vma; |
db114b83 HD |
1955 | if (rmap_item->address < vma->vm_start || |
1956 | rmap_item->address >= vma->vm_end) | |
1957 | continue; | |
1958 | /* | |
1959 | * Initially we examine only the vma which covers this | |
1960 | * rmap_item; but later, if there is still work to do, | |
1961 | * we examine covering vmas in other mms: in case they | |
1962 | * were forked from the original since ksmd passed. | |
1963 | */ | |
1964 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
1965 | continue; | |
1966 | ||
1967 | ret = try_to_unmap_one(page, vma, | |
1968 | rmap_item->address, flags); | |
1969 | if (ret != SWAP_AGAIN || !page_mapped(page)) { | |
b6b19f25 | 1970 | anon_vma_unlock_read(anon_vma); |
db114b83 HD |
1971 | goto out; |
1972 | } | |
1973 | } | |
b6b19f25 | 1974 | anon_vma_unlock_read(anon_vma); |
5ad64688 | 1975 | } |
db114b83 HD |
1976 | if (!search_new_forks++) |
1977 | goto again; | |
5ad64688 | 1978 | out: |
5ad64688 HD |
1979 | return ret; |
1980 | } | |
1981 | ||
e9995ef9 HD |
1982 | #ifdef CONFIG_MIGRATION |
1983 | int rmap_walk_ksm(struct page *page, int (*rmap_one)(struct page *, | |
1984 | struct vm_area_struct *, unsigned long, void *), void *arg) | |
1985 | { | |
1986 | struct stable_node *stable_node; | |
1987 | struct hlist_node *hlist; | |
1988 | struct rmap_item *rmap_item; | |
1989 | int ret = SWAP_AGAIN; | |
1990 | int search_new_forks = 0; | |
1991 | ||
1992 | VM_BUG_ON(!PageKsm(page)); | |
1993 | VM_BUG_ON(!PageLocked(page)); | |
1994 | ||
1995 | stable_node = page_stable_node(page); | |
1996 | if (!stable_node) | |
1997 | return ret; | |
1998 | again: | |
1999 | hlist_for_each_entry(rmap_item, hlist, &stable_node->hlist, hlist) { | |
2000 | struct anon_vma *anon_vma = rmap_item->anon_vma; | |
5beb4930 | 2001 | struct anon_vma_chain *vmac; |
e9995ef9 HD |
2002 | struct vm_area_struct *vma; |
2003 | ||
b6b19f25 | 2004 | anon_vma_lock_read(anon_vma); |
bf181b9f ML |
2005 | anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, |
2006 | 0, ULONG_MAX) { | |
5beb4930 | 2007 | vma = vmac->vma; |
e9995ef9 HD |
2008 | if (rmap_item->address < vma->vm_start || |
2009 | rmap_item->address >= vma->vm_end) | |
2010 | continue; | |
2011 | /* | |
2012 | * Initially we examine only the vma which covers this | |
2013 | * rmap_item; but later, if there is still work to do, | |
2014 | * we examine covering vmas in other mms: in case they | |
2015 | * were forked from the original since ksmd passed. | |
2016 | */ | |
2017 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
2018 | continue; | |
2019 | ||
2020 | ret = rmap_one(page, vma, rmap_item->address, arg); | |
2021 | if (ret != SWAP_AGAIN) { | |
b6b19f25 | 2022 | anon_vma_unlock_read(anon_vma); |
e9995ef9 HD |
2023 | goto out; |
2024 | } | |
2025 | } | |
b6b19f25 | 2026 | anon_vma_unlock_read(anon_vma); |
e9995ef9 HD |
2027 | } |
2028 | if (!search_new_forks++) | |
2029 | goto again; | |
2030 | out: | |
2031 | return ret; | |
2032 | } | |
2033 | ||
2034 | void ksm_migrate_page(struct page *newpage, struct page *oldpage) | |
2035 | { | |
2036 | struct stable_node *stable_node; | |
2037 | ||
2038 | VM_BUG_ON(!PageLocked(oldpage)); | |
2039 | VM_BUG_ON(!PageLocked(newpage)); | |
2040 | VM_BUG_ON(newpage->mapping != oldpage->mapping); | |
2041 | ||
2042 | stable_node = page_stable_node(newpage); | |
2043 | if (stable_node) { | |
62b61f61 HD |
2044 | VM_BUG_ON(stable_node->kpfn != page_to_pfn(oldpage)); |
2045 | stable_node->kpfn = page_to_pfn(newpage); | |
c8d6553b HD |
2046 | /* |
2047 | * newpage->mapping was set in advance; now we need smp_wmb() | |
2048 | * to make sure that the new stable_node->kpfn is visible | |
2049 | * to get_ksm_page() before it can see that oldpage->mapping | |
2050 | * has gone stale (or that PageSwapCache has been cleared). | |
2051 | */ | |
2052 | smp_wmb(); | |
2053 | set_page_stable_node(oldpage, NULL); | |
e9995ef9 HD |
2054 | } |
2055 | } | |
2056 | #endif /* CONFIG_MIGRATION */ | |
2057 | ||
62b61f61 | 2058 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ee0ea59c HD |
2059 | static void ksm_check_stable_tree(unsigned long start_pfn, |
2060 | unsigned long end_pfn) | |
62b61f61 | 2061 | { |
ee0ea59c | 2062 | struct stable_node *stable_node; |
4146d2d6 | 2063 | struct list_head *this, *next; |
62b61f61 | 2064 | struct rb_node *node; |
90bd6fd3 | 2065 | int nid; |
62b61f61 | 2066 | |
ee0ea59c HD |
2067 | for (nid = 0; nid < nr_node_ids; nid++) { |
2068 | node = rb_first(&root_stable_tree[nid]); | |
2069 | while (node) { | |
90bd6fd3 PH |
2070 | stable_node = rb_entry(node, struct stable_node, node); |
2071 | if (stable_node->kpfn >= start_pfn && | |
ee0ea59c HD |
2072 | stable_node->kpfn < end_pfn) { |
2073 | /* | |
2074 | * Don't get_ksm_page, page has already gone: | |
2075 | * which is why we keep kpfn instead of page* | |
2076 | */ | |
2077 | remove_node_from_stable_tree(stable_node); | |
2078 | node = rb_first(&root_stable_tree[nid]); | |
2079 | } else | |
2080 | node = rb_next(node); | |
2081 | cond_resched(); | |
90bd6fd3 | 2082 | } |
ee0ea59c | 2083 | } |
4146d2d6 HD |
2084 | list_for_each_safe(this, next, &migrate_nodes) { |
2085 | stable_node = list_entry(this, struct stable_node, list); | |
2086 | if (stable_node->kpfn >= start_pfn && | |
2087 | stable_node->kpfn < end_pfn) | |
2088 | remove_node_from_stable_tree(stable_node); | |
2089 | cond_resched(); | |
2090 | } | |
62b61f61 HD |
2091 | } |
2092 | ||
2093 | static int ksm_memory_callback(struct notifier_block *self, | |
2094 | unsigned long action, void *arg) | |
2095 | { | |
2096 | struct memory_notify *mn = arg; | |
62b61f61 HD |
2097 | |
2098 | switch (action) { | |
2099 | case MEM_GOING_OFFLINE: | |
2100 | /* | |
2101 | * Keep it very simple for now: just lock out ksmd and | |
2102 | * MADV_UNMERGEABLE while any memory is going offline. | |
a0b0f58c KM |
2103 | * mutex_lock_nested() is necessary because lockdep was alarmed |
2104 | * that here we take ksm_thread_mutex inside notifier chain | |
2105 | * mutex, and later take notifier chain mutex inside | |
2106 | * ksm_thread_mutex to unlock it. But that's safe because both | |
2107 | * are inside mem_hotplug_mutex. | |
62b61f61 | 2108 | */ |
a0b0f58c | 2109 | mutex_lock_nested(&ksm_thread_mutex, SINGLE_DEPTH_NESTING); |
62b61f61 HD |
2110 | break; |
2111 | ||
2112 | case MEM_OFFLINE: | |
2113 | /* | |
2114 | * Most of the work is done by page migration; but there might | |
2115 | * be a few stable_nodes left over, still pointing to struct | |
ee0ea59c HD |
2116 | * pages which have been offlined: prune those from the tree, |
2117 | * otherwise get_ksm_page() might later try to access a | |
2118 | * non-existent struct page. | |
62b61f61 | 2119 | */ |
ee0ea59c HD |
2120 | ksm_check_stable_tree(mn->start_pfn, |
2121 | mn->start_pfn + mn->nr_pages); | |
62b61f61 HD |
2122 | /* fallthrough */ |
2123 | ||
2124 | case MEM_CANCEL_OFFLINE: | |
2125 | mutex_unlock(&ksm_thread_mutex); | |
2126 | break; | |
2127 | } | |
2128 | return NOTIFY_OK; | |
2129 | } | |
2130 | #endif /* CONFIG_MEMORY_HOTREMOVE */ | |
2131 | ||
2ffd8679 HD |
2132 | #ifdef CONFIG_SYSFS |
2133 | /* | |
2134 | * This all compiles without CONFIG_SYSFS, but is a waste of space. | |
2135 | */ | |
2136 | ||
31dbd01f IE |
2137 | #define KSM_ATTR_RO(_name) \ |
2138 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2139 | #define KSM_ATTR(_name) \ | |
2140 | static struct kobj_attribute _name##_attr = \ | |
2141 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2142 | ||
2143 | static ssize_t sleep_millisecs_show(struct kobject *kobj, | |
2144 | struct kobj_attribute *attr, char *buf) | |
2145 | { | |
2146 | return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs); | |
2147 | } | |
2148 | ||
2149 | static ssize_t sleep_millisecs_store(struct kobject *kobj, | |
2150 | struct kobj_attribute *attr, | |
2151 | const char *buf, size_t count) | |
2152 | { | |
2153 | unsigned long msecs; | |
2154 | int err; | |
2155 | ||
2156 | err = strict_strtoul(buf, 10, &msecs); | |
2157 | if (err || msecs > UINT_MAX) | |
2158 | return -EINVAL; | |
2159 | ||
2160 | ksm_thread_sleep_millisecs = msecs; | |
2161 | ||
2162 | return count; | |
2163 | } | |
2164 | KSM_ATTR(sleep_millisecs); | |
2165 | ||
2166 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
2167 | struct kobj_attribute *attr, char *buf) | |
2168 | { | |
2169 | return sprintf(buf, "%u\n", ksm_thread_pages_to_scan); | |
2170 | } | |
2171 | ||
2172 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
2173 | struct kobj_attribute *attr, | |
2174 | const char *buf, size_t count) | |
2175 | { | |
2176 | int err; | |
2177 | unsigned long nr_pages; | |
2178 | ||
2179 | err = strict_strtoul(buf, 10, &nr_pages); | |
2180 | if (err || nr_pages > UINT_MAX) | |
2181 | return -EINVAL; | |
2182 | ||
2183 | ksm_thread_pages_to_scan = nr_pages; | |
2184 | ||
2185 | return count; | |
2186 | } | |
2187 | KSM_ATTR(pages_to_scan); | |
2188 | ||
2189 | static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, | |
2190 | char *buf) | |
2191 | { | |
2192 | return sprintf(buf, "%u\n", ksm_run); | |
2193 | } | |
2194 | ||
2195 | static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, | |
2196 | const char *buf, size_t count) | |
2197 | { | |
2198 | int err; | |
2199 | unsigned long flags; | |
2200 | ||
2201 | err = strict_strtoul(buf, 10, &flags); | |
2202 | if (err || flags > UINT_MAX) | |
2203 | return -EINVAL; | |
2204 | if (flags > KSM_RUN_UNMERGE) | |
2205 | return -EINVAL; | |
2206 | ||
2207 | /* | |
2208 | * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. | |
2209 | * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, | |
d0f209f6 HD |
2210 | * breaking COW to free the pages_shared (but leaves mm_slots |
2211 | * on the list for when ksmd may be set running again). | |
31dbd01f IE |
2212 | */ |
2213 | ||
2214 | mutex_lock(&ksm_thread_mutex); | |
2215 | if (ksm_run != flags) { | |
2216 | ksm_run = flags; | |
d952b791 | 2217 | if (flags & KSM_RUN_UNMERGE) { |
e1e12d2f | 2218 | set_current_oom_origin(); |
d952b791 | 2219 | err = unmerge_and_remove_all_rmap_items(); |
e1e12d2f | 2220 | clear_current_oom_origin(); |
d952b791 HD |
2221 | if (err) { |
2222 | ksm_run = KSM_RUN_STOP; | |
2223 | count = err; | |
2224 | } | |
2225 | } | |
31dbd01f IE |
2226 | } |
2227 | mutex_unlock(&ksm_thread_mutex); | |
2228 | ||
2229 | if (flags & KSM_RUN_MERGE) | |
2230 | wake_up_interruptible(&ksm_thread_wait); | |
2231 | ||
2232 | return count; | |
2233 | } | |
2234 | KSM_ATTR(run); | |
2235 | ||
90bd6fd3 PH |
2236 | #ifdef CONFIG_NUMA |
2237 | static ssize_t merge_across_nodes_show(struct kobject *kobj, | |
2238 | struct kobj_attribute *attr, char *buf) | |
2239 | { | |
2240 | return sprintf(buf, "%u\n", ksm_merge_across_nodes); | |
2241 | } | |
2242 | ||
2243 | static ssize_t merge_across_nodes_store(struct kobject *kobj, | |
2244 | struct kobj_attribute *attr, | |
2245 | const char *buf, size_t count) | |
2246 | { | |
2247 | int err; | |
2248 | unsigned long knob; | |
2249 | ||
2250 | err = kstrtoul(buf, 10, &knob); | |
2251 | if (err) | |
2252 | return err; | |
2253 | if (knob > 1) | |
2254 | return -EINVAL; | |
2255 | ||
2256 | mutex_lock(&ksm_thread_mutex); | |
2257 | if (ksm_merge_across_nodes != knob) { | |
cbf86cfe | 2258 | if (ksm_pages_shared || remove_all_stable_nodes()) |
90bd6fd3 PH |
2259 | err = -EBUSY; |
2260 | else | |
2261 | ksm_merge_across_nodes = knob; | |
2262 | } | |
2263 | mutex_unlock(&ksm_thread_mutex); | |
2264 | ||
2265 | return err ? err : count; | |
2266 | } | |
2267 | KSM_ATTR(merge_across_nodes); | |
2268 | #endif | |
2269 | ||
b4028260 HD |
2270 | static ssize_t pages_shared_show(struct kobject *kobj, |
2271 | struct kobj_attribute *attr, char *buf) | |
2272 | { | |
2273 | return sprintf(buf, "%lu\n", ksm_pages_shared); | |
2274 | } | |
2275 | KSM_ATTR_RO(pages_shared); | |
2276 | ||
2277 | static ssize_t pages_sharing_show(struct kobject *kobj, | |
2278 | struct kobj_attribute *attr, char *buf) | |
2279 | { | |
e178dfde | 2280 | return sprintf(buf, "%lu\n", ksm_pages_sharing); |
b4028260 HD |
2281 | } |
2282 | KSM_ATTR_RO(pages_sharing); | |
2283 | ||
473b0ce4 HD |
2284 | static ssize_t pages_unshared_show(struct kobject *kobj, |
2285 | struct kobj_attribute *attr, char *buf) | |
2286 | { | |
2287 | return sprintf(buf, "%lu\n", ksm_pages_unshared); | |
2288 | } | |
2289 | KSM_ATTR_RO(pages_unshared); | |
2290 | ||
2291 | static ssize_t pages_volatile_show(struct kobject *kobj, | |
2292 | struct kobj_attribute *attr, char *buf) | |
2293 | { | |
2294 | long ksm_pages_volatile; | |
2295 | ||
2296 | ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared | |
2297 | - ksm_pages_sharing - ksm_pages_unshared; | |
2298 | /* | |
2299 | * It was not worth any locking to calculate that statistic, | |
2300 | * but it might therefore sometimes be negative: conceal that. | |
2301 | */ | |
2302 | if (ksm_pages_volatile < 0) | |
2303 | ksm_pages_volatile = 0; | |
2304 | return sprintf(buf, "%ld\n", ksm_pages_volatile); | |
2305 | } | |
2306 | KSM_ATTR_RO(pages_volatile); | |
2307 | ||
2308 | static ssize_t full_scans_show(struct kobject *kobj, | |
2309 | struct kobj_attribute *attr, char *buf) | |
2310 | { | |
2311 | return sprintf(buf, "%lu\n", ksm_scan.seqnr); | |
2312 | } | |
2313 | KSM_ATTR_RO(full_scans); | |
2314 | ||
31dbd01f IE |
2315 | static struct attribute *ksm_attrs[] = { |
2316 | &sleep_millisecs_attr.attr, | |
2317 | &pages_to_scan_attr.attr, | |
2318 | &run_attr.attr, | |
b4028260 HD |
2319 | &pages_shared_attr.attr, |
2320 | &pages_sharing_attr.attr, | |
473b0ce4 HD |
2321 | &pages_unshared_attr.attr, |
2322 | &pages_volatile_attr.attr, | |
2323 | &full_scans_attr.attr, | |
90bd6fd3 PH |
2324 | #ifdef CONFIG_NUMA |
2325 | &merge_across_nodes_attr.attr, | |
2326 | #endif | |
31dbd01f IE |
2327 | NULL, |
2328 | }; | |
2329 | ||
2330 | static struct attribute_group ksm_attr_group = { | |
2331 | .attrs = ksm_attrs, | |
2332 | .name = "ksm", | |
2333 | }; | |
2ffd8679 | 2334 | #endif /* CONFIG_SYSFS */ |
31dbd01f IE |
2335 | |
2336 | static int __init ksm_init(void) | |
2337 | { | |
2338 | struct task_struct *ksm_thread; | |
2339 | int err; | |
90bd6fd3 | 2340 | int nid; |
31dbd01f IE |
2341 | |
2342 | err = ksm_slab_init(); | |
2343 | if (err) | |
2344 | goto out; | |
2345 | ||
90bd6fd3 PH |
2346 | for (nid = 0; nid < nr_node_ids; nid++) |
2347 | root_stable_tree[nid] = RB_ROOT; | |
2348 | ||
31dbd01f IE |
2349 | ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); |
2350 | if (IS_ERR(ksm_thread)) { | |
2351 | printk(KERN_ERR "ksm: creating kthread failed\n"); | |
2352 | err = PTR_ERR(ksm_thread); | |
d9f8984c | 2353 | goto out_free; |
31dbd01f IE |
2354 | } |
2355 | ||
2ffd8679 | 2356 | #ifdef CONFIG_SYSFS |
31dbd01f IE |
2357 | err = sysfs_create_group(mm_kobj, &ksm_attr_group); |
2358 | if (err) { | |
2359 | printk(KERN_ERR "ksm: register sysfs failed\n"); | |
2ffd8679 | 2360 | kthread_stop(ksm_thread); |
d9f8984c | 2361 | goto out_free; |
31dbd01f | 2362 | } |
c73602ad HD |
2363 | #else |
2364 | ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ | |
2365 | ||
2ffd8679 | 2366 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 2367 | |
62b61f61 HD |
2368 | #ifdef CONFIG_MEMORY_HOTREMOVE |
2369 | /* | |
2370 | * Choose a high priority since the callback takes ksm_thread_mutex: | |
2371 | * later callbacks could only be taking locks which nest within that. | |
2372 | */ | |
2373 | hotplug_memory_notifier(ksm_memory_callback, 100); | |
2374 | #endif | |
31dbd01f IE |
2375 | return 0; |
2376 | ||
d9f8984c | 2377 | out_free: |
31dbd01f IE |
2378 | ksm_slab_free(); |
2379 | out: | |
2380 | return err; | |
f8af4da3 | 2381 | } |
31dbd01f | 2382 | module_init(ksm_init) |