]>
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.) | |
8fdb3dbf HD |
90 | * |
91 | * If the merge_across_nodes tunable is unset, then KSM maintains multiple | |
92 | * stable trees and multiple unstable trees: one of each for each NUMA node. | |
31dbd01f IE |
93 | */ |
94 | ||
95 | /** | |
96 | * struct mm_slot - ksm information per mm that is being scanned | |
97 | * @link: link to the mm_slots hash list | |
98 | * @mm_list: link into the mm_slots list, rooted in ksm_mm_head | |
6514d511 | 99 | * @rmap_list: head for this mm_slot's singly-linked list of rmap_items |
31dbd01f IE |
100 | * @mm: the mm that this information is valid for |
101 | */ | |
102 | struct mm_slot { | |
103 | struct hlist_node link; | |
104 | struct list_head mm_list; | |
6514d511 | 105 | struct rmap_item *rmap_list; |
31dbd01f IE |
106 | struct mm_struct *mm; |
107 | }; | |
108 | ||
109 | /** | |
110 | * struct ksm_scan - cursor for scanning | |
111 | * @mm_slot: the current mm_slot we are scanning | |
112 | * @address: the next address inside that to be scanned | |
6514d511 | 113 | * @rmap_list: link to the next rmap to be scanned in the rmap_list |
31dbd01f IE |
114 | * @seqnr: count of completed full scans (needed when removing unstable node) |
115 | * | |
116 | * There is only the one ksm_scan instance of this cursor structure. | |
117 | */ | |
118 | struct ksm_scan { | |
119 | struct mm_slot *mm_slot; | |
120 | unsigned long address; | |
6514d511 | 121 | struct rmap_item **rmap_list; |
31dbd01f IE |
122 | unsigned long seqnr; |
123 | }; | |
124 | ||
7b6ba2c7 HD |
125 | /** |
126 | * struct stable_node - node of the stable rbtree | |
127 | * @node: rb node of this ksm page in the stable tree | |
4146d2d6 | 128 | * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list |
7253473c | 129 | * @hlist_dup: linked into the stable_node->hlist with a stable_node chain |
4146d2d6 | 130 | * @list: linked into migrate_nodes, pending placement in the proper node tree |
7b6ba2c7 | 131 | * @hlist: hlist head of rmap_items using this ksm page |
4146d2d6 | 132 | * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) |
7253473c AA |
133 | * @chain_prune_time: time of the last full garbage collection |
134 | * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN | |
4146d2d6 | 135 | * @nid: NUMA node id of stable tree in which linked (may not match kpfn) |
7b6ba2c7 HD |
136 | */ |
137 | struct stable_node { | |
4146d2d6 HD |
138 | union { |
139 | struct rb_node node; /* when node of stable tree */ | |
140 | struct { /* when listed for migration */ | |
141 | struct list_head *head; | |
7253473c AA |
142 | struct { |
143 | struct hlist_node hlist_dup; | |
144 | struct list_head list; | |
145 | }; | |
4146d2d6 HD |
146 | }; |
147 | }; | |
7b6ba2c7 | 148 | struct hlist_head hlist; |
7253473c AA |
149 | union { |
150 | unsigned long kpfn; | |
151 | unsigned long chain_prune_time; | |
152 | }; | |
153 | /* | |
154 | * STABLE_NODE_CHAIN can be any negative number in | |
155 | * rmap_hlist_len negative range, but better not -1 to be able | |
156 | * to reliably detect underflows. | |
157 | */ | |
158 | #define STABLE_NODE_CHAIN -1024 | |
159 | int rmap_hlist_len; | |
4146d2d6 HD |
160 | #ifdef CONFIG_NUMA |
161 | int nid; | |
162 | #endif | |
7b6ba2c7 HD |
163 | }; |
164 | ||
31dbd01f IE |
165 | /** |
166 | * struct rmap_item - reverse mapping item for virtual addresses | |
6514d511 | 167 | * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list |
db114b83 | 168 | * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree |
bc56620b | 169 | * @nid: NUMA node id of unstable tree in which linked (may not match page) |
31dbd01f IE |
170 | * @mm: the memory structure this rmap_item is pointing into |
171 | * @address: the virtual address this rmap_item tracks (+ flags in low bits) | |
172 | * @oldchecksum: previous checksum of the page at that virtual address | |
7b6ba2c7 HD |
173 | * @node: rb node of this rmap_item in the unstable tree |
174 | * @head: pointer to stable_node heading this list in the stable tree | |
175 | * @hlist: link into hlist of rmap_items hanging off that stable_node | |
31dbd01f IE |
176 | */ |
177 | struct rmap_item { | |
6514d511 | 178 | struct rmap_item *rmap_list; |
bc56620b HD |
179 | union { |
180 | struct anon_vma *anon_vma; /* when stable */ | |
181 | #ifdef CONFIG_NUMA | |
182 | int nid; /* when node of unstable tree */ | |
183 | #endif | |
184 | }; | |
31dbd01f IE |
185 | struct mm_struct *mm; |
186 | unsigned long address; /* + low bits used for flags below */ | |
7b6ba2c7 | 187 | unsigned int oldchecksum; /* when unstable */ |
31dbd01f | 188 | union { |
7b6ba2c7 HD |
189 | struct rb_node node; /* when node of unstable tree */ |
190 | struct { /* when listed from stable tree */ | |
191 | struct stable_node *head; | |
192 | struct hlist_node hlist; | |
193 | }; | |
31dbd01f IE |
194 | }; |
195 | }; | |
196 | ||
197 | #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ | |
7b6ba2c7 HD |
198 | #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ |
199 | #define STABLE_FLAG 0x200 /* is listed from the stable tree */ | |
31dbd01f IE |
200 | |
201 | /* The stable and unstable tree heads */ | |
ef53d16c HD |
202 | static struct rb_root one_stable_tree[1] = { RB_ROOT }; |
203 | static struct rb_root one_unstable_tree[1] = { RB_ROOT }; | |
204 | static struct rb_root *root_stable_tree = one_stable_tree; | |
205 | static struct rb_root *root_unstable_tree = one_unstable_tree; | |
31dbd01f | 206 | |
4146d2d6 HD |
207 | /* Recently migrated nodes of stable tree, pending proper placement */ |
208 | static LIST_HEAD(migrate_nodes); | |
7253473c | 209 | #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev) |
4146d2d6 | 210 | |
4ca3a69b SL |
211 | #define MM_SLOTS_HASH_BITS 10 |
212 | static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
31dbd01f IE |
213 | |
214 | static struct mm_slot ksm_mm_head = { | |
215 | .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list), | |
216 | }; | |
217 | static struct ksm_scan ksm_scan = { | |
218 | .mm_slot = &ksm_mm_head, | |
219 | }; | |
220 | ||
221 | static struct kmem_cache *rmap_item_cache; | |
7b6ba2c7 | 222 | static struct kmem_cache *stable_node_cache; |
31dbd01f IE |
223 | static struct kmem_cache *mm_slot_cache; |
224 | ||
225 | /* The number of nodes in the stable tree */ | |
b4028260 | 226 | static unsigned long ksm_pages_shared; |
31dbd01f | 227 | |
e178dfde | 228 | /* The number of page slots additionally sharing those nodes */ |
b4028260 | 229 | static unsigned long ksm_pages_sharing; |
31dbd01f | 230 | |
473b0ce4 HD |
231 | /* The number of nodes in the unstable tree */ |
232 | static unsigned long ksm_pages_unshared; | |
233 | ||
234 | /* The number of rmap_items in use: to calculate pages_volatile */ | |
235 | static unsigned long ksm_rmap_items; | |
236 | ||
7253473c AA |
237 | /* The number of stable_node chains */ |
238 | static unsigned long ksm_stable_node_chains; | |
239 | ||
240 | /* The number of stable_node dups linked to the stable_node chains */ | |
241 | static unsigned long ksm_stable_node_dups; | |
242 | ||
243 | /* Delay in pruning stale stable_node_dups in the stable_node_chains */ | |
244 | static int ksm_stable_node_chains_prune_millisecs = 2000; | |
245 | ||
246 | /* Maximum number of page slots sharing a stable node */ | |
247 | static int ksm_max_page_sharing = 256; | |
248 | ||
31dbd01f | 249 | /* Number of pages ksmd should scan in one batch */ |
2c6854fd | 250 | static unsigned int ksm_thread_pages_to_scan = 100; |
31dbd01f IE |
251 | |
252 | /* Milliseconds ksmd should sleep between batches */ | |
2ffd8679 | 253 | static unsigned int ksm_thread_sleep_millisecs = 20; |
31dbd01f | 254 | |
fa2aa0fc CI |
255 | /* Checksum of an empty (zeroed) page */ |
256 | static unsigned int zero_checksum __read_mostly; | |
257 | ||
258 | /* Whether to merge empty (zeroed) pages with actual zero pages */ | |
259 | static bool ksm_use_zero_pages __read_mostly; | |
260 | ||
e850dcf5 | 261 | #ifdef CONFIG_NUMA |
90bd6fd3 PH |
262 | /* Zeroed when merging across nodes is not allowed */ |
263 | static unsigned int ksm_merge_across_nodes = 1; | |
ef53d16c | 264 | static int ksm_nr_node_ids = 1; |
e850dcf5 HD |
265 | #else |
266 | #define ksm_merge_across_nodes 1U | |
ef53d16c | 267 | #define ksm_nr_node_ids 1 |
e850dcf5 | 268 | #endif |
90bd6fd3 | 269 | |
31dbd01f IE |
270 | #define KSM_RUN_STOP 0 |
271 | #define KSM_RUN_MERGE 1 | |
272 | #define KSM_RUN_UNMERGE 2 | |
ef4d43a8 HD |
273 | #define KSM_RUN_OFFLINE 4 |
274 | static unsigned long ksm_run = KSM_RUN_STOP; | |
275 | static void wait_while_offlining(void); | |
31dbd01f IE |
276 | |
277 | static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); | |
278 | static DEFINE_MUTEX(ksm_thread_mutex); | |
279 | static DEFINE_SPINLOCK(ksm_mmlist_lock); | |
280 | ||
281 | #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\ | |
282 | sizeof(struct __struct), __alignof__(struct __struct),\ | |
283 | (__flags), NULL) | |
284 | ||
285 | static int __init ksm_slab_init(void) | |
286 | { | |
287 | rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0); | |
288 | if (!rmap_item_cache) | |
289 | goto out; | |
290 | ||
7b6ba2c7 HD |
291 | stable_node_cache = KSM_KMEM_CACHE(stable_node, 0); |
292 | if (!stable_node_cache) | |
293 | goto out_free1; | |
294 | ||
31dbd01f IE |
295 | mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0); |
296 | if (!mm_slot_cache) | |
7b6ba2c7 | 297 | goto out_free2; |
31dbd01f IE |
298 | |
299 | return 0; | |
300 | ||
7b6ba2c7 HD |
301 | out_free2: |
302 | kmem_cache_destroy(stable_node_cache); | |
303 | out_free1: | |
31dbd01f IE |
304 | kmem_cache_destroy(rmap_item_cache); |
305 | out: | |
306 | return -ENOMEM; | |
307 | } | |
308 | ||
309 | static void __init ksm_slab_free(void) | |
310 | { | |
311 | kmem_cache_destroy(mm_slot_cache); | |
7b6ba2c7 | 312 | kmem_cache_destroy(stable_node_cache); |
31dbd01f IE |
313 | kmem_cache_destroy(rmap_item_cache); |
314 | mm_slot_cache = NULL; | |
315 | } | |
316 | ||
7253473c AA |
317 | static __always_inline bool is_stable_node_chain(struct stable_node *chain) |
318 | { | |
319 | return chain->rmap_hlist_len == STABLE_NODE_CHAIN; | |
320 | } | |
321 | ||
322 | static __always_inline bool is_stable_node_dup(struct stable_node *dup) | |
323 | { | |
324 | return dup->head == STABLE_NODE_DUP_HEAD; | |
325 | } | |
326 | ||
327 | static inline void stable_node_chain_add_dup(struct stable_node *dup, | |
328 | struct stable_node *chain) | |
329 | { | |
330 | VM_BUG_ON(is_stable_node_dup(dup)); | |
331 | dup->head = STABLE_NODE_DUP_HEAD; | |
332 | VM_BUG_ON(!is_stable_node_chain(chain)); | |
333 | hlist_add_head(&dup->hlist_dup, &chain->hlist); | |
334 | ksm_stable_node_dups++; | |
335 | } | |
336 | ||
337 | static inline void __stable_node_dup_del(struct stable_node *dup) | |
338 | { | |
b8d20a18 | 339 | VM_BUG_ON(!is_stable_node_dup(dup)); |
7253473c AA |
340 | hlist_del(&dup->hlist_dup); |
341 | ksm_stable_node_dups--; | |
342 | } | |
343 | ||
344 | static inline void stable_node_dup_del(struct stable_node *dup) | |
345 | { | |
346 | VM_BUG_ON(is_stable_node_chain(dup)); | |
347 | if (is_stable_node_dup(dup)) | |
348 | __stable_node_dup_del(dup); | |
349 | else | |
350 | rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid)); | |
351 | #ifdef CONFIG_DEBUG_VM | |
352 | dup->head = NULL; | |
353 | #endif | |
354 | } | |
355 | ||
31dbd01f IE |
356 | static inline struct rmap_item *alloc_rmap_item(void) |
357 | { | |
473b0ce4 HD |
358 | struct rmap_item *rmap_item; |
359 | ||
5b398e41 | 360 | rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | |
361 | __GFP_NORETRY | __GFP_NOWARN); | |
473b0ce4 HD |
362 | if (rmap_item) |
363 | ksm_rmap_items++; | |
364 | return rmap_item; | |
31dbd01f IE |
365 | } |
366 | ||
367 | static inline void free_rmap_item(struct rmap_item *rmap_item) | |
368 | { | |
473b0ce4 | 369 | ksm_rmap_items--; |
31dbd01f IE |
370 | rmap_item->mm = NULL; /* debug safety */ |
371 | kmem_cache_free(rmap_item_cache, rmap_item); | |
372 | } | |
373 | ||
7b6ba2c7 HD |
374 | static inline struct stable_node *alloc_stable_node(void) |
375 | { | |
6213055f | 376 | /* |
377 | * The allocation can take too long with GFP_KERNEL when memory is under | |
378 | * pressure, which may lead to hung task warnings. Adding __GFP_HIGH | |
379 | * grants access to memory reserves, helping to avoid this problem. | |
380 | */ | |
381 | return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); | |
7b6ba2c7 HD |
382 | } |
383 | ||
384 | static inline void free_stable_node(struct stable_node *stable_node) | |
385 | { | |
7253473c AA |
386 | VM_BUG_ON(stable_node->rmap_hlist_len && |
387 | !is_stable_node_chain(stable_node)); | |
7b6ba2c7 HD |
388 | kmem_cache_free(stable_node_cache, stable_node); |
389 | } | |
390 | ||
31dbd01f IE |
391 | static inline struct mm_slot *alloc_mm_slot(void) |
392 | { | |
393 | if (!mm_slot_cache) /* initialization failed */ | |
394 | return NULL; | |
395 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
396 | } | |
397 | ||
398 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
399 | { | |
400 | kmem_cache_free(mm_slot_cache, mm_slot); | |
401 | } | |
402 | ||
31dbd01f IE |
403 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) |
404 | { | |
4ca3a69b SL |
405 | struct mm_slot *slot; |
406 | ||
b67bfe0d | 407 | hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm) |
4ca3a69b SL |
408 | if (slot->mm == mm) |
409 | return slot; | |
31dbd01f | 410 | |
31dbd01f IE |
411 | return NULL; |
412 | } | |
413 | ||
414 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
415 | struct mm_slot *mm_slot) | |
416 | { | |
31dbd01f | 417 | mm_slot->mm = mm; |
4ca3a69b | 418 | hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm); |
31dbd01f IE |
419 | } |
420 | ||
a913e182 HD |
421 | /* |
422 | * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's | |
423 | * page tables after it has passed through ksm_exit() - which, if necessary, | |
424 | * takes mmap_sem briefly to serialize against them. ksm_exit() does not set | |
425 | * a special flag: they can just back out as soon as mm_users goes to zero. | |
426 | * ksm_test_exit() is used throughout to make this test for exit: in some | |
427 | * places for correctness, in some places just to avoid unnecessary work. | |
428 | */ | |
429 | static inline bool ksm_test_exit(struct mm_struct *mm) | |
430 | { | |
431 | return atomic_read(&mm->mm_users) == 0; | |
432 | } | |
433 | ||
31dbd01f IE |
434 | /* |
435 | * We use break_ksm to break COW on a ksm page: it's a stripped down | |
436 | * | |
d4edcf0d | 437 | * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1) |
31dbd01f IE |
438 | * put_page(page); |
439 | * | |
440 | * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma, | |
441 | * in case the application has unmapped and remapped mm,addr meanwhile. | |
442 | * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP | |
443 | * mmap of /dev/mem or /dev/kmem, where we would not want to touch it. | |
1b2ee126 DH |
444 | * |
445 | * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context | |
446 | * of the process that owns 'vma'. We also do not want to enforce | |
447 | * protection keys here anyway. | |
31dbd01f | 448 | */ |
d952b791 | 449 | static int break_ksm(struct vm_area_struct *vma, unsigned long addr) |
31dbd01f IE |
450 | { |
451 | struct page *page; | |
d952b791 | 452 | int ret = 0; |
31dbd01f IE |
453 | |
454 | do { | |
455 | cond_resched(); | |
1b2ee126 DH |
456 | page = follow_page(vma, addr, |
457 | FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE); | |
22eccdd7 | 458 | if (IS_ERR_OR_NULL(page)) |
31dbd01f IE |
459 | break; |
460 | if (PageKsm(page)) | |
dcddffd4 KS |
461 | ret = handle_mm_fault(vma, addr, |
462 | FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE); | |
31dbd01f IE |
463 | else |
464 | ret = VM_FAULT_WRITE; | |
465 | put_page(page); | |
33692f27 | 466 | } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); |
d952b791 HD |
467 | /* |
468 | * We must loop because handle_mm_fault() may back out if there's | |
469 | * any difficulty e.g. if pte accessed bit gets updated concurrently. | |
470 | * | |
471 | * VM_FAULT_WRITE is what we have been hoping for: it indicates that | |
472 | * COW has been broken, even if the vma does not permit VM_WRITE; | |
473 | * but note that a concurrent fault might break PageKsm for us. | |
474 | * | |
475 | * VM_FAULT_SIGBUS could occur if we race with truncation of the | |
476 | * backing file, which also invalidates anonymous pages: that's | |
477 | * okay, that truncation will have unmapped the PageKsm for us. | |
478 | * | |
479 | * VM_FAULT_OOM: at the time of writing (late July 2009), setting | |
480 | * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the | |
481 | * current task has TIF_MEMDIE set, and will be OOM killed on return | |
482 | * to user; and ksmd, having no mm, would never be chosen for that. | |
483 | * | |
484 | * But if the mm is in a limited mem_cgroup, then the fault may fail | |
485 | * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and | |
486 | * even ksmd can fail in this way - though it's usually breaking ksm | |
487 | * just to undo a merge it made a moment before, so unlikely to oom. | |
488 | * | |
489 | * That's a pity: we might therefore have more kernel pages allocated | |
490 | * than we're counting as nodes in the stable tree; but ksm_do_scan | |
491 | * will retry to break_cow on each pass, so should recover the page | |
492 | * in due course. The important thing is to not let VM_MERGEABLE | |
493 | * be cleared while any such pages might remain in the area. | |
494 | */ | |
495 | return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; | |
31dbd01f IE |
496 | } |
497 | ||
ef694222 BL |
498 | static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, |
499 | unsigned long addr) | |
500 | { | |
501 | struct vm_area_struct *vma; | |
502 | if (ksm_test_exit(mm)) | |
503 | return NULL; | |
504 | vma = find_vma(mm, addr); | |
505 | if (!vma || vma->vm_start > addr) | |
506 | return NULL; | |
507 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) | |
508 | return NULL; | |
509 | return vma; | |
510 | } | |
511 | ||
8dd3557a | 512 | static void break_cow(struct rmap_item *rmap_item) |
31dbd01f | 513 | { |
8dd3557a HD |
514 | struct mm_struct *mm = rmap_item->mm; |
515 | unsigned long addr = rmap_item->address; | |
31dbd01f IE |
516 | struct vm_area_struct *vma; |
517 | ||
4035c07a HD |
518 | /* |
519 | * It is not an accident that whenever we want to break COW | |
520 | * to undo, we also need to drop a reference to the anon_vma. | |
521 | */ | |
9e60109f | 522 | put_anon_vma(rmap_item->anon_vma); |
4035c07a | 523 | |
81464e30 | 524 | down_read(&mm->mmap_sem); |
ef694222 BL |
525 | vma = find_mergeable_vma(mm, addr); |
526 | if (vma) | |
527 | break_ksm(vma, addr); | |
31dbd01f IE |
528 | up_read(&mm->mmap_sem); |
529 | } | |
530 | ||
531 | static struct page *get_mergeable_page(struct rmap_item *rmap_item) | |
532 | { | |
533 | struct mm_struct *mm = rmap_item->mm; | |
534 | unsigned long addr = rmap_item->address; | |
535 | struct vm_area_struct *vma; | |
536 | struct page *page; | |
537 | ||
538 | down_read(&mm->mmap_sem); | |
ef694222 BL |
539 | vma = find_mergeable_vma(mm, addr); |
540 | if (!vma) | |
31dbd01f IE |
541 | goto out; |
542 | ||
543 | page = follow_page(vma, addr, FOLL_GET); | |
22eccdd7 | 544 | if (IS_ERR_OR_NULL(page)) |
31dbd01f | 545 | goto out; |
f765f540 | 546 | if (PageAnon(page)) { |
31dbd01f IE |
547 | flush_anon_page(vma, page, addr); |
548 | flush_dcache_page(page); | |
549 | } else { | |
550 | put_page(page); | |
c8f95ed1 AA |
551 | out: |
552 | page = NULL; | |
31dbd01f IE |
553 | } |
554 | up_read(&mm->mmap_sem); | |
555 | return page; | |
556 | } | |
557 | ||
90bd6fd3 PH |
558 | /* |
559 | * This helper is used for getting right index into array of tree roots. | |
560 | * When merge_across_nodes knob is set to 1, there are only two rb-trees for | |
561 | * stable and unstable pages from all nodes with roots in index 0. Otherwise, | |
562 | * every node has its own stable and unstable tree. | |
563 | */ | |
564 | static inline int get_kpfn_nid(unsigned long kpfn) | |
565 | { | |
d8fc16a8 | 566 | return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); |
90bd6fd3 PH |
567 | } |
568 | ||
7253473c AA |
569 | static struct stable_node *alloc_stable_node_chain(struct stable_node *dup, |
570 | struct rb_root *root) | |
571 | { | |
572 | struct stable_node *chain = alloc_stable_node(); | |
573 | VM_BUG_ON(is_stable_node_chain(dup)); | |
574 | if (likely(chain)) { | |
575 | INIT_HLIST_HEAD(&chain->hlist); | |
576 | chain->chain_prune_time = jiffies; | |
577 | chain->rmap_hlist_len = STABLE_NODE_CHAIN; | |
578 | #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA) | |
579 | chain->nid = -1; /* debug */ | |
580 | #endif | |
581 | ksm_stable_node_chains++; | |
582 | ||
583 | /* | |
584 | * Put the stable node chain in the first dimension of | |
585 | * the stable tree and at the same time remove the old | |
586 | * stable node. | |
587 | */ | |
588 | rb_replace_node(&dup->node, &chain->node, root); | |
589 | ||
590 | /* | |
591 | * Move the old stable node to the second dimension | |
592 | * queued in the hlist_dup. The invariant is that all | |
593 | * dup stable_nodes in the chain->hlist point to pages | |
594 | * that are wrprotected and have the exact same | |
595 | * content. | |
596 | */ | |
597 | stable_node_chain_add_dup(dup, chain); | |
598 | } | |
599 | return chain; | |
600 | } | |
601 | ||
602 | static inline void free_stable_node_chain(struct stable_node *chain, | |
603 | struct rb_root *root) | |
604 | { | |
605 | rb_erase(&chain->node, root); | |
606 | free_stable_node(chain); | |
607 | ksm_stable_node_chains--; | |
608 | } | |
609 | ||
4035c07a HD |
610 | static void remove_node_from_stable_tree(struct stable_node *stable_node) |
611 | { | |
612 | struct rmap_item *rmap_item; | |
4035c07a | 613 | |
7253473c AA |
614 | /* check it's not STABLE_NODE_CHAIN or negative */ |
615 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
616 | ||
b67bfe0d | 617 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
4035c07a HD |
618 | if (rmap_item->hlist.next) |
619 | ksm_pages_sharing--; | |
620 | else | |
621 | ksm_pages_shared--; | |
7253473c AA |
622 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
623 | stable_node->rmap_hlist_len--; | |
9e60109f | 624 | put_anon_vma(rmap_item->anon_vma); |
4035c07a HD |
625 | rmap_item->address &= PAGE_MASK; |
626 | cond_resched(); | |
627 | } | |
628 | ||
7253473c AA |
629 | /* |
630 | * We need the second aligned pointer of the migrate_nodes | |
631 | * list_head to stay clear from the rb_parent_color union | |
632 | * (aligned and different than any node) and also different | |
633 | * from &migrate_nodes. This will verify that future list.h changes | |
634 | * don't break STABLE_NODE_DUP_HEAD. | |
635 | */ | |
636 | #if GCC_VERSION >= 40903 /* only recent gcc can handle it */ | |
637 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes); | |
638 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1); | |
639 | #endif | |
640 | ||
4146d2d6 HD |
641 | if (stable_node->head == &migrate_nodes) |
642 | list_del(&stable_node->list); | |
643 | else | |
7253473c | 644 | stable_node_dup_del(stable_node); |
4035c07a HD |
645 | free_stable_node(stable_node); |
646 | } | |
647 | ||
648 | /* | |
649 | * get_ksm_page: checks if the page indicated by the stable node | |
650 | * is still its ksm page, despite having held no reference to it. | |
651 | * In which case we can trust the content of the page, and it | |
652 | * returns the gotten page; but if the page has now been zapped, | |
653 | * remove the stale node from the stable tree and return NULL. | |
c8d6553b | 654 | * But beware, the stable node's page might be being migrated. |
4035c07a HD |
655 | * |
656 | * You would expect the stable_node to hold a reference to the ksm page. | |
657 | * But if it increments the page's count, swapping out has to wait for | |
658 | * ksmd to come around again before it can free the page, which may take | |
659 | * seconds or even minutes: much too unresponsive. So instead we use a | |
660 | * "keyhole reference": access to the ksm page from the stable node peeps | |
661 | * out through its keyhole to see if that page still holds the right key, | |
662 | * pointing back to this stable node. This relies on freeing a PageAnon | |
663 | * page to reset its page->mapping to NULL, and relies on no other use of | |
664 | * a page to put something that might look like our key in page->mapping. | |
4035c07a HD |
665 | * is on its way to being freed; but it is an anomaly to bear in mind. |
666 | */ | |
8fdb3dbf | 667 | static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it) |
4035c07a HD |
668 | { |
669 | struct page *page; | |
670 | void *expected_mapping; | |
c8d6553b | 671 | unsigned long kpfn; |
4035c07a | 672 | |
bda807d4 MK |
673 | expected_mapping = (void *)((unsigned long)stable_node | |
674 | PAGE_MAPPING_KSM); | |
c8d6553b | 675 | again: |
4db0c3c2 | 676 | kpfn = READ_ONCE(stable_node->kpfn); |
c8d6553b HD |
677 | page = pfn_to_page(kpfn); |
678 | ||
679 | /* | |
680 | * page is computed from kpfn, so on most architectures reading | |
681 | * page->mapping is naturally ordered after reading node->kpfn, | |
682 | * but on Alpha we need to be more careful. | |
683 | */ | |
684 | smp_read_barrier_depends(); | |
4db0c3c2 | 685 | if (READ_ONCE(page->mapping) != expected_mapping) |
4035c07a | 686 | goto stale; |
c8d6553b HD |
687 | |
688 | /* | |
689 | * We cannot do anything with the page while its refcount is 0. | |
690 | * Usually 0 means free, or tail of a higher-order page: in which | |
691 | * case this node is no longer referenced, and should be freed; | |
692 | * however, it might mean that the page is under page_freeze_refs(). | |
693 | * The __remove_mapping() case is easy, again the node is now stale; | |
694 | * but if page is swapcache in migrate_page_move_mapping(), it might | |
695 | * still be our page, in which case it's essential to keep the node. | |
696 | */ | |
697 | while (!get_page_unless_zero(page)) { | |
698 | /* | |
699 | * Another check for page->mapping != expected_mapping would | |
700 | * work here too. We have chosen the !PageSwapCache test to | |
701 | * optimize the common case, when the page is or is about to | |
702 | * be freed: PageSwapCache is cleared (under spin_lock_irq) | |
703 | * in the freeze_refs section of __remove_mapping(); but Anon | |
704 | * page->mapping reset to NULL later, in free_pages_prepare(). | |
705 | */ | |
706 | if (!PageSwapCache(page)) | |
707 | goto stale; | |
708 | cpu_relax(); | |
709 | } | |
710 | ||
4db0c3c2 | 711 | if (READ_ONCE(page->mapping) != expected_mapping) { |
4035c07a HD |
712 | put_page(page); |
713 | goto stale; | |
714 | } | |
c8d6553b | 715 | |
8fdb3dbf | 716 | if (lock_it) { |
8aafa6a4 | 717 | lock_page(page); |
4db0c3c2 | 718 | if (READ_ONCE(page->mapping) != expected_mapping) { |
8aafa6a4 HD |
719 | unlock_page(page); |
720 | put_page(page); | |
721 | goto stale; | |
722 | } | |
723 | } | |
4035c07a | 724 | return page; |
c8d6553b | 725 | |
4035c07a | 726 | stale: |
c8d6553b HD |
727 | /* |
728 | * We come here from above when page->mapping or !PageSwapCache | |
729 | * suggests that the node is stale; but it might be under migration. | |
730 | * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(), | |
731 | * before checking whether node->kpfn has been changed. | |
732 | */ | |
733 | smp_rmb(); | |
4db0c3c2 | 734 | if (READ_ONCE(stable_node->kpfn) != kpfn) |
c8d6553b | 735 | goto again; |
4035c07a HD |
736 | remove_node_from_stable_tree(stable_node); |
737 | return NULL; | |
738 | } | |
739 | ||
31dbd01f IE |
740 | /* |
741 | * Removing rmap_item from stable or unstable tree. | |
742 | * This function will clean the information from the stable/unstable tree. | |
743 | */ | |
744 | static void remove_rmap_item_from_tree(struct rmap_item *rmap_item) | |
745 | { | |
7b6ba2c7 HD |
746 | if (rmap_item->address & STABLE_FLAG) { |
747 | struct stable_node *stable_node; | |
5ad64688 | 748 | struct page *page; |
31dbd01f | 749 | |
7b6ba2c7 | 750 | stable_node = rmap_item->head; |
8aafa6a4 | 751 | page = get_ksm_page(stable_node, true); |
4035c07a HD |
752 | if (!page) |
753 | goto out; | |
5ad64688 | 754 | |
7b6ba2c7 | 755 | hlist_del(&rmap_item->hlist); |
4035c07a HD |
756 | unlock_page(page); |
757 | put_page(page); | |
08beca44 | 758 | |
98666f8a | 759 | if (!hlist_empty(&stable_node->hlist)) |
4035c07a HD |
760 | ksm_pages_sharing--; |
761 | else | |
7b6ba2c7 | 762 | ksm_pages_shared--; |
7253473c AA |
763 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
764 | stable_node->rmap_hlist_len--; | |
31dbd01f | 765 | |
9e60109f | 766 | put_anon_vma(rmap_item->anon_vma); |
93d17715 | 767 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 768 | |
7b6ba2c7 | 769 | } else if (rmap_item->address & UNSTABLE_FLAG) { |
31dbd01f IE |
770 | unsigned char age; |
771 | /* | |
9ba69294 | 772 | * Usually ksmd can and must skip the rb_erase, because |
31dbd01f | 773 | * root_unstable_tree was already reset to RB_ROOT. |
9ba69294 HD |
774 | * But be careful when an mm is exiting: do the rb_erase |
775 | * if this rmap_item was inserted by this scan, rather | |
776 | * than left over from before. | |
31dbd01f IE |
777 | */ |
778 | age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); | |
cd551f97 | 779 | BUG_ON(age > 1); |
31dbd01f | 780 | if (!age) |
90bd6fd3 | 781 | rb_erase(&rmap_item->node, |
ef53d16c | 782 | root_unstable_tree + NUMA(rmap_item->nid)); |
473b0ce4 | 783 | ksm_pages_unshared--; |
93d17715 | 784 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 785 | } |
4035c07a | 786 | out: |
31dbd01f IE |
787 | cond_resched(); /* we're called from many long loops */ |
788 | } | |
789 | ||
31dbd01f | 790 | static void remove_trailing_rmap_items(struct mm_slot *mm_slot, |
6514d511 | 791 | struct rmap_item **rmap_list) |
31dbd01f | 792 | { |
6514d511 HD |
793 | while (*rmap_list) { |
794 | struct rmap_item *rmap_item = *rmap_list; | |
795 | *rmap_list = rmap_item->rmap_list; | |
31dbd01f | 796 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
797 | free_rmap_item(rmap_item); |
798 | } | |
799 | } | |
800 | ||
801 | /* | |
e850dcf5 | 802 | * Though it's very tempting to unmerge rmap_items from stable tree rather |
31dbd01f IE |
803 | * than check every pte of a given vma, the locking doesn't quite work for |
804 | * that - an rmap_item is assigned to the stable tree after inserting ksm | |
805 | * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing | |
806 | * rmap_items from parent to child at fork time (so as not to waste time | |
807 | * if exit comes before the next scan reaches it). | |
81464e30 HD |
808 | * |
809 | * Similarly, although we'd like to remove rmap_items (so updating counts | |
810 | * and freeing memory) when unmerging an area, it's easier to leave that | |
811 | * to the next pass of ksmd - consider, for example, how ksmd might be | |
812 | * in cmp_and_merge_page on one of the rmap_items we would be removing. | |
31dbd01f | 813 | */ |
d952b791 HD |
814 | static int unmerge_ksm_pages(struct vm_area_struct *vma, |
815 | unsigned long start, unsigned long end) | |
31dbd01f IE |
816 | { |
817 | unsigned long addr; | |
d952b791 | 818 | int err = 0; |
31dbd01f | 819 | |
d952b791 | 820 | for (addr = start; addr < end && !err; addr += PAGE_SIZE) { |
9ba69294 HD |
821 | if (ksm_test_exit(vma->vm_mm)) |
822 | break; | |
d952b791 HD |
823 | if (signal_pending(current)) |
824 | err = -ERESTARTSYS; | |
825 | else | |
826 | err = break_ksm(vma, addr); | |
827 | } | |
828 | return err; | |
31dbd01f IE |
829 | } |
830 | ||
2ffd8679 HD |
831 | #ifdef CONFIG_SYSFS |
832 | /* | |
833 | * Only called through the sysfs control interface: | |
834 | */ | |
cbf86cfe HD |
835 | static int remove_stable_node(struct stable_node *stable_node) |
836 | { | |
837 | struct page *page; | |
838 | int err; | |
839 | ||
840 | page = get_ksm_page(stable_node, true); | |
841 | if (!page) { | |
842 | /* | |
843 | * get_ksm_page did remove_node_from_stable_tree itself. | |
844 | */ | |
845 | return 0; | |
846 | } | |
847 | ||
8fdb3dbf HD |
848 | if (WARN_ON_ONCE(page_mapped(page))) { |
849 | /* | |
850 | * This should not happen: but if it does, just refuse to let | |
851 | * merge_across_nodes be switched - there is no need to panic. | |
852 | */ | |
cbf86cfe | 853 | err = -EBUSY; |
8fdb3dbf | 854 | } else { |
cbf86cfe | 855 | /* |
8fdb3dbf HD |
856 | * The stable node did not yet appear stale to get_ksm_page(), |
857 | * since that allows for an unmapped ksm page to be recognized | |
858 | * right up until it is freed; but the node is safe to remove. | |
cbf86cfe HD |
859 | * This page might be in a pagevec waiting to be freed, |
860 | * or it might be PageSwapCache (perhaps under writeback), | |
861 | * or it might have been removed from swapcache a moment ago. | |
862 | */ | |
863 | set_page_stable_node(page, NULL); | |
864 | remove_node_from_stable_tree(stable_node); | |
865 | err = 0; | |
866 | } | |
867 | ||
868 | unlock_page(page); | |
869 | put_page(page); | |
870 | return err; | |
871 | } | |
872 | ||
7253473c AA |
873 | static int remove_stable_node_chain(struct stable_node *stable_node, |
874 | struct rb_root *root) | |
875 | { | |
876 | struct stable_node *dup; | |
877 | struct hlist_node *hlist_safe; | |
878 | ||
879 | if (!is_stable_node_chain(stable_node)) { | |
880 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
881 | if (remove_stable_node(stable_node)) | |
882 | return true; | |
883 | else | |
884 | return false; | |
885 | } | |
886 | ||
887 | hlist_for_each_entry_safe(dup, hlist_safe, | |
888 | &stable_node->hlist, hlist_dup) { | |
889 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
890 | if (remove_stable_node(dup)) | |
891 | return true; | |
892 | } | |
893 | BUG_ON(!hlist_empty(&stable_node->hlist)); | |
894 | free_stable_node_chain(stable_node, root); | |
895 | return false; | |
896 | } | |
897 | ||
cbf86cfe HD |
898 | static int remove_all_stable_nodes(void) |
899 | { | |
03640418 | 900 | struct stable_node *stable_node, *next; |
cbf86cfe HD |
901 | int nid; |
902 | int err = 0; | |
903 | ||
ef53d16c | 904 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
cbf86cfe HD |
905 | while (root_stable_tree[nid].rb_node) { |
906 | stable_node = rb_entry(root_stable_tree[nid].rb_node, | |
907 | struct stable_node, node); | |
7253473c AA |
908 | if (remove_stable_node_chain(stable_node, |
909 | root_stable_tree + nid)) { | |
cbf86cfe HD |
910 | err = -EBUSY; |
911 | break; /* proceed to next nid */ | |
912 | } | |
913 | cond_resched(); | |
914 | } | |
915 | } | |
03640418 | 916 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
917 | if (remove_stable_node(stable_node)) |
918 | err = -EBUSY; | |
919 | cond_resched(); | |
920 | } | |
cbf86cfe HD |
921 | return err; |
922 | } | |
923 | ||
d952b791 | 924 | static int unmerge_and_remove_all_rmap_items(void) |
31dbd01f IE |
925 | { |
926 | struct mm_slot *mm_slot; | |
927 | struct mm_struct *mm; | |
928 | struct vm_area_struct *vma; | |
d952b791 HD |
929 | int err = 0; |
930 | ||
931 | spin_lock(&ksm_mmlist_lock); | |
9ba69294 | 932 | ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next, |
d952b791 HD |
933 | struct mm_slot, mm_list); |
934 | spin_unlock(&ksm_mmlist_lock); | |
31dbd01f | 935 | |
9ba69294 HD |
936 | for (mm_slot = ksm_scan.mm_slot; |
937 | mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) { | |
31dbd01f IE |
938 | mm = mm_slot->mm; |
939 | down_read(&mm->mmap_sem); | |
940 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
9ba69294 HD |
941 | if (ksm_test_exit(mm)) |
942 | break; | |
31dbd01f IE |
943 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
944 | continue; | |
d952b791 HD |
945 | err = unmerge_ksm_pages(vma, |
946 | vma->vm_start, vma->vm_end); | |
9ba69294 HD |
947 | if (err) |
948 | goto error; | |
31dbd01f | 949 | } |
9ba69294 | 950 | |
6514d511 | 951 | remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list); |
7496fea9 | 952 | up_read(&mm->mmap_sem); |
d952b791 HD |
953 | |
954 | spin_lock(&ksm_mmlist_lock); | |
9ba69294 | 955 | ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next, |
d952b791 | 956 | struct mm_slot, mm_list); |
9ba69294 | 957 | if (ksm_test_exit(mm)) { |
4ca3a69b | 958 | hash_del(&mm_slot->link); |
9ba69294 HD |
959 | list_del(&mm_slot->mm_list); |
960 | spin_unlock(&ksm_mmlist_lock); | |
961 | ||
962 | free_mm_slot(mm_slot); | |
963 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
9ba69294 | 964 | mmdrop(mm); |
7496fea9 | 965 | } else |
9ba69294 | 966 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f IE |
967 | } |
968 | ||
cbf86cfe HD |
969 | /* Clean up stable nodes, but don't worry if some are still busy */ |
970 | remove_all_stable_nodes(); | |
d952b791 | 971 | ksm_scan.seqnr = 0; |
9ba69294 HD |
972 | return 0; |
973 | ||
974 | error: | |
975 | up_read(&mm->mmap_sem); | |
31dbd01f | 976 | spin_lock(&ksm_mmlist_lock); |
d952b791 | 977 | ksm_scan.mm_slot = &ksm_mm_head; |
31dbd01f | 978 | spin_unlock(&ksm_mmlist_lock); |
d952b791 | 979 | return err; |
31dbd01f | 980 | } |
2ffd8679 | 981 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 982 | |
31dbd01f IE |
983 | static u32 calc_checksum(struct page *page) |
984 | { | |
985 | u32 checksum; | |
9b04c5fe | 986 | void *addr = kmap_atomic(page); |
31dbd01f | 987 | checksum = jhash2(addr, PAGE_SIZE / 4, 17); |
9b04c5fe | 988 | kunmap_atomic(addr); |
31dbd01f IE |
989 | return checksum; |
990 | } | |
991 | ||
992 | static int memcmp_pages(struct page *page1, struct page *page2) | |
993 | { | |
994 | char *addr1, *addr2; | |
995 | int ret; | |
996 | ||
9b04c5fe CW |
997 | addr1 = kmap_atomic(page1); |
998 | addr2 = kmap_atomic(page2); | |
31dbd01f | 999 | ret = memcmp(addr1, addr2, PAGE_SIZE); |
9b04c5fe CW |
1000 | kunmap_atomic(addr2); |
1001 | kunmap_atomic(addr1); | |
31dbd01f IE |
1002 | return ret; |
1003 | } | |
1004 | ||
1005 | static inline int pages_identical(struct page *page1, struct page *page2) | |
1006 | { | |
1007 | return !memcmp_pages(page1, page2); | |
1008 | } | |
1009 | ||
1010 | static int write_protect_page(struct vm_area_struct *vma, struct page *page, | |
1011 | pte_t *orig_pte) | |
1012 | { | |
1013 | struct mm_struct *mm = vma->vm_mm; | |
421ec457 SF |
1014 | unsigned long addr; |
1015 | pte_t *ptep; | |
1016 | spinlock_t *ptl; | |
31dbd01f IE |
1017 | int swapped; |
1018 | int err = -EFAULT; | |
6bdb913f HE |
1019 | unsigned long mmun_start; /* For mmu_notifiers */ |
1020 | unsigned long mmun_end; /* For mmu_notifiers */ | |
31dbd01f | 1021 | |
421ec457 SF |
1022 | addr = page_address_in_vma(page, vma); |
1023 | if (addr == -EFAULT) | |
31dbd01f IE |
1024 | goto out; |
1025 | ||
29ad768c | 1026 | BUG_ON(PageTransCompound(page)); |
6bdb913f | 1027 | |
421ec457 SF |
1028 | mmun_start = addr; |
1029 | mmun_end = addr + PAGE_SIZE; | |
6bdb913f HE |
1030 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
1031 | ||
421ec457 SF |
1032 | ptep = page_check_address(page, mm, addr, &ptl, 0); |
1033 | if (!ptep) | |
6bdb913f | 1034 | goto out_mn; |
31dbd01f | 1035 | |
737f3a90 AK |
1036 | if (pte_write(*ptep) || pte_dirty(*ptep) || |
1037 | (pte_protnone(*ptep) && pte_savedwrite(*ptep))) { | |
31dbd01f IE |
1038 | pte_t entry; |
1039 | ||
1040 | swapped = PageSwapCache(page); | |
421ec457 | 1041 | flush_cache_page(vma, addr, page_to_pfn(page)); |
31dbd01f | 1042 | /* |
25985edc | 1043 | * Ok this is tricky, when get_user_pages_fast() run it doesn't |
31dbd01f IE |
1044 | * take any lock, therefore the check that we are going to make |
1045 | * with the pagecount against the mapcount is racey and | |
1046 | * O_DIRECT can happen right after the check. | |
1047 | * So we clear the pte and flush the tlb before the check | |
1048 | * this assure us that no O_DIRECT can happen after the check | |
1049 | * or in the middle of the check. | |
1050 | */ | |
421ec457 | 1051 | entry = ptep_clear_flush_notify(vma, addr, ptep); |
31dbd01f IE |
1052 | /* |
1053 | * Check that no O_DIRECT or similar I/O is in progress on the | |
1054 | * page | |
1055 | */ | |
31e855ea | 1056 | if (page_mapcount(page) + 1 + swapped != page_count(page)) { |
421ec457 | 1057 | set_pte_at(mm, addr, ptep, entry); |
31dbd01f IE |
1058 | goto out_unlock; |
1059 | } | |
4e31635c HD |
1060 | if (pte_dirty(entry)) |
1061 | set_page_dirty(page); | |
737f3a90 AK |
1062 | |
1063 | if (pte_protnone(entry)) | |
1064 | entry = pte_mkclean(pte_clear_savedwrite(entry)); | |
1065 | else | |
1066 | entry = pte_mkclean(pte_wrprotect(entry)); | |
421ec457 | 1067 | set_pte_at_notify(mm, addr, ptep, entry); |
31dbd01f | 1068 | } |
421ec457 | 1069 | *orig_pte = *ptep; |
31dbd01f IE |
1070 | err = 0; |
1071 | ||
1072 | out_unlock: | |
421ec457 | 1073 | pte_unmap_unlock(ptep, ptl); |
6bdb913f HE |
1074 | out_mn: |
1075 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
31dbd01f IE |
1076 | out: |
1077 | return err; | |
1078 | } | |
1079 | ||
1080 | /** | |
1081 | * replace_page - replace page in vma by new ksm page | |
8dd3557a HD |
1082 | * @vma: vma that holds the pte pointing to page |
1083 | * @page: the page we are replacing by kpage | |
1084 | * @kpage: the ksm page we replace page by | |
31dbd01f IE |
1085 | * @orig_pte: the original value of the pte |
1086 | * | |
1087 | * Returns 0 on success, -EFAULT on failure. | |
1088 | */ | |
8dd3557a HD |
1089 | static int replace_page(struct vm_area_struct *vma, struct page *page, |
1090 | struct page *kpage, pte_t orig_pte) | |
31dbd01f IE |
1091 | { |
1092 | struct mm_struct *mm = vma->vm_mm; | |
31dbd01f IE |
1093 | pmd_t *pmd; |
1094 | pte_t *ptep; | |
fa2aa0fc | 1095 | pte_t newpte; |
31dbd01f IE |
1096 | spinlock_t *ptl; |
1097 | unsigned long addr; | |
31dbd01f | 1098 | int err = -EFAULT; |
6bdb913f HE |
1099 | unsigned long mmun_start; /* For mmu_notifiers */ |
1100 | unsigned long mmun_end; /* For mmu_notifiers */ | |
31dbd01f | 1101 | |
8dd3557a | 1102 | addr = page_address_in_vma(page, vma); |
31dbd01f IE |
1103 | if (addr == -EFAULT) |
1104 | goto out; | |
1105 | ||
6219049a BL |
1106 | pmd = mm_find_pmd(mm, addr); |
1107 | if (!pmd) | |
31dbd01f | 1108 | goto out; |
31dbd01f | 1109 | |
6bdb913f HE |
1110 | mmun_start = addr; |
1111 | mmun_end = addr + PAGE_SIZE; | |
1112 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
1113 | ||
31dbd01f IE |
1114 | ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); |
1115 | if (!pte_same(*ptep, orig_pte)) { | |
1116 | pte_unmap_unlock(ptep, ptl); | |
6bdb913f | 1117 | goto out_mn; |
31dbd01f IE |
1118 | } |
1119 | ||
fa2aa0fc CI |
1120 | /* |
1121 | * No need to check ksm_use_zero_pages here: we can only have a | |
1122 | * zero_page here if ksm_use_zero_pages was enabled alreaady. | |
1123 | */ | |
1124 | if (!is_zero_pfn(page_to_pfn(kpage))) { | |
1125 | get_page(kpage); | |
1126 | page_add_anon_rmap(kpage, vma, addr, false); | |
1127 | newpte = mk_pte(kpage, vma->vm_page_prot); | |
1128 | } else { | |
1129 | newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage), | |
1130 | vma->vm_page_prot)); | |
1131 | } | |
31dbd01f IE |
1132 | |
1133 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
34ee645e | 1134 | ptep_clear_flush_notify(vma, addr, ptep); |
fa2aa0fc | 1135 | set_pte_at_notify(mm, addr, ptep, newpte); |
31dbd01f | 1136 | |
d281ee61 | 1137 | page_remove_rmap(page, false); |
ae52a2ad HD |
1138 | if (!page_mapped(page)) |
1139 | try_to_free_swap(page); | |
8dd3557a | 1140 | put_page(page); |
31dbd01f IE |
1141 | |
1142 | pte_unmap_unlock(ptep, ptl); | |
1143 | err = 0; | |
6bdb913f HE |
1144 | out_mn: |
1145 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
31dbd01f IE |
1146 | out: |
1147 | return err; | |
1148 | } | |
1149 | ||
1150 | /* | |
1151 | * try_to_merge_one_page - take two pages and merge them into one | |
8dd3557a HD |
1152 | * @vma: the vma that holds the pte pointing to page |
1153 | * @page: the PageAnon page that we want to replace with kpage | |
80e14822 HD |
1154 | * @kpage: the PageKsm page that we want to map instead of page, |
1155 | * or NULL the first time when we want to use page as kpage. | |
31dbd01f IE |
1156 | * |
1157 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
1158 | */ | |
1159 | static int try_to_merge_one_page(struct vm_area_struct *vma, | |
8dd3557a | 1160 | struct page *page, struct page *kpage) |
31dbd01f IE |
1161 | { |
1162 | pte_t orig_pte = __pte(0); | |
1163 | int err = -EFAULT; | |
1164 | ||
db114b83 HD |
1165 | if (page == kpage) /* ksm page forked */ |
1166 | return 0; | |
1167 | ||
8dd3557a | 1168 | if (!PageAnon(page)) |
31dbd01f IE |
1169 | goto out; |
1170 | ||
31dbd01f IE |
1171 | /* |
1172 | * We need the page lock to read a stable PageSwapCache in | |
1173 | * write_protect_page(). We use trylock_page() instead of | |
1174 | * lock_page() because we don't want to wait here - we | |
1175 | * prefer to continue scanning and merging different pages, | |
1176 | * then come back to this page when it is unlocked. | |
1177 | */ | |
8dd3557a | 1178 | if (!trylock_page(page)) |
31e855ea | 1179 | goto out; |
f765f540 KS |
1180 | |
1181 | if (PageTransCompound(page)) { | |
1182 | err = split_huge_page(page); | |
1183 | if (err) | |
1184 | goto out_unlock; | |
1185 | } | |
1186 | ||
31dbd01f IE |
1187 | /* |
1188 | * If this anonymous page is mapped only here, its pte may need | |
1189 | * to be write-protected. If it's mapped elsewhere, all of its | |
1190 | * ptes are necessarily already write-protected. But in either | |
1191 | * case, we need to lock and check page_count is not raised. | |
1192 | */ | |
80e14822 HD |
1193 | if (write_protect_page(vma, page, &orig_pte) == 0) { |
1194 | if (!kpage) { | |
1195 | /* | |
1196 | * While we hold page lock, upgrade page from | |
1197 | * PageAnon+anon_vma to PageKsm+NULL stable_node: | |
1198 | * stable_tree_insert() will update stable_node. | |
1199 | */ | |
1200 | set_page_stable_node(page, NULL); | |
1201 | mark_page_accessed(page); | |
337ed7eb MK |
1202 | /* |
1203 | * Page reclaim just frees a clean page with no dirty | |
1204 | * ptes: make sure that the ksm page would be swapped. | |
1205 | */ | |
1206 | if (!PageDirty(page)) | |
1207 | SetPageDirty(page); | |
80e14822 HD |
1208 | err = 0; |
1209 | } else if (pages_identical(page, kpage)) | |
1210 | err = replace_page(vma, page, kpage, orig_pte); | |
1211 | } | |
31dbd01f | 1212 | |
80e14822 | 1213 | if ((vma->vm_flags & VM_LOCKED) && kpage && !err) { |
73848b46 | 1214 | munlock_vma_page(page); |
5ad64688 HD |
1215 | if (!PageMlocked(kpage)) { |
1216 | unlock_page(page); | |
5ad64688 HD |
1217 | lock_page(kpage); |
1218 | mlock_vma_page(kpage); | |
1219 | page = kpage; /* for final unlock */ | |
1220 | } | |
1221 | } | |
73848b46 | 1222 | |
f765f540 | 1223 | out_unlock: |
8dd3557a | 1224 | unlock_page(page); |
31dbd01f IE |
1225 | out: |
1226 | return err; | |
1227 | } | |
1228 | ||
81464e30 HD |
1229 | /* |
1230 | * try_to_merge_with_ksm_page - like try_to_merge_two_pages, | |
1231 | * but no new kernel page is allocated: kpage must already be a ksm page. | |
8dd3557a HD |
1232 | * |
1233 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
81464e30 | 1234 | */ |
8dd3557a HD |
1235 | static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item, |
1236 | struct page *page, struct page *kpage) | |
81464e30 | 1237 | { |
8dd3557a | 1238 | struct mm_struct *mm = rmap_item->mm; |
81464e30 HD |
1239 | struct vm_area_struct *vma; |
1240 | int err = -EFAULT; | |
1241 | ||
8dd3557a | 1242 | down_read(&mm->mmap_sem); |
85c6e8dd AA |
1243 | vma = find_mergeable_vma(mm, rmap_item->address); |
1244 | if (!vma) | |
81464e30 HD |
1245 | goto out; |
1246 | ||
8dd3557a | 1247 | err = try_to_merge_one_page(vma, page, kpage); |
db114b83 HD |
1248 | if (err) |
1249 | goto out; | |
1250 | ||
bc56620b HD |
1251 | /* Unstable nid is in union with stable anon_vma: remove first */ |
1252 | remove_rmap_item_from_tree(rmap_item); | |
1253 | ||
db114b83 | 1254 | /* Must get reference to anon_vma while still holding mmap_sem */ |
9e60109f PZ |
1255 | rmap_item->anon_vma = vma->anon_vma; |
1256 | get_anon_vma(vma->anon_vma); | |
81464e30 | 1257 | out: |
8dd3557a | 1258 | up_read(&mm->mmap_sem); |
81464e30 HD |
1259 | return err; |
1260 | } | |
1261 | ||
31dbd01f IE |
1262 | /* |
1263 | * try_to_merge_two_pages - take two identical pages and prepare them | |
1264 | * to be merged into one page. | |
1265 | * | |
8dd3557a HD |
1266 | * This function returns the kpage if we successfully merged two identical |
1267 | * pages into one ksm page, NULL otherwise. | |
31dbd01f | 1268 | * |
80e14822 | 1269 | * Note that this function upgrades page to ksm page: if one of the pages |
31dbd01f IE |
1270 | * is already a ksm page, try_to_merge_with_ksm_page should be used. |
1271 | */ | |
8dd3557a HD |
1272 | static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item, |
1273 | struct page *page, | |
1274 | struct rmap_item *tree_rmap_item, | |
1275 | struct page *tree_page) | |
31dbd01f | 1276 | { |
80e14822 | 1277 | int err; |
31dbd01f | 1278 | |
80e14822 | 1279 | err = try_to_merge_with_ksm_page(rmap_item, page, NULL); |
31dbd01f | 1280 | if (!err) { |
8dd3557a | 1281 | err = try_to_merge_with_ksm_page(tree_rmap_item, |
80e14822 | 1282 | tree_page, page); |
31dbd01f | 1283 | /* |
81464e30 HD |
1284 | * If that fails, we have a ksm page with only one pte |
1285 | * pointing to it: so break it. | |
31dbd01f | 1286 | */ |
4035c07a | 1287 | if (err) |
8dd3557a | 1288 | break_cow(rmap_item); |
31dbd01f | 1289 | } |
80e14822 | 1290 | return err ? NULL : page; |
31dbd01f IE |
1291 | } |
1292 | ||
7253473c AA |
1293 | static __always_inline |
1294 | bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset) | |
1295 | { | |
1296 | VM_BUG_ON(stable_node->rmap_hlist_len < 0); | |
1297 | /* | |
1298 | * Check that at least one mapping still exists, otherwise | |
1299 | * there's no much point to merge and share with this | |
1300 | * stable_node, as the underlying tree_page of the other | |
1301 | * sharer is going to be freed soon. | |
1302 | */ | |
1303 | return stable_node->rmap_hlist_len && | |
1304 | stable_node->rmap_hlist_len + offset < ksm_max_page_sharing; | |
1305 | } | |
1306 | ||
1307 | static __always_inline | |
1308 | bool is_page_sharing_candidate(struct stable_node *stable_node) | |
1309 | { | |
1310 | return __is_page_sharing_candidate(stable_node, 0); | |
1311 | } | |
1312 | ||
6457f49b AA |
1313 | struct page *stable_node_dup(struct stable_node **_stable_node_dup, |
1314 | struct stable_node **_stable_node, | |
1315 | struct rb_root *root, | |
1316 | bool prune_stale_stable_nodes) | |
7253473c | 1317 | { |
b8d20a18 | 1318 | struct stable_node *dup, *found = NULL, *stable_node = *_stable_node; |
7253473c | 1319 | struct hlist_node *hlist_safe; |
6457f49b | 1320 | struct page *_tree_page, *tree_page = NULL; |
7253473c AA |
1321 | int nr = 0; |
1322 | int found_rmap_hlist_len; | |
1323 | ||
1324 | if (!prune_stale_stable_nodes || | |
1325 | time_before(jiffies, stable_node->chain_prune_time + | |
1326 | msecs_to_jiffies( | |
1327 | ksm_stable_node_chains_prune_millisecs))) | |
1328 | prune_stale_stable_nodes = false; | |
1329 | else | |
1330 | stable_node->chain_prune_time = jiffies; | |
1331 | ||
1332 | hlist_for_each_entry_safe(dup, hlist_safe, | |
1333 | &stable_node->hlist, hlist_dup) { | |
1334 | cond_resched(); | |
1335 | /* | |
1336 | * We must walk all stable_node_dup to prune the stale | |
1337 | * stable nodes during lookup. | |
1338 | * | |
1339 | * get_ksm_page can drop the nodes from the | |
1340 | * stable_node->hlist if they point to freed pages | |
1341 | * (that's why we do a _safe walk). The "dup" | |
1342 | * stable_node parameter itself will be freed from | |
1343 | * under us if it returns NULL. | |
1344 | */ | |
1345 | _tree_page = get_ksm_page(dup, false); | |
1346 | if (!_tree_page) | |
1347 | continue; | |
1348 | nr += 1; | |
1349 | if (is_page_sharing_candidate(dup)) { | |
1350 | if (!found || | |
1351 | dup->rmap_hlist_len > found_rmap_hlist_len) { | |
1352 | if (found) | |
6457f49b | 1353 | put_page(tree_page); |
7253473c AA |
1354 | found = dup; |
1355 | found_rmap_hlist_len = found->rmap_hlist_len; | |
6457f49b | 1356 | tree_page = _tree_page; |
7253473c | 1357 | |
6457f49b | 1358 | /* skip put_page for found dup */ |
7253473c AA |
1359 | if (!prune_stale_stable_nodes) |
1360 | break; | |
7253473c AA |
1361 | continue; |
1362 | } | |
1363 | } | |
1364 | put_page(_tree_page); | |
1365 | } | |
1366 | ||
8832c76f AA |
1367 | if (found) { |
1368 | /* | |
1369 | * nr is counting all dups in the chain only if | |
1370 | * prune_stale_stable_nodes is true, otherwise we may | |
1371 | * break the loop at nr == 1 even if there are | |
1372 | * multiple entries. | |
1373 | */ | |
1374 | if (prune_stale_stable_nodes && nr == 1) { | |
7253473c AA |
1375 | /* |
1376 | * If there's not just one entry it would | |
1377 | * corrupt memory, better BUG_ON. In KSM | |
1378 | * context with no lock held it's not even | |
1379 | * fatal. | |
1380 | */ | |
1381 | BUG_ON(stable_node->hlist.first->next); | |
1382 | ||
1383 | /* | |
1384 | * There's just one entry and it is below the | |
1385 | * deduplication limit so drop the chain. | |
1386 | */ | |
1387 | rb_replace_node(&stable_node->node, &found->node, | |
1388 | root); | |
1389 | free_stable_node(stable_node); | |
1390 | ksm_stable_node_chains--; | |
1391 | ksm_stable_node_dups--; | |
b8d20a18 | 1392 | /* |
3bef4484 AA |
1393 | * NOTE: the caller depends on the stable_node |
1394 | * to be equal to stable_node_dup if the chain | |
1395 | * was collapsed. | |
b8d20a18 | 1396 | */ |
3bef4484 AA |
1397 | *_stable_node = found; |
1398 | /* | |
1399 | * Just for robustneess as stable_node is | |
1400 | * otherwise left as a stable pointer, the | |
1401 | * compiler shall optimize it away at build | |
1402 | * time. | |
1403 | */ | |
1404 | stable_node = NULL; | |
8832c76f AA |
1405 | } else if (stable_node->hlist.first != &found->hlist_dup && |
1406 | __is_page_sharing_candidate(found, 1)) { | |
7253473c | 1407 | /* |
8832c76f AA |
1408 | * If the found stable_node dup can accept one |
1409 | * more future merge (in addition to the one | |
1410 | * that is underway) and is not at the head of | |
1411 | * the chain, put it there so next search will | |
1412 | * be quicker in the !prune_stale_stable_nodes | |
1413 | * case. | |
1414 | * | |
1415 | * NOTE: it would be inaccurate to use nr > 1 | |
1416 | * instead of checking the hlist.first pointer | |
1417 | * directly, because in the | |
1418 | * prune_stale_stable_nodes case "nr" isn't | |
1419 | * the position of the found dup in the chain, | |
1420 | * but the total number of dups in the chain. | |
7253473c AA |
1421 | */ |
1422 | hlist_del(&found->hlist_dup); | |
1423 | hlist_add_head(&found->hlist_dup, | |
1424 | &stable_node->hlist); | |
1425 | } | |
1426 | } | |
1427 | ||
6457f49b AA |
1428 | *_stable_node_dup = found; |
1429 | return tree_page; | |
7253473c AA |
1430 | } |
1431 | ||
1432 | static struct stable_node *stable_node_dup_any(struct stable_node *stable_node, | |
1433 | struct rb_root *root) | |
1434 | { | |
1435 | if (!is_stable_node_chain(stable_node)) | |
1436 | return stable_node; | |
1437 | if (hlist_empty(&stable_node->hlist)) { | |
1438 | free_stable_node_chain(stable_node, root); | |
1439 | return NULL; | |
1440 | } | |
1441 | return hlist_entry(stable_node->hlist.first, | |
1442 | typeof(*stable_node), hlist_dup); | |
1443 | } | |
1444 | ||
6457f49b AA |
1445 | /* |
1446 | * Like for get_ksm_page, this function can free the *_stable_node and | |
1447 | * *_stable_node_dup if the returned tree_page is NULL. | |
1448 | * | |
1449 | * It can also free and overwrite *_stable_node with the found | |
1450 | * stable_node_dup if the chain is collapsed (in which case | |
1451 | * *_stable_node will be equal to *_stable_node_dup like if the chain | |
1452 | * never existed). It's up to the caller to verify tree_page is not | |
1453 | * NULL before dereferencing *_stable_node or *_stable_node_dup. | |
1454 | * | |
1455 | * *_stable_node_dup is really a second output parameter of this | |
1456 | * function and will be overwritten in all cases, the caller doesn't | |
1457 | * need to initialize it. | |
1458 | */ | |
1459 | static struct page *__stable_node_chain(struct stable_node **_stable_node_dup, | |
1460 | struct stable_node **_stable_node, | |
1461 | struct rb_root *root, | |
1462 | bool prune_stale_stable_nodes) | |
7253473c | 1463 | { |
b8d20a18 | 1464 | struct stable_node *stable_node = *_stable_node; |
7253473c AA |
1465 | if (!is_stable_node_chain(stable_node)) { |
1466 | if (is_page_sharing_candidate(stable_node)) { | |
6457f49b AA |
1467 | *_stable_node_dup = stable_node; |
1468 | return get_ksm_page(stable_node, false); | |
7253473c | 1469 | } |
6457f49b AA |
1470 | /* |
1471 | * _stable_node_dup set to NULL means the stable_node | |
1472 | * reached the ksm_max_page_sharing limit. | |
1473 | */ | |
1474 | *_stable_node_dup = NULL; | |
7253473c AA |
1475 | return NULL; |
1476 | } | |
6457f49b | 1477 | return stable_node_dup(_stable_node_dup, _stable_node, root, |
7253473c AA |
1478 | prune_stale_stable_nodes); |
1479 | } | |
1480 | ||
6457f49b AA |
1481 | static __always_inline struct page *chain_prune(struct stable_node **s_n_d, |
1482 | struct stable_node **s_n, | |
1483 | struct rb_root *root) | |
7253473c | 1484 | { |
6457f49b | 1485 | return __stable_node_chain(s_n_d, s_n, root, true); |
7253473c AA |
1486 | } |
1487 | ||
6457f49b AA |
1488 | static __always_inline struct page *chain(struct stable_node **s_n_d, |
1489 | struct stable_node *s_n, | |
1490 | struct rb_root *root) | |
7253473c | 1491 | { |
6457f49b AA |
1492 | struct stable_node *old_stable_node = s_n; |
1493 | struct page *tree_page; | |
1494 | ||
1495 | tree_page = __stable_node_chain(s_n_d, &s_n, root, false); | |
1496 | /* not pruning dups so s_n cannot have changed */ | |
1497 | VM_BUG_ON(s_n != old_stable_node); | |
1498 | return tree_page; | |
7253473c AA |
1499 | } |
1500 | ||
31dbd01f | 1501 | /* |
8dd3557a | 1502 | * stable_tree_search - search for page inside the stable tree |
31dbd01f IE |
1503 | * |
1504 | * This function checks if there is a page inside the stable tree | |
1505 | * with identical content to the page that we are scanning right now. | |
1506 | * | |
7b6ba2c7 | 1507 | * This function returns the stable tree node of identical content if found, |
31dbd01f IE |
1508 | * NULL otherwise. |
1509 | */ | |
62b61f61 | 1510 | static struct page *stable_tree_search(struct page *page) |
31dbd01f | 1511 | { |
90bd6fd3 | 1512 | int nid; |
ef53d16c | 1513 | struct rb_root *root; |
4146d2d6 HD |
1514 | struct rb_node **new; |
1515 | struct rb_node *parent; | |
7253473c | 1516 | struct stable_node *stable_node, *stable_node_dup, *stable_node_any; |
4146d2d6 | 1517 | struct stable_node *page_node; |
31dbd01f | 1518 | |
4146d2d6 HD |
1519 | page_node = page_stable_node(page); |
1520 | if (page_node && page_node->head != &migrate_nodes) { | |
1521 | /* ksm page forked */ | |
08beca44 | 1522 | get_page(page); |
62b61f61 | 1523 | return page; |
08beca44 HD |
1524 | } |
1525 | ||
90bd6fd3 | 1526 | nid = get_kpfn_nid(page_to_pfn(page)); |
ef53d16c | 1527 | root = root_stable_tree + nid; |
4146d2d6 | 1528 | again: |
ef53d16c | 1529 | new = &root->rb_node; |
4146d2d6 | 1530 | parent = NULL; |
90bd6fd3 | 1531 | |
4146d2d6 | 1532 | while (*new) { |
4035c07a | 1533 | struct page *tree_page; |
31dbd01f IE |
1534 | int ret; |
1535 | ||
08beca44 | 1536 | cond_resched(); |
4146d2d6 | 1537 | stable_node = rb_entry(*new, struct stable_node, node); |
7253473c | 1538 | stable_node_any = NULL; |
6457f49b | 1539 | tree_page = chain_prune(&stable_node_dup, &stable_node, root); |
b8d20a18 AA |
1540 | /* |
1541 | * NOTE: stable_node may have been freed by | |
1542 | * chain_prune() if the returned stable_node_dup is | |
1543 | * not NULL. stable_node_dup may have been inserted in | |
1544 | * the rbtree instead as a regular stable_node (in | |
1545 | * order to collapse the stable_node chain if a single | |
3bef4484 AA |
1546 | * stable_node dup was found in it). In such case the |
1547 | * stable_node is overwritten by the calleee to point | |
1548 | * to the stable_node_dup that was collapsed in the | |
1549 | * stable rbtree and stable_node will be equal to | |
1550 | * stable_node_dup like if the chain never existed. | |
b8d20a18 | 1551 | */ |
7253473c AA |
1552 | if (!stable_node_dup) { |
1553 | /* | |
1554 | * Either all stable_node dups were full in | |
1555 | * this stable_node chain, or this chain was | |
1556 | * empty and should be rb_erased. | |
1557 | */ | |
1558 | stable_node_any = stable_node_dup_any(stable_node, | |
1559 | root); | |
1560 | if (!stable_node_any) { | |
1561 | /* rb_erase just run */ | |
1562 | goto again; | |
1563 | } | |
1564 | /* | |
1565 | * Take any of the stable_node dups page of | |
1566 | * this stable_node chain to let the tree walk | |
1567 | * continue. All KSM pages belonging to the | |
1568 | * stable_node dups in a stable_node chain | |
1569 | * have the same content and they're | |
1570 | * wrprotected at all times. Any will work | |
1571 | * fine to continue the walk. | |
1572 | */ | |
1573 | tree_page = get_ksm_page(stable_node_any, false); | |
1574 | } | |
1575 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1576 | if (!tree_page) { |
1577 | /* | |
1578 | * If we walked over a stale stable_node, | |
1579 | * get_ksm_page() will call rb_erase() and it | |
1580 | * may rebalance the tree from under us. So | |
1581 | * restart the search from scratch. Returning | |
1582 | * NULL would be safe too, but we'd generate | |
1583 | * false negative insertions just because some | |
1584 | * stable_node was stale. | |
1585 | */ | |
1586 | goto again; | |
1587 | } | |
31dbd01f | 1588 | |
4035c07a | 1589 | ret = memcmp_pages(page, tree_page); |
c8d6553b | 1590 | put_page(tree_page); |
31dbd01f | 1591 | |
4146d2d6 | 1592 | parent = *new; |
c8d6553b | 1593 | if (ret < 0) |
4146d2d6 | 1594 | new = &parent->rb_left; |
c8d6553b | 1595 | else if (ret > 0) |
4146d2d6 | 1596 | new = &parent->rb_right; |
c8d6553b | 1597 | else { |
7253473c AA |
1598 | if (page_node) { |
1599 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1600 | /* | |
1601 | * Test if the migrated page should be merged | |
1602 | * into a stable node dup. If the mapcount is | |
1603 | * 1 we can migrate it with another KSM page | |
1604 | * without adding it to the chain. | |
1605 | */ | |
1606 | if (page_mapcount(page) > 1) | |
1607 | goto chain_append; | |
1608 | } | |
1609 | ||
1610 | if (!stable_node_dup) { | |
1611 | /* | |
1612 | * If the stable_node is a chain and | |
1613 | * we got a payload match in memcmp | |
1614 | * but we cannot merge the scanned | |
1615 | * page in any of the existing | |
1616 | * stable_node dups because they're | |
1617 | * all full, we need to wait the | |
1618 | * scanned page to find itself a match | |
1619 | * in the unstable tree to create a | |
1620 | * brand new KSM page to add later to | |
1621 | * the dups of this stable_node. | |
1622 | */ | |
1623 | return NULL; | |
1624 | } | |
1625 | ||
c8d6553b HD |
1626 | /* |
1627 | * Lock and unlock the stable_node's page (which | |
1628 | * might already have been migrated) so that page | |
1629 | * migration is sure to notice its raised count. | |
1630 | * It would be more elegant to return stable_node | |
1631 | * than kpage, but that involves more changes. | |
1632 | */ | |
7253473c AA |
1633 | tree_page = get_ksm_page(stable_node_dup, true); |
1634 | if (unlikely(!tree_page)) | |
1635 | /* | |
1636 | * The tree may have been rebalanced, | |
1637 | * so re-evaluate parent and new. | |
1638 | */ | |
4146d2d6 | 1639 | goto again; |
7253473c AA |
1640 | unlock_page(tree_page); |
1641 | ||
1642 | if (get_kpfn_nid(stable_node_dup->kpfn) != | |
1643 | NUMA(stable_node_dup->nid)) { | |
1644 | put_page(tree_page); | |
1645 | goto replace; | |
1646 | } | |
1647 | return tree_page; | |
c8d6553b | 1648 | } |
31dbd01f IE |
1649 | } |
1650 | ||
4146d2d6 HD |
1651 | if (!page_node) |
1652 | return NULL; | |
1653 | ||
1654 | list_del(&page_node->list); | |
1655 | DO_NUMA(page_node->nid = nid); | |
1656 | rb_link_node(&page_node->node, parent, new); | |
ef53d16c | 1657 | rb_insert_color(&page_node->node, root); |
7253473c AA |
1658 | out: |
1659 | if (is_page_sharing_candidate(page_node)) { | |
1660 | get_page(page); | |
1661 | return page; | |
1662 | } else | |
1663 | return NULL; | |
4146d2d6 HD |
1664 | |
1665 | replace: | |
b8d20a18 AA |
1666 | /* |
1667 | * If stable_node was a chain and chain_prune collapsed it, | |
3bef4484 AA |
1668 | * stable_node has been updated to be the new regular |
1669 | * stable_node. A collapse of the chain is indistinguishable | |
1670 | * from the case there was no chain in the stable | |
1671 | * rbtree. Otherwise stable_node is the chain and | |
1672 | * stable_node_dup is the dup to replace. | |
b8d20a18 | 1673 | */ |
3bef4484 | 1674 | if (stable_node_dup == stable_node) { |
b8d20a18 AA |
1675 | VM_BUG_ON(is_stable_node_chain(stable_node_dup)); |
1676 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); | |
7253473c AA |
1677 | /* there is no chain */ |
1678 | if (page_node) { | |
1679 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1680 | list_del(&page_node->list); | |
1681 | DO_NUMA(page_node->nid = nid); | |
b8d20a18 AA |
1682 | rb_replace_node(&stable_node_dup->node, |
1683 | &page_node->node, | |
7253473c AA |
1684 | root); |
1685 | if (is_page_sharing_candidate(page_node)) | |
1686 | get_page(page); | |
1687 | else | |
1688 | page = NULL; | |
1689 | } else { | |
b8d20a18 | 1690 | rb_erase(&stable_node_dup->node, root); |
7253473c AA |
1691 | page = NULL; |
1692 | } | |
4146d2d6 | 1693 | } else { |
7253473c AA |
1694 | VM_BUG_ON(!is_stable_node_chain(stable_node)); |
1695 | __stable_node_dup_del(stable_node_dup); | |
1696 | if (page_node) { | |
1697 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1698 | list_del(&page_node->list); | |
1699 | DO_NUMA(page_node->nid = nid); | |
1700 | stable_node_chain_add_dup(page_node, stable_node); | |
1701 | if (is_page_sharing_candidate(page_node)) | |
1702 | get_page(page); | |
1703 | else | |
1704 | page = NULL; | |
1705 | } else { | |
1706 | page = NULL; | |
1707 | } | |
4146d2d6 | 1708 | } |
7253473c AA |
1709 | stable_node_dup->head = &migrate_nodes; |
1710 | list_add(&stable_node_dup->list, stable_node_dup->head); | |
4146d2d6 | 1711 | return page; |
7253473c AA |
1712 | |
1713 | chain_append: | |
1714 | /* stable_node_dup could be null if it reached the limit */ | |
1715 | if (!stable_node_dup) | |
1716 | stable_node_dup = stable_node_any; | |
b8d20a18 AA |
1717 | /* |
1718 | * If stable_node was a chain and chain_prune collapsed it, | |
3bef4484 AA |
1719 | * stable_node has been updated to be the new regular |
1720 | * stable_node. A collapse of the chain is indistinguishable | |
1721 | * from the case there was no chain in the stable | |
1722 | * rbtree. Otherwise stable_node is the chain and | |
1723 | * stable_node_dup is the dup to replace. | |
b8d20a18 | 1724 | */ |
3bef4484 | 1725 | if (stable_node_dup == stable_node) { |
b8d20a18 AA |
1726 | VM_BUG_ON(is_stable_node_chain(stable_node_dup)); |
1727 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); | |
7253473c AA |
1728 | /* chain is missing so create it */ |
1729 | stable_node = alloc_stable_node_chain(stable_node_dup, | |
1730 | root); | |
1731 | if (!stable_node) | |
1732 | return NULL; | |
1733 | } | |
1734 | /* | |
1735 | * Add this stable_node dup that was | |
1736 | * migrated to the stable_node chain | |
1737 | * of the current nid for this page | |
1738 | * content. | |
1739 | */ | |
b8d20a18 AA |
1740 | VM_BUG_ON(!is_stable_node_chain(stable_node)); |
1741 | VM_BUG_ON(!is_stable_node_dup(stable_node_dup)); | |
7253473c AA |
1742 | VM_BUG_ON(page_node->head != &migrate_nodes); |
1743 | list_del(&page_node->list); | |
1744 | DO_NUMA(page_node->nid = nid); | |
1745 | stable_node_chain_add_dup(page_node, stable_node); | |
1746 | goto out; | |
31dbd01f IE |
1747 | } |
1748 | ||
1749 | /* | |
e850dcf5 | 1750 | * stable_tree_insert - insert stable tree node pointing to new ksm page |
31dbd01f IE |
1751 | * into the stable tree. |
1752 | * | |
7b6ba2c7 HD |
1753 | * This function returns the stable tree node just allocated on success, |
1754 | * NULL otherwise. | |
31dbd01f | 1755 | */ |
7b6ba2c7 | 1756 | static struct stable_node *stable_tree_insert(struct page *kpage) |
31dbd01f | 1757 | { |
90bd6fd3 PH |
1758 | int nid; |
1759 | unsigned long kpfn; | |
ef53d16c | 1760 | struct rb_root *root; |
90bd6fd3 | 1761 | struct rb_node **new; |
f2e5ff85 | 1762 | struct rb_node *parent; |
7253473c AA |
1763 | struct stable_node *stable_node, *stable_node_dup, *stable_node_any; |
1764 | bool need_chain = false; | |
31dbd01f | 1765 | |
90bd6fd3 PH |
1766 | kpfn = page_to_pfn(kpage); |
1767 | nid = get_kpfn_nid(kpfn); | |
ef53d16c | 1768 | root = root_stable_tree + nid; |
f2e5ff85 AA |
1769 | again: |
1770 | parent = NULL; | |
ef53d16c | 1771 | new = &root->rb_node; |
90bd6fd3 | 1772 | |
31dbd01f | 1773 | while (*new) { |
4035c07a | 1774 | struct page *tree_page; |
31dbd01f IE |
1775 | int ret; |
1776 | ||
08beca44 | 1777 | cond_resched(); |
7b6ba2c7 | 1778 | stable_node = rb_entry(*new, struct stable_node, node); |
7253473c | 1779 | stable_node_any = NULL; |
6457f49b | 1780 | tree_page = chain(&stable_node_dup, stable_node, root); |
7253473c AA |
1781 | if (!stable_node_dup) { |
1782 | /* | |
1783 | * Either all stable_node dups were full in | |
1784 | * this stable_node chain, or this chain was | |
1785 | * empty and should be rb_erased. | |
1786 | */ | |
1787 | stable_node_any = stable_node_dup_any(stable_node, | |
1788 | root); | |
1789 | if (!stable_node_any) { | |
1790 | /* rb_erase just run */ | |
1791 | goto again; | |
1792 | } | |
1793 | /* | |
1794 | * Take any of the stable_node dups page of | |
1795 | * this stable_node chain to let the tree walk | |
1796 | * continue. All KSM pages belonging to the | |
1797 | * stable_node dups in a stable_node chain | |
1798 | * have the same content and they're | |
1799 | * wrprotected at all times. Any will work | |
1800 | * fine to continue the walk. | |
1801 | */ | |
1802 | tree_page = get_ksm_page(stable_node_any, false); | |
1803 | } | |
1804 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1805 | if (!tree_page) { |
1806 | /* | |
1807 | * If we walked over a stale stable_node, | |
1808 | * get_ksm_page() will call rb_erase() and it | |
1809 | * may rebalance the tree from under us. So | |
1810 | * restart the search from scratch. Returning | |
1811 | * NULL would be safe too, but we'd generate | |
1812 | * false negative insertions just because some | |
1813 | * stable_node was stale. | |
1814 | */ | |
1815 | goto again; | |
1816 | } | |
31dbd01f | 1817 | |
4035c07a HD |
1818 | ret = memcmp_pages(kpage, tree_page); |
1819 | put_page(tree_page); | |
31dbd01f IE |
1820 | |
1821 | parent = *new; | |
1822 | if (ret < 0) | |
1823 | new = &parent->rb_left; | |
1824 | else if (ret > 0) | |
1825 | new = &parent->rb_right; | |
1826 | else { | |
7253473c AA |
1827 | need_chain = true; |
1828 | break; | |
31dbd01f IE |
1829 | } |
1830 | } | |
1831 | ||
7253473c AA |
1832 | stable_node_dup = alloc_stable_node(); |
1833 | if (!stable_node_dup) | |
7b6ba2c7 | 1834 | return NULL; |
31dbd01f | 1835 | |
7253473c AA |
1836 | INIT_HLIST_HEAD(&stable_node_dup->hlist); |
1837 | stable_node_dup->kpfn = kpfn; | |
1838 | set_page_stable_node(kpage, stable_node_dup); | |
1839 | stable_node_dup->rmap_hlist_len = 0; | |
1840 | DO_NUMA(stable_node_dup->nid = nid); | |
1841 | if (!need_chain) { | |
1842 | rb_link_node(&stable_node_dup->node, parent, new); | |
1843 | rb_insert_color(&stable_node_dup->node, root); | |
1844 | } else { | |
1845 | if (!is_stable_node_chain(stable_node)) { | |
1846 | struct stable_node *orig = stable_node; | |
1847 | /* chain is missing so create it */ | |
1848 | stable_node = alloc_stable_node_chain(orig, root); | |
1849 | if (!stable_node) { | |
1850 | free_stable_node(stable_node_dup); | |
1851 | return NULL; | |
1852 | } | |
1853 | } | |
1854 | stable_node_chain_add_dup(stable_node_dup, stable_node); | |
1855 | } | |
08beca44 | 1856 | |
7253473c | 1857 | return stable_node_dup; |
31dbd01f IE |
1858 | } |
1859 | ||
1860 | /* | |
8dd3557a HD |
1861 | * unstable_tree_search_insert - search for identical page, |
1862 | * else insert rmap_item into the unstable tree. | |
31dbd01f IE |
1863 | * |
1864 | * This function searches for a page in the unstable tree identical to the | |
1865 | * page currently being scanned; and if no identical page is found in the | |
1866 | * tree, we insert rmap_item as a new object into the unstable tree. | |
1867 | * | |
1868 | * This function returns pointer to rmap_item found to be identical | |
1869 | * to the currently scanned page, NULL otherwise. | |
1870 | * | |
1871 | * This function does both searching and inserting, because they share | |
1872 | * the same walking algorithm in an rbtree. | |
1873 | */ | |
8dd3557a HD |
1874 | static |
1875 | struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item, | |
1876 | struct page *page, | |
1877 | struct page **tree_pagep) | |
31dbd01f | 1878 | { |
90bd6fd3 PH |
1879 | struct rb_node **new; |
1880 | struct rb_root *root; | |
31dbd01f | 1881 | struct rb_node *parent = NULL; |
90bd6fd3 PH |
1882 | int nid; |
1883 | ||
1884 | nid = get_kpfn_nid(page_to_pfn(page)); | |
ef53d16c | 1885 | root = root_unstable_tree + nid; |
90bd6fd3 | 1886 | new = &root->rb_node; |
31dbd01f IE |
1887 | |
1888 | while (*new) { | |
1889 | struct rmap_item *tree_rmap_item; | |
8dd3557a | 1890 | struct page *tree_page; |
31dbd01f IE |
1891 | int ret; |
1892 | ||
d178f27f | 1893 | cond_resched(); |
31dbd01f | 1894 | tree_rmap_item = rb_entry(*new, struct rmap_item, node); |
8dd3557a | 1895 | tree_page = get_mergeable_page(tree_rmap_item); |
c8f95ed1 | 1896 | if (!tree_page) |
31dbd01f IE |
1897 | return NULL; |
1898 | ||
1899 | /* | |
8dd3557a | 1900 | * Don't substitute a ksm page for a forked page. |
31dbd01f | 1901 | */ |
8dd3557a HD |
1902 | if (page == tree_page) { |
1903 | put_page(tree_page); | |
31dbd01f IE |
1904 | return NULL; |
1905 | } | |
1906 | ||
8dd3557a | 1907 | ret = memcmp_pages(page, tree_page); |
31dbd01f IE |
1908 | |
1909 | parent = *new; | |
1910 | if (ret < 0) { | |
8dd3557a | 1911 | put_page(tree_page); |
31dbd01f IE |
1912 | new = &parent->rb_left; |
1913 | } else if (ret > 0) { | |
8dd3557a | 1914 | put_page(tree_page); |
31dbd01f | 1915 | new = &parent->rb_right; |
b599cbdf HD |
1916 | } else if (!ksm_merge_across_nodes && |
1917 | page_to_nid(tree_page) != nid) { | |
1918 | /* | |
1919 | * If tree_page has been migrated to another NUMA node, | |
1920 | * it will be flushed out and put in the right unstable | |
1921 | * tree next time: only merge with it when across_nodes. | |
1922 | */ | |
1923 | put_page(tree_page); | |
1924 | return NULL; | |
31dbd01f | 1925 | } else { |
8dd3557a | 1926 | *tree_pagep = tree_page; |
31dbd01f IE |
1927 | return tree_rmap_item; |
1928 | } | |
1929 | } | |
1930 | ||
7b6ba2c7 | 1931 | rmap_item->address |= UNSTABLE_FLAG; |
31dbd01f | 1932 | rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); |
e850dcf5 | 1933 | DO_NUMA(rmap_item->nid = nid); |
31dbd01f | 1934 | rb_link_node(&rmap_item->node, parent, new); |
90bd6fd3 | 1935 | rb_insert_color(&rmap_item->node, root); |
31dbd01f | 1936 | |
473b0ce4 | 1937 | ksm_pages_unshared++; |
31dbd01f IE |
1938 | return NULL; |
1939 | } | |
1940 | ||
1941 | /* | |
1942 | * stable_tree_append - add another rmap_item to the linked list of | |
1943 | * rmap_items hanging off a given node of the stable tree, all sharing | |
1944 | * the same ksm page. | |
1945 | */ | |
1946 | static void stable_tree_append(struct rmap_item *rmap_item, | |
7253473c AA |
1947 | struct stable_node *stable_node, |
1948 | bool max_page_sharing_bypass) | |
31dbd01f | 1949 | { |
7253473c AA |
1950 | /* |
1951 | * rmap won't find this mapping if we don't insert the | |
1952 | * rmap_item in the right stable_node | |
1953 | * duplicate. page_migration could break later if rmap breaks, | |
1954 | * so we can as well crash here. We really need to check for | |
1955 | * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check | |
1956 | * for other negative values as an undeflow if detected here | |
1957 | * for the first time (and not when decreasing rmap_hlist_len) | |
1958 | * would be sign of memory corruption in the stable_node. | |
1959 | */ | |
1960 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
1961 | ||
1962 | stable_node->rmap_hlist_len++; | |
1963 | if (!max_page_sharing_bypass) | |
1964 | /* possibly non fatal but unexpected overflow, only warn */ | |
1965 | WARN_ON_ONCE(stable_node->rmap_hlist_len > | |
1966 | ksm_max_page_sharing); | |
1967 | ||
7b6ba2c7 | 1968 | rmap_item->head = stable_node; |
31dbd01f | 1969 | rmap_item->address |= STABLE_FLAG; |
7b6ba2c7 | 1970 | hlist_add_head(&rmap_item->hlist, &stable_node->hlist); |
e178dfde | 1971 | |
7b6ba2c7 HD |
1972 | if (rmap_item->hlist.next) |
1973 | ksm_pages_sharing++; | |
1974 | else | |
1975 | ksm_pages_shared++; | |
31dbd01f IE |
1976 | } |
1977 | ||
1978 | /* | |
81464e30 HD |
1979 | * cmp_and_merge_page - first see if page can be merged into the stable tree; |
1980 | * if not, compare checksum to previous and if it's the same, see if page can | |
1981 | * be inserted into the unstable tree, or merged with a page already there and | |
1982 | * both transferred to the stable tree. | |
31dbd01f IE |
1983 | * |
1984 | * @page: the page that we are searching identical page to. | |
1985 | * @rmap_item: the reverse mapping into the virtual address of this page | |
1986 | */ | |
1987 | static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item) | |
1988 | { | |
31dbd01f | 1989 | struct rmap_item *tree_rmap_item; |
8dd3557a | 1990 | struct page *tree_page = NULL; |
7b6ba2c7 | 1991 | struct stable_node *stable_node; |
8dd3557a | 1992 | struct page *kpage; |
31dbd01f IE |
1993 | unsigned int checksum; |
1994 | int err; | |
7253473c | 1995 | bool max_page_sharing_bypass = false; |
31dbd01f | 1996 | |
4146d2d6 HD |
1997 | stable_node = page_stable_node(page); |
1998 | if (stable_node) { | |
1999 | if (stable_node->head != &migrate_nodes && | |
7253473c AA |
2000 | get_kpfn_nid(READ_ONCE(stable_node->kpfn)) != |
2001 | NUMA(stable_node->nid)) { | |
2002 | stable_node_dup_del(stable_node); | |
4146d2d6 HD |
2003 | stable_node->head = &migrate_nodes; |
2004 | list_add(&stable_node->list, stable_node->head); | |
2005 | } | |
2006 | if (stable_node->head != &migrate_nodes && | |
2007 | rmap_item->head == stable_node) | |
2008 | return; | |
7253473c AA |
2009 | /* |
2010 | * If it's a KSM fork, allow it to go over the sharing limit | |
2011 | * without warnings. | |
2012 | */ | |
2013 | if (!is_page_sharing_candidate(stable_node)) | |
2014 | max_page_sharing_bypass = true; | |
4146d2d6 | 2015 | } |
31dbd01f IE |
2016 | |
2017 | /* We first start with searching the page inside the stable tree */ | |
62b61f61 | 2018 | kpage = stable_tree_search(page); |
4146d2d6 HD |
2019 | if (kpage == page && rmap_item->head == stable_node) { |
2020 | put_page(kpage); | |
2021 | return; | |
2022 | } | |
2023 | ||
2024 | remove_rmap_item_from_tree(rmap_item); | |
2025 | ||
62b61f61 | 2026 | if (kpage) { |
08beca44 | 2027 | err = try_to_merge_with_ksm_page(rmap_item, page, kpage); |
31dbd01f IE |
2028 | if (!err) { |
2029 | /* | |
2030 | * The page was successfully merged: | |
2031 | * add its rmap_item to the stable tree. | |
2032 | */ | |
5ad64688 | 2033 | lock_page(kpage); |
7253473c AA |
2034 | stable_tree_append(rmap_item, page_stable_node(kpage), |
2035 | max_page_sharing_bypass); | |
5ad64688 | 2036 | unlock_page(kpage); |
31dbd01f | 2037 | } |
8dd3557a | 2038 | put_page(kpage); |
31dbd01f IE |
2039 | return; |
2040 | } | |
2041 | ||
2042 | /* | |
4035c07a HD |
2043 | * If the hash value of the page has changed from the last time |
2044 | * we calculated it, this page is changing frequently: therefore we | |
2045 | * don't want to insert it in the unstable tree, and we don't want | |
2046 | * to waste our time searching for something identical to it there. | |
31dbd01f IE |
2047 | */ |
2048 | checksum = calc_checksum(page); | |
2049 | if (rmap_item->oldchecksum != checksum) { | |
2050 | rmap_item->oldchecksum = checksum; | |
2051 | return; | |
2052 | } | |
2053 | ||
fa2aa0fc CI |
2054 | /* |
2055 | * Same checksum as an empty page. We attempt to merge it with the | |
2056 | * appropriate zero page if the user enabled this via sysfs. | |
2057 | */ | |
2058 | if (ksm_use_zero_pages && (checksum == zero_checksum)) { | |
2059 | struct vm_area_struct *vma; | |
2060 | ||
2061 | vma = find_mergeable_vma(rmap_item->mm, rmap_item->address); | |
2062 | err = try_to_merge_one_page(vma, page, | |
2063 | ZERO_PAGE(rmap_item->address)); | |
2064 | /* | |
2065 | * In case of failure, the page was not really empty, so we | |
2066 | * need to continue. Otherwise we're done. | |
2067 | */ | |
2068 | if (!err) | |
2069 | return; | |
2070 | } | |
8dd3557a HD |
2071 | tree_rmap_item = |
2072 | unstable_tree_search_insert(rmap_item, page, &tree_page); | |
31dbd01f | 2073 | if (tree_rmap_item) { |
8dd3557a HD |
2074 | kpage = try_to_merge_two_pages(rmap_item, page, |
2075 | tree_rmap_item, tree_page); | |
2076 | put_page(tree_page); | |
8dd3557a | 2077 | if (kpage) { |
bc56620b HD |
2078 | /* |
2079 | * The pages were successfully merged: insert new | |
2080 | * node in the stable tree and add both rmap_items. | |
2081 | */ | |
5ad64688 | 2082 | lock_page(kpage); |
7b6ba2c7 HD |
2083 | stable_node = stable_tree_insert(kpage); |
2084 | if (stable_node) { | |
7253473c AA |
2085 | stable_tree_append(tree_rmap_item, stable_node, |
2086 | false); | |
2087 | stable_tree_append(rmap_item, stable_node, | |
2088 | false); | |
7b6ba2c7 | 2089 | } |
5ad64688 | 2090 | unlock_page(kpage); |
7b6ba2c7 | 2091 | |
31dbd01f IE |
2092 | /* |
2093 | * If we fail to insert the page into the stable tree, | |
2094 | * we will have 2 virtual addresses that are pointing | |
2095 | * to a ksm page left outside the stable tree, | |
2096 | * in which case we need to break_cow on both. | |
2097 | */ | |
7b6ba2c7 | 2098 | if (!stable_node) { |
8dd3557a HD |
2099 | break_cow(tree_rmap_item); |
2100 | break_cow(rmap_item); | |
31dbd01f IE |
2101 | } |
2102 | } | |
31dbd01f IE |
2103 | } |
2104 | } | |
2105 | ||
2106 | static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot, | |
6514d511 | 2107 | struct rmap_item **rmap_list, |
31dbd01f IE |
2108 | unsigned long addr) |
2109 | { | |
2110 | struct rmap_item *rmap_item; | |
2111 | ||
6514d511 HD |
2112 | while (*rmap_list) { |
2113 | rmap_item = *rmap_list; | |
93d17715 | 2114 | if ((rmap_item->address & PAGE_MASK) == addr) |
31dbd01f | 2115 | return rmap_item; |
31dbd01f IE |
2116 | if (rmap_item->address > addr) |
2117 | break; | |
6514d511 | 2118 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 2119 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
2120 | free_rmap_item(rmap_item); |
2121 | } | |
2122 | ||
2123 | rmap_item = alloc_rmap_item(); | |
2124 | if (rmap_item) { | |
2125 | /* It has already been zeroed */ | |
2126 | rmap_item->mm = mm_slot->mm; | |
2127 | rmap_item->address = addr; | |
6514d511 HD |
2128 | rmap_item->rmap_list = *rmap_list; |
2129 | *rmap_list = rmap_item; | |
31dbd01f IE |
2130 | } |
2131 | return rmap_item; | |
2132 | } | |
2133 | ||
2134 | static struct rmap_item *scan_get_next_rmap_item(struct page **page) | |
2135 | { | |
2136 | struct mm_struct *mm; | |
2137 | struct mm_slot *slot; | |
2138 | struct vm_area_struct *vma; | |
2139 | struct rmap_item *rmap_item; | |
90bd6fd3 | 2140 | int nid; |
31dbd01f IE |
2141 | |
2142 | if (list_empty(&ksm_mm_head.mm_list)) | |
2143 | return NULL; | |
2144 | ||
2145 | slot = ksm_scan.mm_slot; | |
2146 | if (slot == &ksm_mm_head) { | |
2919bfd0 HD |
2147 | /* |
2148 | * A number of pages can hang around indefinitely on per-cpu | |
2149 | * pagevecs, raised page count preventing write_protect_page | |
2150 | * from merging them. Though it doesn't really matter much, | |
2151 | * it is puzzling to see some stuck in pages_volatile until | |
2152 | * other activity jostles them out, and they also prevented | |
2153 | * LTP's KSM test from succeeding deterministically; so drain | |
2154 | * them here (here rather than on entry to ksm_do_scan(), | |
2155 | * so we don't IPI too often when pages_to_scan is set low). | |
2156 | */ | |
2157 | lru_add_drain_all(); | |
2158 | ||
4146d2d6 HD |
2159 | /* |
2160 | * Whereas stale stable_nodes on the stable_tree itself | |
2161 | * get pruned in the regular course of stable_tree_search(), | |
2162 | * those moved out to the migrate_nodes list can accumulate: | |
2163 | * so prune them once before each full scan. | |
2164 | */ | |
2165 | if (!ksm_merge_across_nodes) { | |
03640418 | 2166 | struct stable_node *stable_node, *next; |
4146d2d6 HD |
2167 | struct page *page; |
2168 | ||
03640418 GT |
2169 | list_for_each_entry_safe(stable_node, next, |
2170 | &migrate_nodes, list) { | |
4146d2d6 HD |
2171 | page = get_ksm_page(stable_node, false); |
2172 | if (page) | |
2173 | put_page(page); | |
2174 | cond_resched(); | |
2175 | } | |
2176 | } | |
2177 | ||
ef53d16c | 2178 | for (nid = 0; nid < ksm_nr_node_ids; nid++) |
90bd6fd3 | 2179 | root_unstable_tree[nid] = RB_ROOT; |
31dbd01f IE |
2180 | |
2181 | spin_lock(&ksm_mmlist_lock); | |
2182 | slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list); | |
2183 | ksm_scan.mm_slot = slot; | |
2184 | spin_unlock(&ksm_mmlist_lock); | |
2b472611 HD |
2185 | /* |
2186 | * Although we tested list_empty() above, a racing __ksm_exit | |
2187 | * of the last mm on the list may have removed it since then. | |
2188 | */ | |
2189 | if (slot == &ksm_mm_head) | |
2190 | return NULL; | |
31dbd01f IE |
2191 | next_mm: |
2192 | ksm_scan.address = 0; | |
6514d511 | 2193 | ksm_scan.rmap_list = &slot->rmap_list; |
31dbd01f IE |
2194 | } |
2195 | ||
2196 | mm = slot->mm; | |
2197 | down_read(&mm->mmap_sem); | |
9ba69294 HD |
2198 | if (ksm_test_exit(mm)) |
2199 | vma = NULL; | |
2200 | else | |
2201 | vma = find_vma(mm, ksm_scan.address); | |
2202 | ||
2203 | for (; vma; vma = vma->vm_next) { | |
31dbd01f IE |
2204 | if (!(vma->vm_flags & VM_MERGEABLE)) |
2205 | continue; | |
2206 | if (ksm_scan.address < vma->vm_start) | |
2207 | ksm_scan.address = vma->vm_start; | |
2208 | if (!vma->anon_vma) | |
2209 | ksm_scan.address = vma->vm_end; | |
2210 | ||
2211 | while (ksm_scan.address < vma->vm_end) { | |
9ba69294 HD |
2212 | if (ksm_test_exit(mm)) |
2213 | break; | |
31dbd01f | 2214 | *page = follow_page(vma, ksm_scan.address, FOLL_GET); |
21ae5b01 AA |
2215 | if (IS_ERR_OR_NULL(*page)) { |
2216 | ksm_scan.address += PAGE_SIZE; | |
2217 | cond_resched(); | |
2218 | continue; | |
2219 | } | |
f765f540 | 2220 | if (PageAnon(*page)) { |
31dbd01f IE |
2221 | flush_anon_page(vma, *page, ksm_scan.address); |
2222 | flush_dcache_page(*page); | |
2223 | rmap_item = get_next_rmap_item(slot, | |
6514d511 | 2224 | ksm_scan.rmap_list, ksm_scan.address); |
31dbd01f | 2225 | if (rmap_item) { |
6514d511 HD |
2226 | ksm_scan.rmap_list = |
2227 | &rmap_item->rmap_list; | |
31dbd01f IE |
2228 | ksm_scan.address += PAGE_SIZE; |
2229 | } else | |
2230 | put_page(*page); | |
2231 | up_read(&mm->mmap_sem); | |
2232 | return rmap_item; | |
2233 | } | |
21ae5b01 | 2234 | put_page(*page); |
31dbd01f IE |
2235 | ksm_scan.address += PAGE_SIZE; |
2236 | cond_resched(); | |
2237 | } | |
2238 | } | |
2239 | ||
9ba69294 HD |
2240 | if (ksm_test_exit(mm)) { |
2241 | ksm_scan.address = 0; | |
6514d511 | 2242 | ksm_scan.rmap_list = &slot->rmap_list; |
9ba69294 | 2243 | } |
31dbd01f IE |
2244 | /* |
2245 | * Nuke all the rmap_items that are above this current rmap: | |
2246 | * because there were no VM_MERGEABLE vmas with such addresses. | |
2247 | */ | |
6514d511 | 2248 | remove_trailing_rmap_items(slot, ksm_scan.rmap_list); |
31dbd01f IE |
2249 | |
2250 | spin_lock(&ksm_mmlist_lock); | |
cd551f97 HD |
2251 | ksm_scan.mm_slot = list_entry(slot->mm_list.next, |
2252 | struct mm_slot, mm_list); | |
2253 | if (ksm_scan.address == 0) { | |
2254 | /* | |
2255 | * We've completed a full scan of all vmas, holding mmap_sem | |
2256 | * throughout, and found no VM_MERGEABLE: so do the same as | |
2257 | * __ksm_exit does to remove this mm from all our lists now. | |
9ba69294 HD |
2258 | * This applies either when cleaning up after __ksm_exit |
2259 | * (but beware: we can reach here even before __ksm_exit), | |
2260 | * or when all VM_MERGEABLE areas have been unmapped (and | |
2261 | * mmap_sem then protects against race with MADV_MERGEABLE). | |
cd551f97 | 2262 | */ |
4ca3a69b | 2263 | hash_del(&slot->link); |
cd551f97 | 2264 | list_del(&slot->mm_list); |
9ba69294 HD |
2265 | spin_unlock(&ksm_mmlist_lock); |
2266 | ||
cd551f97 HD |
2267 | free_mm_slot(slot); |
2268 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
9ba69294 HD |
2269 | up_read(&mm->mmap_sem); |
2270 | mmdrop(mm); | |
2271 | } else { | |
9ba69294 | 2272 | up_read(&mm->mmap_sem); |
7496fea9 ZC |
2273 | /* |
2274 | * up_read(&mm->mmap_sem) first because after | |
2275 | * spin_unlock(&ksm_mmlist_lock) run, the "mm" may | |
2276 | * already have been freed under us by __ksm_exit() | |
2277 | * because the "mm_slot" is still hashed and | |
2278 | * ksm_scan.mm_slot doesn't point to it anymore. | |
2279 | */ | |
2280 | spin_unlock(&ksm_mmlist_lock); | |
cd551f97 | 2281 | } |
31dbd01f IE |
2282 | |
2283 | /* Repeat until we've completed scanning the whole list */ | |
cd551f97 | 2284 | slot = ksm_scan.mm_slot; |
31dbd01f IE |
2285 | if (slot != &ksm_mm_head) |
2286 | goto next_mm; | |
2287 | ||
31dbd01f IE |
2288 | ksm_scan.seqnr++; |
2289 | return NULL; | |
2290 | } | |
2291 | ||
2292 | /** | |
2293 | * ksm_do_scan - the ksm scanner main worker function. | |
2294 | * @scan_npages - number of pages we want to scan before we return. | |
2295 | */ | |
2296 | static void ksm_do_scan(unsigned int scan_npages) | |
2297 | { | |
2298 | struct rmap_item *rmap_item; | |
22eccdd7 | 2299 | struct page *uninitialized_var(page); |
31dbd01f | 2300 | |
878aee7d | 2301 | while (scan_npages-- && likely(!freezing(current))) { |
31dbd01f IE |
2302 | cond_resched(); |
2303 | rmap_item = scan_get_next_rmap_item(&page); | |
2304 | if (!rmap_item) | |
2305 | return; | |
4146d2d6 | 2306 | cmp_and_merge_page(page, rmap_item); |
31dbd01f IE |
2307 | put_page(page); |
2308 | } | |
2309 | } | |
2310 | ||
6e158384 HD |
2311 | static int ksmd_should_run(void) |
2312 | { | |
2313 | return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list); | |
2314 | } | |
2315 | ||
31dbd01f IE |
2316 | static int ksm_scan_thread(void *nothing) |
2317 | { | |
878aee7d | 2318 | set_freezable(); |
339aa624 | 2319 | set_user_nice(current, 5); |
31dbd01f IE |
2320 | |
2321 | while (!kthread_should_stop()) { | |
6e158384 | 2322 | mutex_lock(&ksm_thread_mutex); |
ef4d43a8 | 2323 | wait_while_offlining(); |
6e158384 | 2324 | if (ksmd_should_run()) |
31dbd01f | 2325 | ksm_do_scan(ksm_thread_pages_to_scan); |
6e158384 HD |
2326 | mutex_unlock(&ksm_thread_mutex); |
2327 | ||
878aee7d AA |
2328 | try_to_freeze(); |
2329 | ||
6e158384 | 2330 | if (ksmd_should_run()) { |
23519606 AV |
2331 | if (ksm_thread_sleep_millisecs >= 1000) |
2332 | schedule_timeout_interruptible( | |
2333 | msecs_to_jiffies(round_jiffies_relative(ksm_thread_sleep_millisecs))); | |
2334 | else | |
2335 | schedule_timeout_interruptible( | |
2336 | msecs_to_jiffies(ksm_thread_sleep_millisecs)); | |
31dbd01f | 2337 | } else { |
878aee7d | 2338 | wait_event_freezable(ksm_thread_wait, |
6e158384 | 2339 | ksmd_should_run() || kthread_should_stop()); |
31dbd01f IE |
2340 | } |
2341 | } | |
2342 | return 0; | |
2343 | } | |
2344 | ||
f8af4da3 HD |
2345 | int ksm_madvise(struct vm_area_struct *vma, unsigned long start, |
2346 | unsigned long end, int advice, unsigned long *vm_flags) | |
2347 | { | |
2348 | struct mm_struct *mm = vma->vm_mm; | |
d952b791 | 2349 | int err; |
f8af4da3 HD |
2350 | |
2351 | switch (advice) { | |
2352 | case MADV_MERGEABLE: | |
2353 | /* | |
2354 | * Be somewhat over-protective for now! | |
2355 | */ | |
2356 | if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE | | |
2357 | VM_PFNMAP | VM_IO | VM_DONTEXPAND | | |
0661a336 | 2358 | VM_HUGETLB | VM_MIXEDMAP)) |
f8af4da3 HD |
2359 | return 0; /* just ignore the advice */ |
2360 | ||
cc2383ec KK |
2361 | #ifdef VM_SAO |
2362 | if (*vm_flags & VM_SAO) | |
2363 | return 0; | |
2364 | #endif | |
2365 | ||
d952b791 HD |
2366 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { |
2367 | err = __ksm_enter(mm); | |
2368 | if (err) | |
2369 | return err; | |
2370 | } | |
f8af4da3 HD |
2371 | |
2372 | *vm_flags |= VM_MERGEABLE; | |
2373 | break; | |
2374 | ||
2375 | case MADV_UNMERGEABLE: | |
2376 | if (!(*vm_flags & VM_MERGEABLE)) | |
2377 | return 0; /* just ignore the advice */ | |
2378 | ||
d952b791 HD |
2379 | if (vma->anon_vma) { |
2380 | err = unmerge_ksm_pages(vma, start, end); | |
2381 | if (err) | |
2382 | return err; | |
2383 | } | |
f8af4da3 HD |
2384 | |
2385 | *vm_flags &= ~VM_MERGEABLE; | |
2386 | break; | |
2387 | } | |
2388 | ||
2389 | return 0; | |
2390 | } | |
2391 | ||
2392 | int __ksm_enter(struct mm_struct *mm) | |
2393 | { | |
6e158384 HD |
2394 | struct mm_slot *mm_slot; |
2395 | int needs_wakeup; | |
2396 | ||
2397 | mm_slot = alloc_mm_slot(); | |
31dbd01f IE |
2398 | if (!mm_slot) |
2399 | return -ENOMEM; | |
2400 | ||
6e158384 HD |
2401 | /* Check ksm_run too? Would need tighter locking */ |
2402 | needs_wakeup = list_empty(&ksm_mm_head.mm_list); | |
2403 | ||
31dbd01f IE |
2404 | spin_lock(&ksm_mmlist_lock); |
2405 | insert_to_mm_slots_hash(mm, mm_slot); | |
2406 | /* | |
cbf86cfe HD |
2407 | * When KSM_RUN_MERGE (or KSM_RUN_STOP), |
2408 | * insert just behind the scanning cursor, to let the area settle | |
31dbd01f IE |
2409 | * down a little; when fork is followed by immediate exec, we don't |
2410 | * want ksmd to waste time setting up and tearing down an rmap_list. | |
cbf86cfe HD |
2411 | * |
2412 | * But when KSM_RUN_UNMERGE, it's important to insert ahead of its | |
2413 | * scanning cursor, otherwise KSM pages in newly forked mms will be | |
2414 | * missed: then we might as well insert at the end of the list. | |
31dbd01f | 2415 | */ |
cbf86cfe HD |
2416 | if (ksm_run & KSM_RUN_UNMERGE) |
2417 | list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list); | |
2418 | else | |
2419 | list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list); | |
31dbd01f IE |
2420 | spin_unlock(&ksm_mmlist_lock); |
2421 | ||
f8af4da3 | 2422 | set_bit(MMF_VM_MERGEABLE, &mm->flags); |
9ba69294 | 2423 | atomic_inc(&mm->mm_count); |
6e158384 HD |
2424 | |
2425 | if (needs_wakeup) | |
2426 | wake_up_interruptible(&ksm_thread_wait); | |
2427 | ||
f8af4da3 HD |
2428 | return 0; |
2429 | } | |
2430 | ||
1c2fb7a4 | 2431 | void __ksm_exit(struct mm_struct *mm) |
f8af4da3 | 2432 | { |
cd551f97 | 2433 | struct mm_slot *mm_slot; |
9ba69294 | 2434 | int easy_to_free = 0; |
cd551f97 | 2435 | |
31dbd01f | 2436 | /* |
9ba69294 HD |
2437 | * This process is exiting: if it's straightforward (as is the |
2438 | * case when ksmd was never running), free mm_slot immediately. | |
2439 | * But if it's at the cursor or has rmap_items linked to it, use | |
2440 | * mmap_sem to synchronize with any break_cows before pagetables | |
2441 | * are freed, and leave the mm_slot on the list for ksmd to free. | |
2442 | * Beware: ksm may already have noticed it exiting and freed the slot. | |
31dbd01f | 2443 | */ |
9ba69294 | 2444 | |
cd551f97 HD |
2445 | spin_lock(&ksm_mmlist_lock); |
2446 | mm_slot = get_mm_slot(mm); | |
9ba69294 | 2447 | if (mm_slot && ksm_scan.mm_slot != mm_slot) { |
6514d511 | 2448 | if (!mm_slot->rmap_list) { |
4ca3a69b | 2449 | hash_del(&mm_slot->link); |
9ba69294 HD |
2450 | list_del(&mm_slot->mm_list); |
2451 | easy_to_free = 1; | |
2452 | } else { | |
2453 | list_move(&mm_slot->mm_list, | |
2454 | &ksm_scan.mm_slot->mm_list); | |
2455 | } | |
cd551f97 | 2456 | } |
cd551f97 HD |
2457 | spin_unlock(&ksm_mmlist_lock); |
2458 | ||
9ba69294 HD |
2459 | if (easy_to_free) { |
2460 | free_mm_slot(mm_slot); | |
2461 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
2462 | mmdrop(mm); | |
2463 | } else if (mm_slot) { | |
9ba69294 HD |
2464 | down_write(&mm->mmap_sem); |
2465 | up_write(&mm->mmap_sem); | |
9ba69294 | 2466 | } |
31dbd01f IE |
2467 | } |
2468 | ||
cbf86cfe | 2469 | struct page *ksm_might_need_to_copy(struct page *page, |
5ad64688 HD |
2470 | struct vm_area_struct *vma, unsigned long address) |
2471 | { | |
cbf86cfe | 2472 | struct anon_vma *anon_vma = page_anon_vma(page); |
5ad64688 HD |
2473 | struct page *new_page; |
2474 | ||
cbf86cfe HD |
2475 | if (PageKsm(page)) { |
2476 | if (page_stable_node(page) && | |
2477 | !(ksm_run & KSM_RUN_UNMERGE)) | |
2478 | return page; /* no need to copy it */ | |
2479 | } else if (!anon_vma) { | |
2480 | return page; /* no need to copy it */ | |
2481 | } else if (anon_vma->root == vma->anon_vma->root && | |
2482 | page->index == linear_page_index(vma, address)) { | |
2483 | return page; /* still no need to copy it */ | |
2484 | } | |
2485 | if (!PageUptodate(page)) | |
2486 | return page; /* let do_swap_page report the error */ | |
2487 | ||
5ad64688 HD |
2488 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
2489 | if (new_page) { | |
2490 | copy_user_highpage(new_page, page, address, vma); | |
2491 | ||
2492 | SetPageDirty(new_page); | |
2493 | __SetPageUptodate(new_page); | |
48c935ad | 2494 | __SetPageLocked(new_page); |
5ad64688 HD |
2495 | } |
2496 | ||
5ad64688 HD |
2497 | return new_page; |
2498 | } | |
2499 | ||
051ac83a | 2500 | int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc) |
e9995ef9 HD |
2501 | { |
2502 | struct stable_node *stable_node; | |
e9995ef9 HD |
2503 | struct rmap_item *rmap_item; |
2504 | int ret = SWAP_AGAIN; | |
2505 | int search_new_forks = 0; | |
2506 | ||
309381fe | 2507 | VM_BUG_ON_PAGE(!PageKsm(page), page); |
9f32624b JK |
2508 | |
2509 | /* | |
2510 | * Rely on the page lock to protect against concurrent modifications | |
2511 | * to that page's node of the stable tree. | |
2512 | */ | |
309381fe | 2513 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
e9995ef9 HD |
2514 | |
2515 | stable_node = page_stable_node(page); | |
2516 | if (!stable_node) | |
2517 | return ret; | |
2518 | again: | |
b67bfe0d | 2519 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
e9995ef9 | 2520 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
5beb4930 | 2521 | struct anon_vma_chain *vmac; |
e9995ef9 HD |
2522 | struct vm_area_struct *vma; |
2523 | ||
ad12695f | 2524 | cond_resched(); |
b6b19f25 | 2525 | anon_vma_lock_read(anon_vma); |
bf181b9f ML |
2526 | anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, |
2527 | 0, ULONG_MAX) { | |
ad12695f | 2528 | cond_resched(); |
5beb4930 | 2529 | vma = vmac->vma; |
e9995ef9 HD |
2530 | if (rmap_item->address < vma->vm_start || |
2531 | rmap_item->address >= vma->vm_end) | |
2532 | continue; | |
2533 | /* | |
2534 | * Initially we examine only the vma which covers this | |
2535 | * rmap_item; but later, if there is still work to do, | |
2536 | * we examine covering vmas in other mms: in case they | |
2537 | * were forked from the original since ksmd passed. | |
2538 | */ | |
2539 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
2540 | continue; | |
2541 | ||
0dd1c7bb JK |
2542 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
2543 | continue; | |
2544 | ||
051ac83a JK |
2545 | ret = rwc->rmap_one(page, vma, |
2546 | rmap_item->address, rwc->arg); | |
e9995ef9 | 2547 | if (ret != SWAP_AGAIN) { |
b6b19f25 | 2548 | anon_vma_unlock_read(anon_vma); |
e9995ef9 HD |
2549 | goto out; |
2550 | } | |
0dd1c7bb JK |
2551 | if (rwc->done && rwc->done(page)) { |
2552 | anon_vma_unlock_read(anon_vma); | |
2553 | goto out; | |
2554 | } | |
e9995ef9 | 2555 | } |
b6b19f25 | 2556 | anon_vma_unlock_read(anon_vma); |
e9995ef9 HD |
2557 | } |
2558 | if (!search_new_forks++) | |
2559 | goto again; | |
2560 | out: | |
2561 | return ret; | |
2562 | } | |
2563 | ||
52629506 | 2564 | #ifdef CONFIG_MIGRATION |
e9995ef9 HD |
2565 | void ksm_migrate_page(struct page *newpage, struct page *oldpage) |
2566 | { | |
2567 | struct stable_node *stable_node; | |
2568 | ||
309381fe SL |
2569 | VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage); |
2570 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
2571 | VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage); | |
e9995ef9 HD |
2572 | |
2573 | stable_node = page_stable_node(newpage); | |
2574 | if (stable_node) { | |
309381fe | 2575 | VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage); |
62b61f61 | 2576 | stable_node->kpfn = page_to_pfn(newpage); |
c8d6553b HD |
2577 | /* |
2578 | * newpage->mapping was set in advance; now we need smp_wmb() | |
2579 | * to make sure that the new stable_node->kpfn is visible | |
2580 | * to get_ksm_page() before it can see that oldpage->mapping | |
2581 | * has gone stale (or that PageSwapCache has been cleared). | |
2582 | */ | |
2583 | smp_wmb(); | |
2584 | set_page_stable_node(oldpage, NULL); | |
e9995ef9 HD |
2585 | } |
2586 | } | |
2587 | #endif /* CONFIG_MIGRATION */ | |
2588 | ||
62b61f61 | 2589 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 HD |
2590 | static void wait_while_offlining(void) |
2591 | { | |
2592 | while (ksm_run & KSM_RUN_OFFLINE) { | |
2593 | mutex_unlock(&ksm_thread_mutex); | |
2594 | wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), | |
74316201 | 2595 | TASK_UNINTERRUPTIBLE); |
ef4d43a8 HD |
2596 | mutex_lock(&ksm_thread_mutex); |
2597 | } | |
2598 | } | |
2599 | ||
7253473c AA |
2600 | static bool stable_node_dup_remove_range(struct stable_node *stable_node, |
2601 | unsigned long start_pfn, | |
2602 | unsigned long end_pfn) | |
2603 | { | |
2604 | if (stable_node->kpfn >= start_pfn && | |
2605 | stable_node->kpfn < end_pfn) { | |
2606 | /* | |
2607 | * Don't get_ksm_page, page has already gone: | |
2608 | * which is why we keep kpfn instead of page* | |
2609 | */ | |
2610 | remove_node_from_stable_tree(stable_node); | |
2611 | return true; | |
2612 | } | |
2613 | return false; | |
2614 | } | |
2615 | ||
2616 | static bool stable_node_chain_remove_range(struct stable_node *stable_node, | |
2617 | unsigned long start_pfn, | |
2618 | unsigned long end_pfn, | |
2619 | struct rb_root *root) | |
2620 | { | |
2621 | struct stable_node *dup; | |
2622 | struct hlist_node *hlist_safe; | |
2623 | ||
2624 | if (!is_stable_node_chain(stable_node)) { | |
2625 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
2626 | return stable_node_dup_remove_range(stable_node, start_pfn, | |
2627 | end_pfn); | |
2628 | } | |
2629 | ||
2630 | hlist_for_each_entry_safe(dup, hlist_safe, | |
2631 | &stable_node->hlist, hlist_dup) { | |
2632 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
2633 | stable_node_dup_remove_range(dup, start_pfn, end_pfn); | |
2634 | } | |
2635 | if (hlist_empty(&stable_node->hlist)) { | |
2636 | free_stable_node_chain(stable_node, root); | |
2637 | return true; /* notify caller that tree was rebalanced */ | |
2638 | } else | |
2639 | return false; | |
2640 | } | |
2641 | ||
ee0ea59c HD |
2642 | static void ksm_check_stable_tree(unsigned long start_pfn, |
2643 | unsigned long end_pfn) | |
62b61f61 | 2644 | { |
03640418 | 2645 | struct stable_node *stable_node, *next; |
62b61f61 | 2646 | struct rb_node *node; |
90bd6fd3 | 2647 | int nid; |
62b61f61 | 2648 | |
ef53d16c HD |
2649 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
2650 | node = rb_first(root_stable_tree + nid); | |
ee0ea59c | 2651 | while (node) { |
90bd6fd3 | 2652 | stable_node = rb_entry(node, struct stable_node, node); |
7253473c AA |
2653 | if (stable_node_chain_remove_range(stable_node, |
2654 | start_pfn, end_pfn, | |
2655 | root_stable_tree + | |
2656 | nid)) | |
ef53d16c | 2657 | node = rb_first(root_stable_tree + nid); |
7253473c | 2658 | else |
ee0ea59c HD |
2659 | node = rb_next(node); |
2660 | cond_resched(); | |
90bd6fd3 | 2661 | } |
ee0ea59c | 2662 | } |
03640418 | 2663 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
2664 | if (stable_node->kpfn >= start_pfn && |
2665 | stable_node->kpfn < end_pfn) | |
2666 | remove_node_from_stable_tree(stable_node); | |
2667 | cond_resched(); | |
2668 | } | |
62b61f61 HD |
2669 | } |
2670 | ||
2671 | static int ksm_memory_callback(struct notifier_block *self, | |
2672 | unsigned long action, void *arg) | |
2673 | { | |
2674 | struct memory_notify *mn = arg; | |
62b61f61 HD |
2675 | |
2676 | switch (action) { | |
2677 | case MEM_GOING_OFFLINE: | |
2678 | /* | |
ef4d43a8 HD |
2679 | * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() |
2680 | * and remove_all_stable_nodes() while memory is going offline: | |
2681 | * it is unsafe for them to touch the stable tree at this time. | |
2682 | * But unmerge_ksm_pages(), rmap lookups and other entry points | |
2683 | * which do not need the ksm_thread_mutex are all safe. | |
62b61f61 | 2684 | */ |
ef4d43a8 HD |
2685 | mutex_lock(&ksm_thread_mutex); |
2686 | ksm_run |= KSM_RUN_OFFLINE; | |
2687 | mutex_unlock(&ksm_thread_mutex); | |
62b61f61 HD |
2688 | break; |
2689 | ||
2690 | case MEM_OFFLINE: | |
2691 | /* | |
2692 | * Most of the work is done by page migration; but there might | |
2693 | * be a few stable_nodes left over, still pointing to struct | |
ee0ea59c HD |
2694 | * pages which have been offlined: prune those from the tree, |
2695 | * otherwise get_ksm_page() might later try to access a | |
2696 | * non-existent struct page. | |
62b61f61 | 2697 | */ |
ee0ea59c HD |
2698 | ksm_check_stable_tree(mn->start_pfn, |
2699 | mn->start_pfn + mn->nr_pages); | |
62b61f61 HD |
2700 | /* fallthrough */ |
2701 | ||
2702 | case MEM_CANCEL_OFFLINE: | |
ef4d43a8 HD |
2703 | mutex_lock(&ksm_thread_mutex); |
2704 | ksm_run &= ~KSM_RUN_OFFLINE; | |
62b61f61 | 2705 | mutex_unlock(&ksm_thread_mutex); |
ef4d43a8 HD |
2706 | |
2707 | smp_mb(); /* wake_up_bit advises this */ | |
2708 | wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); | |
62b61f61 HD |
2709 | break; |
2710 | } | |
2711 | return NOTIFY_OK; | |
2712 | } | |
ef4d43a8 HD |
2713 | #else |
2714 | static void wait_while_offlining(void) | |
2715 | { | |
2716 | } | |
62b61f61 HD |
2717 | #endif /* CONFIG_MEMORY_HOTREMOVE */ |
2718 | ||
2ffd8679 HD |
2719 | #ifdef CONFIG_SYSFS |
2720 | /* | |
2721 | * This all compiles without CONFIG_SYSFS, but is a waste of space. | |
2722 | */ | |
2723 | ||
31dbd01f IE |
2724 | #define KSM_ATTR_RO(_name) \ |
2725 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2726 | #define KSM_ATTR(_name) \ | |
2727 | static struct kobj_attribute _name##_attr = \ | |
2728 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2729 | ||
2730 | static ssize_t sleep_millisecs_show(struct kobject *kobj, | |
2731 | struct kobj_attribute *attr, char *buf) | |
2732 | { | |
2733 | return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs); | |
2734 | } | |
2735 | ||
2736 | static ssize_t sleep_millisecs_store(struct kobject *kobj, | |
2737 | struct kobj_attribute *attr, | |
2738 | const char *buf, size_t count) | |
2739 | { | |
2740 | unsigned long msecs; | |
2741 | int err; | |
2742 | ||
3dbb95f7 | 2743 | err = kstrtoul(buf, 10, &msecs); |
31dbd01f IE |
2744 | if (err || msecs > UINT_MAX) |
2745 | return -EINVAL; | |
2746 | ||
2747 | ksm_thread_sleep_millisecs = msecs; | |
2748 | ||
2749 | return count; | |
2750 | } | |
2751 | KSM_ATTR(sleep_millisecs); | |
2752 | ||
2753 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
2754 | struct kobj_attribute *attr, char *buf) | |
2755 | { | |
2756 | return sprintf(buf, "%u\n", ksm_thread_pages_to_scan); | |
2757 | } | |
2758 | ||
2759 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
2760 | struct kobj_attribute *attr, | |
2761 | const char *buf, size_t count) | |
2762 | { | |
2763 | int err; | |
2764 | unsigned long nr_pages; | |
2765 | ||
3dbb95f7 | 2766 | err = kstrtoul(buf, 10, &nr_pages); |
31dbd01f IE |
2767 | if (err || nr_pages > UINT_MAX) |
2768 | return -EINVAL; | |
2769 | ||
2770 | ksm_thread_pages_to_scan = nr_pages; | |
2771 | ||
2772 | return count; | |
2773 | } | |
2774 | KSM_ATTR(pages_to_scan); | |
2775 | ||
2776 | static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, | |
2777 | char *buf) | |
2778 | { | |
ef4d43a8 | 2779 | return sprintf(buf, "%lu\n", ksm_run); |
31dbd01f IE |
2780 | } |
2781 | ||
2782 | static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, | |
2783 | const char *buf, size_t count) | |
2784 | { | |
2785 | int err; | |
2786 | unsigned long flags; | |
2787 | ||
3dbb95f7 | 2788 | err = kstrtoul(buf, 10, &flags); |
31dbd01f IE |
2789 | if (err || flags > UINT_MAX) |
2790 | return -EINVAL; | |
2791 | if (flags > KSM_RUN_UNMERGE) | |
2792 | return -EINVAL; | |
2793 | ||
2794 | /* | |
2795 | * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. | |
2796 | * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, | |
d0f209f6 HD |
2797 | * breaking COW to free the pages_shared (but leaves mm_slots |
2798 | * on the list for when ksmd may be set running again). | |
31dbd01f IE |
2799 | */ |
2800 | ||
2801 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 2802 | wait_while_offlining(); |
31dbd01f IE |
2803 | if (ksm_run != flags) { |
2804 | ksm_run = flags; | |
d952b791 | 2805 | if (flags & KSM_RUN_UNMERGE) { |
e1e12d2f | 2806 | set_current_oom_origin(); |
d952b791 | 2807 | err = unmerge_and_remove_all_rmap_items(); |
e1e12d2f | 2808 | clear_current_oom_origin(); |
d952b791 HD |
2809 | if (err) { |
2810 | ksm_run = KSM_RUN_STOP; | |
2811 | count = err; | |
2812 | } | |
2813 | } | |
31dbd01f IE |
2814 | } |
2815 | mutex_unlock(&ksm_thread_mutex); | |
2816 | ||
2817 | if (flags & KSM_RUN_MERGE) | |
2818 | wake_up_interruptible(&ksm_thread_wait); | |
2819 | ||
2820 | return count; | |
2821 | } | |
2822 | KSM_ATTR(run); | |
2823 | ||
90bd6fd3 PH |
2824 | #ifdef CONFIG_NUMA |
2825 | static ssize_t merge_across_nodes_show(struct kobject *kobj, | |
2826 | struct kobj_attribute *attr, char *buf) | |
2827 | { | |
2828 | return sprintf(buf, "%u\n", ksm_merge_across_nodes); | |
2829 | } | |
2830 | ||
2831 | static ssize_t merge_across_nodes_store(struct kobject *kobj, | |
2832 | struct kobj_attribute *attr, | |
2833 | const char *buf, size_t count) | |
2834 | { | |
2835 | int err; | |
2836 | unsigned long knob; | |
2837 | ||
2838 | err = kstrtoul(buf, 10, &knob); | |
2839 | if (err) | |
2840 | return err; | |
2841 | if (knob > 1) | |
2842 | return -EINVAL; | |
2843 | ||
2844 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 2845 | wait_while_offlining(); |
90bd6fd3 | 2846 | if (ksm_merge_across_nodes != knob) { |
cbf86cfe | 2847 | if (ksm_pages_shared || remove_all_stable_nodes()) |
90bd6fd3 | 2848 | err = -EBUSY; |
ef53d16c HD |
2849 | else if (root_stable_tree == one_stable_tree) { |
2850 | struct rb_root *buf; | |
2851 | /* | |
2852 | * This is the first time that we switch away from the | |
2853 | * default of merging across nodes: must now allocate | |
2854 | * a buffer to hold as many roots as may be needed. | |
2855 | * Allocate stable and unstable together: | |
2856 | * MAXSMP NODES_SHIFT 10 will use 16kB. | |
2857 | */ | |
bafe1e14 JP |
2858 | buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), |
2859 | GFP_KERNEL); | |
ef53d16c HD |
2860 | /* Let us assume that RB_ROOT is NULL is zero */ |
2861 | if (!buf) | |
2862 | err = -ENOMEM; | |
2863 | else { | |
2864 | root_stable_tree = buf; | |
2865 | root_unstable_tree = buf + nr_node_ids; | |
2866 | /* Stable tree is empty but not the unstable */ | |
2867 | root_unstable_tree[0] = one_unstable_tree[0]; | |
2868 | } | |
2869 | } | |
2870 | if (!err) { | |
90bd6fd3 | 2871 | ksm_merge_across_nodes = knob; |
ef53d16c HD |
2872 | ksm_nr_node_ids = knob ? 1 : nr_node_ids; |
2873 | } | |
90bd6fd3 PH |
2874 | } |
2875 | mutex_unlock(&ksm_thread_mutex); | |
2876 | ||
2877 | return err ? err : count; | |
2878 | } | |
2879 | KSM_ATTR(merge_across_nodes); | |
2880 | #endif | |
2881 | ||
fa2aa0fc CI |
2882 | static ssize_t use_zero_pages_show(struct kobject *kobj, |
2883 | struct kobj_attribute *attr, char *buf) | |
2884 | { | |
2885 | return sprintf(buf, "%u\n", ksm_use_zero_pages); | |
2886 | } | |
2887 | static ssize_t use_zero_pages_store(struct kobject *kobj, | |
2888 | struct kobj_attribute *attr, | |
2889 | const char *buf, size_t count) | |
2890 | { | |
2891 | int err; | |
2892 | bool value; | |
2893 | ||
2894 | err = kstrtobool(buf, &value); | |
2895 | if (err) | |
2896 | return -EINVAL; | |
2897 | ||
2898 | ksm_use_zero_pages = value; | |
2899 | ||
2900 | return count; | |
2901 | } | |
2902 | KSM_ATTR(use_zero_pages); | |
2903 | ||
7253473c AA |
2904 | static ssize_t max_page_sharing_show(struct kobject *kobj, |
2905 | struct kobj_attribute *attr, char *buf) | |
2906 | { | |
2907 | return sprintf(buf, "%u\n", ksm_max_page_sharing); | |
2908 | } | |
2909 | ||
2910 | static ssize_t max_page_sharing_store(struct kobject *kobj, | |
2911 | struct kobj_attribute *attr, | |
2912 | const char *buf, size_t count) | |
2913 | { | |
2914 | int err; | |
2915 | int knob; | |
2916 | ||
2917 | err = kstrtoint(buf, 10, &knob); | |
2918 | if (err) | |
2919 | return err; | |
2920 | /* | |
2921 | * When a KSM page is created it is shared by 2 mappings. This | |
2922 | * being a signed comparison, it implicitly verifies it's not | |
2923 | * negative. | |
2924 | */ | |
2925 | if (knob < 2) | |
2926 | return -EINVAL; | |
2927 | ||
2928 | if (READ_ONCE(ksm_max_page_sharing) == knob) | |
2929 | return count; | |
2930 | ||
2931 | mutex_lock(&ksm_thread_mutex); | |
2932 | wait_while_offlining(); | |
2933 | if (ksm_max_page_sharing != knob) { | |
2934 | if (ksm_pages_shared || remove_all_stable_nodes()) | |
2935 | err = -EBUSY; | |
2936 | else | |
2937 | ksm_max_page_sharing = knob; | |
2938 | } | |
2939 | mutex_unlock(&ksm_thread_mutex); | |
2940 | ||
2941 | return err ? err : count; | |
2942 | } | |
2943 | KSM_ATTR(max_page_sharing); | |
2944 | ||
b4028260 HD |
2945 | static ssize_t pages_shared_show(struct kobject *kobj, |
2946 | struct kobj_attribute *attr, char *buf) | |
2947 | { | |
2948 | return sprintf(buf, "%lu\n", ksm_pages_shared); | |
2949 | } | |
2950 | KSM_ATTR_RO(pages_shared); | |
2951 | ||
2952 | static ssize_t pages_sharing_show(struct kobject *kobj, | |
2953 | struct kobj_attribute *attr, char *buf) | |
2954 | { | |
e178dfde | 2955 | return sprintf(buf, "%lu\n", ksm_pages_sharing); |
b4028260 HD |
2956 | } |
2957 | KSM_ATTR_RO(pages_sharing); | |
2958 | ||
473b0ce4 HD |
2959 | static ssize_t pages_unshared_show(struct kobject *kobj, |
2960 | struct kobj_attribute *attr, char *buf) | |
2961 | { | |
2962 | return sprintf(buf, "%lu\n", ksm_pages_unshared); | |
2963 | } | |
2964 | KSM_ATTR_RO(pages_unshared); | |
2965 | ||
2966 | static ssize_t pages_volatile_show(struct kobject *kobj, | |
2967 | struct kobj_attribute *attr, char *buf) | |
2968 | { | |
2969 | long ksm_pages_volatile; | |
2970 | ||
2971 | ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared | |
2972 | - ksm_pages_sharing - ksm_pages_unshared; | |
2973 | /* | |
2974 | * It was not worth any locking to calculate that statistic, | |
2975 | * but it might therefore sometimes be negative: conceal that. | |
2976 | */ | |
2977 | if (ksm_pages_volatile < 0) | |
2978 | ksm_pages_volatile = 0; | |
2979 | return sprintf(buf, "%ld\n", ksm_pages_volatile); | |
2980 | } | |
2981 | KSM_ATTR_RO(pages_volatile); | |
2982 | ||
7253473c AA |
2983 | static ssize_t stable_node_dups_show(struct kobject *kobj, |
2984 | struct kobj_attribute *attr, char *buf) | |
2985 | { | |
2986 | return sprintf(buf, "%lu\n", ksm_stable_node_dups); | |
2987 | } | |
2988 | KSM_ATTR_RO(stable_node_dups); | |
2989 | ||
2990 | static ssize_t stable_node_chains_show(struct kobject *kobj, | |
2991 | struct kobj_attribute *attr, char *buf) | |
2992 | { | |
2993 | return sprintf(buf, "%lu\n", ksm_stable_node_chains); | |
2994 | } | |
2995 | KSM_ATTR_RO(stable_node_chains); | |
2996 | ||
2997 | static ssize_t | |
2998 | stable_node_chains_prune_millisecs_show(struct kobject *kobj, | |
2999 | struct kobj_attribute *attr, | |
3000 | char *buf) | |
3001 | { | |
3002 | return sprintf(buf, "%u\n", ksm_stable_node_chains_prune_millisecs); | |
3003 | } | |
3004 | ||
3005 | static ssize_t | |
3006 | stable_node_chains_prune_millisecs_store(struct kobject *kobj, | |
3007 | struct kobj_attribute *attr, | |
3008 | const char *buf, size_t count) | |
3009 | { | |
3010 | unsigned long msecs; | |
3011 | int err; | |
3012 | ||
3013 | err = kstrtoul(buf, 10, &msecs); | |
3014 | if (err || msecs > UINT_MAX) | |
3015 | return -EINVAL; | |
3016 | ||
3017 | ksm_stable_node_chains_prune_millisecs = msecs; | |
3018 | ||
3019 | return count; | |
3020 | } | |
3021 | KSM_ATTR(stable_node_chains_prune_millisecs); | |
3022 | ||
473b0ce4 HD |
3023 | static ssize_t full_scans_show(struct kobject *kobj, |
3024 | struct kobj_attribute *attr, char *buf) | |
3025 | { | |
3026 | return sprintf(buf, "%lu\n", ksm_scan.seqnr); | |
3027 | } | |
3028 | KSM_ATTR_RO(full_scans); | |
3029 | ||
31dbd01f IE |
3030 | static struct attribute *ksm_attrs[] = { |
3031 | &sleep_millisecs_attr.attr, | |
3032 | &pages_to_scan_attr.attr, | |
3033 | &run_attr.attr, | |
b4028260 HD |
3034 | &pages_shared_attr.attr, |
3035 | &pages_sharing_attr.attr, | |
473b0ce4 HD |
3036 | &pages_unshared_attr.attr, |
3037 | &pages_volatile_attr.attr, | |
3038 | &full_scans_attr.attr, | |
90bd6fd3 PH |
3039 | #ifdef CONFIG_NUMA |
3040 | &merge_across_nodes_attr.attr, | |
3041 | #endif | |
7253473c AA |
3042 | &max_page_sharing_attr.attr, |
3043 | &stable_node_chains_attr.attr, | |
3044 | &stable_node_dups_attr.attr, | |
3045 | &stable_node_chains_prune_millisecs_attr.attr, | |
fa2aa0fc | 3046 | &use_zero_pages_attr.attr, |
31dbd01f IE |
3047 | NULL, |
3048 | }; | |
3049 | ||
3050 | static struct attribute_group ksm_attr_group = { | |
3051 | .attrs = ksm_attrs, | |
3052 | .name = "ksm", | |
3053 | }; | |
2ffd8679 | 3054 | #endif /* CONFIG_SYSFS */ |
31dbd01f IE |
3055 | |
3056 | static int __init ksm_init(void) | |
3057 | { | |
3058 | struct task_struct *ksm_thread; | |
3059 | int err; | |
3060 | ||
fa2aa0fc CI |
3061 | /* The correct value depends on page size and endianness */ |
3062 | zero_checksum = calc_checksum(ZERO_PAGE(0)); | |
3063 | /* Default to false for backwards compatibility */ | |
3064 | ksm_use_zero_pages = false; | |
3065 | ||
31dbd01f IE |
3066 | err = ksm_slab_init(); |
3067 | if (err) | |
3068 | goto out; | |
3069 | ||
31dbd01f IE |
3070 | ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); |
3071 | if (IS_ERR(ksm_thread)) { | |
25acde31 | 3072 | pr_err("ksm: creating kthread failed\n"); |
31dbd01f | 3073 | err = PTR_ERR(ksm_thread); |
d9f8984c | 3074 | goto out_free; |
31dbd01f IE |
3075 | } |
3076 | ||
2ffd8679 | 3077 | #ifdef CONFIG_SYSFS |
31dbd01f IE |
3078 | err = sysfs_create_group(mm_kobj, &ksm_attr_group); |
3079 | if (err) { | |
25acde31 | 3080 | pr_err("ksm: register sysfs failed\n"); |
2ffd8679 | 3081 | kthread_stop(ksm_thread); |
d9f8984c | 3082 | goto out_free; |
31dbd01f | 3083 | } |
c73602ad HD |
3084 | #else |
3085 | ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ | |
3086 | ||
2ffd8679 | 3087 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 3088 | |
62b61f61 | 3089 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 | 3090 | /* There is no significance to this priority 100 */ |
62b61f61 HD |
3091 | hotplug_memory_notifier(ksm_memory_callback, 100); |
3092 | #endif | |
31dbd01f IE |
3093 | return 0; |
3094 | ||
d9f8984c | 3095 | out_free: |
31dbd01f IE |
3096 | ksm_slab_free(); |
3097 | out: | |
3098 | return err; | |
f8af4da3 | 3099 | } |
a64fb3cd | 3100 | subsys_initcall(ksm_init); |