]>
Commit | Line | Data |
---|---|---|
7a338472 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
f8af4da3 | 2 | /* |
31dbd01f IE |
3 | * Memory merging support. |
4 | * | |
5 | * This code enables dynamic sharing of identical pages found in different | |
6 | * memory areas, even if they are not shared by fork() | |
7 | * | |
36b2528d | 8 | * Copyright (C) 2008-2009 Red Hat, Inc. |
31dbd01f IE |
9 | * Authors: |
10 | * Izik Eidus | |
11 | * Andrea Arcangeli | |
12 | * Chris Wright | |
36b2528d | 13 | * Hugh Dickins |
f8af4da3 HD |
14 | */ |
15 | ||
16 | #include <linux/errno.h> | |
31dbd01f IE |
17 | #include <linux/mm.h> |
18 | #include <linux/fs.h> | |
f8af4da3 | 19 | #include <linux/mman.h> |
31dbd01f | 20 | #include <linux/sched.h> |
6e84f315 | 21 | #include <linux/sched/mm.h> |
f7ccbae4 | 22 | #include <linux/sched/coredump.h> |
31dbd01f IE |
23 | #include <linux/rwsem.h> |
24 | #include <linux/pagemap.h> | |
25 | #include <linux/rmap.h> | |
26 | #include <linux/spinlock.h> | |
59e1a2f4 | 27 | #include <linux/xxhash.h> |
31dbd01f IE |
28 | #include <linux/delay.h> |
29 | #include <linux/kthread.h> | |
30 | #include <linux/wait.h> | |
31 | #include <linux/slab.h> | |
32 | #include <linux/rbtree.h> | |
62b61f61 | 33 | #include <linux/memory.h> |
31dbd01f | 34 | #include <linux/mmu_notifier.h> |
2c6854fd | 35 | #include <linux/swap.h> |
f8af4da3 | 36 | #include <linux/ksm.h> |
4ca3a69b | 37 | #include <linux/hashtable.h> |
878aee7d | 38 | #include <linux/freezer.h> |
72788c38 | 39 | #include <linux/oom.h> |
90bd6fd3 | 40 | #include <linux/numa.h> |
f8af4da3 | 41 | |
31dbd01f | 42 | #include <asm/tlbflush.h> |
73848b46 | 43 | #include "internal.h" |
31dbd01f | 44 | |
e850dcf5 HD |
45 | #ifdef CONFIG_NUMA |
46 | #define NUMA(x) (x) | |
47 | #define DO_NUMA(x) do { (x); } while (0) | |
48 | #else | |
49 | #define NUMA(x) (0) | |
50 | #define DO_NUMA(x) do { } while (0) | |
51 | #endif | |
52 | ||
5a2ca3ef MR |
53 | /** |
54 | * DOC: Overview | |
55 | * | |
31dbd01f IE |
56 | * A few notes about the KSM scanning process, |
57 | * to make it easier to understand the data structures below: | |
58 | * | |
59 | * In order to reduce excessive scanning, KSM sorts the memory pages by their | |
60 | * contents into a data structure that holds pointers to the pages' locations. | |
61 | * | |
62 | * Since the contents of the pages may change at any moment, KSM cannot just | |
63 | * insert the pages into a normal sorted tree and expect it to find anything. | |
64 | * Therefore KSM uses two data structures - the stable and the unstable tree. | |
65 | * | |
66 | * The stable tree holds pointers to all the merged pages (ksm pages), sorted | |
67 | * by their contents. Because each such page is write-protected, searching on | |
68 | * this tree is fully assured to be working (except when pages are unmapped), | |
69 | * and therefore this tree is called the stable tree. | |
70 | * | |
5a2ca3ef MR |
71 | * The stable tree node includes information required for reverse |
72 | * mapping from a KSM page to virtual addresses that map this page. | |
73 | * | |
74 | * In order to avoid large latencies of the rmap walks on KSM pages, | |
75 | * KSM maintains two types of nodes in the stable tree: | |
76 | * | |
77 | * * the regular nodes that keep the reverse mapping structures in a | |
78 | * linked list | |
79 | * * the "chains" that link nodes ("dups") that represent the same | |
80 | * write protected memory content, but each "dup" corresponds to a | |
81 | * different KSM page copy of that content | |
82 | * | |
83 | * Internally, the regular nodes, "dups" and "chains" are represented | |
9303c9d5 | 84 | * using the same struct stable_node structure. |
5a2ca3ef | 85 | * |
31dbd01f IE |
86 | * In addition to the stable tree, KSM uses a second data structure called the |
87 | * unstable tree: this tree holds pointers to pages which have been found to | |
88 | * be "unchanged for a period of time". The unstable tree sorts these pages | |
89 | * by their contents, but since they are not write-protected, KSM cannot rely | |
90 | * upon the unstable tree to work correctly - the unstable tree is liable to | |
91 | * be corrupted as its contents are modified, and so it is called unstable. | |
92 | * | |
93 | * KSM solves this problem by several techniques: | |
94 | * | |
95 | * 1) The unstable tree is flushed every time KSM completes scanning all | |
96 | * memory areas, and then the tree is rebuilt again from the beginning. | |
97 | * 2) KSM will only insert into the unstable tree, pages whose hash value | |
98 | * has not changed since the previous scan of all memory areas. | |
99 | * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the | |
100 | * colors of the nodes and not on their contents, assuring that even when | |
101 | * the tree gets "corrupted" it won't get out of balance, so scanning time | |
102 | * remains the same (also, searching and inserting nodes in an rbtree uses | |
103 | * the same algorithm, so we have no overhead when we flush and rebuild). | |
104 | * 4) KSM never flushes the stable tree, which means that even if it were to | |
105 | * take 10 attempts to find a page in the unstable tree, once it is found, | |
106 | * it is secured in the stable tree. (When we scan a new page, we first | |
107 | * compare it against the stable tree, and then against the unstable tree.) | |
8fdb3dbf HD |
108 | * |
109 | * If the merge_across_nodes tunable is unset, then KSM maintains multiple | |
110 | * stable trees and multiple unstable trees: one of each for each NUMA node. | |
31dbd01f IE |
111 | */ |
112 | ||
113 | /** | |
114 | * struct mm_slot - ksm information per mm that is being scanned | |
115 | * @link: link to the mm_slots hash list | |
116 | * @mm_list: link into the mm_slots list, rooted in ksm_mm_head | |
6514d511 | 117 | * @rmap_list: head for this mm_slot's singly-linked list of rmap_items |
31dbd01f IE |
118 | * @mm: the mm that this information is valid for |
119 | */ | |
120 | struct mm_slot { | |
121 | struct hlist_node link; | |
122 | struct list_head mm_list; | |
6514d511 | 123 | struct rmap_item *rmap_list; |
31dbd01f IE |
124 | struct mm_struct *mm; |
125 | }; | |
126 | ||
127 | /** | |
128 | * struct ksm_scan - cursor for scanning | |
129 | * @mm_slot: the current mm_slot we are scanning | |
130 | * @address: the next address inside that to be scanned | |
6514d511 | 131 | * @rmap_list: link to the next rmap to be scanned in the rmap_list |
31dbd01f IE |
132 | * @seqnr: count of completed full scans (needed when removing unstable node) |
133 | * | |
134 | * There is only the one ksm_scan instance of this cursor structure. | |
135 | */ | |
136 | struct ksm_scan { | |
137 | struct mm_slot *mm_slot; | |
138 | unsigned long address; | |
6514d511 | 139 | struct rmap_item **rmap_list; |
31dbd01f IE |
140 | unsigned long seqnr; |
141 | }; | |
142 | ||
7b6ba2c7 HD |
143 | /** |
144 | * struct stable_node - node of the stable rbtree | |
145 | * @node: rb node of this ksm page in the stable tree | |
4146d2d6 | 146 | * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list |
2c653d0e | 147 | * @hlist_dup: linked into the stable_node->hlist with a stable_node chain |
4146d2d6 | 148 | * @list: linked into migrate_nodes, pending placement in the proper node tree |
7b6ba2c7 | 149 | * @hlist: hlist head of rmap_items using this ksm page |
4146d2d6 | 150 | * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) |
2c653d0e AA |
151 | * @chain_prune_time: time of the last full garbage collection |
152 | * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN | |
4146d2d6 | 153 | * @nid: NUMA node id of stable tree in which linked (may not match kpfn) |
7b6ba2c7 HD |
154 | */ |
155 | struct stable_node { | |
4146d2d6 HD |
156 | union { |
157 | struct rb_node node; /* when node of stable tree */ | |
158 | struct { /* when listed for migration */ | |
159 | struct list_head *head; | |
2c653d0e AA |
160 | struct { |
161 | struct hlist_node hlist_dup; | |
162 | struct list_head list; | |
163 | }; | |
4146d2d6 HD |
164 | }; |
165 | }; | |
7b6ba2c7 | 166 | struct hlist_head hlist; |
2c653d0e AA |
167 | union { |
168 | unsigned long kpfn; | |
169 | unsigned long chain_prune_time; | |
170 | }; | |
171 | /* | |
172 | * STABLE_NODE_CHAIN can be any negative number in | |
173 | * rmap_hlist_len negative range, but better not -1 to be able | |
174 | * to reliably detect underflows. | |
175 | */ | |
176 | #define STABLE_NODE_CHAIN -1024 | |
177 | int rmap_hlist_len; | |
4146d2d6 HD |
178 | #ifdef CONFIG_NUMA |
179 | int nid; | |
180 | #endif | |
7b6ba2c7 HD |
181 | }; |
182 | ||
31dbd01f IE |
183 | /** |
184 | * struct rmap_item - reverse mapping item for virtual addresses | |
6514d511 | 185 | * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list |
db114b83 | 186 | * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree |
bc56620b | 187 | * @nid: NUMA node id of unstable tree in which linked (may not match page) |
31dbd01f IE |
188 | * @mm: the memory structure this rmap_item is pointing into |
189 | * @address: the virtual address this rmap_item tracks (+ flags in low bits) | |
190 | * @oldchecksum: previous checksum of the page at that virtual address | |
7b6ba2c7 HD |
191 | * @node: rb node of this rmap_item in the unstable tree |
192 | * @head: pointer to stable_node heading this list in the stable tree | |
193 | * @hlist: link into hlist of rmap_items hanging off that stable_node | |
31dbd01f IE |
194 | */ |
195 | struct rmap_item { | |
6514d511 | 196 | struct rmap_item *rmap_list; |
bc56620b HD |
197 | union { |
198 | struct anon_vma *anon_vma; /* when stable */ | |
199 | #ifdef CONFIG_NUMA | |
200 | int nid; /* when node of unstable tree */ | |
201 | #endif | |
202 | }; | |
31dbd01f IE |
203 | struct mm_struct *mm; |
204 | unsigned long address; /* + low bits used for flags below */ | |
7b6ba2c7 | 205 | unsigned int oldchecksum; /* when unstable */ |
31dbd01f | 206 | union { |
7b6ba2c7 HD |
207 | struct rb_node node; /* when node of unstable tree */ |
208 | struct { /* when listed from stable tree */ | |
209 | struct stable_node *head; | |
210 | struct hlist_node hlist; | |
211 | }; | |
31dbd01f IE |
212 | }; |
213 | }; | |
214 | ||
215 | #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ | |
7b6ba2c7 HD |
216 | #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ |
217 | #define STABLE_FLAG 0x200 /* is listed from the stable tree */ | |
31dbd01f IE |
218 | |
219 | /* The stable and unstable tree heads */ | |
ef53d16c HD |
220 | static struct rb_root one_stable_tree[1] = { RB_ROOT }; |
221 | static struct rb_root one_unstable_tree[1] = { RB_ROOT }; | |
222 | static struct rb_root *root_stable_tree = one_stable_tree; | |
223 | static struct rb_root *root_unstable_tree = one_unstable_tree; | |
31dbd01f | 224 | |
4146d2d6 HD |
225 | /* Recently migrated nodes of stable tree, pending proper placement */ |
226 | static LIST_HEAD(migrate_nodes); | |
2c653d0e | 227 | #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev) |
4146d2d6 | 228 | |
4ca3a69b SL |
229 | #define MM_SLOTS_HASH_BITS 10 |
230 | static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
31dbd01f IE |
231 | |
232 | static struct mm_slot ksm_mm_head = { | |
233 | .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list), | |
234 | }; | |
235 | static struct ksm_scan ksm_scan = { | |
236 | .mm_slot = &ksm_mm_head, | |
237 | }; | |
238 | ||
239 | static struct kmem_cache *rmap_item_cache; | |
7b6ba2c7 | 240 | static struct kmem_cache *stable_node_cache; |
31dbd01f IE |
241 | static struct kmem_cache *mm_slot_cache; |
242 | ||
243 | /* The number of nodes in the stable tree */ | |
b4028260 | 244 | static unsigned long ksm_pages_shared; |
31dbd01f | 245 | |
e178dfde | 246 | /* The number of page slots additionally sharing those nodes */ |
b4028260 | 247 | static unsigned long ksm_pages_sharing; |
31dbd01f | 248 | |
473b0ce4 HD |
249 | /* The number of nodes in the unstable tree */ |
250 | static unsigned long ksm_pages_unshared; | |
251 | ||
252 | /* The number of rmap_items in use: to calculate pages_volatile */ | |
253 | static unsigned long ksm_rmap_items; | |
254 | ||
2c653d0e AA |
255 | /* The number of stable_node chains */ |
256 | static unsigned long ksm_stable_node_chains; | |
257 | ||
258 | /* The number of stable_node dups linked to the stable_node chains */ | |
259 | static unsigned long ksm_stable_node_dups; | |
260 | ||
261 | /* Delay in pruning stale stable_node_dups in the stable_node_chains */ | |
584ff0df | 262 | static unsigned int ksm_stable_node_chains_prune_millisecs = 2000; |
2c653d0e AA |
263 | |
264 | /* Maximum number of page slots sharing a stable node */ | |
265 | static int ksm_max_page_sharing = 256; | |
266 | ||
31dbd01f | 267 | /* Number of pages ksmd should scan in one batch */ |
2c6854fd | 268 | static unsigned int ksm_thread_pages_to_scan = 100; |
31dbd01f IE |
269 | |
270 | /* Milliseconds ksmd should sleep between batches */ | |
2ffd8679 | 271 | static unsigned int ksm_thread_sleep_millisecs = 20; |
31dbd01f | 272 | |
e86c59b1 CI |
273 | /* Checksum of an empty (zeroed) page */ |
274 | static unsigned int zero_checksum __read_mostly; | |
275 | ||
276 | /* Whether to merge empty (zeroed) pages with actual zero pages */ | |
277 | static bool ksm_use_zero_pages __read_mostly; | |
278 | ||
e850dcf5 | 279 | #ifdef CONFIG_NUMA |
90bd6fd3 PH |
280 | /* Zeroed when merging across nodes is not allowed */ |
281 | static unsigned int ksm_merge_across_nodes = 1; | |
ef53d16c | 282 | static int ksm_nr_node_ids = 1; |
e850dcf5 HD |
283 | #else |
284 | #define ksm_merge_across_nodes 1U | |
ef53d16c | 285 | #define ksm_nr_node_ids 1 |
e850dcf5 | 286 | #endif |
90bd6fd3 | 287 | |
31dbd01f IE |
288 | #define KSM_RUN_STOP 0 |
289 | #define KSM_RUN_MERGE 1 | |
290 | #define KSM_RUN_UNMERGE 2 | |
ef4d43a8 HD |
291 | #define KSM_RUN_OFFLINE 4 |
292 | static unsigned long ksm_run = KSM_RUN_STOP; | |
293 | static void wait_while_offlining(void); | |
31dbd01f IE |
294 | |
295 | static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); | |
fcf9a0ef | 296 | static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait); |
31dbd01f IE |
297 | static DEFINE_MUTEX(ksm_thread_mutex); |
298 | static DEFINE_SPINLOCK(ksm_mmlist_lock); | |
299 | ||
300 | #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\ | |
301 | sizeof(struct __struct), __alignof__(struct __struct),\ | |
302 | (__flags), NULL) | |
303 | ||
304 | static int __init ksm_slab_init(void) | |
305 | { | |
306 | rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0); | |
307 | if (!rmap_item_cache) | |
308 | goto out; | |
309 | ||
7b6ba2c7 HD |
310 | stable_node_cache = KSM_KMEM_CACHE(stable_node, 0); |
311 | if (!stable_node_cache) | |
312 | goto out_free1; | |
313 | ||
31dbd01f IE |
314 | mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0); |
315 | if (!mm_slot_cache) | |
7b6ba2c7 | 316 | goto out_free2; |
31dbd01f IE |
317 | |
318 | return 0; | |
319 | ||
7b6ba2c7 HD |
320 | out_free2: |
321 | kmem_cache_destroy(stable_node_cache); | |
322 | out_free1: | |
31dbd01f IE |
323 | kmem_cache_destroy(rmap_item_cache); |
324 | out: | |
325 | return -ENOMEM; | |
326 | } | |
327 | ||
328 | static void __init ksm_slab_free(void) | |
329 | { | |
330 | kmem_cache_destroy(mm_slot_cache); | |
7b6ba2c7 | 331 | kmem_cache_destroy(stable_node_cache); |
31dbd01f IE |
332 | kmem_cache_destroy(rmap_item_cache); |
333 | mm_slot_cache = NULL; | |
334 | } | |
335 | ||
2c653d0e AA |
336 | static __always_inline bool is_stable_node_chain(struct stable_node *chain) |
337 | { | |
338 | return chain->rmap_hlist_len == STABLE_NODE_CHAIN; | |
339 | } | |
340 | ||
341 | static __always_inline bool is_stable_node_dup(struct stable_node *dup) | |
342 | { | |
343 | return dup->head == STABLE_NODE_DUP_HEAD; | |
344 | } | |
345 | ||
346 | static inline void stable_node_chain_add_dup(struct stable_node *dup, | |
347 | struct stable_node *chain) | |
348 | { | |
349 | VM_BUG_ON(is_stable_node_dup(dup)); | |
350 | dup->head = STABLE_NODE_DUP_HEAD; | |
351 | VM_BUG_ON(!is_stable_node_chain(chain)); | |
352 | hlist_add_head(&dup->hlist_dup, &chain->hlist); | |
353 | ksm_stable_node_dups++; | |
354 | } | |
355 | ||
356 | static inline void __stable_node_dup_del(struct stable_node *dup) | |
357 | { | |
b4fecc67 | 358 | VM_BUG_ON(!is_stable_node_dup(dup)); |
2c653d0e AA |
359 | hlist_del(&dup->hlist_dup); |
360 | ksm_stable_node_dups--; | |
361 | } | |
362 | ||
363 | static inline void stable_node_dup_del(struct stable_node *dup) | |
364 | { | |
365 | VM_BUG_ON(is_stable_node_chain(dup)); | |
366 | if (is_stable_node_dup(dup)) | |
367 | __stable_node_dup_del(dup); | |
368 | else | |
369 | rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid)); | |
370 | #ifdef CONFIG_DEBUG_VM | |
371 | dup->head = NULL; | |
372 | #endif | |
373 | } | |
374 | ||
31dbd01f IE |
375 | static inline struct rmap_item *alloc_rmap_item(void) |
376 | { | |
473b0ce4 HD |
377 | struct rmap_item *rmap_item; |
378 | ||
5b398e41 | 379 | rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | |
380 | __GFP_NORETRY | __GFP_NOWARN); | |
473b0ce4 HD |
381 | if (rmap_item) |
382 | ksm_rmap_items++; | |
383 | return rmap_item; | |
31dbd01f IE |
384 | } |
385 | ||
386 | static inline void free_rmap_item(struct rmap_item *rmap_item) | |
387 | { | |
473b0ce4 | 388 | ksm_rmap_items--; |
31dbd01f IE |
389 | rmap_item->mm = NULL; /* debug safety */ |
390 | kmem_cache_free(rmap_item_cache, rmap_item); | |
391 | } | |
392 | ||
7b6ba2c7 HD |
393 | static inline struct stable_node *alloc_stable_node(void) |
394 | { | |
6213055f | 395 | /* |
396 | * The allocation can take too long with GFP_KERNEL when memory is under | |
397 | * pressure, which may lead to hung task warnings. Adding __GFP_HIGH | |
398 | * grants access to memory reserves, helping to avoid this problem. | |
399 | */ | |
400 | return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); | |
7b6ba2c7 HD |
401 | } |
402 | ||
403 | static inline void free_stable_node(struct stable_node *stable_node) | |
404 | { | |
2c653d0e AA |
405 | VM_BUG_ON(stable_node->rmap_hlist_len && |
406 | !is_stable_node_chain(stable_node)); | |
7b6ba2c7 HD |
407 | kmem_cache_free(stable_node_cache, stable_node); |
408 | } | |
409 | ||
31dbd01f IE |
410 | static inline struct mm_slot *alloc_mm_slot(void) |
411 | { | |
412 | if (!mm_slot_cache) /* initialization failed */ | |
413 | return NULL; | |
414 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
415 | } | |
416 | ||
417 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
418 | { | |
419 | kmem_cache_free(mm_slot_cache, mm_slot); | |
420 | } | |
421 | ||
31dbd01f IE |
422 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) |
423 | { | |
4ca3a69b SL |
424 | struct mm_slot *slot; |
425 | ||
b67bfe0d | 426 | hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm) |
4ca3a69b SL |
427 | if (slot->mm == mm) |
428 | return slot; | |
31dbd01f | 429 | |
31dbd01f IE |
430 | return NULL; |
431 | } | |
432 | ||
433 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
434 | struct mm_slot *mm_slot) | |
435 | { | |
31dbd01f | 436 | mm_slot->mm = mm; |
4ca3a69b | 437 | hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm); |
31dbd01f IE |
438 | } |
439 | ||
a913e182 HD |
440 | /* |
441 | * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's | |
442 | * page tables after it has passed through ksm_exit() - which, if necessary, | |
c1e8d7c6 | 443 | * takes mmap_lock briefly to serialize against them. ksm_exit() does not set |
a913e182 HD |
444 | * a special flag: they can just back out as soon as mm_users goes to zero. |
445 | * ksm_test_exit() is used throughout to make this test for exit: in some | |
446 | * places for correctness, in some places just to avoid unnecessary work. | |
447 | */ | |
448 | static inline bool ksm_test_exit(struct mm_struct *mm) | |
449 | { | |
450 | return atomic_read(&mm->mm_users) == 0; | |
451 | } | |
452 | ||
31dbd01f IE |
453 | /* |
454 | * We use break_ksm to break COW on a ksm page: it's a stripped down | |
455 | * | |
7a9547fd | 456 | * if (get_user_pages(addr, 1, FOLL_WRITE, &page, NULL) == 1) |
31dbd01f IE |
457 | * put_page(page); |
458 | * | |
459 | * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma, | |
460 | * in case the application has unmapped and remapped mm,addr meanwhile. | |
461 | * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP | |
bbcd53c9 | 462 | * mmap of /dev/mem, where we would not want to touch it. |
1b2ee126 DH |
463 | * |
464 | * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context | |
465 | * of the process that owns 'vma'. We also do not want to enforce | |
466 | * protection keys here anyway. | |
31dbd01f | 467 | */ |
d952b791 | 468 | static int break_ksm(struct vm_area_struct *vma, unsigned long addr) |
31dbd01f IE |
469 | { |
470 | struct page *page; | |
50a7ca3c | 471 | vm_fault_t ret = 0; |
31dbd01f IE |
472 | |
473 | do { | |
474 | cond_resched(); | |
1b2ee126 DH |
475 | page = follow_page(vma, addr, |
476 | FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE); | |
22eccdd7 | 477 | if (IS_ERR_OR_NULL(page)) |
31dbd01f IE |
478 | break; |
479 | if (PageKsm(page)) | |
dcddffd4 | 480 | ret = handle_mm_fault(vma, addr, |
bce617ed PX |
481 | FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE, |
482 | NULL); | |
31dbd01f IE |
483 | else |
484 | ret = VM_FAULT_WRITE; | |
485 | put_page(page); | |
33692f27 | 486 | } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); |
d952b791 HD |
487 | /* |
488 | * We must loop because handle_mm_fault() may back out if there's | |
489 | * any difficulty e.g. if pte accessed bit gets updated concurrently. | |
490 | * | |
491 | * VM_FAULT_WRITE is what we have been hoping for: it indicates that | |
492 | * COW has been broken, even if the vma does not permit VM_WRITE; | |
493 | * but note that a concurrent fault might break PageKsm for us. | |
494 | * | |
495 | * VM_FAULT_SIGBUS could occur if we race with truncation of the | |
496 | * backing file, which also invalidates anonymous pages: that's | |
497 | * okay, that truncation will have unmapped the PageKsm for us. | |
498 | * | |
499 | * VM_FAULT_OOM: at the time of writing (late July 2009), setting | |
500 | * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the | |
501 | * current task has TIF_MEMDIE set, and will be OOM killed on return | |
502 | * to user; and ksmd, having no mm, would never be chosen for that. | |
503 | * | |
504 | * But if the mm is in a limited mem_cgroup, then the fault may fail | |
505 | * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and | |
506 | * even ksmd can fail in this way - though it's usually breaking ksm | |
507 | * just to undo a merge it made a moment before, so unlikely to oom. | |
508 | * | |
509 | * That's a pity: we might therefore have more kernel pages allocated | |
510 | * than we're counting as nodes in the stable tree; but ksm_do_scan | |
511 | * will retry to break_cow on each pass, so should recover the page | |
512 | * in due course. The important thing is to not let VM_MERGEABLE | |
513 | * be cleared while any such pages might remain in the area. | |
514 | */ | |
515 | return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; | |
31dbd01f IE |
516 | } |
517 | ||
ef694222 BL |
518 | static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, |
519 | unsigned long addr) | |
520 | { | |
521 | struct vm_area_struct *vma; | |
522 | if (ksm_test_exit(mm)) | |
523 | return NULL; | |
ff69fb81 LH |
524 | vma = vma_lookup(mm, addr); |
525 | if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) | |
ef694222 BL |
526 | return NULL; |
527 | return vma; | |
528 | } | |
529 | ||
8dd3557a | 530 | static void break_cow(struct rmap_item *rmap_item) |
31dbd01f | 531 | { |
8dd3557a HD |
532 | struct mm_struct *mm = rmap_item->mm; |
533 | unsigned long addr = rmap_item->address; | |
31dbd01f IE |
534 | struct vm_area_struct *vma; |
535 | ||
4035c07a HD |
536 | /* |
537 | * It is not an accident that whenever we want to break COW | |
538 | * to undo, we also need to drop a reference to the anon_vma. | |
539 | */ | |
9e60109f | 540 | put_anon_vma(rmap_item->anon_vma); |
4035c07a | 541 | |
d8ed45c5 | 542 | mmap_read_lock(mm); |
ef694222 BL |
543 | vma = find_mergeable_vma(mm, addr); |
544 | if (vma) | |
545 | break_ksm(vma, addr); | |
d8ed45c5 | 546 | mmap_read_unlock(mm); |
31dbd01f IE |
547 | } |
548 | ||
549 | static struct page *get_mergeable_page(struct rmap_item *rmap_item) | |
550 | { | |
551 | struct mm_struct *mm = rmap_item->mm; | |
552 | unsigned long addr = rmap_item->address; | |
553 | struct vm_area_struct *vma; | |
554 | struct page *page; | |
555 | ||
d8ed45c5 | 556 | mmap_read_lock(mm); |
ef694222 BL |
557 | vma = find_mergeable_vma(mm, addr); |
558 | if (!vma) | |
31dbd01f IE |
559 | goto out; |
560 | ||
561 | page = follow_page(vma, addr, FOLL_GET); | |
22eccdd7 | 562 | if (IS_ERR_OR_NULL(page)) |
31dbd01f | 563 | goto out; |
f765f540 | 564 | if (PageAnon(page)) { |
31dbd01f IE |
565 | flush_anon_page(vma, page, addr); |
566 | flush_dcache_page(page); | |
567 | } else { | |
568 | put_page(page); | |
c8f95ed1 AA |
569 | out: |
570 | page = NULL; | |
31dbd01f | 571 | } |
d8ed45c5 | 572 | mmap_read_unlock(mm); |
31dbd01f IE |
573 | return page; |
574 | } | |
575 | ||
90bd6fd3 PH |
576 | /* |
577 | * This helper is used for getting right index into array of tree roots. | |
578 | * When merge_across_nodes knob is set to 1, there are only two rb-trees for | |
579 | * stable and unstable pages from all nodes with roots in index 0. Otherwise, | |
580 | * every node has its own stable and unstable tree. | |
581 | */ | |
582 | static inline int get_kpfn_nid(unsigned long kpfn) | |
583 | { | |
d8fc16a8 | 584 | return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); |
90bd6fd3 PH |
585 | } |
586 | ||
2c653d0e AA |
587 | static struct stable_node *alloc_stable_node_chain(struct stable_node *dup, |
588 | struct rb_root *root) | |
589 | { | |
590 | struct stable_node *chain = alloc_stable_node(); | |
591 | VM_BUG_ON(is_stable_node_chain(dup)); | |
592 | if (likely(chain)) { | |
593 | INIT_HLIST_HEAD(&chain->hlist); | |
594 | chain->chain_prune_time = jiffies; | |
595 | chain->rmap_hlist_len = STABLE_NODE_CHAIN; | |
596 | #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA) | |
98fa15f3 | 597 | chain->nid = NUMA_NO_NODE; /* debug */ |
2c653d0e AA |
598 | #endif |
599 | ksm_stable_node_chains++; | |
600 | ||
601 | /* | |
602 | * Put the stable node chain in the first dimension of | |
603 | * the stable tree and at the same time remove the old | |
604 | * stable node. | |
605 | */ | |
606 | rb_replace_node(&dup->node, &chain->node, root); | |
607 | ||
608 | /* | |
609 | * Move the old stable node to the second dimension | |
610 | * queued in the hlist_dup. The invariant is that all | |
611 | * dup stable_nodes in the chain->hlist point to pages | |
457aef94 | 612 | * that are write protected and have the exact same |
2c653d0e AA |
613 | * content. |
614 | */ | |
615 | stable_node_chain_add_dup(dup, chain); | |
616 | } | |
617 | return chain; | |
618 | } | |
619 | ||
620 | static inline void free_stable_node_chain(struct stable_node *chain, | |
621 | struct rb_root *root) | |
622 | { | |
623 | rb_erase(&chain->node, root); | |
624 | free_stable_node(chain); | |
625 | ksm_stable_node_chains--; | |
626 | } | |
627 | ||
4035c07a HD |
628 | static void remove_node_from_stable_tree(struct stable_node *stable_node) |
629 | { | |
630 | struct rmap_item *rmap_item; | |
4035c07a | 631 | |
2c653d0e AA |
632 | /* check it's not STABLE_NODE_CHAIN or negative */ |
633 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
634 | ||
b67bfe0d | 635 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
4035c07a HD |
636 | if (rmap_item->hlist.next) |
637 | ksm_pages_sharing--; | |
638 | else | |
639 | ksm_pages_shared--; | |
2c653d0e AA |
640 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
641 | stable_node->rmap_hlist_len--; | |
9e60109f | 642 | put_anon_vma(rmap_item->anon_vma); |
4035c07a HD |
643 | rmap_item->address &= PAGE_MASK; |
644 | cond_resched(); | |
645 | } | |
646 | ||
2c653d0e AA |
647 | /* |
648 | * We need the second aligned pointer of the migrate_nodes | |
649 | * list_head to stay clear from the rb_parent_color union | |
650 | * (aligned and different than any node) and also different | |
651 | * from &migrate_nodes. This will verify that future list.h changes | |
815f0ddb | 652 | * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it. |
2c653d0e | 653 | */ |
815f0ddb | 654 | #if defined(GCC_VERSION) && GCC_VERSION >= 40903 |
2c653d0e AA |
655 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes); |
656 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1); | |
657 | #endif | |
658 | ||
4146d2d6 HD |
659 | if (stable_node->head == &migrate_nodes) |
660 | list_del(&stable_node->list); | |
661 | else | |
2c653d0e | 662 | stable_node_dup_del(stable_node); |
4035c07a HD |
663 | free_stable_node(stable_node); |
664 | } | |
665 | ||
2cee57d1 YS |
666 | enum get_ksm_page_flags { |
667 | GET_KSM_PAGE_NOLOCK, | |
668 | GET_KSM_PAGE_LOCK, | |
669 | GET_KSM_PAGE_TRYLOCK | |
670 | }; | |
671 | ||
4035c07a HD |
672 | /* |
673 | * get_ksm_page: checks if the page indicated by the stable node | |
674 | * is still its ksm page, despite having held no reference to it. | |
675 | * In which case we can trust the content of the page, and it | |
676 | * returns the gotten page; but if the page has now been zapped, | |
677 | * remove the stale node from the stable tree and return NULL. | |
c8d6553b | 678 | * But beware, the stable node's page might be being migrated. |
4035c07a HD |
679 | * |
680 | * You would expect the stable_node to hold a reference to the ksm page. | |
681 | * But if it increments the page's count, swapping out has to wait for | |
682 | * ksmd to come around again before it can free the page, which may take | |
683 | * seconds or even minutes: much too unresponsive. So instead we use a | |
684 | * "keyhole reference": access to the ksm page from the stable node peeps | |
685 | * out through its keyhole to see if that page still holds the right key, | |
686 | * pointing back to this stable node. This relies on freeing a PageAnon | |
687 | * page to reset its page->mapping to NULL, and relies on no other use of | |
688 | * a page to put something that might look like our key in page->mapping. | |
4035c07a HD |
689 | * is on its way to being freed; but it is an anomaly to bear in mind. |
690 | */ | |
2cee57d1 YS |
691 | static struct page *get_ksm_page(struct stable_node *stable_node, |
692 | enum get_ksm_page_flags flags) | |
4035c07a HD |
693 | { |
694 | struct page *page; | |
695 | void *expected_mapping; | |
c8d6553b | 696 | unsigned long kpfn; |
4035c07a | 697 | |
bda807d4 MK |
698 | expected_mapping = (void *)((unsigned long)stable_node | |
699 | PAGE_MAPPING_KSM); | |
c8d6553b | 700 | again: |
08df4774 | 701 | kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */ |
c8d6553b | 702 | page = pfn_to_page(kpfn); |
4db0c3c2 | 703 | if (READ_ONCE(page->mapping) != expected_mapping) |
4035c07a | 704 | goto stale; |
c8d6553b HD |
705 | |
706 | /* | |
707 | * We cannot do anything with the page while its refcount is 0. | |
708 | * Usually 0 means free, or tail of a higher-order page: in which | |
709 | * case this node is no longer referenced, and should be freed; | |
1c4c3b99 | 710 | * however, it might mean that the page is under page_ref_freeze(). |
c8d6553b | 711 | * The __remove_mapping() case is easy, again the node is now stale; |
52d1e606 KT |
712 | * the same is in reuse_ksm_page() case; but if page is swapcache |
713 | * in migrate_page_move_mapping(), it might still be our page, | |
714 | * in which case it's essential to keep the node. | |
c8d6553b HD |
715 | */ |
716 | while (!get_page_unless_zero(page)) { | |
717 | /* | |
718 | * Another check for page->mapping != expected_mapping would | |
719 | * work here too. We have chosen the !PageSwapCache test to | |
720 | * optimize the common case, when the page is or is about to | |
721 | * be freed: PageSwapCache is cleared (under spin_lock_irq) | |
1c4c3b99 | 722 | * in the ref_freeze section of __remove_mapping(); but Anon |
c8d6553b HD |
723 | * page->mapping reset to NULL later, in free_pages_prepare(). |
724 | */ | |
725 | if (!PageSwapCache(page)) | |
726 | goto stale; | |
727 | cpu_relax(); | |
728 | } | |
729 | ||
4db0c3c2 | 730 | if (READ_ONCE(page->mapping) != expected_mapping) { |
4035c07a HD |
731 | put_page(page); |
732 | goto stale; | |
733 | } | |
c8d6553b | 734 | |
2cee57d1 YS |
735 | if (flags == GET_KSM_PAGE_TRYLOCK) { |
736 | if (!trylock_page(page)) { | |
737 | put_page(page); | |
738 | return ERR_PTR(-EBUSY); | |
739 | } | |
740 | } else if (flags == GET_KSM_PAGE_LOCK) | |
8aafa6a4 | 741 | lock_page(page); |
2cee57d1 YS |
742 | |
743 | if (flags != GET_KSM_PAGE_NOLOCK) { | |
4db0c3c2 | 744 | if (READ_ONCE(page->mapping) != expected_mapping) { |
8aafa6a4 HD |
745 | unlock_page(page); |
746 | put_page(page); | |
747 | goto stale; | |
748 | } | |
749 | } | |
4035c07a | 750 | return page; |
c8d6553b | 751 | |
4035c07a | 752 | stale: |
c8d6553b HD |
753 | /* |
754 | * We come here from above when page->mapping or !PageSwapCache | |
755 | * suggests that the node is stale; but it might be under migration. | |
756 | * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(), | |
757 | * before checking whether node->kpfn has been changed. | |
758 | */ | |
759 | smp_rmb(); | |
4db0c3c2 | 760 | if (READ_ONCE(stable_node->kpfn) != kpfn) |
c8d6553b | 761 | goto again; |
4035c07a HD |
762 | remove_node_from_stable_tree(stable_node); |
763 | return NULL; | |
764 | } | |
765 | ||
31dbd01f IE |
766 | /* |
767 | * Removing rmap_item from stable or unstable tree. | |
768 | * This function will clean the information from the stable/unstable tree. | |
769 | */ | |
770 | static void remove_rmap_item_from_tree(struct rmap_item *rmap_item) | |
771 | { | |
7b6ba2c7 HD |
772 | if (rmap_item->address & STABLE_FLAG) { |
773 | struct stable_node *stable_node; | |
5ad64688 | 774 | struct page *page; |
31dbd01f | 775 | |
7b6ba2c7 | 776 | stable_node = rmap_item->head; |
62862290 | 777 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
4035c07a HD |
778 | if (!page) |
779 | goto out; | |
5ad64688 | 780 | |
7b6ba2c7 | 781 | hlist_del(&rmap_item->hlist); |
62862290 | 782 | unlock_page(page); |
4035c07a | 783 | put_page(page); |
08beca44 | 784 | |
98666f8a | 785 | if (!hlist_empty(&stable_node->hlist)) |
4035c07a HD |
786 | ksm_pages_sharing--; |
787 | else | |
7b6ba2c7 | 788 | ksm_pages_shared--; |
2c653d0e AA |
789 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
790 | stable_node->rmap_hlist_len--; | |
31dbd01f | 791 | |
9e60109f | 792 | put_anon_vma(rmap_item->anon_vma); |
c89a384e | 793 | rmap_item->head = NULL; |
93d17715 | 794 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 795 | |
7b6ba2c7 | 796 | } else if (rmap_item->address & UNSTABLE_FLAG) { |
31dbd01f IE |
797 | unsigned char age; |
798 | /* | |
9ba69294 | 799 | * Usually ksmd can and must skip the rb_erase, because |
31dbd01f | 800 | * root_unstable_tree was already reset to RB_ROOT. |
9ba69294 HD |
801 | * But be careful when an mm is exiting: do the rb_erase |
802 | * if this rmap_item was inserted by this scan, rather | |
803 | * than left over from before. | |
31dbd01f IE |
804 | */ |
805 | age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); | |
cd551f97 | 806 | BUG_ON(age > 1); |
31dbd01f | 807 | if (!age) |
90bd6fd3 | 808 | rb_erase(&rmap_item->node, |
ef53d16c | 809 | root_unstable_tree + NUMA(rmap_item->nid)); |
473b0ce4 | 810 | ksm_pages_unshared--; |
93d17715 | 811 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 812 | } |
4035c07a | 813 | out: |
31dbd01f IE |
814 | cond_resched(); /* we're called from many long loops */ |
815 | } | |
816 | ||
420be4ed | 817 | static void remove_trailing_rmap_items(struct rmap_item **rmap_list) |
31dbd01f | 818 | { |
6514d511 HD |
819 | while (*rmap_list) { |
820 | struct rmap_item *rmap_item = *rmap_list; | |
821 | *rmap_list = rmap_item->rmap_list; | |
31dbd01f | 822 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
823 | free_rmap_item(rmap_item); |
824 | } | |
825 | } | |
826 | ||
827 | /* | |
e850dcf5 | 828 | * Though it's very tempting to unmerge rmap_items from stable tree rather |
31dbd01f IE |
829 | * than check every pte of a given vma, the locking doesn't quite work for |
830 | * that - an rmap_item is assigned to the stable tree after inserting ksm | |
c1e8d7c6 | 831 | * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing |
31dbd01f IE |
832 | * rmap_items from parent to child at fork time (so as not to waste time |
833 | * if exit comes before the next scan reaches it). | |
81464e30 HD |
834 | * |
835 | * Similarly, although we'd like to remove rmap_items (so updating counts | |
836 | * and freeing memory) when unmerging an area, it's easier to leave that | |
837 | * to the next pass of ksmd - consider, for example, how ksmd might be | |
838 | * in cmp_and_merge_page on one of the rmap_items we would be removing. | |
31dbd01f | 839 | */ |
d952b791 HD |
840 | static int unmerge_ksm_pages(struct vm_area_struct *vma, |
841 | unsigned long start, unsigned long end) | |
31dbd01f IE |
842 | { |
843 | unsigned long addr; | |
d952b791 | 844 | int err = 0; |
31dbd01f | 845 | |
d952b791 | 846 | for (addr = start; addr < end && !err; addr += PAGE_SIZE) { |
9ba69294 HD |
847 | if (ksm_test_exit(vma->vm_mm)) |
848 | break; | |
d952b791 HD |
849 | if (signal_pending(current)) |
850 | err = -ERESTARTSYS; | |
851 | else | |
852 | err = break_ksm(vma, addr); | |
853 | } | |
854 | return err; | |
31dbd01f IE |
855 | } |
856 | ||
88484826 MR |
857 | static inline struct stable_node *page_stable_node(struct page *page) |
858 | { | |
859 | return PageKsm(page) ? page_rmapping(page) : NULL; | |
860 | } | |
861 | ||
862 | static inline void set_page_stable_node(struct page *page, | |
863 | struct stable_node *stable_node) | |
864 | { | |
865 | page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM); | |
866 | } | |
867 | ||
2ffd8679 HD |
868 | #ifdef CONFIG_SYSFS |
869 | /* | |
870 | * Only called through the sysfs control interface: | |
871 | */ | |
cbf86cfe HD |
872 | static int remove_stable_node(struct stable_node *stable_node) |
873 | { | |
874 | struct page *page; | |
875 | int err; | |
876 | ||
2cee57d1 | 877 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
cbf86cfe HD |
878 | if (!page) { |
879 | /* | |
880 | * get_ksm_page did remove_node_from_stable_tree itself. | |
881 | */ | |
882 | return 0; | |
883 | } | |
884 | ||
9a63236f AR |
885 | /* |
886 | * Page could be still mapped if this races with __mmput() running in | |
887 | * between ksm_exit() and exit_mmap(). Just refuse to let | |
888 | * merge_across_nodes/max_page_sharing be switched. | |
889 | */ | |
890 | err = -EBUSY; | |
891 | if (!page_mapped(page)) { | |
cbf86cfe | 892 | /* |
8fdb3dbf HD |
893 | * The stable node did not yet appear stale to get_ksm_page(), |
894 | * since that allows for an unmapped ksm page to be recognized | |
895 | * right up until it is freed; but the node is safe to remove. | |
cbf86cfe HD |
896 | * This page might be in a pagevec waiting to be freed, |
897 | * or it might be PageSwapCache (perhaps under writeback), | |
898 | * or it might have been removed from swapcache a moment ago. | |
899 | */ | |
900 | set_page_stable_node(page, NULL); | |
901 | remove_node_from_stable_tree(stable_node); | |
902 | err = 0; | |
903 | } | |
904 | ||
905 | unlock_page(page); | |
906 | put_page(page); | |
907 | return err; | |
908 | } | |
909 | ||
2c653d0e AA |
910 | static int remove_stable_node_chain(struct stable_node *stable_node, |
911 | struct rb_root *root) | |
912 | { | |
913 | struct stable_node *dup; | |
914 | struct hlist_node *hlist_safe; | |
915 | ||
916 | if (!is_stable_node_chain(stable_node)) { | |
917 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
918 | if (remove_stable_node(stable_node)) | |
919 | return true; | |
920 | else | |
921 | return false; | |
922 | } | |
923 | ||
924 | hlist_for_each_entry_safe(dup, hlist_safe, | |
925 | &stable_node->hlist, hlist_dup) { | |
926 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
927 | if (remove_stable_node(dup)) | |
928 | return true; | |
929 | } | |
930 | BUG_ON(!hlist_empty(&stable_node->hlist)); | |
931 | free_stable_node_chain(stable_node, root); | |
932 | return false; | |
933 | } | |
934 | ||
cbf86cfe HD |
935 | static int remove_all_stable_nodes(void) |
936 | { | |
03640418 | 937 | struct stable_node *stable_node, *next; |
cbf86cfe HD |
938 | int nid; |
939 | int err = 0; | |
940 | ||
ef53d16c | 941 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
cbf86cfe HD |
942 | while (root_stable_tree[nid].rb_node) { |
943 | stable_node = rb_entry(root_stable_tree[nid].rb_node, | |
944 | struct stable_node, node); | |
2c653d0e AA |
945 | if (remove_stable_node_chain(stable_node, |
946 | root_stable_tree + nid)) { | |
cbf86cfe HD |
947 | err = -EBUSY; |
948 | break; /* proceed to next nid */ | |
949 | } | |
950 | cond_resched(); | |
951 | } | |
952 | } | |
03640418 | 953 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
954 | if (remove_stable_node(stable_node)) |
955 | err = -EBUSY; | |
956 | cond_resched(); | |
957 | } | |
cbf86cfe HD |
958 | return err; |
959 | } | |
960 | ||
d952b791 | 961 | static int unmerge_and_remove_all_rmap_items(void) |
31dbd01f IE |
962 | { |
963 | struct mm_slot *mm_slot; | |
964 | struct mm_struct *mm; | |
965 | struct vm_area_struct *vma; | |
d952b791 HD |
966 | int err = 0; |
967 | ||
968 | spin_lock(&ksm_mmlist_lock); | |
9ba69294 | 969 | ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next, |
d952b791 HD |
970 | struct mm_slot, mm_list); |
971 | spin_unlock(&ksm_mmlist_lock); | |
31dbd01f | 972 | |
9ba69294 HD |
973 | for (mm_slot = ksm_scan.mm_slot; |
974 | mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) { | |
31dbd01f | 975 | mm = mm_slot->mm; |
d8ed45c5 | 976 | mmap_read_lock(mm); |
31dbd01f | 977 | for (vma = mm->mmap; vma; vma = vma->vm_next) { |
9ba69294 HD |
978 | if (ksm_test_exit(mm)) |
979 | break; | |
31dbd01f IE |
980 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
981 | continue; | |
d952b791 HD |
982 | err = unmerge_ksm_pages(vma, |
983 | vma->vm_start, vma->vm_end); | |
9ba69294 HD |
984 | if (err) |
985 | goto error; | |
31dbd01f | 986 | } |
9ba69294 | 987 | |
420be4ed | 988 | remove_trailing_rmap_items(&mm_slot->rmap_list); |
d8ed45c5 | 989 | mmap_read_unlock(mm); |
d952b791 HD |
990 | |
991 | spin_lock(&ksm_mmlist_lock); | |
9ba69294 | 992 | ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next, |
d952b791 | 993 | struct mm_slot, mm_list); |
9ba69294 | 994 | if (ksm_test_exit(mm)) { |
4ca3a69b | 995 | hash_del(&mm_slot->link); |
9ba69294 HD |
996 | list_del(&mm_slot->mm_list); |
997 | spin_unlock(&ksm_mmlist_lock); | |
998 | ||
999 | free_mm_slot(mm_slot); | |
1000 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
9ba69294 | 1001 | mmdrop(mm); |
7496fea9 | 1002 | } else |
9ba69294 | 1003 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f IE |
1004 | } |
1005 | ||
cbf86cfe HD |
1006 | /* Clean up stable nodes, but don't worry if some are still busy */ |
1007 | remove_all_stable_nodes(); | |
d952b791 | 1008 | ksm_scan.seqnr = 0; |
9ba69294 HD |
1009 | return 0; |
1010 | ||
1011 | error: | |
d8ed45c5 | 1012 | mmap_read_unlock(mm); |
31dbd01f | 1013 | spin_lock(&ksm_mmlist_lock); |
d952b791 | 1014 | ksm_scan.mm_slot = &ksm_mm_head; |
31dbd01f | 1015 | spin_unlock(&ksm_mmlist_lock); |
d952b791 | 1016 | return err; |
31dbd01f | 1017 | } |
2ffd8679 | 1018 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 1019 | |
31dbd01f IE |
1020 | static u32 calc_checksum(struct page *page) |
1021 | { | |
1022 | u32 checksum; | |
9b04c5fe | 1023 | void *addr = kmap_atomic(page); |
59e1a2f4 | 1024 | checksum = xxhash(addr, PAGE_SIZE, 0); |
9b04c5fe | 1025 | kunmap_atomic(addr); |
31dbd01f IE |
1026 | return checksum; |
1027 | } | |
1028 | ||
31dbd01f IE |
1029 | static int write_protect_page(struct vm_area_struct *vma, struct page *page, |
1030 | pte_t *orig_pte) | |
1031 | { | |
1032 | struct mm_struct *mm = vma->vm_mm; | |
36eaff33 KS |
1033 | struct page_vma_mapped_walk pvmw = { |
1034 | .page = page, | |
1035 | .vma = vma, | |
1036 | }; | |
31dbd01f IE |
1037 | int swapped; |
1038 | int err = -EFAULT; | |
ac46d4f3 | 1039 | struct mmu_notifier_range range; |
31dbd01f | 1040 | |
36eaff33 KS |
1041 | pvmw.address = page_address_in_vma(page, vma); |
1042 | if (pvmw.address == -EFAULT) | |
31dbd01f IE |
1043 | goto out; |
1044 | ||
29ad768c | 1045 | BUG_ON(PageTransCompound(page)); |
6bdb913f | 1046 | |
7269f999 | 1047 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, |
6f4f13e8 | 1048 | pvmw.address, |
ac46d4f3 JG |
1049 | pvmw.address + PAGE_SIZE); |
1050 | mmu_notifier_invalidate_range_start(&range); | |
6bdb913f | 1051 | |
36eaff33 | 1052 | if (!page_vma_mapped_walk(&pvmw)) |
6bdb913f | 1053 | goto out_mn; |
36eaff33 KS |
1054 | if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?")) |
1055 | goto out_unlock; | |
31dbd01f | 1056 | |
595cd8f2 | 1057 | if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) || |
b3a81d08 MK |
1058 | (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) || |
1059 | mm_tlb_flush_pending(mm)) { | |
31dbd01f IE |
1060 | pte_t entry; |
1061 | ||
1062 | swapped = PageSwapCache(page); | |
36eaff33 | 1063 | flush_cache_page(vma, pvmw.address, page_to_pfn(page)); |
31dbd01f | 1064 | /* |
25985edc | 1065 | * Ok this is tricky, when get_user_pages_fast() run it doesn't |
31dbd01f | 1066 | * take any lock, therefore the check that we are going to make |
f0953a1b | 1067 | * with the pagecount against the mapcount is racy and |
31dbd01f IE |
1068 | * O_DIRECT can happen right after the check. |
1069 | * So we clear the pte and flush the tlb before the check | |
1070 | * this assure us that no O_DIRECT can happen after the check | |
1071 | * or in the middle of the check. | |
0f10851e JG |
1072 | * |
1073 | * No need to notify as we are downgrading page table to read | |
1074 | * only not changing it to point to a new page. | |
1075 | * | |
ad56b738 | 1076 | * See Documentation/vm/mmu_notifier.rst |
31dbd01f | 1077 | */ |
0f10851e | 1078 | entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte); |
31dbd01f IE |
1079 | /* |
1080 | * Check that no O_DIRECT or similar I/O is in progress on the | |
1081 | * page | |
1082 | */ | |
31e855ea | 1083 | if (page_mapcount(page) + 1 + swapped != page_count(page)) { |
36eaff33 | 1084 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f IE |
1085 | goto out_unlock; |
1086 | } | |
4e31635c HD |
1087 | if (pte_dirty(entry)) |
1088 | set_page_dirty(page); | |
595cd8f2 AK |
1089 | |
1090 | if (pte_protnone(entry)) | |
1091 | entry = pte_mkclean(pte_clear_savedwrite(entry)); | |
1092 | else | |
1093 | entry = pte_mkclean(pte_wrprotect(entry)); | |
36eaff33 | 1094 | set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f | 1095 | } |
36eaff33 | 1096 | *orig_pte = *pvmw.pte; |
31dbd01f IE |
1097 | err = 0; |
1098 | ||
1099 | out_unlock: | |
36eaff33 | 1100 | page_vma_mapped_walk_done(&pvmw); |
6bdb913f | 1101 | out_mn: |
ac46d4f3 | 1102 | mmu_notifier_invalidate_range_end(&range); |
31dbd01f IE |
1103 | out: |
1104 | return err; | |
1105 | } | |
1106 | ||
1107 | /** | |
1108 | * replace_page - replace page in vma by new ksm page | |
8dd3557a HD |
1109 | * @vma: vma that holds the pte pointing to page |
1110 | * @page: the page we are replacing by kpage | |
1111 | * @kpage: the ksm page we replace page by | |
31dbd01f IE |
1112 | * @orig_pte: the original value of the pte |
1113 | * | |
1114 | * Returns 0 on success, -EFAULT on failure. | |
1115 | */ | |
8dd3557a HD |
1116 | static int replace_page(struct vm_area_struct *vma, struct page *page, |
1117 | struct page *kpage, pte_t orig_pte) | |
31dbd01f IE |
1118 | { |
1119 | struct mm_struct *mm = vma->vm_mm; | |
31dbd01f IE |
1120 | pmd_t *pmd; |
1121 | pte_t *ptep; | |
e86c59b1 | 1122 | pte_t newpte; |
31dbd01f IE |
1123 | spinlock_t *ptl; |
1124 | unsigned long addr; | |
31dbd01f | 1125 | int err = -EFAULT; |
ac46d4f3 | 1126 | struct mmu_notifier_range range; |
31dbd01f | 1127 | |
8dd3557a | 1128 | addr = page_address_in_vma(page, vma); |
31dbd01f IE |
1129 | if (addr == -EFAULT) |
1130 | goto out; | |
1131 | ||
6219049a BL |
1132 | pmd = mm_find_pmd(mm, addr); |
1133 | if (!pmd) | |
31dbd01f | 1134 | goto out; |
31dbd01f | 1135 | |
7269f999 | 1136 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr, |
6f4f13e8 | 1137 | addr + PAGE_SIZE); |
ac46d4f3 | 1138 | mmu_notifier_invalidate_range_start(&range); |
6bdb913f | 1139 | |
31dbd01f IE |
1140 | ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); |
1141 | if (!pte_same(*ptep, orig_pte)) { | |
1142 | pte_unmap_unlock(ptep, ptl); | |
6bdb913f | 1143 | goto out_mn; |
31dbd01f IE |
1144 | } |
1145 | ||
e86c59b1 CI |
1146 | /* |
1147 | * No need to check ksm_use_zero_pages here: we can only have a | |
457aef94 | 1148 | * zero_page here if ksm_use_zero_pages was enabled already. |
e86c59b1 CI |
1149 | */ |
1150 | if (!is_zero_pfn(page_to_pfn(kpage))) { | |
1151 | get_page(kpage); | |
1152 | page_add_anon_rmap(kpage, vma, addr, false); | |
1153 | newpte = mk_pte(kpage, vma->vm_page_prot); | |
1154 | } else { | |
1155 | newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage), | |
1156 | vma->vm_page_prot)); | |
a38c015f CI |
1157 | /* |
1158 | * We're replacing an anonymous page with a zero page, which is | |
1159 | * not anonymous. We need to do proper accounting otherwise we | |
1160 | * will get wrong values in /proc, and a BUG message in dmesg | |
1161 | * when tearing down the mm. | |
1162 | */ | |
1163 | dec_mm_counter(mm, MM_ANONPAGES); | |
e86c59b1 | 1164 | } |
31dbd01f IE |
1165 | |
1166 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
0f10851e JG |
1167 | /* |
1168 | * No need to notify as we are replacing a read only page with another | |
1169 | * read only page with the same content. | |
1170 | * | |
ad56b738 | 1171 | * See Documentation/vm/mmu_notifier.rst |
0f10851e JG |
1172 | */ |
1173 | ptep_clear_flush(vma, addr, ptep); | |
e86c59b1 | 1174 | set_pte_at_notify(mm, addr, ptep, newpte); |
31dbd01f | 1175 | |
d281ee61 | 1176 | page_remove_rmap(page, false); |
ae52a2ad HD |
1177 | if (!page_mapped(page)) |
1178 | try_to_free_swap(page); | |
8dd3557a | 1179 | put_page(page); |
31dbd01f IE |
1180 | |
1181 | pte_unmap_unlock(ptep, ptl); | |
1182 | err = 0; | |
6bdb913f | 1183 | out_mn: |
ac46d4f3 | 1184 | mmu_notifier_invalidate_range_end(&range); |
31dbd01f IE |
1185 | out: |
1186 | return err; | |
1187 | } | |
1188 | ||
1189 | /* | |
1190 | * try_to_merge_one_page - take two pages and merge them into one | |
8dd3557a HD |
1191 | * @vma: the vma that holds the pte pointing to page |
1192 | * @page: the PageAnon page that we want to replace with kpage | |
80e14822 HD |
1193 | * @kpage: the PageKsm page that we want to map instead of page, |
1194 | * or NULL the first time when we want to use page as kpage. | |
31dbd01f IE |
1195 | * |
1196 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
1197 | */ | |
1198 | static int try_to_merge_one_page(struct vm_area_struct *vma, | |
8dd3557a | 1199 | struct page *page, struct page *kpage) |
31dbd01f IE |
1200 | { |
1201 | pte_t orig_pte = __pte(0); | |
1202 | int err = -EFAULT; | |
1203 | ||
db114b83 HD |
1204 | if (page == kpage) /* ksm page forked */ |
1205 | return 0; | |
1206 | ||
8dd3557a | 1207 | if (!PageAnon(page)) |
31dbd01f IE |
1208 | goto out; |
1209 | ||
31dbd01f IE |
1210 | /* |
1211 | * We need the page lock to read a stable PageSwapCache in | |
1212 | * write_protect_page(). We use trylock_page() instead of | |
1213 | * lock_page() because we don't want to wait here - we | |
1214 | * prefer to continue scanning and merging different pages, | |
1215 | * then come back to this page when it is unlocked. | |
1216 | */ | |
8dd3557a | 1217 | if (!trylock_page(page)) |
31e855ea | 1218 | goto out; |
f765f540 KS |
1219 | |
1220 | if (PageTransCompound(page)) { | |
a7306c34 | 1221 | if (split_huge_page(page)) |
f765f540 KS |
1222 | goto out_unlock; |
1223 | } | |
1224 | ||
31dbd01f IE |
1225 | /* |
1226 | * If this anonymous page is mapped only here, its pte may need | |
1227 | * to be write-protected. If it's mapped elsewhere, all of its | |
1228 | * ptes are necessarily already write-protected. But in either | |
1229 | * case, we need to lock and check page_count is not raised. | |
1230 | */ | |
80e14822 HD |
1231 | if (write_protect_page(vma, page, &orig_pte) == 0) { |
1232 | if (!kpage) { | |
1233 | /* | |
1234 | * While we hold page lock, upgrade page from | |
1235 | * PageAnon+anon_vma to PageKsm+NULL stable_node: | |
1236 | * stable_tree_insert() will update stable_node. | |
1237 | */ | |
1238 | set_page_stable_node(page, NULL); | |
1239 | mark_page_accessed(page); | |
337ed7eb MK |
1240 | /* |
1241 | * Page reclaim just frees a clean page with no dirty | |
1242 | * ptes: make sure that the ksm page would be swapped. | |
1243 | */ | |
1244 | if (!PageDirty(page)) | |
1245 | SetPageDirty(page); | |
80e14822 HD |
1246 | err = 0; |
1247 | } else if (pages_identical(page, kpage)) | |
1248 | err = replace_page(vma, page, kpage, orig_pte); | |
1249 | } | |
31dbd01f | 1250 | |
80e14822 | 1251 | if ((vma->vm_flags & VM_LOCKED) && kpage && !err) { |
73848b46 | 1252 | munlock_vma_page(page); |
5ad64688 HD |
1253 | if (!PageMlocked(kpage)) { |
1254 | unlock_page(page); | |
5ad64688 HD |
1255 | lock_page(kpage); |
1256 | mlock_vma_page(kpage); | |
1257 | page = kpage; /* for final unlock */ | |
1258 | } | |
1259 | } | |
73848b46 | 1260 | |
f765f540 | 1261 | out_unlock: |
8dd3557a | 1262 | unlock_page(page); |
31dbd01f IE |
1263 | out: |
1264 | return err; | |
1265 | } | |
1266 | ||
81464e30 HD |
1267 | /* |
1268 | * try_to_merge_with_ksm_page - like try_to_merge_two_pages, | |
1269 | * but no new kernel page is allocated: kpage must already be a ksm page. | |
8dd3557a HD |
1270 | * |
1271 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
81464e30 | 1272 | */ |
8dd3557a HD |
1273 | static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item, |
1274 | struct page *page, struct page *kpage) | |
81464e30 | 1275 | { |
8dd3557a | 1276 | struct mm_struct *mm = rmap_item->mm; |
81464e30 HD |
1277 | struct vm_area_struct *vma; |
1278 | int err = -EFAULT; | |
1279 | ||
d8ed45c5 | 1280 | mmap_read_lock(mm); |
85c6e8dd AA |
1281 | vma = find_mergeable_vma(mm, rmap_item->address); |
1282 | if (!vma) | |
81464e30 HD |
1283 | goto out; |
1284 | ||
8dd3557a | 1285 | err = try_to_merge_one_page(vma, page, kpage); |
db114b83 HD |
1286 | if (err) |
1287 | goto out; | |
1288 | ||
bc56620b HD |
1289 | /* Unstable nid is in union with stable anon_vma: remove first */ |
1290 | remove_rmap_item_from_tree(rmap_item); | |
1291 | ||
c1e8d7c6 | 1292 | /* Must get reference to anon_vma while still holding mmap_lock */ |
9e60109f PZ |
1293 | rmap_item->anon_vma = vma->anon_vma; |
1294 | get_anon_vma(vma->anon_vma); | |
81464e30 | 1295 | out: |
d8ed45c5 | 1296 | mmap_read_unlock(mm); |
81464e30 HD |
1297 | return err; |
1298 | } | |
1299 | ||
31dbd01f IE |
1300 | /* |
1301 | * try_to_merge_two_pages - take two identical pages and prepare them | |
1302 | * to be merged into one page. | |
1303 | * | |
8dd3557a HD |
1304 | * This function returns the kpage if we successfully merged two identical |
1305 | * pages into one ksm page, NULL otherwise. | |
31dbd01f | 1306 | * |
80e14822 | 1307 | * Note that this function upgrades page to ksm page: if one of the pages |
31dbd01f IE |
1308 | * is already a ksm page, try_to_merge_with_ksm_page should be used. |
1309 | */ | |
8dd3557a HD |
1310 | static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item, |
1311 | struct page *page, | |
1312 | struct rmap_item *tree_rmap_item, | |
1313 | struct page *tree_page) | |
31dbd01f | 1314 | { |
80e14822 | 1315 | int err; |
31dbd01f | 1316 | |
80e14822 | 1317 | err = try_to_merge_with_ksm_page(rmap_item, page, NULL); |
31dbd01f | 1318 | if (!err) { |
8dd3557a | 1319 | err = try_to_merge_with_ksm_page(tree_rmap_item, |
80e14822 | 1320 | tree_page, page); |
31dbd01f | 1321 | /* |
81464e30 HD |
1322 | * If that fails, we have a ksm page with only one pte |
1323 | * pointing to it: so break it. | |
31dbd01f | 1324 | */ |
4035c07a | 1325 | if (err) |
8dd3557a | 1326 | break_cow(rmap_item); |
31dbd01f | 1327 | } |
80e14822 | 1328 | return err ? NULL : page; |
31dbd01f IE |
1329 | } |
1330 | ||
2c653d0e AA |
1331 | static __always_inline |
1332 | bool __is_page_sharing_candidate(struct stable_node *stable_node, int offset) | |
1333 | { | |
1334 | VM_BUG_ON(stable_node->rmap_hlist_len < 0); | |
1335 | /* | |
1336 | * Check that at least one mapping still exists, otherwise | |
1337 | * there's no much point to merge and share with this | |
1338 | * stable_node, as the underlying tree_page of the other | |
1339 | * sharer is going to be freed soon. | |
1340 | */ | |
1341 | return stable_node->rmap_hlist_len && | |
1342 | stable_node->rmap_hlist_len + offset < ksm_max_page_sharing; | |
1343 | } | |
1344 | ||
1345 | static __always_inline | |
1346 | bool is_page_sharing_candidate(struct stable_node *stable_node) | |
1347 | { | |
1348 | return __is_page_sharing_candidate(stable_node, 0); | |
1349 | } | |
1350 | ||
c01f0b54 CIK |
1351 | static struct page *stable_node_dup(struct stable_node **_stable_node_dup, |
1352 | struct stable_node **_stable_node, | |
1353 | struct rb_root *root, | |
1354 | bool prune_stale_stable_nodes) | |
2c653d0e | 1355 | { |
b4fecc67 | 1356 | struct stable_node *dup, *found = NULL, *stable_node = *_stable_node; |
2c653d0e | 1357 | struct hlist_node *hlist_safe; |
8dc5ffcd | 1358 | struct page *_tree_page, *tree_page = NULL; |
2c653d0e AA |
1359 | int nr = 0; |
1360 | int found_rmap_hlist_len; | |
1361 | ||
1362 | if (!prune_stale_stable_nodes || | |
1363 | time_before(jiffies, stable_node->chain_prune_time + | |
1364 | msecs_to_jiffies( | |
1365 | ksm_stable_node_chains_prune_millisecs))) | |
1366 | prune_stale_stable_nodes = false; | |
1367 | else | |
1368 | stable_node->chain_prune_time = jiffies; | |
1369 | ||
1370 | hlist_for_each_entry_safe(dup, hlist_safe, | |
1371 | &stable_node->hlist, hlist_dup) { | |
1372 | cond_resched(); | |
1373 | /* | |
1374 | * We must walk all stable_node_dup to prune the stale | |
1375 | * stable nodes during lookup. | |
1376 | * | |
1377 | * get_ksm_page can drop the nodes from the | |
1378 | * stable_node->hlist if they point to freed pages | |
1379 | * (that's why we do a _safe walk). The "dup" | |
1380 | * stable_node parameter itself will be freed from | |
1381 | * under us if it returns NULL. | |
1382 | */ | |
2cee57d1 | 1383 | _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK); |
2c653d0e AA |
1384 | if (!_tree_page) |
1385 | continue; | |
1386 | nr += 1; | |
1387 | if (is_page_sharing_candidate(dup)) { | |
1388 | if (!found || | |
1389 | dup->rmap_hlist_len > found_rmap_hlist_len) { | |
1390 | if (found) | |
8dc5ffcd | 1391 | put_page(tree_page); |
2c653d0e AA |
1392 | found = dup; |
1393 | found_rmap_hlist_len = found->rmap_hlist_len; | |
8dc5ffcd | 1394 | tree_page = _tree_page; |
2c653d0e | 1395 | |
8dc5ffcd | 1396 | /* skip put_page for found dup */ |
2c653d0e AA |
1397 | if (!prune_stale_stable_nodes) |
1398 | break; | |
2c653d0e AA |
1399 | continue; |
1400 | } | |
1401 | } | |
1402 | put_page(_tree_page); | |
1403 | } | |
1404 | ||
80b18dfa AA |
1405 | if (found) { |
1406 | /* | |
1407 | * nr is counting all dups in the chain only if | |
1408 | * prune_stale_stable_nodes is true, otherwise we may | |
1409 | * break the loop at nr == 1 even if there are | |
1410 | * multiple entries. | |
1411 | */ | |
1412 | if (prune_stale_stable_nodes && nr == 1) { | |
2c653d0e AA |
1413 | /* |
1414 | * If there's not just one entry it would | |
1415 | * corrupt memory, better BUG_ON. In KSM | |
1416 | * context with no lock held it's not even | |
1417 | * fatal. | |
1418 | */ | |
1419 | BUG_ON(stable_node->hlist.first->next); | |
1420 | ||
1421 | /* | |
1422 | * There's just one entry and it is below the | |
1423 | * deduplication limit so drop the chain. | |
1424 | */ | |
1425 | rb_replace_node(&stable_node->node, &found->node, | |
1426 | root); | |
1427 | free_stable_node(stable_node); | |
1428 | ksm_stable_node_chains--; | |
1429 | ksm_stable_node_dups--; | |
b4fecc67 | 1430 | /* |
0ba1d0f7 AA |
1431 | * NOTE: the caller depends on the stable_node |
1432 | * to be equal to stable_node_dup if the chain | |
1433 | * was collapsed. | |
b4fecc67 | 1434 | */ |
0ba1d0f7 AA |
1435 | *_stable_node = found; |
1436 | /* | |
f0953a1b | 1437 | * Just for robustness, as stable_node is |
0ba1d0f7 AA |
1438 | * otherwise left as a stable pointer, the |
1439 | * compiler shall optimize it away at build | |
1440 | * time. | |
1441 | */ | |
1442 | stable_node = NULL; | |
80b18dfa AA |
1443 | } else if (stable_node->hlist.first != &found->hlist_dup && |
1444 | __is_page_sharing_candidate(found, 1)) { | |
2c653d0e | 1445 | /* |
80b18dfa AA |
1446 | * If the found stable_node dup can accept one |
1447 | * more future merge (in addition to the one | |
1448 | * that is underway) and is not at the head of | |
1449 | * the chain, put it there so next search will | |
1450 | * be quicker in the !prune_stale_stable_nodes | |
1451 | * case. | |
1452 | * | |
1453 | * NOTE: it would be inaccurate to use nr > 1 | |
1454 | * instead of checking the hlist.first pointer | |
1455 | * directly, because in the | |
1456 | * prune_stale_stable_nodes case "nr" isn't | |
1457 | * the position of the found dup in the chain, | |
1458 | * but the total number of dups in the chain. | |
2c653d0e AA |
1459 | */ |
1460 | hlist_del(&found->hlist_dup); | |
1461 | hlist_add_head(&found->hlist_dup, | |
1462 | &stable_node->hlist); | |
1463 | } | |
1464 | } | |
1465 | ||
8dc5ffcd AA |
1466 | *_stable_node_dup = found; |
1467 | return tree_page; | |
2c653d0e AA |
1468 | } |
1469 | ||
1470 | static struct stable_node *stable_node_dup_any(struct stable_node *stable_node, | |
1471 | struct rb_root *root) | |
1472 | { | |
1473 | if (!is_stable_node_chain(stable_node)) | |
1474 | return stable_node; | |
1475 | if (hlist_empty(&stable_node->hlist)) { | |
1476 | free_stable_node_chain(stable_node, root); | |
1477 | return NULL; | |
1478 | } | |
1479 | return hlist_entry(stable_node->hlist.first, | |
1480 | typeof(*stable_node), hlist_dup); | |
1481 | } | |
1482 | ||
8dc5ffcd AA |
1483 | /* |
1484 | * Like for get_ksm_page, this function can free the *_stable_node and | |
1485 | * *_stable_node_dup if the returned tree_page is NULL. | |
1486 | * | |
1487 | * It can also free and overwrite *_stable_node with the found | |
1488 | * stable_node_dup if the chain is collapsed (in which case | |
1489 | * *_stable_node will be equal to *_stable_node_dup like if the chain | |
1490 | * never existed). It's up to the caller to verify tree_page is not | |
1491 | * NULL before dereferencing *_stable_node or *_stable_node_dup. | |
1492 | * | |
1493 | * *_stable_node_dup is really a second output parameter of this | |
1494 | * function and will be overwritten in all cases, the caller doesn't | |
1495 | * need to initialize it. | |
1496 | */ | |
1497 | static struct page *__stable_node_chain(struct stable_node **_stable_node_dup, | |
1498 | struct stable_node **_stable_node, | |
1499 | struct rb_root *root, | |
1500 | bool prune_stale_stable_nodes) | |
2c653d0e | 1501 | { |
b4fecc67 | 1502 | struct stable_node *stable_node = *_stable_node; |
2c653d0e AA |
1503 | if (!is_stable_node_chain(stable_node)) { |
1504 | if (is_page_sharing_candidate(stable_node)) { | |
8dc5ffcd | 1505 | *_stable_node_dup = stable_node; |
2cee57d1 | 1506 | return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK); |
2c653d0e | 1507 | } |
8dc5ffcd AA |
1508 | /* |
1509 | * _stable_node_dup set to NULL means the stable_node | |
1510 | * reached the ksm_max_page_sharing limit. | |
1511 | */ | |
1512 | *_stable_node_dup = NULL; | |
2c653d0e AA |
1513 | return NULL; |
1514 | } | |
8dc5ffcd | 1515 | return stable_node_dup(_stable_node_dup, _stable_node, root, |
2c653d0e AA |
1516 | prune_stale_stable_nodes); |
1517 | } | |
1518 | ||
8dc5ffcd AA |
1519 | static __always_inline struct page *chain_prune(struct stable_node **s_n_d, |
1520 | struct stable_node **s_n, | |
1521 | struct rb_root *root) | |
2c653d0e | 1522 | { |
8dc5ffcd | 1523 | return __stable_node_chain(s_n_d, s_n, root, true); |
2c653d0e AA |
1524 | } |
1525 | ||
8dc5ffcd AA |
1526 | static __always_inline struct page *chain(struct stable_node **s_n_d, |
1527 | struct stable_node *s_n, | |
1528 | struct rb_root *root) | |
2c653d0e | 1529 | { |
8dc5ffcd AA |
1530 | struct stable_node *old_stable_node = s_n; |
1531 | struct page *tree_page; | |
1532 | ||
1533 | tree_page = __stable_node_chain(s_n_d, &s_n, root, false); | |
1534 | /* not pruning dups so s_n cannot have changed */ | |
1535 | VM_BUG_ON(s_n != old_stable_node); | |
1536 | return tree_page; | |
2c653d0e AA |
1537 | } |
1538 | ||
31dbd01f | 1539 | /* |
8dd3557a | 1540 | * stable_tree_search - search for page inside the stable tree |
31dbd01f IE |
1541 | * |
1542 | * This function checks if there is a page inside the stable tree | |
1543 | * with identical content to the page that we are scanning right now. | |
1544 | * | |
7b6ba2c7 | 1545 | * This function returns the stable tree node of identical content if found, |
31dbd01f IE |
1546 | * NULL otherwise. |
1547 | */ | |
62b61f61 | 1548 | static struct page *stable_tree_search(struct page *page) |
31dbd01f | 1549 | { |
90bd6fd3 | 1550 | int nid; |
ef53d16c | 1551 | struct rb_root *root; |
4146d2d6 HD |
1552 | struct rb_node **new; |
1553 | struct rb_node *parent; | |
2c653d0e | 1554 | struct stable_node *stable_node, *stable_node_dup, *stable_node_any; |
4146d2d6 | 1555 | struct stable_node *page_node; |
31dbd01f | 1556 | |
4146d2d6 HD |
1557 | page_node = page_stable_node(page); |
1558 | if (page_node && page_node->head != &migrate_nodes) { | |
1559 | /* ksm page forked */ | |
08beca44 | 1560 | get_page(page); |
62b61f61 | 1561 | return page; |
08beca44 HD |
1562 | } |
1563 | ||
90bd6fd3 | 1564 | nid = get_kpfn_nid(page_to_pfn(page)); |
ef53d16c | 1565 | root = root_stable_tree + nid; |
4146d2d6 | 1566 | again: |
ef53d16c | 1567 | new = &root->rb_node; |
4146d2d6 | 1568 | parent = NULL; |
90bd6fd3 | 1569 | |
4146d2d6 | 1570 | while (*new) { |
4035c07a | 1571 | struct page *tree_page; |
31dbd01f IE |
1572 | int ret; |
1573 | ||
08beca44 | 1574 | cond_resched(); |
4146d2d6 | 1575 | stable_node = rb_entry(*new, struct stable_node, node); |
2c653d0e | 1576 | stable_node_any = NULL; |
8dc5ffcd | 1577 | tree_page = chain_prune(&stable_node_dup, &stable_node, root); |
b4fecc67 AA |
1578 | /* |
1579 | * NOTE: stable_node may have been freed by | |
1580 | * chain_prune() if the returned stable_node_dup is | |
1581 | * not NULL. stable_node_dup may have been inserted in | |
1582 | * the rbtree instead as a regular stable_node (in | |
1583 | * order to collapse the stable_node chain if a single | |
0ba1d0f7 AA |
1584 | * stable_node dup was found in it). In such case the |
1585 | * stable_node is overwritten by the calleee to point | |
1586 | * to the stable_node_dup that was collapsed in the | |
1587 | * stable rbtree and stable_node will be equal to | |
1588 | * stable_node_dup like if the chain never existed. | |
b4fecc67 | 1589 | */ |
2c653d0e AA |
1590 | if (!stable_node_dup) { |
1591 | /* | |
1592 | * Either all stable_node dups were full in | |
1593 | * this stable_node chain, or this chain was | |
1594 | * empty and should be rb_erased. | |
1595 | */ | |
1596 | stable_node_any = stable_node_dup_any(stable_node, | |
1597 | root); | |
1598 | if (!stable_node_any) { | |
1599 | /* rb_erase just run */ | |
1600 | goto again; | |
1601 | } | |
1602 | /* | |
1603 | * Take any of the stable_node dups page of | |
1604 | * this stable_node chain to let the tree walk | |
1605 | * continue. All KSM pages belonging to the | |
1606 | * stable_node dups in a stable_node chain | |
1607 | * have the same content and they're | |
457aef94 | 1608 | * write protected at all times. Any will work |
2c653d0e AA |
1609 | * fine to continue the walk. |
1610 | */ | |
2cee57d1 YS |
1611 | tree_page = get_ksm_page(stable_node_any, |
1612 | GET_KSM_PAGE_NOLOCK); | |
2c653d0e AA |
1613 | } |
1614 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1615 | if (!tree_page) { |
1616 | /* | |
1617 | * If we walked over a stale stable_node, | |
1618 | * get_ksm_page() will call rb_erase() and it | |
1619 | * may rebalance the tree from under us. So | |
1620 | * restart the search from scratch. Returning | |
1621 | * NULL would be safe too, but we'd generate | |
1622 | * false negative insertions just because some | |
1623 | * stable_node was stale. | |
1624 | */ | |
1625 | goto again; | |
1626 | } | |
31dbd01f | 1627 | |
4035c07a | 1628 | ret = memcmp_pages(page, tree_page); |
c8d6553b | 1629 | put_page(tree_page); |
31dbd01f | 1630 | |
4146d2d6 | 1631 | parent = *new; |
c8d6553b | 1632 | if (ret < 0) |
4146d2d6 | 1633 | new = &parent->rb_left; |
c8d6553b | 1634 | else if (ret > 0) |
4146d2d6 | 1635 | new = &parent->rb_right; |
c8d6553b | 1636 | else { |
2c653d0e AA |
1637 | if (page_node) { |
1638 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1639 | /* | |
1640 | * Test if the migrated page should be merged | |
1641 | * into a stable node dup. If the mapcount is | |
1642 | * 1 we can migrate it with another KSM page | |
1643 | * without adding it to the chain. | |
1644 | */ | |
1645 | if (page_mapcount(page) > 1) | |
1646 | goto chain_append; | |
1647 | } | |
1648 | ||
1649 | if (!stable_node_dup) { | |
1650 | /* | |
1651 | * If the stable_node is a chain and | |
1652 | * we got a payload match in memcmp | |
1653 | * but we cannot merge the scanned | |
1654 | * page in any of the existing | |
1655 | * stable_node dups because they're | |
1656 | * all full, we need to wait the | |
1657 | * scanned page to find itself a match | |
1658 | * in the unstable tree to create a | |
1659 | * brand new KSM page to add later to | |
1660 | * the dups of this stable_node. | |
1661 | */ | |
1662 | return NULL; | |
1663 | } | |
1664 | ||
c8d6553b HD |
1665 | /* |
1666 | * Lock and unlock the stable_node's page (which | |
1667 | * might already have been migrated) so that page | |
1668 | * migration is sure to notice its raised count. | |
1669 | * It would be more elegant to return stable_node | |
1670 | * than kpage, but that involves more changes. | |
1671 | */ | |
2cee57d1 YS |
1672 | tree_page = get_ksm_page(stable_node_dup, |
1673 | GET_KSM_PAGE_TRYLOCK); | |
1674 | ||
1675 | if (PTR_ERR(tree_page) == -EBUSY) | |
1676 | return ERR_PTR(-EBUSY); | |
1677 | ||
2c653d0e AA |
1678 | if (unlikely(!tree_page)) |
1679 | /* | |
1680 | * The tree may have been rebalanced, | |
1681 | * so re-evaluate parent and new. | |
1682 | */ | |
4146d2d6 | 1683 | goto again; |
2c653d0e AA |
1684 | unlock_page(tree_page); |
1685 | ||
1686 | if (get_kpfn_nid(stable_node_dup->kpfn) != | |
1687 | NUMA(stable_node_dup->nid)) { | |
1688 | put_page(tree_page); | |
1689 | goto replace; | |
1690 | } | |
1691 | return tree_page; | |
c8d6553b | 1692 | } |
31dbd01f IE |
1693 | } |
1694 | ||
4146d2d6 HD |
1695 | if (!page_node) |
1696 | return NULL; | |
1697 | ||
1698 | list_del(&page_node->list); | |
1699 | DO_NUMA(page_node->nid = nid); | |
1700 | rb_link_node(&page_node->node, parent, new); | |
ef53d16c | 1701 | rb_insert_color(&page_node->node, root); |
2c653d0e AA |
1702 | out: |
1703 | if (is_page_sharing_candidate(page_node)) { | |
1704 | get_page(page); | |
1705 | return page; | |
1706 | } else | |
1707 | return NULL; | |
4146d2d6 HD |
1708 | |
1709 | replace: | |
b4fecc67 AA |
1710 | /* |
1711 | * If stable_node was a chain and chain_prune collapsed it, | |
0ba1d0f7 AA |
1712 | * stable_node has been updated to be the new regular |
1713 | * stable_node. A collapse of the chain is indistinguishable | |
1714 | * from the case there was no chain in the stable | |
1715 | * rbtree. Otherwise stable_node is the chain and | |
1716 | * stable_node_dup is the dup to replace. | |
b4fecc67 | 1717 | */ |
0ba1d0f7 | 1718 | if (stable_node_dup == stable_node) { |
b4fecc67 AA |
1719 | VM_BUG_ON(is_stable_node_chain(stable_node_dup)); |
1720 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); | |
2c653d0e AA |
1721 | /* there is no chain */ |
1722 | if (page_node) { | |
1723 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1724 | list_del(&page_node->list); | |
1725 | DO_NUMA(page_node->nid = nid); | |
b4fecc67 AA |
1726 | rb_replace_node(&stable_node_dup->node, |
1727 | &page_node->node, | |
2c653d0e AA |
1728 | root); |
1729 | if (is_page_sharing_candidate(page_node)) | |
1730 | get_page(page); | |
1731 | else | |
1732 | page = NULL; | |
1733 | } else { | |
b4fecc67 | 1734 | rb_erase(&stable_node_dup->node, root); |
2c653d0e AA |
1735 | page = NULL; |
1736 | } | |
4146d2d6 | 1737 | } else { |
2c653d0e AA |
1738 | VM_BUG_ON(!is_stable_node_chain(stable_node)); |
1739 | __stable_node_dup_del(stable_node_dup); | |
1740 | if (page_node) { | |
1741 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1742 | list_del(&page_node->list); | |
1743 | DO_NUMA(page_node->nid = nid); | |
1744 | stable_node_chain_add_dup(page_node, stable_node); | |
1745 | if (is_page_sharing_candidate(page_node)) | |
1746 | get_page(page); | |
1747 | else | |
1748 | page = NULL; | |
1749 | } else { | |
1750 | page = NULL; | |
1751 | } | |
4146d2d6 | 1752 | } |
2c653d0e AA |
1753 | stable_node_dup->head = &migrate_nodes; |
1754 | list_add(&stable_node_dup->list, stable_node_dup->head); | |
4146d2d6 | 1755 | return page; |
2c653d0e AA |
1756 | |
1757 | chain_append: | |
1758 | /* stable_node_dup could be null if it reached the limit */ | |
1759 | if (!stable_node_dup) | |
1760 | stable_node_dup = stable_node_any; | |
b4fecc67 AA |
1761 | /* |
1762 | * If stable_node was a chain and chain_prune collapsed it, | |
0ba1d0f7 AA |
1763 | * stable_node has been updated to be the new regular |
1764 | * stable_node. A collapse of the chain is indistinguishable | |
1765 | * from the case there was no chain in the stable | |
1766 | * rbtree. Otherwise stable_node is the chain and | |
1767 | * stable_node_dup is the dup to replace. | |
b4fecc67 | 1768 | */ |
0ba1d0f7 | 1769 | if (stable_node_dup == stable_node) { |
b4fecc67 | 1770 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); |
2c653d0e AA |
1771 | /* chain is missing so create it */ |
1772 | stable_node = alloc_stable_node_chain(stable_node_dup, | |
1773 | root); | |
1774 | if (!stable_node) | |
1775 | return NULL; | |
1776 | } | |
1777 | /* | |
1778 | * Add this stable_node dup that was | |
1779 | * migrated to the stable_node chain | |
1780 | * of the current nid for this page | |
1781 | * content. | |
1782 | */ | |
b4fecc67 | 1783 | VM_BUG_ON(!is_stable_node_dup(stable_node_dup)); |
2c653d0e AA |
1784 | VM_BUG_ON(page_node->head != &migrate_nodes); |
1785 | list_del(&page_node->list); | |
1786 | DO_NUMA(page_node->nid = nid); | |
1787 | stable_node_chain_add_dup(page_node, stable_node); | |
1788 | goto out; | |
31dbd01f IE |
1789 | } |
1790 | ||
1791 | /* | |
e850dcf5 | 1792 | * stable_tree_insert - insert stable tree node pointing to new ksm page |
31dbd01f IE |
1793 | * into the stable tree. |
1794 | * | |
7b6ba2c7 HD |
1795 | * This function returns the stable tree node just allocated on success, |
1796 | * NULL otherwise. | |
31dbd01f | 1797 | */ |
7b6ba2c7 | 1798 | static struct stable_node *stable_tree_insert(struct page *kpage) |
31dbd01f | 1799 | { |
90bd6fd3 PH |
1800 | int nid; |
1801 | unsigned long kpfn; | |
ef53d16c | 1802 | struct rb_root *root; |
90bd6fd3 | 1803 | struct rb_node **new; |
f2e5ff85 | 1804 | struct rb_node *parent; |
2c653d0e AA |
1805 | struct stable_node *stable_node, *stable_node_dup, *stable_node_any; |
1806 | bool need_chain = false; | |
31dbd01f | 1807 | |
90bd6fd3 PH |
1808 | kpfn = page_to_pfn(kpage); |
1809 | nid = get_kpfn_nid(kpfn); | |
ef53d16c | 1810 | root = root_stable_tree + nid; |
f2e5ff85 AA |
1811 | again: |
1812 | parent = NULL; | |
ef53d16c | 1813 | new = &root->rb_node; |
90bd6fd3 | 1814 | |
31dbd01f | 1815 | while (*new) { |
4035c07a | 1816 | struct page *tree_page; |
31dbd01f IE |
1817 | int ret; |
1818 | ||
08beca44 | 1819 | cond_resched(); |
7b6ba2c7 | 1820 | stable_node = rb_entry(*new, struct stable_node, node); |
2c653d0e | 1821 | stable_node_any = NULL; |
8dc5ffcd | 1822 | tree_page = chain(&stable_node_dup, stable_node, root); |
2c653d0e AA |
1823 | if (!stable_node_dup) { |
1824 | /* | |
1825 | * Either all stable_node dups were full in | |
1826 | * this stable_node chain, or this chain was | |
1827 | * empty and should be rb_erased. | |
1828 | */ | |
1829 | stable_node_any = stable_node_dup_any(stable_node, | |
1830 | root); | |
1831 | if (!stable_node_any) { | |
1832 | /* rb_erase just run */ | |
1833 | goto again; | |
1834 | } | |
1835 | /* | |
1836 | * Take any of the stable_node dups page of | |
1837 | * this stable_node chain to let the tree walk | |
1838 | * continue. All KSM pages belonging to the | |
1839 | * stable_node dups in a stable_node chain | |
1840 | * have the same content and they're | |
457aef94 | 1841 | * write protected at all times. Any will work |
2c653d0e AA |
1842 | * fine to continue the walk. |
1843 | */ | |
2cee57d1 YS |
1844 | tree_page = get_ksm_page(stable_node_any, |
1845 | GET_KSM_PAGE_NOLOCK); | |
2c653d0e AA |
1846 | } |
1847 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1848 | if (!tree_page) { |
1849 | /* | |
1850 | * If we walked over a stale stable_node, | |
1851 | * get_ksm_page() will call rb_erase() and it | |
1852 | * may rebalance the tree from under us. So | |
1853 | * restart the search from scratch. Returning | |
1854 | * NULL would be safe too, but we'd generate | |
1855 | * false negative insertions just because some | |
1856 | * stable_node was stale. | |
1857 | */ | |
1858 | goto again; | |
1859 | } | |
31dbd01f | 1860 | |
4035c07a HD |
1861 | ret = memcmp_pages(kpage, tree_page); |
1862 | put_page(tree_page); | |
31dbd01f IE |
1863 | |
1864 | parent = *new; | |
1865 | if (ret < 0) | |
1866 | new = &parent->rb_left; | |
1867 | else if (ret > 0) | |
1868 | new = &parent->rb_right; | |
1869 | else { | |
2c653d0e AA |
1870 | need_chain = true; |
1871 | break; | |
31dbd01f IE |
1872 | } |
1873 | } | |
1874 | ||
2c653d0e AA |
1875 | stable_node_dup = alloc_stable_node(); |
1876 | if (!stable_node_dup) | |
7b6ba2c7 | 1877 | return NULL; |
31dbd01f | 1878 | |
2c653d0e AA |
1879 | INIT_HLIST_HEAD(&stable_node_dup->hlist); |
1880 | stable_node_dup->kpfn = kpfn; | |
1881 | set_page_stable_node(kpage, stable_node_dup); | |
1882 | stable_node_dup->rmap_hlist_len = 0; | |
1883 | DO_NUMA(stable_node_dup->nid = nid); | |
1884 | if (!need_chain) { | |
1885 | rb_link_node(&stable_node_dup->node, parent, new); | |
1886 | rb_insert_color(&stable_node_dup->node, root); | |
1887 | } else { | |
1888 | if (!is_stable_node_chain(stable_node)) { | |
1889 | struct stable_node *orig = stable_node; | |
1890 | /* chain is missing so create it */ | |
1891 | stable_node = alloc_stable_node_chain(orig, root); | |
1892 | if (!stable_node) { | |
1893 | free_stable_node(stable_node_dup); | |
1894 | return NULL; | |
1895 | } | |
1896 | } | |
1897 | stable_node_chain_add_dup(stable_node_dup, stable_node); | |
1898 | } | |
08beca44 | 1899 | |
2c653d0e | 1900 | return stable_node_dup; |
31dbd01f IE |
1901 | } |
1902 | ||
1903 | /* | |
8dd3557a HD |
1904 | * unstable_tree_search_insert - search for identical page, |
1905 | * else insert rmap_item into the unstable tree. | |
31dbd01f IE |
1906 | * |
1907 | * This function searches for a page in the unstable tree identical to the | |
1908 | * page currently being scanned; and if no identical page is found in the | |
1909 | * tree, we insert rmap_item as a new object into the unstable tree. | |
1910 | * | |
1911 | * This function returns pointer to rmap_item found to be identical | |
1912 | * to the currently scanned page, NULL otherwise. | |
1913 | * | |
1914 | * This function does both searching and inserting, because they share | |
1915 | * the same walking algorithm in an rbtree. | |
1916 | */ | |
8dd3557a HD |
1917 | static |
1918 | struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item, | |
1919 | struct page *page, | |
1920 | struct page **tree_pagep) | |
31dbd01f | 1921 | { |
90bd6fd3 PH |
1922 | struct rb_node **new; |
1923 | struct rb_root *root; | |
31dbd01f | 1924 | struct rb_node *parent = NULL; |
90bd6fd3 PH |
1925 | int nid; |
1926 | ||
1927 | nid = get_kpfn_nid(page_to_pfn(page)); | |
ef53d16c | 1928 | root = root_unstable_tree + nid; |
90bd6fd3 | 1929 | new = &root->rb_node; |
31dbd01f IE |
1930 | |
1931 | while (*new) { | |
1932 | struct rmap_item *tree_rmap_item; | |
8dd3557a | 1933 | struct page *tree_page; |
31dbd01f IE |
1934 | int ret; |
1935 | ||
d178f27f | 1936 | cond_resched(); |
31dbd01f | 1937 | tree_rmap_item = rb_entry(*new, struct rmap_item, node); |
8dd3557a | 1938 | tree_page = get_mergeable_page(tree_rmap_item); |
c8f95ed1 | 1939 | if (!tree_page) |
31dbd01f IE |
1940 | return NULL; |
1941 | ||
1942 | /* | |
8dd3557a | 1943 | * Don't substitute a ksm page for a forked page. |
31dbd01f | 1944 | */ |
8dd3557a HD |
1945 | if (page == tree_page) { |
1946 | put_page(tree_page); | |
31dbd01f IE |
1947 | return NULL; |
1948 | } | |
1949 | ||
8dd3557a | 1950 | ret = memcmp_pages(page, tree_page); |
31dbd01f IE |
1951 | |
1952 | parent = *new; | |
1953 | if (ret < 0) { | |
8dd3557a | 1954 | put_page(tree_page); |
31dbd01f IE |
1955 | new = &parent->rb_left; |
1956 | } else if (ret > 0) { | |
8dd3557a | 1957 | put_page(tree_page); |
31dbd01f | 1958 | new = &parent->rb_right; |
b599cbdf HD |
1959 | } else if (!ksm_merge_across_nodes && |
1960 | page_to_nid(tree_page) != nid) { | |
1961 | /* | |
1962 | * If tree_page has been migrated to another NUMA node, | |
1963 | * it will be flushed out and put in the right unstable | |
1964 | * tree next time: only merge with it when across_nodes. | |
1965 | */ | |
1966 | put_page(tree_page); | |
1967 | return NULL; | |
31dbd01f | 1968 | } else { |
8dd3557a | 1969 | *tree_pagep = tree_page; |
31dbd01f IE |
1970 | return tree_rmap_item; |
1971 | } | |
1972 | } | |
1973 | ||
7b6ba2c7 | 1974 | rmap_item->address |= UNSTABLE_FLAG; |
31dbd01f | 1975 | rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); |
e850dcf5 | 1976 | DO_NUMA(rmap_item->nid = nid); |
31dbd01f | 1977 | rb_link_node(&rmap_item->node, parent, new); |
90bd6fd3 | 1978 | rb_insert_color(&rmap_item->node, root); |
31dbd01f | 1979 | |
473b0ce4 | 1980 | ksm_pages_unshared++; |
31dbd01f IE |
1981 | return NULL; |
1982 | } | |
1983 | ||
1984 | /* | |
1985 | * stable_tree_append - add another rmap_item to the linked list of | |
1986 | * rmap_items hanging off a given node of the stable tree, all sharing | |
1987 | * the same ksm page. | |
1988 | */ | |
1989 | static void stable_tree_append(struct rmap_item *rmap_item, | |
2c653d0e AA |
1990 | struct stable_node *stable_node, |
1991 | bool max_page_sharing_bypass) | |
31dbd01f | 1992 | { |
2c653d0e AA |
1993 | /* |
1994 | * rmap won't find this mapping if we don't insert the | |
1995 | * rmap_item in the right stable_node | |
1996 | * duplicate. page_migration could break later if rmap breaks, | |
1997 | * so we can as well crash here. We really need to check for | |
1998 | * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check | |
457aef94 | 1999 | * for other negative values as an underflow if detected here |
2c653d0e AA |
2000 | * for the first time (and not when decreasing rmap_hlist_len) |
2001 | * would be sign of memory corruption in the stable_node. | |
2002 | */ | |
2003 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
2004 | ||
2005 | stable_node->rmap_hlist_len++; | |
2006 | if (!max_page_sharing_bypass) | |
2007 | /* possibly non fatal but unexpected overflow, only warn */ | |
2008 | WARN_ON_ONCE(stable_node->rmap_hlist_len > | |
2009 | ksm_max_page_sharing); | |
2010 | ||
7b6ba2c7 | 2011 | rmap_item->head = stable_node; |
31dbd01f | 2012 | rmap_item->address |= STABLE_FLAG; |
7b6ba2c7 | 2013 | hlist_add_head(&rmap_item->hlist, &stable_node->hlist); |
e178dfde | 2014 | |
7b6ba2c7 HD |
2015 | if (rmap_item->hlist.next) |
2016 | ksm_pages_sharing++; | |
2017 | else | |
2018 | ksm_pages_shared++; | |
31dbd01f IE |
2019 | } |
2020 | ||
2021 | /* | |
81464e30 HD |
2022 | * cmp_and_merge_page - first see if page can be merged into the stable tree; |
2023 | * if not, compare checksum to previous and if it's the same, see if page can | |
2024 | * be inserted into the unstable tree, or merged with a page already there and | |
2025 | * both transferred to the stable tree. | |
31dbd01f IE |
2026 | * |
2027 | * @page: the page that we are searching identical page to. | |
2028 | * @rmap_item: the reverse mapping into the virtual address of this page | |
2029 | */ | |
2030 | static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item) | |
2031 | { | |
4b22927f | 2032 | struct mm_struct *mm = rmap_item->mm; |
31dbd01f | 2033 | struct rmap_item *tree_rmap_item; |
8dd3557a | 2034 | struct page *tree_page = NULL; |
7b6ba2c7 | 2035 | struct stable_node *stable_node; |
8dd3557a | 2036 | struct page *kpage; |
31dbd01f IE |
2037 | unsigned int checksum; |
2038 | int err; | |
2c653d0e | 2039 | bool max_page_sharing_bypass = false; |
31dbd01f | 2040 | |
4146d2d6 HD |
2041 | stable_node = page_stable_node(page); |
2042 | if (stable_node) { | |
2043 | if (stable_node->head != &migrate_nodes && | |
2c653d0e AA |
2044 | get_kpfn_nid(READ_ONCE(stable_node->kpfn)) != |
2045 | NUMA(stable_node->nid)) { | |
2046 | stable_node_dup_del(stable_node); | |
4146d2d6 HD |
2047 | stable_node->head = &migrate_nodes; |
2048 | list_add(&stable_node->list, stable_node->head); | |
2049 | } | |
2050 | if (stable_node->head != &migrate_nodes && | |
2051 | rmap_item->head == stable_node) | |
2052 | return; | |
2c653d0e AA |
2053 | /* |
2054 | * If it's a KSM fork, allow it to go over the sharing limit | |
2055 | * without warnings. | |
2056 | */ | |
2057 | if (!is_page_sharing_candidate(stable_node)) | |
2058 | max_page_sharing_bypass = true; | |
4146d2d6 | 2059 | } |
31dbd01f IE |
2060 | |
2061 | /* We first start with searching the page inside the stable tree */ | |
62b61f61 | 2062 | kpage = stable_tree_search(page); |
4146d2d6 HD |
2063 | if (kpage == page && rmap_item->head == stable_node) { |
2064 | put_page(kpage); | |
2065 | return; | |
2066 | } | |
2067 | ||
2068 | remove_rmap_item_from_tree(rmap_item); | |
2069 | ||
62b61f61 | 2070 | if (kpage) { |
2cee57d1 YS |
2071 | if (PTR_ERR(kpage) == -EBUSY) |
2072 | return; | |
2073 | ||
08beca44 | 2074 | err = try_to_merge_with_ksm_page(rmap_item, page, kpage); |
31dbd01f IE |
2075 | if (!err) { |
2076 | /* | |
2077 | * The page was successfully merged: | |
2078 | * add its rmap_item to the stable tree. | |
2079 | */ | |
5ad64688 | 2080 | lock_page(kpage); |
2c653d0e AA |
2081 | stable_tree_append(rmap_item, page_stable_node(kpage), |
2082 | max_page_sharing_bypass); | |
5ad64688 | 2083 | unlock_page(kpage); |
31dbd01f | 2084 | } |
8dd3557a | 2085 | put_page(kpage); |
31dbd01f IE |
2086 | return; |
2087 | } | |
2088 | ||
2089 | /* | |
4035c07a HD |
2090 | * If the hash value of the page has changed from the last time |
2091 | * we calculated it, this page is changing frequently: therefore we | |
2092 | * don't want to insert it in the unstable tree, and we don't want | |
2093 | * to waste our time searching for something identical to it there. | |
31dbd01f IE |
2094 | */ |
2095 | checksum = calc_checksum(page); | |
2096 | if (rmap_item->oldchecksum != checksum) { | |
2097 | rmap_item->oldchecksum = checksum; | |
2098 | return; | |
2099 | } | |
2100 | ||
e86c59b1 CI |
2101 | /* |
2102 | * Same checksum as an empty page. We attempt to merge it with the | |
2103 | * appropriate zero page if the user enabled this via sysfs. | |
2104 | */ | |
2105 | if (ksm_use_zero_pages && (checksum == zero_checksum)) { | |
2106 | struct vm_area_struct *vma; | |
2107 | ||
d8ed45c5 | 2108 | mmap_read_lock(mm); |
4b22927f | 2109 | vma = find_mergeable_vma(mm, rmap_item->address); |
56df70a6 MS |
2110 | if (vma) { |
2111 | err = try_to_merge_one_page(vma, page, | |
2112 | ZERO_PAGE(rmap_item->address)); | |
2113 | } else { | |
2114 | /* | |
2115 | * If the vma is out of date, we do not need to | |
2116 | * continue. | |
2117 | */ | |
2118 | err = 0; | |
2119 | } | |
d8ed45c5 | 2120 | mmap_read_unlock(mm); |
e86c59b1 CI |
2121 | /* |
2122 | * In case of failure, the page was not really empty, so we | |
2123 | * need to continue. Otherwise we're done. | |
2124 | */ | |
2125 | if (!err) | |
2126 | return; | |
2127 | } | |
8dd3557a HD |
2128 | tree_rmap_item = |
2129 | unstable_tree_search_insert(rmap_item, page, &tree_page); | |
31dbd01f | 2130 | if (tree_rmap_item) { |
77da2ba0 CI |
2131 | bool split; |
2132 | ||
8dd3557a HD |
2133 | kpage = try_to_merge_two_pages(rmap_item, page, |
2134 | tree_rmap_item, tree_page); | |
77da2ba0 CI |
2135 | /* |
2136 | * If both pages we tried to merge belong to the same compound | |
2137 | * page, then we actually ended up increasing the reference | |
2138 | * count of the same compound page twice, and split_huge_page | |
2139 | * failed. | |
2140 | * Here we set a flag if that happened, and we use it later to | |
2141 | * try split_huge_page again. Since we call put_page right | |
2142 | * afterwards, the reference count will be correct and | |
2143 | * split_huge_page should succeed. | |
2144 | */ | |
2145 | split = PageTransCompound(page) | |
2146 | && compound_head(page) == compound_head(tree_page); | |
8dd3557a | 2147 | put_page(tree_page); |
8dd3557a | 2148 | if (kpage) { |
bc56620b HD |
2149 | /* |
2150 | * The pages were successfully merged: insert new | |
2151 | * node in the stable tree and add both rmap_items. | |
2152 | */ | |
5ad64688 | 2153 | lock_page(kpage); |
7b6ba2c7 HD |
2154 | stable_node = stable_tree_insert(kpage); |
2155 | if (stable_node) { | |
2c653d0e AA |
2156 | stable_tree_append(tree_rmap_item, stable_node, |
2157 | false); | |
2158 | stable_tree_append(rmap_item, stable_node, | |
2159 | false); | |
7b6ba2c7 | 2160 | } |
5ad64688 | 2161 | unlock_page(kpage); |
7b6ba2c7 | 2162 | |
31dbd01f IE |
2163 | /* |
2164 | * If we fail to insert the page into the stable tree, | |
2165 | * we will have 2 virtual addresses that are pointing | |
2166 | * to a ksm page left outside the stable tree, | |
2167 | * in which case we need to break_cow on both. | |
2168 | */ | |
7b6ba2c7 | 2169 | if (!stable_node) { |
8dd3557a HD |
2170 | break_cow(tree_rmap_item); |
2171 | break_cow(rmap_item); | |
31dbd01f | 2172 | } |
77da2ba0 CI |
2173 | } else if (split) { |
2174 | /* | |
2175 | * We are here if we tried to merge two pages and | |
2176 | * failed because they both belonged to the same | |
2177 | * compound page. We will split the page now, but no | |
2178 | * merging will take place. | |
2179 | * We do not want to add the cost of a full lock; if | |
2180 | * the page is locked, it is better to skip it and | |
2181 | * perhaps try again later. | |
2182 | */ | |
2183 | if (!trylock_page(page)) | |
2184 | return; | |
2185 | split_huge_page(page); | |
2186 | unlock_page(page); | |
31dbd01f | 2187 | } |
31dbd01f IE |
2188 | } |
2189 | } | |
2190 | ||
2191 | static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot, | |
6514d511 | 2192 | struct rmap_item **rmap_list, |
31dbd01f IE |
2193 | unsigned long addr) |
2194 | { | |
2195 | struct rmap_item *rmap_item; | |
2196 | ||
6514d511 HD |
2197 | while (*rmap_list) { |
2198 | rmap_item = *rmap_list; | |
93d17715 | 2199 | if ((rmap_item->address & PAGE_MASK) == addr) |
31dbd01f | 2200 | return rmap_item; |
31dbd01f IE |
2201 | if (rmap_item->address > addr) |
2202 | break; | |
6514d511 | 2203 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 2204 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
2205 | free_rmap_item(rmap_item); |
2206 | } | |
2207 | ||
2208 | rmap_item = alloc_rmap_item(); | |
2209 | if (rmap_item) { | |
2210 | /* It has already been zeroed */ | |
2211 | rmap_item->mm = mm_slot->mm; | |
2212 | rmap_item->address = addr; | |
6514d511 HD |
2213 | rmap_item->rmap_list = *rmap_list; |
2214 | *rmap_list = rmap_item; | |
31dbd01f IE |
2215 | } |
2216 | return rmap_item; | |
2217 | } | |
2218 | ||
2219 | static struct rmap_item *scan_get_next_rmap_item(struct page **page) | |
2220 | { | |
2221 | struct mm_struct *mm; | |
2222 | struct mm_slot *slot; | |
2223 | struct vm_area_struct *vma; | |
2224 | struct rmap_item *rmap_item; | |
90bd6fd3 | 2225 | int nid; |
31dbd01f IE |
2226 | |
2227 | if (list_empty(&ksm_mm_head.mm_list)) | |
2228 | return NULL; | |
2229 | ||
2230 | slot = ksm_scan.mm_slot; | |
2231 | if (slot == &ksm_mm_head) { | |
2919bfd0 HD |
2232 | /* |
2233 | * A number of pages can hang around indefinitely on per-cpu | |
2234 | * pagevecs, raised page count preventing write_protect_page | |
2235 | * from merging them. Though it doesn't really matter much, | |
2236 | * it is puzzling to see some stuck in pages_volatile until | |
2237 | * other activity jostles them out, and they also prevented | |
2238 | * LTP's KSM test from succeeding deterministically; so drain | |
2239 | * them here (here rather than on entry to ksm_do_scan(), | |
2240 | * so we don't IPI too often when pages_to_scan is set low). | |
2241 | */ | |
2242 | lru_add_drain_all(); | |
2243 | ||
4146d2d6 HD |
2244 | /* |
2245 | * Whereas stale stable_nodes on the stable_tree itself | |
2246 | * get pruned in the regular course of stable_tree_search(), | |
2247 | * those moved out to the migrate_nodes list can accumulate: | |
2248 | * so prune them once before each full scan. | |
2249 | */ | |
2250 | if (!ksm_merge_across_nodes) { | |
03640418 | 2251 | struct stable_node *stable_node, *next; |
4146d2d6 HD |
2252 | struct page *page; |
2253 | ||
03640418 GT |
2254 | list_for_each_entry_safe(stable_node, next, |
2255 | &migrate_nodes, list) { | |
2cee57d1 YS |
2256 | page = get_ksm_page(stable_node, |
2257 | GET_KSM_PAGE_NOLOCK); | |
4146d2d6 HD |
2258 | if (page) |
2259 | put_page(page); | |
2260 | cond_resched(); | |
2261 | } | |
2262 | } | |
2263 | ||
ef53d16c | 2264 | for (nid = 0; nid < ksm_nr_node_ids; nid++) |
90bd6fd3 | 2265 | root_unstable_tree[nid] = RB_ROOT; |
31dbd01f IE |
2266 | |
2267 | spin_lock(&ksm_mmlist_lock); | |
2268 | slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list); | |
2269 | ksm_scan.mm_slot = slot; | |
2270 | spin_unlock(&ksm_mmlist_lock); | |
2b472611 HD |
2271 | /* |
2272 | * Although we tested list_empty() above, a racing __ksm_exit | |
2273 | * of the last mm on the list may have removed it since then. | |
2274 | */ | |
2275 | if (slot == &ksm_mm_head) | |
2276 | return NULL; | |
31dbd01f IE |
2277 | next_mm: |
2278 | ksm_scan.address = 0; | |
6514d511 | 2279 | ksm_scan.rmap_list = &slot->rmap_list; |
31dbd01f IE |
2280 | } |
2281 | ||
2282 | mm = slot->mm; | |
d8ed45c5 | 2283 | mmap_read_lock(mm); |
9ba69294 HD |
2284 | if (ksm_test_exit(mm)) |
2285 | vma = NULL; | |
2286 | else | |
2287 | vma = find_vma(mm, ksm_scan.address); | |
2288 | ||
2289 | for (; vma; vma = vma->vm_next) { | |
31dbd01f IE |
2290 | if (!(vma->vm_flags & VM_MERGEABLE)) |
2291 | continue; | |
2292 | if (ksm_scan.address < vma->vm_start) | |
2293 | ksm_scan.address = vma->vm_start; | |
2294 | if (!vma->anon_vma) | |
2295 | ksm_scan.address = vma->vm_end; | |
2296 | ||
2297 | while (ksm_scan.address < vma->vm_end) { | |
9ba69294 HD |
2298 | if (ksm_test_exit(mm)) |
2299 | break; | |
31dbd01f | 2300 | *page = follow_page(vma, ksm_scan.address, FOLL_GET); |
21ae5b01 AA |
2301 | if (IS_ERR_OR_NULL(*page)) { |
2302 | ksm_scan.address += PAGE_SIZE; | |
2303 | cond_resched(); | |
2304 | continue; | |
2305 | } | |
f765f540 | 2306 | if (PageAnon(*page)) { |
31dbd01f IE |
2307 | flush_anon_page(vma, *page, ksm_scan.address); |
2308 | flush_dcache_page(*page); | |
2309 | rmap_item = get_next_rmap_item(slot, | |
6514d511 | 2310 | ksm_scan.rmap_list, ksm_scan.address); |
31dbd01f | 2311 | if (rmap_item) { |
6514d511 HD |
2312 | ksm_scan.rmap_list = |
2313 | &rmap_item->rmap_list; | |
31dbd01f IE |
2314 | ksm_scan.address += PAGE_SIZE; |
2315 | } else | |
2316 | put_page(*page); | |
d8ed45c5 | 2317 | mmap_read_unlock(mm); |
31dbd01f IE |
2318 | return rmap_item; |
2319 | } | |
21ae5b01 | 2320 | put_page(*page); |
31dbd01f IE |
2321 | ksm_scan.address += PAGE_SIZE; |
2322 | cond_resched(); | |
2323 | } | |
2324 | } | |
2325 | ||
9ba69294 HD |
2326 | if (ksm_test_exit(mm)) { |
2327 | ksm_scan.address = 0; | |
6514d511 | 2328 | ksm_scan.rmap_list = &slot->rmap_list; |
9ba69294 | 2329 | } |
31dbd01f IE |
2330 | /* |
2331 | * Nuke all the rmap_items that are above this current rmap: | |
2332 | * because there were no VM_MERGEABLE vmas with such addresses. | |
2333 | */ | |
420be4ed | 2334 | remove_trailing_rmap_items(ksm_scan.rmap_list); |
31dbd01f IE |
2335 | |
2336 | spin_lock(&ksm_mmlist_lock); | |
cd551f97 HD |
2337 | ksm_scan.mm_slot = list_entry(slot->mm_list.next, |
2338 | struct mm_slot, mm_list); | |
2339 | if (ksm_scan.address == 0) { | |
2340 | /* | |
c1e8d7c6 | 2341 | * We've completed a full scan of all vmas, holding mmap_lock |
cd551f97 HD |
2342 | * throughout, and found no VM_MERGEABLE: so do the same as |
2343 | * __ksm_exit does to remove this mm from all our lists now. | |
9ba69294 HD |
2344 | * This applies either when cleaning up after __ksm_exit |
2345 | * (but beware: we can reach here even before __ksm_exit), | |
2346 | * or when all VM_MERGEABLE areas have been unmapped (and | |
c1e8d7c6 | 2347 | * mmap_lock then protects against race with MADV_MERGEABLE). |
cd551f97 | 2348 | */ |
4ca3a69b | 2349 | hash_del(&slot->link); |
cd551f97 | 2350 | list_del(&slot->mm_list); |
9ba69294 HD |
2351 | spin_unlock(&ksm_mmlist_lock); |
2352 | ||
cd551f97 HD |
2353 | free_mm_slot(slot); |
2354 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
d8ed45c5 | 2355 | mmap_read_unlock(mm); |
9ba69294 HD |
2356 | mmdrop(mm); |
2357 | } else { | |
d8ed45c5 | 2358 | mmap_read_unlock(mm); |
7496fea9 | 2359 | /* |
3e4e28c5 | 2360 | * mmap_read_unlock(mm) first because after |
7496fea9 ZC |
2361 | * spin_unlock(&ksm_mmlist_lock) run, the "mm" may |
2362 | * already have been freed under us by __ksm_exit() | |
2363 | * because the "mm_slot" is still hashed and | |
2364 | * ksm_scan.mm_slot doesn't point to it anymore. | |
2365 | */ | |
2366 | spin_unlock(&ksm_mmlist_lock); | |
cd551f97 | 2367 | } |
31dbd01f IE |
2368 | |
2369 | /* Repeat until we've completed scanning the whole list */ | |
cd551f97 | 2370 | slot = ksm_scan.mm_slot; |
31dbd01f IE |
2371 | if (slot != &ksm_mm_head) |
2372 | goto next_mm; | |
2373 | ||
31dbd01f IE |
2374 | ksm_scan.seqnr++; |
2375 | return NULL; | |
2376 | } | |
2377 | ||
2378 | /** | |
2379 | * ksm_do_scan - the ksm scanner main worker function. | |
b7701a5f | 2380 | * @scan_npages: number of pages we want to scan before we return. |
31dbd01f IE |
2381 | */ |
2382 | static void ksm_do_scan(unsigned int scan_npages) | |
2383 | { | |
2384 | struct rmap_item *rmap_item; | |
3f649ab7 | 2385 | struct page *page; |
31dbd01f | 2386 | |
878aee7d | 2387 | while (scan_npages-- && likely(!freezing(current))) { |
31dbd01f IE |
2388 | cond_resched(); |
2389 | rmap_item = scan_get_next_rmap_item(&page); | |
2390 | if (!rmap_item) | |
2391 | return; | |
4146d2d6 | 2392 | cmp_and_merge_page(page, rmap_item); |
31dbd01f IE |
2393 | put_page(page); |
2394 | } | |
2395 | } | |
2396 | ||
6e158384 HD |
2397 | static int ksmd_should_run(void) |
2398 | { | |
2399 | return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list); | |
2400 | } | |
2401 | ||
31dbd01f IE |
2402 | static int ksm_scan_thread(void *nothing) |
2403 | { | |
fcf9a0ef KT |
2404 | unsigned int sleep_ms; |
2405 | ||
878aee7d | 2406 | set_freezable(); |
339aa624 | 2407 | set_user_nice(current, 5); |
31dbd01f IE |
2408 | |
2409 | while (!kthread_should_stop()) { | |
6e158384 | 2410 | mutex_lock(&ksm_thread_mutex); |
ef4d43a8 | 2411 | wait_while_offlining(); |
6e158384 | 2412 | if (ksmd_should_run()) |
31dbd01f | 2413 | ksm_do_scan(ksm_thread_pages_to_scan); |
6e158384 HD |
2414 | mutex_unlock(&ksm_thread_mutex); |
2415 | ||
878aee7d AA |
2416 | try_to_freeze(); |
2417 | ||
6e158384 | 2418 | if (ksmd_should_run()) { |
fcf9a0ef KT |
2419 | sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs); |
2420 | wait_event_interruptible_timeout(ksm_iter_wait, | |
2421 | sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs), | |
2422 | msecs_to_jiffies(sleep_ms)); | |
31dbd01f | 2423 | } else { |
878aee7d | 2424 | wait_event_freezable(ksm_thread_wait, |
6e158384 | 2425 | ksmd_should_run() || kthread_should_stop()); |
31dbd01f IE |
2426 | } |
2427 | } | |
2428 | return 0; | |
2429 | } | |
2430 | ||
f8af4da3 HD |
2431 | int ksm_madvise(struct vm_area_struct *vma, unsigned long start, |
2432 | unsigned long end, int advice, unsigned long *vm_flags) | |
2433 | { | |
2434 | struct mm_struct *mm = vma->vm_mm; | |
d952b791 | 2435 | int err; |
f8af4da3 HD |
2436 | |
2437 | switch (advice) { | |
2438 | case MADV_MERGEABLE: | |
2439 | /* | |
2440 | * Be somewhat over-protective for now! | |
2441 | */ | |
2442 | if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE | | |
2443 | VM_PFNMAP | VM_IO | VM_DONTEXPAND | | |
0661a336 | 2444 | VM_HUGETLB | VM_MIXEDMAP)) |
f8af4da3 HD |
2445 | return 0; /* just ignore the advice */ |
2446 | ||
e1fb4a08 DJ |
2447 | if (vma_is_dax(vma)) |
2448 | return 0; | |
2449 | ||
12564485 SA |
2450 | #ifdef VM_SAO |
2451 | if (*vm_flags & VM_SAO) | |
2452 | return 0; | |
2453 | #endif | |
74a04967 KA |
2454 | #ifdef VM_SPARC_ADI |
2455 | if (*vm_flags & VM_SPARC_ADI) | |
2456 | return 0; | |
2457 | #endif | |
cc2383ec | 2458 | |
d952b791 HD |
2459 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { |
2460 | err = __ksm_enter(mm); | |
2461 | if (err) | |
2462 | return err; | |
2463 | } | |
f8af4da3 HD |
2464 | |
2465 | *vm_flags |= VM_MERGEABLE; | |
2466 | break; | |
2467 | ||
2468 | case MADV_UNMERGEABLE: | |
2469 | if (!(*vm_flags & VM_MERGEABLE)) | |
2470 | return 0; /* just ignore the advice */ | |
2471 | ||
d952b791 HD |
2472 | if (vma->anon_vma) { |
2473 | err = unmerge_ksm_pages(vma, start, end); | |
2474 | if (err) | |
2475 | return err; | |
2476 | } | |
f8af4da3 HD |
2477 | |
2478 | *vm_flags &= ~VM_MERGEABLE; | |
2479 | break; | |
2480 | } | |
2481 | ||
2482 | return 0; | |
2483 | } | |
33cf1707 | 2484 | EXPORT_SYMBOL_GPL(ksm_madvise); |
f8af4da3 HD |
2485 | |
2486 | int __ksm_enter(struct mm_struct *mm) | |
2487 | { | |
6e158384 HD |
2488 | struct mm_slot *mm_slot; |
2489 | int needs_wakeup; | |
2490 | ||
2491 | mm_slot = alloc_mm_slot(); | |
31dbd01f IE |
2492 | if (!mm_slot) |
2493 | return -ENOMEM; | |
2494 | ||
6e158384 HD |
2495 | /* Check ksm_run too? Would need tighter locking */ |
2496 | needs_wakeup = list_empty(&ksm_mm_head.mm_list); | |
2497 | ||
31dbd01f IE |
2498 | spin_lock(&ksm_mmlist_lock); |
2499 | insert_to_mm_slots_hash(mm, mm_slot); | |
2500 | /* | |
cbf86cfe HD |
2501 | * When KSM_RUN_MERGE (or KSM_RUN_STOP), |
2502 | * insert just behind the scanning cursor, to let the area settle | |
31dbd01f IE |
2503 | * down a little; when fork is followed by immediate exec, we don't |
2504 | * want ksmd to waste time setting up and tearing down an rmap_list. | |
cbf86cfe HD |
2505 | * |
2506 | * But when KSM_RUN_UNMERGE, it's important to insert ahead of its | |
2507 | * scanning cursor, otherwise KSM pages in newly forked mms will be | |
2508 | * missed: then we might as well insert at the end of the list. | |
31dbd01f | 2509 | */ |
cbf86cfe HD |
2510 | if (ksm_run & KSM_RUN_UNMERGE) |
2511 | list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list); | |
2512 | else | |
2513 | list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list); | |
31dbd01f IE |
2514 | spin_unlock(&ksm_mmlist_lock); |
2515 | ||
f8af4da3 | 2516 | set_bit(MMF_VM_MERGEABLE, &mm->flags); |
f1f10076 | 2517 | mmgrab(mm); |
6e158384 HD |
2518 | |
2519 | if (needs_wakeup) | |
2520 | wake_up_interruptible(&ksm_thread_wait); | |
2521 | ||
f8af4da3 HD |
2522 | return 0; |
2523 | } | |
2524 | ||
1c2fb7a4 | 2525 | void __ksm_exit(struct mm_struct *mm) |
f8af4da3 | 2526 | { |
cd551f97 | 2527 | struct mm_slot *mm_slot; |
9ba69294 | 2528 | int easy_to_free = 0; |
cd551f97 | 2529 | |
31dbd01f | 2530 | /* |
9ba69294 HD |
2531 | * This process is exiting: if it's straightforward (as is the |
2532 | * case when ksmd was never running), free mm_slot immediately. | |
2533 | * But if it's at the cursor or has rmap_items linked to it, use | |
c1e8d7c6 | 2534 | * mmap_lock to synchronize with any break_cows before pagetables |
9ba69294 HD |
2535 | * are freed, and leave the mm_slot on the list for ksmd to free. |
2536 | * Beware: ksm may already have noticed it exiting and freed the slot. | |
31dbd01f | 2537 | */ |
9ba69294 | 2538 | |
cd551f97 HD |
2539 | spin_lock(&ksm_mmlist_lock); |
2540 | mm_slot = get_mm_slot(mm); | |
9ba69294 | 2541 | if (mm_slot && ksm_scan.mm_slot != mm_slot) { |
6514d511 | 2542 | if (!mm_slot->rmap_list) { |
4ca3a69b | 2543 | hash_del(&mm_slot->link); |
9ba69294 HD |
2544 | list_del(&mm_slot->mm_list); |
2545 | easy_to_free = 1; | |
2546 | } else { | |
2547 | list_move(&mm_slot->mm_list, | |
2548 | &ksm_scan.mm_slot->mm_list); | |
2549 | } | |
cd551f97 | 2550 | } |
cd551f97 HD |
2551 | spin_unlock(&ksm_mmlist_lock); |
2552 | ||
9ba69294 HD |
2553 | if (easy_to_free) { |
2554 | free_mm_slot(mm_slot); | |
2555 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
2556 | mmdrop(mm); | |
2557 | } else if (mm_slot) { | |
d8ed45c5 ML |
2558 | mmap_write_lock(mm); |
2559 | mmap_write_unlock(mm); | |
9ba69294 | 2560 | } |
31dbd01f IE |
2561 | } |
2562 | ||
cbf86cfe | 2563 | struct page *ksm_might_need_to_copy(struct page *page, |
5ad64688 HD |
2564 | struct vm_area_struct *vma, unsigned long address) |
2565 | { | |
cbf86cfe | 2566 | struct anon_vma *anon_vma = page_anon_vma(page); |
5ad64688 HD |
2567 | struct page *new_page; |
2568 | ||
cbf86cfe HD |
2569 | if (PageKsm(page)) { |
2570 | if (page_stable_node(page) && | |
2571 | !(ksm_run & KSM_RUN_UNMERGE)) | |
2572 | return page; /* no need to copy it */ | |
2573 | } else if (!anon_vma) { | |
2574 | return page; /* no need to copy it */ | |
2575 | } else if (anon_vma->root == vma->anon_vma->root && | |
2576 | page->index == linear_page_index(vma, address)) { | |
2577 | return page; /* still no need to copy it */ | |
2578 | } | |
2579 | if (!PageUptodate(page)) | |
2580 | return page; /* let do_swap_page report the error */ | |
2581 | ||
5ad64688 | 2582 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
62fdb163 HD |
2583 | if (new_page && mem_cgroup_charge(new_page, vma->vm_mm, GFP_KERNEL)) { |
2584 | put_page(new_page); | |
2585 | new_page = NULL; | |
2586 | } | |
5ad64688 HD |
2587 | if (new_page) { |
2588 | copy_user_highpage(new_page, page, address, vma); | |
2589 | ||
2590 | SetPageDirty(new_page); | |
2591 | __SetPageUptodate(new_page); | |
48c935ad | 2592 | __SetPageLocked(new_page); |
5ad64688 HD |
2593 | } |
2594 | ||
5ad64688 HD |
2595 | return new_page; |
2596 | } | |
2597 | ||
1df631ae | 2598 | void rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc) |
e9995ef9 HD |
2599 | { |
2600 | struct stable_node *stable_node; | |
e9995ef9 | 2601 | struct rmap_item *rmap_item; |
e9995ef9 HD |
2602 | int search_new_forks = 0; |
2603 | ||
309381fe | 2604 | VM_BUG_ON_PAGE(!PageKsm(page), page); |
9f32624b JK |
2605 | |
2606 | /* | |
2607 | * Rely on the page lock to protect against concurrent modifications | |
2608 | * to that page's node of the stable tree. | |
2609 | */ | |
309381fe | 2610 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
e9995ef9 HD |
2611 | |
2612 | stable_node = page_stable_node(page); | |
2613 | if (!stable_node) | |
1df631ae | 2614 | return; |
e9995ef9 | 2615 | again: |
b67bfe0d | 2616 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
e9995ef9 | 2617 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
5beb4930 | 2618 | struct anon_vma_chain *vmac; |
e9995ef9 HD |
2619 | struct vm_area_struct *vma; |
2620 | ||
ad12695f | 2621 | cond_resched(); |
b6b19f25 | 2622 | anon_vma_lock_read(anon_vma); |
bf181b9f ML |
2623 | anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, |
2624 | 0, ULONG_MAX) { | |
1105a2fc JH |
2625 | unsigned long addr; |
2626 | ||
ad12695f | 2627 | cond_resched(); |
5beb4930 | 2628 | vma = vmac->vma; |
1105a2fc JH |
2629 | |
2630 | /* Ignore the stable/unstable/sqnr flags */ | |
cd7fae26 | 2631 | addr = rmap_item->address & PAGE_MASK; |
1105a2fc JH |
2632 | |
2633 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
e9995ef9 HD |
2634 | continue; |
2635 | /* | |
2636 | * Initially we examine only the vma which covers this | |
2637 | * rmap_item; but later, if there is still work to do, | |
2638 | * we examine covering vmas in other mms: in case they | |
2639 | * were forked from the original since ksmd passed. | |
2640 | */ | |
2641 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
2642 | continue; | |
2643 | ||
0dd1c7bb JK |
2644 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
2645 | continue; | |
2646 | ||
1105a2fc | 2647 | if (!rwc->rmap_one(page, vma, addr, rwc->arg)) { |
b6b19f25 | 2648 | anon_vma_unlock_read(anon_vma); |
1df631ae | 2649 | return; |
e9995ef9 | 2650 | } |
0dd1c7bb JK |
2651 | if (rwc->done && rwc->done(page)) { |
2652 | anon_vma_unlock_read(anon_vma); | |
1df631ae | 2653 | return; |
0dd1c7bb | 2654 | } |
e9995ef9 | 2655 | } |
b6b19f25 | 2656 | anon_vma_unlock_read(anon_vma); |
e9995ef9 HD |
2657 | } |
2658 | if (!search_new_forks++) | |
2659 | goto again; | |
e9995ef9 HD |
2660 | } |
2661 | ||
52629506 | 2662 | #ifdef CONFIG_MIGRATION |
e9995ef9 HD |
2663 | void ksm_migrate_page(struct page *newpage, struct page *oldpage) |
2664 | { | |
2665 | struct stable_node *stable_node; | |
2666 | ||
309381fe SL |
2667 | VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage); |
2668 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
2669 | VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage); | |
e9995ef9 HD |
2670 | |
2671 | stable_node = page_stable_node(newpage); | |
2672 | if (stable_node) { | |
309381fe | 2673 | VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage); |
62b61f61 | 2674 | stable_node->kpfn = page_to_pfn(newpage); |
c8d6553b HD |
2675 | /* |
2676 | * newpage->mapping was set in advance; now we need smp_wmb() | |
2677 | * to make sure that the new stable_node->kpfn is visible | |
2678 | * to get_ksm_page() before it can see that oldpage->mapping | |
2679 | * has gone stale (or that PageSwapCache has been cleared). | |
2680 | */ | |
2681 | smp_wmb(); | |
2682 | set_page_stable_node(oldpage, NULL); | |
e9995ef9 HD |
2683 | } |
2684 | } | |
2685 | #endif /* CONFIG_MIGRATION */ | |
2686 | ||
62b61f61 | 2687 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 HD |
2688 | static void wait_while_offlining(void) |
2689 | { | |
2690 | while (ksm_run & KSM_RUN_OFFLINE) { | |
2691 | mutex_unlock(&ksm_thread_mutex); | |
2692 | wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), | |
74316201 | 2693 | TASK_UNINTERRUPTIBLE); |
ef4d43a8 HD |
2694 | mutex_lock(&ksm_thread_mutex); |
2695 | } | |
2696 | } | |
2697 | ||
2c653d0e AA |
2698 | static bool stable_node_dup_remove_range(struct stable_node *stable_node, |
2699 | unsigned long start_pfn, | |
2700 | unsigned long end_pfn) | |
2701 | { | |
2702 | if (stable_node->kpfn >= start_pfn && | |
2703 | stable_node->kpfn < end_pfn) { | |
2704 | /* | |
2705 | * Don't get_ksm_page, page has already gone: | |
2706 | * which is why we keep kpfn instead of page* | |
2707 | */ | |
2708 | remove_node_from_stable_tree(stable_node); | |
2709 | return true; | |
2710 | } | |
2711 | return false; | |
2712 | } | |
2713 | ||
2714 | static bool stable_node_chain_remove_range(struct stable_node *stable_node, | |
2715 | unsigned long start_pfn, | |
2716 | unsigned long end_pfn, | |
2717 | struct rb_root *root) | |
2718 | { | |
2719 | struct stable_node *dup; | |
2720 | struct hlist_node *hlist_safe; | |
2721 | ||
2722 | if (!is_stable_node_chain(stable_node)) { | |
2723 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
2724 | return stable_node_dup_remove_range(stable_node, start_pfn, | |
2725 | end_pfn); | |
2726 | } | |
2727 | ||
2728 | hlist_for_each_entry_safe(dup, hlist_safe, | |
2729 | &stable_node->hlist, hlist_dup) { | |
2730 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
2731 | stable_node_dup_remove_range(dup, start_pfn, end_pfn); | |
2732 | } | |
2733 | if (hlist_empty(&stable_node->hlist)) { | |
2734 | free_stable_node_chain(stable_node, root); | |
2735 | return true; /* notify caller that tree was rebalanced */ | |
2736 | } else | |
2737 | return false; | |
2738 | } | |
2739 | ||
ee0ea59c HD |
2740 | static void ksm_check_stable_tree(unsigned long start_pfn, |
2741 | unsigned long end_pfn) | |
62b61f61 | 2742 | { |
03640418 | 2743 | struct stable_node *stable_node, *next; |
62b61f61 | 2744 | struct rb_node *node; |
90bd6fd3 | 2745 | int nid; |
62b61f61 | 2746 | |
ef53d16c HD |
2747 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
2748 | node = rb_first(root_stable_tree + nid); | |
ee0ea59c | 2749 | while (node) { |
90bd6fd3 | 2750 | stable_node = rb_entry(node, struct stable_node, node); |
2c653d0e AA |
2751 | if (stable_node_chain_remove_range(stable_node, |
2752 | start_pfn, end_pfn, | |
2753 | root_stable_tree + | |
2754 | nid)) | |
ef53d16c | 2755 | node = rb_first(root_stable_tree + nid); |
2c653d0e | 2756 | else |
ee0ea59c HD |
2757 | node = rb_next(node); |
2758 | cond_resched(); | |
90bd6fd3 | 2759 | } |
ee0ea59c | 2760 | } |
03640418 | 2761 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
2762 | if (stable_node->kpfn >= start_pfn && |
2763 | stable_node->kpfn < end_pfn) | |
2764 | remove_node_from_stable_tree(stable_node); | |
2765 | cond_resched(); | |
2766 | } | |
62b61f61 HD |
2767 | } |
2768 | ||
2769 | static int ksm_memory_callback(struct notifier_block *self, | |
2770 | unsigned long action, void *arg) | |
2771 | { | |
2772 | struct memory_notify *mn = arg; | |
62b61f61 HD |
2773 | |
2774 | switch (action) { | |
2775 | case MEM_GOING_OFFLINE: | |
2776 | /* | |
ef4d43a8 HD |
2777 | * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() |
2778 | * and remove_all_stable_nodes() while memory is going offline: | |
2779 | * it is unsafe for them to touch the stable tree at this time. | |
2780 | * But unmerge_ksm_pages(), rmap lookups and other entry points | |
2781 | * which do not need the ksm_thread_mutex are all safe. | |
62b61f61 | 2782 | */ |
ef4d43a8 HD |
2783 | mutex_lock(&ksm_thread_mutex); |
2784 | ksm_run |= KSM_RUN_OFFLINE; | |
2785 | mutex_unlock(&ksm_thread_mutex); | |
62b61f61 HD |
2786 | break; |
2787 | ||
2788 | case MEM_OFFLINE: | |
2789 | /* | |
2790 | * Most of the work is done by page migration; but there might | |
2791 | * be a few stable_nodes left over, still pointing to struct | |
ee0ea59c HD |
2792 | * pages which have been offlined: prune those from the tree, |
2793 | * otherwise get_ksm_page() might later try to access a | |
2794 | * non-existent struct page. | |
62b61f61 | 2795 | */ |
ee0ea59c HD |
2796 | ksm_check_stable_tree(mn->start_pfn, |
2797 | mn->start_pfn + mn->nr_pages); | |
e4a9bc58 | 2798 | fallthrough; |
62b61f61 | 2799 | case MEM_CANCEL_OFFLINE: |
ef4d43a8 HD |
2800 | mutex_lock(&ksm_thread_mutex); |
2801 | ksm_run &= ~KSM_RUN_OFFLINE; | |
62b61f61 | 2802 | mutex_unlock(&ksm_thread_mutex); |
ef4d43a8 HD |
2803 | |
2804 | smp_mb(); /* wake_up_bit advises this */ | |
2805 | wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); | |
62b61f61 HD |
2806 | break; |
2807 | } | |
2808 | return NOTIFY_OK; | |
2809 | } | |
ef4d43a8 HD |
2810 | #else |
2811 | static void wait_while_offlining(void) | |
2812 | { | |
2813 | } | |
62b61f61 HD |
2814 | #endif /* CONFIG_MEMORY_HOTREMOVE */ |
2815 | ||
2ffd8679 HD |
2816 | #ifdef CONFIG_SYSFS |
2817 | /* | |
2818 | * This all compiles without CONFIG_SYSFS, but is a waste of space. | |
2819 | */ | |
2820 | ||
31dbd01f IE |
2821 | #define KSM_ATTR_RO(_name) \ |
2822 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2823 | #define KSM_ATTR(_name) \ | |
2824 | static struct kobj_attribute _name##_attr = \ | |
2825 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2826 | ||
2827 | static ssize_t sleep_millisecs_show(struct kobject *kobj, | |
2828 | struct kobj_attribute *attr, char *buf) | |
2829 | { | |
ae7a927d | 2830 | return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs); |
31dbd01f IE |
2831 | } |
2832 | ||
2833 | static ssize_t sleep_millisecs_store(struct kobject *kobj, | |
2834 | struct kobj_attribute *attr, | |
2835 | const char *buf, size_t count) | |
2836 | { | |
dfefd226 | 2837 | unsigned int msecs; |
31dbd01f IE |
2838 | int err; |
2839 | ||
dfefd226 AD |
2840 | err = kstrtouint(buf, 10, &msecs); |
2841 | if (err) | |
31dbd01f IE |
2842 | return -EINVAL; |
2843 | ||
2844 | ksm_thread_sleep_millisecs = msecs; | |
fcf9a0ef | 2845 | wake_up_interruptible(&ksm_iter_wait); |
31dbd01f IE |
2846 | |
2847 | return count; | |
2848 | } | |
2849 | KSM_ATTR(sleep_millisecs); | |
2850 | ||
2851 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
2852 | struct kobj_attribute *attr, char *buf) | |
2853 | { | |
ae7a927d | 2854 | return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan); |
31dbd01f IE |
2855 | } |
2856 | ||
2857 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
2858 | struct kobj_attribute *attr, | |
2859 | const char *buf, size_t count) | |
2860 | { | |
dfefd226 | 2861 | unsigned int nr_pages; |
31dbd01f | 2862 | int err; |
31dbd01f | 2863 | |
dfefd226 AD |
2864 | err = kstrtouint(buf, 10, &nr_pages); |
2865 | if (err) | |
31dbd01f IE |
2866 | return -EINVAL; |
2867 | ||
2868 | ksm_thread_pages_to_scan = nr_pages; | |
2869 | ||
2870 | return count; | |
2871 | } | |
2872 | KSM_ATTR(pages_to_scan); | |
2873 | ||
2874 | static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, | |
2875 | char *buf) | |
2876 | { | |
ae7a927d | 2877 | return sysfs_emit(buf, "%lu\n", ksm_run); |
31dbd01f IE |
2878 | } |
2879 | ||
2880 | static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, | |
2881 | const char *buf, size_t count) | |
2882 | { | |
dfefd226 | 2883 | unsigned int flags; |
31dbd01f | 2884 | int err; |
31dbd01f | 2885 | |
dfefd226 AD |
2886 | err = kstrtouint(buf, 10, &flags); |
2887 | if (err) | |
31dbd01f IE |
2888 | return -EINVAL; |
2889 | if (flags > KSM_RUN_UNMERGE) | |
2890 | return -EINVAL; | |
2891 | ||
2892 | /* | |
2893 | * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. | |
2894 | * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, | |
d0f209f6 HD |
2895 | * breaking COW to free the pages_shared (but leaves mm_slots |
2896 | * on the list for when ksmd may be set running again). | |
31dbd01f IE |
2897 | */ |
2898 | ||
2899 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 2900 | wait_while_offlining(); |
31dbd01f IE |
2901 | if (ksm_run != flags) { |
2902 | ksm_run = flags; | |
d952b791 | 2903 | if (flags & KSM_RUN_UNMERGE) { |
e1e12d2f | 2904 | set_current_oom_origin(); |
d952b791 | 2905 | err = unmerge_and_remove_all_rmap_items(); |
e1e12d2f | 2906 | clear_current_oom_origin(); |
d952b791 HD |
2907 | if (err) { |
2908 | ksm_run = KSM_RUN_STOP; | |
2909 | count = err; | |
2910 | } | |
2911 | } | |
31dbd01f IE |
2912 | } |
2913 | mutex_unlock(&ksm_thread_mutex); | |
2914 | ||
2915 | if (flags & KSM_RUN_MERGE) | |
2916 | wake_up_interruptible(&ksm_thread_wait); | |
2917 | ||
2918 | return count; | |
2919 | } | |
2920 | KSM_ATTR(run); | |
2921 | ||
90bd6fd3 PH |
2922 | #ifdef CONFIG_NUMA |
2923 | static ssize_t merge_across_nodes_show(struct kobject *kobj, | |
ae7a927d | 2924 | struct kobj_attribute *attr, char *buf) |
90bd6fd3 | 2925 | { |
ae7a927d | 2926 | return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes); |
90bd6fd3 PH |
2927 | } |
2928 | ||
2929 | static ssize_t merge_across_nodes_store(struct kobject *kobj, | |
2930 | struct kobj_attribute *attr, | |
2931 | const char *buf, size_t count) | |
2932 | { | |
2933 | int err; | |
2934 | unsigned long knob; | |
2935 | ||
2936 | err = kstrtoul(buf, 10, &knob); | |
2937 | if (err) | |
2938 | return err; | |
2939 | if (knob > 1) | |
2940 | return -EINVAL; | |
2941 | ||
2942 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 2943 | wait_while_offlining(); |
90bd6fd3 | 2944 | if (ksm_merge_across_nodes != knob) { |
cbf86cfe | 2945 | if (ksm_pages_shared || remove_all_stable_nodes()) |
90bd6fd3 | 2946 | err = -EBUSY; |
ef53d16c HD |
2947 | else if (root_stable_tree == one_stable_tree) { |
2948 | struct rb_root *buf; | |
2949 | /* | |
2950 | * This is the first time that we switch away from the | |
2951 | * default of merging across nodes: must now allocate | |
2952 | * a buffer to hold as many roots as may be needed. | |
2953 | * Allocate stable and unstable together: | |
2954 | * MAXSMP NODES_SHIFT 10 will use 16kB. | |
2955 | */ | |
bafe1e14 JP |
2956 | buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), |
2957 | GFP_KERNEL); | |
ef53d16c HD |
2958 | /* Let us assume that RB_ROOT is NULL is zero */ |
2959 | if (!buf) | |
2960 | err = -ENOMEM; | |
2961 | else { | |
2962 | root_stable_tree = buf; | |
2963 | root_unstable_tree = buf + nr_node_ids; | |
2964 | /* Stable tree is empty but not the unstable */ | |
2965 | root_unstable_tree[0] = one_unstable_tree[0]; | |
2966 | } | |
2967 | } | |
2968 | if (!err) { | |
90bd6fd3 | 2969 | ksm_merge_across_nodes = knob; |
ef53d16c HD |
2970 | ksm_nr_node_ids = knob ? 1 : nr_node_ids; |
2971 | } | |
90bd6fd3 PH |
2972 | } |
2973 | mutex_unlock(&ksm_thread_mutex); | |
2974 | ||
2975 | return err ? err : count; | |
2976 | } | |
2977 | KSM_ATTR(merge_across_nodes); | |
2978 | #endif | |
2979 | ||
e86c59b1 | 2980 | static ssize_t use_zero_pages_show(struct kobject *kobj, |
ae7a927d | 2981 | struct kobj_attribute *attr, char *buf) |
e86c59b1 | 2982 | { |
ae7a927d | 2983 | return sysfs_emit(buf, "%u\n", ksm_use_zero_pages); |
e86c59b1 CI |
2984 | } |
2985 | static ssize_t use_zero_pages_store(struct kobject *kobj, | |
2986 | struct kobj_attribute *attr, | |
2987 | const char *buf, size_t count) | |
2988 | { | |
2989 | int err; | |
2990 | bool value; | |
2991 | ||
2992 | err = kstrtobool(buf, &value); | |
2993 | if (err) | |
2994 | return -EINVAL; | |
2995 | ||
2996 | ksm_use_zero_pages = value; | |
2997 | ||
2998 | return count; | |
2999 | } | |
3000 | KSM_ATTR(use_zero_pages); | |
3001 | ||
2c653d0e AA |
3002 | static ssize_t max_page_sharing_show(struct kobject *kobj, |
3003 | struct kobj_attribute *attr, char *buf) | |
3004 | { | |
ae7a927d | 3005 | return sysfs_emit(buf, "%u\n", ksm_max_page_sharing); |
2c653d0e AA |
3006 | } |
3007 | ||
3008 | static ssize_t max_page_sharing_store(struct kobject *kobj, | |
3009 | struct kobj_attribute *attr, | |
3010 | const char *buf, size_t count) | |
3011 | { | |
3012 | int err; | |
3013 | int knob; | |
3014 | ||
3015 | err = kstrtoint(buf, 10, &knob); | |
3016 | if (err) | |
3017 | return err; | |
3018 | /* | |
3019 | * When a KSM page is created it is shared by 2 mappings. This | |
3020 | * being a signed comparison, it implicitly verifies it's not | |
3021 | * negative. | |
3022 | */ | |
3023 | if (knob < 2) | |
3024 | return -EINVAL; | |
3025 | ||
3026 | if (READ_ONCE(ksm_max_page_sharing) == knob) | |
3027 | return count; | |
3028 | ||
3029 | mutex_lock(&ksm_thread_mutex); | |
3030 | wait_while_offlining(); | |
3031 | if (ksm_max_page_sharing != knob) { | |
3032 | if (ksm_pages_shared || remove_all_stable_nodes()) | |
3033 | err = -EBUSY; | |
3034 | else | |
3035 | ksm_max_page_sharing = knob; | |
3036 | } | |
3037 | mutex_unlock(&ksm_thread_mutex); | |
3038 | ||
3039 | return err ? err : count; | |
3040 | } | |
3041 | KSM_ATTR(max_page_sharing); | |
3042 | ||
b4028260 HD |
3043 | static ssize_t pages_shared_show(struct kobject *kobj, |
3044 | struct kobj_attribute *attr, char *buf) | |
3045 | { | |
ae7a927d | 3046 | return sysfs_emit(buf, "%lu\n", ksm_pages_shared); |
b4028260 HD |
3047 | } |
3048 | KSM_ATTR_RO(pages_shared); | |
3049 | ||
3050 | static ssize_t pages_sharing_show(struct kobject *kobj, | |
3051 | struct kobj_attribute *attr, char *buf) | |
3052 | { | |
ae7a927d | 3053 | return sysfs_emit(buf, "%lu\n", ksm_pages_sharing); |
b4028260 HD |
3054 | } |
3055 | KSM_ATTR_RO(pages_sharing); | |
3056 | ||
473b0ce4 HD |
3057 | static ssize_t pages_unshared_show(struct kobject *kobj, |
3058 | struct kobj_attribute *attr, char *buf) | |
3059 | { | |
ae7a927d | 3060 | return sysfs_emit(buf, "%lu\n", ksm_pages_unshared); |
473b0ce4 HD |
3061 | } |
3062 | KSM_ATTR_RO(pages_unshared); | |
3063 | ||
3064 | static ssize_t pages_volatile_show(struct kobject *kobj, | |
3065 | struct kobj_attribute *attr, char *buf) | |
3066 | { | |
3067 | long ksm_pages_volatile; | |
3068 | ||
3069 | ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared | |
3070 | - ksm_pages_sharing - ksm_pages_unshared; | |
3071 | /* | |
3072 | * It was not worth any locking to calculate that statistic, | |
3073 | * but it might therefore sometimes be negative: conceal that. | |
3074 | */ | |
3075 | if (ksm_pages_volatile < 0) | |
3076 | ksm_pages_volatile = 0; | |
ae7a927d | 3077 | return sysfs_emit(buf, "%ld\n", ksm_pages_volatile); |
473b0ce4 HD |
3078 | } |
3079 | KSM_ATTR_RO(pages_volatile); | |
3080 | ||
2c653d0e AA |
3081 | static ssize_t stable_node_dups_show(struct kobject *kobj, |
3082 | struct kobj_attribute *attr, char *buf) | |
3083 | { | |
ae7a927d | 3084 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups); |
2c653d0e AA |
3085 | } |
3086 | KSM_ATTR_RO(stable_node_dups); | |
3087 | ||
3088 | static ssize_t stable_node_chains_show(struct kobject *kobj, | |
3089 | struct kobj_attribute *attr, char *buf) | |
3090 | { | |
ae7a927d | 3091 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains); |
2c653d0e AA |
3092 | } |
3093 | KSM_ATTR_RO(stable_node_chains); | |
3094 | ||
3095 | static ssize_t | |
3096 | stable_node_chains_prune_millisecs_show(struct kobject *kobj, | |
3097 | struct kobj_attribute *attr, | |
3098 | char *buf) | |
3099 | { | |
ae7a927d | 3100 | return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs); |
2c653d0e AA |
3101 | } |
3102 | ||
3103 | static ssize_t | |
3104 | stable_node_chains_prune_millisecs_store(struct kobject *kobj, | |
3105 | struct kobj_attribute *attr, | |
3106 | const char *buf, size_t count) | |
3107 | { | |
584ff0df | 3108 | unsigned int msecs; |
2c653d0e AA |
3109 | int err; |
3110 | ||
584ff0df ZB |
3111 | err = kstrtouint(buf, 10, &msecs); |
3112 | if (err) | |
2c653d0e AA |
3113 | return -EINVAL; |
3114 | ||
3115 | ksm_stable_node_chains_prune_millisecs = msecs; | |
3116 | ||
3117 | return count; | |
3118 | } | |
3119 | KSM_ATTR(stable_node_chains_prune_millisecs); | |
3120 | ||
473b0ce4 HD |
3121 | static ssize_t full_scans_show(struct kobject *kobj, |
3122 | struct kobj_attribute *attr, char *buf) | |
3123 | { | |
ae7a927d | 3124 | return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr); |
473b0ce4 HD |
3125 | } |
3126 | KSM_ATTR_RO(full_scans); | |
3127 | ||
31dbd01f IE |
3128 | static struct attribute *ksm_attrs[] = { |
3129 | &sleep_millisecs_attr.attr, | |
3130 | &pages_to_scan_attr.attr, | |
3131 | &run_attr.attr, | |
b4028260 HD |
3132 | &pages_shared_attr.attr, |
3133 | &pages_sharing_attr.attr, | |
473b0ce4 HD |
3134 | &pages_unshared_attr.attr, |
3135 | &pages_volatile_attr.attr, | |
3136 | &full_scans_attr.attr, | |
90bd6fd3 PH |
3137 | #ifdef CONFIG_NUMA |
3138 | &merge_across_nodes_attr.attr, | |
3139 | #endif | |
2c653d0e AA |
3140 | &max_page_sharing_attr.attr, |
3141 | &stable_node_chains_attr.attr, | |
3142 | &stable_node_dups_attr.attr, | |
3143 | &stable_node_chains_prune_millisecs_attr.attr, | |
e86c59b1 | 3144 | &use_zero_pages_attr.attr, |
31dbd01f IE |
3145 | NULL, |
3146 | }; | |
3147 | ||
f907c26a | 3148 | static const struct attribute_group ksm_attr_group = { |
31dbd01f IE |
3149 | .attrs = ksm_attrs, |
3150 | .name = "ksm", | |
3151 | }; | |
2ffd8679 | 3152 | #endif /* CONFIG_SYSFS */ |
31dbd01f IE |
3153 | |
3154 | static int __init ksm_init(void) | |
3155 | { | |
3156 | struct task_struct *ksm_thread; | |
3157 | int err; | |
3158 | ||
e86c59b1 CI |
3159 | /* The correct value depends on page size and endianness */ |
3160 | zero_checksum = calc_checksum(ZERO_PAGE(0)); | |
3161 | /* Default to false for backwards compatibility */ | |
3162 | ksm_use_zero_pages = false; | |
3163 | ||
31dbd01f IE |
3164 | err = ksm_slab_init(); |
3165 | if (err) | |
3166 | goto out; | |
3167 | ||
31dbd01f IE |
3168 | ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); |
3169 | if (IS_ERR(ksm_thread)) { | |
25acde31 | 3170 | pr_err("ksm: creating kthread failed\n"); |
31dbd01f | 3171 | err = PTR_ERR(ksm_thread); |
d9f8984c | 3172 | goto out_free; |
31dbd01f IE |
3173 | } |
3174 | ||
2ffd8679 | 3175 | #ifdef CONFIG_SYSFS |
31dbd01f IE |
3176 | err = sysfs_create_group(mm_kobj, &ksm_attr_group); |
3177 | if (err) { | |
25acde31 | 3178 | pr_err("ksm: register sysfs failed\n"); |
2ffd8679 | 3179 | kthread_stop(ksm_thread); |
d9f8984c | 3180 | goto out_free; |
31dbd01f | 3181 | } |
c73602ad HD |
3182 | #else |
3183 | ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ | |
3184 | ||
2ffd8679 | 3185 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 3186 | |
62b61f61 | 3187 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 | 3188 | /* There is no significance to this priority 100 */ |
62b61f61 HD |
3189 | hotplug_memory_notifier(ksm_memory_callback, 100); |
3190 | #endif | |
31dbd01f IE |
3191 | return 0; |
3192 | ||
d9f8984c | 3193 | out_free: |
31dbd01f IE |
3194 | ksm_slab_free(); |
3195 | out: | |
3196 | return err; | |
f8af4da3 | 3197 | } |
a64fb3cd | 3198 | subsys_initcall(ksm_init); |