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 | 17 | #include <linux/mm.h> |
36090def | 18 | #include <linux/mm_inline.h> |
31dbd01f | 19 | #include <linux/fs.h> |
f8af4da3 | 20 | #include <linux/mman.h> |
31dbd01f | 21 | #include <linux/sched.h> |
6e84f315 | 22 | #include <linux/sched/mm.h> |
f7ccbae4 | 23 | #include <linux/sched/coredump.h> |
31dbd01f IE |
24 | #include <linux/rwsem.h> |
25 | #include <linux/pagemap.h> | |
26 | #include <linux/rmap.h> | |
27 | #include <linux/spinlock.h> | |
59e1a2f4 | 28 | #include <linux/xxhash.h> |
31dbd01f IE |
29 | #include <linux/delay.h> |
30 | #include <linux/kthread.h> | |
31 | #include <linux/wait.h> | |
32 | #include <linux/slab.h> | |
33 | #include <linux/rbtree.h> | |
62b61f61 | 34 | #include <linux/memory.h> |
31dbd01f | 35 | #include <linux/mmu_notifier.h> |
2c6854fd | 36 | #include <linux/swap.h> |
f8af4da3 | 37 | #include <linux/ksm.h> |
4ca3a69b | 38 | #include <linux/hashtable.h> |
878aee7d | 39 | #include <linux/freezer.h> |
72788c38 | 40 | #include <linux/oom.h> |
90bd6fd3 | 41 | #include <linux/numa.h> |
d7c0e68d | 42 | #include <linux/pagewalk.h> |
f8af4da3 | 43 | |
31dbd01f | 44 | #include <asm/tlbflush.h> |
73848b46 | 45 | #include "internal.h" |
58730ab6 | 46 | #include "mm_slot.h" |
31dbd01f | 47 | |
739100c8 SR |
48 | #define CREATE_TRACE_POINTS |
49 | #include <trace/events/ksm.h> | |
50 | ||
e850dcf5 HD |
51 | #ifdef CONFIG_NUMA |
52 | #define NUMA(x) (x) | |
53 | #define DO_NUMA(x) do { (x); } while (0) | |
54 | #else | |
55 | #define NUMA(x) (0) | |
56 | #define DO_NUMA(x) do { } while (0) | |
57 | #endif | |
58 | ||
5a2ca3ef MR |
59 | /** |
60 | * DOC: Overview | |
61 | * | |
31dbd01f IE |
62 | * A few notes about the KSM scanning process, |
63 | * to make it easier to understand the data structures below: | |
64 | * | |
65 | * In order to reduce excessive scanning, KSM sorts the memory pages by their | |
66 | * contents into a data structure that holds pointers to the pages' locations. | |
67 | * | |
68 | * Since the contents of the pages may change at any moment, KSM cannot just | |
69 | * insert the pages into a normal sorted tree and expect it to find anything. | |
70 | * Therefore KSM uses two data structures - the stable and the unstable tree. | |
71 | * | |
72 | * The stable tree holds pointers to all the merged pages (ksm pages), sorted | |
73 | * by their contents. Because each such page is write-protected, searching on | |
74 | * this tree is fully assured to be working (except when pages are unmapped), | |
75 | * and therefore this tree is called the stable tree. | |
76 | * | |
5a2ca3ef MR |
77 | * The stable tree node includes information required for reverse |
78 | * mapping from a KSM page to virtual addresses that map this page. | |
79 | * | |
80 | * In order to avoid large latencies of the rmap walks on KSM pages, | |
81 | * KSM maintains two types of nodes in the stable tree: | |
82 | * | |
83 | * * the regular nodes that keep the reverse mapping structures in a | |
84 | * linked list | |
85 | * * the "chains" that link nodes ("dups") that represent the same | |
86 | * write protected memory content, but each "dup" corresponds to a | |
87 | * different KSM page copy of that content | |
88 | * | |
89 | * Internally, the regular nodes, "dups" and "chains" are represented | |
21fbd591 | 90 | * using the same struct ksm_stable_node structure. |
5a2ca3ef | 91 | * |
31dbd01f IE |
92 | * In addition to the stable tree, KSM uses a second data structure called the |
93 | * unstable tree: this tree holds pointers to pages which have been found to | |
94 | * be "unchanged for a period of time". The unstable tree sorts these pages | |
95 | * by their contents, but since they are not write-protected, KSM cannot rely | |
96 | * upon the unstable tree to work correctly - the unstable tree is liable to | |
97 | * be corrupted as its contents are modified, and so it is called unstable. | |
98 | * | |
99 | * KSM solves this problem by several techniques: | |
100 | * | |
101 | * 1) The unstable tree is flushed every time KSM completes scanning all | |
102 | * memory areas, and then the tree is rebuilt again from the beginning. | |
103 | * 2) KSM will only insert into the unstable tree, pages whose hash value | |
104 | * has not changed since the previous scan of all memory areas. | |
105 | * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the | |
106 | * colors of the nodes and not on their contents, assuring that even when | |
107 | * the tree gets "corrupted" it won't get out of balance, so scanning time | |
108 | * remains the same (also, searching and inserting nodes in an rbtree uses | |
109 | * the same algorithm, so we have no overhead when we flush and rebuild). | |
110 | * 4) KSM never flushes the stable tree, which means that even if it were to | |
111 | * take 10 attempts to find a page in the unstable tree, once it is found, | |
112 | * it is secured in the stable tree. (When we scan a new page, we first | |
113 | * compare it against the stable tree, and then against the unstable tree.) | |
8fdb3dbf HD |
114 | * |
115 | * If the merge_across_nodes tunable is unset, then KSM maintains multiple | |
116 | * stable trees and multiple unstable trees: one of each for each NUMA node. | |
31dbd01f IE |
117 | */ |
118 | ||
119 | /** | |
21fbd591 | 120 | * struct ksm_mm_slot - ksm information per mm that is being scanned |
58730ab6 | 121 | * @slot: hash lookup from mm to mm_slot |
6514d511 | 122 | * @rmap_list: head for this mm_slot's singly-linked list of rmap_items |
31dbd01f | 123 | */ |
21fbd591 | 124 | struct ksm_mm_slot { |
58730ab6 | 125 | struct mm_slot slot; |
21fbd591 | 126 | struct ksm_rmap_item *rmap_list; |
31dbd01f IE |
127 | }; |
128 | ||
129 | /** | |
130 | * struct ksm_scan - cursor for scanning | |
131 | * @mm_slot: the current mm_slot we are scanning | |
132 | * @address: the next address inside that to be scanned | |
6514d511 | 133 | * @rmap_list: link to the next rmap to be scanned in the rmap_list |
31dbd01f IE |
134 | * @seqnr: count of completed full scans (needed when removing unstable node) |
135 | * | |
136 | * There is only the one ksm_scan instance of this cursor structure. | |
137 | */ | |
138 | struct ksm_scan { | |
21fbd591 | 139 | struct ksm_mm_slot *mm_slot; |
31dbd01f | 140 | unsigned long address; |
21fbd591 | 141 | struct ksm_rmap_item **rmap_list; |
31dbd01f IE |
142 | unsigned long seqnr; |
143 | }; | |
144 | ||
7b6ba2c7 | 145 | /** |
21fbd591 | 146 | * struct ksm_stable_node - node of the stable rbtree |
7b6ba2c7 | 147 | * @node: rb node of this ksm page in the stable tree |
4146d2d6 | 148 | * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list |
2c653d0e | 149 | * @hlist_dup: linked into the stable_node->hlist with a stable_node chain |
4146d2d6 | 150 | * @list: linked into migrate_nodes, pending placement in the proper node tree |
7b6ba2c7 | 151 | * @hlist: hlist head of rmap_items using this ksm page |
4146d2d6 | 152 | * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) |
2c653d0e AA |
153 | * @chain_prune_time: time of the last full garbage collection |
154 | * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN | |
4146d2d6 | 155 | * @nid: NUMA node id of stable tree in which linked (may not match kpfn) |
7b6ba2c7 | 156 | */ |
21fbd591 | 157 | struct ksm_stable_node { |
4146d2d6 HD |
158 | union { |
159 | struct rb_node node; /* when node of stable tree */ | |
160 | struct { /* when listed for migration */ | |
161 | struct list_head *head; | |
2c653d0e AA |
162 | struct { |
163 | struct hlist_node hlist_dup; | |
164 | struct list_head list; | |
165 | }; | |
4146d2d6 HD |
166 | }; |
167 | }; | |
7b6ba2c7 | 168 | struct hlist_head hlist; |
2c653d0e AA |
169 | union { |
170 | unsigned long kpfn; | |
171 | unsigned long chain_prune_time; | |
172 | }; | |
173 | /* | |
174 | * STABLE_NODE_CHAIN can be any negative number in | |
175 | * rmap_hlist_len negative range, but better not -1 to be able | |
176 | * to reliably detect underflows. | |
177 | */ | |
178 | #define STABLE_NODE_CHAIN -1024 | |
179 | int rmap_hlist_len; | |
4146d2d6 HD |
180 | #ifdef CONFIG_NUMA |
181 | int nid; | |
182 | #endif | |
7b6ba2c7 HD |
183 | }; |
184 | ||
31dbd01f | 185 | /** |
21fbd591 | 186 | * struct ksm_rmap_item - reverse mapping item for virtual addresses |
6514d511 | 187 | * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list |
db114b83 | 188 | * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree |
bc56620b | 189 | * @nid: NUMA node id of unstable tree in which linked (may not match page) |
31dbd01f IE |
190 | * @mm: the memory structure this rmap_item is pointing into |
191 | * @address: the virtual address this rmap_item tracks (+ flags in low bits) | |
192 | * @oldchecksum: previous checksum of the page at that virtual address | |
7b6ba2c7 HD |
193 | * @node: rb node of this rmap_item in the unstable tree |
194 | * @head: pointer to stable_node heading this list in the stable tree | |
195 | * @hlist: link into hlist of rmap_items hanging off that stable_node | |
31dbd01f | 196 | */ |
21fbd591 QZ |
197 | struct ksm_rmap_item { |
198 | struct ksm_rmap_item *rmap_list; | |
bc56620b HD |
199 | union { |
200 | struct anon_vma *anon_vma; /* when stable */ | |
201 | #ifdef CONFIG_NUMA | |
202 | int nid; /* when node of unstable tree */ | |
203 | #endif | |
204 | }; | |
31dbd01f IE |
205 | struct mm_struct *mm; |
206 | unsigned long address; /* + low bits used for flags below */ | |
7b6ba2c7 | 207 | unsigned int oldchecksum; /* when unstable */ |
31dbd01f | 208 | union { |
7b6ba2c7 HD |
209 | struct rb_node node; /* when node of unstable tree */ |
210 | struct { /* when listed from stable tree */ | |
21fbd591 | 211 | struct ksm_stable_node *head; |
7b6ba2c7 HD |
212 | struct hlist_node hlist; |
213 | }; | |
31dbd01f IE |
214 | }; |
215 | }; | |
216 | ||
217 | #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ | |
7b6ba2c7 HD |
218 | #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ |
219 | #define STABLE_FLAG 0x200 /* is listed from the stable tree */ | |
31dbd01f IE |
220 | |
221 | /* The stable and unstable tree heads */ | |
ef53d16c HD |
222 | static struct rb_root one_stable_tree[1] = { RB_ROOT }; |
223 | static struct rb_root one_unstable_tree[1] = { RB_ROOT }; | |
224 | static struct rb_root *root_stable_tree = one_stable_tree; | |
225 | static struct rb_root *root_unstable_tree = one_unstable_tree; | |
31dbd01f | 226 | |
4146d2d6 HD |
227 | /* Recently migrated nodes of stable tree, pending proper placement */ |
228 | static LIST_HEAD(migrate_nodes); | |
2c653d0e | 229 | #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev) |
4146d2d6 | 230 | |
4ca3a69b SL |
231 | #define MM_SLOTS_HASH_BITS 10 |
232 | static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
31dbd01f | 233 | |
21fbd591 | 234 | static struct ksm_mm_slot ksm_mm_head = { |
58730ab6 | 235 | .slot.mm_node = LIST_HEAD_INIT(ksm_mm_head.slot.mm_node), |
31dbd01f IE |
236 | }; |
237 | static struct ksm_scan ksm_scan = { | |
238 | .mm_slot = &ksm_mm_head, | |
239 | }; | |
240 | ||
241 | static struct kmem_cache *rmap_item_cache; | |
7b6ba2c7 | 242 | static struct kmem_cache *stable_node_cache; |
31dbd01f IE |
243 | static struct kmem_cache *mm_slot_cache; |
244 | ||
245 | /* The number of nodes in the stable tree */ | |
b4028260 | 246 | static unsigned long ksm_pages_shared; |
31dbd01f | 247 | |
e178dfde | 248 | /* The number of page slots additionally sharing those nodes */ |
b4028260 | 249 | static unsigned long ksm_pages_sharing; |
31dbd01f | 250 | |
473b0ce4 HD |
251 | /* The number of nodes in the unstable tree */ |
252 | static unsigned long ksm_pages_unshared; | |
253 | ||
254 | /* The number of rmap_items in use: to calculate pages_volatile */ | |
255 | static unsigned long ksm_rmap_items; | |
256 | ||
2c653d0e AA |
257 | /* The number of stable_node chains */ |
258 | static unsigned long ksm_stable_node_chains; | |
259 | ||
260 | /* The number of stable_node dups linked to the stable_node chains */ | |
261 | static unsigned long ksm_stable_node_dups; | |
262 | ||
263 | /* Delay in pruning stale stable_node_dups in the stable_node_chains */ | |
584ff0df | 264 | static unsigned int ksm_stable_node_chains_prune_millisecs = 2000; |
2c653d0e AA |
265 | |
266 | /* Maximum number of page slots sharing a stable node */ | |
267 | static int ksm_max_page_sharing = 256; | |
268 | ||
31dbd01f | 269 | /* Number of pages ksmd should scan in one batch */ |
2c6854fd | 270 | static unsigned int ksm_thread_pages_to_scan = 100; |
31dbd01f IE |
271 | |
272 | /* Milliseconds ksmd should sleep between batches */ | |
2ffd8679 | 273 | static unsigned int ksm_thread_sleep_millisecs = 20; |
31dbd01f | 274 | |
e86c59b1 CI |
275 | /* Checksum of an empty (zeroed) page */ |
276 | static unsigned int zero_checksum __read_mostly; | |
277 | ||
278 | /* Whether to merge empty (zeroed) pages with actual zero pages */ | |
279 | static bool ksm_use_zero_pages __read_mostly; | |
280 | ||
e850dcf5 | 281 | #ifdef CONFIG_NUMA |
90bd6fd3 PH |
282 | /* Zeroed when merging across nodes is not allowed */ |
283 | static unsigned int ksm_merge_across_nodes = 1; | |
ef53d16c | 284 | static int ksm_nr_node_ids = 1; |
e850dcf5 HD |
285 | #else |
286 | #define ksm_merge_across_nodes 1U | |
ef53d16c | 287 | #define ksm_nr_node_ids 1 |
e850dcf5 | 288 | #endif |
90bd6fd3 | 289 | |
31dbd01f IE |
290 | #define KSM_RUN_STOP 0 |
291 | #define KSM_RUN_MERGE 1 | |
292 | #define KSM_RUN_UNMERGE 2 | |
ef4d43a8 HD |
293 | #define KSM_RUN_OFFLINE 4 |
294 | static unsigned long ksm_run = KSM_RUN_STOP; | |
295 | static void wait_while_offlining(void); | |
31dbd01f IE |
296 | |
297 | static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); | |
fcf9a0ef | 298 | static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait); |
31dbd01f IE |
299 | static DEFINE_MUTEX(ksm_thread_mutex); |
300 | static DEFINE_SPINLOCK(ksm_mmlist_lock); | |
301 | ||
21fbd591 | 302 | #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create(#__struct,\ |
31dbd01f IE |
303 | sizeof(struct __struct), __alignof__(struct __struct),\ |
304 | (__flags), NULL) | |
305 | ||
306 | static int __init ksm_slab_init(void) | |
307 | { | |
21fbd591 | 308 | rmap_item_cache = KSM_KMEM_CACHE(ksm_rmap_item, 0); |
31dbd01f IE |
309 | if (!rmap_item_cache) |
310 | goto out; | |
311 | ||
21fbd591 | 312 | stable_node_cache = KSM_KMEM_CACHE(ksm_stable_node, 0); |
7b6ba2c7 HD |
313 | if (!stable_node_cache) |
314 | goto out_free1; | |
315 | ||
21fbd591 | 316 | mm_slot_cache = KSM_KMEM_CACHE(ksm_mm_slot, 0); |
31dbd01f | 317 | if (!mm_slot_cache) |
7b6ba2c7 | 318 | goto out_free2; |
31dbd01f IE |
319 | |
320 | return 0; | |
321 | ||
7b6ba2c7 HD |
322 | out_free2: |
323 | kmem_cache_destroy(stable_node_cache); | |
324 | out_free1: | |
31dbd01f IE |
325 | kmem_cache_destroy(rmap_item_cache); |
326 | out: | |
327 | return -ENOMEM; | |
328 | } | |
329 | ||
330 | static void __init ksm_slab_free(void) | |
331 | { | |
332 | kmem_cache_destroy(mm_slot_cache); | |
7b6ba2c7 | 333 | kmem_cache_destroy(stable_node_cache); |
31dbd01f IE |
334 | kmem_cache_destroy(rmap_item_cache); |
335 | mm_slot_cache = NULL; | |
336 | } | |
337 | ||
21fbd591 | 338 | static __always_inline bool is_stable_node_chain(struct ksm_stable_node *chain) |
2c653d0e AA |
339 | { |
340 | return chain->rmap_hlist_len == STABLE_NODE_CHAIN; | |
341 | } | |
342 | ||
21fbd591 | 343 | static __always_inline bool is_stable_node_dup(struct ksm_stable_node *dup) |
2c653d0e AA |
344 | { |
345 | return dup->head == STABLE_NODE_DUP_HEAD; | |
346 | } | |
347 | ||
21fbd591 QZ |
348 | static inline void stable_node_chain_add_dup(struct ksm_stable_node *dup, |
349 | struct ksm_stable_node *chain) | |
2c653d0e AA |
350 | { |
351 | VM_BUG_ON(is_stable_node_dup(dup)); | |
352 | dup->head = STABLE_NODE_DUP_HEAD; | |
353 | VM_BUG_ON(!is_stable_node_chain(chain)); | |
354 | hlist_add_head(&dup->hlist_dup, &chain->hlist); | |
355 | ksm_stable_node_dups++; | |
356 | } | |
357 | ||
21fbd591 | 358 | static inline void __stable_node_dup_del(struct ksm_stable_node *dup) |
2c653d0e | 359 | { |
b4fecc67 | 360 | VM_BUG_ON(!is_stable_node_dup(dup)); |
2c653d0e AA |
361 | hlist_del(&dup->hlist_dup); |
362 | ksm_stable_node_dups--; | |
363 | } | |
364 | ||
21fbd591 | 365 | static inline void stable_node_dup_del(struct ksm_stable_node *dup) |
2c653d0e AA |
366 | { |
367 | VM_BUG_ON(is_stable_node_chain(dup)); | |
368 | if (is_stable_node_dup(dup)) | |
369 | __stable_node_dup_del(dup); | |
370 | else | |
371 | rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid)); | |
372 | #ifdef CONFIG_DEBUG_VM | |
373 | dup->head = NULL; | |
374 | #endif | |
375 | } | |
376 | ||
21fbd591 | 377 | static inline struct ksm_rmap_item *alloc_rmap_item(void) |
31dbd01f | 378 | { |
21fbd591 | 379 | struct ksm_rmap_item *rmap_item; |
473b0ce4 | 380 | |
5b398e41 | 381 | rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | |
382 | __GFP_NORETRY | __GFP_NOWARN); | |
473b0ce4 HD |
383 | if (rmap_item) |
384 | ksm_rmap_items++; | |
385 | return rmap_item; | |
31dbd01f IE |
386 | } |
387 | ||
21fbd591 | 388 | static inline void free_rmap_item(struct ksm_rmap_item *rmap_item) |
31dbd01f | 389 | { |
473b0ce4 | 390 | ksm_rmap_items--; |
cb4df4ca | 391 | rmap_item->mm->ksm_rmap_items--; |
31dbd01f IE |
392 | rmap_item->mm = NULL; /* debug safety */ |
393 | kmem_cache_free(rmap_item_cache, rmap_item); | |
394 | } | |
395 | ||
21fbd591 | 396 | static inline struct ksm_stable_node *alloc_stable_node(void) |
7b6ba2c7 | 397 | { |
6213055f | 398 | /* |
399 | * The allocation can take too long with GFP_KERNEL when memory is under | |
400 | * pressure, which may lead to hung task warnings. Adding __GFP_HIGH | |
401 | * grants access to memory reserves, helping to avoid this problem. | |
402 | */ | |
403 | return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); | |
7b6ba2c7 HD |
404 | } |
405 | ||
21fbd591 | 406 | static inline void free_stable_node(struct ksm_stable_node *stable_node) |
7b6ba2c7 | 407 | { |
2c653d0e AA |
408 | VM_BUG_ON(stable_node->rmap_hlist_len && |
409 | !is_stable_node_chain(stable_node)); | |
7b6ba2c7 HD |
410 | kmem_cache_free(stable_node_cache, stable_node); |
411 | } | |
412 | ||
a913e182 HD |
413 | /* |
414 | * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's | |
415 | * page tables after it has passed through ksm_exit() - which, if necessary, | |
c1e8d7c6 | 416 | * takes mmap_lock briefly to serialize against them. ksm_exit() does not set |
a913e182 HD |
417 | * a special flag: they can just back out as soon as mm_users goes to zero. |
418 | * ksm_test_exit() is used throughout to make this test for exit: in some | |
419 | * places for correctness, in some places just to avoid unnecessary work. | |
420 | */ | |
421 | static inline bool ksm_test_exit(struct mm_struct *mm) | |
422 | { | |
423 | return atomic_read(&mm->mm_users) == 0; | |
424 | } | |
425 | ||
d7c0e68d DH |
426 | static int break_ksm_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long next, |
427 | struct mm_walk *walk) | |
428 | { | |
429 | struct page *page = NULL; | |
430 | spinlock_t *ptl; | |
431 | pte_t *pte; | |
432 | int ret; | |
433 | ||
434 | if (pmd_leaf(*pmd) || !pmd_present(*pmd)) | |
435 | return 0; | |
436 | ||
437 | pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); | |
438 | if (pte_present(*pte)) { | |
439 | page = vm_normal_page(walk->vma, addr, *pte); | |
440 | } else if (!pte_none(*pte)) { | |
441 | swp_entry_t entry = pte_to_swp_entry(*pte); | |
442 | ||
443 | /* | |
444 | * As KSM pages remain KSM pages until freed, no need to wait | |
445 | * here for migration to end. | |
446 | */ | |
447 | if (is_migration_entry(entry)) | |
448 | page = pfn_swap_entry_to_page(entry); | |
449 | } | |
450 | ret = page && PageKsm(page); | |
451 | pte_unmap_unlock(pte, ptl); | |
452 | return ret; | |
453 | } | |
454 | ||
455 | static const struct mm_walk_ops break_ksm_ops = { | |
456 | .pmd_entry = break_ksm_pmd_entry, | |
457 | }; | |
458 | ||
31dbd01f | 459 | /* |
6cce3314 DH |
460 | * We use break_ksm to break COW on a ksm page by triggering unsharing, |
461 | * such that the ksm page will get replaced by an exclusive anonymous page. | |
31dbd01f | 462 | * |
6cce3314 | 463 | * We take great care only to touch a ksm page, in a VM_MERGEABLE vma, |
31dbd01f IE |
464 | * in case the application has unmapped and remapped mm,addr meanwhile. |
465 | * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP | |
bbcd53c9 | 466 | * mmap of /dev/mem, where we would not want to touch it. |
1b2ee126 | 467 | * |
6cce3314 | 468 | * FAULT_FLAG_REMOTE/FOLL_REMOTE are because we do this outside the context |
1b2ee126 DH |
469 | * of the process that owns 'vma'. We also do not want to enforce |
470 | * protection keys here anyway. | |
31dbd01f | 471 | */ |
d952b791 | 472 | static int break_ksm(struct vm_area_struct *vma, unsigned long addr) |
31dbd01f | 473 | { |
50a7ca3c | 474 | vm_fault_t ret = 0; |
31dbd01f IE |
475 | |
476 | do { | |
d7c0e68d | 477 | int ksm_page; |
58f595c6 | 478 | |
31dbd01f | 479 | cond_resched(); |
d7c0e68d DH |
480 | ksm_page = walk_page_range_vma(vma, addr, addr + 1, |
481 | &break_ksm_ops, NULL); | |
482 | if (WARN_ON_ONCE(ksm_page < 0)) | |
483 | return ksm_page; | |
58f595c6 DH |
484 | if (!ksm_page) |
485 | return 0; | |
486 | ret = handle_mm_fault(vma, addr, | |
6cce3314 | 487 | FAULT_FLAG_UNSHARE | FAULT_FLAG_REMOTE, |
58f595c6 DH |
488 | NULL); |
489 | } while (!(ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); | |
d952b791 | 490 | /* |
58f595c6 DH |
491 | * We must loop until we no longer find a KSM page because |
492 | * handle_mm_fault() may back out if there's any difficulty e.g. if | |
493 | * pte accessed bit gets updated concurrently. | |
d952b791 HD |
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 | ||
21fbd591 | 530 | static void break_cow(struct ksm_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 | ||
21fbd591 | 549 | static struct page *get_mergeable_page(struct ksm_rmap_item *rmap_item) |
31dbd01f IE |
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); | |
f7091ed6 | 562 | if (IS_ERR_OR_NULL(page)) |
31dbd01f | 563 | goto out; |
f7091ed6 HW |
564 | if (is_zone_device_page(page)) |
565 | goto out_putpage; | |
f765f540 | 566 | if (PageAnon(page)) { |
31dbd01f IE |
567 | flush_anon_page(vma, page, addr); |
568 | flush_dcache_page(page); | |
569 | } else { | |
f7091ed6 | 570 | out_putpage: |
31dbd01f | 571 | put_page(page); |
c8f95ed1 AA |
572 | out: |
573 | page = NULL; | |
31dbd01f | 574 | } |
d8ed45c5 | 575 | mmap_read_unlock(mm); |
31dbd01f IE |
576 | return page; |
577 | } | |
578 | ||
90bd6fd3 PH |
579 | /* |
580 | * This helper is used for getting right index into array of tree roots. | |
581 | * When merge_across_nodes knob is set to 1, there are only two rb-trees for | |
582 | * stable and unstable pages from all nodes with roots in index 0. Otherwise, | |
583 | * every node has its own stable and unstable tree. | |
584 | */ | |
585 | static inline int get_kpfn_nid(unsigned long kpfn) | |
586 | { | |
d8fc16a8 | 587 | return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); |
90bd6fd3 PH |
588 | } |
589 | ||
21fbd591 | 590 | static struct ksm_stable_node *alloc_stable_node_chain(struct ksm_stable_node *dup, |
2c653d0e AA |
591 | struct rb_root *root) |
592 | { | |
21fbd591 | 593 | struct ksm_stable_node *chain = alloc_stable_node(); |
2c653d0e AA |
594 | VM_BUG_ON(is_stable_node_chain(dup)); |
595 | if (likely(chain)) { | |
596 | INIT_HLIST_HEAD(&chain->hlist); | |
597 | chain->chain_prune_time = jiffies; | |
598 | chain->rmap_hlist_len = STABLE_NODE_CHAIN; | |
599 | #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA) | |
98fa15f3 | 600 | chain->nid = NUMA_NO_NODE; /* debug */ |
2c653d0e AA |
601 | #endif |
602 | ksm_stable_node_chains++; | |
603 | ||
604 | /* | |
605 | * Put the stable node chain in the first dimension of | |
606 | * the stable tree and at the same time remove the old | |
607 | * stable node. | |
608 | */ | |
609 | rb_replace_node(&dup->node, &chain->node, root); | |
610 | ||
611 | /* | |
612 | * Move the old stable node to the second dimension | |
613 | * queued in the hlist_dup. The invariant is that all | |
614 | * dup stable_nodes in the chain->hlist point to pages | |
457aef94 | 615 | * that are write protected and have the exact same |
2c653d0e AA |
616 | * content. |
617 | */ | |
618 | stable_node_chain_add_dup(dup, chain); | |
619 | } | |
620 | return chain; | |
621 | } | |
622 | ||
21fbd591 | 623 | static inline void free_stable_node_chain(struct ksm_stable_node *chain, |
2c653d0e AA |
624 | struct rb_root *root) |
625 | { | |
626 | rb_erase(&chain->node, root); | |
627 | free_stable_node(chain); | |
628 | ksm_stable_node_chains--; | |
629 | } | |
630 | ||
21fbd591 | 631 | static void remove_node_from_stable_tree(struct ksm_stable_node *stable_node) |
4035c07a | 632 | { |
21fbd591 | 633 | struct ksm_rmap_item *rmap_item; |
4035c07a | 634 | |
2c653d0e AA |
635 | /* check it's not STABLE_NODE_CHAIN or negative */ |
636 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
637 | ||
b67bfe0d | 638 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
739100c8 | 639 | if (rmap_item->hlist.next) { |
4035c07a | 640 | ksm_pages_sharing--; |
739100c8 SR |
641 | trace_ksm_remove_rmap_item(stable_node->kpfn, rmap_item, rmap_item->mm); |
642 | } else { | |
4035c07a | 643 | ksm_pages_shared--; |
739100c8 | 644 | } |
76093853 | 645 | |
646 | rmap_item->mm->ksm_merging_pages--; | |
647 | ||
2c653d0e AA |
648 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
649 | stable_node->rmap_hlist_len--; | |
9e60109f | 650 | put_anon_vma(rmap_item->anon_vma); |
4035c07a HD |
651 | rmap_item->address &= PAGE_MASK; |
652 | cond_resched(); | |
653 | } | |
654 | ||
2c653d0e AA |
655 | /* |
656 | * We need the second aligned pointer of the migrate_nodes | |
657 | * list_head to stay clear from the rb_parent_color union | |
658 | * (aligned and different than any node) and also different | |
659 | * from &migrate_nodes. This will verify that future list.h changes | |
815f0ddb | 660 | * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it. |
2c653d0e | 661 | */ |
2c653d0e AA |
662 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes); |
663 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1); | |
2c653d0e | 664 | |
739100c8 | 665 | trace_ksm_remove_ksm_page(stable_node->kpfn); |
4146d2d6 HD |
666 | if (stable_node->head == &migrate_nodes) |
667 | list_del(&stable_node->list); | |
668 | else | |
2c653d0e | 669 | stable_node_dup_del(stable_node); |
4035c07a HD |
670 | free_stable_node(stable_node); |
671 | } | |
672 | ||
2cee57d1 YS |
673 | enum get_ksm_page_flags { |
674 | GET_KSM_PAGE_NOLOCK, | |
675 | GET_KSM_PAGE_LOCK, | |
676 | GET_KSM_PAGE_TRYLOCK | |
677 | }; | |
678 | ||
4035c07a HD |
679 | /* |
680 | * get_ksm_page: checks if the page indicated by the stable node | |
681 | * is still its ksm page, despite having held no reference to it. | |
682 | * In which case we can trust the content of the page, and it | |
683 | * returns the gotten page; but if the page has now been zapped, | |
684 | * remove the stale node from the stable tree and return NULL. | |
c8d6553b | 685 | * But beware, the stable node's page might be being migrated. |
4035c07a HD |
686 | * |
687 | * You would expect the stable_node to hold a reference to the ksm page. | |
688 | * But if it increments the page's count, swapping out has to wait for | |
689 | * ksmd to come around again before it can free the page, which may take | |
690 | * seconds or even minutes: much too unresponsive. So instead we use a | |
691 | * "keyhole reference": access to the ksm page from the stable node peeps | |
692 | * out through its keyhole to see if that page still holds the right key, | |
693 | * pointing back to this stable node. This relies on freeing a PageAnon | |
694 | * page to reset its page->mapping to NULL, and relies on no other use of | |
695 | * a page to put something that might look like our key in page->mapping. | |
4035c07a HD |
696 | * is on its way to being freed; but it is an anomaly to bear in mind. |
697 | */ | |
21fbd591 | 698 | static struct page *get_ksm_page(struct ksm_stable_node *stable_node, |
2cee57d1 | 699 | enum get_ksm_page_flags flags) |
4035c07a HD |
700 | { |
701 | struct page *page; | |
702 | void *expected_mapping; | |
c8d6553b | 703 | unsigned long kpfn; |
4035c07a | 704 | |
bda807d4 MK |
705 | expected_mapping = (void *)((unsigned long)stable_node | |
706 | PAGE_MAPPING_KSM); | |
c8d6553b | 707 | again: |
08df4774 | 708 | kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */ |
c8d6553b | 709 | page = pfn_to_page(kpfn); |
4db0c3c2 | 710 | if (READ_ONCE(page->mapping) != expected_mapping) |
4035c07a | 711 | goto stale; |
c8d6553b HD |
712 | |
713 | /* | |
714 | * We cannot do anything with the page while its refcount is 0. | |
715 | * Usually 0 means free, or tail of a higher-order page: in which | |
716 | * case this node is no longer referenced, and should be freed; | |
1c4c3b99 | 717 | * however, it might mean that the page is under page_ref_freeze(). |
c8d6553b | 718 | * The __remove_mapping() case is easy, again the node is now stale; |
52d1e606 | 719 | * the same is in reuse_ksm_page() case; but if page is swapcache |
9800562f | 720 | * in folio_migrate_mapping(), it might still be our page, |
52d1e606 | 721 | * in which case it's essential to keep the node. |
c8d6553b HD |
722 | */ |
723 | while (!get_page_unless_zero(page)) { | |
724 | /* | |
725 | * Another check for page->mapping != expected_mapping would | |
726 | * work here too. We have chosen the !PageSwapCache test to | |
727 | * optimize the common case, when the page is or is about to | |
728 | * be freed: PageSwapCache is cleared (under spin_lock_irq) | |
1c4c3b99 | 729 | * in the ref_freeze section of __remove_mapping(); but Anon |
c8d6553b HD |
730 | * page->mapping reset to NULL later, in free_pages_prepare(). |
731 | */ | |
732 | if (!PageSwapCache(page)) | |
733 | goto stale; | |
734 | cpu_relax(); | |
735 | } | |
736 | ||
4db0c3c2 | 737 | if (READ_ONCE(page->mapping) != expected_mapping) { |
4035c07a HD |
738 | put_page(page); |
739 | goto stale; | |
740 | } | |
c8d6553b | 741 | |
2cee57d1 YS |
742 | if (flags == GET_KSM_PAGE_TRYLOCK) { |
743 | if (!trylock_page(page)) { | |
744 | put_page(page); | |
745 | return ERR_PTR(-EBUSY); | |
746 | } | |
747 | } else if (flags == GET_KSM_PAGE_LOCK) | |
8aafa6a4 | 748 | lock_page(page); |
2cee57d1 YS |
749 | |
750 | if (flags != GET_KSM_PAGE_NOLOCK) { | |
4db0c3c2 | 751 | if (READ_ONCE(page->mapping) != expected_mapping) { |
8aafa6a4 HD |
752 | unlock_page(page); |
753 | put_page(page); | |
754 | goto stale; | |
755 | } | |
756 | } | |
4035c07a | 757 | return page; |
c8d6553b | 758 | |
4035c07a | 759 | stale: |
c8d6553b HD |
760 | /* |
761 | * We come here from above when page->mapping or !PageSwapCache | |
762 | * suggests that the node is stale; but it might be under migration. | |
19138349 | 763 | * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(), |
c8d6553b HD |
764 | * before checking whether node->kpfn has been changed. |
765 | */ | |
766 | smp_rmb(); | |
4db0c3c2 | 767 | if (READ_ONCE(stable_node->kpfn) != kpfn) |
c8d6553b | 768 | goto again; |
4035c07a HD |
769 | remove_node_from_stable_tree(stable_node); |
770 | return NULL; | |
771 | } | |
772 | ||
31dbd01f IE |
773 | /* |
774 | * Removing rmap_item from stable or unstable tree. | |
775 | * This function will clean the information from the stable/unstable tree. | |
776 | */ | |
21fbd591 | 777 | static void remove_rmap_item_from_tree(struct ksm_rmap_item *rmap_item) |
31dbd01f | 778 | { |
7b6ba2c7 | 779 | if (rmap_item->address & STABLE_FLAG) { |
21fbd591 | 780 | struct ksm_stable_node *stable_node; |
5ad64688 | 781 | struct page *page; |
31dbd01f | 782 | |
7b6ba2c7 | 783 | stable_node = rmap_item->head; |
62862290 | 784 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
4035c07a HD |
785 | if (!page) |
786 | goto out; | |
5ad64688 | 787 | |
7b6ba2c7 | 788 | hlist_del(&rmap_item->hlist); |
62862290 | 789 | unlock_page(page); |
4035c07a | 790 | put_page(page); |
08beca44 | 791 | |
98666f8a | 792 | if (!hlist_empty(&stable_node->hlist)) |
4035c07a HD |
793 | ksm_pages_sharing--; |
794 | else | |
7b6ba2c7 | 795 | ksm_pages_shared--; |
76093853 | 796 | |
797 | rmap_item->mm->ksm_merging_pages--; | |
798 | ||
2c653d0e AA |
799 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
800 | stable_node->rmap_hlist_len--; | |
31dbd01f | 801 | |
9e60109f | 802 | put_anon_vma(rmap_item->anon_vma); |
c89a384e | 803 | rmap_item->head = NULL; |
93d17715 | 804 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 805 | |
7b6ba2c7 | 806 | } else if (rmap_item->address & UNSTABLE_FLAG) { |
31dbd01f IE |
807 | unsigned char age; |
808 | /* | |
9ba69294 | 809 | * Usually ksmd can and must skip the rb_erase, because |
31dbd01f | 810 | * root_unstable_tree was already reset to RB_ROOT. |
9ba69294 HD |
811 | * But be careful when an mm is exiting: do the rb_erase |
812 | * if this rmap_item was inserted by this scan, rather | |
813 | * than left over from before. | |
31dbd01f IE |
814 | */ |
815 | age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); | |
cd551f97 | 816 | BUG_ON(age > 1); |
31dbd01f | 817 | if (!age) |
90bd6fd3 | 818 | rb_erase(&rmap_item->node, |
ef53d16c | 819 | root_unstable_tree + NUMA(rmap_item->nid)); |
473b0ce4 | 820 | ksm_pages_unshared--; |
93d17715 | 821 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 822 | } |
4035c07a | 823 | out: |
31dbd01f IE |
824 | cond_resched(); /* we're called from many long loops */ |
825 | } | |
826 | ||
21fbd591 | 827 | static void remove_trailing_rmap_items(struct ksm_rmap_item **rmap_list) |
31dbd01f | 828 | { |
6514d511 | 829 | while (*rmap_list) { |
21fbd591 | 830 | struct ksm_rmap_item *rmap_item = *rmap_list; |
6514d511 | 831 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 832 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
833 | free_rmap_item(rmap_item); |
834 | } | |
835 | } | |
836 | ||
837 | /* | |
e850dcf5 | 838 | * Though it's very tempting to unmerge rmap_items from stable tree rather |
31dbd01f IE |
839 | * than check every pte of a given vma, the locking doesn't quite work for |
840 | * that - an rmap_item is assigned to the stable tree after inserting ksm | |
c1e8d7c6 | 841 | * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing |
31dbd01f IE |
842 | * rmap_items from parent to child at fork time (so as not to waste time |
843 | * if exit comes before the next scan reaches it). | |
81464e30 HD |
844 | * |
845 | * Similarly, although we'd like to remove rmap_items (so updating counts | |
846 | * and freeing memory) when unmerging an area, it's easier to leave that | |
847 | * to the next pass of ksmd - consider, for example, how ksmd might be | |
848 | * in cmp_and_merge_page on one of the rmap_items we would be removing. | |
31dbd01f | 849 | */ |
d952b791 HD |
850 | static int unmerge_ksm_pages(struct vm_area_struct *vma, |
851 | unsigned long start, unsigned long end) | |
31dbd01f IE |
852 | { |
853 | unsigned long addr; | |
d952b791 | 854 | int err = 0; |
31dbd01f | 855 | |
d952b791 | 856 | for (addr = start; addr < end && !err; addr += PAGE_SIZE) { |
9ba69294 HD |
857 | if (ksm_test_exit(vma->vm_mm)) |
858 | break; | |
d952b791 HD |
859 | if (signal_pending(current)) |
860 | err = -ERESTARTSYS; | |
861 | else | |
862 | err = break_ksm(vma, addr); | |
863 | } | |
864 | return err; | |
31dbd01f IE |
865 | } |
866 | ||
21fbd591 | 867 | static inline struct ksm_stable_node *folio_stable_node(struct folio *folio) |
19138349 MWO |
868 | { |
869 | return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL; | |
870 | } | |
871 | ||
21fbd591 | 872 | static inline struct ksm_stable_node *page_stable_node(struct page *page) |
88484826 | 873 | { |
19138349 | 874 | return folio_stable_node(page_folio(page)); |
88484826 MR |
875 | } |
876 | ||
877 | static inline void set_page_stable_node(struct page *page, | |
21fbd591 | 878 | struct ksm_stable_node *stable_node) |
88484826 | 879 | { |
6c287605 | 880 | VM_BUG_ON_PAGE(PageAnon(page) && PageAnonExclusive(page), page); |
88484826 MR |
881 | page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM); |
882 | } | |
883 | ||
2ffd8679 HD |
884 | #ifdef CONFIG_SYSFS |
885 | /* | |
886 | * Only called through the sysfs control interface: | |
887 | */ | |
21fbd591 | 888 | static int remove_stable_node(struct ksm_stable_node *stable_node) |
cbf86cfe HD |
889 | { |
890 | struct page *page; | |
891 | int err; | |
892 | ||
2cee57d1 | 893 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
cbf86cfe HD |
894 | if (!page) { |
895 | /* | |
896 | * get_ksm_page did remove_node_from_stable_tree itself. | |
897 | */ | |
898 | return 0; | |
899 | } | |
900 | ||
9a63236f AR |
901 | /* |
902 | * Page could be still mapped if this races with __mmput() running in | |
903 | * between ksm_exit() and exit_mmap(). Just refuse to let | |
904 | * merge_across_nodes/max_page_sharing be switched. | |
905 | */ | |
906 | err = -EBUSY; | |
907 | if (!page_mapped(page)) { | |
cbf86cfe | 908 | /* |
8fdb3dbf HD |
909 | * The stable node did not yet appear stale to get_ksm_page(), |
910 | * since that allows for an unmapped ksm page to be recognized | |
911 | * right up until it is freed; but the node is safe to remove. | |
cbf86cfe HD |
912 | * This page might be in a pagevec waiting to be freed, |
913 | * or it might be PageSwapCache (perhaps under writeback), | |
914 | * or it might have been removed from swapcache a moment ago. | |
915 | */ | |
916 | set_page_stable_node(page, NULL); | |
917 | remove_node_from_stable_tree(stable_node); | |
918 | err = 0; | |
919 | } | |
920 | ||
921 | unlock_page(page); | |
922 | put_page(page); | |
923 | return err; | |
924 | } | |
925 | ||
21fbd591 | 926 | static int remove_stable_node_chain(struct ksm_stable_node *stable_node, |
2c653d0e AA |
927 | struct rb_root *root) |
928 | { | |
21fbd591 | 929 | struct ksm_stable_node *dup; |
2c653d0e AA |
930 | struct hlist_node *hlist_safe; |
931 | ||
932 | if (!is_stable_node_chain(stable_node)) { | |
933 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
934 | if (remove_stable_node(stable_node)) | |
935 | return true; | |
936 | else | |
937 | return false; | |
938 | } | |
939 | ||
940 | hlist_for_each_entry_safe(dup, hlist_safe, | |
941 | &stable_node->hlist, hlist_dup) { | |
942 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
943 | if (remove_stable_node(dup)) | |
944 | return true; | |
945 | } | |
946 | BUG_ON(!hlist_empty(&stable_node->hlist)); | |
947 | free_stable_node_chain(stable_node, root); | |
948 | return false; | |
949 | } | |
950 | ||
cbf86cfe HD |
951 | static int remove_all_stable_nodes(void) |
952 | { | |
21fbd591 | 953 | struct ksm_stable_node *stable_node, *next; |
cbf86cfe HD |
954 | int nid; |
955 | int err = 0; | |
956 | ||
ef53d16c | 957 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
cbf86cfe HD |
958 | while (root_stable_tree[nid].rb_node) { |
959 | stable_node = rb_entry(root_stable_tree[nid].rb_node, | |
21fbd591 | 960 | struct ksm_stable_node, node); |
2c653d0e AA |
961 | if (remove_stable_node_chain(stable_node, |
962 | root_stable_tree + nid)) { | |
cbf86cfe HD |
963 | err = -EBUSY; |
964 | break; /* proceed to next nid */ | |
965 | } | |
966 | cond_resched(); | |
967 | } | |
968 | } | |
03640418 | 969 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
970 | if (remove_stable_node(stable_node)) |
971 | err = -EBUSY; | |
972 | cond_resched(); | |
973 | } | |
cbf86cfe HD |
974 | return err; |
975 | } | |
976 | ||
d952b791 | 977 | static int unmerge_and_remove_all_rmap_items(void) |
31dbd01f | 978 | { |
21fbd591 | 979 | struct ksm_mm_slot *mm_slot; |
58730ab6 | 980 | struct mm_slot *slot; |
31dbd01f IE |
981 | struct mm_struct *mm; |
982 | struct vm_area_struct *vma; | |
d952b791 HD |
983 | int err = 0; |
984 | ||
985 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
986 | slot = list_entry(ksm_mm_head.slot.mm_node.next, |
987 | struct mm_slot, mm_node); | |
988 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
d952b791 | 989 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f | 990 | |
a5f18ba0 MWO |
991 | for (mm_slot = ksm_scan.mm_slot; mm_slot != &ksm_mm_head; |
992 | mm_slot = ksm_scan.mm_slot) { | |
58730ab6 | 993 | VMA_ITERATOR(vmi, mm_slot->slot.mm, 0); |
a5f18ba0 | 994 | |
58730ab6 | 995 | mm = mm_slot->slot.mm; |
d8ed45c5 | 996 | mmap_read_lock(mm); |
6db504ce LH |
997 | |
998 | /* | |
999 | * Exit right away if mm is exiting to avoid lockdep issue in | |
1000 | * the maple tree | |
1001 | */ | |
1002 | if (ksm_test_exit(mm)) | |
1003 | goto mm_exiting; | |
1004 | ||
a5f18ba0 | 1005 | for_each_vma(vmi, vma) { |
31dbd01f IE |
1006 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
1007 | continue; | |
d952b791 HD |
1008 | err = unmerge_ksm_pages(vma, |
1009 | vma->vm_start, vma->vm_end); | |
9ba69294 HD |
1010 | if (err) |
1011 | goto error; | |
31dbd01f | 1012 | } |
9ba69294 | 1013 | |
6db504ce | 1014 | mm_exiting: |
420be4ed | 1015 | remove_trailing_rmap_items(&mm_slot->rmap_list); |
d8ed45c5 | 1016 | mmap_read_unlock(mm); |
d952b791 HD |
1017 | |
1018 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
1019 | slot = list_entry(mm_slot->slot.mm_node.next, |
1020 | struct mm_slot, mm_node); | |
1021 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
9ba69294 | 1022 | if (ksm_test_exit(mm)) { |
58730ab6 QZ |
1023 | hash_del(&mm_slot->slot.hash); |
1024 | list_del(&mm_slot->slot.mm_node); | |
9ba69294 HD |
1025 | spin_unlock(&ksm_mmlist_lock); |
1026 | ||
58730ab6 | 1027 | mm_slot_free(mm_slot_cache, mm_slot); |
9ba69294 | 1028 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
9ba69294 | 1029 | mmdrop(mm); |
7496fea9 | 1030 | } else |
9ba69294 | 1031 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f IE |
1032 | } |
1033 | ||
cbf86cfe HD |
1034 | /* Clean up stable nodes, but don't worry if some are still busy */ |
1035 | remove_all_stable_nodes(); | |
d952b791 | 1036 | ksm_scan.seqnr = 0; |
9ba69294 HD |
1037 | return 0; |
1038 | ||
1039 | error: | |
d8ed45c5 | 1040 | mmap_read_unlock(mm); |
31dbd01f | 1041 | spin_lock(&ksm_mmlist_lock); |
d952b791 | 1042 | ksm_scan.mm_slot = &ksm_mm_head; |
31dbd01f | 1043 | spin_unlock(&ksm_mmlist_lock); |
d952b791 | 1044 | return err; |
31dbd01f | 1045 | } |
2ffd8679 | 1046 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 1047 | |
31dbd01f IE |
1048 | static u32 calc_checksum(struct page *page) |
1049 | { | |
1050 | u32 checksum; | |
9b04c5fe | 1051 | void *addr = kmap_atomic(page); |
59e1a2f4 | 1052 | checksum = xxhash(addr, PAGE_SIZE, 0); |
9b04c5fe | 1053 | kunmap_atomic(addr); |
31dbd01f IE |
1054 | return checksum; |
1055 | } | |
1056 | ||
31dbd01f IE |
1057 | static int write_protect_page(struct vm_area_struct *vma, struct page *page, |
1058 | pte_t *orig_pte) | |
1059 | { | |
1060 | struct mm_struct *mm = vma->vm_mm; | |
eed05e54 | 1061 | DEFINE_PAGE_VMA_WALK(pvmw, page, vma, 0, 0); |
31dbd01f IE |
1062 | int swapped; |
1063 | int err = -EFAULT; | |
ac46d4f3 | 1064 | struct mmu_notifier_range range; |
6c287605 | 1065 | bool anon_exclusive; |
31dbd01f | 1066 | |
36eaff33 KS |
1067 | pvmw.address = page_address_in_vma(page, vma); |
1068 | if (pvmw.address == -EFAULT) | |
31dbd01f IE |
1069 | goto out; |
1070 | ||
29ad768c | 1071 | BUG_ON(PageTransCompound(page)); |
6bdb913f | 1072 | |
7d4a8be0 | 1073 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, pvmw.address, |
ac46d4f3 JG |
1074 | pvmw.address + PAGE_SIZE); |
1075 | mmu_notifier_invalidate_range_start(&range); | |
6bdb913f | 1076 | |
36eaff33 | 1077 | if (!page_vma_mapped_walk(&pvmw)) |
6bdb913f | 1078 | goto out_mn; |
36eaff33 KS |
1079 | if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?")) |
1080 | goto out_unlock; | |
31dbd01f | 1081 | |
6c287605 | 1082 | anon_exclusive = PageAnonExclusive(page); |
595cd8f2 | 1083 | if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) || |
6c287605 | 1084 | anon_exclusive || mm_tlb_flush_pending(mm)) { |
31dbd01f IE |
1085 | pte_t entry; |
1086 | ||
1087 | swapped = PageSwapCache(page); | |
36eaff33 | 1088 | flush_cache_page(vma, pvmw.address, page_to_pfn(page)); |
31dbd01f | 1089 | /* |
25985edc | 1090 | * Ok this is tricky, when get_user_pages_fast() run it doesn't |
31dbd01f | 1091 | * take any lock, therefore the check that we are going to make |
f0953a1b | 1092 | * with the pagecount against the mapcount is racy and |
31dbd01f IE |
1093 | * O_DIRECT can happen right after the check. |
1094 | * So we clear the pte and flush the tlb before the check | |
1095 | * this assure us that no O_DIRECT can happen after the check | |
1096 | * or in the middle of the check. | |
0f10851e JG |
1097 | * |
1098 | * No need to notify as we are downgrading page table to read | |
1099 | * only not changing it to point to a new page. | |
1100 | * | |
ee65728e | 1101 | * See Documentation/mm/mmu_notifier.rst |
31dbd01f | 1102 | */ |
0f10851e | 1103 | entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte); |
31dbd01f IE |
1104 | /* |
1105 | * Check that no O_DIRECT or similar I/O is in progress on the | |
1106 | * page | |
1107 | */ | |
31e855ea | 1108 | if (page_mapcount(page) + 1 + swapped != page_count(page)) { |
36eaff33 | 1109 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f IE |
1110 | goto out_unlock; |
1111 | } | |
6c287605 | 1112 | |
088b8aa5 | 1113 | /* See page_try_share_anon_rmap(): clear PTE first. */ |
6c287605 DH |
1114 | if (anon_exclusive && page_try_share_anon_rmap(page)) { |
1115 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); | |
1116 | goto out_unlock; | |
1117 | } | |
1118 | ||
4e31635c HD |
1119 | if (pte_dirty(entry)) |
1120 | set_page_dirty(page); | |
6a56ccbc DH |
1121 | entry = pte_mkclean(entry); |
1122 | ||
1123 | if (pte_write(entry)) | |
1124 | entry = pte_wrprotect(entry); | |
595cd8f2 | 1125 | |
36eaff33 | 1126 | set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f | 1127 | } |
36eaff33 | 1128 | *orig_pte = *pvmw.pte; |
31dbd01f IE |
1129 | err = 0; |
1130 | ||
1131 | out_unlock: | |
36eaff33 | 1132 | page_vma_mapped_walk_done(&pvmw); |
6bdb913f | 1133 | out_mn: |
ac46d4f3 | 1134 | mmu_notifier_invalidate_range_end(&range); |
31dbd01f IE |
1135 | out: |
1136 | return err; | |
1137 | } | |
1138 | ||
1139 | /** | |
1140 | * replace_page - replace page in vma by new ksm page | |
8dd3557a HD |
1141 | * @vma: vma that holds the pte pointing to page |
1142 | * @page: the page we are replacing by kpage | |
1143 | * @kpage: the ksm page we replace page by | |
31dbd01f IE |
1144 | * @orig_pte: the original value of the pte |
1145 | * | |
1146 | * Returns 0 on success, -EFAULT on failure. | |
1147 | */ | |
8dd3557a HD |
1148 | static int replace_page(struct vm_area_struct *vma, struct page *page, |
1149 | struct page *kpage, pte_t orig_pte) | |
31dbd01f IE |
1150 | { |
1151 | struct mm_struct *mm = vma->vm_mm; | |
b4e6f66e | 1152 | struct folio *folio; |
31dbd01f | 1153 | pmd_t *pmd; |
50722804 | 1154 | pmd_t pmde; |
31dbd01f | 1155 | pte_t *ptep; |
e86c59b1 | 1156 | pte_t newpte; |
31dbd01f IE |
1157 | spinlock_t *ptl; |
1158 | unsigned long addr; | |
31dbd01f | 1159 | int err = -EFAULT; |
ac46d4f3 | 1160 | struct mmu_notifier_range range; |
31dbd01f | 1161 | |
8dd3557a | 1162 | addr = page_address_in_vma(page, vma); |
31dbd01f IE |
1163 | if (addr == -EFAULT) |
1164 | goto out; | |
1165 | ||
6219049a BL |
1166 | pmd = mm_find_pmd(mm, addr); |
1167 | if (!pmd) | |
31dbd01f | 1168 | goto out; |
50722804 ZK |
1169 | /* |
1170 | * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() | |
1171 | * without holding anon_vma lock for write. So when looking for a | |
1172 | * genuine pmde (in which to find pte), test present and !THP together. | |
1173 | */ | |
1174 | pmde = *pmd; | |
1175 | barrier(); | |
1176 | if (!pmd_present(pmde) || pmd_trans_huge(pmde)) | |
1177 | goto out; | |
31dbd01f | 1178 | |
7d4a8be0 | 1179 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, addr, |
6f4f13e8 | 1180 | addr + PAGE_SIZE); |
ac46d4f3 | 1181 | mmu_notifier_invalidate_range_start(&range); |
6bdb913f | 1182 | |
31dbd01f IE |
1183 | ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); |
1184 | if (!pte_same(*ptep, orig_pte)) { | |
1185 | pte_unmap_unlock(ptep, ptl); | |
6bdb913f | 1186 | goto out_mn; |
31dbd01f | 1187 | } |
6c287605 DH |
1188 | VM_BUG_ON_PAGE(PageAnonExclusive(page), page); |
1189 | VM_BUG_ON_PAGE(PageAnon(kpage) && PageAnonExclusive(kpage), kpage); | |
31dbd01f | 1190 | |
e86c59b1 CI |
1191 | /* |
1192 | * No need to check ksm_use_zero_pages here: we can only have a | |
457aef94 | 1193 | * zero_page here if ksm_use_zero_pages was enabled already. |
e86c59b1 CI |
1194 | */ |
1195 | if (!is_zero_pfn(page_to_pfn(kpage))) { | |
1196 | get_page(kpage); | |
f1e2db12 | 1197 | page_add_anon_rmap(kpage, vma, addr, RMAP_NONE); |
e86c59b1 CI |
1198 | newpte = mk_pte(kpage, vma->vm_page_prot); |
1199 | } else { | |
1200 | newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage), | |
1201 | vma->vm_page_prot)); | |
a38c015f CI |
1202 | /* |
1203 | * We're replacing an anonymous page with a zero page, which is | |
1204 | * not anonymous. We need to do proper accounting otherwise we | |
1205 | * will get wrong values in /proc, and a BUG message in dmesg | |
1206 | * when tearing down the mm. | |
1207 | */ | |
1208 | dec_mm_counter(mm, MM_ANONPAGES); | |
e86c59b1 | 1209 | } |
31dbd01f IE |
1210 | |
1211 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
0f10851e JG |
1212 | /* |
1213 | * No need to notify as we are replacing a read only page with another | |
1214 | * read only page with the same content. | |
1215 | * | |
ee65728e | 1216 | * See Documentation/mm/mmu_notifier.rst |
0f10851e JG |
1217 | */ |
1218 | ptep_clear_flush(vma, addr, ptep); | |
e86c59b1 | 1219 | set_pte_at_notify(mm, addr, ptep, newpte); |
31dbd01f | 1220 | |
b4e6f66e | 1221 | folio = page_folio(page); |
cea86fe2 | 1222 | page_remove_rmap(page, vma, false); |
b4e6f66e MWO |
1223 | if (!folio_mapped(folio)) |
1224 | folio_free_swap(folio); | |
1225 | folio_put(folio); | |
31dbd01f IE |
1226 | |
1227 | pte_unmap_unlock(ptep, ptl); | |
1228 | err = 0; | |
6bdb913f | 1229 | out_mn: |
ac46d4f3 | 1230 | mmu_notifier_invalidate_range_end(&range); |
31dbd01f IE |
1231 | out: |
1232 | return err; | |
1233 | } | |
1234 | ||
1235 | /* | |
1236 | * try_to_merge_one_page - take two pages and merge them into one | |
8dd3557a HD |
1237 | * @vma: the vma that holds the pte pointing to page |
1238 | * @page: the PageAnon page that we want to replace with kpage | |
80e14822 HD |
1239 | * @kpage: the PageKsm page that we want to map instead of page, |
1240 | * or NULL the first time when we want to use page as kpage. | |
31dbd01f IE |
1241 | * |
1242 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
1243 | */ | |
1244 | static int try_to_merge_one_page(struct vm_area_struct *vma, | |
8dd3557a | 1245 | struct page *page, struct page *kpage) |
31dbd01f IE |
1246 | { |
1247 | pte_t orig_pte = __pte(0); | |
1248 | int err = -EFAULT; | |
1249 | ||
db114b83 HD |
1250 | if (page == kpage) /* ksm page forked */ |
1251 | return 0; | |
1252 | ||
8dd3557a | 1253 | if (!PageAnon(page)) |
31dbd01f IE |
1254 | goto out; |
1255 | ||
31dbd01f IE |
1256 | /* |
1257 | * We need the page lock to read a stable PageSwapCache in | |
1258 | * write_protect_page(). We use trylock_page() instead of | |
1259 | * lock_page() because we don't want to wait here - we | |
1260 | * prefer to continue scanning and merging different pages, | |
1261 | * then come back to this page when it is unlocked. | |
1262 | */ | |
8dd3557a | 1263 | if (!trylock_page(page)) |
31e855ea | 1264 | goto out; |
f765f540 KS |
1265 | |
1266 | if (PageTransCompound(page)) { | |
a7306c34 | 1267 | if (split_huge_page(page)) |
f765f540 KS |
1268 | goto out_unlock; |
1269 | } | |
1270 | ||
31dbd01f IE |
1271 | /* |
1272 | * If this anonymous page is mapped only here, its pte may need | |
1273 | * to be write-protected. If it's mapped elsewhere, all of its | |
1274 | * ptes are necessarily already write-protected. But in either | |
1275 | * case, we need to lock and check page_count is not raised. | |
1276 | */ | |
80e14822 HD |
1277 | if (write_protect_page(vma, page, &orig_pte) == 0) { |
1278 | if (!kpage) { | |
1279 | /* | |
1280 | * While we hold page lock, upgrade page from | |
1281 | * PageAnon+anon_vma to PageKsm+NULL stable_node: | |
1282 | * stable_tree_insert() will update stable_node. | |
1283 | */ | |
1284 | set_page_stable_node(page, NULL); | |
1285 | mark_page_accessed(page); | |
337ed7eb MK |
1286 | /* |
1287 | * Page reclaim just frees a clean page with no dirty | |
1288 | * ptes: make sure that the ksm page would be swapped. | |
1289 | */ | |
1290 | if (!PageDirty(page)) | |
1291 | SetPageDirty(page); | |
80e14822 HD |
1292 | err = 0; |
1293 | } else if (pages_identical(page, kpage)) | |
1294 | err = replace_page(vma, page, kpage, orig_pte); | |
1295 | } | |
31dbd01f | 1296 | |
f765f540 | 1297 | out_unlock: |
8dd3557a | 1298 | unlock_page(page); |
31dbd01f IE |
1299 | out: |
1300 | return err; | |
1301 | } | |
1302 | ||
81464e30 HD |
1303 | /* |
1304 | * try_to_merge_with_ksm_page - like try_to_merge_two_pages, | |
1305 | * but no new kernel page is allocated: kpage must already be a ksm page. | |
8dd3557a HD |
1306 | * |
1307 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
81464e30 | 1308 | */ |
21fbd591 | 1309 | static int try_to_merge_with_ksm_page(struct ksm_rmap_item *rmap_item, |
8dd3557a | 1310 | struct page *page, struct page *kpage) |
81464e30 | 1311 | { |
8dd3557a | 1312 | struct mm_struct *mm = rmap_item->mm; |
81464e30 HD |
1313 | struct vm_area_struct *vma; |
1314 | int err = -EFAULT; | |
1315 | ||
d8ed45c5 | 1316 | mmap_read_lock(mm); |
85c6e8dd AA |
1317 | vma = find_mergeable_vma(mm, rmap_item->address); |
1318 | if (!vma) | |
81464e30 HD |
1319 | goto out; |
1320 | ||
8dd3557a | 1321 | err = try_to_merge_one_page(vma, page, kpage); |
db114b83 HD |
1322 | if (err) |
1323 | goto out; | |
1324 | ||
bc56620b HD |
1325 | /* Unstable nid is in union with stable anon_vma: remove first */ |
1326 | remove_rmap_item_from_tree(rmap_item); | |
1327 | ||
c1e8d7c6 | 1328 | /* Must get reference to anon_vma while still holding mmap_lock */ |
9e60109f PZ |
1329 | rmap_item->anon_vma = vma->anon_vma; |
1330 | get_anon_vma(vma->anon_vma); | |
81464e30 | 1331 | out: |
d8ed45c5 | 1332 | mmap_read_unlock(mm); |
739100c8 SR |
1333 | trace_ksm_merge_with_ksm_page(kpage, page_to_pfn(kpage ? kpage : page), |
1334 | rmap_item, mm, err); | |
81464e30 HD |
1335 | return err; |
1336 | } | |
1337 | ||
31dbd01f IE |
1338 | /* |
1339 | * try_to_merge_two_pages - take two identical pages and prepare them | |
1340 | * to be merged into one page. | |
1341 | * | |
8dd3557a HD |
1342 | * This function returns the kpage if we successfully merged two identical |
1343 | * pages into one ksm page, NULL otherwise. | |
31dbd01f | 1344 | * |
80e14822 | 1345 | * Note that this function upgrades page to ksm page: if one of the pages |
31dbd01f IE |
1346 | * is already a ksm page, try_to_merge_with_ksm_page should be used. |
1347 | */ | |
21fbd591 | 1348 | static struct page *try_to_merge_two_pages(struct ksm_rmap_item *rmap_item, |
8dd3557a | 1349 | struct page *page, |
21fbd591 | 1350 | struct ksm_rmap_item *tree_rmap_item, |
8dd3557a | 1351 | struct page *tree_page) |
31dbd01f | 1352 | { |
80e14822 | 1353 | int err; |
31dbd01f | 1354 | |
80e14822 | 1355 | err = try_to_merge_with_ksm_page(rmap_item, page, NULL); |
31dbd01f | 1356 | if (!err) { |
8dd3557a | 1357 | err = try_to_merge_with_ksm_page(tree_rmap_item, |
80e14822 | 1358 | tree_page, page); |
31dbd01f | 1359 | /* |
81464e30 HD |
1360 | * If that fails, we have a ksm page with only one pte |
1361 | * pointing to it: so break it. | |
31dbd01f | 1362 | */ |
4035c07a | 1363 | if (err) |
8dd3557a | 1364 | break_cow(rmap_item); |
31dbd01f | 1365 | } |
80e14822 | 1366 | return err ? NULL : page; |
31dbd01f IE |
1367 | } |
1368 | ||
2c653d0e | 1369 | static __always_inline |
21fbd591 | 1370 | bool __is_page_sharing_candidate(struct ksm_stable_node *stable_node, int offset) |
2c653d0e AA |
1371 | { |
1372 | VM_BUG_ON(stable_node->rmap_hlist_len < 0); | |
1373 | /* | |
1374 | * Check that at least one mapping still exists, otherwise | |
1375 | * there's no much point to merge and share with this | |
1376 | * stable_node, as the underlying tree_page of the other | |
1377 | * sharer is going to be freed soon. | |
1378 | */ | |
1379 | return stable_node->rmap_hlist_len && | |
1380 | stable_node->rmap_hlist_len + offset < ksm_max_page_sharing; | |
1381 | } | |
1382 | ||
1383 | static __always_inline | |
21fbd591 | 1384 | bool is_page_sharing_candidate(struct ksm_stable_node *stable_node) |
2c653d0e AA |
1385 | { |
1386 | return __is_page_sharing_candidate(stable_node, 0); | |
1387 | } | |
1388 | ||
21fbd591 QZ |
1389 | static struct page *stable_node_dup(struct ksm_stable_node **_stable_node_dup, |
1390 | struct ksm_stable_node **_stable_node, | |
c01f0b54 CIK |
1391 | struct rb_root *root, |
1392 | bool prune_stale_stable_nodes) | |
2c653d0e | 1393 | { |
21fbd591 | 1394 | struct ksm_stable_node *dup, *found = NULL, *stable_node = *_stable_node; |
2c653d0e | 1395 | struct hlist_node *hlist_safe; |
8dc5ffcd | 1396 | struct page *_tree_page, *tree_page = NULL; |
2c653d0e AA |
1397 | int nr = 0; |
1398 | int found_rmap_hlist_len; | |
1399 | ||
1400 | if (!prune_stale_stable_nodes || | |
1401 | time_before(jiffies, stable_node->chain_prune_time + | |
1402 | msecs_to_jiffies( | |
1403 | ksm_stable_node_chains_prune_millisecs))) | |
1404 | prune_stale_stable_nodes = false; | |
1405 | else | |
1406 | stable_node->chain_prune_time = jiffies; | |
1407 | ||
1408 | hlist_for_each_entry_safe(dup, hlist_safe, | |
1409 | &stable_node->hlist, hlist_dup) { | |
1410 | cond_resched(); | |
1411 | /* | |
1412 | * We must walk all stable_node_dup to prune the stale | |
1413 | * stable nodes during lookup. | |
1414 | * | |
1415 | * get_ksm_page can drop the nodes from the | |
1416 | * stable_node->hlist if they point to freed pages | |
1417 | * (that's why we do a _safe walk). The "dup" | |
1418 | * stable_node parameter itself will be freed from | |
1419 | * under us if it returns NULL. | |
1420 | */ | |
2cee57d1 | 1421 | _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK); |
2c653d0e AA |
1422 | if (!_tree_page) |
1423 | continue; | |
1424 | nr += 1; | |
1425 | if (is_page_sharing_candidate(dup)) { | |
1426 | if (!found || | |
1427 | dup->rmap_hlist_len > found_rmap_hlist_len) { | |
1428 | if (found) | |
8dc5ffcd | 1429 | put_page(tree_page); |
2c653d0e AA |
1430 | found = dup; |
1431 | found_rmap_hlist_len = found->rmap_hlist_len; | |
8dc5ffcd | 1432 | tree_page = _tree_page; |
2c653d0e | 1433 | |
8dc5ffcd | 1434 | /* skip put_page for found dup */ |
2c653d0e AA |
1435 | if (!prune_stale_stable_nodes) |
1436 | break; | |
2c653d0e AA |
1437 | continue; |
1438 | } | |
1439 | } | |
1440 | put_page(_tree_page); | |
1441 | } | |
1442 | ||
80b18dfa AA |
1443 | if (found) { |
1444 | /* | |
1445 | * nr is counting all dups in the chain only if | |
1446 | * prune_stale_stable_nodes is true, otherwise we may | |
1447 | * break the loop at nr == 1 even if there are | |
1448 | * multiple entries. | |
1449 | */ | |
1450 | if (prune_stale_stable_nodes && nr == 1) { | |
2c653d0e AA |
1451 | /* |
1452 | * If there's not just one entry it would | |
1453 | * corrupt memory, better BUG_ON. In KSM | |
1454 | * context with no lock held it's not even | |
1455 | * fatal. | |
1456 | */ | |
1457 | BUG_ON(stable_node->hlist.first->next); | |
1458 | ||
1459 | /* | |
1460 | * There's just one entry and it is below the | |
1461 | * deduplication limit so drop the chain. | |
1462 | */ | |
1463 | rb_replace_node(&stable_node->node, &found->node, | |
1464 | root); | |
1465 | free_stable_node(stable_node); | |
1466 | ksm_stable_node_chains--; | |
1467 | ksm_stable_node_dups--; | |
b4fecc67 | 1468 | /* |
0ba1d0f7 AA |
1469 | * NOTE: the caller depends on the stable_node |
1470 | * to be equal to stable_node_dup if the chain | |
1471 | * was collapsed. | |
b4fecc67 | 1472 | */ |
0ba1d0f7 AA |
1473 | *_stable_node = found; |
1474 | /* | |
f0953a1b | 1475 | * Just for robustness, as stable_node is |
0ba1d0f7 AA |
1476 | * otherwise left as a stable pointer, the |
1477 | * compiler shall optimize it away at build | |
1478 | * time. | |
1479 | */ | |
1480 | stable_node = NULL; | |
80b18dfa AA |
1481 | } else if (stable_node->hlist.first != &found->hlist_dup && |
1482 | __is_page_sharing_candidate(found, 1)) { | |
2c653d0e | 1483 | /* |
80b18dfa AA |
1484 | * If the found stable_node dup can accept one |
1485 | * more future merge (in addition to the one | |
1486 | * that is underway) and is not at the head of | |
1487 | * the chain, put it there so next search will | |
1488 | * be quicker in the !prune_stale_stable_nodes | |
1489 | * case. | |
1490 | * | |
1491 | * NOTE: it would be inaccurate to use nr > 1 | |
1492 | * instead of checking the hlist.first pointer | |
1493 | * directly, because in the | |
1494 | * prune_stale_stable_nodes case "nr" isn't | |
1495 | * the position of the found dup in the chain, | |
1496 | * but the total number of dups in the chain. | |
2c653d0e AA |
1497 | */ |
1498 | hlist_del(&found->hlist_dup); | |
1499 | hlist_add_head(&found->hlist_dup, | |
1500 | &stable_node->hlist); | |
1501 | } | |
1502 | } | |
1503 | ||
8dc5ffcd AA |
1504 | *_stable_node_dup = found; |
1505 | return tree_page; | |
2c653d0e AA |
1506 | } |
1507 | ||
21fbd591 | 1508 | static struct ksm_stable_node *stable_node_dup_any(struct ksm_stable_node *stable_node, |
2c653d0e AA |
1509 | struct rb_root *root) |
1510 | { | |
1511 | if (!is_stable_node_chain(stable_node)) | |
1512 | return stable_node; | |
1513 | if (hlist_empty(&stable_node->hlist)) { | |
1514 | free_stable_node_chain(stable_node, root); | |
1515 | return NULL; | |
1516 | } | |
1517 | return hlist_entry(stable_node->hlist.first, | |
1518 | typeof(*stable_node), hlist_dup); | |
1519 | } | |
1520 | ||
8dc5ffcd AA |
1521 | /* |
1522 | * Like for get_ksm_page, this function can free the *_stable_node and | |
1523 | * *_stable_node_dup if the returned tree_page is NULL. | |
1524 | * | |
1525 | * It can also free and overwrite *_stable_node with the found | |
1526 | * stable_node_dup if the chain is collapsed (in which case | |
1527 | * *_stable_node will be equal to *_stable_node_dup like if the chain | |
1528 | * never existed). It's up to the caller to verify tree_page is not | |
1529 | * NULL before dereferencing *_stable_node or *_stable_node_dup. | |
1530 | * | |
1531 | * *_stable_node_dup is really a second output parameter of this | |
1532 | * function and will be overwritten in all cases, the caller doesn't | |
1533 | * need to initialize it. | |
1534 | */ | |
21fbd591 QZ |
1535 | static struct page *__stable_node_chain(struct ksm_stable_node **_stable_node_dup, |
1536 | struct ksm_stable_node **_stable_node, | |
8dc5ffcd AA |
1537 | struct rb_root *root, |
1538 | bool prune_stale_stable_nodes) | |
2c653d0e | 1539 | { |
21fbd591 | 1540 | struct ksm_stable_node *stable_node = *_stable_node; |
2c653d0e AA |
1541 | if (!is_stable_node_chain(stable_node)) { |
1542 | if (is_page_sharing_candidate(stable_node)) { | |
8dc5ffcd | 1543 | *_stable_node_dup = stable_node; |
2cee57d1 | 1544 | return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK); |
2c653d0e | 1545 | } |
8dc5ffcd AA |
1546 | /* |
1547 | * _stable_node_dup set to NULL means the stable_node | |
1548 | * reached the ksm_max_page_sharing limit. | |
1549 | */ | |
1550 | *_stable_node_dup = NULL; | |
2c653d0e AA |
1551 | return NULL; |
1552 | } | |
8dc5ffcd | 1553 | return stable_node_dup(_stable_node_dup, _stable_node, root, |
2c653d0e AA |
1554 | prune_stale_stable_nodes); |
1555 | } | |
1556 | ||
21fbd591 QZ |
1557 | static __always_inline struct page *chain_prune(struct ksm_stable_node **s_n_d, |
1558 | struct ksm_stable_node **s_n, | |
8dc5ffcd | 1559 | struct rb_root *root) |
2c653d0e | 1560 | { |
8dc5ffcd | 1561 | return __stable_node_chain(s_n_d, s_n, root, true); |
2c653d0e AA |
1562 | } |
1563 | ||
21fbd591 QZ |
1564 | static __always_inline struct page *chain(struct ksm_stable_node **s_n_d, |
1565 | struct ksm_stable_node *s_n, | |
8dc5ffcd | 1566 | struct rb_root *root) |
2c653d0e | 1567 | { |
21fbd591 | 1568 | struct ksm_stable_node *old_stable_node = s_n; |
8dc5ffcd AA |
1569 | struct page *tree_page; |
1570 | ||
1571 | tree_page = __stable_node_chain(s_n_d, &s_n, root, false); | |
1572 | /* not pruning dups so s_n cannot have changed */ | |
1573 | VM_BUG_ON(s_n != old_stable_node); | |
1574 | return tree_page; | |
2c653d0e AA |
1575 | } |
1576 | ||
31dbd01f | 1577 | /* |
8dd3557a | 1578 | * stable_tree_search - search for page inside the stable tree |
31dbd01f IE |
1579 | * |
1580 | * This function checks if there is a page inside the stable tree | |
1581 | * with identical content to the page that we are scanning right now. | |
1582 | * | |
7b6ba2c7 | 1583 | * This function returns the stable tree node of identical content if found, |
31dbd01f IE |
1584 | * NULL otherwise. |
1585 | */ | |
62b61f61 | 1586 | static struct page *stable_tree_search(struct page *page) |
31dbd01f | 1587 | { |
90bd6fd3 | 1588 | int nid; |
ef53d16c | 1589 | struct rb_root *root; |
4146d2d6 HD |
1590 | struct rb_node **new; |
1591 | struct rb_node *parent; | |
21fbd591 QZ |
1592 | struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; |
1593 | struct ksm_stable_node *page_node; | |
31dbd01f | 1594 | |
4146d2d6 HD |
1595 | page_node = page_stable_node(page); |
1596 | if (page_node && page_node->head != &migrate_nodes) { | |
1597 | /* ksm page forked */ | |
08beca44 | 1598 | get_page(page); |
62b61f61 | 1599 | return page; |
08beca44 HD |
1600 | } |
1601 | ||
90bd6fd3 | 1602 | nid = get_kpfn_nid(page_to_pfn(page)); |
ef53d16c | 1603 | root = root_stable_tree + nid; |
4146d2d6 | 1604 | again: |
ef53d16c | 1605 | new = &root->rb_node; |
4146d2d6 | 1606 | parent = NULL; |
90bd6fd3 | 1607 | |
4146d2d6 | 1608 | while (*new) { |
4035c07a | 1609 | struct page *tree_page; |
31dbd01f IE |
1610 | int ret; |
1611 | ||
08beca44 | 1612 | cond_resched(); |
21fbd591 | 1613 | stable_node = rb_entry(*new, struct ksm_stable_node, node); |
2c653d0e | 1614 | stable_node_any = NULL; |
8dc5ffcd | 1615 | tree_page = chain_prune(&stable_node_dup, &stable_node, root); |
b4fecc67 AA |
1616 | /* |
1617 | * NOTE: stable_node may have been freed by | |
1618 | * chain_prune() if the returned stable_node_dup is | |
1619 | * not NULL. stable_node_dup may have been inserted in | |
1620 | * the rbtree instead as a regular stable_node (in | |
1621 | * order to collapse the stable_node chain if a single | |
0ba1d0f7 | 1622 | * stable_node dup was found in it). In such case the |
3413b2c8 | 1623 | * stable_node is overwritten by the callee to point |
0ba1d0f7 AA |
1624 | * to the stable_node_dup that was collapsed in the |
1625 | * stable rbtree and stable_node will be equal to | |
1626 | * stable_node_dup like if the chain never existed. | |
b4fecc67 | 1627 | */ |
2c653d0e AA |
1628 | if (!stable_node_dup) { |
1629 | /* | |
1630 | * Either all stable_node dups were full in | |
1631 | * this stable_node chain, or this chain was | |
1632 | * empty and should be rb_erased. | |
1633 | */ | |
1634 | stable_node_any = stable_node_dup_any(stable_node, | |
1635 | root); | |
1636 | if (!stable_node_any) { | |
1637 | /* rb_erase just run */ | |
1638 | goto again; | |
1639 | } | |
1640 | /* | |
1641 | * Take any of the stable_node dups page of | |
1642 | * this stable_node chain to let the tree walk | |
1643 | * continue. All KSM pages belonging to the | |
1644 | * stable_node dups in a stable_node chain | |
1645 | * have the same content and they're | |
457aef94 | 1646 | * write protected at all times. Any will work |
2c653d0e AA |
1647 | * fine to continue the walk. |
1648 | */ | |
2cee57d1 YS |
1649 | tree_page = get_ksm_page(stable_node_any, |
1650 | GET_KSM_PAGE_NOLOCK); | |
2c653d0e AA |
1651 | } |
1652 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1653 | if (!tree_page) { |
1654 | /* | |
1655 | * If we walked over a stale stable_node, | |
1656 | * get_ksm_page() will call rb_erase() and it | |
1657 | * may rebalance the tree from under us. So | |
1658 | * restart the search from scratch. Returning | |
1659 | * NULL would be safe too, but we'd generate | |
1660 | * false negative insertions just because some | |
1661 | * stable_node was stale. | |
1662 | */ | |
1663 | goto again; | |
1664 | } | |
31dbd01f | 1665 | |
4035c07a | 1666 | ret = memcmp_pages(page, tree_page); |
c8d6553b | 1667 | put_page(tree_page); |
31dbd01f | 1668 | |
4146d2d6 | 1669 | parent = *new; |
c8d6553b | 1670 | if (ret < 0) |
4146d2d6 | 1671 | new = &parent->rb_left; |
c8d6553b | 1672 | else if (ret > 0) |
4146d2d6 | 1673 | new = &parent->rb_right; |
c8d6553b | 1674 | else { |
2c653d0e AA |
1675 | if (page_node) { |
1676 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1677 | /* | |
1678 | * Test if the migrated page should be merged | |
1679 | * into a stable node dup. If the mapcount is | |
1680 | * 1 we can migrate it with another KSM page | |
1681 | * without adding it to the chain. | |
1682 | */ | |
1683 | if (page_mapcount(page) > 1) | |
1684 | goto chain_append; | |
1685 | } | |
1686 | ||
1687 | if (!stable_node_dup) { | |
1688 | /* | |
1689 | * If the stable_node is a chain and | |
1690 | * we got a payload match in memcmp | |
1691 | * but we cannot merge the scanned | |
1692 | * page in any of the existing | |
1693 | * stable_node dups because they're | |
1694 | * all full, we need to wait the | |
1695 | * scanned page to find itself a match | |
1696 | * in the unstable tree to create a | |
1697 | * brand new KSM page to add later to | |
1698 | * the dups of this stable_node. | |
1699 | */ | |
1700 | return NULL; | |
1701 | } | |
1702 | ||
c8d6553b HD |
1703 | /* |
1704 | * Lock and unlock the stable_node's page (which | |
1705 | * might already have been migrated) so that page | |
1706 | * migration is sure to notice its raised count. | |
1707 | * It would be more elegant to return stable_node | |
1708 | * than kpage, but that involves more changes. | |
1709 | */ | |
2cee57d1 YS |
1710 | tree_page = get_ksm_page(stable_node_dup, |
1711 | GET_KSM_PAGE_TRYLOCK); | |
1712 | ||
1713 | if (PTR_ERR(tree_page) == -EBUSY) | |
1714 | return ERR_PTR(-EBUSY); | |
1715 | ||
2c653d0e AA |
1716 | if (unlikely(!tree_page)) |
1717 | /* | |
1718 | * The tree may have been rebalanced, | |
1719 | * so re-evaluate parent and new. | |
1720 | */ | |
4146d2d6 | 1721 | goto again; |
2c653d0e AA |
1722 | unlock_page(tree_page); |
1723 | ||
1724 | if (get_kpfn_nid(stable_node_dup->kpfn) != | |
1725 | NUMA(stable_node_dup->nid)) { | |
1726 | put_page(tree_page); | |
1727 | goto replace; | |
1728 | } | |
1729 | return tree_page; | |
c8d6553b | 1730 | } |
31dbd01f IE |
1731 | } |
1732 | ||
4146d2d6 HD |
1733 | if (!page_node) |
1734 | return NULL; | |
1735 | ||
1736 | list_del(&page_node->list); | |
1737 | DO_NUMA(page_node->nid = nid); | |
1738 | rb_link_node(&page_node->node, parent, new); | |
ef53d16c | 1739 | rb_insert_color(&page_node->node, root); |
2c653d0e AA |
1740 | out: |
1741 | if (is_page_sharing_candidate(page_node)) { | |
1742 | get_page(page); | |
1743 | return page; | |
1744 | } else | |
1745 | return NULL; | |
4146d2d6 HD |
1746 | |
1747 | replace: | |
b4fecc67 AA |
1748 | /* |
1749 | * If stable_node was a chain and chain_prune collapsed it, | |
0ba1d0f7 AA |
1750 | * stable_node has been updated to be the new regular |
1751 | * stable_node. A collapse of the chain is indistinguishable | |
1752 | * from the case there was no chain in the stable | |
1753 | * rbtree. Otherwise stable_node is the chain and | |
1754 | * stable_node_dup is the dup to replace. | |
b4fecc67 | 1755 | */ |
0ba1d0f7 | 1756 | if (stable_node_dup == stable_node) { |
b4fecc67 AA |
1757 | VM_BUG_ON(is_stable_node_chain(stable_node_dup)); |
1758 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); | |
2c653d0e AA |
1759 | /* there is no chain */ |
1760 | if (page_node) { | |
1761 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1762 | list_del(&page_node->list); | |
1763 | DO_NUMA(page_node->nid = nid); | |
b4fecc67 AA |
1764 | rb_replace_node(&stable_node_dup->node, |
1765 | &page_node->node, | |
2c653d0e AA |
1766 | root); |
1767 | if (is_page_sharing_candidate(page_node)) | |
1768 | get_page(page); | |
1769 | else | |
1770 | page = NULL; | |
1771 | } else { | |
b4fecc67 | 1772 | rb_erase(&stable_node_dup->node, root); |
2c653d0e AA |
1773 | page = NULL; |
1774 | } | |
4146d2d6 | 1775 | } else { |
2c653d0e AA |
1776 | VM_BUG_ON(!is_stable_node_chain(stable_node)); |
1777 | __stable_node_dup_del(stable_node_dup); | |
1778 | if (page_node) { | |
1779 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1780 | list_del(&page_node->list); | |
1781 | DO_NUMA(page_node->nid = nid); | |
1782 | stable_node_chain_add_dup(page_node, stable_node); | |
1783 | if (is_page_sharing_candidate(page_node)) | |
1784 | get_page(page); | |
1785 | else | |
1786 | page = NULL; | |
1787 | } else { | |
1788 | page = NULL; | |
1789 | } | |
4146d2d6 | 1790 | } |
2c653d0e AA |
1791 | stable_node_dup->head = &migrate_nodes; |
1792 | list_add(&stable_node_dup->list, stable_node_dup->head); | |
4146d2d6 | 1793 | return page; |
2c653d0e AA |
1794 | |
1795 | chain_append: | |
1796 | /* stable_node_dup could be null if it reached the limit */ | |
1797 | if (!stable_node_dup) | |
1798 | stable_node_dup = stable_node_any; | |
b4fecc67 AA |
1799 | /* |
1800 | * If stable_node was a chain and chain_prune collapsed it, | |
0ba1d0f7 AA |
1801 | * stable_node has been updated to be the new regular |
1802 | * stable_node. A collapse of the chain is indistinguishable | |
1803 | * from the case there was no chain in the stable | |
1804 | * rbtree. Otherwise stable_node is the chain and | |
1805 | * stable_node_dup is the dup to replace. | |
b4fecc67 | 1806 | */ |
0ba1d0f7 | 1807 | if (stable_node_dup == stable_node) { |
b4fecc67 | 1808 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); |
2c653d0e AA |
1809 | /* chain is missing so create it */ |
1810 | stable_node = alloc_stable_node_chain(stable_node_dup, | |
1811 | root); | |
1812 | if (!stable_node) | |
1813 | return NULL; | |
1814 | } | |
1815 | /* | |
1816 | * Add this stable_node dup that was | |
1817 | * migrated to the stable_node chain | |
1818 | * of the current nid for this page | |
1819 | * content. | |
1820 | */ | |
b4fecc67 | 1821 | VM_BUG_ON(!is_stable_node_dup(stable_node_dup)); |
2c653d0e AA |
1822 | VM_BUG_ON(page_node->head != &migrate_nodes); |
1823 | list_del(&page_node->list); | |
1824 | DO_NUMA(page_node->nid = nid); | |
1825 | stable_node_chain_add_dup(page_node, stable_node); | |
1826 | goto out; | |
31dbd01f IE |
1827 | } |
1828 | ||
1829 | /* | |
e850dcf5 | 1830 | * stable_tree_insert - insert stable tree node pointing to new ksm page |
31dbd01f IE |
1831 | * into the stable tree. |
1832 | * | |
7b6ba2c7 HD |
1833 | * This function returns the stable tree node just allocated on success, |
1834 | * NULL otherwise. | |
31dbd01f | 1835 | */ |
21fbd591 | 1836 | static struct ksm_stable_node *stable_tree_insert(struct page *kpage) |
31dbd01f | 1837 | { |
90bd6fd3 PH |
1838 | int nid; |
1839 | unsigned long kpfn; | |
ef53d16c | 1840 | struct rb_root *root; |
90bd6fd3 | 1841 | struct rb_node **new; |
f2e5ff85 | 1842 | struct rb_node *parent; |
21fbd591 | 1843 | struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; |
2c653d0e | 1844 | bool need_chain = false; |
31dbd01f | 1845 | |
90bd6fd3 PH |
1846 | kpfn = page_to_pfn(kpage); |
1847 | nid = get_kpfn_nid(kpfn); | |
ef53d16c | 1848 | root = root_stable_tree + nid; |
f2e5ff85 AA |
1849 | again: |
1850 | parent = NULL; | |
ef53d16c | 1851 | new = &root->rb_node; |
90bd6fd3 | 1852 | |
31dbd01f | 1853 | while (*new) { |
4035c07a | 1854 | struct page *tree_page; |
31dbd01f IE |
1855 | int ret; |
1856 | ||
08beca44 | 1857 | cond_resched(); |
21fbd591 | 1858 | stable_node = rb_entry(*new, struct ksm_stable_node, node); |
2c653d0e | 1859 | stable_node_any = NULL; |
8dc5ffcd | 1860 | tree_page = chain(&stable_node_dup, stable_node, root); |
2c653d0e AA |
1861 | if (!stable_node_dup) { |
1862 | /* | |
1863 | * Either all stable_node dups were full in | |
1864 | * this stable_node chain, or this chain was | |
1865 | * empty and should be rb_erased. | |
1866 | */ | |
1867 | stable_node_any = stable_node_dup_any(stable_node, | |
1868 | root); | |
1869 | if (!stable_node_any) { | |
1870 | /* rb_erase just run */ | |
1871 | goto again; | |
1872 | } | |
1873 | /* | |
1874 | * Take any of the stable_node dups page of | |
1875 | * this stable_node chain to let the tree walk | |
1876 | * continue. All KSM pages belonging to the | |
1877 | * stable_node dups in a stable_node chain | |
1878 | * have the same content and they're | |
457aef94 | 1879 | * write protected at all times. Any will work |
2c653d0e AA |
1880 | * fine to continue the walk. |
1881 | */ | |
2cee57d1 YS |
1882 | tree_page = get_ksm_page(stable_node_any, |
1883 | GET_KSM_PAGE_NOLOCK); | |
2c653d0e AA |
1884 | } |
1885 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1886 | if (!tree_page) { |
1887 | /* | |
1888 | * If we walked over a stale stable_node, | |
1889 | * get_ksm_page() will call rb_erase() and it | |
1890 | * may rebalance the tree from under us. So | |
1891 | * restart the search from scratch. Returning | |
1892 | * NULL would be safe too, but we'd generate | |
1893 | * false negative insertions just because some | |
1894 | * stable_node was stale. | |
1895 | */ | |
1896 | goto again; | |
1897 | } | |
31dbd01f | 1898 | |
4035c07a HD |
1899 | ret = memcmp_pages(kpage, tree_page); |
1900 | put_page(tree_page); | |
31dbd01f IE |
1901 | |
1902 | parent = *new; | |
1903 | if (ret < 0) | |
1904 | new = &parent->rb_left; | |
1905 | else if (ret > 0) | |
1906 | new = &parent->rb_right; | |
1907 | else { | |
2c653d0e AA |
1908 | need_chain = true; |
1909 | break; | |
31dbd01f IE |
1910 | } |
1911 | } | |
1912 | ||
2c653d0e AA |
1913 | stable_node_dup = alloc_stable_node(); |
1914 | if (!stable_node_dup) | |
7b6ba2c7 | 1915 | return NULL; |
31dbd01f | 1916 | |
2c653d0e AA |
1917 | INIT_HLIST_HEAD(&stable_node_dup->hlist); |
1918 | stable_node_dup->kpfn = kpfn; | |
1919 | set_page_stable_node(kpage, stable_node_dup); | |
1920 | stable_node_dup->rmap_hlist_len = 0; | |
1921 | DO_NUMA(stable_node_dup->nid = nid); | |
1922 | if (!need_chain) { | |
1923 | rb_link_node(&stable_node_dup->node, parent, new); | |
1924 | rb_insert_color(&stable_node_dup->node, root); | |
1925 | } else { | |
1926 | if (!is_stable_node_chain(stable_node)) { | |
21fbd591 | 1927 | struct ksm_stable_node *orig = stable_node; |
2c653d0e AA |
1928 | /* chain is missing so create it */ |
1929 | stable_node = alloc_stable_node_chain(orig, root); | |
1930 | if (!stable_node) { | |
1931 | free_stable_node(stable_node_dup); | |
1932 | return NULL; | |
1933 | } | |
1934 | } | |
1935 | stable_node_chain_add_dup(stable_node_dup, stable_node); | |
1936 | } | |
08beca44 | 1937 | |
2c653d0e | 1938 | return stable_node_dup; |
31dbd01f IE |
1939 | } |
1940 | ||
1941 | /* | |
8dd3557a HD |
1942 | * unstable_tree_search_insert - search for identical page, |
1943 | * else insert rmap_item into the unstable tree. | |
31dbd01f IE |
1944 | * |
1945 | * This function searches for a page in the unstable tree identical to the | |
1946 | * page currently being scanned; and if no identical page is found in the | |
1947 | * tree, we insert rmap_item as a new object into the unstable tree. | |
1948 | * | |
1949 | * This function returns pointer to rmap_item found to be identical | |
1950 | * to the currently scanned page, NULL otherwise. | |
1951 | * | |
1952 | * This function does both searching and inserting, because they share | |
1953 | * the same walking algorithm in an rbtree. | |
1954 | */ | |
8dd3557a | 1955 | static |
21fbd591 | 1956 | struct ksm_rmap_item *unstable_tree_search_insert(struct ksm_rmap_item *rmap_item, |
8dd3557a HD |
1957 | struct page *page, |
1958 | struct page **tree_pagep) | |
31dbd01f | 1959 | { |
90bd6fd3 PH |
1960 | struct rb_node **new; |
1961 | struct rb_root *root; | |
31dbd01f | 1962 | struct rb_node *parent = NULL; |
90bd6fd3 PH |
1963 | int nid; |
1964 | ||
1965 | nid = get_kpfn_nid(page_to_pfn(page)); | |
ef53d16c | 1966 | root = root_unstable_tree + nid; |
90bd6fd3 | 1967 | new = &root->rb_node; |
31dbd01f IE |
1968 | |
1969 | while (*new) { | |
21fbd591 | 1970 | struct ksm_rmap_item *tree_rmap_item; |
8dd3557a | 1971 | struct page *tree_page; |
31dbd01f IE |
1972 | int ret; |
1973 | ||
d178f27f | 1974 | cond_resched(); |
21fbd591 | 1975 | tree_rmap_item = rb_entry(*new, struct ksm_rmap_item, node); |
8dd3557a | 1976 | tree_page = get_mergeable_page(tree_rmap_item); |
c8f95ed1 | 1977 | if (!tree_page) |
31dbd01f IE |
1978 | return NULL; |
1979 | ||
1980 | /* | |
8dd3557a | 1981 | * Don't substitute a ksm page for a forked page. |
31dbd01f | 1982 | */ |
8dd3557a HD |
1983 | if (page == tree_page) { |
1984 | put_page(tree_page); | |
31dbd01f IE |
1985 | return NULL; |
1986 | } | |
1987 | ||
8dd3557a | 1988 | ret = memcmp_pages(page, tree_page); |
31dbd01f IE |
1989 | |
1990 | parent = *new; | |
1991 | if (ret < 0) { | |
8dd3557a | 1992 | put_page(tree_page); |
31dbd01f IE |
1993 | new = &parent->rb_left; |
1994 | } else if (ret > 0) { | |
8dd3557a | 1995 | put_page(tree_page); |
31dbd01f | 1996 | new = &parent->rb_right; |
b599cbdf HD |
1997 | } else if (!ksm_merge_across_nodes && |
1998 | page_to_nid(tree_page) != nid) { | |
1999 | /* | |
2000 | * If tree_page has been migrated to another NUMA node, | |
2001 | * it will be flushed out and put in the right unstable | |
2002 | * tree next time: only merge with it when across_nodes. | |
2003 | */ | |
2004 | put_page(tree_page); | |
2005 | return NULL; | |
31dbd01f | 2006 | } else { |
8dd3557a | 2007 | *tree_pagep = tree_page; |
31dbd01f IE |
2008 | return tree_rmap_item; |
2009 | } | |
2010 | } | |
2011 | ||
7b6ba2c7 | 2012 | rmap_item->address |= UNSTABLE_FLAG; |
31dbd01f | 2013 | rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); |
e850dcf5 | 2014 | DO_NUMA(rmap_item->nid = nid); |
31dbd01f | 2015 | rb_link_node(&rmap_item->node, parent, new); |
90bd6fd3 | 2016 | rb_insert_color(&rmap_item->node, root); |
31dbd01f | 2017 | |
473b0ce4 | 2018 | ksm_pages_unshared++; |
31dbd01f IE |
2019 | return NULL; |
2020 | } | |
2021 | ||
2022 | /* | |
2023 | * stable_tree_append - add another rmap_item to the linked list of | |
2024 | * rmap_items hanging off a given node of the stable tree, all sharing | |
2025 | * the same ksm page. | |
2026 | */ | |
21fbd591 QZ |
2027 | static void stable_tree_append(struct ksm_rmap_item *rmap_item, |
2028 | struct ksm_stable_node *stable_node, | |
2c653d0e | 2029 | bool max_page_sharing_bypass) |
31dbd01f | 2030 | { |
2c653d0e AA |
2031 | /* |
2032 | * rmap won't find this mapping if we don't insert the | |
2033 | * rmap_item in the right stable_node | |
2034 | * duplicate. page_migration could break later if rmap breaks, | |
2035 | * so we can as well crash here. We really need to check for | |
2036 | * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check | |
457aef94 | 2037 | * for other negative values as an underflow if detected here |
2c653d0e AA |
2038 | * for the first time (and not when decreasing rmap_hlist_len) |
2039 | * would be sign of memory corruption in the stable_node. | |
2040 | */ | |
2041 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
2042 | ||
2043 | stable_node->rmap_hlist_len++; | |
2044 | if (!max_page_sharing_bypass) | |
2045 | /* possibly non fatal but unexpected overflow, only warn */ | |
2046 | WARN_ON_ONCE(stable_node->rmap_hlist_len > | |
2047 | ksm_max_page_sharing); | |
2048 | ||
7b6ba2c7 | 2049 | rmap_item->head = stable_node; |
31dbd01f | 2050 | rmap_item->address |= STABLE_FLAG; |
7b6ba2c7 | 2051 | hlist_add_head(&rmap_item->hlist, &stable_node->hlist); |
e178dfde | 2052 | |
7b6ba2c7 HD |
2053 | if (rmap_item->hlist.next) |
2054 | ksm_pages_sharing++; | |
2055 | else | |
2056 | ksm_pages_shared++; | |
76093853 | 2057 | |
2058 | rmap_item->mm->ksm_merging_pages++; | |
31dbd01f IE |
2059 | } |
2060 | ||
2061 | /* | |
81464e30 HD |
2062 | * cmp_and_merge_page - first see if page can be merged into the stable tree; |
2063 | * if not, compare checksum to previous and if it's the same, see if page can | |
2064 | * be inserted into the unstable tree, or merged with a page already there and | |
2065 | * both transferred to the stable tree. | |
31dbd01f IE |
2066 | * |
2067 | * @page: the page that we are searching identical page to. | |
2068 | * @rmap_item: the reverse mapping into the virtual address of this page | |
2069 | */ | |
21fbd591 | 2070 | static void cmp_and_merge_page(struct page *page, struct ksm_rmap_item *rmap_item) |
31dbd01f | 2071 | { |
4b22927f | 2072 | struct mm_struct *mm = rmap_item->mm; |
21fbd591 | 2073 | struct ksm_rmap_item *tree_rmap_item; |
8dd3557a | 2074 | struct page *tree_page = NULL; |
21fbd591 | 2075 | struct ksm_stable_node *stable_node; |
8dd3557a | 2076 | struct page *kpage; |
31dbd01f IE |
2077 | unsigned int checksum; |
2078 | int err; | |
2c653d0e | 2079 | bool max_page_sharing_bypass = false; |
31dbd01f | 2080 | |
4146d2d6 HD |
2081 | stable_node = page_stable_node(page); |
2082 | if (stable_node) { | |
2083 | if (stable_node->head != &migrate_nodes && | |
2c653d0e AA |
2084 | get_kpfn_nid(READ_ONCE(stable_node->kpfn)) != |
2085 | NUMA(stable_node->nid)) { | |
2086 | stable_node_dup_del(stable_node); | |
4146d2d6 HD |
2087 | stable_node->head = &migrate_nodes; |
2088 | list_add(&stable_node->list, stable_node->head); | |
2089 | } | |
2090 | if (stable_node->head != &migrate_nodes && | |
2091 | rmap_item->head == stable_node) | |
2092 | return; | |
2c653d0e AA |
2093 | /* |
2094 | * If it's a KSM fork, allow it to go over the sharing limit | |
2095 | * without warnings. | |
2096 | */ | |
2097 | if (!is_page_sharing_candidate(stable_node)) | |
2098 | max_page_sharing_bypass = true; | |
4146d2d6 | 2099 | } |
31dbd01f IE |
2100 | |
2101 | /* We first start with searching the page inside the stable tree */ | |
62b61f61 | 2102 | kpage = stable_tree_search(page); |
4146d2d6 HD |
2103 | if (kpage == page && rmap_item->head == stable_node) { |
2104 | put_page(kpage); | |
2105 | return; | |
2106 | } | |
2107 | ||
2108 | remove_rmap_item_from_tree(rmap_item); | |
2109 | ||
62b61f61 | 2110 | if (kpage) { |
2cee57d1 YS |
2111 | if (PTR_ERR(kpage) == -EBUSY) |
2112 | return; | |
2113 | ||
08beca44 | 2114 | err = try_to_merge_with_ksm_page(rmap_item, page, kpage); |
31dbd01f IE |
2115 | if (!err) { |
2116 | /* | |
2117 | * The page was successfully merged: | |
2118 | * add its rmap_item to the stable tree. | |
2119 | */ | |
5ad64688 | 2120 | lock_page(kpage); |
2c653d0e AA |
2121 | stable_tree_append(rmap_item, page_stable_node(kpage), |
2122 | max_page_sharing_bypass); | |
5ad64688 | 2123 | unlock_page(kpage); |
31dbd01f | 2124 | } |
8dd3557a | 2125 | put_page(kpage); |
31dbd01f IE |
2126 | return; |
2127 | } | |
2128 | ||
2129 | /* | |
4035c07a HD |
2130 | * If the hash value of the page has changed from the last time |
2131 | * we calculated it, this page is changing frequently: therefore we | |
2132 | * don't want to insert it in the unstable tree, and we don't want | |
2133 | * to waste our time searching for something identical to it there. | |
31dbd01f IE |
2134 | */ |
2135 | checksum = calc_checksum(page); | |
2136 | if (rmap_item->oldchecksum != checksum) { | |
2137 | rmap_item->oldchecksum = checksum; | |
2138 | return; | |
2139 | } | |
2140 | ||
e86c59b1 CI |
2141 | /* |
2142 | * Same checksum as an empty page. We attempt to merge it with the | |
2143 | * appropriate zero page if the user enabled this via sysfs. | |
2144 | */ | |
2145 | if (ksm_use_zero_pages && (checksum == zero_checksum)) { | |
2146 | struct vm_area_struct *vma; | |
2147 | ||
d8ed45c5 | 2148 | mmap_read_lock(mm); |
4b22927f | 2149 | vma = find_mergeable_vma(mm, rmap_item->address); |
56df70a6 MS |
2150 | if (vma) { |
2151 | err = try_to_merge_one_page(vma, page, | |
2152 | ZERO_PAGE(rmap_item->address)); | |
739100c8 SR |
2153 | trace_ksm_merge_one_page( |
2154 | page_to_pfn(ZERO_PAGE(rmap_item->address)), | |
2155 | rmap_item, mm, err); | |
56df70a6 MS |
2156 | } else { |
2157 | /* | |
2158 | * If the vma is out of date, we do not need to | |
2159 | * continue. | |
2160 | */ | |
2161 | err = 0; | |
2162 | } | |
d8ed45c5 | 2163 | mmap_read_unlock(mm); |
e86c59b1 CI |
2164 | /* |
2165 | * In case of failure, the page was not really empty, so we | |
2166 | * need to continue. Otherwise we're done. | |
2167 | */ | |
2168 | if (!err) | |
2169 | return; | |
2170 | } | |
8dd3557a HD |
2171 | tree_rmap_item = |
2172 | unstable_tree_search_insert(rmap_item, page, &tree_page); | |
31dbd01f | 2173 | if (tree_rmap_item) { |
77da2ba0 CI |
2174 | bool split; |
2175 | ||
8dd3557a HD |
2176 | kpage = try_to_merge_two_pages(rmap_item, page, |
2177 | tree_rmap_item, tree_page); | |
77da2ba0 CI |
2178 | /* |
2179 | * If both pages we tried to merge belong to the same compound | |
2180 | * page, then we actually ended up increasing the reference | |
2181 | * count of the same compound page twice, and split_huge_page | |
2182 | * failed. | |
2183 | * Here we set a flag if that happened, and we use it later to | |
2184 | * try split_huge_page again. Since we call put_page right | |
2185 | * afterwards, the reference count will be correct and | |
2186 | * split_huge_page should succeed. | |
2187 | */ | |
2188 | split = PageTransCompound(page) | |
2189 | && compound_head(page) == compound_head(tree_page); | |
8dd3557a | 2190 | put_page(tree_page); |
8dd3557a | 2191 | if (kpage) { |
bc56620b HD |
2192 | /* |
2193 | * The pages were successfully merged: insert new | |
2194 | * node in the stable tree and add both rmap_items. | |
2195 | */ | |
5ad64688 | 2196 | lock_page(kpage); |
7b6ba2c7 HD |
2197 | stable_node = stable_tree_insert(kpage); |
2198 | if (stable_node) { | |
2c653d0e AA |
2199 | stable_tree_append(tree_rmap_item, stable_node, |
2200 | false); | |
2201 | stable_tree_append(rmap_item, stable_node, | |
2202 | false); | |
7b6ba2c7 | 2203 | } |
5ad64688 | 2204 | unlock_page(kpage); |
7b6ba2c7 | 2205 | |
31dbd01f IE |
2206 | /* |
2207 | * If we fail to insert the page into the stable tree, | |
2208 | * we will have 2 virtual addresses that are pointing | |
2209 | * to a ksm page left outside the stable tree, | |
2210 | * in which case we need to break_cow on both. | |
2211 | */ | |
7b6ba2c7 | 2212 | if (!stable_node) { |
8dd3557a HD |
2213 | break_cow(tree_rmap_item); |
2214 | break_cow(rmap_item); | |
31dbd01f | 2215 | } |
77da2ba0 CI |
2216 | } else if (split) { |
2217 | /* | |
2218 | * We are here if we tried to merge two pages and | |
2219 | * failed because they both belonged to the same | |
2220 | * compound page. We will split the page now, but no | |
2221 | * merging will take place. | |
2222 | * We do not want to add the cost of a full lock; if | |
2223 | * the page is locked, it is better to skip it and | |
2224 | * perhaps try again later. | |
2225 | */ | |
2226 | if (!trylock_page(page)) | |
2227 | return; | |
2228 | split_huge_page(page); | |
2229 | unlock_page(page); | |
31dbd01f | 2230 | } |
31dbd01f IE |
2231 | } |
2232 | } | |
2233 | ||
21fbd591 QZ |
2234 | static struct ksm_rmap_item *get_next_rmap_item(struct ksm_mm_slot *mm_slot, |
2235 | struct ksm_rmap_item **rmap_list, | |
31dbd01f IE |
2236 | unsigned long addr) |
2237 | { | |
21fbd591 | 2238 | struct ksm_rmap_item *rmap_item; |
31dbd01f | 2239 | |
6514d511 HD |
2240 | while (*rmap_list) { |
2241 | rmap_item = *rmap_list; | |
93d17715 | 2242 | if ((rmap_item->address & PAGE_MASK) == addr) |
31dbd01f | 2243 | return rmap_item; |
31dbd01f IE |
2244 | if (rmap_item->address > addr) |
2245 | break; | |
6514d511 | 2246 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 2247 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
2248 | free_rmap_item(rmap_item); |
2249 | } | |
2250 | ||
2251 | rmap_item = alloc_rmap_item(); | |
2252 | if (rmap_item) { | |
2253 | /* It has already been zeroed */ | |
58730ab6 | 2254 | rmap_item->mm = mm_slot->slot.mm; |
cb4df4ca | 2255 | rmap_item->mm->ksm_rmap_items++; |
31dbd01f | 2256 | rmap_item->address = addr; |
6514d511 HD |
2257 | rmap_item->rmap_list = *rmap_list; |
2258 | *rmap_list = rmap_item; | |
31dbd01f IE |
2259 | } |
2260 | return rmap_item; | |
2261 | } | |
2262 | ||
21fbd591 | 2263 | static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page) |
31dbd01f IE |
2264 | { |
2265 | struct mm_struct *mm; | |
58730ab6 QZ |
2266 | struct ksm_mm_slot *mm_slot; |
2267 | struct mm_slot *slot; | |
31dbd01f | 2268 | struct vm_area_struct *vma; |
21fbd591 | 2269 | struct ksm_rmap_item *rmap_item; |
a5f18ba0 | 2270 | struct vma_iterator vmi; |
90bd6fd3 | 2271 | int nid; |
31dbd01f | 2272 | |
58730ab6 | 2273 | if (list_empty(&ksm_mm_head.slot.mm_node)) |
31dbd01f IE |
2274 | return NULL; |
2275 | ||
58730ab6 QZ |
2276 | mm_slot = ksm_scan.mm_slot; |
2277 | if (mm_slot == &ksm_mm_head) { | |
739100c8 SR |
2278 | trace_ksm_start_scan(ksm_scan.seqnr, ksm_rmap_items); |
2279 | ||
2919bfd0 HD |
2280 | /* |
2281 | * A number of pages can hang around indefinitely on per-cpu | |
2282 | * pagevecs, raised page count preventing write_protect_page | |
2283 | * from merging them. Though it doesn't really matter much, | |
2284 | * it is puzzling to see some stuck in pages_volatile until | |
2285 | * other activity jostles them out, and they also prevented | |
2286 | * LTP's KSM test from succeeding deterministically; so drain | |
2287 | * them here (here rather than on entry to ksm_do_scan(), | |
2288 | * so we don't IPI too often when pages_to_scan is set low). | |
2289 | */ | |
2290 | lru_add_drain_all(); | |
2291 | ||
4146d2d6 HD |
2292 | /* |
2293 | * Whereas stale stable_nodes on the stable_tree itself | |
2294 | * get pruned in the regular course of stable_tree_search(), | |
2295 | * those moved out to the migrate_nodes list can accumulate: | |
2296 | * so prune them once before each full scan. | |
2297 | */ | |
2298 | if (!ksm_merge_across_nodes) { | |
21fbd591 | 2299 | struct ksm_stable_node *stable_node, *next; |
4146d2d6 HD |
2300 | struct page *page; |
2301 | ||
03640418 GT |
2302 | list_for_each_entry_safe(stable_node, next, |
2303 | &migrate_nodes, list) { | |
2cee57d1 YS |
2304 | page = get_ksm_page(stable_node, |
2305 | GET_KSM_PAGE_NOLOCK); | |
4146d2d6 HD |
2306 | if (page) |
2307 | put_page(page); | |
2308 | cond_resched(); | |
2309 | } | |
2310 | } | |
2311 | ||
ef53d16c | 2312 | for (nid = 0; nid < ksm_nr_node_ids; nid++) |
90bd6fd3 | 2313 | root_unstable_tree[nid] = RB_ROOT; |
31dbd01f IE |
2314 | |
2315 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
2316 | slot = list_entry(mm_slot->slot.mm_node.next, |
2317 | struct mm_slot, mm_node); | |
2318 | mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
2319 | ksm_scan.mm_slot = mm_slot; | |
31dbd01f | 2320 | spin_unlock(&ksm_mmlist_lock); |
2b472611 HD |
2321 | /* |
2322 | * Although we tested list_empty() above, a racing __ksm_exit | |
2323 | * of the last mm on the list may have removed it since then. | |
2324 | */ | |
58730ab6 | 2325 | if (mm_slot == &ksm_mm_head) |
2b472611 | 2326 | return NULL; |
31dbd01f IE |
2327 | next_mm: |
2328 | ksm_scan.address = 0; | |
58730ab6 | 2329 | ksm_scan.rmap_list = &mm_slot->rmap_list; |
31dbd01f IE |
2330 | } |
2331 | ||
58730ab6 | 2332 | slot = &mm_slot->slot; |
31dbd01f | 2333 | mm = slot->mm; |
a5f18ba0 MWO |
2334 | vma_iter_init(&vmi, mm, ksm_scan.address); |
2335 | ||
d8ed45c5 | 2336 | mmap_read_lock(mm); |
9ba69294 | 2337 | if (ksm_test_exit(mm)) |
a5f18ba0 | 2338 | goto no_vmas; |
9ba69294 | 2339 | |
a5f18ba0 | 2340 | for_each_vma(vmi, vma) { |
31dbd01f IE |
2341 | if (!(vma->vm_flags & VM_MERGEABLE)) |
2342 | continue; | |
2343 | if (ksm_scan.address < vma->vm_start) | |
2344 | ksm_scan.address = vma->vm_start; | |
2345 | if (!vma->anon_vma) | |
2346 | ksm_scan.address = vma->vm_end; | |
2347 | ||
2348 | while (ksm_scan.address < vma->vm_end) { | |
9ba69294 HD |
2349 | if (ksm_test_exit(mm)) |
2350 | break; | |
31dbd01f | 2351 | *page = follow_page(vma, ksm_scan.address, FOLL_GET); |
f7091ed6 | 2352 | if (IS_ERR_OR_NULL(*page)) { |
21ae5b01 AA |
2353 | ksm_scan.address += PAGE_SIZE; |
2354 | cond_resched(); | |
2355 | continue; | |
2356 | } | |
f7091ed6 HW |
2357 | if (is_zone_device_page(*page)) |
2358 | goto next_page; | |
f765f540 | 2359 | if (PageAnon(*page)) { |
31dbd01f IE |
2360 | flush_anon_page(vma, *page, ksm_scan.address); |
2361 | flush_dcache_page(*page); | |
58730ab6 | 2362 | rmap_item = get_next_rmap_item(mm_slot, |
6514d511 | 2363 | ksm_scan.rmap_list, ksm_scan.address); |
31dbd01f | 2364 | if (rmap_item) { |
6514d511 HD |
2365 | ksm_scan.rmap_list = |
2366 | &rmap_item->rmap_list; | |
31dbd01f IE |
2367 | ksm_scan.address += PAGE_SIZE; |
2368 | } else | |
2369 | put_page(*page); | |
d8ed45c5 | 2370 | mmap_read_unlock(mm); |
31dbd01f IE |
2371 | return rmap_item; |
2372 | } | |
f7091ed6 | 2373 | next_page: |
21ae5b01 | 2374 | put_page(*page); |
31dbd01f IE |
2375 | ksm_scan.address += PAGE_SIZE; |
2376 | cond_resched(); | |
2377 | } | |
2378 | } | |
2379 | ||
9ba69294 | 2380 | if (ksm_test_exit(mm)) { |
a5f18ba0 | 2381 | no_vmas: |
9ba69294 | 2382 | ksm_scan.address = 0; |
58730ab6 | 2383 | ksm_scan.rmap_list = &mm_slot->rmap_list; |
9ba69294 | 2384 | } |
31dbd01f IE |
2385 | /* |
2386 | * Nuke all the rmap_items that are above this current rmap: | |
2387 | * because there were no VM_MERGEABLE vmas with such addresses. | |
2388 | */ | |
420be4ed | 2389 | remove_trailing_rmap_items(ksm_scan.rmap_list); |
31dbd01f IE |
2390 | |
2391 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
2392 | slot = list_entry(mm_slot->slot.mm_node.next, |
2393 | struct mm_slot, mm_node); | |
2394 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
cd551f97 HD |
2395 | if (ksm_scan.address == 0) { |
2396 | /* | |
c1e8d7c6 | 2397 | * We've completed a full scan of all vmas, holding mmap_lock |
cd551f97 HD |
2398 | * throughout, and found no VM_MERGEABLE: so do the same as |
2399 | * __ksm_exit does to remove this mm from all our lists now. | |
9ba69294 HD |
2400 | * This applies either when cleaning up after __ksm_exit |
2401 | * (but beware: we can reach here even before __ksm_exit), | |
2402 | * or when all VM_MERGEABLE areas have been unmapped (and | |
c1e8d7c6 | 2403 | * mmap_lock then protects against race with MADV_MERGEABLE). |
cd551f97 | 2404 | */ |
58730ab6 QZ |
2405 | hash_del(&mm_slot->slot.hash); |
2406 | list_del(&mm_slot->slot.mm_node); | |
9ba69294 HD |
2407 | spin_unlock(&ksm_mmlist_lock); |
2408 | ||
58730ab6 | 2409 | mm_slot_free(mm_slot_cache, mm_slot); |
cd551f97 | 2410 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
d8ed45c5 | 2411 | mmap_read_unlock(mm); |
9ba69294 HD |
2412 | mmdrop(mm); |
2413 | } else { | |
d8ed45c5 | 2414 | mmap_read_unlock(mm); |
7496fea9 | 2415 | /* |
3e4e28c5 | 2416 | * mmap_read_unlock(mm) first because after |
7496fea9 ZC |
2417 | * spin_unlock(&ksm_mmlist_lock) run, the "mm" may |
2418 | * already have been freed under us by __ksm_exit() | |
2419 | * because the "mm_slot" is still hashed and | |
2420 | * ksm_scan.mm_slot doesn't point to it anymore. | |
2421 | */ | |
2422 | spin_unlock(&ksm_mmlist_lock); | |
cd551f97 | 2423 | } |
31dbd01f IE |
2424 | |
2425 | /* Repeat until we've completed scanning the whole list */ | |
58730ab6 QZ |
2426 | mm_slot = ksm_scan.mm_slot; |
2427 | if (mm_slot != &ksm_mm_head) | |
31dbd01f IE |
2428 | goto next_mm; |
2429 | ||
739100c8 | 2430 | trace_ksm_stop_scan(ksm_scan.seqnr, ksm_rmap_items); |
31dbd01f IE |
2431 | ksm_scan.seqnr++; |
2432 | return NULL; | |
2433 | } | |
2434 | ||
2435 | /** | |
2436 | * ksm_do_scan - the ksm scanner main worker function. | |
b7701a5f | 2437 | * @scan_npages: number of pages we want to scan before we return. |
31dbd01f IE |
2438 | */ |
2439 | static void ksm_do_scan(unsigned int scan_npages) | |
2440 | { | |
21fbd591 | 2441 | struct ksm_rmap_item *rmap_item; |
3f649ab7 | 2442 | struct page *page; |
31dbd01f | 2443 | |
878aee7d | 2444 | while (scan_npages-- && likely(!freezing(current))) { |
31dbd01f IE |
2445 | cond_resched(); |
2446 | rmap_item = scan_get_next_rmap_item(&page); | |
2447 | if (!rmap_item) | |
2448 | return; | |
4146d2d6 | 2449 | cmp_and_merge_page(page, rmap_item); |
31dbd01f IE |
2450 | put_page(page); |
2451 | } | |
2452 | } | |
2453 | ||
6e158384 HD |
2454 | static int ksmd_should_run(void) |
2455 | { | |
58730ab6 | 2456 | return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.slot.mm_node); |
6e158384 HD |
2457 | } |
2458 | ||
31dbd01f IE |
2459 | static int ksm_scan_thread(void *nothing) |
2460 | { | |
fcf9a0ef KT |
2461 | unsigned int sleep_ms; |
2462 | ||
878aee7d | 2463 | set_freezable(); |
339aa624 | 2464 | set_user_nice(current, 5); |
31dbd01f IE |
2465 | |
2466 | while (!kthread_should_stop()) { | |
6e158384 | 2467 | mutex_lock(&ksm_thread_mutex); |
ef4d43a8 | 2468 | wait_while_offlining(); |
6e158384 | 2469 | if (ksmd_should_run()) |
31dbd01f | 2470 | ksm_do_scan(ksm_thread_pages_to_scan); |
6e158384 HD |
2471 | mutex_unlock(&ksm_thread_mutex); |
2472 | ||
878aee7d AA |
2473 | try_to_freeze(); |
2474 | ||
6e158384 | 2475 | if (ksmd_should_run()) { |
fcf9a0ef KT |
2476 | sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs); |
2477 | wait_event_interruptible_timeout(ksm_iter_wait, | |
2478 | sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs), | |
2479 | msecs_to_jiffies(sleep_ms)); | |
31dbd01f | 2480 | } else { |
878aee7d | 2481 | wait_event_freezable(ksm_thread_wait, |
6e158384 | 2482 | ksmd_should_run() || kthread_should_stop()); |
31dbd01f IE |
2483 | } |
2484 | } | |
2485 | return 0; | |
2486 | } | |
2487 | ||
f8af4da3 HD |
2488 | int ksm_madvise(struct vm_area_struct *vma, unsigned long start, |
2489 | unsigned long end, int advice, unsigned long *vm_flags) | |
2490 | { | |
2491 | struct mm_struct *mm = vma->vm_mm; | |
d952b791 | 2492 | int err; |
f8af4da3 HD |
2493 | |
2494 | switch (advice) { | |
2495 | case MADV_MERGEABLE: | |
2496 | /* | |
2497 | * Be somewhat over-protective for now! | |
2498 | */ | |
2499 | if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE | | |
2500 | VM_PFNMAP | VM_IO | VM_DONTEXPAND | | |
0661a336 | 2501 | VM_HUGETLB | VM_MIXEDMAP)) |
f8af4da3 HD |
2502 | return 0; /* just ignore the advice */ |
2503 | ||
e1fb4a08 DJ |
2504 | if (vma_is_dax(vma)) |
2505 | return 0; | |
2506 | ||
12564485 SA |
2507 | #ifdef VM_SAO |
2508 | if (*vm_flags & VM_SAO) | |
2509 | return 0; | |
2510 | #endif | |
74a04967 KA |
2511 | #ifdef VM_SPARC_ADI |
2512 | if (*vm_flags & VM_SPARC_ADI) | |
2513 | return 0; | |
2514 | #endif | |
cc2383ec | 2515 | |
d952b791 HD |
2516 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { |
2517 | err = __ksm_enter(mm); | |
2518 | if (err) | |
2519 | return err; | |
2520 | } | |
f8af4da3 HD |
2521 | |
2522 | *vm_flags |= VM_MERGEABLE; | |
2523 | break; | |
2524 | ||
2525 | case MADV_UNMERGEABLE: | |
2526 | if (!(*vm_flags & VM_MERGEABLE)) | |
2527 | return 0; /* just ignore the advice */ | |
2528 | ||
d952b791 HD |
2529 | if (vma->anon_vma) { |
2530 | err = unmerge_ksm_pages(vma, start, end); | |
2531 | if (err) | |
2532 | return err; | |
2533 | } | |
f8af4da3 HD |
2534 | |
2535 | *vm_flags &= ~VM_MERGEABLE; | |
2536 | break; | |
2537 | } | |
2538 | ||
2539 | return 0; | |
2540 | } | |
33cf1707 | 2541 | EXPORT_SYMBOL_GPL(ksm_madvise); |
f8af4da3 HD |
2542 | |
2543 | int __ksm_enter(struct mm_struct *mm) | |
2544 | { | |
21fbd591 | 2545 | struct ksm_mm_slot *mm_slot; |
58730ab6 | 2546 | struct mm_slot *slot; |
6e158384 HD |
2547 | int needs_wakeup; |
2548 | ||
58730ab6 | 2549 | mm_slot = mm_slot_alloc(mm_slot_cache); |
31dbd01f IE |
2550 | if (!mm_slot) |
2551 | return -ENOMEM; | |
2552 | ||
58730ab6 QZ |
2553 | slot = &mm_slot->slot; |
2554 | ||
6e158384 | 2555 | /* Check ksm_run too? Would need tighter locking */ |
58730ab6 | 2556 | needs_wakeup = list_empty(&ksm_mm_head.slot.mm_node); |
6e158384 | 2557 | |
31dbd01f | 2558 | spin_lock(&ksm_mmlist_lock); |
58730ab6 | 2559 | mm_slot_insert(mm_slots_hash, mm, slot); |
31dbd01f | 2560 | /* |
cbf86cfe HD |
2561 | * When KSM_RUN_MERGE (or KSM_RUN_STOP), |
2562 | * insert just behind the scanning cursor, to let the area settle | |
31dbd01f IE |
2563 | * down a little; when fork is followed by immediate exec, we don't |
2564 | * want ksmd to waste time setting up and tearing down an rmap_list. | |
cbf86cfe HD |
2565 | * |
2566 | * But when KSM_RUN_UNMERGE, it's important to insert ahead of its | |
2567 | * scanning cursor, otherwise KSM pages in newly forked mms will be | |
2568 | * missed: then we might as well insert at the end of the list. | |
31dbd01f | 2569 | */ |
cbf86cfe | 2570 | if (ksm_run & KSM_RUN_UNMERGE) |
58730ab6 | 2571 | list_add_tail(&slot->mm_node, &ksm_mm_head.slot.mm_node); |
cbf86cfe | 2572 | else |
58730ab6 | 2573 | list_add_tail(&slot->mm_node, &ksm_scan.mm_slot->slot.mm_node); |
31dbd01f IE |
2574 | spin_unlock(&ksm_mmlist_lock); |
2575 | ||
f8af4da3 | 2576 | set_bit(MMF_VM_MERGEABLE, &mm->flags); |
f1f10076 | 2577 | mmgrab(mm); |
6e158384 HD |
2578 | |
2579 | if (needs_wakeup) | |
2580 | wake_up_interruptible(&ksm_thread_wait); | |
2581 | ||
739100c8 | 2582 | trace_ksm_enter(mm); |
f8af4da3 HD |
2583 | return 0; |
2584 | } | |
2585 | ||
1c2fb7a4 | 2586 | void __ksm_exit(struct mm_struct *mm) |
f8af4da3 | 2587 | { |
21fbd591 | 2588 | struct ksm_mm_slot *mm_slot; |
58730ab6 | 2589 | struct mm_slot *slot; |
9ba69294 | 2590 | int easy_to_free = 0; |
cd551f97 | 2591 | |
31dbd01f | 2592 | /* |
9ba69294 HD |
2593 | * This process is exiting: if it's straightforward (as is the |
2594 | * case when ksmd was never running), free mm_slot immediately. | |
2595 | * But if it's at the cursor or has rmap_items linked to it, use | |
c1e8d7c6 | 2596 | * mmap_lock to synchronize with any break_cows before pagetables |
9ba69294 HD |
2597 | * are freed, and leave the mm_slot on the list for ksmd to free. |
2598 | * Beware: ksm may already have noticed it exiting and freed the slot. | |
31dbd01f | 2599 | */ |
9ba69294 | 2600 | |
cd551f97 | 2601 | spin_lock(&ksm_mmlist_lock); |
58730ab6 QZ |
2602 | slot = mm_slot_lookup(mm_slots_hash, mm); |
2603 | mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
9ba69294 | 2604 | if (mm_slot && ksm_scan.mm_slot != mm_slot) { |
6514d511 | 2605 | if (!mm_slot->rmap_list) { |
58730ab6 QZ |
2606 | hash_del(&slot->hash); |
2607 | list_del(&slot->mm_node); | |
9ba69294 HD |
2608 | easy_to_free = 1; |
2609 | } else { | |
58730ab6 QZ |
2610 | list_move(&slot->mm_node, |
2611 | &ksm_scan.mm_slot->slot.mm_node); | |
9ba69294 | 2612 | } |
cd551f97 | 2613 | } |
cd551f97 HD |
2614 | spin_unlock(&ksm_mmlist_lock); |
2615 | ||
9ba69294 | 2616 | if (easy_to_free) { |
58730ab6 | 2617 | mm_slot_free(mm_slot_cache, mm_slot); |
9ba69294 HD |
2618 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
2619 | mmdrop(mm); | |
2620 | } else if (mm_slot) { | |
d8ed45c5 ML |
2621 | mmap_write_lock(mm); |
2622 | mmap_write_unlock(mm); | |
9ba69294 | 2623 | } |
739100c8 SR |
2624 | |
2625 | trace_ksm_exit(mm); | |
31dbd01f IE |
2626 | } |
2627 | ||
cbf86cfe | 2628 | struct page *ksm_might_need_to_copy(struct page *page, |
5ad64688 HD |
2629 | struct vm_area_struct *vma, unsigned long address) |
2630 | { | |
e05b3453 MWO |
2631 | struct folio *folio = page_folio(page); |
2632 | struct anon_vma *anon_vma = folio_anon_vma(folio); | |
5ad64688 HD |
2633 | struct page *new_page; |
2634 | ||
cbf86cfe HD |
2635 | if (PageKsm(page)) { |
2636 | if (page_stable_node(page) && | |
2637 | !(ksm_run & KSM_RUN_UNMERGE)) | |
2638 | return page; /* no need to copy it */ | |
2639 | } else if (!anon_vma) { | |
2640 | return page; /* no need to copy it */ | |
e1c63e11 NS |
2641 | } else if (page->index == linear_page_index(vma, address) && |
2642 | anon_vma->root == vma->anon_vma->root) { | |
cbf86cfe HD |
2643 | return page; /* still no need to copy it */ |
2644 | } | |
2645 | if (!PageUptodate(page)) | |
2646 | return page; /* let do_swap_page report the error */ | |
2647 | ||
5ad64688 | 2648 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
8f425e4e MWO |
2649 | if (new_page && |
2650 | mem_cgroup_charge(page_folio(new_page), vma->vm_mm, GFP_KERNEL)) { | |
62fdb163 HD |
2651 | put_page(new_page); |
2652 | new_page = NULL; | |
2653 | } | |
5ad64688 | 2654 | if (new_page) { |
6b970599 KW |
2655 | if (copy_mc_user_highpage(new_page, page, address, vma)) { |
2656 | put_page(new_page); | |
2657 | memory_failure_queue(page_to_pfn(page), 0); | |
2658 | return ERR_PTR(-EHWPOISON); | |
2659 | } | |
5ad64688 HD |
2660 | SetPageDirty(new_page); |
2661 | __SetPageUptodate(new_page); | |
48c935ad | 2662 | __SetPageLocked(new_page); |
4d45c3af YY |
2663 | #ifdef CONFIG_SWAP |
2664 | count_vm_event(KSM_SWPIN_COPY); | |
2665 | #endif | |
5ad64688 HD |
2666 | } |
2667 | ||
5ad64688 HD |
2668 | return new_page; |
2669 | } | |
2670 | ||
6d4675e6 | 2671 | void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc) |
e9995ef9 | 2672 | { |
21fbd591 QZ |
2673 | struct ksm_stable_node *stable_node; |
2674 | struct ksm_rmap_item *rmap_item; | |
e9995ef9 HD |
2675 | int search_new_forks = 0; |
2676 | ||
2f031c6f | 2677 | VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio); |
9f32624b JK |
2678 | |
2679 | /* | |
2680 | * Rely on the page lock to protect against concurrent modifications | |
2681 | * to that page's node of the stable tree. | |
2682 | */ | |
2f031c6f | 2683 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
e9995ef9 | 2684 | |
2f031c6f | 2685 | stable_node = folio_stable_node(folio); |
e9995ef9 | 2686 | if (!stable_node) |
1df631ae | 2687 | return; |
e9995ef9 | 2688 | again: |
b67bfe0d | 2689 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
e9995ef9 | 2690 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
5beb4930 | 2691 | struct anon_vma_chain *vmac; |
e9995ef9 HD |
2692 | struct vm_area_struct *vma; |
2693 | ||
ad12695f | 2694 | cond_resched(); |
6d4675e6 MK |
2695 | if (!anon_vma_trylock_read(anon_vma)) { |
2696 | if (rwc->try_lock) { | |
2697 | rwc->contended = true; | |
2698 | return; | |
2699 | } | |
2700 | anon_vma_lock_read(anon_vma); | |
2701 | } | |
bf181b9f ML |
2702 | anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, |
2703 | 0, ULONG_MAX) { | |
1105a2fc JH |
2704 | unsigned long addr; |
2705 | ||
ad12695f | 2706 | cond_resched(); |
5beb4930 | 2707 | vma = vmac->vma; |
1105a2fc JH |
2708 | |
2709 | /* Ignore the stable/unstable/sqnr flags */ | |
cd7fae26 | 2710 | addr = rmap_item->address & PAGE_MASK; |
1105a2fc JH |
2711 | |
2712 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
e9995ef9 HD |
2713 | continue; |
2714 | /* | |
2715 | * Initially we examine only the vma which covers this | |
2716 | * rmap_item; but later, if there is still work to do, | |
2717 | * we examine covering vmas in other mms: in case they | |
2718 | * were forked from the original since ksmd passed. | |
2719 | */ | |
2720 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
2721 | continue; | |
2722 | ||
0dd1c7bb JK |
2723 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
2724 | continue; | |
2725 | ||
2f031c6f | 2726 | if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) { |
b6b19f25 | 2727 | anon_vma_unlock_read(anon_vma); |
1df631ae | 2728 | return; |
e9995ef9 | 2729 | } |
2f031c6f | 2730 | if (rwc->done && rwc->done(folio)) { |
0dd1c7bb | 2731 | anon_vma_unlock_read(anon_vma); |
1df631ae | 2732 | return; |
0dd1c7bb | 2733 | } |
e9995ef9 | 2734 | } |
b6b19f25 | 2735 | anon_vma_unlock_read(anon_vma); |
e9995ef9 HD |
2736 | } |
2737 | if (!search_new_forks++) | |
2738 | goto again; | |
e9995ef9 HD |
2739 | } |
2740 | ||
52629506 | 2741 | #ifdef CONFIG_MIGRATION |
19138349 | 2742 | void folio_migrate_ksm(struct folio *newfolio, struct folio *folio) |
e9995ef9 | 2743 | { |
21fbd591 | 2744 | struct ksm_stable_node *stable_node; |
e9995ef9 | 2745 | |
19138349 MWO |
2746 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
2747 | VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio); | |
2748 | VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio); | |
e9995ef9 | 2749 | |
19138349 | 2750 | stable_node = folio_stable_node(folio); |
e9995ef9 | 2751 | if (stable_node) { |
19138349 MWO |
2752 | VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio); |
2753 | stable_node->kpfn = folio_pfn(newfolio); | |
c8d6553b | 2754 | /* |
19138349 | 2755 | * newfolio->mapping was set in advance; now we need smp_wmb() |
c8d6553b | 2756 | * to make sure that the new stable_node->kpfn is visible |
19138349 MWO |
2757 | * to get_ksm_page() before it can see that folio->mapping |
2758 | * has gone stale (or that folio_test_swapcache has been cleared). | |
c8d6553b HD |
2759 | */ |
2760 | smp_wmb(); | |
19138349 | 2761 | set_page_stable_node(&folio->page, NULL); |
e9995ef9 HD |
2762 | } |
2763 | } | |
2764 | #endif /* CONFIG_MIGRATION */ | |
2765 | ||
62b61f61 | 2766 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 HD |
2767 | static void wait_while_offlining(void) |
2768 | { | |
2769 | while (ksm_run & KSM_RUN_OFFLINE) { | |
2770 | mutex_unlock(&ksm_thread_mutex); | |
2771 | wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), | |
74316201 | 2772 | TASK_UNINTERRUPTIBLE); |
ef4d43a8 HD |
2773 | mutex_lock(&ksm_thread_mutex); |
2774 | } | |
2775 | } | |
2776 | ||
21fbd591 | 2777 | static bool stable_node_dup_remove_range(struct ksm_stable_node *stable_node, |
2c653d0e AA |
2778 | unsigned long start_pfn, |
2779 | unsigned long end_pfn) | |
2780 | { | |
2781 | if (stable_node->kpfn >= start_pfn && | |
2782 | stable_node->kpfn < end_pfn) { | |
2783 | /* | |
2784 | * Don't get_ksm_page, page has already gone: | |
2785 | * which is why we keep kpfn instead of page* | |
2786 | */ | |
2787 | remove_node_from_stable_tree(stable_node); | |
2788 | return true; | |
2789 | } | |
2790 | return false; | |
2791 | } | |
2792 | ||
21fbd591 | 2793 | static bool stable_node_chain_remove_range(struct ksm_stable_node *stable_node, |
2c653d0e AA |
2794 | unsigned long start_pfn, |
2795 | unsigned long end_pfn, | |
2796 | struct rb_root *root) | |
2797 | { | |
21fbd591 | 2798 | struct ksm_stable_node *dup; |
2c653d0e AA |
2799 | struct hlist_node *hlist_safe; |
2800 | ||
2801 | if (!is_stable_node_chain(stable_node)) { | |
2802 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
2803 | return stable_node_dup_remove_range(stable_node, start_pfn, | |
2804 | end_pfn); | |
2805 | } | |
2806 | ||
2807 | hlist_for_each_entry_safe(dup, hlist_safe, | |
2808 | &stable_node->hlist, hlist_dup) { | |
2809 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
2810 | stable_node_dup_remove_range(dup, start_pfn, end_pfn); | |
2811 | } | |
2812 | if (hlist_empty(&stable_node->hlist)) { | |
2813 | free_stable_node_chain(stable_node, root); | |
2814 | return true; /* notify caller that tree was rebalanced */ | |
2815 | } else | |
2816 | return false; | |
2817 | } | |
2818 | ||
ee0ea59c HD |
2819 | static void ksm_check_stable_tree(unsigned long start_pfn, |
2820 | unsigned long end_pfn) | |
62b61f61 | 2821 | { |
21fbd591 | 2822 | struct ksm_stable_node *stable_node, *next; |
62b61f61 | 2823 | struct rb_node *node; |
90bd6fd3 | 2824 | int nid; |
62b61f61 | 2825 | |
ef53d16c HD |
2826 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
2827 | node = rb_first(root_stable_tree + nid); | |
ee0ea59c | 2828 | while (node) { |
21fbd591 | 2829 | stable_node = rb_entry(node, struct ksm_stable_node, node); |
2c653d0e AA |
2830 | if (stable_node_chain_remove_range(stable_node, |
2831 | start_pfn, end_pfn, | |
2832 | root_stable_tree + | |
2833 | nid)) | |
ef53d16c | 2834 | node = rb_first(root_stable_tree + nid); |
2c653d0e | 2835 | else |
ee0ea59c HD |
2836 | node = rb_next(node); |
2837 | cond_resched(); | |
90bd6fd3 | 2838 | } |
ee0ea59c | 2839 | } |
03640418 | 2840 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
2841 | if (stable_node->kpfn >= start_pfn && |
2842 | stable_node->kpfn < end_pfn) | |
2843 | remove_node_from_stable_tree(stable_node); | |
2844 | cond_resched(); | |
2845 | } | |
62b61f61 HD |
2846 | } |
2847 | ||
2848 | static int ksm_memory_callback(struct notifier_block *self, | |
2849 | unsigned long action, void *arg) | |
2850 | { | |
2851 | struct memory_notify *mn = arg; | |
62b61f61 HD |
2852 | |
2853 | switch (action) { | |
2854 | case MEM_GOING_OFFLINE: | |
2855 | /* | |
ef4d43a8 HD |
2856 | * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() |
2857 | * and remove_all_stable_nodes() while memory is going offline: | |
2858 | * it is unsafe for them to touch the stable tree at this time. | |
2859 | * But unmerge_ksm_pages(), rmap lookups and other entry points | |
2860 | * which do not need the ksm_thread_mutex are all safe. | |
62b61f61 | 2861 | */ |
ef4d43a8 HD |
2862 | mutex_lock(&ksm_thread_mutex); |
2863 | ksm_run |= KSM_RUN_OFFLINE; | |
2864 | mutex_unlock(&ksm_thread_mutex); | |
62b61f61 HD |
2865 | break; |
2866 | ||
2867 | case MEM_OFFLINE: | |
2868 | /* | |
2869 | * Most of the work is done by page migration; but there might | |
2870 | * be a few stable_nodes left over, still pointing to struct | |
ee0ea59c HD |
2871 | * pages which have been offlined: prune those from the tree, |
2872 | * otherwise get_ksm_page() might later try to access a | |
2873 | * non-existent struct page. | |
62b61f61 | 2874 | */ |
ee0ea59c HD |
2875 | ksm_check_stable_tree(mn->start_pfn, |
2876 | mn->start_pfn + mn->nr_pages); | |
e4a9bc58 | 2877 | fallthrough; |
62b61f61 | 2878 | case MEM_CANCEL_OFFLINE: |
ef4d43a8 HD |
2879 | mutex_lock(&ksm_thread_mutex); |
2880 | ksm_run &= ~KSM_RUN_OFFLINE; | |
62b61f61 | 2881 | mutex_unlock(&ksm_thread_mutex); |
ef4d43a8 HD |
2882 | |
2883 | smp_mb(); /* wake_up_bit advises this */ | |
2884 | wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); | |
62b61f61 HD |
2885 | break; |
2886 | } | |
2887 | return NOTIFY_OK; | |
2888 | } | |
ef4d43a8 HD |
2889 | #else |
2890 | static void wait_while_offlining(void) | |
2891 | { | |
2892 | } | |
62b61f61 HD |
2893 | #endif /* CONFIG_MEMORY_HOTREMOVE */ |
2894 | ||
2ffd8679 HD |
2895 | #ifdef CONFIG_SYSFS |
2896 | /* | |
2897 | * This all compiles without CONFIG_SYSFS, but is a waste of space. | |
2898 | */ | |
2899 | ||
31dbd01f IE |
2900 | #define KSM_ATTR_RO(_name) \ |
2901 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2902 | #define KSM_ATTR(_name) \ | |
1bad2e5c | 2903 | static struct kobj_attribute _name##_attr = __ATTR_RW(_name) |
31dbd01f IE |
2904 | |
2905 | static ssize_t sleep_millisecs_show(struct kobject *kobj, | |
2906 | struct kobj_attribute *attr, char *buf) | |
2907 | { | |
ae7a927d | 2908 | return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs); |
31dbd01f IE |
2909 | } |
2910 | ||
2911 | static ssize_t sleep_millisecs_store(struct kobject *kobj, | |
2912 | struct kobj_attribute *attr, | |
2913 | const char *buf, size_t count) | |
2914 | { | |
dfefd226 | 2915 | unsigned int msecs; |
31dbd01f IE |
2916 | int err; |
2917 | ||
dfefd226 AD |
2918 | err = kstrtouint(buf, 10, &msecs); |
2919 | if (err) | |
31dbd01f IE |
2920 | return -EINVAL; |
2921 | ||
2922 | ksm_thread_sleep_millisecs = msecs; | |
fcf9a0ef | 2923 | wake_up_interruptible(&ksm_iter_wait); |
31dbd01f IE |
2924 | |
2925 | return count; | |
2926 | } | |
2927 | KSM_ATTR(sleep_millisecs); | |
2928 | ||
2929 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
2930 | struct kobj_attribute *attr, char *buf) | |
2931 | { | |
ae7a927d | 2932 | return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan); |
31dbd01f IE |
2933 | } |
2934 | ||
2935 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
2936 | struct kobj_attribute *attr, | |
2937 | const char *buf, size_t count) | |
2938 | { | |
dfefd226 | 2939 | unsigned int nr_pages; |
31dbd01f | 2940 | int err; |
31dbd01f | 2941 | |
dfefd226 AD |
2942 | err = kstrtouint(buf, 10, &nr_pages); |
2943 | if (err) | |
31dbd01f IE |
2944 | return -EINVAL; |
2945 | ||
2946 | ksm_thread_pages_to_scan = nr_pages; | |
2947 | ||
2948 | return count; | |
2949 | } | |
2950 | KSM_ATTR(pages_to_scan); | |
2951 | ||
2952 | static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, | |
2953 | char *buf) | |
2954 | { | |
ae7a927d | 2955 | return sysfs_emit(buf, "%lu\n", ksm_run); |
31dbd01f IE |
2956 | } |
2957 | ||
2958 | static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, | |
2959 | const char *buf, size_t count) | |
2960 | { | |
dfefd226 | 2961 | unsigned int flags; |
31dbd01f | 2962 | int err; |
31dbd01f | 2963 | |
dfefd226 AD |
2964 | err = kstrtouint(buf, 10, &flags); |
2965 | if (err) | |
31dbd01f IE |
2966 | return -EINVAL; |
2967 | if (flags > KSM_RUN_UNMERGE) | |
2968 | return -EINVAL; | |
2969 | ||
2970 | /* | |
2971 | * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. | |
2972 | * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, | |
d0f209f6 HD |
2973 | * breaking COW to free the pages_shared (but leaves mm_slots |
2974 | * on the list for when ksmd may be set running again). | |
31dbd01f IE |
2975 | */ |
2976 | ||
2977 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 2978 | wait_while_offlining(); |
31dbd01f IE |
2979 | if (ksm_run != flags) { |
2980 | ksm_run = flags; | |
d952b791 | 2981 | if (flags & KSM_RUN_UNMERGE) { |
e1e12d2f | 2982 | set_current_oom_origin(); |
d952b791 | 2983 | err = unmerge_and_remove_all_rmap_items(); |
e1e12d2f | 2984 | clear_current_oom_origin(); |
d952b791 HD |
2985 | if (err) { |
2986 | ksm_run = KSM_RUN_STOP; | |
2987 | count = err; | |
2988 | } | |
2989 | } | |
31dbd01f IE |
2990 | } |
2991 | mutex_unlock(&ksm_thread_mutex); | |
2992 | ||
2993 | if (flags & KSM_RUN_MERGE) | |
2994 | wake_up_interruptible(&ksm_thread_wait); | |
2995 | ||
2996 | return count; | |
2997 | } | |
2998 | KSM_ATTR(run); | |
2999 | ||
90bd6fd3 PH |
3000 | #ifdef CONFIG_NUMA |
3001 | static ssize_t merge_across_nodes_show(struct kobject *kobj, | |
ae7a927d | 3002 | struct kobj_attribute *attr, char *buf) |
90bd6fd3 | 3003 | { |
ae7a927d | 3004 | return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes); |
90bd6fd3 PH |
3005 | } |
3006 | ||
3007 | static ssize_t merge_across_nodes_store(struct kobject *kobj, | |
3008 | struct kobj_attribute *attr, | |
3009 | const char *buf, size_t count) | |
3010 | { | |
3011 | int err; | |
3012 | unsigned long knob; | |
3013 | ||
3014 | err = kstrtoul(buf, 10, &knob); | |
3015 | if (err) | |
3016 | return err; | |
3017 | if (knob > 1) | |
3018 | return -EINVAL; | |
3019 | ||
3020 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 3021 | wait_while_offlining(); |
90bd6fd3 | 3022 | if (ksm_merge_across_nodes != knob) { |
cbf86cfe | 3023 | if (ksm_pages_shared || remove_all_stable_nodes()) |
90bd6fd3 | 3024 | err = -EBUSY; |
ef53d16c HD |
3025 | else if (root_stable_tree == one_stable_tree) { |
3026 | struct rb_root *buf; | |
3027 | /* | |
3028 | * This is the first time that we switch away from the | |
3029 | * default of merging across nodes: must now allocate | |
3030 | * a buffer to hold as many roots as may be needed. | |
3031 | * Allocate stable and unstable together: | |
3032 | * MAXSMP NODES_SHIFT 10 will use 16kB. | |
3033 | */ | |
bafe1e14 JP |
3034 | buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), |
3035 | GFP_KERNEL); | |
ef53d16c HD |
3036 | /* Let us assume that RB_ROOT is NULL is zero */ |
3037 | if (!buf) | |
3038 | err = -ENOMEM; | |
3039 | else { | |
3040 | root_stable_tree = buf; | |
3041 | root_unstable_tree = buf + nr_node_ids; | |
3042 | /* Stable tree is empty but not the unstable */ | |
3043 | root_unstable_tree[0] = one_unstable_tree[0]; | |
3044 | } | |
3045 | } | |
3046 | if (!err) { | |
90bd6fd3 | 3047 | ksm_merge_across_nodes = knob; |
ef53d16c HD |
3048 | ksm_nr_node_ids = knob ? 1 : nr_node_ids; |
3049 | } | |
90bd6fd3 PH |
3050 | } |
3051 | mutex_unlock(&ksm_thread_mutex); | |
3052 | ||
3053 | return err ? err : count; | |
3054 | } | |
3055 | KSM_ATTR(merge_across_nodes); | |
3056 | #endif | |
3057 | ||
e86c59b1 | 3058 | static ssize_t use_zero_pages_show(struct kobject *kobj, |
ae7a927d | 3059 | struct kobj_attribute *attr, char *buf) |
e86c59b1 | 3060 | { |
ae7a927d | 3061 | return sysfs_emit(buf, "%u\n", ksm_use_zero_pages); |
e86c59b1 CI |
3062 | } |
3063 | static ssize_t use_zero_pages_store(struct kobject *kobj, | |
3064 | struct kobj_attribute *attr, | |
3065 | const char *buf, size_t count) | |
3066 | { | |
3067 | int err; | |
3068 | bool value; | |
3069 | ||
3070 | err = kstrtobool(buf, &value); | |
3071 | if (err) | |
3072 | return -EINVAL; | |
3073 | ||
3074 | ksm_use_zero_pages = value; | |
3075 | ||
3076 | return count; | |
3077 | } | |
3078 | KSM_ATTR(use_zero_pages); | |
3079 | ||
2c653d0e AA |
3080 | static ssize_t max_page_sharing_show(struct kobject *kobj, |
3081 | struct kobj_attribute *attr, char *buf) | |
3082 | { | |
ae7a927d | 3083 | return sysfs_emit(buf, "%u\n", ksm_max_page_sharing); |
2c653d0e AA |
3084 | } |
3085 | ||
3086 | static ssize_t max_page_sharing_store(struct kobject *kobj, | |
3087 | struct kobj_attribute *attr, | |
3088 | const char *buf, size_t count) | |
3089 | { | |
3090 | int err; | |
3091 | int knob; | |
3092 | ||
3093 | err = kstrtoint(buf, 10, &knob); | |
3094 | if (err) | |
3095 | return err; | |
3096 | /* | |
3097 | * When a KSM page is created it is shared by 2 mappings. This | |
3098 | * being a signed comparison, it implicitly verifies it's not | |
3099 | * negative. | |
3100 | */ | |
3101 | if (knob < 2) | |
3102 | return -EINVAL; | |
3103 | ||
3104 | if (READ_ONCE(ksm_max_page_sharing) == knob) | |
3105 | return count; | |
3106 | ||
3107 | mutex_lock(&ksm_thread_mutex); | |
3108 | wait_while_offlining(); | |
3109 | if (ksm_max_page_sharing != knob) { | |
3110 | if (ksm_pages_shared || remove_all_stable_nodes()) | |
3111 | err = -EBUSY; | |
3112 | else | |
3113 | ksm_max_page_sharing = knob; | |
3114 | } | |
3115 | mutex_unlock(&ksm_thread_mutex); | |
3116 | ||
3117 | return err ? err : count; | |
3118 | } | |
3119 | KSM_ATTR(max_page_sharing); | |
3120 | ||
b4028260 HD |
3121 | static ssize_t pages_shared_show(struct kobject *kobj, |
3122 | struct kobj_attribute *attr, char *buf) | |
3123 | { | |
ae7a927d | 3124 | return sysfs_emit(buf, "%lu\n", ksm_pages_shared); |
b4028260 HD |
3125 | } |
3126 | KSM_ATTR_RO(pages_shared); | |
3127 | ||
3128 | static ssize_t pages_sharing_show(struct kobject *kobj, | |
3129 | struct kobj_attribute *attr, char *buf) | |
3130 | { | |
ae7a927d | 3131 | return sysfs_emit(buf, "%lu\n", ksm_pages_sharing); |
b4028260 HD |
3132 | } |
3133 | KSM_ATTR_RO(pages_sharing); | |
3134 | ||
473b0ce4 HD |
3135 | static ssize_t pages_unshared_show(struct kobject *kobj, |
3136 | struct kobj_attribute *attr, char *buf) | |
3137 | { | |
ae7a927d | 3138 | return sysfs_emit(buf, "%lu\n", ksm_pages_unshared); |
473b0ce4 HD |
3139 | } |
3140 | KSM_ATTR_RO(pages_unshared); | |
3141 | ||
3142 | static ssize_t pages_volatile_show(struct kobject *kobj, | |
3143 | struct kobj_attribute *attr, char *buf) | |
3144 | { | |
3145 | long ksm_pages_volatile; | |
3146 | ||
3147 | ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared | |
3148 | - ksm_pages_sharing - ksm_pages_unshared; | |
3149 | /* | |
3150 | * It was not worth any locking to calculate that statistic, | |
3151 | * but it might therefore sometimes be negative: conceal that. | |
3152 | */ | |
3153 | if (ksm_pages_volatile < 0) | |
3154 | ksm_pages_volatile = 0; | |
ae7a927d | 3155 | return sysfs_emit(buf, "%ld\n", ksm_pages_volatile); |
473b0ce4 HD |
3156 | } |
3157 | KSM_ATTR_RO(pages_volatile); | |
3158 | ||
2c653d0e AA |
3159 | static ssize_t stable_node_dups_show(struct kobject *kobj, |
3160 | struct kobj_attribute *attr, char *buf) | |
3161 | { | |
ae7a927d | 3162 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups); |
2c653d0e AA |
3163 | } |
3164 | KSM_ATTR_RO(stable_node_dups); | |
3165 | ||
3166 | static ssize_t stable_node_chains_show(struct kobject *kobj, | |
3167 | struct kobj_attribute *attr, char *buf) | |
3168 | { | |
ae7a927d | 3169 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains); |
2c653d0e AA |
3170 | } |
3171 | KSM_ATTR_RO(stable_node_chains); | |
3172 | ||
3173 | static ssize_t | |
3174 | stable_node_chains_prune_millisecs_show(struct kobject *kobj, | |
3175 | struct kobj_attribute *attr, | |
3176 | char *buf) | |
3177 | { | |
ae7a927d | 3178 | return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs); |
2c653d0e AA |
3179 | } |
3180 | ||
3181 | static ssize_t | |
3182 | stable_node_chains_prune_millisecs_store(struct kobject *kobj, | |
3183 | struct kobj_attribute *attr, | |
3184 | const char *buf, size_t count) | |
3185 | { | |
584ff0df | 3186 | unsigned int msecs; |
2c653d0e AA |
3187 | int err; |
3188 | ||
584ff0df ZB |
3189 | err = kstrtouint(buf, 10, &msecs); |
3190 | if (err) | |
2c653d0e AA |
3191 | return -EINVAL; |
3192 | ||
3193 | ksm_stable_node_chains_prune_millisecs = msecs; | |
3194 | ||
3195 | return count; | |
3196 | } | |
3197 | KSM_ATTR(stable_node_chains_prune_millisecs); | |
3198 | ||
473b0ce4 HD |
3199 | static ssize_t full_scans_show(struct kobject *kobj, |
3200 | struct kobj_attribute *attr, char *buf) | |
3201 | { | |
ae7a927d | 3202 | return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr); |
473b0ce4 HD |
3203 | } |
3204 | KSM_ATTR_RO(full_scans); | |
3205 | ||
31dbd01f IE |
3206 | static struct attribute *ksm_attrs[] = { |
3207 | &sleep_millisecs_attr.attr, | |
3208 | &pages_to_scan_attr.attr, | |
3209 | &run_attr.attr, | |
b4028260 HD |
3210 | &pages_shared_attr.attr, |
3211 | &pages_sharing_attr.attr, | |
473b0ce4 HD |
3212 | &pages_unshared_attr.attr, |
3213 | &pages_volatile_attr.attr, | |
3214 | &full_scans_attr.attr, | |
90bd6fd3 PH |
3215 | #ifdef CONFIG_NUMA |
3216 | &merge_across_nodes_attr.attr, | |
3217 | #endif | |
2c653d0e AA |
3218 | &max_page_sharing_attr.attr, |
3219 | &stable_node_chains_attr.attr, | |
3220 | &stable_node_dups_attr.attr, | |
3221 | &stable_node_chains_prune_millisecs_attr.attr, | |
e86c59b1 | 3222 | &use_zero_pages_attr.attr, |
31dbd01f IE |
3223 | NULL, |
3224 | }; | |
3225 | ||
f907c26a | 3226 | static const struct attribute_group ksm_attr_group = { |
31dbd01f IE |
3227 | .attrs = ksm_attrs, |
3228 | .name = "ksm", | |
3229 | }; | |
2ffd8679 | 3230 | #endif /* CONFIG_SYSFS */ |
31dbd01f IE |
3231 | |
3232 | static int __init ksm_init(void) | |
3233 | { | |
3234 | struct task_struct *ksm_thread; | |
3235 | int err; | |
3236 | ||
e86c59b1 CI |
3237 | /* The correct value depends on page size and endianness */ |
3238 | zero_checksum = calc_checksum(ZERO_PAGE(0)); | |
3239 | /* Default to false for backwards compatibility */ | |
3240 | ksm_use_zero_pages = false; | |
3241 | ||
31dbd01f IE |
3242 | err = ksm_slab_init(); |
3243 | if (err) | |
3244 | goto out; | |
3245 | ||
31dbd01f IE |
3246 | ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); |
3247 | if (IS_ERR(ksm_thread)) { | |
25acde31 | 3248 | pr_err("ksm: creating kthread failed\n"); |
31dbd01f | 3249 | err = PTR_ERR(ksm_thread); |
d9f8984c | 3250 | goto out_free; |
31dbd01f IE |
3251 | } |
3252 | ||
2ffd8679 | 3253 | #ifdef CONFIG_SYSFS |
31dbd01f IE |
3254 | err = sysfs_create_group(mm_kobj, &ksm_attr_group); |
3255 | if (err) { | |
25acde31 | 3256 | pr_err("ksm: register sysfs failed\n"); |
2ffd8679 | 3257 | kthread_stop(ksm_thread); |
d9f8984c | 3258 | goto out_free; |
31dbd01f | 3259 | } |
c73602ad HD |
3260 | #else |
3261 | ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ | |
3262 | ||
2ffd8679 | 3263 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 3264 | |
62b61f61 | 3265 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 | 3266 | /* There is no significance to this priority 100 */ |
1eeaa4fd | 3267 | hotplug_memory_notifier(ksm_memory_callback, KSM_CALLBACK_PRI); |
62b61f61 | 3268 | #endif |
31dbd01f IE |
3269 | return 0; |
3270 | ||
d9f8984c | 3271 | out_free: |
31dbd01f IE |
3272 | ksm_slab_free(); |
3273 | out: | |
3274 | return err; | |
f8af4da3 | 3275 | } |
a64fb3cd | 3276 | subsys_initcall(ksm_init); |