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> |
f8af4da3 | 42 | |
31dbd01f | 43 | #include <asm/tlbflush.h> |
73848b46 | 44 | #include "internal.h" |
31dbd01f | 45 | |
e850dcf5 HD |
46 | #ifdef CONFIG_NUMA |
47 | #define NUMA(x) (x) | |
48 | #define DO_NUMA(x) do { (x); } while (0) | |
49 | #else | |
50 | #define NUMA(x) (0) | |
51 | #define DO_NUMA(x) do { } while (0) | |
52 | #endif | |
53 | ||
5a2ca3ef MR |
54 | /** |
55 | * DOC: Overview | |
56 | * | |
31dbd01f IE |
57 | * A few notes about the KSM scanning process, |
58 | * to make it easier to understand the data structures below: | |
59 | * | |
60 | * In order to reduce excessive scanning, KSM sorts the memory pages by their | |
61 | * contents into a data structure that holds pointers to the pages' locations. | |
62 | * | |
63 | * Since the contents of the pages may change at any moment, KSM cannot just | |
64 | * insert the pages into a normal sorted tree and expect it to find anything. | |
65 | * Therefore KSM uses two data structures - the stable and the unstable tree. | |
66 | * | |
67 | * The stable tree holds pointers to all the merged pages (ksm pages), sorted | |
68 | * by their contents. Because each such page is write-protected, searching on | |
69 | * this tree is fully assured to be working (except when pages are unmapped), | |
70 | * and therefore this tree is called the stable tree. | |
71 | * | |
5a2ca3ef MR |
72 | * The stable tree node includes information required for reverse |
73 | * mapping from a KSM page to virtual addresses that map this page. | |
74 | * | |
75 | * In order to avoid large latencies of the rmap walks on KSM pages, | |
76 | * KSM maintains two types of nodes in the stable tree: | |
77 | * | |
78 | * * the regular nodes that keep the reverse mapping structures in a | |
79 | * linked list | |
80 | * * the "chains" that link nodes ("dups") that represent the same | |
81 | * write protected memory content, but each "dup" corresponds to a | |
82 | * different KSM page copy of that content | |
83 | * | |
84 | * Internally, the regular nodes, "dups" and "chains" are represented | |
21fbd591 | 85 | * using the same struct ksm_stable_node structure. |
5a2ca3ef | 86 | * |
31dbd01f IE |
87 | * In addition to the stable tree, KSM uses a second data structure called the |
88 | * unstable tree: this tree holds pointers to pages which have been found to | |
89 | * be "unchanged for a period of time". The unstable tree sorts these pages | |
90 | * by their contents, but since they are not write-protected, KSM cannot rely | |
91 | * upon the unstable tree to work correctly - the unstable tree is liable to | |
92 | * be corrupted as its contents are modified, and so it is called unstable. | |
93 | * | |
94 | * KSM solves this problem by several techniques: | |
95 | * | |
96 | * 1) The unstable tree is flushed every time KSM completes scanning all | |
97 | * memory areas, and then the tree is rebuilt again from the beginning. | |
98 | * 2) KSM will only insert into the unstable tree, pages whose hash value | |
99 | * has not changed since the previous scan of all memory areas. | |
100 | * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the | |
101 | * colors of the nodes and not on their contents, assuring that even when | |
102 | * the tree gets "corrupted" it won't get out of balance, so scanning time | |
103 | * remains the same (also, searching and inserting nodes in an rbtree uses | |
104 | * the same algorithm, so we have no overhead when we flush and rebuild). | |
105 | * 4) KSM never flushes the stable tree, which means that even if it were to | |
106 | * take 10 attempts to find a page in the unstable tree, once it is found, | |
107 | * it is secured in the stable tree. (When we scan a new page, we first | |
108 | * compare it against the stable tree, and then against the unstable tree.) | |
8fdb3dbf HD |
109 | * |
110 | * If the merge_across_nodes tunable is unset, then KSM maintains multiple | |
111 | * stable trees and multiple unstable trees: one of each for each NUMA node. | |
31dbd01f IE |
112 | */ |
113 | ||
114 | /** | |
21fbd591 | 115 | * struct ksm_mm_slot - ksm information per mm that is being scanned |
79b09941 | 116 | * @hash: link to the mm_slots hash list |
23f746e4 | 117 | * @mm_node: link into the mm_slots list, rooted in ksm_mm_head |
6514d511 | 118 | * @rmap_list: head for this mm_slot's singly-linked list of rmap_items |
31dbd01f IE |
119 | * @mm: the mm that this information is valid for |
120 | */ | |
21fbd591 | 121 | struct ksm_mm_slot { |
79b09941 | 122 | struct hlist_node hash; |
23f746e4 | 123 | struct list_head mm_node; |
21fbd591 | 124 | struct ksm_rmap_item *rmap_list; |
31dbd01f IE |
125 | struct mm_struct *mm; |
126 | }; | |
127 | ||
128 | /** | |
129 | * struct ksm_scan - cursor for scanning | |
130 | * @mm_slot: the current mm_slot we are scanning | |
131 | * @address: the next address inside that to be scanned | |
6514d511 | 132 | * @rmap_list: link to the next rmap to be scanned in the rmap_list |
31dbd01f IE |
133 | * @seqnr: count of completed full scans (needed when removing unstable node) |
134 | * | |
135 | * There is only the one ksm_scan instance of this cursor structure. | |
136 | */ | |
137 | struct ksm_scan { | |
21fbd591 | 138 | struct ksm_mm_slot *mm_slot; |
31dbd01f | 139 | unsigned long address; |
21fbd591 | 140 | struct ksm_rmap_item **rmap_list; |
31dbd01f IE |
141 | unsigned long seqnr; |
142 | }; | |
143 | ||
7b6ba2c7 | 144 | /** |
21fbd591 | 145 | * struct ksm_stable_node - node of the stable rbtree |
7b6ba2c7 | 146 | * @node: rb node of this ksm page in the stable tree |
4146d2d6 | 147 | * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list |
2c653d0e | 148 | * @hlist_dup: linked into the stable_node->hlist with a stable_node chain |
4146d2d6 | 149 | * @list: linked into migrate_nodes, pending placement in the proper node tree |
7b6ba2c7 | 150 | * @hlist: hlist head of rmap_items using this ksm page |
4146d2d6 | 151 | * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) |
2c653d0e AA |
152 | * @chain_prune_time: time of the last full garbage collection |
153 | * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN | |
4146d2d6 | 154 | * @nid: NUMA node id of stable tree in which linked (may not match kpfn) |
7b6ba2c7 | 155 | */ |
21fbd591 | 156 | struct ksm_stable_node { |
4146d2d6 HD |
157 | union { |
158 | struct rb_node node; /* when node of stable tree */ | |
159 | struct { /* when listed for migration */ | |
160 | struct list_head *head; | |
2c653d0e AA |
161 | struct { |
162 | struct hlist_node hlist_dup; | |
163 | struct list_head list; | |
164 | }; | |
4146d2d6 HD |
165 | }; |
166 | }; | |
7b6ba2c7 | 167 | struct hlist_head hlist; |
2c653d0e AA |
168 | union { |
169 | unsigned long kpfn; | |
170 | unsigned long chain_prune_time; | |
171 | }; | |
172 | /* | |
173 | * STABLE_NODE_CHAIN can be any negative number in | |
174 | * rmap_hlist_len negative range, but better not -1 to be able | |
175 | * to reliably detect underflows. | |
176 | */ | |
177 | #define STABLE_NODE_CHAIN -1024 | |
178 | int rmap_hlist_len; | |
4146d2d6 HD |
179 | #ifdef CONFIG_NUMA |
180 | int nid; | |
181 | #endif | |
7b6ba2c7 HD |
182 | }; |
183 | ||
31dbd01f | 184 | /** |
21fbd591 | 185 | * struct ksm_rmap_item - reverse mapping item for virtual addresses |
6514d511 | 186 | * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list |
db114b83 | 187 | * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree |
bc56620b | 188 | * @nid: NUMA node id of unstable tree in which linked (may not match page) |
31dbd01f IE |
189 | * @mm: the memory structure this rmap_item is pointing into |
190 | * @address: the virtual address this rmap_item tracks (+ flags in low bits) | |
191 | * @oldchecksum: previous checksum of the page at that virtual address | |
7b6ba2c7 HD |
192 | * @node: rb node of this rmap_item in the unstable tree |
193 | * @head: pointer to stable_node heading this list in the stable tree | |
194 | * @hlist: link into hlist of rmap_items hanging off that stable_node | |
31dbd01f | 195 | */ |
21fbd591 QZ |
196 | struct ksm_rmap_item { |
197 | struct ksm_rmap_item *rmap_list; | |
bc56620b HD |
198 | union { |
199 | struct anon_vma *anon_vma; /* when stable */ | |
200 | #ifdef CONFIG_NUMA | |
201 | int nid; /* when node of unstable tree */ | |
202 | #endif | |
203 | }; | |
31dbd01f IE |
204 | struct mm_struct *mm; |
205 | unsigned long address; /* + low bits used for flags below */ | |
7b6ba2c7 | 206 | unsigned int oldchecksum; /* when unstable */ |
31dbd01f | 207 | union { |
7b6ba2c7 HD |
208 | struct rb_node node; /* when node of unstable tree */ |
209 | struct { /* when listed from stable tree */ | |
21fbd591 | 210 | struct ksm_stable_node *head; |
7b6ba2c7 HD |
211 | struct hlist_node hlist; |
212 | }; | |
31dbd01f IE |
213 | }; |
214 | }; | |
215 | ||
216 | #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ | |
7b6ba2c7 HD |
217 | #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ |
218 | #define STABLE_FLAG 0x200 /* is listed from the stable tree */ | |
31dbd01f IE |
219 | |
220 | /* The stable and unstable tree heads */ | |
ef53d16c HD |
221 | static struct rb_root one_stable_tree[1] = { RB_ROOT }; |
222 | static struct rb_root one_unstable_tree[1] = { RB_ROOT }; | |
223 | static struct rb_root *root_stable_tree = one_stable_tree; | |
224 | static struct rb_root *root_unstable_tree = one_unstable_tree; | |
31dbd01f | 225 | |
4146d2d6 HD |
226 | /* Recently migrated nodes of stable tree, pending proper placement */ |
227 | static LIST_HEAD(migrate_nodes); | |
2c653d0e | 228 | #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev) |
4146d2d6 | 229 | |
4ca3a69b SL |
230 | #define MM_SLOTS_HASH_BITS 10 |
231 | static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
31dbd01f | 232 | |
21fbd591 | 233 | static struct ksm_mm_slot ksm_mm_head = { |
23f746e4 | 234 | .mm_node = LIST_HEAD_INIT(ksm_mm_head.mm_node), |
31dbd01f IE |
235 | }; |
236 | static struct ksm_scan ksm_scan = { | |
237 | .mm_slot = &ksm_mm_head, | |
238 | }; | |
239 | ||
240 | static struct kmem_cache *rmap_item_cache; | |
7b6ba2c7 | 241 | static struct kmem_cache *stable_node_cache; |
31dbd01f IE |
242 | static struct kmem_cache *mm_slot_cache; |
243 | ||
244 | /* The number of nodes in the stable tree */ | |
b4028260 | 245 | static unsigned long ksm_pages_shared; |
31dbd01f | 246 | |
e178dfde | 247 | /* The number of page slots additionally sharing those nodes */ |
b4028260 | 248 | static unsigned long ksm_pages_sharing; |
31dbd01f | 249 | |
473b0ce4 HD |
250 | /* The number of nodes in the unstable tree */ |
251 | static unsigned long ksm_pages_unshared; | |
252 | ||
253 | /* The number of rmap_items in use: to calculate pages_volatile */ | |
254 | static unsigned long ksm_rmap_items; | |
255 | ||
2c653d0e AA |
256 | /* The number of stable_node chains */ |
257 | static unsigned long ksm_stable_node_chains; | |
258 | ||
259 | /* The number of stable_node dups linked to the stable_node chains */ | |
260 | static unsigned long ksm_stable_node_dups; | |
261 | ||
262 | /* Delay in pruning stale stable_node_dups in the stable_node_chains */ | |
584ff0df | 263 | static unsigned int ksm_stable_node_chains_prune_millisecs = 2000; |
2c653d0e AA |
264 | |
265 | /* Maximum number of page slots sharing a stable node */ | |
266 | static int ksm_max_page_sharing = 256; | |
267 | ||
31dbd01f | 268 | /* Number of pages ksmd should scan in one batch */ |
2c6854fd | 269 | static unsigned int ksm_thread_pages_to_scan = 100; |
31dbd01f IE |
270 | |
271 | /* Milliseconds ksmd should sleep between batches */ | |
2ffd8679 | 272 | static unsigned int ksm_thread_sleep_millisecs = 20; |
31dbd01f | 273 | |
e86c59b1 CI |
274 | /* Checksum of an empty (zeroed) page */ |
275 | static unsigned int zero_checksum __read_mostly; | |
276 | ||
277 | /* Whether to merge empty (zeroed) pages with actual zero pages */ | |
278 | static bool ksm_use_zero_pages __read_mostly; | |
279 | ||
e850dcf5 | 280 | #ifdef CONFIG_NUMA |
90bd6fd3 PH |
281 | /* Zeroed when merging across nodes is not allowed */ |
282 | static unsigned int ksm_merge_across_nodes = 1; | |
ef53d16c | 283 | static int ksm_nr_node_ids = 1; |
e850dcf5 HD |
284 | #else |
285 | #define ksm_merge_across_nodes 1U | |
ef53d16c | 286 | #define ksm_nr_node_ids 1 |
e850dcf5 | 287 | #endif |
90bd6fd3 | 288 | |
31dbd01f IE |
289 | #define KSM_RUN_STOP 0 |
290 | #define KSM_RUN_MERGE 1 | |
291 | #define KSM_RUN_UNMERGE 2 | |
ef4d43a8 HD |
292 | #define KSM_RUN_OFFLINE 4 |
293 | static unsigned long ksm_run = KSM_RUN_STOP; | |
294 | static void wait_while_offlining(void); | |
31dbd01f IE |
295 | |
296 | static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); | |
fcf9a0ef | 297 | static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait); |
31dbd01f IE |
298 | static DEFINE_MUTEX(ksm_thread_mutex); |
299 | static DEFINE_SPINLOCK(ksm_mmlist_lock); | |
300 | ||
21fbd591 | 301 | #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create(#__struct,\ |
31dbd01f IE |
302 | sizeof(struct __struct), __alignof__(struct __struct),\ |
303 | (__flags), NULL) | |
304 | ||
305 | static int __init ksm_slab_init(void) | |
306 | { | |
21fbd591 | 307 | rmap_item_cache = KSM_KMEM_CACHE(ksm_rmap_item, 0); |
31dbd01f IE |
308 | if (!rmap_item_cache) |
309 | goto out; | |
310 | ||
21fbd591 | 311 | stable_node_cache = KSM_KMEM_CACHE(ksm_stable_node, 0); |
7b6ba2c7 HD |
312 | if (!stable_node_cache) |
313 | goto out_free1; | |
314 | ||
21fbd591 | 315 | mm_slot_cache = KSM_KMEM_CACHE(ksm_mm_slot, 0); |
31dbd01f | 316 | if (!mm_slot_cache) |
7b6ba2c7 | 317 | goto out_free2; |
31dbd01f IE |
318 | |
319 | return 0; | |
320 | ||
7b6ba2c7 HD |
321 | out_free2: |
322 | kmem_cache_destroy(stable_node_cache); | |
323 | out_free1: | |
31dbd01f IE |
324 | kmem_cache_destroy(rmap_item_cache); |
325 | out: | |
326 | return -ENOMEM; | |
327 | } | |
328 | ||
329 | static void __init ksm_slab_free(void) | |
330 | { | |
331 | kmem_cache_destroy(mm_slot_cache); | |
7b6ba2c7 | 332 | kmem_cache_destroy(stable_node_cache); |
31dbd01f IE |
333 | kmem_cache_destroy(rmap_item_cache); |
334 | mm_slot_cache = NULL; | |
335 | } | |
336 | ||
21fbd591 | 337 | static __always_inline bool is_stable_node_chain(struct ksm_stable_node *chain) |
2c653d0e AA |
338 | { |
339 | return chain->rmap_hlist_len == STABLE_NODE_CHAIN; | |
340 | } | |
341 | ||
21fbd591 | 342 | static __always_inline bool is_stable_node_dup(struct ksm_stable_node *dup) |
2c653d0e AA |
343 | { |
344 | return dup->head == STABLE_NODE_DUP_HEAD; | |
345 | } | |
346 | ||
21fbd591 QZ |
347 | static inline void stable_node_chain_add_dup(struct ksm_stable_node *dup, |
348 | struct ksm_stable_node *chain) | |
2c653d0e AA |
349 | { |
350 | VM_BUG_ON(is_stable_node_dup(dup)); | |
351 | dup->head = STABLE_NODE_DUP_HEAD; | |
352 | VM_BUG_ON(!is_stable_node_chain(chain)); | |
353 | hlist_add_head(&dup->hlist_dup, &chain->hlist); | |
354 | ksm_stable_node_dups++; | |
355 | } | |
356 | ||
21fbd591 | 357 | static inline void __stable_node_dup_del(struct ksm_stable_node *dup) |
2c653d0e | 358 | { |
b4fecc67 | 359 | VM_BUG_ON(!is_stable_node_dup(dup)); |
2c653d0e AA |
360 | hlist_del(&dup->hlist_dup); |
361 | ksm_stable_node_dups--; | |
362 | } | |
363 | ||
21fbd591 | 364 | static inline void stable_node_dup_del(struct ksm_stable_node *dup) |
2c653d0e AA |
365 | { |
366 | VM_BUG_ON(is_stable_node_chain(dup)); | |
367 | if (is_stable_node_dup(dup)) | |
368 | __stable_node_dup_del(dup); | |
369 | else | |
370 | rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid)); | |
371 | #ifdef CONFIG_DEBUG_VM | |
372 | dup->head = NULL; | |
373 | #endif | |
374 | } | |
375 | ||
21fbd591 | 376 | static inline struct ksm_rmap_item *alloc_rmap_item(void) |
31dbd01f | 377 | { |
21fbd591 | 378 | struct ksm_rmap_item *rmap_item; |
473b0ce4 | 379 | |
5b398e41 | 380 | rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | |
381 | __GFP_NORETRY | __GFP_NOWARN); | |
473b0ce4 HD |
382 | if (rmap_item) |
383 | ksm_rmap_items++; | |
384 | return rmap_item; | |
31dbd01f IE |
385 | } |
386 | ||
21fbd591 | 387 | static inline void free_rmap_item(struct ksm_rmap_item *rmap_item) |
31dbd01f | 388 | { |
473b0ce4 | 389 | ksm_rmap_items--; |
cb4df4ca | 390 | rmap_item->mm->ksm_rmap_items--; |
31dbd01f IE |
391 | rmap_item->mm = NULL; /* debug safety */ |
392 | kmem_cache_free(rmap_item_cache, rmap_item); | |
393 | } | |
394 | ||
21fbd591 | 395 | static inline struct ksm_stable_node *alloc_stable_node(void) |
7b6ba2c7 | 396 | { |
6213055f | 397 | /* |
398 | * The allocation can take too long with GFP_KERNEL when memory is under | |
399 | * pressure, which may lead to hung task warnings. Adding __GFP_HIGH | |
400 | * grants access to memory reserves, helping to avoid this problem. | |
401 | */ | |
402 | return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); | |
7b6ba2c7 HD |
403 | } |
404 | ||
21fbd591 | 405 | static inline void free_stable_node(struct ksm_stable_node *stable_node) |
7b6ba2c7 | 406 | { |
2c653d0e AA |
407 | VM_BUG_ON(stable_node->rmap_hlist_len && |
408 | !is_stable_node_chain(stable_node)); | |
7b6ba2c7 HD |
409 | kmem_cache_free(stable_node_cache, stable_node); |
410 | } | |
411 | ||
21fbd591 | 412 | static inline struct ksm_mm_slot *alloc_mm_slot(void) |
31dbd01f IE |
413 | { |
414 | if (!mm_slot_cache) /* initialization failed */ | |
415 | return NULL; | |
416 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
417 | } | |
418 | ||
21fbd591 | 419 | static inline void free_mm_slot(struct ksm_mm_slot *mm_slot) |
31dbd01f IE |
420 | { |
421 | kmem_cache_free(mm_slot_cache, mm_slot); | |
422 | } | |
423 | ||
21fbd591 | 424 | static struct ksm_mm_slot *get_mm_slot(struct mm_struct *mm) |
31dbd01f | 425 | { |
21fbd591 | 426 | struct ksm_mm_slot *slot; |
4ca3a69b | 427 | |
79b09941 | 428 | hash_for_each_possible(mm_slots_hash, slot, hash, (unsigned long)mm) |
4ca3a69b SL |
429 | if (slot->mm == mm) |
430 | return slot; | |
31dbd01f | 431 | |
31dbd01f IE |
432 | return NULL; |
433 | } | |
434 | ||
435 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
21fbd591 | 436 | struct ksm_mm_slot *mm_slot) |
31dbd01f | 437 | { |
31dbd01f | 438 | mm_slot->mm = mm; |
79b09941 | 439 | hash_add(mm_slots_hash, &mm_slot->hash, (unsigned long)mm); |
31dbd01f IE |
440 | } |
441 | ||
a913e182 HD |
442 | /* |
443 | * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's | |
444 | * page tables after it has passed through ksm_exit() - which, if necessary, | |
c1e8d7c6 | 445 | * takes mmap_lock briefly to serialize against them. ksm_exit() does not set |
a913e182 HD |
446 | * a special flag: they can just back out as soon as mm_users goes to zero. |
447 | * ksm_test_exit() is used throughout to make this test for exit: in some | |
448 | * places for correctness, in some places just to avoid unnecessary work. | |
449 | */ | |
450 | static inline bool ksm_test_exit(struct mm_struct *mm) | |
451 | { | |
452 | return atomic_read(&mm->mm_users) == 0; | |
453 | } | |
454 | ||
31dbd01f IE |
455 | /* |
456 | * We use break_ksm to break COW on a ksm page: it's a stripped down | |
457 | * | |
7a9547fd | 458 | * if (get_user_pages(addr, 1, FOLL_WRITE, &page, NULL) == 1) |
31dbd01f IE |
459 | * put_page(page); |
460 | * | |
461 | * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma, | |
462 | * in case the application has unmapped and remapped mm,addr meanwhile. | |
463 | * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP | |
bbcd53c9 | 464 | * mmap of /dev/mem, where we would not want to touch it. |
1b2ee126 DH |
465 | * |
466 | * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context | |
467 | * of the process that owns 'vma'. We also do not want to enforce | |
468 | * protection keys here anyway. | |
31dbd01f | 469 | */ |
d952b791 | 470 | static int break_ksm(struct vm_area_struct *vma, unsigned long addr) |
31dbd01f IE |
471 | { |
472 | struct page *page; | |
50a7ca3c | 473 | vm_fault_t ret = 0; |
31dbd01f IE |
474 | |
475 | do { | |
476 | cond_resched(); | |
1b2ee126 DH |
477 | page = follow_page(vma, addr, |
478 | FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE); | |
f7091ed6 | 479 | if (IS_ERR_OR_NULL(page)) |
31dbd01f IE |
480 | break; |
481 | if (PageKsm(page)) | |
dcddffd4 | 482 | ret = handle_mm_fault(vma, addr, |
bce617ed PX |
483 | FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE, |
484 | NULL); | |
31dbd01f IE |
485 | else |
486 | ret = VM_FAULT_WRITE; | |
487 | put_page(page); | |
33692f27 | 488 | } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); |
d952b791 HD |
489 | /* |
490 | * We must loop because handle_mm_fault() may back out if there's | |
491 | * any difficulty e.g. if pte accessed bit gets updated concurrently. | |
492 | * | |
493 | * VM_FAULT_WRITE is what we have been hoping for: it indicates that | |
494 | * COW has been broken, even if the vma does not permit VM_WRITE; | |
495 | * but note that a concurrent fault might break PageKsm for us. | |
496 | * | |
497 | * VM_FAULT_SIGBUS could occur if we race with truncation of the | |
498 | * backing file, which also invalidates anonymous pages: that's | |
499 | * okay, that truncation will have unmapped the PageKsm for us. | |
500 | * | |
501 | * VM_FAULT_OOM: at the time of writing (late July 2009), setting | |
502 | * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the | |
503 | * current task has TIF_MEMDIE set, and will be OOM killed on return | |
504 | * to user; and ksmd, having no mm, would never be chosen for that. | |
505 | * | |
506 | * But if the mm is in a limited mem_cgroup, then the fault may fail | |
507 | * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and | |
508 | * even ksmd can fail in this way - though it's usually breaking ksm | |
509 | * just to undo a merge it made a moment before, so unlikely to oom. | |
510 | * | |
511 | * That's a pity: we might therefore have more kernel pages allocated | |
512 | * than we're counting as nodes in the stable tree; but ksm_do_scan | |
513 | * will retry to break_cow on each pass, so should recover the page | |
514 | * in due course. The important thing is to not let VM_MERGEABLE | |
515 | * be cleared while any such pages might remain in the area. | |
516 | */ | |
517 | return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; | |
31dbd01f IE |
518 | } |
519 | ||
ef694222 BL |
520 | static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, |
521 | unsigned long addr) | |
522 | { | |
523 | struct vm_area_struct *vma; | |
524 | if (ksm_test_exit(mm)) | |
525 | return NULL; | |
ff69fb81 LH |
526 | vma = vma_lookup(mm, addr); |
527 | if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) | |
ef694222 BL |
528 | return NULL; |
529 | return vma; | |
530 | } | |
531 | ||
21fbd591 | 532 | static void break_cow(struct ksm_rmap_item *rmap_item) |
31dbd01f | 533 | { |
8dd3557a HD |
534 | struct mm_struct *mm = rmap_item->mm; |
535 | unsigned long addr = rmap_item->address; | |
31dbd01f IE |
536 | struct vm_area_struct *vma; |
537 | ||
4035c07a HD |
538 | /* |
539 | * It is not an accident that whenever we want to break COW | |
540 | * to undo, we also need to drop a reference to the anon_vma. | |
541 | */ | |
9e60109f | 542 | put_anon_vma(rmap_item->anon_vma); |
4035c07a | 543 | |
d8ed45c5 | 544 | mmap_read_lock(mm); |
ef694222 BL |
545 | vma = find_mergeable_vma(mm, addr); |
546 | if (vma) | |
547 | break_ksm(vma, addr); | |
d8ed45c5 | 548 | mmap_read_unlock(mm); |
31dbd01f IE |
549 | } |
550 | ||
21fbd591 | 551 | static struct page *get_mergeable_page(struct ksm_rmap_item *rmap_item) |
31dbd01f IE |
552 | { |
553 | struct mm_struct *mm = rmap_item->mm; | |
554 | unsigned long addr = rmap_item->address; | |
555 | struct vm_area_struct *vma; | |
556 | struct page *page; | |
557 | ||
d8ed45c5 | 558 | mmap_read_lock(mm); |
ef694222 BL |
559 | vma = find_mergeable_vma(mm, addr); |
560 | if (!vma) | |
31dbd01f IE |
561 | goto out; |
562 | ||
563 | page = follow_page(vma, addr, FOLL_GET); | |
f7091ed6 | 564 | if (IS_ERR_OR_NULL(page)) |
31dbd01f | 565 | goto out; |
f7091ed6 HW |
566 | if (is_zone_device_page(page)) |
567 | goto out_putpage; | |
f765f540 | 568 | if (PageAnon(page)) { |
31dbd01f IE |
569 | flush_anon_page(vma, page, addr); |
570 | flush_dcache_page(page); | |
571 | } else { | |
f7091ed6 | 572 | out_putpage: |
31dbd01f | 573 | put_page(page); |
c8f95ed1 AA |
574 | out: |
575 | page = NULL; | |
31dbd01f | 576 | } |
d8ed45c5 | 577 | mmap_read_unlock(mm); |
31dbd01f IE |
578 | return page; |
579 | } | |
580 | ||
90bd6fd3 PH |
581 | /* |
582 | * This helper is used for getting right index into array of tree roots. | |
583 | * When merge_across_nodes knob is set to 1, there are only two rb-trees for | |
584 | * stable and unstable pages from all nodes with roots in index 0. Otherwise, | |
585 | * every node has its own stable and unstable tree. | |
586 | */ | |
587 | static inline int get_kpfn_nid(unsigned long kpfn) | |
588 | { | |
d8fc16a8 | 589 | return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); |
90bd6fd3 PH |
590 | } |
591 | ||
21fbd591 | 592 | static struct ksm_stable_node *alloc_stable_node_chain(struct ksm_stable_node *dup, |
2c653d0e AA |
593 | struct rb_root *root) |
594 | { | |
21fbd591 | 595 | struct ksm_stable_node *chain = alloc_stable_node(); |
2c653d0e AA |
596 | VM_BUG_ON(is_stable_node_chain(dup)); |
597 | if (likely(chain)) { | |
598 | INIT_HLIST_HEAD(&chain->hlist); | |
599 | chain->chain_prune_time = jiffies; | |
600 | chain->rmap_hlist_len = STABLE_NODE_CHAIN; | |
601 | #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA) | |
98fa15f3 | 602 | chain->nid = NUMA_NO_NODE; /* debug */ |
2c653d0e AA |
603 | #endif |
604 | ksm_stable_node_chains++; | |
605 | ||
606 | /* | |
607 | * Put the stable node chain in the first dimension of | |
608 | * the stable tree and at the same time remove the old | |
609 | * stable node. | |
610 | */ | |
611 | rb_replace_node(&dup->node, &chain->node, root); | |
612 | ||
613 | /* | |
614 | * Move the old stable node to the second dimension | |
615 | * queued in the hlist_dup. The invariant is that all | |
616 | * dup stable_nodes in the chain->hlist point to pages | |
457aef94 | 617 | * that are write protected and have the exact same |
2c653d0e AA |
618 | * content. |
619 | */ | |
620 | stable_node_chain_add_dup(dup, chain); | |
621 | } | |
622 | return chain; | |
623 | } | |
624 | ||
21fbd591 | 625 | static inline void free_stable_node_chain(struct ksm_stable_node *chain, |
2c653d0e AA |
626 | struct rb_root *root) |
627 | { | |
628 | rb_erase(&chain->node, root); | |
629 | free_stable_node(chain); | |
630 | ksm_stable_node_chains--; | |
631 | } | |
632 | ||
21fbd591 | 633 | static void remove_node_from_stable_tree(struct ksm_stable_node *stable_node) |
4035c07a | 634 | { |
21fbd591 | 635 | struct ksm_rmap_item *rmap_item; |
4035c07a | 636 | |
2c653d0e AA |
637 | /* check it's not STABLE_NODE_CHAIN or negative */ |
638 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
639 | ||
b67bfe0d | 640 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
4035c07a HD |
641 | if (rmap_item->hlist.next) |
642 | ksm_pages_sharing--; | |
643 | else | |
644 | ksm_pages_shared--; | |
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 | |
4146d2d6 HD |
665 | if (stable_node->head == &migrate_nodes) |
666 | list_del(&stable_node->list); | |
667 | else | |
2c653d0e | 668 | stable_node_dup_del(stable_node); |
4035c07a HD |
669 | free_stable_node(stable_node); |
670 | } | |
671 | ||
2cee57d1 YS |
672 | enum get_ksm_page_flags { |
673 | GET_KSM_PAGE_NOLOCK, | |
674 | GET_KSM_PAGE_LOCK, | |
675 | GET_KSM_PAGE_TRYLOCK | |
676 | }; | |
677 | ||
4035c07a HD |
678 | /* |
679 | * get_ksm_page: checks if the page indicated by the stable node | |
680 | * is still its ksm page, despite having held no reference to it. | |
681 | * In which case we can trust the content of the page, and it | |
682 | * returns the gotten page; but if the page has now been zapped, | |
683 | * remove the stale node from the stable tree and return NULL. | |
c8d6553b | 684 | * But beware, the stable node's page might be being migrated. |
4035c07a HD |
685 | * |
686 | * You would expect the stable_node to hold a reference to the ksm page. | |
687 | * But if it increments the page's count, swapping out has to wait for | |
688 | * ksmd to come around again before it can free the page, which may take | |
689 | * seconds or even minutes: much too unresponsive. So instead we use a | |
690 | * "keyhole reference": access to the ksm page from the stable node peeps | |
691 | * out through its keyhole to see if that page still holds the right key, | |
692 | * pointing back to this stable node. This relies on freeing a PageAnon | |
693 | * page to reset its page->mapping to NULL, and relies on no other use of | |
694 | * a page to put something that might look like our key in page->mapping. | |
4035c07a HD |
695 | * is on its way to being freed; but it is an anomaly to bear in mind. |
696 | */ | |
21fbd591 | 697 | static struct page *get_ksm_page(struct ksm_stable_node *stable_node, |
2cee57d1 | 698 | enum get_ksm_page_flags flags) |
4035c07a HD |
699 | { |
700 | struct page *page; | |
701 | void *expected_mapping; | |
c8d6553b | 702 | unsigned long kpfn; |
4035c07a | 703 | |
bda807d4 MK |
704 | expected_mapping = (void *)((unsigned long)stable_node | |
705 | PAGE_MAPPING_KSM); | |
c8d6553b | 706 | again: |
08df4774 | 707 | kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */ |
c8d6553b | 708 | page = pfn_to_page(kpfn); |
4db0c3c2 | 709 | if (READ_ONCE(page->mapping) != expected_mapping) |
4035c07a | 710 | goto stale; |
c8d6553b HD |
711 | |
712 | /* | |
713 | * We cannot do anything with the page while its refcount is 0. | |
714 | * Usually 0 means free, or tail of a higher-order page: in which | |
715 | * case this node is no longer referenced, and should be freed; | |
1c4c3b99 | 716 | * however, it might mean that the page is under page_ref_freeze(). |
c8d6553b | 717 | * The __remove_mapping() case is easy, again the node is now stale; |
52d1e606 | 718 | * the same is in reuse_ksm_page() case; but if page is swapcache |
9800562f | 719 | * in folio_migrate_mapping(), it might still be our page, |
52d1e606 | 720 | * in which case it's essential to keep the node. |
c8d6553b HD |
721 | */ |
722 | while (!get_page_unless_zero(page)) { | |
723 | /* | |
724 | * Another check for page->mapping != expected_mapping would | |
725 | * work here too. We have chosen the !PageSwapCache test to | |
726 | * optimize the common case, when the page is or is about to | |
727 | * be freed: PageSwapCache is cleared (under spin_lock_irq) | |
1c4c3b99 | 728 | * in the ref_freeze section of __remove_mapping(); but Anon |
c8d6553b HD |
729 | * page->mapping reset to NULL later, in free_pages_prepare(). |
730 | */ | |
731 | if (!PageSwapCache(page)) | |
732 | goto stale; | |
733 | cpu_relax(); | |
734 | } | |
735 | ||
4db0c3c2 | 736 | if (READ_ONCE(page->mapping) != expected_mapping) { |
4035c07a HD |
737 | put_page(page); |
738 | goto stale; | |
739 | } | |
c8d6553b | 740 | |
2cee57d1 YS |
741 | if (flags == GET_KSM_PAGE_TRYLOCK) { |
742 | if (!trylock_page(page)) { | |
743 | put_page(page); | |
744 | return ERR_PTR(-EBUSY); | |
745 | } | |
746 | } else if (flags == GET_KSM_PAGE_LOCK) | |
8aafa6a4 | 747 | lock_page(page); |
2cee57d1 YS |
748 | |
749 | if (flags != GET_KSM_PAGE_NOLOCK) { | |
4db0c3c2 | 750 | if (READ_ONCE(page->mapping) != expected_mapping) { |
8aafa6a4 HD |
751 | unlock_page(page); |
752 | put_page(page); | |
753 | goto stale; | |
754 | } | |
755 | } | |
4035c07a | 756 | return page; |
c8d6553b | 757 | |
4035c07a | 758 | stale: |
c8d6553b HD |
759 | /* |
760 | * We come here from above when page->mapping or !PageSwapCache | |
761 | * suggests that the node is stale; but it might be under migration. | |
19138349 | 762 | * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(), |
c8d6553b HD |
763 | * before checking whether node->kpfn has been changed. |
764 | */ | |
765 | smp_rmb(); | |
4db0c3c2 | 766 | if (READ_ONCE(stable_node->kpfn) != kpfn) |
c8d6553b | 767 | goto again; |
4035c07a HD |
768 | remove_node_from_stable_tree(stable_node); |
769 | return NULL; | |
770 | } | |
771 | ||
31dbd01f IE |
772 | /* |
773 | * Removing rmap_item from stable or unstable tree. | |
774 | * This function will clean the information from the stable/unstable tree. | |
775 | */ | |
21fbd591 | 776 | static void remove_rmap_item_from_tree(struct ksm_rmap_item *rmap_item) |
31dbd01f | 777 | { |
7b6ba2c7 | 778 | if (rmap_item->address & STABLE_FLAG) { |
21fbd591 | 779 | struct ksm_stable_node *stable_node; |
5ad64688 | 780 | struct page *page; |
31dbd01f | 781 | |
7b6ba2c7 | 782 | stable_node = rmap_item->head; |
62862290 | 783 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
4035c07a HD |
784 | if (!page) |
785 | goto out; | |
5ad64688 | 786 | |
7b6ba2c7 | 787 | hlist_del(&rmap_item->hlist); |
62862290 | 788 | unlock_page(page); |
4035c07a | 789 | put_page(page); |
08beca44 | 790 | |
98666f8a | 791 | if (!hlist_empty(&stable_node->hlist)) |
4035c07a HD |
792 | ksm_pages_sharing--; |
793 | else | |
7b6ba2c7 | 794 | ksm_pages_shared--; |
76093853 | 795 | |
796 | rmap_item->mm->ksm_merging_pages--; | |
797 | ||
2c653d0e AA |
798 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
799 | stable_node->rmap_hlist_len--; | |
31dbd01f | 800 | |
9e60109f | 801 | put_anon_vma(rmap_item->anon_vma); |
c89a384e | 802 | rmap_item->head = NULL; |
93d17715 | 803 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 804 | |
7b6ba2c7 | 805 | } else if (rmap_item->address & UNSTABLE_FLAG) { |
31dbd01f IE |
806 | unsigned char age; |
807 | /* | |
9ba69294 | 808 | * Usually ksmd can and must skip the rb_erase, because |
31dbd01f | 809 | * root_unstable_tree was already reset to RB_ROOT. |
9ba69294 HD |
810 | * But be careful when an mm is exiting: do the rb_erase |
811 | * if this rmap_item was inserted by this scan, rather | |
812 | * than left over from before. | |
31dbd01f IE |
813 | */ |
814 | age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); | |
cd551f97 | 815 | BUG_ON(age > 1); |
31dbd01f | 816 | if (!age) |
90bd6fd3 | 817 | rb_erase(&rmap_item->node, |
ef53d16c | 818 | root_unstable_tree + NUMA(rmap_item->nid)); |
473b0ce4 | 819 | ksm_pages_unshared--; |
93d17715 | 820 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 821 | } |
4035c07a | 822 | out: |
31dbd01f IE |
823 | cond_resched(); /* we're called from many long loops */ |
824 | } | |
825 | ||
21fbd591 | 826 | static void remove_trailing_rmap_items(struct ksm_rmap_item **rmap_list) |
31dbd01f | 827 | { |
6514d511 | 828 | while (*rmap_list) { |
21fbd591 | 829 | struct ksm_rmap_item *rmap_item = *rmap_list; |
6514d511 | 830 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 831 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
832 | free_rmap_item(rmap_item); |
833 | } | |
834 | } | |
835 | ||
836 | /* | |
e850dcf5 | 837 | * Though it's very tempting to unmerge rmap_items from stable tree rather |
31dbd01f IE |
838 | * than check every pte of a given vma, the locking doesn't quite work for |
839 | * that - an rmap_item is assigned to the stable tree after inserting ksm | |
c1e8d7c6 | 840 | * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing |
31dbd01f IE |
841 | * rmap_items from parent to child at fork time (so as not to waste time |
842 | * if exit comes before the next scan reaches it). | |
81464e30 HD |
843 | * |
844 | * Similarly, although we'd like to remove rmap_items (so updating counts | |
845 | * and freeing memory) when unmerging an area, it's easier to leave that | |
846 | * to the next pass of ksmd - consider, for example, how ksmd might be | |
847 | * in cmp_and_merge_page on one of the rmap_items we would be removing. | |
31dbd01f | 848 | */ |
d952b791 HD |
849 | static int unmerge_ksm_pages(struct vm_area_struct *vma, |
850 | unsigned long start, unsigned long end) | |
31dbd01f IE |
851 | { |
852 | unsigned long addr; | |
d952b791 | 853 | int err = 0; |
31dbd01f | 854 | |
d952b791 | 855 | for (addr = start; addr < end && !err; addr += PAGE_SIZE) { |
9ba69294 HD |
856 | if (ksm_test_exit(vma->vm_mm)) |
857 | break; | |
d952b791 HD |
858 | if (signal_pending(current)) |
859 | err = -ERESTARTSYS; | |
860 | else | |
861 | err = break_ksm(vma, addr); | |
862 | } | |
863 | return err; | |
31dbd01f IE |
864 | } |
865 | ||
21fbd591 | 866 | static inline struct ksm_stable_node *folio_stable_node(struct folio *folio) |
19138349 MWO |
867 | { |
868 | return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL; | |
869 | } | |
870 | ||
21fbd591 | 871 | static inline struct ksm_stable_node *page_stable_node(struct page *page) |
88484826 | 872 | { |
19138349 | 873 | return folio_stable_node(page_folio(page)); |
88484826 MR |
874 | } |
875 | ||
876 | static inline void set_page_stable_node(struct page *page, | |
21fbd591 | 877 | struct ksm_stable_node *stable_node) |
88484826 | 878 | { |
6c287605 | 879 | VM_BUG_ON_PAGE(PageAnon(page) && PageAnonExclusive(page), page); |
88484826 MR |
880 | page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM); |
881 | } | |
882 | ||
2ffd8679 HD |
883 | #ifdef CONFIG_SYSFS |
884 | /* | |
885 | * Only called through the sysfs control interface: | |
886 | */ | |
21fbd591 | 887 | static int remove_stable_node(struct ksm_stable_node *stable_node) |
cbf86cfe HD |
888 | { |
889 | struct page *page; | |
890 | int err; | |
891 | ||
2cee57d1 | 892 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
cbf86cfe HD |
893 | if (!page) { |
894 | /* | |
895 | * get_ksm_page did remove_node_from_stable_tree itself. | |
896 | */ | |
897 | return 0; | |
898 | } | |
899 | ||
9a63236f AR |
900 | /* |
901 | * Page could be still mapped if this races with __mmput() running in | |
902 | * between ksm_exit() and exit_mmap(). Just refuse to let | |
903 | * merge_across_nodes/max_page_sharing be switched. | |
904 | */ | |
905 | err = -EBUSY; | |
906 | if (!page_mapped(page)) { | |
cbf86cfe | 907 | /* |
8fdb3dbf HD |
908 | * The stable node did not yet appear stale to get_ksm_page(), |
909 | * since that allows for an unmapped ksm page to be recognized | |
910 | * right up until it is freed; but the node is safe to remove. | |
cbf86cfe HD |
911 | * This page might be in a pagevec waiting to be freed, |
912 | * or it might be PageSwapCache (perhaps under writeback), | |
913 | * or it might have been removed from swapcache a moment ago. | |
914 | */ | |
915 | set_page_stable_node(page, NULL); | |
916 | remove_node_from_stable_tree(stable_node); | |
917 | err = 0; | |
918 | } | |
919 | ||
920 | unlock_page(page); | |
921 | put_page(page); | |
922 | return err; | |
923 | } | |
924 | ||
21fbd591 | 925 | static int remove_stable_node_chain(struct ksm_stable_node *stable_node, |
2c653d0e AA |
926 | struct rb_root *root) |
927 | { | |
21fbd591 | 928 | struct ksm_stable_node *dup; |
2c653d0e AA |
929 | struct hlist_node *hlist_safe; |
930 | ||
931 | if (!is_stable_node_chain(stable_node)) { | |
932 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
933 | if (remove_stable_node(stable_node)) | |
934 | return true; | |
935 | else | |
936 | return false; | |
937 | } | |
938 | ||
939 | hlist_for_each_entry_safe(dup, hlist_safe, | |
940 | &stable_node->hlist, hlist_dup) { | |
941 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
942 | if (remove_stable_node(dup)) | |
943 | return true; | |
944 | } | |
945 | BUG_ON(!hlist_empty(&stable_node->hlist)); | |
946 | free_stable_node_chain(stable_node, root); | |
947 | return false; | |
948 | } | |
949 | ||
cbf86cfe HD |
950 | static int remove_all_stable_nodes(void) |
951 | { | |
21fbd591 | 952 | struct ksm_stable_node *stable_node, *next; |
cbf86cfe HD |
953 | int nid; |
954 | int err = 0; | |
955 | ||
ef53d16c | 956 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
cbf86cfe HD |
957 | while (root_stable_tree[nid].rb_node) { |
958 | stable_node = rb_entry(root_stable_tree[nid].rb_node, | |
21fbd591 | 959 | struct ksm_stable_node, node); |
2c653d0e AA |
960 | if (remove_stable_node_chain(stable_node, |
961 | root_stable_tree + nid)) { | |
cbf86cfe HD |
962 | err = -EBUSY; |
963 | break; /* proceed to next nid */ | |
964 | } | |
965 | cond_resched(); | |
966 | } | |
967 | } | |
03640418 | 968 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
969 | if (remove_stable_node(stable_node)) |
970 | err = -EBUSY; | |
971 | cond_resched(); | |
972 | } | |
cbf86cfe HD |
973 | return err; |
974 | } | |
975 | ||
d952b791 | 976 | static int unmerge_and_remove_all_rmap_items(void) |
31dbd01f | 977 | { |
21fbd591 | 978 | struct ksm_mm_slot *mm_slot; |
31dbd01f IE |
979 | struct mm_struct *mm; |
980 | struct vm_area_struct *vma; | |
d952b791 HD |
981 | int err = 0; |
982 | ||
983 | spin_lock(&ksm_mmlist_lock); | |
23f746e4 QZ |
984 | ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_node.next, |
985 | struct ksm_mm_slot, mm_node); | |
d952b791 | 986 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f | 987 | |
a5f18ba0 MWO |
988 | for (mm_slot = ksm_scan.mm_slot; mm_slot != &ksm_mm_head; |
989 | mm_slot = ksm_scan.mm_slot) { | |
990 | VMA_ITERATOR(vmi, mm_slot->mm, 0); | |
991 | ||
31dbd01f | 992 | mm = mm_slot->mm; |
d8ed45c5 | 993 | mmap_read_lock(mm); |
a5f18ba0 | 994 | for_each_vma(vmi, vma) { |
9ba69294 HD |
995 | if (ksm_test_exit(mm)) |
996 | break; | |
31dbd01f IE |
997 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
998 | continue; | |
d952b791 HD |
999 | err = unmerge_ksm_pages(vma, |
1000 | vma->vm_start, vma->vm_end); | |
9ba69294 HD |
1001 | if (err) |
1002 | goto error; | |
31dbd01f | 1003 | } |
9ba69294 | 1004 | |
420be4ed | 1005 | remove_trailing_rmap_items(&mm_slot->rmap_list); |
d8ed45c5 | 1006 | mmap_read_unlock(mm); |
d952b791 HD |
1007 | |
1008 | spin_lock(&ksm_mmlist_lock); | |
23f746e4 QZ |
1009 | ksm_scan.mm_slot = list_entry(mm_slot->mm_node.next, |
1010 | struct ksm_mm_slot, mm_node); | |
9ba69294 | 1011 | if (ksm_test_exit(mm)) { |
79b09941 | 1012 | hash_del(&mm_slot->hash); |
23f746e4 | 1013 | list_del(&mm_slot->mm_node); |
9ba69294 HD |
1014 | spin_unlock(&ksm_mmlist_lock); |
1015 | ||
1016 | free_mm_slot(mm_slot); | |
1017 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
9ba69294 | 1018 | mmdrop(mm); |
7496fea9 | 1019 | } else |
9ba69294 | 1020 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f IE |
1021 | } |
1022 | ||
cbf86cfe HD |
1023 | /* Clean up stable nodes, but don't worry if some are still busy */ |
1024 | remove_all_stable_nodes(); | |
d952b791 | 1025 | ksm_scan.seqnr = 0; |
9ba69294 HD |
1026 | return 0; |
1027 | ||
1028 | error: | |
d8ed45c5 | 1029 | mmap_read_unlock(mm); |
31dbd01f | 1030 | spin_lock(&ksm_mmlist_lock); |
d952b791 | 1031 | ksm_scan.mm_slot = &ksm_mm_head; |
31dbd01f | 1032 | spin_unlock(&ksm_mmlist_lock); |
d952b791 | 1033 | return err; |
31dbd01f | 1034 | } |
2ffd8679 | 1035 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 1036 | |
31dbd01f IE |
1037 | static u32 calc_checksum(struct page *page) |
1038 | { | |
1039 | u32 checksum; | |
9b04c5fe | 1040 | void *addr = kmap_atomic(page); |
59e1a2f4 | 1041 | checksum = xxhash(addr, PAGE_SIZE, 0); |
9b04c5fe | 1042 | kunmap_atomic(addr); |
31dbd01f IE |
1043 | return checksum; |
1044 | } | |
1045 | ||
31dbd01f IE |
1046 | static int write_protect_page(struct vm_area_struct *vma, struct page *page, |
1047 | pte_t *orig_pte) | |
1048 | { | |
1049 | struct mm_struct *mm = vma->vm_mm; | |
eed05e54 | 1050 | DEFINE_PAGE_VMA_WALK(pvmw, page, vma, 0, 0); |
31dbd01f IE |
1051 | int swapped; |
1052 | int err = -EFAULT; | |
ac46d4f3 | 1053 | struct mmu_notifier_range range; |
6c287605 | 1054 | bool anon_exclusive; |
31dbd01f | 1055 | |
36eaff33 KS |
1056 | pvmw.address = page_address_in_vma(page, vma); |
1057 | if (pvmw.address == -EFAULT) | |
31dbd01f IE |
1058 | goto out; |
1059 | ||
29ad768c | 1060 | BUG_ON(PageTransCompound(page)); |
6bdb913f | 1061 | |
7269f999 | 1062 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, |
6f4f13e8 | 1063 | pvmw.address, |
ac46d4f3 JG |
1064 | pvmw.address + PAGE_SIZE); |
1065 | mmu_notifier_invalidate_range_start(&range); | |
6bdb913f | 1066 | |
36eaff33 | 1067 | if (!page_vma_mapped_walk(&pvmw)) |
6bdb913f | 1068 | goto out_mn; |
36eaff33 KS |
1069 | if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?")) |
1070 | goto out_unlock; | |
31dbd01f | 1071 | |
6c287605 | 1072 | anon_exclusive = PageAnonExclusive(page); |
595cd8f2 | 1073 | if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) || |
b3a81d08 | 1074 | (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte)) || |
6c287605 | 1075 | anon_exclusive || mm_tlb_flush_pending(mm)) { |
31dbd01f IE |
1076 | pte_t entry; |
1077 | ||
1078 | swapped = PageSwapCache(page); | |
36eaff33 | 1079 | flush_cache_page(vma, pvmw.address, page_to_pfn(page)); |
31dbd01f | 1080 | /* |
25985edc | 1081 | * Ok this is tricky, when get_user_pages_fast() run it doesn't |
31dbd01f | 1082 | * take any lock, therefore the check that we are going to make |
f0953a1b | 1083 | * with the pagecount against the mapcount is racy and |
31dbd01f IE |
1084 | * O_DIRECT can happen right after the check. |
1085 | * So we clear the pte and flush the tlb before the check | |
1086 | * this assure us that no O_DIRECT can happen after the check | |
1087 | * or in the middle of the check. | |
0f10851e JG |
1088 | * |
1089 | * No need to notify as we are downgrading page table to read | |
1090 | * only not changing it to point to a new page. | |
1091 | * | |
ee65728e | 1092 | * See Documentation/mm/mmu_notifier.rst |
31dbd01f | 1093 | */ |
0f10851e | 1094 | entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte); |
31dbd01f IE |
1095 | /* |
1096 | * Check that no O_DIRECT or similar I/O is in progress on the | |
1097 | * page | |
1098 | */ | |
31e855ea | 1099 | if (page_mapcount(page) + 1 + swapped != page_count(page)) { |
36eaff33 | 1100 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f IE |
1101 | goto out_unlock; |
1102 | } | |
6c287605 | 1103 | |
088b8aa5 | 1104 | /* See page_try_share_anon_rmap(): clear PTE first. */ |
6c287605 DH |
1105 | if (anon_exclusive && page_try_share_anon_rmap(page)) { |
1106 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); | |
1107 | goto out_unlock; | |
1108 | } | |
1109 | ||
4e31635c HD |
1110 | if (pte_dirty(entry)) |
1111 | set_page_dirty(page); | |
595cd8f2 AK |
1112 | |
1113 | if (pte_protnone(entry)) | |
1114 | entry = pte_mkclean(pte_clear_savedwrite(entry)); | |
1115 | else | |
1116 | entry = pte_mkclean(pte_wrprotect(entry)); | |
36eaff33 | 1117 | set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f | 1118 | } |
36eaff33 | 1119 | *orig_pte = *pvmw.pte; |
31dbd01f IE |
1120 | err = 0; |
1121 | ||
1122 | out_unlock: | |
36eaff33 | 1123 | page_vma_mapped_walk_done(&pvmw); |
6bdb913f | 1124 | out_mn: |
ac46d4f3 | 1125 | mmu_notifier_invalidate_range_end(&range); |
31dbd01f IE |
1126 | out: |
1127 | return err; | |
1128 | } | |
1129 | ||
1130 | /** | |
1131 | * replace_page - replace page in vma by new ksm page | |
8dd3557a HD |
1132 | * @vma: vma that holds the pte pointing to page |
1133 | * @page: the page we are replacing by kpage | |
1134 | * @kpage: the ksm page we replace page by | |
31dbd01f IE |
1135 | * @orig_pte: the original value of the pte |
1136 | * | |
1137 | * Returns 0 on success, -EFAULT on failure. | |
1138 | */ | |
8dd3557a HD |
1139 | static int replace_page(struct vm_area_struct *vma, struct page *page, |
1140 | struct page *kpage, pte_t orig_pte) | |
31dbd01f IE |
1141 | { |
1142 | struct mm_struct *mm = vma->vm_mm; | |
31dbd01f | 1143 | pmd_t *pmd; |
50722804 | 1144 | pmd_t pmde; |
31dbd01f | 1145 | pte_t *ptep; |
e86c59b1 | 1146 | pte_t newpte; |
31dbd01f IE |
1147 | spinlock_t *ptl; |
1148 | unsigned long addr; | |
31dbd01f | 1149 | int err = -EFAULT; |
ac46d4f3 | 1150 | struct mmu_notifier_range range; |
31dbd01f | 1151 | |
8dd3557a | 1152 | addr = page_address_in_vma(page, vma); |
31dbd01f IE |
1153 | if (addr == -EFAULT) |
1154 | goto out; | |
1155 | ||
6219049a BL |
1156 | pmd = mm_find_pmd(mm, addr); |
1157 | if (!pmd) | |
31dbd01f | 1158 | goto out; |
50722804 ZK |
1159 | /* |
1160 | * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() | |
1161 | * without holding anon_vma lock for write. So when looking for a | |
1162 | * genuine pmde (in which to find pte), test present and !THP together. | |
1163 | */ | |
1164 | pmde = *pmd; | |
1165 | barrier(); | |
1166 | if (!pmd_present(pmde) || pmd_trans_huge(pmde)) | |
1167 | goto out; | |
31dbd01f | 1168 | |
7269f999 | 1169 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr, |
6f4f13e8 | 1170 | addr + PAGE_SIZE); |
ac46d4f3 | 1171 | mmu_notifier_invalidate_range_start(&range); |
6bdb913f | 1172 | |
31dbd01f IE |
1173 | ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); |
1174 | if (!pte_same(*ptep, orig_pte)) { | |
1175 | pte_unmap_unlock(ptep, ptl); | |
6bdb913f | 1176 | goto out_mn; |
31dbd01f | 1177 | } |
6c287605 DH |
1178 | VM_BUG_ON_PAGE(PageAnonExclusive(page), page); |
1179 | VM_BUG_ON_PAGE(PageAnon(kpage) && PageAnonExclusive(kpage), kpage); | |
31dbd01f | 1180 | |
e86c59b1 CI |
1181 | /* |
1182 | * No need to check ksm_use_zero_pages here: we can only have a | |
457aef94 | 1183 | * zero_page here if ksm_use_zero_pages was enabled already. |
e86c59b1 CI |
1184 | */ |
1185 | if (!is_zero_pfn(page_to_pfn(kpage))) { | |
1186 | get_page(kpage); | |
f1e2db12 | 1187 | page_add_anon_rmap(kpage, vma, addr, RMAP_NONE); |
e86c59b1 CI |
1188 | newpte = mk_pte(kpage, vma->vm_page_prot); |
1189 | } else { | |
1190 | newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage), | |
1191 | vma->vm_page_prot)); | |
a38c015f CI |
1192 | /* |
1193 | * We're replacing an anonymous page with a zero page, which is | |
1194 | * not anonymous. We need to do proper accounting otherwise we | |
1195 | * will get wrong values in /proc, and a BUG message in dmesg | |
1196 | * when tearing down the mm. | |
1197 | */ | |
1198 | dec_mm_counter(mm, MM_ANONPAGES); | |
e86c59b1 | 1199 | } |
31dbd01f IE |
1200 | |
1201 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
0f10851e JG |
1202 | /* |
1203 | * No need to notify as we are replacing a read only page with another | |
1204 | * read only page with the same content. | |
1205 | * | |
ee65728e | 1206 | * See Documentation/mm/mmu_notifier.rst |
0f10851e JG |
1207 | */ |
1208 | ptep_clear_flush(vma, addr, ptep); | |
e86c59b1 | 1209 | set_pte_at_notify(mm, addr, ptep, newpte); |
31dbd01f | 1210 | |
cea86fe2 | 1211 | page_remove_rmap(page, vma, false); |
ae52a2ad HD |
1212 | if (!page_mapped(page)) |
1213 | try_to_free_swap(page); | |
8dd3557a | 1214 | put_page(page); |
31dbd01f IE |
1215 | |
1216 | pte_unmap_unlock(ptep, ptl); | |
1217 | err = 0; | |
6bdb913f | 1218 | out_mn: |
ac46d4f3 | 1219 | mmu_notifier_invalidate_range_end(&range); |
31dbd01f IE |
1220 | out: |
1221 | return err; | |
1222 | } | |
1223 | ||
1224 | /* | |
1225 | * try_to_merge_one_page - take two pages and merge them into one | |
8dd3557a HD |
1226 | * @vma: the vma that holds the pte pointing to page |
1227 | * @page: the PageAnon page that we want to replace with kpage | |
80e14822 HD |
1228 | * @kpage: the PageKsm page that we want to map instead of page, |
1229 | * or NULL the first time when we want to use page as kpage. | |
31dbd01f IE |
1230 | * |
1231 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
1232 | */ | |
1233 | static int try_to_merge_one_page(struct vm_area_struct *vma, | |
8dd3557a | 1234 | struct page *page, struct page *kpage) |
31dbd01f IE |
1235 | { |
1236 | pte_t orig_pte = __pte(0); | |
1237 | int err = -EFAULT; | |
1238 | ||
db114b83 HD |
1239 | if (page == kpage) /* ksm page forked */ |
1240 | return 0; | |
1241 | ||
8dd3557a | 1242 | if (!PageAnon(page)) |
31dbd01f IE |
1243 | goto out; |
1244 | ||
31dbd01f IE |
1245 | /* |
1246 | * We need the page lock to read a stable PageSwapCache in | |
1247 | * write_protect_page(). We use trylock_page() instead of | |
1248 | * lock_page() because we don't want to wait here - we | |
1249 | * prefer to continue scanning and merging different pages, | |
1250 | * then come back to this page when it is unlocked. | |
1251 | */ | |
8dd3557a | 1252 | if (!trylock_page(page)) |
31e855ea | 1253 | goto out; |
f765f540 KS |
1254 | |
1255 | if (PageTransCompound(page)) { | |
a7306c34 | 1256 | if (split_huge_page(page)) |
f765f540 KS |
1257 | goto out_unlock; |
1258 | } | |
1259 | ||
31dbd01f IE |
1260 | /* |
1261 | * If this anonymous page is mapped only here, its pte may need | |
1262 | * to be write-protected. If it's mapped elsewhere, all of its | |
1263 | * ptes are necessarily already write-protected. But in either | |
1264 | * case, we need to lock and check page_count is not raised. | |
1265 | */ | |
80e14822 HD |
1266 | if (write_protect_page(vma, page, &orig_pte) == 0) { |
1267 | if (!kpage) { | |
1268 | /* | |
1269 | * While we hold page lock, upgrade page from | |
1270 | * PageAnon+anon_vma to PageKsm+NULL stable_node: | |
1271 | * stable_tree_insert() will update stable_node. | |
1272 | */ | |
1273 | set_page_stable_node(page, NULL); | |
1274 | mark_page_accessed(page); | |
337ed7eb MK |
1275 | /* |
1276 | * Page reclaim just frees a clean page with no dirty | |
1277 | * ptes: make sure that the ksm page would be swapped. | |
1278 | */ | |
1279 | if (!PageDirty(page)) | |
1280 | SetPageDirty(page); | |
80e14822 HD |
1281 | err = 0; |
1282 | } else if (pages_identical(page, kpage)) | |
1283 | err = replace_page(vma, page, kpage, orig_pte); | |
1284 | } | |
31dbd01f | 1285 | |
f765f540 | 1286 | out_unlock: |
8dd3557a | 1287 | unlock_page(page); |
31dbd01f IE |
1288 | out: |
1289 | return err; | |
1290 | } | |
1291 | ||
81464e30 HD |
1292 | /* |
1293 | * try_to_merge_with_ksm_page - like try_to_merge_two_pages, | |
1294 | * but no new kernel page is allocated: kpage must already be a ksm page. | |
8dd3557a HD |
1295 | * |
1296 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
81464e30 | 1297 | */ |
21fbd591 | 1298 | static int try_to_merge_with_ksm_page(struct ksm_rmap_item *rmap_item, |
8dd3557a | 1299 | struct page *page, struct page *kpage) |
81464e30 | 1300 | { |
8dd3557a | 1301 | struct mm_struct *mm = rmap_item->mm; |
81464e30 HD |
1302 | struct vm_area_struct *vma; |
1303 | int err = -EFAULT; | |
1304 | ||
d8ed45c5 | 1305 | mmap_read_lock(mm); |
85c6e8dd AA |
1306 | vma = find_mergeable_vma(mm, rmap_item->address); |
1307 | if (!vma) | |
81464e30 HD |
1308 | goto out; |
1309 | ||
8dd3557a | 1310 | err = try_to_merge_one_page(vma, page, kpage); |
db114b83 HD |
1311 | if (err) |
1312 | goto out; | |
1313 | ||
bc56620b HD |
1314 | /* Unstable nid is in union with stable anon_vma: remove first */ |
1315 | remove_rmap_item_from_tree(rmap_item); | |
1316 | ||
c1e8d7c6 | 1317 | /* Must get reference to anon_vma while still holding mmap_lock */ |
9e60109f PZ |
1318 | rmap_item->anon_vma = vma->anon_vma; |
1319 | get_anon_vma(vma->anon_vma); | |
81464e30 | 1320 | out: |
d8ed45c5 | 1321 | mmap_read_unlock(mm); |
81464e30 HD |
1322 | return err; |
1323 | } | |
1324 | ||
31dbd01f IE |
1325 | /* |
1326 | * try_to_merge_two_pages - take two identical pages and prepare them | |
1327 | * to be merged into one page. | |
1328 | * | |
8dd3557a HD |
1329 | * This function returns the kpage if we successfully merged two identical |
1330 | * pages into one ksm page, NULL otherwise. | |
31dbd01f | 1331 | * |
80e14822 | 1332 | * Note that this function upgrades page to ksm page: if one of the pages |
31dbd01f IE |
1333 | * is already a ksm page, try_to_merge_with_ksm_page should be used. |
1334 | */ | |
21fbd591 | 1335 | static struct page *try_to_merge_two_pages(struct ksm_rmap_item *rmap_item, |
8dd3557a | 1336 | struct page *page, |
21fbd591 | 1337 | struct ksm_rmap_item *tree_rmap_item, |
8dd3557a | 1338 | struct page *tree_page) |
31dbd01f | 1339 | { |
80e14822 | 1340 | int err; |
31dbd01f | 1341 | |
80e14822 | 1342 | err = try_to_merge_with_ksm_page(rmap_item, page, NULL); |
31dbd01f | 1343 | if (!err) { |
8dd3557a | 1344 | err = try_to_merge_with_ksm_page(tree_rmap_item, |
80e14822 | 1345 | tree_page, page); |
31dbd01f | 1346 | /* |
81464e30 HD |
1347 | * If that fails, we have a ksm page with only one pte |
1348 | * pointing to it: so break it. | |
31dbd01f | 1349 | */ |
4035c07a | 1350 | if (err) |
8dd3557a | 1351 | break_cow(rmap_item); |
31dbd01f | 1352 | } |
80e14822 | 1353 | return err ? NULL : page; |
31dbd01f IE |
1354 | } |
1355 | ||
2c653d0e | 1356 | static __always_inline |
21fbd591 | 1357 | bool __is_page_sharing_candidate(struct ksm_stable_node *stable_node, int offset) |
2c653d0e AA |
1358 | { |
1359 | VM_BUG_ON(stable_node->rmap_hlist_len < 0); | |
1360 | /* | |
1361 | * Check that at least one mapping still exists, otherwise | |
1362 | * there's no much point to merge and share with this | |
1363 | * stable_node, as the underlying tree_page of the other | |
1364 | * sharer is going to be freed soon. | |
1365 | */ | |
1366 | return stable_node->rmap_hlist_len && | |
1367 | stable_node->rmap_hlist_len + offset < ksm_max_page_sharing; | |
1368 | } | |
1369 | ||
1370 | static __always_inline | |
21fbd591 | 1371 | bool is_page_sharing_candidate(struct ksm_stable_node *stable_node) |
2c653d0e AA |
1372 | { |
1373 | return __is_page_sharing_candidate(stable_node, 0); | |
1374 | } | |
1375 | ||
21fbd591 QZ |
1376 | static struct page *stable_node_dup(struct ksm_stable_node **_stable_node_dup, |
1377 | struct ksm_stable_node **_stable_node, | |
c01f0b54 CIK |
1378 | struct rb_root *root, |
1379 | bool prune_stale_stable_nodes) | |
2c653d0e | 1380 | { |
21fbd591 | 1381 | struct ksm_stable_node *dup, *found = NULL, *stable_node = *_stable_node; |
2c653d0e | 1382 | struct hlist_node *hlist_safe; |
8dc5ffcd | 1383 | struct page *_tree_page, *tree_page = NULL; |
2c653d0e AA |
1384 | int nr = 0; |
1385 | int found_rmap_hlist_len; | |
1386 | ||
1387 | if (!prune_stale_stable_nodes || | |
1388 | time_before(jiffies, stable_node->chain_prune_time + | |
1389 | msecs_to_jiffies( | |
1390 | ksm_stable_node_chains_prune_millisecs))) | |
1391 | prune_stale_stable_nodes = false; | |
1392 | else | |
1393 | stable_node->chain_prune_time = jiffies; | |
1394 | ||
1395 | hlist_for_each_entry_safe(dup, hlist_safe, | |
1396 | &stable_node->hlist, hlist_dup) { | |
1397 | cond_resched(); | |
1398 | /* | |
1399 | * We must walk all stable_node_dup to prune the stale | |
1400 | * stable nodes during lookup. | |
1401 | * | |
1402 | * get_ksm_page can drop the nodes from the | |
1403 | * stable_node->hlist if they point to freed pages | |
1404 | * (that's why we do a _safe walk). The "dup" | |
1405 | * stable_node parameter itself will be freed from | |
1406 | * under us if it returns NULL. | |
1407 | */ | |
2cee57d1 | 1408 | _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK); |
2c653d0e AA |
1409 | if (!_tree_page) |
1410 | continue; | |
1411 | nr += 1; | |
1412 | if (is_page_sharing_candidate(dup)) { | |
1413 | if (!found || | |
1414 | dup->rmap_hlist_len > found_rmap_hlist_len) { | |
1415 | if (found) | |
8dc5ffcd | 1416 | put_page(tree_page); |
2c653d0e AA |
1417 | found = dup; |
1418 | found_rmap_hlist_len = found->rmap_hlist_len; | |
8dc5ffcd | 1419 | tree_page = _tree_page; |
2c653d0e | 1420 | |
8dc5ffcd | 1421 | /* skip put_page for found dup */ |
2c653d0e AA |
1422 | if (!prune_stale_stable_nodes) |
1423 | break; | |
2c653d0e AA |
1424 | continue; |
1425 | } | |
1426 | } | |
1427 | put_page(_tree_page); | |
1428 | } | |
1429 | ||
80b18dfa AA |
1430 | if (found) { |
1431 | /* | |
1432 | * nr is counting all dups in the chain only if | |
1433 | * prune_stale_stable_nodes is true, otherwise we may | |
1434 | * break the loop at nr == 1 even if there are | |
1435 | * multiple entries. | |
1436 | */ | |
1437 | if (prune_stale_stable_nodes && nr == 1) { | |
2c653d0e AA |
1438 | /* |
1439 | * If there's not just one entry it would | |
1440 | * corrupt memory, better BUG_ON. In KSM | |
1441 | * context with no lock held it's not even | |
1442 | * fatal. | |
1443 | */ | |
1444 | BUG_ON(stable_node->hlist.first->next); | |
1445 | ||
1446 | /* | |
1447 | * There's just one entry and it is below the | |
1448 | * deduplication limit so drop the chain. | |
1449 | */ | |
1450 | rb_replace_node(&stable_node->node, &found->node, | |
1451 | root); | |
1452 | free_stable_node(stable_node); | |
1453 | ksm_stable_node_chains--; | |
1454 | ksm_stable_node_dups--; | |
b4fecc67 | 1455 | /* |
0ba1d0f7 AA |
1456 | * NOTE: the caller depends on the stable_node |
1457 | * to be equal to stable_node_dup if the chain | |
1458 | * was collapsed. | |
b4fecc67 | 1459 | */ |
0ba1d0f7 AA |
1460 | *_stable_node = found; |
1461 | /* | |
f0953a1b | 1462 | * Just for robustness, as stable_node is |
0ba1d0f7 AA |
1463 | * otherwise left as a stable pointer, the |
1464 | * compiler shall optimize it away at build | |
1465 | * time. | |
1466 | */ | |
1467 | stable_node = NULL; | |
80b18dfa AA |
1468 | } else if (stable_node->hlist.first != &found->hlist_dup && |
1469 | __is_page_sharing_candidate(found, 1)) { | |
2c653d0e | 1470 | /* |
80b18dfa AA |
1471 | * If the found stable_node dup can accept one |
1472 | * more future merge (in addition to the one | |
1473 | * that is underway) and is not at the head of | |
1474 | * the chain, put it there so next search will | |
1475 | * be quicker in the !prune_stale_stable_nodes | |
1476 | * case. | |
1477 | * | |
1478 | * NOTE: it would be inaccurate to use nr > 1 | |
1479 | * instead of checking the hlist.first pointer | |
1480 | * directly, because in the | |
1481 | * prune_stale_stable_nodes case "nr" isn't | |
1482 | * the position of the found dup in the chain, | |
1483 | * but the total number of dups in the chain. | |
2c653d0e AA |
1484 | */ |
1485 | hlist_del(&found->hlist_dup); | |
1486 | hlist_add_head(&found->hlist_dup, | |
1487 | &stable_node->hlist); | |
1488 | } | |
1489 | } | |
1490 | ||
8dc5ffcd AA |
1491 | *_stable_node_dup = found; |
1492 | return tree_page; | |
2c653d0e AA |
1493 | } |
1494 | ||
21fbd591 | 1495 | static struct ksm_stable_node *stable_node_dup_any(struct ksm_stable_node *stable_node, |
2c653d0e AA |
1496 | struct rb_root *root) |
1497 | { | |
1498 | if (!is_stable_node_chain(stable_node)) | |
1499 | return stable_node; | |
1500 | if (hlist_empty(&stable_node->hlist)) { | |
1501 | free_stable_node_chain(stable_node, root); | |
1502 | return NULL; | |
1503 | } | |
1504 | return hlist_entry(stable_node->hlist.first, | |
1505 | typeof(*stable_node), hlist_dup); | |
1506 | } | |
1507 | ||
8dc5ffcd AA |
1508 | /* |
1509 | * Like for get_ksm_page, this function can free the *_stable_node and | |
1510 | * *_stable_node_dup if the returned tree_page is NULL. | |
1511 | * | |
1512 | * It can also free and overwrite *_stable_node with the found | |
1513 | * stable_node_dup if the chain is collapsed (in which case | |
1514 | * *_stable_node will be equal to *_stable_node_dup like if the chain | |
1515 | * never existed). It's up to the caller to verify tree_page is not | |
1516 | * NULL before dereferencing *_stable_node or *_stable_node_dup. | |
1517 | * | |
1518 | * *_stable_node_dup is really a second output parameter of this | |
1519 | * function and will be overwritten in all cases, the caller doesn't | |
1520 | * need to initialize it. | |
1521 | */ | |
21fbd591 QZ |
1522 | static struct page *__stable_node_chain(struct ksm_stable_node **_stable_node_dup, |
1523 | struct ksm_stable_node **_stable_node, | |
8dc5ffcd AA |
1524 | struct rb_root *root, |
1525 | bool prune_stale_stable_nodes) | |
2c653d0e | 1526 | { |
21fbd591 | 1527 | struct ksm_stable_node *stable_node = *_stable_node; |
2c653d0e AA |
1528 | if (!is_stable_node_chain(stable_node)) { |
1529 | if (is_page_sharing_candidate(stable_node)) { | |
8dc5ffcd | 1530 | *_stable_node_dup = stable_node; |
2cee57d1 | 1531 | return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK); |
2c653d0e | 1532 | } |
8dc5ffcd AA |
1533 | /* |
1534 | * _stable_node_dup set to NULL means the stable_node | |
1535 | * reached the ksm_max_page_sharing limit. | |
1536 | */ | |
1537 | *_stable_node_dup = NULL; | |
2c653d0e AA |
1538 | return NULL; |
1539 | } | |
8dc5ffcd | 1540 | return stable_node_dup(_stable_node_dup, _stable_node, root, |
2c653d0e AA |
1541 | prune_stale_stable_nodes); |
1542 | } | |
1543 | ||
21fbd591 QZ |
1544 | static __always_inline struct page *chain_prune(struct ksm_stable_node **s_n_d, |
1545 | struct ksm_stable_node **s_n, | |
8dc5ffcd | 1546 | struct rb_root *root) |
2c653d0e | 1547 | { |
8dc5ffcd | 1548 | return __stable_node_chain(s_n_d, s_n, root, true); |
2c653d0e AA |
1549 | } |
1550 | ||
21fbd591 QZ |
1551 | static __always_inline struct page *chain(struct ksm_stable_node **s_n_d, |
1552 | struct ksm_stable_node *s_n, | |
8dc5ffcd | 1553 | struct rb_root *root) |
2c653d0e | 1554 | { |
21fbd591 | 1555 | struct ksm_stable_node *old_stable_node = s_n; |
8dc5ffcd AA |
1556 | struct page *tree_page; |
1557 | ||
1558 | tree_page = __stable_node_chain(s_n_d, &s_n, root, false); | |
1559 | /* not pruning dups so s_n cannot have changed */ | |
1560 | VM_BUG_ON(s_n != old_stable_node); | |
1561 | return tree_page; | |
2c653d0e AA |
1562 | } |
1563 | ||
31dbd01f | 1564 | /* |
8dd3557a | 1565 | * stable_tree_search - search for page inside the stable tree |
31dbd01f IE |
1566 | * |
1567 | * This function checks if there is a page inside the stable tree | |
1568 | * with identical content to the page that we are scanning right now. | |
1569 | * | |
7b6ba2c7 | 1570 | * This function returns the stable tree node of identical content if found, |
31dbd01f IE |
1571 | * NULL otherwise. |
1572 | */ | |
62b61f61 | 1573 | static struct page *stable_tree_search(struct page *page) |
31dbd01f | 1574 | { |
90bd6fd3 | 1575 | int nid; |
ef53d16c | 1576 | struct rb_root *root; |
4146d2d6 HD |
1577 | struct rb_node **new; |
1578 | struct rb_node *parent; | |
21fbd591 QZ |
1579 | struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; |
1580 | struct ksm_stable_node *page_node; | |
31dbd01f | 1581 | |
4146d2d6 HD |
1582 | page_node = page_stable_node(page); |
1583 | if (page_node && page_node->head != &migrate_nodes) { | |
1584 | /* ksm page forked */ | |
08beca44 | 1585 | get_page(page); |
62b61f61 | 1586 | return page; |
08beca44 HD |
1587 | } |
1588 | ||
90bd6fd3 | 1589 | nid = get_kpfn_nid(page_to_pfn(page)); |
ef53d16c | 1590 | root = root_stable_tree + nid; |
4146d2d6 | 1591 | again: |
ef53d16c | 1592 | new = &root->rb_node; |
4146d2d6 | 1593 | parent = NULL; |
90bd6fd3 | 1594 | |
4146d2d6 | 1595 | while (*new) { |
4035c07a | 1596 | struct page *tree_page; |
31dbd01f IE |
1597 | int ret; |
1598 | ||
08beca44 | 1599 | cond_resched(); |
21fbd591 | 1600 | stable_node = rb_entry(*new, struct ksm_stable_node, node); |
2c653d0e | 1601 | stable_node_any = NULL; |
8dc5ffcd | 1602 | tree_page = chain_prune(&stable_node_dup, &stable_node, root); |
b4fecc67 AA |
1603 | /* |
1604 | * NOTE: stable_node may have been freed by | |
1605 | * chain_prune() if the returned stable_node_dup is | |
1606 | * not NULL. stable_node_dup may have been inserted in | |
1607 | * the rbtree instead as a regular stable_node (in | |
1608 | * order to collapse the stable_node chain if a single | |
0ba1d0f7 | 1609 | * stable_node dup was found in it). In such case the |
3413b2c8 | 1610 | * stable_node is overwritten by the callee to point |
0ba1d0f7 AA |
1611 | * to the stable_node_dup that was collapsed in the |
1612 | * stable rbtree and stable_node will be equal to | |
1613 | * stable_node_dup like if the chain never existed. | |
b4fecc67 | 1614 | */ |
2c653d0e AA |
1615 | if (!stable_node_dup) { |
1616 | /* | |
1617 | * Either all stable_node dups were full in | |
1618 | * this stable_node chain, or this chain was | |
1619 | * empty and should be rb_erased. | |
1620 | */ | |
1621 | stable_node_any = stable_node_dup_any(stable_node, | |
1622 | root); | |
1623 | if (!stable_node_any) { | |
1624 | /* rb_erase just run */ | |
1625 | goto again; | |
1626 | } | |
1627 | /* | |
1628 | * Take any of the stable_node dups page of | |
1629 | * this stable_node chain to let the tree walk | |
1630 | * continue. All KSM pages belonging to the | |
1631 | * stable_node dups in a stable_node chain | |
1632 | * have the same content and they're | |
457aef94 | 1633 | * write protected at all times. Any will work |
2c653d0e AA |
1634 | * fine to continue the walk. |
1635 | */ | |
2cee57d1 YS |
1636 | tree_page = get_ksm_page(stable_node_any, |
1637 | GET_KSM_PAGE_NOLOCK); | |
2c653d0e AA |
1638 | } |
1639 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1640 | if (!tree_page) { |
1641 | /* | |
1642 | * If we walked over a stale stable_node, | |
1643 | * get_ksm_page() will call rb_erase() and it | |
1644 | * may rebalance the tree from under us. So | |
1645 | * restart the search from scratch. Returning | |
1646 | * NULL would be safe too, but we'd generate | |
1647 | * false negative insertions just because some | |
1648 | * stable_node was stale. | |
1649 | */ | |
1650 | goto again; | |
1651 | } | |
31dbd01f | 1652 | |
4035c07a | 1653 | ret = memcmp_pages(page, tree_page); |
c8d6553b | 1654 | put_page(tree_page); |
31dbd01f | 1655 | |
4146d2d6 | 1656 | parent = *new; |
c8d6553b | 1657 | if (ret < 0) |
4146d2d6 | 1658 | new = &parent->rb_left; |
c8d6553b | 1659 | else if (ret > 0) |
4146d2d6 | 1660 | new = &parent->rb_right; |
c8d6553b | 1661 | else { |
2c653d0e AA |
1662 | if (page_node) { |
1663 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1664 | /* | |
1665 | * Test if the migrated page should be merged | |
1666 | * into a stable node dup. If the mapcount is | |
1667 | * 1 we can migrate it with another KSM page | |
1668 | * without adding it to the chain. | |
1669 | */ | |
1670 | if (page_mapcount(page) > 1) | |
1671 | goto chain_append; | |
1672 | } | |
1673 | ||
1674 | if (!stable_node_dup) { | |
1675 | /* | |
1676 | * If the stable_node is a chain and | |
1677 | * we got a payload match in memcmp | |
1678 | * but we cannot merge the scanned | |
1679 | * page in any of the existing | |
1680 | * stable_node dups because they're | |
1681 | * all full, we need to wait the | |
1682 | * scanned page to find itself a match | |
1683 | * in the unstable tree to create a | |
1684 | * brand new KSM page to add later to | |
1685 | * the dups of this stable_node. | |
1686 | */ | |
1687 | return NULL; | |
1688 | } | |
1689 | ||
c8d6553b HD |
1690 | /* |
1691 | * Lock and unlock the stable_node's page (which | |
1692 | * might already have been migrated) so that page | |
1693 | * migration is sure to notice its raised count. | |
1694 | * It would be more elegant to return stable_node | |
1695 | * than kpage, but that involves more changes. | |
1696 | */ | |
2cee57d1 YS |
1697 | tree_page = get_ksm_page(stable_node_dup, |
1698 | GET_KSM_PAGE_TRYLOCK); | |
1699 | ||
1700 | if (PTR_ERR(tree_page) == -EBUSY) | |
1701 | return ERR_PTR(-EBUSY); | |
1702 | ||
2c653d0e AA |
1703 | if (unlikely(!tree_page)) |
1704 | /* | |
1705 | * The tree may have been rebalanced, | |
1706 | * so re-evaluate parent and new. | |
1707 | */ | |
4146d2d6 | 1708 | goto again; |
2c653d0e AA |
1709 | unlock_page(tree_page); |
1710 | ||
1711 | if (get_kpfn_nid(stable_node_dup->kpfn) != | |
1712 | NUMA(stable_node_dup->nid)) { | |
1713 | put_page(tree_page); | |
1714 | goto replace; | |
1715 | } | |
1716 | return tree_page; | |
c8d6553b | 1717 | } |
31dbd01f IE |
1718 | } |
1719 | ||
4146d2d6 HD |
1720 | if (!page_node) |
1721 | return NULL; | |
1722 | ||
1723 | list_del(&page_node->list); | |
1724 | DO_NUMA(page_node->nid = nid); | |
1725 | rb_link_node(&page_node->node, parent, new); | |
ef53d16c | 1726 | rb_insert_color(&page_node->node, root); |
2c653d0e AA |
1727 | out: |
1728 | if (is_page_sharing_candidate(page_node)) { | |
1729 | get_page(page); | |
1730 | return page; | |
1731 | } else | |
1732 | return NULL; | |
4146d2d6 HD |
1733 | |
1734 | replace: | |
b4fecc67 AA |
1735 | /* |
1736 | * If stable_node was a chain and chain_prune collapsed it, | |
0ba1d0f7 AA |
1737 | * stable_node has been updated to be the new regular |
1738 | * stable_node. A collapse of the chain is indistinguishable | |
1739 | * from the case there was no chain in the stable | |
1740 | * rbtree. Otherwise stable_node is the chain and | |
1741 | * stable_node_dup is the dup to replace. | |
b4fecc67 | 1742 | */ |
0ba1d0f7 | 1743 | if (stable_node_dup == stable_node) { |
b4fecc67 AA |
1744 | VM_BUG_ON(is_stable_node_chain(stable_node_dup)); |
1745 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); | |
2c653d0e AA |
1746 | /* there is no chain */ |
1747 | if (page_node) { | |
1748 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1749 | list_del(&page_node->list); | |
1750 | DO_NUMA(page_node->nid = nid); | |
b4fecc67 AA |
1751 | rb_replace_node(&stable_node_dup->node, |
1752 | &page_node->node, | |
2c653d0e AA |
1753 | root); |
1754 | if (is_page_sharing_candidate(page_node)) | |
1755 | get_page(page); | |
1756 | else | |
1757 | page = NULL; | |
1758 | } else { | |
b4fecc67 | 1759 | rb_erase(&stable_node_dup->node, root); |
2c653d0e AA |
1760 | page = NULL; |
1761 | } | |
4146d2d6 | 1762 | } else { |
2c653d0e AA |
1763 | VM_BUG_ON(!is_stable_node_chain(stable_node)); |
1764 | __stable_node_dup_del(stable_node_dup); | |
1765 | if (page_node) { | |
1766 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1767 | list_del(&page_node->list); | |
1768 | DO_NUMA(page_node->nid = nid); | |
1769 | stable_node_chain_add_dup(page_node, stable_node); | |
1770 | if (is_page_sharing_candidate(page_node)) | |
1771 | get_page(page); | |
1772 | else | |
1773 | page = NULL; | |
1774 | } else { | |
1775 | page = NULL; | |
1776 | } | |
4146d2d6 | 1777 | } |
2c653d0e AA |
1778 | stable_node_dup->head = &migrate_nodes; |
1779 | list_add(&stable_node_dup->list, stable_node_dup->head); | |
4146d2d6 | 1780 | return page; |
2c653d0e AA |
1781 | |
1782 | chain_append: | |
1783 | /* stable_node_dup could be null if it reached the limit */ | |
1784 | if (!stable_node_dup) | |
1785 | stable_node_dup = stable_node_any; | |
b4fecc67 AA |
1786 | /* |
1787 | * If stable_node was a chain and chain_prune collapsed it, | |
0ba1d0f7 AA |
1788 | * stable_node has been updated to be the new regular |
1789 | * stable_node. A collapse of the chain is indistinguishable | |
1790 | * from the case there was no chain in the stable | |
1791 | * rbtree. Otherwise stable_node is the chain and | |
1792 | * stable_node_dup is the dup to replace. | |
b4fecc67 | 1793 | */ |
0ba1d0f7 | 1794 | if (stable_node_dup == stable_node) { |
b4fecc67 | 1795 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); |
2c653d0e AA |
1796 | /* chain is missing so create it */ |
1797 | stable_node = alloc_stable_node_chain(stable_node_dup, | |
1798 | root); | |
1799 | if (!stable_node) | |
1800 | return NULL; | |
1801 | } | |
1802 | /* | |
1803 | * Add this stable_node dup that was | |
1804 | * migrated to the stable_node chain | |
1805 | * of the current nid for this page | |
1806 | * content. | |
1807 | */ | |
b4fecc67 | 1808 | VM_BUG_ON(!is_stable_node_dup(stable_node_dup)); |
2c653d0e AA |
1809 | VM_BUG_ON(page_node->head != &migrate_nodes); |
1810 | list_del(&page_node->list); | |
1811 | DO_NUMA(page_node->nid = nid); | |
1812 | stable_node_chain_add_dup(page_node, stable_node); | |
1813 | goto out; | |
31dbd01f IE |
1814 | } |
1815 | ||
1816 | /* | |
e850dcf5 | 1817 | * stable_tree_insert - insert stable tree node pointing to new ksm page |
31dbd01f IE |
1818 | * into the stable tree. |
1819 | * | |
7b6ba2c7 HD |
1820 | * This function returns the stable tree node just allocated on success, |
1821 | * NULL otherwise. | |
31dbd01f | 1822 | */ |
21fbd591 | 1823 | static struct ksm_stable_node *stable_tree_insert(struct page *kpage) |
31dbd01f | 1824 | { |
90bd6fd3 PH |
1825 | int nid; |
1826 | unsigned long kpfn; | |
ef53d16c | 1827 | struct rb_root *root; |
90bd6fd3 | 1828 | struct rb_node **new; |
f2e5ff85 | 1829 | struct rb_node *parent; |
21fbd591 | 1830 | struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; |
2c653d0e | 1831 | bool need_chain = false; |
31dbd01f | 1832 | |
90bd6fd3 PH |
1833 | kpfn = page_to_pfn(kpage); |
1834 | nid = get_kpfn_nid(kpfn); | |
ef53d16c | 1835 | root = root_stable_tree + nid; |
f2e5ff85 AA |
1836 | again: |
1837 | parent = NULL; | |
ef53d16c | 1838 | new = &root->rb_node; |
90bd6fd3 | 1839 | |
31dbd01f | 1840 | while (*new) { |
4035c07a | 1841 | struct page *tree_page; |
31dbd01f IE |
1842 | int ret; |
1843 | ||
08beca44 | 1844 | cond_resched(); |
21fbd591 | 1845 | stable_node = rb_entry(*new, struct ksm_stable_node, node); |
2c653d0e | 1846 | stable_node_any = NULL; |
8dc5ffcd | 1847 | tree_page = chain(&stable_node_dup, stable_node, root); |
2c653d0e AA |
1848 | if (!stable_node_dup) { |
1849 | /* | |
1850 | * Either all stable_node dups were full in | |
1851 | * this stable_node chain, or this chain was | |
1852 | * empty and should be rb_erased. | |
1853 | */ | |
1854 | stable_node_any = stable_node_dup_any(stable_node, | |
1855 | root); | |
1856 | if (!stable_node_any) { | |
1857 | /* rb_erase just run */ | |
1858 | goto again; | |
1859 | } | |
1860 | /* | |
1861 | * Take any of the stable_node dups page of | |
1862 | * this stable_node chain to let the tree walk | |
1863 | * continue. All KSM pages belonging to the | |
1864 | * stable_node dups in a stable_node chain | |
1865 | * have the same content and they're | |
457aef94 | 1866 | * write protected at all times. Any will work |
2c653d0e AA |
1867 | * fine to continue the walk. |
1868 | */ | |
2cee57d1 YS |
1869 | tree_page = get_ksm_page(stable_node_any, |
1870 | GET_KSM_PAGE_NOLOCK); | |
2c653d0e AA |
1871 | } |
1872 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1873 | if (!tree_page) { |
1874 | /* | |
1875 | * If we walked over a stale stable_node, | |
1876 | * get_ksm_page() will call rb_erase() and it | |
1877 | * may rebalance the tree from under us. So | |
1878 | * restart the search from scratch. Returning | |
1879 | * NULL would be safe too, but we'd generate | |
1880 | * false negative insertions just because some | |
1881 | * stable_node was stale. | |
1882 | */ | |
1883 | goto again; | |
1884 | } | |
31dbd01f | 1885 | |
4035c07a HD |
1886 | ret = memcmp_pages(kpage, tree_page); |
1887 | put_page(tree_page); | |
31dbd01f IE |
1888 | |
1889 | parent = *new; | |
1890 | if (ret < 0) | |
1891 | new = &parent->rb_left; | |
1892 | else if (ret > 0) | |
1893 | new = &parent->rb_right; | |
1894 | else { | |
2c653d0e AA |
1895 | need_chain = true; |
1896 | break; | |
31dbd01f IE |
1897 | } |
1898 | } | |
1899 | ||
2c653d0e AA |
1900 | stable_node_dup = alloc_stable_node(); |
1901 | if (!stable_node_dup) | |
7b6ba2c7 | 1902 | return NULL; |
31dbd01f | 1903 | |
2c653d0e AA |
1904 | INIT_HLIST_HEAD(&stable_node_dup->hlist); |
1905 | stable_node_dup->kpfn = kpfn; | |
1906 | set_page_stable_node(kpage, stable_node_dup); | |
1907 | stable_node_dup->rmap_hlist_len = 0; | |
1908 | DO_NUMA(stable_node_dup->nid = nid); | |
1909 | if (!need_chain) { | |
1910 | rb_link_node(&stable_node_dup->node, parent, new); | |
1911 | rb_insert_color(&stable_node_dup->node, root); | |
1912 | } else { | |
1913 | if (!is_stable_node_chain(stable_node)) { | |
21fbd591 | 1914 | struct ksm_stable_node *orig = stable_node; |
2c653d0e AA |
1915 | /* chain is missing so create it */ |
1916 | stable_node = alloc_stable_node_chain(orig, root); | |
1917 | if (!stable_node) { | |
1918 | free_stable_node(stable_node_dup); | |
1919 | return NULL; | |
1920 | } | |
1921 | } | |
1922 | stable_node_chain_add_dup(stable_node_dup, stable_node); | |
1923 | } | |
08beca44 | 1924 | |
2c653d0e | 1925 | return stable_node_dup; |
31dbd01f IE |
1926 | } |
1927 | ||
1928 | /* | |
8dd3557a HD |
1929 | * unstable_tree_search_insert - search for identical page, |
1930 | * else insert rmap_item into the unstable tree. | |
31dbd01f IE |
1931 | * |
1932 | * This function searches for a page in the unstable tree identical to the | |
1933 | * page currently being scanned; and if no identical page is found in the | |
1934 | * tree, we insert rmap_item as a new object into the unstable tree. | |
1935 | * | |
1936 | * This function returns pointer to rmap_item found to be identical | |
1937 | * to the currently scanned page, NULL otherwise. | |
1938 | * | |
1939 | * This function does both searching and inserting, because they share | |
1940 | * the same walking algorithm in an rbtree. | |
1941 | */ | |
8dd3557a | 1942 | static |
21fbd591 | 1943 | struct ksm_rmap_item *unstable_tree_search_insert(struct ksm_rmap_item *rmap_item, |
8dd3557a HD |
1944 | struct page *page, |
1945 | struct page **tree_pagep) | |
31dbd01f | 1946 | { |
90bd6fd3 PH |
1947 | struct rb_node **new; |
1948 | struct rb_root *root; | |
31dbd01f | 1949 | struct rb_node *parent = NULL; |
90bd6fd3 PH |
1950 | int nid; |
1951 | ||
1952 | nid = get_kpfn_nid(page_to_pfn(page)); | |
ef53d16c | 1953 | root = root_unstable_tree + nid; |
90bd6fd3 | 1954 | new = &root->rb_node; |
31dbd01f IE |
1955 | |
1956 | while (*new) { | |
21fbd591 | 1957 | struct ksm_rmap_item *tree_rmap_item; |
8dd3557a | 1958 | struct page *tree_page; |
31dbd01f IE |
1959 | int ret; |
1960 | ||
d178f27f | 1961 | cond_resched(); |
21fbd591 | 1962 | tree_rmap_item = rb_entry(*new, struct ksm_rmap_item, node); |
8dd3557a | 1963 | tree_page = get_mergeable_page(tree_rmap_item); |
c8f95ed1 | 1964 | if (!tree_page) |
31dbd01f IE |
1965 | return NULL; |
1966 | ||
1967 | /* | |
8dd3557a | 1968 | * Don't substitute a ksm page for a forked page. |
31dbd01f | 1969 | */ |
8dd3557a HD |
1970 | if (page == tree_page) { |
1971 | put_page(tree_page); | |
31dbd01f IE |
1972 | return NULL; |
1973 | } | |
1974 | ||
8dd3557a | 1975 | ret = memcmp_pages(page, tree_page); |
31dbd01f IE |
1976 | |
1977 | parent = *new; | |
1978 | if (ret < 0) { | |
8dd3557a | 1979 | put_page(tree_page); |
31dbd01f IE |
1980 | new = &parent->rb_left; |
1981 | } else if (ret > 0) { | |
8dd3557a | 1982 | put_page(tree_page); |
31dbd01f | 1983 | new = &parent->rb_right; |
b599cbdf HD |
1984 | } else if (!ksm_merge_across_nodes && |
1985 | page_to_nid(tree_page) != nid) { | |
1986 | /* | |
1987 | * If tree_page has been migrated to another NUMA node, | |
1988 | * it will be flushed out and put in the right unstable | |
1989 | * tree next time: only merge with it when across_nodes. | |
1990 | */ | |
1991 | put_page(tree_page); | |
1992 | return NULL; | |
31dbd01f | 1993 | } else { |
8dd3557a | 1994 | *tree_pagep = tree_page; |
31dbd01f IE |
1995 | return tree_rmap_item; |
1996 | } | |
1997 | } | |
1998 | ||
7b6ba2c7 | 1999 | rmap_item->address |= UNSTABLE_FLAG; |
31dbd01f | 2000 | rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); |
e850dcf5 | 2001 | DO_NUMA(rmap_item->nid = nid); |
31dbd01f | 2002 | rb_link_node(&rmap_item->node, parent, new); |
90bd6fd3 | 2003 | rb_insert_color(&rmap_item->node, root); |
31dbd01f | 2004 | |
473b0ce4 | 2005 | ksm_pages_unshared++; |
31dbd01f IE |
2006 | return NULL; |
2007 | } | |
2008 | ||
2009 | /* | |
2010 | * stable_tree_append - add another rmap_item to the linked list of | |
2011 | * rmap_items hanging off a given node of the stable tree, all sharing | |
2012 | * the same ksm page. | |
2013 | */ | |
21fbd591 QZ |
2014 | static void stable_tree_append(struct ksm_rmap_item *rmap_item, |
2015 | struct ksm_stable_node *stable_node, | |
2c653d0e | 2016 | bool max_page_sharing_bypass) |
31dbd01f | 2017 | { |
2c653d0e AA |
2018 | /* |
2019 | * rmap won't find this mapping if we don't insert the | |
2020 | * rmap_item in the right stable_node | |
2021 | * duplicate. page_migration could break later if rmap breaks, | |
2022 | * so we can as well crash here. We really need to check for | |
2023 | * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check | |
457aef94 | 2024 | * for other negative values as an underflow if detected here |
2c653d0e AA |
2025 | * for the first time (and not when decreasing rmap_hlist_len) |
2026 | * would be sign of memory corruption in the stable_node. | |
2027 | */ | |
2028 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
2029 | ||
2030 | stable_node->rmap_hlist_len++; | |
2031 | if (!max_page_sharing_bypass) | |
2032 | /* possibly non fatal but unexpected overflow, only warn */ | |
2033 | WARN_ON_ONCE(stable_node->rmap_hlist_len > | |
2034 | ksm_max_page_sharing); | |
2035 | ||
7b6ba2c7 | 2036 | rmap_item->head = stable_node; |
31dbd01f | 2037 | rmap_item->address |= STABLE_FLAG; |
7b6ba2c7 | 2038 | hlist_add_head(&rmap_item->hlist, &stable_node->hlist); |
e178dfde | 2039 | |
7b6ba2c7 HD |
2040 | if (rmap_item->hlist.next) |
2041 | ksm_pages_sharing++; | |
2042 | else | |
2043 | ksm_pages_shared++; | |
76093853 | 2044 | |
2045 | rmap_item->mm->ksm_merging_pages++; | |
31dbd01f IE |
2046 | } |
2047 | ||
2048 | /* | |
81464e30 HD |
2049 | * cmp_and_merge_page - first see if page can be merged into the stable tree; |
2050 | * if not, compare checksum to previous and if it's the same, see if page can | |
2051 | * be inserted into the unstable tree, or merged with a page already there and | |
2052 | * both transferred to the stable tree. | |
31dbd01f IE |
2053 | * |
2054 | * @page: the page that we are searching identical page to. | |
2055 | * @rmap_item: the reverse mapping into the virtual address of this page | |
2056 | */ | |
21fbd591 | 2057 | static void cmp_and_merge_page(struct page *page, struct ksm_rmap_item *rmap_item) |
31dbd01f | 2058 | { |
4b22927f | 2059 | struct mm_struct *mm = rmap_item->mm; |
21fbd591 | 2060 | struct ksm_rmap_item *tree_rmap_item; |
8dd3557a | 2061 | struct page *tree_page = NULL; |
21fbd591 | 2062 | struct ksm_stable_node *stable_node; |
8dd3557a | 2063 | struct page *kpage; |
31dbd01f IE |
2064 | unsigned int checksum; |
2065 | int err; | |
2c653d0e | 2066 | bool max_page_sharing_bypass = false; |
31dbd01f | 2067 | |
4146d2d6 HD |
2068 | stable_node = page_stable_node(page); |
2069 | if (stable_node) { | |
2070 | if (stable_node->head != &migrate_nodes && | |
2c653d0e AA |
2071 | get_kpfn_nid(READ_ONCE(stable_node->kpfn)) != |
2072 | NUMA(stable_node->nid)) { | |
2073 | stable_node_dup_del(stable_node); | |
4146d2d6 HD |
2074 | stable_node->head = &migrate_nodes; |
2075 | list_add(&stable_node->list, stable_node->head); | |
2076 | } | |
2077 | if (stable_node->head != &migrate_nodes && | |
2078 | rmap_item->head == stable_node) | |
2079 | return; | |
2c653d0e AA |
2080 | /* |
2081 | * If it's a KSM fork, allow it to go over the sharing limit | |
2082 | * without warnings. | |
2083 | */ | |
2084 | if (!is_page_sharing_candidate(stable_node)) | |
2085 | max_page_sharing_bypass = true; | |
4146d2d6 | 2086 | } |
31dbd01f IE |
2087 | |
2088 | /* We first start with searching the page inside the stable tree */ | |
62b61f61 | 2089 | kpage = stable_tree_search(page); |
4146d2d6 HD |
2090 | if (kpage == page && rmap_item->head == stable_node) { |
2091 | put_page(kpage); | |
2092 | return; | |
2093 | } | |
2094 | ||
2095 | remove_rmap_item_from_tree(rmap_item); | |
2096 | ||
62b61f61 | 2097 | if (kpage) { |
2cee57d1 YS |
2098 | if (PTR_ERR(kpage) == -EBUSY) |
2099 | return; | |
2100 | ||
08beca44 | 2101 | err = try_to_merge_with_ksm_page(rmap_item, page, kpage); |
31dbd01f IE |
2102 | if (!err) { |
2103 | /* | |
2104 | * The page was successfully merged: | |
2105 | * add its rmap_item to the stable tree. | |
2106 | */ | |
5ad64688 | 2107 | lock_page(kpage); |
2c653d0e AA |
2108 | stable_tree_append(rmap_item, page_stable_node(kpage), |
2109 | max_page_sharing_bypass); | |
5ad64688 | 2110 | unlock_page(kpage); |
31dbd01f | 2111 | } |
8dd3557a | 2112 | put_page(kpage); |
31dbd01f IE |
2113 | return; |
2114 | } | |
2115 | ||
2116 | /* | |
4035c07a HD |
2117 | * If the hash value of the page has changed from the last time |
2118 | * we calculated it, this page is changing frequently: therefore we | |
2119 | * don't want to insert it in the unstable tree, and we don't want | |
2120 | * to waste our time searching for something identical to it there. | |
31dbd01f IE |
2121 | */ |
2122 | checksum = calc_checksum(page); | |
2123 | if (rmap_item->oldchecksum != checksum) { | |
2124 | rmap_item->oldchecksum = checksum; | |
2125 | return; | |
2126 | } | |
2127 | ||
e86c59b1 CI |
2128 | /* |
2129 | * Same checksum as an empty page. We attempt to merge it with the | |
2130 | * appropriate zero page if the user enabled this via sysfs. | |
2131 | */ | |
2132 | if (ksm_use_zero_pages && (checksum == zero_checksum)) { | |
2133 | struct vm_area_struct *vma; | |
2134 | ||
d8ed45c5 | 2135 | mmap_read_lock(mm); |
4b22927f | 2136 | vma = find_mergeable_vma(mm, rmap_item->address); |
56df70a6 MS |
2137 | if (vma) { |
2138 | err = try_to_merge_one_page(vma, page, | |
2139 | ZERO_PAGE(rmap_item->address)); | |
2140 | } else { | |
2141 | /* | |
2142 | * If the vma is out of date, we do not need to | |
2143 | * continue. | |
2144 | */ | |
2145 | err = 0; | |
2146 | } | |
d8ed45c5 | 2147 | mmap_read_unlock(mm); |
e86c59b1 CI |
2148 | /* |
2149 | * In case of failure, the page was not really empty, so we | |
2150 | * need to continue. Otherwise we're done. | |
2151 | */ | |
2152 | if (!err) | |
2153 | return; | |
2154 | } | |
8dd3557a HD |
2155 | tree_rmap_item = |
2156 | unstable_tree_search_insert(rmap_item, page, &tree_page); | |
31dbd01f | 2157 | if (tree_rmap_item) { |
77da2ba0 CI |
2158 | bool split; |
2159 | ||
8dd3557a HD |
2160 | kpage = try_to_merge_two_pages(rmap_item, page, |
2161 | tree_rmap_item, tree_page); | |
77da2ba0 CI |
2162 | /* |
2163 | * If both pages we tried to merge belong to the same compound | |
2164 | * page, then we actually ended up increasing the reference | |
2165 | * count of the same compound page twice, and split_huge_page | |
2166 | * failed. | |
2167 | * Here we set a flag if that happened, and we use it later to | |
2168 | * try split_huge_page again. Since we call put_page right | |
2169 | * afterwards, the reference count will be correct and | |
2170 | * split_huge_page should succeed. | |
2171 | */ | |
2172 | split = PageTransCompound(page) | |
2173 | && compound_head(page) == compound_head(tree_page); | |
8dd3557a | 2174 | put_page(tree_page); |
8dd3557a | 2175 | if (kpage) { |
bc56620b HD |
2176 | /* |
2177 | * The pages were successfully merged: insert new | |
2178 | * node in the stable tree and add both rmap_items. | |
2179 | */ | |
5ad64688 | 2180 | lock_page(kpage); |
7b6ba2c7 HD |
2181 | stable_node = stable_tree_insert(kpage); |
2182 | if (stable_node) { | |
2c653d0e AA |
2183 | stable_tree_append(tree_rmap_item, stable_node, |
2184 | false); | |
2185 | stable_tree_append(rmap_item, stable_node, | |
2186 | false); | |
7b6ba2c7 | 2187 | } |
5ad64688 | 2188 | unlock_page(kpage); |
7b6ba2c7 | 2189 | |
31dbd01f IE |
2190 | /* |
2191 | * If we fail to insert the page into the stable tree, | |
2192 | * we will have 2 virtual addresses that are pointing | |
2193 | * to a ksm page left outside the stable tree, | |
2194 | * in which case we need to break_cow on both. | |
2195 | */ | |
7b6ba2c7 | 2196 | if (!stable_node) { |
8dd3557a HD |
2197 | break_cow(tree_rmap_item); |
2198 | break_cow(rmap_item); | |
31dbd01f | 2199 | } |
77da2ba0 CI |
2200 | } else if (split) { |
2201 | /* | |
2202 | * We are here if we tried to merge two pages and | |
2203 | * failed because they both belonged to the same | |
2204 | * compound page. We will split the page now, but no | |
2205 | * merging will take place. | |
2206 | * We do not want to add the cost of a full lock; if | |
2207 | * the page is locked, it is better to skip it and | |
2208 | * perhaps try again later. | |
2209 | */ | |
2210 | if (!trylock_page(page)) | |
2211 | return; | |
2212 | split_huge_page(page); | |
2213 | unlock_page(page); | |
31dbd01f | 2214 | } |
31dbd01f IE |
2215 | } |
2216 | } | |
2217 | ||
21fbd591 QZ |
2218 | static struct ksm_rmap_item *get_next_rmap_item(struct ksm_mm_slot *mm_slot, |
2219 | struct ksm_rmap_item **rmap_list, | |
31dbd01f IE |
2220 | unsigned long addr) |
2221 | { | |
21fbd591 | 2222 | struct ksm_rmap_item *rmap_item; |
31dbd01f | 2223 | |
6514d511 HD |
2224 | while (*rmap_list) { |
2225 | rmap_item = *rmap_list; | |
93d17715 | 2226 | if ((rmap_item->address & PAGE_MASK) == addr) |
31dbd01f | 2227 | return rmap_item; |
31dbd01f IE |
2228 | if (rmap_item->address > addr) |
2229 | break; | |
6514d511 | 2230 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 2231 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
2232 | free_rmap_item(rmap_item); |
2233 | } | |
2234 | ||
2235 | rmap_item = alloc_rmap_item(); | |
2236 | if (rmap_item) { | |
2237 | /* It has already been zeroed */ | |
2238 | rmap_item->mm = mm_slot->mm; | |
cb4df4ca | 2239 | rmap_item->mm->ksm_rmap_items++; |
31dbd01f | 2240 | rmap_item->address = addr; |
6514d511 HD |
2241 | rmap_item->rmap_list = *rmap_list; |
2242 | *rmap_list = rmap_item; | |
31dbd01f IE |
2243 | } |
2244 | return rmap_item; | |
2245 | } | |
2246 | ||
21fbd591 | 2247 | static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page) |
31dbd01f IE |
2248 | { |
2249 | struct mm_struct *mm; | |
21fbd591 | 2250 | struct ksm_mm_slot *slot; |
31dbd01f | 2251 | struct vm_area_struct *vma; |
21fbd591 | 2252 | struct ksm_rmap_item *rmap_item; |
a5f18ba0 | 2253 | struct vma_iterator vmi; |
90bd6fd3 | 2254 | int nid; |
31dbd01f | 2255 | |
23f746e4 | 2256 | if (list_empty(&ksm_mm_head.mm_node)) |
31dbd01f IE |
2257 | return NULL; |
2258 | ||
2259 | slot = ksm_scan.mm_slot; | |
2260 | if (slot == &ksm_mm_head) { | |
2919bfd0 HD |
2261 | /* |
2262 | * A number of pages can hang around indefinitely on per-cpu | |
2263 | * pagevecs, raised page count preventing write_protect_page | |
2264 | * from merging them. Though it doesn't really matter much, | |
2265 | * it is puzzling to see some stuck in pages_volatile until | |
2266 | * other activity jostles them out, and they also prevented | |
2267 | * LTP's KSM test from succeeding deterministically; so drain | |
2268 | * them here (here rather than on entry to ksm_do_scan(), | |
2269 | * so we don't IPI too often when pages_to_scan is set low). | |
2270 | */ | |
2271 | lru_add_drain_all(); | |
2272 | ||
4146d2d6 HD |
2273 | /* |
2274 | * Whereas stale stable_nodes on the stable_tree itself | |
2275 | * get pruned in the regular course of stable_tree_search(), | |
2276 | * those moved out to the migrate_nodes list can accumulate: | |
2277 | * so prune them once before each full scan. | |
2278 | */ | |
2279 | if (!ksm_merge_across_nodes) { | |
21fbd591 | 2280 | struct ksm_stable_node *stable_node, *next; |
4146d2d6 HD |
2281 | struct page *page; |
2282 | ||
03640418 GT |
2283 | list_for_each_entry_safe(stable_node, next, |
2284 | &migrate_nodes, list) { | |
2cee57d1 YS |
2285 | page = get_ksm_page(stable_node, |
2286 | GET_KSM_PAGE_NOLOCK); | |
4146d2d6 HD |
2287 | if (page) |
2288 | put_page(page); | |
2289 | cond_resched(); | |
2290 | } | |
2291 | } | |
2292 | ||
ef53d16c | 2293 | for (nid = 0; nid < ksm_nr_node_ids; nid++) |
90bd6fd3 | 2294 | root_unstable_tree[nid] = RB_ROOT; |
31dbd01f IE |
2295 | |
2296 | spin_lock(&ksm_mmlist_lock); | |
23f746e4 | 2297 | slot = list_entry(slot->mm_node.next, struct ksm_mm_slot, mm_node); |
31dbd01f IE |
2298 | ksm_scan.mm_slot = slot; |
2299 | spin_unlock(&ksm_mmlist_lock); | |
2b472611 HD |
2300 | /* |
2301 | * Although we tested list_empty() above, a racing __ksm_exit | |
2302 | * of the last mm on the list may have removed it since then. | |
2303 | */ | |
2304 | if (slot == &ksm_mm_head) | |
2305 | return NULL; | |
31dbd01f IE |
2306 | next_mm: |
2307 | ksm_scan.address = 0; | |
6514d511 | 2308 | ksm_scan.rmap_list = &slot->rmap_list; |
31dbd01f IE |
2309 | } |
2310 | ||
2311 | mm = slot->mm; | |
a5f18ba0 MWO |
2312 | vma_iter_init(&vmi, mm, ksm_scan.address); |
2313 | ||
d8ed45c5 | 2314 | mmap_read_lock(mm); |
9ba69294 | 2315 | if (ksm_test_exit(mm)) |
a5f18ba0 | 2316 | goto no_vmas; |
9ba69294 | 2317 | |
a5f18ba0 | 2318 | for_each_vma(vmi, vma) { |
31dbd01f IE |
2319 | if (!(vma->vm_flags & VM_MERGEABLE)) |
2320 | continue; | |
2321 | if (ksm_scan.address < vma->vm_start) | |
2322 | ksm_scan.address = vma->vm_start; | |
2323 | if (!vma->anon_vma) | |
2324 | ksm_scan.address = vma->vm_end; | |
2325 | ||
2326 | while (ksm_scan.address < vma->vm_end) { | |
9ba69294 HD |
2327 | if (ksm_test_exit(mm)) |
2328 | break; | |
31dbd01f | 2329 | *page = follow_page(vma, ksm_scan.address, FOLL_GET); |
f7091ed6 | 2330 | if (IS_ERR_OR_NULL(*page)) { |
21ae5b01 AA |
2331 | ksm_scan.address += PAGE_SIZE; |
2332 | cond_resched(); | |
2333 | continue; | |
2334 | } | |
f7091ed6 HW |
2335 | if (is_zone_device_page(*page)) |
2336 | goto next_page; | |
f765f540 | 2337 | if (PageAnon(*page)) { |
31dbd01f IE |
2338 | flush_anon_page(vma, *page, ksm_scan.address); |
2339 | flush_dcache_page(*page); | |
2340 | rmap_item = get_next_rmap_item(slot, | |
6514d511 | 2341 | ksm_scan.rmap_list, ksm_scan.address); |
31dbd01f | 2342 | if (rmap_item) { |
6514d511 HD |
2343 | ksm_scan.rmap_list = |
2344 | &rmap_item->rmap_list; | |
31dbd01f IE |
2345 | ksm_scan.address += PAGE_SIZE; |
2346 | } else | |
2347 | put_page(*page); | |
d8ed45c5 | 2348 | mmap_read_unlock(mm); |
31dbd01f IE |
2349 | return rmap_item; |
2350 | } | |
f7091ed6 | 2351 | next_page: |
21ae5b01 | 2352 | put_page(*page); |
31dbd01f IE |
2353 | ksm_scan.address += PAGE_SIZE; |
2354 | cond_resched(); | |
2355 | } | |
2356 | } | |
2357 | ||
9ba69294 | 2358 | if (ksm_test_exit(mm)) { |
a5f18ba0 | 2359 | no_vmas: |
9ba69294 | 2360 | ksm_scan.address = 0; |
6514d511 | 2361 | ksm_scan.rmap_list = &slot->rmap_list; |
9ba69294 | 2362 | } |
31dbd01f IE |
2363 | /* |
2364 | * Nuke all the rmap_items that are above this current rmap: | |
2365 | * because there were no VM_MERGEABLE vmas with such addresses. | |
2366 | */ | |
420be4ed | 2367 | remove_trailing_rmap_items(ksm_scan.rmap_list); |
31dbd01f IE |
2368 | |
2369 | spin_lock(&ksm_mmlist_lock); | |
23f746e4 QZ |
2370 | ksm_scan.mm_slot = list_entry(slot->mm_node.next, |
2371 | struct ksm_mm_slot, mm_node); | |
cd551f97 HD |
2372 | if (ksm_scan.address == 0) { |
2373 | /* | |
c1e8d7c6 | 2374 | * We've completed a full scan of all vmas, holding mmap_lock |
cd551f97 HD |
2375 | * throughout, and found no VM_MERGEABLE: so do the same as |
2376 | * __ksm_exit does to remove this mm from all our lists now. | |
9ba69294 HD |
2377 | * This applies either when cleaning up after __ksm_exit |
2378 | * (but beware: we can reach here even before __ksm_exit), | |
2379 | * or when all VM_MERGEABLE areas have been unmapped (and | |
c1e8d7c6 | 2380 | * mmap_lock then protects against race with MADV_MERGEABLE). |
cd551f97 | 2381 | */ |
79b09941 | 2382 | hash_del(&slot->hash); |
23f746e4 | 2383 | list_del(&slot->mm_node); |
9ba69294 HD |
2384 | spin_unlock(&ksm_mmlist_lock); |
2385 | ||
cd551f97 HD |
2386 | free_mm_slot(slot); |
2387 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
d8ed45c5 | 2388 | mmap_read_unlock(mm); |
9ba69294 HD |
2389 | mmdrop(mm); |
2390 | } else { | |
d8ed45c5 | 2391 | mmap_read_unlock(mm); |
7496fea9 | 2392 | /* |
3e4e28c5 | 2393 | * mmap_read_unlock(mm) first because after |
7496fea9 ZC |
2394 | * spin_unlock(&ksm_mmlist_lock) run, the "mm" may |
2395 | * already have been freed under us by __ksm_exit() | |
2396 | * because the "mm_slot" is still hashed and | |
2397 | * ksm_scan.mm_slot doesn't point to it anymore. | |
2398 | */ | |
2399 | spin_unlock(&ksm_mmlist_lock); | |
cd551f97 | 2400 | } |
31dbd01f IE |
2401 | |
2402 | /* Repeat until we've completed scanning the whole list */ | |
cd551f97 | 2403 | slot = ksm_scan.mm_slot; |
31dbd01f IE |
2404 | if (slot != &ksm_mm_head) |
2405 | goto next_mm; | |
2406 | ||
31dbd01f IE |
2407 | ksm_scan.seqnr++; |
2408 | return NULL; | |
2409 | } | |
2410 | ||
2411 | /** | |
2412 | * ksm_do_scan - the ksm scanner main worker function. | |
b7701a5f | 2413 | * @scan_npages: number of pages we want to scan before we return. |
31dbd01f IE |
2414 | */ |
2415 | static void ksm_do_scan(unsigned int scan_npages) | |
2416 | { | |
21fbd591 | 2417 | struct ksm_rmap_item *rmap_item; |
3f649ab7 | 2418 | struct page *page; |
31dbd01f | 2419 | |
878aee7d | 2420 | while (scan_npages-- && likely(!freezing(current))) { |
31dbd01f IE |
2421 | cond_resched(); |
2422 | rmap_item = scan_get_next_rmap_item(&page); | |
2423 | if (!rmap_item) | |
2424 | return; | |
4146d2d6 | 2425 | cmp_and_merge_page(page, rmap_item); |
31dbd01f IE |
2426 | put_page(page); |
2427 | } | |
2428 | } | |
2429 | ||
6e158384 HD |
2430 | static int ksmd_should_run(void) |
2431 | { | |
23f746e4 | 2432 | return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_node); |
6e158384 HD |
2433 | } |
2434 | ||
31dbd01f IE |
2435 | static int ksm_scan_thread(void *nothing) |
2436 | { | |
fcf9a0ef KT |
2437 | unsigned int sleep_ms; |
2438 | ||
878aee7d | 2439 | set_freezable(); |
339aa624 | 2440 | set_user_nice(current, 5); |
31dbd01f IE |
2441 | |
2442 | while (!kthread_should_stop()) { | |
6e158384 | 2443 | mutex_lock(&ksm_thread_mutex); |
ef4d43a8 | 2444 | wait_while_offlining(); |
6e158384 | 2445 | if (ksmd_should_run()) |
31dbd01f | 2446 | ksm_do_scan(ksm_thread_pages_to_scan); |
6e158384 HD |
2447 | mutex_unlock(&ksm_thread_mutex); |
2448 | ||
878aee7d AA |
2449 | try_to_freeze(); |
2450 | ||
6e158384 | 2451 | if (ksmd_should_run()) { |
fcf9a0ef KT |
2452 | sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs); |
2453 | wait_event_interruptible_timeout(ksm_iter_wait, | |
2454 | sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs), | |
2455 | msecs_to_jiffies(sleep_ms)); | |
31dbd01f | 2456 | } else { |
878aee7d | 2457 | wait_event_freezable(ksm_thread_wait, |
6e158384 | 2458 | ksmd_should_run() || kthread_should_stop()); |
31dbd01f IE |
2459 | } |
2460 | } | |
2461 | return 0; | |
2462 | } | |
2463 | ||
f8af4da3 HD |
2464 | int ksm_madvise(struct vm_area_struct *vma, unsigned long start, |
2465 | unsigned long end, int advice, unsigned long *vm_flags) | |
2466 | { | |
2467 | struct mm_struct *mm = vma->vm_mm; | |
d952b791 | 2468 | int err; |
f8af4da3 HD |
2469 | |
2470 | switch (advice) { | |
2471 | case MADV_MERGEABLE: | |
2472 | /* | |
2473 | * Be somewhat over-protective for now! | |
2474 | */ | |
2475 | if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE | | |
2476 | VM_PFNMAP | VM_IO | VM_DONTEXPAND | | |
0661a336 | 2477 | VM_HUGETLB | VM_MIXEDMAP)) |
f8af4da3 HD |
2478 | return 0; /* just ignore the advice */ |
2479 | ||
e1fb4a08 DJ |
2480 | if (vma_is_dax(vma)) |
2481 | return 0; | |
2482 | ||
12564485 SA |
2483 | #ifdef VM_SAO |
2484 | if (*vm_flags & VM_SAO) | |
2485 | return 0; | |
2486 | #endif | |
74a04967 KA |
2487 | #ifdef VM_SPARC_ADI |
2488 | if (*vm_flags & VM_SPARC_ADI) | |
2489 | return 0; | |
2490 | #endif | |
cc2383ec | 2491 | |
d952b791 HD |
2492 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { |
2493 | err = __ksm_enter(mm); | |
2494 | if (err) | |
2495 | return err; | |
2496 | } | |
f8af4da3 HD |
2497 | |
2498 | *vm_flags |= VM_MERGEABLE; | |
2499 | break; | |
2500 | ||
2501 | case MADV_UNMERGEABLE: | |
2502 | if (!(*vm_flags & VM_MERGEABLE)) | |
2503 | return 0; /* just ignore the advice */ | |
2504 | ||
d952b791 HD |
2505 | if (vma->anon_vma) { |
2506 | err = unmerge_ksm_pages(vma, start, end); | |
2507 | if (err) | |
2508 | return err; | |
2509 | } | |
f8af4da3 HD |
2510 | |
2511 | *vm_flags &= ~VM_MERGEABLE; | |
2512 | break; | |
2513 | } | |
2514 | ||
2515 | return 0; | |
2516 | } | |
33cf1707 | 2517 | EXPORT_SYMBOL_GPL(ksm_madvise); |
f8af4da3 HD |
2518 | |
2519 | int __ksm_enter(struct mm_struct *mm) | |
2520 | { | |
21fbd591 | 2521 | struct ksm_mm_slot *mm_slot; |
6e158384 HD |
2522 | int needs_wakeup; |
2523 | ||
2524 | mm_slot = alloc_mm_slot(); | |
31dbd01f IE |
2525 | if (!mm_slot) |
2526 | return -ENOMEM; | |
2527 | ||
6e158384 | 2528 | /* Check ksm_run too? Would need tighter locking */ |
23f746e4 | 2529 | needs_wakeup = list_empty(&ksm_mm_head.mm_node); |
6e158384 | 2530 | |
31dbd01f IE |
2531 | spin_lock(&ksm_mmlist_lock); |
2532 | insert_to_mm_slots_hash(mm, mm_slot); | |
2533 | /* | |
cbf86cfe HD |
2534 | * When KSM_RUN_MERGE (or KSM_RUN_STOP), |
2535 | * insert just behind the scanning cursor, to let the area settle | |
31dbd01f IE |
2536 | * down a little; when fork is followed by immediate exec, we don't |
2537 | * want ksmd to waste time setting up and tearing down an rmap_list. | |
cbf86cfe HD |
2538 | * |
2539 | * But when KSM_RUN_UNMERGE, it's important to insert ahead of its | |
2540 | * scanning cursor, otherwise KSM pages in newly forked mms will be | |
2541 | * missed: then we might as well insert at the end of the list. | |
31dbd01f | 2542 | */ |
cbf86cfe | 2543 | if (ksm_run & KSM_RUN_UNMERGE) |
23f746e4 | 2544 | list_add_tail(&mm_slot->mm_node, &ksm_mm_head.mm_node); |
cbf86cfe | 2545 | else |
23f746e4 | 2546 | list_add_tail(&mm_slot->mm_node, &ksm_scan.mm_slot->mm_node); |
31dbd01f IE |
2547 | spin_unlock(&ksm_mmlist_lock); |
2548 | ||
f8af4da3 | 2549 | set_bit(MMF_VM_MERGEABLE, &mm->flags); |
f1f10076 | 2550 | mmgrab(mm); |
6e158384 HD |
2551 | |
2552 | if (needs_wakeup) | |
2553 | wake_up_interruptible(&ksm_thread_wait); | |
2554 | ||
f8af4da3 HD |
2555 | return 0; |
2556 | } | |
2557 | ||
1c2fb7a4 | 2558 | void __ksm_exit(struct mm_struct *mm) |
f8af4da3 | 2559 | { |
21fbd591 | 2560 | struct ksm_mm_slot *mm_slot; |
9ba69294 | 2561 | int easy_to_free = 0; |
cd551f97 | 2562 | |
31dbd01f | 2563 | /* |
9ba69294 HD |
2564 | * This process is exiting: if it's straightforward (as is the |
2565 | * case when ksmd was never running), free mm_slot immediately. | |
2566 | * But if it's at the cursor or has rmap_items linked to it, use | |
c1e8d7c6 | 2567 | * mmap_lock to synchronize with any break_cows before pagetables |
9ba69294 HD |
2568 | * are freed, and leave the mm_slot on the list for ksmd to free. |
2569 | * Beware: ksm may already have noticed it exiting and freed the slot. | |
31dbd01f | 2570 | */ |
9ba69294 | 2571 | |
cd551f97 HD |
2572 | spin_lock(&ksm_mmlist_lock); |
2573 | mm_slot = get_mm_slot(mm); | |
9ba69294 | 2574 | if (mm_slot && ksm_scan.mm_slot != mm_slot) { |
6514d511 | 2575 | if (!mm_slot->rmap_list) { |
79b09941 | 2576 | hash_del(&mm_slot->hash); |
23f746e4 | 2577 | list_del(&mm_slot->mm_node); |
9ba69294 HD |
2578 | easy_to_free = 1; |
2579 | } else { | |
23f746e4 QZ |
2580 | list_move(&mm_slot->mm_node, |
2581 | &ksm_scan.mm_slot->mm_node); | |
9ba69294 | 2582 | } |
cd551f97 | 2583 | } |
cd551f97 HD |
2584 | spin_unlock(&ksm_mmlist_lock); |
2585 | ||
9ba69294 HD |
2586 | if (easy_to_free) { |
2587 | free_mm_slot(mm_slot); | |
2588 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
2589 | mmdrop(mm); | |
2590 | } else if (mm_slot) { | |
d8ed45c5 ML |
2591 | mmap_write_lock(mm); |
2592 | mmap_write_unlock(mm); | |
9ba69294 | 2593 | } |
31dbd01f IE |
2594 | } |
2595 | ||
cbf86cfe | 2596 | struct page *ksm_might_need_to_copy(struct page *page, |
5ad64688 HD |
2597 | struct vm_area_struct *vma, unsigned long address) |
2598 | { | |
e05b3453 MWO |
2599 | struct folio *folio = page_folio(page); |
2600 | struct anon_vma *anon_vma = folio_anon_vma(folio); | |
5ad64688 HD |
2601 | struct page *new_page; |
2602 | ||
cbf86cfe HD |
2603 | if (PageKsm(page)) { |
2604 | if (page_stable_node(page) && | |
2605 | !(ksm_run & KSM_RUN_UNMERGE)) | |
2606 | return page; /* no need to copy it */ | |
2607 | } else if (!anon_vma) { | |
2608 | return page; /* no need to copy it */ | |
e1c63e11 NS |
2609 | } else if (page->index == linear_page_index(vma, address) && |
2610 | anon_vma->root == vma->anon_vma->root) { | |
cbf86cfe HD |
2611 | return page; /* still no need to copy it */ |
2612 | } | |
2613 | if (!PageUptodate(page)) | |
2614 | return page; /* let do_swap_page report the error */ | |
2615 | ||
5ad64688 | 2616 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
8f425e4e MWO |
2617 | if (new_page && |
2618 | mem_cgroup_charge(page_folio(new_page), vma->vm_mm, GFP_KERNEL)) { | |
62fdb163 HD |
2619 | put_page(new_page); |
2620 | new_page = NULL; | |
2621 | } | |
5ad64688 HD |
2622 | if (new_page) { |
2623 | copy_user_highpage(new_page, page, address, vma); | |
2624 | ||
2625 | SetPageDirty(new_page); | |
2626 | __SetPageUptodate(new_page); | |
48c935ad | 2627 | __SetPageLocked(new_page); |
4d45c3af YY |
2628 | #ifdef CONFIG_SWAP |
2629 | count_vm_event(KSM_SWPIN_COPY); | |
2630 | #endif | |
5ad64688 HD |
2631 | } |
2632 | ||
5ad64688 HD |
2633 | return new_page; |
2634 | } | |
2635 | ||
6d4675e6 | 2636 | void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc) |
e9995ef9 | 2637 | { |
21fbd591 QZ |
2638 | struct ksm_stable_node *stable_node; |
2639 | struct ksm_rmap_item *rmap_item; | |
e9995ef9 HD |
2640 | int search_new_forks = 0; |
2641 | ||
2f031c6f | 2642 | VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio); |
9f32624b JK |
2643 | |
2644 | /* | |
2645 | * Rely on the page lock to protect against concurrent modifications | |
2646 | * to that page's node of the stable tree. | |
2647 | */ | |
2f031c6f | 2648 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
e9995ef9 | 2649 | |
2f031c6f | 2650 | stable_node = folio_stable_node(folio); |
e9995ef9 | 2651 | if (!stable_node) |
1df631ae | 2652 | return; |
e9995ef9 | 2653 | again: |
b67bfe0d | 2654 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
e9995ef9 | 2655 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
5beb4930 | 2656 | struct anon_vma_chain *vmac; |
e9995ef9 HD |
2657 | struct vm_area_struct *vma; |
2658 | ||
ad12695f | 2659 | cond_resched(); |
6d4675e6 MK |
2660 | if (!anon_vma_trylock_read(anon_vma)) { |
2661 | if (rwc->try_lock) { | |
2662 | rwc->contended = true; | |
2663 | return; | |
2664 | } | |
2665 | anon_vma_lock_read(anon_vma); | |
2666 | } | |
bf181b9f ML |
2667 | anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, |
2668 | 0, ULONG_MAX) { | |
1105a2fc JH |
2669 | unsigned long addr; |
2670 | ||
ad12695f | 2671 | cond_resched(); |
5beb4930 | 2672 | vma = vmac->vma; |
1105a2fc JH |
2673 | |
2674 | /* Ignore the stable/unstable/sqnr flags */ | |
cd7fae26 | 2675 | addr = rmap_item->address & PAGE_MASK; |
1105a2fc JH |
2676 | |
2677 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
e9995ef9 HD |
2678 | continue; |
2679 | /* | |
2680 | * Initially we examine only the vma which covers this | |
2681 | * rmap_item; but later, if there is still work to do, | |
2682 | * we examine covering vmas in other mms: in case they | |
2683 | * were forked from the original since ksmd passed. | |
2684 | */ | |
2685 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
2686 | continue; | |
2687 | ||
0dd1c7bb JK |
2688 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
2689 | continue; | |
2690 | ||
2f031c6f | 2691 | if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) { |
b6b19f25 | 2692 | anon_vma_unlock_read(anon_vma); |
1df631ae | 2693 | return; |
e9995ef9 | 2694 | } |
2f031c6f | 2695 | if (rwc->done && rwc->done(folio)) { |
0dd1c7bb | 2696 | anon_vma_unlock_read(anon_vma); |
1df631ae | 2697 | return; |
0dd1c7bb | 2698 | } |
e9995ef9 | 2699 | } |
b6b19f25 | 2700 | anon_vma_unlock_read(anon_vma); |
e9995ef9 HD |
2701 | } |
2702 | if (!search_new_forks++) | |
2703 | goto again; | |
e9995ef9 HD |
2704 | } |
2705 | ||
52629506 | 2706 | #ifdef CONFIG_MIGRATION |
19138349 | 2707 | void folio_migrate_ksm(struct folio *newfolio, struct folio *folio) |
e9995ef9 | 2708 | { |
21fbd591 | 2709 | struct ksm_stable_node *stable_node; |
e9995ef9 | 2710 | |
19138349 MWO |
2711 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
2712 | VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio); | |
2713 | VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio); | |
e9995ef9 | 2714 | |
19138349 | 2715 | stable_node = folio_stable_node(folio); |
e9995ef9 | 2716 | if (stable_node) { |
19138349 MWO |
2717 | VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio); |
2718 | stable_node->kpfn = folio_pfn(newfolio); | |
c8d6553b | 2719 | /* |
19138349 | 2720 | * newfolio->mapping was set in advance; now we need smp_wmb() |
c8d6553b | 2721 | * to make sure that the new stable_node->kpfn is visible |
19138349 MWO |
2722 | * to get_ksm_page() before it can see that folio->mapping |
2723 | * has gone stale (or that folio_test_swapcache has been cleared). | |
c8d6553b HD |
2724 | */ |
2725 | smp_wmb(); | |
19138349 | 2726 | set_page_stable_node(&folio->page, NULL); |
e9995ef9 HD |
2727 | } |
2728 | } | |
2729 | #endif /* CONFIG_MIGRATION */ | |
2730 | ||
62b61f61 | 2731 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 HD |
2732 | static void wait_while_offlining(void) |
2733 | { | |
2734 | while (ksm_run & KSM_RUN_OFFLINE) { | |
2735 | mutex_unlock(&ksm_thread_mutex); | |
2736 | wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), | |
74316201 | 2737 | TASK_UNINTERRUPTIBLE); |
ef4d43a8 HD |
2738 | mutex_lock(&ksm_thread_mutex); |
2739 | } | |
2740 | } | |
2741 | ||
21fbd591 | 2742 | static bool stable_node_dup_remove_range(struct ksm_stable_node *stable_node, |
2c653d0e AA |
2743 | unsigned long start_pfn, |
2744 | unsigned long end_pfn) | |
2745 | { | |
2746 | if (stable_node->kpfn >= start_pfn && | |
2747 | stable_node->kpfn < end_pfn) { | |
2748 | /* | |
2749 | * Don't get_ksm_page, page has already gone: | |
2750 | * which is why we keep kpfn instead of page* | |
2751 | */ | |
2752 | remove_node_from_stable_tree(stable_node); | |
2753 | return true; | |
2754 | } | |
2755 | return false; | |
2756 | } | |
2757 | ||
21fbd591 | 2758 | static bool stable_node_chain_remove_range(struct ksm_stable_node *stable_node, |
2c653d0e AA |
2759 | unsigned long start_pfn, |
2760 | unsigned long end_pfn, | |
2761 | struct rb_root *root) | |
2762 | { | |
21fbd591 | 2763 | struct ksm_stable_node *dup; |
2c653d0e AA |
2764 | struct hlist_node *hlist_safe; |
2765 | ||
2766 | if (!is_stable_node_chain(stable_node)) { | |
2767 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
2768 | return stable_node_dup_remove_range(stable_node, start_pfn, | |
2769 | end_pfn); | |
2770 | } | |
2771 | ||
2772 | hlist_for_each_entry_safe(dup, hlist_safe, | |
2773 | &stable_node->hlist, hlist_dup) { | |
2774 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
2775 | stable_node_dup_remove_range(dup, start_pfn, end_pfn); | |
2776 | } | |
2777 | if (hlist_empty(&stable_node->hlist)) { | |
2778 | free_stable_node_chain(stable_node, root); | |
2779 | return true; /* notify caller that tree was rebalanced */ | |
2780 | } else | |
2781 | return false; | |
2782 | } | |
2783 | ||
ee0ea59c HD |
2784 | static void ksm_check_stable_tree(unsigned long start_pfn, |
2785 | unsigned long end_pfn) | |
62b61f61 | 2786 | { |
21fbd591 | 2787 | struct ksm_stable_node *stable_node, *next; |
62b61f61 | 2788 | struct rb_node *node; |
90bd6fd3 | 2789 | int nid; |
62b61f61 | 2790 | |
ef53d16c HD |
2791 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
2792 | node = rb_first(root_stable_tree + nid); | |
ee0ea59c | 2793 | while (node) { |
21fbd591 | 2794 | stable_node = rb_entry(node, struct ksm_stable_node, node); |
2c653d0e AA |
2795 | if (stable_node_chain_remove_range(stable_node, |
2796 | start_pfn, end_pfn, | |
2797 | root_stable_tree + | |
2798 | nid)) | |
ef53d16c | 2799 | node = rb_first(root_stable_tree + nid); |
2c653d0e | 2800 | else |
ee0ea59c HD |
2801 | node = rb_next(node); |
2802 | cond_resched(); | |
90bd6fd3 | 2803 | } |
ee0ea59c | 2804 | } |
03640418 | 2805 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
2806 | if (stable_node->kpfn >= start_pfn && |
2807 | stable_node->kpfn < end_pfn) | |
2808 | remove_node_from_stable_tree(stable_node); | |
2809 | cond_resched(); | |
2810 | } | |
62b61f61 HD |
2811 | } |
2812 | ||
2813 | static int ksm_memory_callback(struct notifier_block *self, | |
2814 | unsigned long action, void *arg) | |
2815 | { | |
2816 | struct memory_notify *mn = arg; | |
62b61f61 HD |
2817 | |
2818 | switch (action) { | |
2819 | case MEM_GOING_OFFLINE: | |
2820 | /* | |
ef4d43a8 HD |
2821 | * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() |
2822 | * and remove_all_stable_nodes() while memory is going offline: | |
2823 | * it is unsafe for them to touch the stable tree at this time. | |
2824 | * But unmerge_ksm_pages(), rmap lookups and other entry points | |
2825 | * which do not need the ksm_thread_mutex are all safe. | |
62b61f61 | 2826 | */ |
ef4d43a8 HD |
2827 | mutex_lock(&ksm_thread_mutex); |
2828 | ksm_run |= KSM_RUN_OFFLINE; | |
2829 | mutex_unlock(&ksm_thread_mutex); | |
62b61f61 HD |
2830 | break; |
2831 | ||
2832 | case MEM_OFFLINE: | |
2833 | /* | |
2834 | * Most of the work is done by page migration; but there might | |
2835 | * be a few stable_nodes left over, still pointing to struct | |
ee0ea59c HD |
2836 | * pages which have been offlined: prune those from the tree, |
2837 | * otherwise get_ksm_page() might later try to access a | |
2838 | * non-existent struct page. | |
62b61f61 | 2839 | */ |
ee0ea59c HD |
2840 | ksm_check_stable_tree(mn->start_pfn, |
2841 | mn->start_pfn + mn->nr_pages); | |
e4a9bc58 | 2842 | fallthrough; |
62b61f61 | 2843 | case MEM_CANCEL_OFFLINE: |
ef4d43a8 HD |
2844 | mutex_lock(&ksm_thread_mutex); |
2845 | ksm_run &= ~KSM_RUN_OFFLINE; | |
62b61f61 | 2846 | mutex_unlock(&ksm_thread_mutex); |
ef4d43a8 HD |
2847 | |
2848 | smp_mb(); /* wake_up_bit advises this */ | |
2849 | wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); | |
62b61f61 HD |
2850 | break; |
2851 | } | |
2852 | return NOTIFY_OK; | |
2853 | } | |
ef4d43a8 HD |
2854 | #else |
2855 | static void wait_while_offlining(void) | |
2856 | { | |
2857 | } | |
62b61f61 HD |
2858 | #endif /* CONFIG_MEMORY_HOTREMOVE */ |
2859 | ||
2ffd8679 HD |
2860 | #ifdef CONFIG_SYSFS |
2861 | /* | |
2862 | * This all compiles without CONFIG_SYSFS, but is a waste of space. | |
2863 | */ | |
2864 | ||
31dbd01f IE |
2865 | #define KSM_ATTR_RO(_name) \ |
2866 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2867 | #define KSM_ATTR(_name) \ | |
1bad2e5c | 2868 | static struct kobj_attribute _name##_attr = __ATTR_RW(_name) |
31dbd01f IE |
2869 | |
2870 | static ssize_t sleep_millisecs_show(struct kobject *kobj, | |
2871 | struct kobj_attribute *attr, char *buf) | |
2872 | { | |
ae7a927d | 2873 | return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs); |
31dbd01f IE |
2874 | } |
2875 | ||
2876 | static ssize_t sleep_millisecs_store(struct kobject *kobj, | |
2877 | struct kobj_attribute *attr, | |
2878 | const char *buf, size_t count) | |
2879 | { | |
dfefd226 | 2880 | unsigned int msecs; |
31dbd01f IE |
2881 | int err; |
2882 | ||
dfefd226 AD |
2883 | err = kstrtouint(buf, 10, &msecs); |
2884 | if (err) | |
31dbd01f IE |
2885 | return -EINVAL; |
2886 | ||
2887 | ksm_thread_sleep_millisecs = msecs; | |
fcf9a0ef | 2888 | wake_up_interruptible(&ksm_iter_wait); |
31dbd01f IE |
2889 | |
2890 | return count; | |
2891 | } | |
2892 | KSM_ATTR(sleep_millisecs); | |
2893 | ||
2894 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
2895 | struct kobj_attribute *attr, char *buf) | |
2896 | { | |
ae7a927d | 2897 | return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan); |
31dbd01f IE |
2898 | } |
2899 | ||
2900 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
2901 | struct kobj_attribute *attr, | |
2902 | const char *buf, size_t count) | |
2903 | { | |
dfefd226 | 2904 | unsigned int nr_pages; |
31dbd01f | 2905 | int err; |
31dbd01f | 2906 | |
dfefd226 AD |
2907 | err = kstrtouint(buf, 10, &nr_pages); |
2908 | if (err) | |
31dbd01f IE |
2909 | return -EINVAL; |
2910 | ||
2911 | ksm_thread_pages_to_scan = nr_pages; | |
2912 | ||
2913 | return count; | |
2914 | } | |
2915 | KSM_ATTR(pages_to_scan); | |
2916 | ||
2917 | static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, | |
2918 | char *buf) | |
2919 | { | |
ae7a927d | 2920 | return sysfs_emit(buf, "%lu\n", ksm_run); |
31dbd01f IE |
2921 | } |
2922 | ||
2923 | static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, | |
2924 | const char *buf, size_t count) | |
2925 | { | |
dfefd226 | 2926 | unsigned int flags; |
31dbd01f | 2927 | int err; |
31dbd01f | 2928 | |
dfefd226 AD |
2929 | err = kstrtouint(buf, 10, &flags); |
2930 | if (err) | |
31dbd01f IE |
2931 | return -EINVAL; |
2932 | if (flags > KSM_RUN_UNMERGE) | |
2933 | return -EINVAL; | |
2934 | ||
2935 | /* | |
2936 | * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. | |
2937 | * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, | |
d0f209f6 HD |
2938 | * breaking COW to free the pages_shared (but leaves mm_slots |
2939 | * on the list for when ksmd may be set running again). | |
31dbd01f IE |
2940 | */ |
2941 | ||
2942 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 2943 | wait_while_offlining(); |
31dbd01f IE |
2944 | if (ksm_run != flags) { |
2945 | ksm_run = flags; | |
d952b791 | 2946 | if (flags & KSM_RUN_UNMERGE) { |
e1e12d2f | 2947 | set_current_oom_origin(); |
d952b791 | 2948 | err = unmerge_and_remove_all_rmap_items(); |
e1e12d2f | 2949 | clear_current_oom_origin(); |
d952b791 HD |
2950 | if (err) { |
2951 | ksm_run = KSM_RUN_STOP; | |
2952 | count = err; | |
2953 | } | |
2954 | } | |
31dbd01f IE |
2955 | } |
2956 | mutex_unlock(&ksm_thread_mutex); | |
2957 | ||
2958 | if (flags & KSM_RUN_MERGE) | |
2959 | wake_up_interruptible(&ksm_thread_wait); | |
2960 | ||
2961 | return count; | |
2962 | } | |
2963 | KSM_ATTR(run); | |
2964 | ||
90bd6fd3 PH |
2965 | #ifdef CONFIG_NUMA |
2966 | static ssize_t merge_across_nodes_show(struct kobject *kobj, | |
ae7a927d | 2967 | struct kobj_attribute *attr, char *buf) |
90bd6fd3 | 2968 | { |
ae7a927d | 2969 | return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes); |
90bd6fd3 PH |
2970 | } |
2971 | ||
2972 | static ssize_t merge_across_nodes_store(struct kobject *kobj, | |
2973 | struct kobj_attribute *attr, | |
2974 | const char *buf, size_t count) | |
2975 | { | |
2976 | int err; | |
2977 | unsigned long knob; | |
2978 | ||
2979 | err = kstrtoul(buf, 10, &knob); | |
2980 | if (err) | |
2981 | return err; | |
2982 | if (knob > 1) | |
2983 | return -EINVAL; | |
2984 | ||
2985 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 2986 | wait_while_offlining(); |
90bd6fd3 | 2987 | if (ksm_merge_across_nodes != knob) { |
cbf86cfe | 2988 | if (ksm_pages_shared || remove_all_stable_nodes()) |
90bd6fd3 | 2989 | err = -EBUSY; |
ef53d16c HD |
2990 | else if (root_stable_tree == one_stable_tree) { |
2991 | struct rb_root *buf; | |
2992 | /* | |
2993 | * This is the first time that we switch away from the | |
2994 | * default of merging across nodes: must now allocate | |
2995 | * a buffer to hold as many roots as may be needed. | |
2996 | * Allocate stable and unstable together: | |
2997 | * MAXSMP NODES_SHIFT 10 will use 16kB. | |
2998 | */ | |
bafe1e14 JP |
2999 | buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), |
3000 | GFP_KERNEL); | |
ef53d16c HD |
3001 | /* Let us assume that RB_ROOT is NULL is zero */ |
3002 | if (!buf) | |
3003 | err = -ENOMEM; | |
3004 | else { | |
3005 | root_stable_tree = buf; | |
3006 | root_unstable_tree = buf + nr_node_ids; | |
3007 | /* Stable tree is empty but not the unstable */ | |
3008 | root_unstable_tree[0] = one_unstable_tree[0]; | |
3009 | } | |
3010 | } | |
3011 | if (!err) { | |
90bd6fd3 | 3012 | ksm_merge_across_nodes = knob; |
ef53d16c HD |
3013 | ksm_nr_node_ids = knob ? 1 : nr_node_ids; |
3014 | } | |
90bd6fd3 PH |
3015 | } |
3016 | mutex_unlock(&ksm_thread_mutex); | |
3017 | ||
3018 | return err ? err : count; | |
3019 | } | |
3020 | KSM_ATTR(merge_across_nodes); | |
3021 | #endif | |
3022 | ||
e86c59b1 | 3023 | static ssize_t use_zero_pages_show(struct kobject *kobj, |
ae7a927d | 3024 | struct kobj_attribute *attr, char *buf) |
e86c59b1 | 3025 | { |
ae7a927d | 3026 | return sysfs_emit(buf, "%u\n", ksm_use_zero_pages); |
e86c59b1 CI |
3027 | } |
3028 | static ssize_t use_zero_pages_store(struct kobject *kobj, | |
3029 | struct kobj_attribute *attr, | |
3030 | const char *buf, size_t count) | |
3031 | { | |
3032 | int err; | |
3033 | bool value; | |
3034 | ||
3035 | err = kstrtobool(buf, &value); | |
3036 | if (err) | |
3037 | return -EINVAL; | |
3038 | ||
3039 | ksm_use_zero_pages = value; | |
3040 | ||
3041 | return count; | |
3042 | } | |
3043 | KSM_ATTR(use_zero_pages); | |
3044 | ||
2c653d0e AA |
3045 | static ssize_t max_page_sharing_show(struct kobject *kobj, |
3046 | struct kobj_attribute *attr, char *buf) | |
3047 | { | |
ae7a927d | 3048 | return sysfs_emit(buf, "%u\n", ksm_max_page_sharing); |
2c653d0e AA |
3049 | } |
3050 | ||
3051 | static ssize_t max_page_sharing_store(struct kobject *kobj, | |
3052 | struct kobj_attribute *attr, | |
3053 | const char *buf, size_t count) | |
3054 | { | |
3055 | int err; | |
3056 | int knob; | |
3057 | ||
3058 | err = kstrtoint(buf, 10, &knob); | |
3059 | if (err) | |
3060 | return err; | |
3061 | /* | |
3062 | * When a KSM page is created it is shared by 2 mappings. This | |
3063 | * being a signed comparison, it implicitly verifies it's not | |
3064 | * negative. | |
3065 | */ | |
3066 | if (knob < 2) | |
3067 | return -EINVAL; | |
3068 | ||
3069 | if (READ_ONCE(ksm_max_page_sharing) == knob) | |
3070 | return count; | |
3071 | ||
3072 | mutex_lock(&ksm_thread_mutex); | |
3073 | wait_while_offlining(); | |
3074 | if (ksm_max_page_sharing != knob) { | |
3075 | if (ksm_pages_shared || remove_all_stable_nodes()) | |
3076 | err = -EBUSY; | |
3077 | else | |
3078 | ksm_max_page_sharing = knob; | |
3079 | } | |
3080 | mutex_unlock(&ksm_thread_mutex); | |
3081 | ||
3082 | return err ? err : count; | |
3083 | } | |
3084 | KSM_ATTR(max_page_sharing); | |
3085 | ||
b4028260 HD |
3086 | static ssize_t pages_shared_show(struct kobject *kobj, |
3087 | struct kobj_attribute *attr, char *buf) | |
3088 | { | |
ae7a927d | 3089 | return sysfs_emit(buf, "%lu\n", ksm_pages_shared); |
b4028260 HD |
3090 | } |
3091 | KSM_ATTR_RO(pages_shared); | |
3092 | ||
3093 | static ssize_t pages_sharing_show(struct kobject *kobj, | |
3094 | struct kobj_attribute *attr, char *buf) | |
3095 | { | |
ae7a927d | 3096 | return sysfs_emit(buf, "%lu\n", ksm_pages_sharing); |
b4028260 HD |
3097 | } |
3098 | KSM_ATTR_RO(pages_sharing); | |
3099 | ||
473b0ce4 HD |
3100 | static ssize_t pages_unshared_show(struct kobject *kobj, |
3101 | struct kobj_attribute *attr, char *buf) | |
3102 | { | |
ae7a927d | 3103 | return sysfs_emit(buf, "%lu\n", ksm_pages_unshared); |
473b0ce4 HD |
3104 | } |
3105 | KSM_ATTR_RO(pages_unshared); | |
3106 | ||
3107 | static ssize_t pages_volatile_show(struct kobject *kobj, | |
3108 | struct kobj_attribute *attr, char *buf) | |
3109 | { | |
3110 | long ksm_pages_volatile; | |
3111 | ||
3112 | ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared | |
3113 | - ksm_pages_sharing - ksm_pages_unshared; | |
3114 | /* | |
3115 | * It was not worth any locking to calculate that statistic, | |
3116 | * but it might therefore sometimes be negative: conceal that. | |
3117 | */ | |
3118 | if (ksm_pages_volatile < 0) | |
3119 | ksm_pages_volatile = 0; | |
ae7a927d | 3120 | return sysfs_emit(buf, "%ld\n", ksm_pages_volatile); |
473b0ce4 HD |
3121 | } |
3122 | KSM_ATTR_RO(pages_volatile); | |
3123 | ||
2c653d0e AA |
3124 | static ssize_t stable_node_dups_show(struct kobject *kobj, |
3125 | struct kobj_attribute *attr, char *buf) | |
3126 | { | |
ae7a927d | 3127 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups); |
2c653d0e AA |
3128 | } |
3129 | KSM_ATTR_RO(stable_node_dups); | |
3130 | ||
3131 | static ssize_t stable_node_chains_show(struct kobject *kobj, | |
3132 | struct kobj_attribute *attr, char *buf) | |
3133 | { | |
ae7a927d | 3134 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains); |
2c653d0e AA |
3135 | } |
3136 | KSM_ATTR_RO(stable_node_chains); | |
3137 | ||
3138 | static ssize_t | |
3139 | stable_node_chains_prune_millisecs_show(struct kobject *kobj, | |
3140 | struct kobj_attribute *attr, | |
3141 | char *buf) | |
3142 | { | |
ae7a927d | 3143 | return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs); |
2c653d0e AA |
3144 | } |
3145 | ||
3146 | static ssize_t | |
3147 | stable_node_chains_prune_millisecs_store(struct kobject *kobj, | |
3148 | struct kobj_attribute *attr, | |
3149 | const char *buf, size_t count) | |
3150 | { | |
584ff0df | 3151 | unsigned int msecs; |
2c653d0e AA |
3152 | int err; |
3153 | ||
584ff0df ZB |
3154 | err = kstrtouint(buf, 10, &msecs); |
3155 | if (err) | |
2c653d0e AA |
3156 | return -EINVAL; |
3157 | ||
3158 | ksm_stable_node_chains_prune_millisecs = msecs; | |
3159 | ||
3160 | return count; | |
3161 | } | |
3162 | KSM_ATTR(stable_node_chains_prune_millisecs); | |
3163 | ||
473b0ce4 HD |
3164 | static ssize_t full_scans_show(struct kobject *kobj, |
3165 | struct kobj_attribute *attr, char *buf) | |
3166 | { | |
ae7a927d | 3167 | return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr); |
473b0ce4 HD |
3168 | } |
3169 | KSM_ATTR_RO(full_scans); | |
3170 | ||
31dbd01f IE |
3171 | static struct attribute *ksm_attrs[] = { |
3172 | &sleep_millisecs_attr.attr, | |
3173 | &pages_to_scan_attr.attr, | |
3174 | &run_attr.attr, | |
b4028260 HD |
3175 | &pages_shared_attr.attr, |
3176 | &pages_sharing_attr.attr, | |
473b0ce4 HD |
3177 | &pages_unshared_attr.attr, |
3178 | &pages_volatile_attr.attr, | |
3179 | &full_scans_attr.attr, | |
90bd6fd3 PH |
3180 | #ifdef CONFIG_NUMA |
3181 | &merge_across_nodes_attr.attr, | |
3182 | #endif | |
2c653d0e AA |
3183 | &max_page_sharing_attr.attr, |
3184 | &stable_node_chains_attr.attr, | |
3185 | &stable_node_dups_attr.attr, | |
3186 | &stable_node_chains_prune_millisecs_attr.attr, | |
e86c59b1 | 3187 | &use_zero_pages_attr.attr, |
31dbd01f IE |
3188 | NULL, |
3189 | }; | |
3190 | ||
f907c26a | 3191 | static const struct attribute_group ksm_attr_group = { |
31dbd01f IE |
3192 | .attrs = ksm_attrs, |
3193 | .name = "ksm", | |
3194 | }; | |
2ffd8679 | 3195 | #endif /* CONFIG_SYSFS */ |
31dbd01f IE |
3196 | |
3197 | static int __init ksm_init(void) | |
3198 | { | |
3199 | struct task_struct *ksm_thread; | |
3200 | int err; | |
3201 | ||
e86c59b1 CI |
3202 | /* The correct value depends on page size and endianness */ |
3203 | zero_checksum = calc_checksum(ZERO_PAGE(0)); | |
3204 | /* Default to false for backwards compatibility */ | |
3205 | ksm_use_zero_pages = false; | |
3206 | ||
31dbd01f IE |
3207 | err = ksm_slab_init(); |
3208 | if (err) | |
3209 | goto out; | |
3210 | ||
31dbd01f IE |
3211 | ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); |
3212 | if (IS_ERR(ksm_thread)) { | |
25acde31 | 3213 | pr_err("ksm: creating kthread failed\n"); |
31dbd01f | 3214 | err = PTR_ERR(ksm_thread); |
d9f8984c | 3215 | goto out_free; |
31dbd01f IE |
3216 | } |
3217 | ||
2ffd8679 | 3218 | #ifdef CONFIG_SYSFS |
31dbd01f IE |
3219 | err = sysfs_create_group(mm_kobj, &ksm_attr_group); |
3220 | if (err) { | |
25acde31 | 3221 | pr_err("ksm: register sysfs failed\n"); |
2ffd8679 | 3222 | kthread_stop(ksm_thread); |
d9f8984c | 3223 | goto out_free; |
31dbd01f | 3224 | } |
c73602ad HD |
3225 | #else |
3226 | ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ | |
3227 | ||
2ffd8679 | 3228 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 3229 | |
62b61f61 | 3230 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 | 3231 | /* There is no significance to this priority 100 */ |
62b61f61 HD |
3232 | hotplug_memory_notifier(ksm_memory_callback, 100); |
3233 | #endif | |
31dbd01f IE |
3234 | return 0; |
3235 | ||
d9f8984c | 3236 | out_free: |
31dbd01f IE |
3237 | ksm_slab_free(); |
3238 | out: | |
3239 | return err; | |
f8af4da3 | 3240 | } |
a64fb3cd | 3241 | subsys_initcall(ksm_init); |