Commit | Line | Data |
---|---|---|
7a338472 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
f8af4da3 | 2 | /* |
31dbd01f IE |
3 | * Memory merging support. |
4 | * | |
5 | * This code enables dynamic sharing of identical pages found in different | |
6 | * memory areas, even if they are not shared by fork() | |
7 | * | |
36b2528d | 8 | * Copyright (C) 2008-2009 Red Hat, Inc. |
31dbd01f IE |
9 | * Authors: |
10 | * Izik Eidus | |
11 | * Andrea Arcangeli | |
12 | * Chris Wright | |
36b2528d | 13 | * Hugh Dickins |
f8af4da3 HD |
14 | */ |
15 | ||
16 | #include <linux/errno.h> | |
31dbd01f | 17 | #include <linux/mm.h> |
36090def | 18 | #include <linux/mm_inline.h> |
31dbd01f | 19 | #include <linux/fs.h> |
f8af4da3 | 20 | #include <linux/mman.h> |
31dbd01f | 21 | #include <linux/sched.h> |
6e84f315 | 22 | #include <linux/sched/mm.h> |
f7ccbae4 | 23 | #include <linux/sched/coredump.h> |
31dbd01f IE |
24 | #include <linux/rwsem.h> |
25 | #include <linux/pagemap.h> | |
26 | #include <linux/rmap.h> | |
27 | #include <linux/spinlock.h> | |
59e1a2f4 | 28 | #include <linux/xxhash.h> |
31dbd01f IE |
29 | #include <linux/delay.h> |
30 | #include <linux/kthread.h> | |
31 | #include <linux/wait.h> | |
32 | #include <linux/slab.h> | |
33 | #include <linux/rbtree.h> | |
62b61f61 | 34 | #include <linux/memory.h> |
31dbd01f | 35 | #include <linux/mmu_notifier.h> |
2c6854fd | 36 | #include <linux/swap.h> |
f8af4da3 | 37 | #include <linux/ksm.h> |
4ca3a69b | 38 | #include <linux/hashtable.h> |
878aee7d | 39 | #include <linux/freezer.h> |
72788c38 | 40 | #include <linux/oom.h> |
90bd6fd3 | 41 | #include <linux/numa.h> |
d7c0e68d | 42 | #include <linux/pagewalk.h> |
f8af4da3 | 43 | |
31dbd01f | 44 | #include <asm/tlbflush.h> |
73848b46 | 45 | #include "internal.h" |
58730ab6 | 46 | #include "mm_slot.h" |
31dbd01f | 47 | |
739100c8 SR |
48 | #define CREATE_TRACE_POINTS |
49 | #include <trace/events/ksm.h> | |
50 | ||
e850dcf5 HD |
51 | #ifdef CONFIG_NUMA |
52 | #define NUMA(x) (x) | |
53 | #define DO_NUMA(x) do { (x); } while (0) | |
54 | #else | |
55 | #define NUMA(x) (0) | |
56 | #define DO_NUMA(x) do { } while (0) | |
57 | #endif | |
58 | ||
5a2ca3ef MR |
59 | /** |
60 | * DOC: Overview | |
61 | * | |
31dbd01f IE |
62 | * A few notes about the KSM scanning process, |
63 | * to make it easier to understand the data structures below: | |
64 | * | |
65 | * In order to reduce excessive scanning, KSM sorts the memory pages by their | |
66 | * contents into a data structure that holds pointers to the pages' locations. | |
67 | * | |
68 | * Since the contents of the pages may change at any moment, KSM cannot just | |
69 | * insert the pages into a normal sorted tree and expect it to find anything. | |
70 | * Therefore KSM uses two data structures - the stable and the unstable tree. | |
71 | * | |
72 | * The stable tree holds pointers to all the merged pages (ksm pages), sorted | |
73 | * by their contents. Because each such page is write-protected, searching on | |
74 | * this tree is fully assured to be working (except when pages are unmapped), | |
75 | * and therefore this tree is called the stable tree. | |
76 | * | |
5a2ca3ef MR |
77 | * The stable tree node includes information required for reverse |
78 | * mapping from a KSM page to virtual addresses that map this page. | |
79 | * | |
80 | * In order to avoid large latencies of the rmap walks on KSM pages, | |
81 | * KSM maintains two types of nodes in the stable tree: | |
82 | * | |
83 | * * the regular nodes that keep the reverse mapping structures in a | |
84 | * linked list | |
85 | * * the "chains" that link nodes ("dups") that represent the same | |
86 | * write protected memory content, but each "dup" corresponds to a | |
87 | * different KSM page copy of that content | |
88 | * | |
89 | * Internally, the regular nodes, "dups" and "chains" are represented | |
21fbd591 | 90 | * using the same struct ksm_stable_node structure. |
5a2ca3ef | 91 | * |
31dbd01f IE |
92 | * In addition to the stable tree, KSM uses a second data structure called the |
93 | * unstable tree: this tree holds pointers to pages which have been found to | |
94 | * be "unchanged for a period of time". The unstable tree sorts these pages | |
95 | * by their contents, but since they are not write-protected, KSM cannot rely | |
96 | * upon the unstable tree to work correctly - the unstable tree is liable to | |
97 | * be corrupted as its contents are modified, and so it is called unstable. | |
98 | * | |
99 | * KSM solves this problem by several techniques: | |
100 | * | |
101 | * 1) The unstable tree is flushed every time KSM completes scanning all | |
102 | * memory areas, and then the tree is rebuilt again from the beginning. | |
103 | * 2) KSM will only insert into the unstable tree, pages whose hash value | |
104 | * has not changed since the previous scan of all memory areas. | |
105 | * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the | |
106 | * colors of the nodes and not on their contents, assuring that even when | |
107 | * the tree gets "corrupted" it won't get out of balance, so scanning time | |
108 | * remains the same (also, searching and inserting nodes in an rbtree uses | |
109 | * the same algorithm, so we have no overhead when we flush and rebuild). | |
110 | * 4) KSM never flushes the stable tree, which means that even if it were to | |
111 | * take 10 attempts to find a page in the unstable tree, once it is found, | |
112 | * it is secured in the stable tree. (When we scan a new page, we first | |
113 | * compare it against the stable tree, and then against the unstable tree.) | |
8fdb3dbf HD |
114 | * |
115 | * If the merge_across_nodes tunable is unset, then KSM maintains multiple | |
116 | * stable trees and multiple unstable trees: one of each for each NUMA node. | |
31dbd01f IE |
117 | */ |
118 | ||
119 | /** | |
21fbd591 | 120 | * struct ksm_mm_slot - ksm information per mm that is being scanned |
58730ab6 | 121 | * @slot: hash lookup from mm to mm_slot |
6514d511 | 122 | * @rmap_list: head for this mm_slot's singly-linked list of rmap_items |
31dbd01f | 123 | */ |
21fbd591 | 124 | struct ksm_mm_slot { |
58730ab6 | 125 | struct mm_slot slot; |
21fbd591 | 126 | struct ksm_rmap_item *rmap_list; |
31dbd01f IE |
127 | }; |
128 | ||
129 | /** | |
130 | * struct ksm_scan - cursor for scanning | |
131 | * @mm_slot: the current mm_slot we are scanning | |
132 | * @address: the next address inside that to be scanned | |
6514d511 | 133 | * @rmap_list: link to the next rmap to be scanned in the rmap_list |
31dbd01f IE |
134 | * @seqnr: count of completed full scans (needed when removing unstable node) |
135 | * | |
136 | * There is only the one ksm_scan instance of this cursor structure. | |
137 | */ | |
138 | struct ksm_scan { | |
21fbd591 | 139 | struct ksm_mm_slot *mm_slot; |
31dbd01f | 140 | unsigned long address; |
21fbd591 | 141 | struct ksm_rmap_item **rmap_list; |
31dbd01f IE |
142 | unsigned long seqnr; |
143 | }; | |
144 | ||
7b6ba2c7 | 145 | /** |
21fbd591 | 146 | * struct ksm_stable_node - node of the stable rbtree |
7b6ba2c7 | 147 | * @node: rb node of this ksm page in the stable tree |
4146d2d6 | 148 | * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list |
2c653d0e | 149 | * @hlist_dup: linked into the stable_node->hlist with a stable_node chain |
4146d2d6 | 150 | * @list: linked into migrate_nodes, pending placement in the proper node tree |
7b6ba2c7 | 151 | * @hlist: hlist head of rmap_items using this ksm page |
4146d2d6 | 152 | * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) |
2c653d0e AA |
153 | * @chain_prune_time: time of the last full garbage collection |
154 | * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN | |
4146d2d6 | 155 | * @nid: NUMA node id of stable tree in which linked (may not match kpfn) |
7b6ba2c7 | 156 | */ |
21fbd591 | 157 | struct ksm_stable_node { |
4146d2d6 HD |
158 | union { |
159 | struct rb_node node; /* when node of stable tree */ | |
160 | struct { /* when listed for migration */ | |
161 | struct list_head *head; | |
2c653d0e AA |
162 | struct { |
163 | struct hlist_node hlist_dup; | |
164 | struct list_head list; | |
165 | }; | |
4146d2d6 HD |
166 | }; |
167 | }; | |
7b6ba2c7 | 168 | struct hlist_head hlist; |
2c653d0e AA |
169 | union { |
170 | unsigned long kpfn; | |
171 | unsigned long chain_prune_time; | |
172 | }; | |
173 | /* | |
174 | * STABLE_NODE_CHAIN can be any negative number in | |
175 | * rmap_hlist_len negative range, but better not -1 to be able | |
176 | * to reliably detect underflows. | |
177 | */ | |
178 | #define STABLE_NODE_CHAIN -1024 | |
179 | int rmap_hlist_len; | |
4146d2d6 HD |
180 | #ifdef CONFIG_NUMA |
181 | int nid; | |
182 | #endif | |
7b6ba2c7 HD |
183 | }; |
184 | ||
31dbd01f | 185 | /** |
21fbd591 | 186 | * struct ksm_rmap_item - reverse mapping item for virtual addresses |
6514d511 | 187 | * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list |
db114b83 | 188 | * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree |
bc56620b | 189 | * @nid: NUMA node id of unstable tree in which linked (may not match page) |
31dbd01f IE |
190 | * @mm: the memory structure this rmap_item is pointing into |
191 | * @address: the virtual address this rmap_item tracks (+ flags in low bits) | |
192 | * @oldchecksum: previous checksum of the page at that virtual address | |
7b6ba2c7 HD |
193 | * @node: rb node of this rmap_item in the unstable tree |
194 | * @head: pointer to stable_node heading this list in the stable tree | |
195 | * @hlist: link into hlist of rmap_items hanging off that stable_node | |
31dbd01f | 196 | */ |
21fbd591 QZ |
197 | struct ksm_rmap_item { |
198 | struct ksm_rmap_item *rmap_list; | |
bc56620b HD |
199 | union { |
200 | struct anon_vma *anon_vma; /* when stable */ | |
201 | #ifdef CONFIG_NUMA | |
202 | int nid; /* when node of unstable tree */ | |
203 | #endif | |
204 | }; | |
31dbd01f IE |
205 | struct mm_struct *mm; |
206 | unsigned long address; /* + low bits used for flags below */ | |
7b6ba2c7 | 207 | unsigned int oldchecksum; /* when unstable */ |
31dbd01f | 208 | union { |
7b6ba2c7 HD |
209 | struct rb_node node; /* when node of unstable tree */ |
210 | struct { /* when listed from stable tree */ | |
21fbd591 | 211 | struct ksm_stable_node *head; |
7b6ba2c7 HD |
212 | struct hlist_node hlist; |
213 | }; | |
31dbd01f IE |
214 | }; |
215 | }; | |
216 | ||
217 | #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ | |
7b6ba2c7 HD |
218 | #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ |
219 | #define STABLE_FLAG 0x200 /* is listed from the stable tree */ | |
31dbd01f IE |
220 | |
221 | /* The stable and unstable tree heads */ | |
ef53d16c HD |
222 | static struct rb_root one_stable_tree[1] = { RB_ROOT }; |
223 | static struct rb_root one_unstable_tree[1] = { RB_ROOT }; | |
224 | static struct rb_root *root_stable_tree = one_stable_tree; | |
225 | static struct rb_root *root_unstable_tree = one_unstable_tree; | |
31dbd01f | 226 | |
4146d2d6 HD |
227 | /* Recently migrated nodes of stable tree, pending proper placement */ |
228 | static LIST_HEAD(migrate_nodes); | |
2c653d0e | 229 | #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev) |
4146d2d6 | 230 | |
4ca3a69b SL |
231 | #define MM_SLOTS_HASH_BITS 10 |
232 | static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
31dbd01f | 233 | |
21fbd591 | 234 | static struct ksm_mm_slot ksm_mm_head = { |
58730ab6 | 235 | .slot.mm_node = LIST_HEAD_INIT(ksm_mm_head.slot.mm_node), |
31dbd01f IE |
236 | }; |
237 | static struct ksm_scan ksm_scan = { | |
238 | .mm_slot = &ksm_mm_head, | |
239 | }; | |
240 | ||
241 | static struct kmem_cache *rmap_item_cache; | |
7b6ba2c7 | 242 | static struct kmem_cache *stable_node_cache; |
31dbd01f IE |
243 | static struct kmem_cache *mm_slot_cache; |
244 | ||
b348b5fe SR |
245 | /* The number of pages scanned */ |
246 | static unsigned long ksm_pages_scanned; | |
247 | ||
31dbd01f | 248 | /* The number of nodes in the stable tree */ |
b4028260 | 249 | static unsigned long ksm_pages_shared; |
31dbd01f | 250 | |
e178dfde | 251 | /* The number of page slots additionally sharing those nodes */ |
b4028260 | 252 | static unsigned long ksm_pages_sharing; |
31dbd01f | 253 | |
473b0ce4 HD |
254 | /* The number of nodes in the unstable tree */ |
255 | static unsigned long ksm_pages_unshared; | |
256 | ||
257 | /* The number of rmap_items in use: to calculate pages_volatile */ | |
258 | static unsigned long ksm_rmap_items; | |
259 | ||
2c653d0e AA |
260 | /* The number of stable_node chains */ |
261 | static unsigned long ksm_stable_node_chains; | |
262 | ||
263 | /* The number of stable_node dups linked to the stable_node chains */ | |
264 | static unsigned long ksm_stable_node_dups; | |
265 | ||
266 | /* Delay in pruning stale stable_node_dups in the stable_node_chains */ | |
584ff0df | 267 | static unsigned int ksm_stable_node_chains_prune_millisecs = 2000; |
2c653d0e AA |
268 | |
269 | /* Maximum number of page slots sharing a stable node */ | |
270 | static int ksm_max_page_sharing = 256; | |
271 | ||
31dbd01f | 272 | /* Number of pages ksmd should scan in one batch */ |
2c6854fd | 273 | static unsigned int ksm_thread_pages_to_scan = 100; |
31dbd01f IE |
274 | |
275 | /* Milliseconds ksmd should sleep between batches */ | |
2ffd8679 | 276 | static unsigned int ksm_thread_sleep_millisecs = 20; |
31dbd01f | 277 | |
e86c59b1 CI |
278 | /* Checksum of an empty (zeroed) page */ |
279 | static unsigned int zero_checksum __read_mostly; | |
280 | ||
281 | /* Whether to merge empty (zeroed) pages with actual zero pages */ | |
282 | static bool ksm_use_zero_pages __read_mostly; | |
283 | ||
e2942062 | 284 | /* The number of zero pages which is placed by KSM */ |
285 | unsigned long ksm_zero_pages; | |
286 | ||
e850dcf5 | 287 | #ifdef CONFIG_NUMA |
90bd6fd3 PH |
288 | /* Zeroed when merging across nodes is not allowed */ |
289 | static unsigned int ksm_merge_across_nodes = 1; | |
ef53d16c | 290 | static int ksm_nr_node_ids = 1; |
e850dcf5 HD |
291 | #else |
292 | #define ksm_merge_across_nodes 1U | |
ef53d16c | 293 | #define ksm_nr_node_ids 1 |
e850dcf5 | 294 | #endif |
90bd6fd3 | 295 | |
31dbd01f IE |
296 | #define KSM_RUN_STOP 0 |
297 | #define KSM_RUN_MERGE 1 | |
298 | #define KSM_RUN_UNMERGE 2 | |
ef4d43a8 HD |
299 | #define KSM_RUN_OFFLINE 4 |
300 | static unsigned long ksm_run = KSM_RUN_STOP; | |
301 | static void wait_while_offlining(void); | |
31dbd01f IE |
302 | |
303 | static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); | |
fcf9a0ef | 304 | static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait); |
31dbd01f IE |
305 | static DEFINE_MUTEX(ksm_thread_mutex); |
306 | static DEFINE_SPINLOCK(ksm_mmlist_lock); | |
307 | ||
21fbd591 | 308 | #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create(#__struct,\ |
31dbd01f IE |
309 | sizeof(struct __struct), __alignof__(struct __struct),\ |
310 | (__flags), NULL) | |
311 | ||
312 | static int __init ksm_slab_init(void) | |
313 | { | |
21fbd591 | 314 | rmap_item_cache = KSM_KMEM_CACHE(ksm_rmap_item, 0); |
31dbd01f IE |
315 | if (!rmap_item_cache) |
316 | goto out; | |
317 | ||
21fbd591 | 318 | stable_node_cache = KSM_KMEM_CACHE(ksm_stable_node, 0); |
7b6ba2c7 HD |
319 | if (!stable_node_cache) |
320 | goto out_free1; | |
321 | ||
21fbd591 | 322 | mm_slot_cache = KSM_KMEM_CACHE(ksm_mm_slot, 0); |
31dbd01f | 323 | if (!mm_slot_cache) |
7b6ba2c7 | 324 | goto out_free2; |
31dbd01f IE |
325 | |
326 | return 0; | |
327 | ||
7b6ba2c7 HD |
328 | out_free2: |
329 | kmem_cache_destroy(stable_node_cache); | |
330 | out_free1: | |
31dbd01f IE |
331 | kmem_cache_destroy(rmap_item_cache); |
332 | out: | |
333 | return -ENOMEM; | |
334 | } | |
335 | ||
336 | static void __init ksm_slab_free(void) | |
337 | { | |
338 | kmem_cache_destroy(mm_slot_cache); | |
7b6ba2c7 | 339 | kmem_cache_destroy(stable_node_cache); |
31dbd01f IE |
340 | kmem_cache_destroy(rmap_item_cache); |
341 | mm_slot_cache = NULL; | |
342 | } | |
343 | ||
21fbd591 | 344 | static __always_inline bool is_stable_node_chain(struct ksm_stable_node *chain) |
2c653d0e AA |
345 | { |
346 | return chain->rmap_hlist_len == STABLE_NODE_CHAIN; | |
347 | } | |
348 | ||
21fbd591 | 349 | static __always_inline bool is_stable_node_dup(struct ksm_stable_node *dup) |
2c653d0e AA |
350 | { |
351 | return dup->head == STABLE_NODE_DUP_HEAD; | |
352 | } | |
353 | ||
21fbd591 QZ |
354 | static inline void stable_node_chain_add_dup(struct ksm_stable_node *dup, |
355 | struct ksm_stable_node *chain) | |
2c653d0e AA |
356 | { |
357 | VM_BUG_ON(is_stable_node_dup(dup)); | |
358 | dup->head = STABLE_NODE_DUP_HEAD; | |
359 | VM_BUG_ON(!is_stable_node_chain(chain)); | |
360 | hlist_add_head(&dup->hlist_dup, &chain->hlist); | |
361 | ksm_stable_node_dups++; | |
362 | } | |
363 | ||
21fbd591 | 364 | static inline void __stable_node_dup_del(struct ksm_stable_node *dup) |
2c653d0e | 365 | { |
b4fecc67 | 366 | VM_BUG_ON(!is_stable_node_dup(dup)); |
2c653d0e AA |
367 | hlist_del(&dup->hlist_dup); |
368 | ksm_stable_node_dups--; | |
369 | } | |
370 | ||
21fbd591 | 371 | static inline void stable_node_dup_del(struct ksm_stable_node *dup) |
2c653d0e AA |
372 | { |
373 | VM_BUG_ON(is_stable_node_chain(dup)); | |
374 | if (is_stable_node_dup(dup)) | |
375 | __stable_node_dup_del(dup); | |
376 | else | |
377 | rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid)); | |
378 | #ifdef CONFIG_DEBUG_VM | |
379 | dup->head = NULL; | |
380 | #endif | |
381 | } | |
382 | ||
21fbd591 | 383 | static inline struct ksm_rmap_item *alloc_rmap_item(void) |
31dbd01f | 384 | { |
21fbd591 | 385 | struct ksm_rmap_item *rmap_item; |
473b0ce4 | 386 | |
5b398e41 | 387 | rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | |
388 | __GFP_NORETRY | __GFP_NOWARN); | |
473b0ce4 HD |
389 | if (rmap_item) |
390 | ksm_rmap_items++; | |
391 | return rmap_item; | |
31dbd01f IE |
392 | } |
393 | ||
21fbd591 | 394 | static inline void free_rmap_item(struct ksm_rmap_item *rmap_item) |
31dbd01f | 395 | { |
473b0ce4 | 396 | ksm_rmap_items--; |
cb4df4ca | 397 | rmap_item->mm->ksm_rmap_items--; |
31dbd01f IE |
398 | rmap_item->mm = NULL; /* debug safety */ |
399 | kmem_cache_free(rmap_item_cache, rmap_item); | |
400 | } | |
401 | ||
21fbd591 | 402 | static inline struct ksm_stable_node *alloc_stable_node(void) |
7b6ba2c7 | 403 | { |
6213055f | 404 | /* |
405 | * The allocation can take too long with GFP_KERNEL when memory is under | |
406 | * pressure, which may lead to hung task warnings. Adding __GFP_HIGH | |
407 | * grants access to memory reserves, helping to avoid this problem. | |
408 | */ | |
409 | return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); | |
7b6ba2c7 HD |
410 | } |
411 | ||
21fbd591 | 412 | static inline void free_stable_node(struct ksm_stable_node *stable_node) |
7b6ba2c7 | 413 | { |
2c653d0e AA |
414 | VM_BUG_ON(stable_node->rmap_hlist_len && |
415 | !is_stable_node_chain(stable_node)); | |
7b6ba2c7 HD |
416 | kmem_cache_free(stable_node_cache, stable_node); |
417 | } | |
418 | ||
a913e182 HD |
419 | /* |
420 | * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's | |
421 | * page tables after it has passed through ksm_exit() - which, if necessary, | |
c1e8d7c6 | 422 | * takes mmap_lock briefly to serialize against them. ksm_exit() does not set |
a913e182 HD |
423 | * a special flag: they can just back out as soon as mm_users goes to zero. |
424 | * ksm_test_exit() is used throughout to make this test for exit: in some | |
425 | * places for correctness, in some places just to avoid unnecessary work. | |
426 | */ | |
427 | static inline bool ksm_test_exit(struct mm_struct *mm) | |
428 | { | |
429 | return atomic_read(&mm->mm_users) == 0; | |
430 | } | |
431 | ||
d7c0e68d DH |
432 | static int break_ksm_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long next, |
433 | struct mm_walk *walk) | |
434 | { | |
435 | struct page *page = NULL; | |
436 | spinlock_t *ptl; | |
437 | pte_t *pte; | |
c33c7948 | 438 | pte_t ptent; |
d7c0e68d DH |
439 | int ret; |
440 | ||
d7c0e68d | 441 | pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); |
04dee9e8 HD |
442 | if (!pte) |
443 | return 0; | |
c33c7948 RR |
444 | ptent = ptep_get(pte); |
445 | if (pte_present(ptent)) { | |
446 | page = vm_normal_page(walk->vma, addr, ptent); | |
447 | } else if (!pte_none(ptent)) { | |
448 | swp_entry_t entry = pte_to_swp_entry(ptent); | |
d7c0e68d DH |
449 | |
450 | /* | |
451 | * As KSM pages remain KSM pages until freed, no need to wait | |
452 | * here for migration to end. | |
453 | */ | |
454 | if (is_migration_entry(entry)) | |
455 | page = pfn_swap_entry_to_page(entry); | |
456 | } | |
79271476 | 457 | /* return 1 if the page is an normal ksm page or KSM-placed zero page */ |
458 | ret = (page && PageKsm(page)) || is_ksm_zero_pte(*pte); | |
d7c0e68d DH |
459 | pte_unmap_unlock(pte, ptl); |
460 | return ret; | |
461 | } | |
462 | ||
463 | static const struct mm_walk_ops break_ksm_ops = { | |
464 | .pmd_entry = break_ksm_pmd_entry, | |
49b06385 SB |
465 | .walk_lock = PGWALK_RDLOCK, |
466 | }; | |
467 | ||
468 | static const struct mm_walk_ops break_ksm_lock_vma_ops = { | |
469 | .pmd_entry = break_ksm_pmd_entry, | |
470 | .walk_lock = PGWALK_WRLOCK, | |
d7c0e68d DH |
471 | }; |
472 | ||
31dbd01f | 473 | /* |
6cce3314 DH |
474 | * We use break_ksm to break COW on a ksm page by triggering unsharing, |
475 | * such that the ksm page will get replaced by an exclusive anonymous page. | |
31dbd01f | 476 | * |
6cce3314 | 477 | * We take great care only to touch a ksm page, in a VM_MERGEABLE vma, |
31dbd01f IE |
478 | * in case the application has unmapped and remapped mm,addr meanwhile. |
479 | * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP | |
bbcd53c9 | 480 | * mmap of /dev/mem, where we would not want to touch it. |
1b2ee126 | 481 | * |
6cce3314 | 482 | * FAULT_FLAG_REMOTE/FOLL_REMOTE are because we do this outside the context |
1b2ee126 DH |
483 | * of the process that owns 'vma'. We also do not want to enforce |
484 | * protection keys here anyway. | |
31dbd01f | 485 | */ |
49b06385 | 486 | static int break_ksm(struct vm_area_struct *vma, unsigned long addr, bool lock_vma) |
31dbd01f | 487 | { |
50a7ca3c | 488 | vm_fault_t ret = 0; |
49b06385 SB |
489 | const struct mm_walk_ops *ops = lock_vma ? |
490 | &break_ksm_lock_vma_ops : &break_ksm_ops; | |
31dbd01f IE |
491 | |
492 | do { | |
d7c0e68d | 493 | int ksm_page; |
58f595c6 | 494 | |
31dbd01f | 495 | cond_resched(); |
49b06385 | 496 | ksm_page = walk_page_range_vma(vma, addr, addr + 1, ops, NULL); |
d7c0e68d DH |
497 | if (WARN_ON_ONCE(ksm_page < 0)) |
498 | return ksm_page; | |
58f595c6 DH |
499 | if (!ksm_page) |
500 | return 0; | |
501 | ret = handle_mm_fault(vma, addr, | |
6cce3314 | 502 | FAULT_FLAG_UNSHARE | FAULT_FLAG_REMOTE, |
58f595c6 DH |
503 | NULL); |
504 | } while (!(ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); | |
d952b791 | 505 | /* |
58f595c6 DH |
506 | * We must loop until we no longer find a KSM page because |
507 | * handle_mm_fault() may back out if there's any difficulty e.g. if | |
508 | * pte accessed bit gets updated concurrently. | |
d952b791 HD |
509 | * |
510 | * VM_FAULT_SIGBUS could occur if we race with truncation of the | |
511 | * backing file, which also invalidates anonymous pages: that's | |
512 | * okay, that truncation will have unmapped the PageKsm for us. | |
513 | * | |
514 | * VM_FAULT_OOM: at the time of writing (late July 2009), setting | |
515 | * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the | |
516 | * current task has TIF_MEMDIE set, and will be OOM killed on return | |
517 | * to user; and ksmd, having no mm, would never be chosen for that. | |
518 | * | |
519 | * But if the mm is in a limited mem_cgroup, then the fault may fail | |
520 | * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and | |
521 | * even ksmd can fail in this way - though it's usually breaking ksm | |
522 | * just to undo a merge it made a moment before, so unlikely to oom. | |
523 | * | |
524 | * That's a pity: we might therefore have more kernel pages allocated | |
525 | * than we're counting as nodes in the stable tree; but ksm_do_scan | |
526 | * will retry to break_cow on each pass, so should recover the page | |
527 | * in due course. The important thing is to not let VM_MERGEABLE | |
528 | * be cleared while any such pages might remain in the area. | |
529 | */ | |
530 | return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; | |
31dbd01f IE |
531 | } |
532 | ||
d7597f59 SR |
533 | static bool vma_ksm_compatible(struct vm_area_struct *vma) |
534 | { | |
535 | if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE | VM_PFNMAP | | |
536 | VM_IO | VM_DONTEXPAND | VM_HUGETLB | | |
537 | VM_MIXEDMAP)) | |
538 | return false; /* just ignore the advice */ | |
539 | ||
540 | if (vma_is_dax(vma)) | |
541 | return false; | |
542 | ||
543 | #ifdef VM_SAO | |
544 | if (vma->vm_flags & VM_SAO) | |
545 | return false; | |
546 | #endif | |
547 | #ifdef VM_SPARC_ADI | |
548 | if (vma->vm_flags & VM_SPARC_ADI) | |
549 | return false; | |
550 | #endif | |
551 | ||
552 | return true; | |
553 | } | |
554 | ||
ef694222 BL |
555 | static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, |
556 | unsigned long addr) | |
557 | { | |
558 | struct vm_area_struct *vma; | |
559 | if (ksm_test_exit(mm)) | |
560 | return NULL; | |
ff69fb81 LH |
561 | vma = vma_lookup(mm, addr); |
562 | if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) | |
ef694222 BL |
563 | return NULL; |
564 | return vma; | |
565 | } | |
566 | ||
21fbd591 | 567 | static void break_cow(struct ksm_rmap_item *rmap_item) |
31dbd01f | 568 | { |
8dd3557a HD |
569 | struct mm_struct *mm = rmap_item->mm; |
570 | unsigned long addr = rmap_item->address; | |
31dbd01f IE |
571 | struct vm_area_struct *vma; |
572 | ||
4035c07a HD |
573 | /* |
574 | * It is not an accident that whenever we want to break COW | |
575 | * to undo, we also need to drop a reference to the anon_vma. | |
576 | */ | |
9e60109f | 577 | put_anon_vma(rmap_item->anon_vma); |
4035c07a | 578 | |
d8ed45c5 | 579 | mmap_read_lock(mm); |
ef694222 BL |
580 | vma = find_mergeable_vma(mm, addr); |
581 | if (vma) | |
49b06385 | 582 | break_ksm(vma, addr, false); |
d8ed45c5 | 583 | mmap_read_unlock(mm); |
31dbd01f IE |
584 | } |
585 | ||
21fbd591 | 586 | static struct page *get_mergeable_page(struct ksm_rmap_item *rmap_item) |
31dbd01f IE |
587 | { |
588 | struct mm_struct *mm = rmap_item->mm; | |
589 | unsigned long addr = rmap_item->address; | |
590 | struct vm_area_struct *vma; | |
591 | struct page *page; | |
592 | ||
d8ed45c5 | 593 | mmap_read_lock(mm); |
ef694222 BL |
594 | vma = find_mergeable_vma(mm, addr); |
595 | if (!vma) | |
31dbd01f IE |
596 | goto out; |
597 | ||
598 | page = follow_page(vma, addr, FOLL_GET); | |
f7091ed6 | 599 | if (IS_ERR_OR_NULL(page)) |
31dbd01f | 600 | goto out; |
f7091ed6 HW |
601 | if (is_zone_device_page(page)) |
602 | goto out_putpage; | |
f765f540 | 603 | if (PageAnon(page)) { |
31dbd01f IE |
604 | flush_anon_page(vma, page, addr); |
605 | flush_dcache_page(page); | |
606 | } else { | |
f7091ed6 | 607 | out_putpage: |
31dbd01f | 608 | put_page(page); |
c8f95ed1 AA |
609 | out: |
610 | page = NULL; | |
31dbd01f | 611 | } |
d8ed45c5 | 612 | mmap_read_unlock(mm); |
31dbd01f IE |
613 | return page; |
614 | } | |
615 | ||
90bd6fd3 PH |
616 | /* |
617 | * This helper is used for getting right index into array of tree roots. | |
618 | * When merge_across_nodes knob is set to 1, there are only two rb-trees for | |
619 | * stable and unstable pages from all nodes with roots in index 0. Otherwise, | |
620 | * every node has its own stable and unstable tree. | |
621 | */ | |
622 | static inline int get_kpfn_nid(unsigned long kpfn) | |
623 | { | |
d8fc16a8 | 624 | return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); |
90bd6fd3 PH |
625 | } |
626 | ||
21fbd591 | 627 | static struct ksm_stable_node *alloc_stable_node_chain(struct ksm_stable_node *dup, |
2c653d0e AA |
628 | struct rb_root *root) |
629 | { | |
21fbd591 | 630 | struct ksm_stable_node *chain = alloc_stable_node(); |
2c653d0e AA |
631 | VM_BUG_ON(is_stable_node_chain(dup)); |
632 | if (likely(chain)) { | |
633 | INIT_HLIST_HEAD(&chain->hlist); | |
634 | chain->chain_prune_time = jiffies; | |
635 | chain->rmap_hlist_len = STABLE_NODE_CHAIN; | |
636 | #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA) | |
98fa15f3 | 637 | chain->nid = NUMA_NO_NODE; /* debug */ |
2c653d0e AA |
638 | #endif |
639 | ksm_stable_node_chains++; | |
640 | ||
641 | /* | |
642 | * Put the stable node chain in the first dimension of | |
643 | * the stable tree and at the same time remove the old | |
644 | * stable node. | |
645 | */ | |
646 | rb_replace_node(&dup->node, &chain->node, root); | |
647 | ||
648 | /* | |
649 | * Move the old stable node to the second dimension | |
650 | * queued in the hlist_dup. The invariant is that all | |
651 | * dup stable_nodes in the chain->hlist point to pages | |
457aef94 | 652 | * that are write protected and have the exact same |
2c653d0e AA |
653 | * content. |
654 | */ | |
655 | stable_node_chain_add_dup(dup, chain); | |
656 | } | |
657 | return chain; | |
658 | } | |
659 | ||
21fbd591 | 660 | static inline void free_stable_node_chain(struct ksm_stable_node *chain, |
2c653d0e AA |
661 | struct rb_root *root) |
662 | { | |
663 | rb_erase(&chain->node, root); | |
664 | free_stable_node(chain); | |
665 | ksm_stable_node_chains--; | |
666 | } | |
667 | ||
21fbd591 | 668 | static void remove_node_from_stable_tree(struct ksm_stable_node *stable_node) |
4035c07a | 669 | { |
21fbd591 | 670 | struct ksm_rmap_item *rmap_item; |
4035c07a | 671 | |
2c653d0e AA |
672 | /* check it's not STABLE_NODE_CHAIN or negative */ |
673 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
674 | ||
b67bfe0d | 675 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
739100c8 | 676 | if (rmap_item->hlist.next) { |
4035c07a | 677 | ksm_pages_sharing--; |
739100c8 SR |
678 | trace_ksm_remove_rmap_item(stable_node->kpfn, rmap_item, rmap_item->mm); |
679 | } else { | |
4035c07a | 680 | ksm_pages_shared--; |
739100c8 | 681 | } |
76093853 | 682 | |
683 | rmap_item->mm->ksm_merging_pages--; | |
684 | ||
2c653d0e AA |
685 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
686 | stable_node->rmap_hlist_len--; | |
9e60109f | 687 | put_anon_vma(rmap_item->anon_vma); |
4035c07a HD |
688 | rmap_item->address &= PAGE_MASK; |
689 | cond_resched(); | |
690 | } | |
691 | ||
2c653d0e AA |
692 | /* |
693 | * We need the second aligned pointer of the migrate_nodes | |
694 | * list_head to stay clear from the rb_parent_color union | |
695 | * (aligned and different than any node) and also different | |
696 | * from &migrate_nodes. This will verify that future list.h changes | |
815f0ddb | 697 | * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it. |
2c653d0e | 698 | */ |
2c653d0e AA |
699 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes); |
700 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1); | |
2c653d0e | 701 | |
739100c8 | 702 | trace_ksm_remove_ksm_page(stable_node->kpfn); |
4146d2d6 HD |
703 | if (stable_node->head == &migrate_nodes) |
704 | list_del(&stable_node->list); | |
705 | else | |
2c653d0e | 706 | stable_node_dup_del(stable_node); |
4035c07a HD |
707 | free_stable_node(stable_node); |
708 | } | |
709 | ||
2cee57d1 YS |
710 | enum get_ksm_page_flags { |
711 | GET_KSM_PAGE_NOLOCK, | |
712 | GET_KSM_PAGE_LOCK, | |
713 | GET_KSM_PAGE_TRYLOCK | |
714 | }; | |
715 | ||
4035c07a HD |
716 | /* |
717 | * get_ksm_page: checks if the page indicated by the stable node | |
718 | * is still its ksm page, despite having held no reference to it. | |
719 | * In which case we can trust the content of the page, and it | |
720 | * returns the gotten page; but if the page has now been zapped, | |
721 | * remove the stale node from the stable tree and return NULL. | |
c8d6553b | 722 | * But beware, the stable node's page might be being migrated. |
4035c07a HD |
723 | * |
724 | * You would expect the stable_node to hold a reference to the ksm page. | |
725 | * But if it increments the page's count, swapping out has to wait for | |
726 | * ksmd to come around again before it can free the page, which may take | |
727 | * seconds or even minutes: much too unresponsive. So instead we use a | |
728 | * "keyhole reference": access to the ksm page from the stable node peeps | |
729 | * out through its keyhole to see if that page still holds the right key, | |
730 | * pointing back to this stable node. This relies on freeing a PageAnon | |
731 | * page to reset its page->mapping to NULL, and relies on no other use of | |
732 | * a page to put something that might look like our key in page->mapping. | |
4035c07a HD |
733 | * is on its way to being freed; but it is an anomaly to bear in mind. |
734 | */ | |
21fbd591 | 735 | static struct page *get_ksm_page(struct ksm_stable_node *stable_node, |
2cee57d1 | 736 | enum get_ksm_page_flags flags) |
4035c07a HD |
737 | { |
738 | struct page *page; | |
739 | void *expected_mapping; | |
c8d6553b | 740 | unsigned long kpfn; |
4035c07a | 741 | |
bda807d4 MK |
742 | expected_mapping = (void *)((unsigned long)stable_node | |
743 | PAGE_MAPPING_KSM); | |
c8d6553b | 744 | again: |
08df4774 | 745 | kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */ |
c8d6553b | 746 | page = pfn_to_page(kpfn); |
4db0c3c2 | 747 | if (READ_ONCE(page->mapping) != expected_mapping) |
4035c07a | 748 | goto stale; |
c8d6553b HD |
749 | |
750 | /* | |
751 | * We cannot do anything with the page while its refcount is 0. | |
752 | * Usually 0 means free, or tail of a higher-order page: in which | |
753 | * case this node is no longer referenced, and should be freed; | |
1c4c3b99 | 754 | * however, it might mean that the page is under page_ref_freeze(). |
c8d6553b | 755 | * The __remove_mapping() case is easy, again the node is now stale; |
52d1e606 | 756 | * the same is in reuse_ksm_page() case; but if page is swapcache |
9800562f | 757 | * in folio_migrate_mapping(), it might still be our page, |
52d1e606 | 758 | * in which case it's essential to keep the node. |
c8d6553b HD |
759 | */ |
760 | while (!get_page_unless_zero(page)) { | |
761 | /* | |
762 | * Another check for page->mapping != expected_mapping would | |
763 | * work here too. We have chosen the !PageSwapCache test to | |
764 | * optimize the common case, when the page is or is about to | |
765 | * be freed: PageSwapCache is cleared (under spin_lock_irq) | |
1c4c3b99 | 766 | * in the ref_freeze section of __remove_mapping(); but Anon |
c8d6553b HD |
767 | * page->mapping reset to NULL later, in free_pages_prepare(). |
768 | */ | |
769 | if (!PageSwapCache(page)) | |
770 | goto stale; | |
771 | cpu_relax(); | |
772 | } | |
773 | ||
4db0c3c2 | 774 | if (READ_ONCE(page->mapping) != expected_mapping) { |
4035c07a HD |
775 | put_page(page); |
776 | goto stale; | |
777 | } | |
c8d6553b | 778 | |
2cee57d1 YS |
779 | if (flags == GET_KSM_PAGE_TRYLOCK) { |
780 | if (!trylock_page(page)) { | |
781 | put_page(page); | |
782 | return ERR_PTR(-EBUSY); | |
783 | } | |
784 | } else if (flags == GET_KSM_PAGE_LOCK) | |
8aafa6a4 | 785 | lock_page(page); |
2cee57d1 YS |
786 | |
787 | if (flags != GET_KSM_PAGE_NOLOCK) { | |
4db0c3c2 | 788 | if (READ_ONCE(page->mapping) != expected_mapping) { |
8aafa6a4 HD |
789 | unlock_page(page); |
790 | put_page(page); | |
791 | goto stale; | |
792 | } | |
793 | } | |
4035c07a | 794 | return page; |
c8d6553b | 795 | |
4035c07a | 796 | stale: |
c8d6553b HD |
797 | /* |
798 | * We come here from above when page->mapping or !PageSwapCache | |
799 | * suggests that the node is stale; but it might be under migration. | |
19138349 | 800 | * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(), |
c8d6553b HD |
801 | * before checking whether node->kpfn has been changed. |
802 | */ | |
803 | smp_rmb(); | |
4db0c3c2 | 804 | if (READ_ONCE(stable_node->kpfn) != kpfn) |
c8d6553b | 805 | goto again; |
4035c07a HD |
806 | remove_node_from_stable_tree(stable_node); |
807 | return NULL; | |
808 | } | |
809 | ||
31dbd01f IE |
810 | /* |
811 | * Removing rmap_item from stable or unstable tree. | |
812 | * This function will clean the information from the stable/unstable tree. | |
813 | */ | |
21fbd591 | 814 | static void remove_rmap_item_from_tree(struct ksm_rmap_item *rmap_item) |
31dbd01f | 815 | { |
7b6ba2c7 | 816 | if (rmap_item->address & STABLE_FLAG) { |
21fbd591 | 817 | struct ksm_stable_node *stable_node; |
5ad64688 | 818 | struct page *page; |
31dbd01f | 819 | |
7b6ba2c7 | 820 | stable_node = rmap_item->head; |
62862290 | 821 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
4035c07a HD |
822 | if (!page) |
823 | goto out; | |
5ad64688 | 824 | |
7b6ba2c7 | 825 | hlist_del(&rmap_item->hlist); |
62862290 | 826 | unlock_page(page); |
4035c07a | 827 | put_page(page); |
08beca44 | 828 | |
98666f8a | 829 | if (!hlist_empty(&stable_node->hlist)) |
4035c07a HD |
830 | ksm_pages_sharing--; |
831 | else | |
7b6ba2c7 | 832 | ksm_pages_shared--; |
76093853 | 833 | |
834 | rmap_item->mm->ksm_merging_pages--; | |
835 | ||
2c653d0e AA |
836 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
837 | stable_node->rmap_hlist_len--; | |
31dbd01f | 838 | |
9e60109f | 839 | put_anon_vma(rmap_item->anon_vma); |
c89a384e | 840 | rmap_item->head = NULL; |
93d17715 | 841 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 842 | |
7b6ba2c7 | 843 | } else if (rmap_item->address & UNSTABLE_FLAG) { |
31dbd01f IE |
844 | unsigned char age; |
845 | /* | |
9ba69294 | 846 | * Usually ksmd can and must skip the rb_erase, because |
31dbd01f | 847 | * root_unstable_tree was already reset to RB_ROOT. |
9ba69294 HD |
848 | * But be careful when an mm is exiting: do the rb_erase |
849 | * if this rmap_item was inserted by this scan, rather | |
850 | * than left over from before. | |
31dbd01f IE |
851 | */ |
852 | age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); | |
cd551f97 | 853 | BUG_ON(age > 1); |
31dbd01f | 854 | if (!age) |
90bd6fd3 | 855 | rb_erase(&rmap_item->node, |
ef53d16c | 856 | root_unstable_tree + NUMA(rmap_item->nid)); |
473b0ce4 | 857 | ksm_pages_unshared--; |
93d17715 | 858 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 859 | } |
4035c07a | 860 | out: |
31dbd01f IE |
861 | cond_resched(); /* we're called from many long loops */ |
862 | } | |
863 | ||
21fbd591 | 864 | static void remove_trailing_rmap_items(struct ksm_rmap_item **rmap_list) |
31dbd01f | 865 | { |
6514d511 | 866 | while (*rmap_list) { |
21fbd591 | 867 | struct ksm_rmap_item *rmap_item = *rmap_list; |
6514d511 | 868 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 869 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
870 | free_rmap_item(rmap_item); |
871 | } | |
872 | } | |
873 | ||
874 | /* | |
e850dcf5 | 875 | * Though it's very tempting to unmerge rmap_items from stable tree rather |
31dbd01f IE |
876 | * than check every pte of a given vma, the locking doesn't quite work for |
877 | * that - an rmap_item is assigned to the stable tree after inserting ksm | |
c1e8d7c6 | 878 | * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing |
31dbd01f IE |
879 | * rmap_items from parent to child at fork time (so as not to waste time |
880 | * if exit comes before the next scan reaches it). | |
81464e30 HD |
881 | * |
882 | * Similarly, although we'd like to remove rmap_items (so updating counts | |
883 | * and freeing memory) when unmerging an area, it's easier to leave that | |
884 | * to the next pass of ksmd - consider, for example, how ksmd might be | |
885 | * in cmp_and_merge_page on one of the rmap_items we would be removing. | |
31dbd01f | 886 | */ |
d952b791 | 887 | static int unmerge_ksm_pages(struct vm_area_struct *vma, |
49b06385 | 888 | unsigned long start, unsigned long end, bool lock_vma) |
31dbd01f IE |
889 | { |
890 | unsigned long addr; | |
d952b791 | 891 | int err = 0; |
31dbd01f | 892 | |
d952b791 | 893 | for (addr = start; addr < end && !err; addr += PAGE_SIZE) { |
9ba69294 HD |
894 | if (ksm_test_exit(vma->vm_mm)) |
895 | break; | |
d952b791 HD |
896 | if (signal_pending(current)) |
897 | err = -ERESTARTSYS; | |
898 | else | |
49b06385 | 899 | err = break_ksm(vma, addr, lock_vma); |
d952b791 HD |
900 | } |
901 | return err; | |
31dbd01f IE |
902 | } |
903 | ||
21fbd591 | 904 | static inline struct ksm_stable_node *folio_stable_node(struct folio *folio) |
19138349 MWO |
905 | { |
906 | return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL; | |
907 | } | |
908 | ||
21fbd591 | 909 | static inline struct ksm_stable_node *page_stable_node(struct page *page) |
88484826 | 910 | { |
19138349 | 911 | return folio_stable_node(page_folio(page)); |
88484826 MR |
912 | } |
913 | ||
914 | static inline void set_page_stable_node(struct page *page, | |
21fbd591 | 915 | struct ksm_stable_node *stable_node) |
88484826 | 916 | { |
6c287605 | 917 | VM_BUG_ON_PAGE(PageAnon(page) && PageAnonExclusive(page), page); |
88484826 MR |
918 | page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM); |
919 | } | |
920 | ||
2ffd8679 HD |
921 | #ifdef CONFIG_SYSFS |
922 | /* | |
923 | * Only called through the sysfs control interface: | |
924 | */ | |
21fbd591 | 925 | static int remove_stable_node(struct ksm_stable_node *stable_node) |
cbf86cfe HD |
926 | { |
927 | struct page *page; | |
928 | int err; | |
929 | ||
2cee57d1 | 930 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
cbf86cfe HD |
931 | if (!page) { |
932 | /* | |
933 | * get_ksm_page did remove_node_from_stable_tree itself. | |
934 | */ | |
935 | return 0; | |
936 | } | |
937 | ||
9a63236f AR |
938 | /* |
939 | * Page could be still mapped if this races with __mmput() running in | |
940 | * between ksm_exit() and exit_mmap(). Just refuse to let | |
941 | * merge_across_nodes/max_page_sharing be switched. | |
942 | */ | |
943 | err = -EBUSY; | |
944 | if (!page_mapped(page)) { | |
cbf86cfe | 945 | /* |
8fdb3dbf HD |
946 | * The stable node did not yet appear stale to get_ksm_page(), |
947 | * since that allows for an unmapped ksm page to be recognized | |
948 | * right up until it is freed; but the node is safe to remove. | |
1fec6890 | 949 | * This page might be in an LRU cache waiting to be freed, |
cbf86cfe HD |
950 | * or it might be PageSwapCache (perhaps under writeback), |
951 | * or it might have been removed from swapcache a moment ago. | |
952 | */ | |
953 | set_page_stable_node(page, NULL); | |
954 | remove_node_from_stable_tree(stable_node); | |
955 | err = 0; | |
956 | } | |
957 | ||
958 | unlock_page(page); | |
959 | put_page(page); | |
960 | return err; | |
961 | } | |
962 | ||
21fbd591 | 963 | static int remove_stable_node_chain(struct ksm_stable_node *stable_node, |
2c653d0e AA |
964 | struct rb_root *root) |
965 | { | |
21fbd591 | 966 | struct ksm_stable_node *dup; |
2c653d0e AA |
967 | struct hlist_node *hlist_safe; |
968 | ||
969 | if (!is_stable_node_chain(stable_node)) { | |
970 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
971 | if (remove_stable_node(stable_node)) | |
972 | return true; | |
973 | else | |
974 | return false; | |
975 | } | |
976 | ||
977 | hlist_for_each_entry_safe(dup, hlist_safe, | |
978 | &stable_node->hlist, hlist_dup) { | |
979 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
980 | if (remove_stable_node(dup)) | |
981 | return true; | |
982 | } | |
983 | BUG_ON(!hlist_empty(&stable_node->hlist)); | |
984 | free_stable_node_chain(stable_node, root); | |
985 | return false; | |
986 | } | |
987 | ||
cbf86cfe HD |
988 | static int remove_all_stable_nodes(void) |
989 | { | |
21fbd591 | 990 | struct ksm_stable_node *stable_node, *next; |
cbf86cfe HD |
991 | int nid; |
992 | int err = 0; | |
993 | ||
ef53d16c | 994 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
cbf86cfe HD |
995 | while (root_stable_tree[nid].rb_node) { |
996 | stable_node = rb_entry(root_stable_tree[nid].rb_node, | |
21fbd591 | 997 | struct ksm_stable_node, node); |
2c653d0e AA |
998 | if (remove_stable_node_chain(stable_node, |
999 | root_stable_tree + nid)) { | |
cbf86cfe HD |
1000 | err = -EBUSY; |
1001 | break; /* proceed to next nid */ | |
1002 | } | |
1003 | cond_resched(); | |
1004 | } | |
1005 | } | |
03640418 | 1006 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
1007 | if (remove_stable_node(stable_node)) |
1008 | err = -EBUSY; | |
1009 | cond_resched(); | |
1010 | } | |
cbf86cfe HD |
1011 | return err; |
1012 | } | |
1013 | ||
d952b791 | 1014 | static int unmerge_and_remove_all_rmap_items(void) |
31dbd01f | 1015 | { |
21fbd591 | 1016 | struct ksm_mm_slot *mm_slot; |
58730ab6 | 1017 | struct mm_slot *slot; |
31dbd01f IE |
1018 | struct mm_struct *mm; |
1019 | struct vm_area_struct *vma; | |
d952b791 HD |
1020 | int err = 0; |
1021 | ||
1022 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
1023 | slot = list_entry(ksm_mm_head.slot.mm_node.next, |
1024 | struct mm_slot, mm_node); | |
1025 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
d952b791 | 1026 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f | 1027 | |
a5f18ba0 MWO |
1028 | for (mm_slot = ksm_scan.mm_slot; mm_slot != &ksm_mm_head; |
1029 | mm_slot = ksm_scan.mm_slot) { | |
58730ab6 | 1030 | VMA_ITERATOR(vmi, mm_slot->slot.mm, 0); |
a5f18ba0 | 1031 | |
58730ab6 | 1032 | mm = mm_slot->slot.mm; |
d8ed45c5 | 1033 | mmap_read_lock(mm); |
6db504ce LH |
1034 | |
1035 | /* | |
1036 | * Exit right away if mm is exiting to avoid lockdep issue in | |
1037 | * the maple tree | |
1038 | */ | |
1039 | if (ksm_test_exit(mm)) | |
1040 | goto mm_exiting; | |
1041 | ||
a5f18ba0 | 1042 | for_each_vma(vmi, vma) { |
31dbd01f IE |
1043 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
1044 | continue; | |
d952b791 | 1045 | err = unmerge_ksm_pages(vma, |
49b06385 | 1046 | vma->vm_start, vma->vm_end, false); |
9ba69294 HD |
1047 | if (err) |
1048 | goto error; | |
31dbd01f | 1049 | } |
9ba69294 | 1050 | |
6db504ce | 1051 | mm_exiting: |
420be4ed | 1052 | remove_trailing_rmap_items(&mm_slot->rmap_list); |
d8ed45c5 | 1053 | mmap_read_unlock(mm); |
d952b791 HD |
1054 | |
1055 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
1056 | slot = list_entry(mm_slot->slot.mm_node.next, |
1057 | struct mm_slot, mm_node); | |
1058 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
9ba69294 | 1059 | if (ksm_test_exit(mm)) { |
58730ab6 QZ |
1060 | hash_del(&mm_slot->slot.hash); |
1061 | list_del(&mm_slot->slot.mm_node); | |
9ba69294 HD |
1062 | spin_unlock(&ksm_mmlist_lock); |
1063 | ||
58730ab6 | 1064 | mm_slot_free(mm_slot_cache, mm_slot); |
9ba69294 | 1065 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
d7597f59 | 1066 | clear_bit(MMF_VM_MERGE_ANY, &mm->flags); |
9ba69294 | 1067 | mmdrop(mm); |
7496fea9 | 1068 | } else |
9ba69294 | 1069 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f IE |
1070 | } |
1071 | ||
cbf86cfe HD |
1072 | /* Clean up stable nodes, but don't worry if some are still busy */ |
1073 | remove_all_stable_nodes(); | |
d952b791 | 1074 | ksm_scan.seqnr = 0; |
9ba69294 HD |
1075 | return 0; |
1076 | ||
1077 | error: | |
d8ed45c5 | 1078 | mmap_read_unlock(mm); |
31dbd01f | 1079 | spin_lock(&ksm_mmlist_lock); |
d952b791 | 1080 | ksm_scan.mm_slot = &ksm_mm_head; |
31dbd01f | 1081 | spin_unlock(&ksm_mmlist_lock); |
d952b791 | 1082 | return err; |
31dbd01f | 1083 | } |
2ffd8679 | 1084 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 1085 | |
31dbd01f IE |
1086 | static u32 calc_checksum(struct page *page) |
1087 | { | |
1088 | u32 checksum; | |
9b04c5fe | 1089 | void *addr = kmap_atomic(page); |
59e1a2f4 | 1090 | checksum = xxhash(addr, PAGE_SIZE, 0); |
9b04c5fe | 1091 | kunmap_atomic(addr); |
31dbd01f IE |
1092 | return checksum; |
1093 | } | |
1094 | ||
31dbd01f IE |
1095 | static int write_protect_page(struct vm_area_struct *vma, struct page *page, |
1096 | pte_t *orig_pte) | |
1097 | { | |
1098 | struct mm_struct *mm = vma->vm_mm; | |
eed05e54 | 1099 | DEFINE_PAGE_VMA_WALK(pvmw, page, vma, 0, 0); |
31dbd01f IE |
1100 | int swapped; |
1101 | int err = -EFAULT; | |
ac46d4f3 | 1102 | struct mmu_notifier_range range; |
6c287605 | 1103 | bool anon_exclusive; |
c33c7948 | 1104 | pte_t entry; |
31dbd01f | 1105 | |
36eaff33 KS |
1106 | pvmw.address = page_address_in_vma(page, vma); |
1107 | if (pvmw.address == -EFAULT) | |
31dbd01f IE |
1108 | goto out; |
1109 | ||
29ad768c | 1110 | BUG_ON(PageTransCompound(page)); |
6bdb913f | 1111 | |
7d4a8be0 | 1112 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, pvmw.address, |
ac46d4f3 JG |
1113 | pvmw.address + PAGE_SIZE); |
1114 | mmu_notifier_invalidate_range_start(&range); | |
6bdb913f | 1115 | |
36eaff33 | 1116 | if (!page_vma_mapped_walk(&pvmw)) |
6bdb913f | 1117 | goto out_mn; |
36eaff33 KS |
1118 | if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?")) |
1119 | goto out_unlock; | |
31dbd01f | 1120 | |
6c287605 | 1121 | anon_exclusive = PageAnonExclusive(page); |
c33c7948 RR |
1122 | entry = ptep_get(pvmw.pte); |
1123 | if (pte_write(entry) || pte_dirty(entry) || | |
6c287605 | 1124 | anon_exclusive || mm_tlb_flush_pending(mm)) { |
31dbd01f | 1125 | swapped = PageSwapCache(page); |
36eaff33 | 1126 | flush_cache_page(vma, pvmw.address, page_to_pfn(page)); |
31dbd01f | 1127 | /* |
25985edc | 1128 | * Ok this is tricky, when get_user_pages_fast() run it doesn't |
31dbd01f | 1129 | * take any lock, therefore the check that we are going to make |
f0953a1b | 1130 | * with the pagecount against the mapcount is racy and |
31dbd01f IE |
1131 | * O_DIRECT can happen right after the check. |
1132 | * So we clear the pte and flush the tlb before the check | |
1133 | * this assure us that no O_DIRECT can happen after the check | |
1134 | * or in the middle of the check. | |
0f10851e JG |
1135 | * |
1136 | * No need to notify as we are downgrading page table to read | |
1137 | * only not changing it to point to a new page. | |
1138 | * | |
ee65728e | 1139 | * See Documentation/mm/mmu_notifier.rst |
31dbd01f | 1140 | */ |
0f10851e | 1141 | entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte); |
31dbd01f IE |
1142 | /* |
1143 | * Check that no O_DIRECT or similar I/O is in progress on the | |
1144 | * page | |
1145 | */ | |
31e855ea | 1146 | if (page_mapcount(page) + 1 + swapped != page_count(page)) { |
36eaff33 | 1147 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f IE |
1148 | goto out_unlock; |
1149 | } | |
6c287605 | 1150 | |
088b8aa5 | 1151 | /* See page_try_share_anon_rmap(): clear PTE first. */ |
6c287605 DH |
1152 | if (anon_exclusive && page_try_share_anon_rmap(page)) { |
1153 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); | |
1154 | goto out_unlock; | |
1155 | } | |
1156 | ||
4e31635c HD |
1157 | if (pte_dirty(entry)) |
1158 | set_page_dirty(page); | |
6a56ccbc DH |
1159 | entry = pte_mkclean(entry); |
1160 | ||
1161 | if (pte_write(entry)) | |
1162 | entry = pte_wrprotect(entry); | |
595cd8f2 | 1163 | |
36eaff33 | 1164 | set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f | 1165 | } |
c33c7948 | 1166 | *orig_pte = entry; |
31dbd01f IE |
1167 | err = 0; |
1168 | ||
1169 | out_unlock: | |
36eaff33 | 1170 | page_vma_mapped_walk_done(&pvmw); |
6bdb913f | 1171 | out_mn: |
ac46d4f3 | 1172 | mmu_notifier_invalidate_range_end(&range); |
31dbd01f IE |
1173 | out: |
1174 | return err; | |
1175 | } | |
1176 | ||
1177 | /** | |
1178 | * replace_page - replace page in vma by new ksm page | |
8dd3557a HD |
1179 | * @vma: vma that holds the pte pointing to page |
1180 | * @page: the page we are replacing by kpage | |
1181 | * @kpage: the ksm page we replace page by | |
31dbd01f IE |
1182 | * @orig_pte: the original value of the pte |
1183 | * | |
1184 | * Returns 0 on success, -EFAULT on failure. | |
1185 | */ | |
8dd3557a HD |
1186 | static int replace_page(struct vm_area_struct *vma, struct page *page, |
1187 | struct page *kpage, pte_t orig_pte) | |
31dbd01f IE |
1188 | { |
1189 | struct mm_struct *mm = vma->vm_mm; | |
b4e6f66e | 1190 | struct folio *folio; |
31dbd01f | 1191 | pmd_t *pmd; |
50722804 | 1192 | pmd_t pmde; |
31dbd01f | 1193 | pte_t *ptep; |
e86c59b1 | 1194 | pte_t newpte; |
31dbd01f IE |
1195 | spinlock_t *ptl; |
1196 | unsigned long addr; | |
31dbd01f | 1197 | int err = -EFAULT; |
ac46d4f3 | 1198 | struct mmu_notifier_range range; |
31dbd01f | 1199 | |
8dd3557a | 1200 | addr = page_address_in_vma(page, vma); |
31dbd01f IE |
1201 | if (addr == -EFAULT) |
1202 | goto out; | |
1203 | ||
6219049a BL |
1204 | pmd = mm_find_pmd(mm, addr); |
1205 | if (!pmd) | |
31dbd01f | 1206 | goto out; |
50722804 ZK |
1207 | /* |
1208 | * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() | |
1209 | * without holding anon_vma lock for write. So when looking for a | |
1210 | * genuine pmde (in which to find pte), test present and !THP together. | |
1211 | */ | |
26e1a0c3 | 1212 | pmde = pmdp_get_lockless(pmd); |
50722804 ZK |
1213 | if (!pmd_present(pmde) || pmd_trans_huge(pmde)) |
1214 | goto out; | |
31dbd01f | 1215 | |
7d4a8be0 | 1216 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, addr, |
6f4f13e8 | 1217 | addr + PAGE_SIZE); |
ac46d4f3 | 1218 | mmu_notifier_invalidate_range_start(&range); |
6bdb913f | 1219 | |
31dbd01f | 1220 | ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); |
04dee9e8 HD |
1221 | if (!ptep) |
1222 | goto out_mn; | |
c33c7948 | 1223 | if (!pte_same(ptep_get(ptep), orig_pte)) { |
31dbd01f | 1224 | pte_unmap_unlock(ptep, ptl); |
6bdb913f | 1225 | goto out_mn; |
31dbd01f | 1226 | } |
6c287605 DH |
1227 | VM_BUG_ON_PAGE(PageAnonExclusive(page), page); |
1228 | VM_BUG_ON_PAGE(PageAnon(kpage) && PageAnonExclusive(kpage), kpage); | |
31dbd01f | 1229 | |
e86c59b1 CI |
1230 | /* |
1231 | * No need to check ksm_use_zero_pages here: we can only have a | |
457aef94 | 1232 | * zero_page here if ksm_use_zero_pages was enabled already. |
e86c59b1 CI |
1233 | */ |
1234 | if (!is_zero_pfn(page_to_pfn(kpage))) { | |
1235 | get_page(kpage); | |
f1e2db12 | 1236 | page_add_anon_rmap(kpage, vma, addr, RMAP_NONE); |
e86c59b1 CI |
1237 | newpte = mk_pte(kpage, vma->vm_page_prot); |
1238 | } else { | |
79271476 | 1239 | /* |
1240 | * Use pte_mkdirty to mark the zero page mapped by KSM, and then | |
1241 | * we can easily track all KSM-placed zero pages by checking if | |
1242 | * the dirty bit in zero page's PTE is set. | |
1243 | */ | |
1244 | newpte = pte_mkdirty(pte_mkspecial(pfn_pte(page_to_pfn(kpage), vma->vm_page_prot))); | |
e2942062 | 1245 | ksm_zero_pages++; |
6080d19f | 1246 | mm->ksm_zero_pages++; |
a38c015f CI |
1247 | /* |
1248 | * We're replacing an anonymous page with a zero page, which is | |
1249 | * not anonymous. We need to do proper accounting otherwise we | |
1250 | * will get wrong values in /proc, and a BUG message in dmesg | |
1251 | * when tearing down the mm. | |
1252 | */ | |
1253 | dec_mm_counter(mm, MM_ANONPAGES); | |
e86c59b1 | 1254 | } |
31dbd01f | 1255 | |
c33c7948 | 1256 | flush_cache_page(vma, addr, pte_pfn(ptep_get(ptep))); |
0f10851e JG |
1257 | /* |
1258 | * No need to notify as we are replacing a read only page with another | |
1259 | * read only page with the same content. | |
1260 | * | |
ee65728e | 1261 | * See Documentation/mm/mmu_notifier.rst |
0f10851e JG |
1262 | */ |
1263 | ptep_clear_flush(vma, addr, ptep); | |
e86c59b1 | 1264 | set_pte_at_notify(mm, addr, ptep, newpte); |
31dbd01f | 1265 | |
b4e6f66e | 1266 | folio = page_folio(page); |
cea86fe2 | 1267 | page_remove_rmap(page, vma, false); |
b4e6f66e MWO |
1268 | if (!folio_mapped(folio)) |
1269 | folio_free_swap(folio); | |
1270 | folio_put(folio); | |
31dbd01f IE |
1271 | |
1272 | pte_unmap_unlock(ptep, ptl); | |
1273 | err = 0; | |
6bdb913f | 1274 | out_mn: |
ac46d4f3 | 1275 | mmu_notifier_invalidate_range_end(&range); |
31dbd01f IE |
1276 | out: |
1277 | return err; | |
1278 | } | |
1279 | ||
1280 | /* | |
1281 | * try_to_merge_one_page - take two pages and merge them into one | |
8dd3557a HD |
1282 | * @vma: the vma that holds the pte pointing to page |
1283 | * @page: the PageAnon page that we want to replace with kpage | |
80e14822 HD |
1284 | * @kpage: the PageKsm page that we want to map instead of page, |
1285 | * or NULL the first time when we want to use page as kpage. | |
31dbd01f IE |
1286 | * |
1287 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
1288 | */ | |
1289 | static int try_to_merge_one_page(struct vm_area_struct *vma, | |
8dd3557a | 1290 | struct page *page, struct page *kpage) |
31dbd01f IE |
1291 | { |
1292 | pte_t orig_pte = __pte(0); | |
1293 | int err = -EFAULT; | |
1294 | ||
db114b83 HD |
1295 | if (page == kpage) /* ksm page forked */ |
1296 | return 0; | |
1297 | ||
8dd3557a | 1298 | if (!PageAnon(page)) |
31dbd01f IE |
1299 | goto out; |
1300 | ||
31dbd01f IE |
1301 | /* |
1302 | * We need the page lock to read a stable PageSwapCache in | |
1303 | * write_protect_page(). We use trylock_page() instead of | |
1304 | * lock_page() because we don't want to wait here - we | |
1305 | * prefer to continue scanning and merging different pages, | |
1306 | * then come back to this page when it is unlocked. | |
1307 | */ | |
8dd3557a | 1308 | if (!trylock_page(page)) |
31e855ea | 1309 | goto out; |
f765f540 KS |
1310 | |
1311 | if (PageTransCompound(page)) { | |
a7306c34 | 1312 | if (split_huge_page(page)) |
f765f540 KS |
1313 | goto out_unlock; |
1314 | } | |
1315 | ||
31dbd01f IE |
1316 | /* |
1317 | * If this anonymous page is mapped only here, its pte may need | |
1318 | * to be write-protected. If it's mapped elsewhere, all of its | |
1319 | * ptes are necessarily already write-protected. But in either | |
1320 | * case, we need to lock and check page_count is not raised. | |
1321 | */ | |
80e14822 HD |
1322 | if (write_protect_page(vma, page, &orig_pte) == 0) { |
1323 | if (!kpage) { | |
1324 | /* | |
1325 | * While we hold page lock, upgrade page from | |
1326 | * PageAnon+anon_vma to PageKsm+NULL stable_node: | |
1327 | * stable_tree_insert() will update stable_node. | |
1328 | */ | |
1329 | set_page_stable_node(page, NULL); | |
1330 | mark_page_accessed(page); | |
337ed7eb MK |
1331 | /* |
1332 | * Page reclaim just frees a clean page with no dirty | |
1333 | * ptes: make sure that the ksm page would be swapped. | |
1334 | */ | |
1335 | if (!PageDirty(page)) | |
1336 | SetPageDirty(page); | |
80e14822 HD |
1337 | err = 0; |
1338 | } else if (pages_identical(page, kpage)) | |
1339 | err = replace_page(vma, page, kpage, orig_pte); | |
1340 | } | |
31dbd01f | 1341 | |
f765f540 | 1342 | out_unlock: |
8dd3557a | 1343 | unlock_page(page); |
31dbd01f IE |
1344 | out: |
1345 | return err; | |
1346 | } | |
1347 | ||
81464e30 HD |
1348 | /* |
1349 | * try_to_merge_with_ksm_page - like try_to_merge_two_pages, | |
1350 | * but no new kernel page is allocated: kpage must already be a ksm page. | |
8dd3557a HD |
1351 | * |
1352 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
81464e30 | 1353 | */ |
21fbd591 | 1354 | static int try_to_merge_with_ksm_page(struct ksm_rmap_item *rmap_item, |
8dd3557a | 1355 | struct page *page, struct page *kpage) |
81464e30 | 1356 | { |
8dd3557a | 1357 | struct mm_struct *mm = rmap_item->mm; |
81464e30 HD |
1358 | struct vm_area_struct *vma; |
1359 | int err = -EFAULT; | |
1360 | ||
d8ed45c5 | 1361 | mmap_read_lock(mm); |
85c6e8dd AA |
1362 | vma = find_mergeable_vma(mm, rmap_item->address); |
1363 | if (!vma) | |
81464e30 HD |
1364 | goto out; |
1365 | ||
8dd3557a | 1366 | err = try_to_merge_one_page(vma, page, kpage); |
db114b83 HD |
1367 | if (err) |
1368 | goto out; | |
1369 | ||
bc56620b HD |
1370 | /* Unstable nid is in union with stable anon_vma: remove first */ |
1371 | remove_rmap_item_from_tree(rmap_item); | |
1372 | ||
c1e8d7c6 | 1373 | /* Must get reference to anon_vma while still holding mmap_lock */ |
9e60109f PZ |
1374 | rmap_item->anon_vma = vma->anon_vma; |
1375 | get_anon_vma(vma->anon_vma); | |
81464e30 | 1376 | out: |
d8ed45c5 | 1377 | mmap_read_unlock(mm); |
739100c8 SR |
1378 | trace_ksm_merge_with_ksm_page(kpage, page_to_pfn(kpage ? kpage : page), |
1379 | rmap_item, mm, err); | |
81464e30 HD |
1380 | return err; |
1381 | } | |
1382 | ||
31dbd01f IE |
1383 | /* |
1384 | * try_to_merge_two_pages - take two identical pages and prepare them | |
1385 | * to be merged into one page. | |
1386 | * | |
8dd3557a HD |
1387 | * This function returns the kpage if we successfully merged two identical |
1388 | * pages into one ksm page, NULL otherwise. | |
31dbd01f | 1389 | * |
80e14822 | 1390 | * Note that this function upgrades page to ksm page: if one of the pages |
31dbd01f IE |
1391 | * is already a ksm page, try_to_merge_with_ksm_page should be used. |
1392 | */ | |
21fbd591 | 1393 | static struct page *try_to_merge_two_pages(struct ksm_rmap_item *rmap_item, |
8dd3557a | 1394 | struct page *page, |
21fbd591 | 1395 | struct ksm_rmap_item *tree_rmap_item, |
8dd3557a | 1396 | struct page *tree_page) |
31dbd01f | 1397 | { |
80e14822 | 1398 | int err; |
31dbd01f | 1399 | |
80e14822 | 1400 | err = try_to_merge_with_ksm_page(rmap_item, page, NULL); |
31dbd01f | 1401 | if (!err) { |
8dd3557a | 1402 | err = try_to_merge_with_ksm_page(tree_rmap_item, |
80e14822 | 1403 | tree_page, page); |
31dbd01f | 1404 | /* |
81464e30 HD |
1405 | * If that fails, we have a ksm page with only one pte |
1406 | * pointing to it: so break it. | |
31dbd01f | 1407 | */ |
4035c07a | 1408 | if (err) |
8dd3557a | 1409 | break_cow(rmap_item); |
31dbd01f | 1410 | } |
80e14822 | 1411 | return err ? NULL : page; |
31dbd01f IE |
1412 | } |
1413 | ||
2c653d0e | 1414 | static __always_inline |
21fbd591 | 1415 | bool __is_page_sharing_candidate(struct ksm_stable_node *stable_node, int offset) |
2c653d0e AA |
1416 | { |
1417 | VM_BUG_ON(stable_node->rmap_hlist_len < 0); | |
1418 | /* | |
1419 | * Check that at least one mapping still exists, otherwise | |
1420 | * there's no much point to merge and share with this | |
1421 | * stable_node, as the underlying tree_page of the other | |
1422 | * sharer is going to be freed soon. | |
1423 | */ | |
1424 | return stable_node->rmap_hlist_len && | |
1425 | stable_node->rmap_hlist_len + offset < ksm_max_page_sharing; | |
1426 | } | |
1427 | ||
1428 | static __always_inline | |
21fbd591 | 1429 | bool is_page_sharing_candidate(struct ksm_stable_node *stable_node) |
2c653d0e AA |
1430 | { |
1431 | return __is_page_sharing_candidate(stable_node, 0); | |
1432 | } | |
1433 | ||
21fbd591 QZ |
1434 | static struct page *stable_node_dup(struct ksm_stable_node **_stable_node_dup, |
1435 | struct ksm_stable_node **_stable_node, | |
c01f0b54 CIK |
1436 | struct rb_root *root, |
1437 | bool prune_stale_stable_nodes) | |
2c653d0e | 1438 | { |
21fbd591 | 1439 | struct ksm_stable_node *dup, *found = NULL, *stable_node = *_stable_node; |
2c653d0e | 1440 | struct hlist_node *hlist_safe; |
8dc5ffcd | 1441 | struct page *_tree_page, *tree_page = NULL; |
2c653d0e AA |
1442 | int nr = 0; |
1443 | int found_rmap_hlist_len; | |
1444 | ||
1445 | if (!prune_stale_stable_nodes || | |
1446 | time_before(jiffies, stable_node->chain_prune_time + | |
1447 | msecs_to_jiffies( | |
1448 | ksm_stable_node_chains_prune_millisecs))) | |
1449 | prune_stale_stable_nodes = false; | |
1450 | else | |
1451 | stable_node->chain_prune_time = jiffies; | |
1452 | ||
1453 | hlist_for_each_entry_safe(dup, hlist_safe, | |
1454 | &stable_node->hlist, hlist_dup) { | |
1455 | cond_resched(); | |
1456 | /* | |
1457 | * We must walk all stable_node_dup to prune the stale | |
1458 | * stable nodes during lookup. | |
1459 | * | |
1460 | * get_ksm_page can drop the nodes from the | |
1461 | * stable_node->hlist if they point to freed pages | |
1462 | * (that's why we do a _safe walk). The "dup" | |
1463 | * stable_node parameter itself will be freed from | |
1464 | * under us if it returns NULL. | |
1465 | */ | |
2cee57d1 | 1466 | _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK); |
2c653d0e AA |
1467 | if (!_tree_page) |
1468 | continue; | |
1469 | nr += 1; | |
1470 | if (is_page_sharing_candidate(dup)) { | |
1471 | if (!found || | |
1472 | dup->rmap_hlist_len > found_rmap_hlist_len) { | |
1473 | if (found) | |
8dc5ffcd | 1474 | put_page(tree_page); |
2c653d0e AA |
1475 | found = dup; |
1476 | found_rmap_hlist_len = found->rmap_hlist_len; | |
8dc5ffcd | 1477 | tree_page = _tree_page; |
2c653d0e | 1478 | |
8dc5ffcd | 1479 | /* skip put_page for found dup */ |
2c653d0e AA |
1480 | if (!prune_stale_stable_nodes) |
1481 | break; | |
2c653d0e AA |
1482 | continue; |
1483 | } | |
1484 | } | |
1485 | put_page(_tree_page); | |
1486 | } | |
1487 | ||
80b18dfa AA |
1488 | if (found) { |
1489 | /* | |
1490 | * nr is counting all dups in the chain only if | |
1491 | * prune_stale_stable_nodes is true, otherwise we may | |
1492 | * break the loop at nr == 1 even if there are | |
1493 | * multiple entries. | |
1494 | */ | |
1495 | if (prune_stale_stable_nodes && nr == 1) { | |
2c653d0e AA |
1496 | /* |
1497 | * If there's not just one entry it would | |
1498 | * corrupt memory, better BUG_ON. In KSM | |
1499 | * context with no lock held it's not even | |
1500 | * fatal. | |
1501 | */ | |
1502 | BUG_ON(stable_node->hlist.first->next); | |
1503 | ||
1504 | /* | |
1505 | * There's just one entry and it is below the | |
1506 | * deduplication limit so drop the chain. | |
1507 | */ | |
1508 | rb_replace_node(&stable_node->node, &found->node, | |
1509 | root); | |
1510 | free_stable_node(stable_node); | |
1511 | ksm_stable_node_chains--; | |
1512 | ksm_stable_node_dups--; | |
b4fecc67 | 1513 | /* |
0ba1d0f7 AA |
1514 | * NOTE: the caller depends on the stable_node |
1515 | * to be equal to stable_node_dup if the chain | |
1516 | * was collapsed. | |
b4fecc67 | 1517 | */ |
0ba1d0f7 AA |
1518 | *_stable_node = found; |
1519 | /* | |
f0953a1b | 1520 | * Just for robustness, as stable_node is |
0ba1d0f7 AA |
1521 | * otherwise left as a stable pointer, the |
1522 | * compiler shall optimize it away at build | |
1523 | * time. | |
1524 | */ | |
1525 | stable_node = NULL; | |
80b18dfa AA |
1526 | } else if (stable_node->hlist.first != &found->hlist_dup && |
1527 | __is_page_sharing_candidate(found, 1)) { | |
2c653d0e | 1528 | /* |
80b18dfa AA |
1529 | * If the found stable_node dup can accept one |
1530 | * more future merge (in addition to the one | |
1531 | * that is underway) and is not at the head of | |
1532 | * the chain, put it there so next search will | |
1533 | * be quicker in the !prune_stale_stable_nodes | |
1534 | * case. | |
1535 | * | |
1536 | * NOTE: it would be inaccurate to use nr > 1 | |
1537 | * instead of checking the hlist.first pointer | |
1538 | * directly, because in the | |
1539 | * prune_stale_stable_nodes case "nr" isn't | |
1540 | * the position of the found dup in the chain, | |
1541 | * but the total number of dups in the chain. | |
2c653d0e AA |
1542 | */ |
1543 | hlist_del(&found->hlist_dup); | |
1544 | hlist_add_head(&found->hlist_dup, | |
1545 | &stable_node->hlist); | |
1546 | } | |
1547 | } | |
1548 | ||
8dc5ffcd AA |
1549 | *_stable_node_dup = found; |
1550 | return tree_page; | |
2c653d0e AA |
1551 | } |
1552 | ||
21fbd591 | 1553 | static struct ksm_stable_node *stable_node_dup_any(struct ksm_stable_node *stable_node, |
2c653d0e AA |
1554 | struct rb_root *root) |
1555 | { | |
1556 | if (!is_stable_node_chain(stable_node)) | |
1557 | return stable_node; | |
1558 | if (hlist_empty(&stable_node->hlist)) { | |
1559 | free_stable_node_chain(stable_node, root); | |
1560 | return NULL; | |
1561 | } | |
1562 | return hlist_entry(stable_node->hlist.first, | |
1563 | typeof(*stable_node), hlist_dup); | |
1564 | } | |
1565 | ||
8dc5ffcd AA |
1566 | /* |
1567 | * Like for get_ksm_page, this function can free the *_stable_node and | |
1568 | * *_stable_node_dup if the returned tree_page is NULL. | |
1569 | * | |
1570 | * It can also free and overwrite *_stable_node with the found | |
1571 | * stable_node_dup if the chain is collapsed (in which case | |
1572 | * *_stable_node will be equal to *_stable_node_dup like if the chain | |
1573 | * never existed). It's up to the caller to verify tree_page is not | |
1574 | * NULL before dereferencing *_stable_node or *_stable_node_dup. | |
1575 | * | |
1576 | * *_stable_node_dup is really a second output parameter of this | |
1577 | * function and will be overwritten in all cases, the caller doesn't | |
1578 | * need to initialize it. | |
1579 | */ | |
21fbd591 QZ |
1580 | static struct page *__stable_node_chain(struct ksm_stable_node **_stable_node_dup, |
1581 | struct ksm_stable_node **_stable_node, | |
8dc5ffcd AA |
1582 | struct rb_root *root, |
1583 | bool prune_stale_stable_nodes) | |
2c653d0e | 1584 | { |
21fbd591 | 1585 | struct ksm_stable_node *stable_node = *_stable_node; |
2c653d0e AA |
1586 | if (!is_stable_node_chain(stable_node)) { |
1587 | if (is_page_sharing_candidate(stable_node)) { | |
8dc5ffcd | 1588 | *_stable_node_dup = stable_node; |
2cee57d1 | 1589 | return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK); |
2c653d0e | 1590 | } |
8dc5ffcd AA |
1591 | /* |
1592 | * _stable_node_dup set to NULL means the stable_node | |
1593 | * reached the ksm_max_page_sharing limit. | |
1594 | */ | |
1595 | *_stable_node_dup = NULL; | |
2c653d0e AA |
1596 | return NULL; |
1597 | } | |
8dc5ffcd | 1598 | return stable_node_dup(_stable_node_dup, _stable_node, root, |
2c653d0e AA |
1599 | prune_stale_stable_nodes); |
1600 | } | |
1601 | ||
21fbd591 QZ |
1602 | static __always_inline struct page *chain_prune(struct ksm_stable_node **s_n_d, |
1603 | struct ksm_stable_node **s_n, | |
8dc5ffcd | 1604 | struct rb_root *root) |
2c653d0e | 1605 | { |
8dc5ffcd | 1606 | return __stable_node_chain(s_n_d, s_n, root, true); |
2c653d0e AA |
1607 | } |
1608 | ||
21fbd591 QZ |
1609 | static __always_inline struct page *chain(struct ksm_stable_node **s_n_d, |
1610 | struct ksm_stable_node *s_n, | |
8dc5ffcd | 1611 | struct rb_root *root) |
2c653d0e | 1612 | { |
21fbd591 | 1613 | struct ksm_stable_node *old_stable_node = s_n; |
8dc5ffcd AA |
1614 | struct page *tree_page; |
1615 | ||
1616 | tree_page = __stable_node_chain(s_n_d, &s_n, root, false); | |
1617 | /* not pruning dups so s_n cannot have changed */ | |
1618 | VM_BUG_ON(s_n != old_stable_node); | |
1619 | return tree_page; | |
2c653d0e AA |
1620 | } |
1621 | ||
31dbd01f | 1622 | /* |
8dd3557a | 1623 | * stable_tree_search - search for page inside the stable tree |
31dbd01f IE |
1624 | * |
1625 | * This function checks if there is a page inside the stable tree | |
1626 | * with identical content to the page that we are scanning right now. | |
1627 | * | |
7b6ba2c7 | 1628 | * This function returns the stable tree node of identical content if found, |
31dbd01f IE |
1629 | * NULL otherwise. |
1630 | */ | |
62b61f61 | 1631 | static struct page *stable_tree_search(struct page *page) |
31dbd01f | 1632 | { |
90bd6fd3 | 1633 | int nid; |
ef53d16c | 1634 | struct rb_root *root; |
4146d2d6 HD |
1635 | struct rb_node **new; |
1636 | struct rb_node *parent; | |
21fbd591 QZ |
1637 | struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; |
1638 | struct ksm_stable_node *page_node; | |
31dbd01f | 1639 | |
4146d2d6 HD |
1640 | page_node = page_stable_node(page); |
1641 | if (page_node && page_node->head != &migrate_nodes) { | |
1642 | /* ksm page forked */ | |
08beca44 | 1643 | get_page(page); |
62b61f61 | 1644 | return page; |
08beca44 HD |
1645 | } |
1646 | ||
90bd6fd3 | 1647 | nid = get_kpfn_nid(page_to_pfn(page)); |
ef53d16c | 1648 | root = root_stable_tree + nid; |
4146d2d6 | 1649 | again: |
ef53d16c | 1650 | new = &root->rb_node; |
4146d2d6 | 1651 | parent = NULL; |
90bd6fd3 | 1652 | |
4146d2d6 | 1653 | while (*new) { |
4035c07a | 1654 | struct page *tree_page; |
31dbd01f IE |
1655 | int ret; |
1656 | ||
08beca44 | 1657 | cond_resched(); |
21fbd591 | 1658 | stable_node = rb_entry(*new, struct ksm_stable_node, node); |
2c653d0e | 1659 | stable_node_any = NULL; |
8dc5ffcd | 1660 | tree_page = chain_prune(&stable_node_dup, &stable_node, root); |
b4fecc67 AA |
1661 | /* |
1662 | * NOTE: stable_node may have been freed by | |
1663 | * chain_prune() if the returned stable_node_dup is | |
1664 | * not NULL. stable_node_dup may have been inserted in | |
1665 | * the rbtree instead as a regular stable_node (in | |
1666 | * order to collapse the stable_node chain if a single | |
0ba1d0f7 | 1667 | * stable_node dup was found in it). In such case the |
3413b2c8 | 1668 | * stable_node is overwritten by the callee to point |
0ba1d0f7 AA |
1669 | * to the stable_node_dup that was collapsed in the |
1670 | * stable rbtree and stable_node will be equal to | |
1671 | * stable_node_dup like if the chain never existed. | |
b4fecc67 | 1672 | */ |
2c653d0e AA |
1673 | if (!stable_node_dup) { |
1674 | /* | |
1675 | * Either all stable_node dups were full in | |
1676 | * this stable_node chain, or this chain was | |
1677 | * empty and should be rb_erased. | |
1678 | */ | |
1679 | stable_node_any = stable_node_dup_any(stable_node, | |
1680 | root); | |
1681 | if (!stable_node_any) { | |
1682 | /* rb_erase just run */ | |
1683 | goto again; | |
1684 | } | |
1685 | /* | |
1686 | * Take any of the stable_node dups page of | |
1687 | * this stable_node chain to let the tree walk | |
1688 | * continue. All KSM pages belonging to the | |
1689 | * stable_node dups in a stable_node chain | |
1690 | * have the same content and they're | |
457aef94 | 1691 | * write protected at all times. Any will work |
2c653d0e AA |
1692 | * fine to continue the walk. |
1693 | */ | |
2cee57d1 YS |
1694 | tree_page = get_ksm_page(stable_node_any, |
1695 | GET_KSM_PAGE_NOLOCK); | |
2c653d0e AA |
1696 | } |
1697 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1698 | if (!tree_page) { |
1699 | /* | |
1700 | * If we walked over a stale stable_node, | |
1701 | * get_ksm_page() will call rb_erase() and it | |
1702 | * may rebalance the tree from under us. So | |
1703 | * restart the search from scratch. Returning | |
1704 | * NULL would be safe too, but we'd generate | |
1705 | * false negative insertions just because some | |
1706 | * stable_node was stale. | |
1707 | */ | |
1708 | goto again; | |
1709 | } | |
31dbd01f | 1710 | |
4035c07a | 1711 | ret = memcmp_pages(page, tree_page); |
c8d6553b | 1712 | put_page(tree_page); |
31dbd01f | 1713 | |
4146d2d6 | 1714 | parent = *new; |
c8d6553b | 1715 | if (ret < 0) |
4146d2d6 | 1716 | new = &parent->rb_left; |
c8d6553b | 1717 | else if (ret > 0) |
4146d2d6 | 1718 | new = &parent->rb_right; |
c8d6553b | 1719 | else { |
2c653d0e AA |
1720 | if (page_node) { |
1721 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1722 | /* | |
1723 | * Test if the migrated page should be merged | |
1724 | * into a stable node dup. If the mapcount is | |
1725 | * 1 we can migrate it with another KSM page | |
1726 | * without adding it to the chain. | |
1727 | */ | |
1728 | if (page_mapcount(page) > 1) | |
1729 | goto chain_append; | |
1730 | } | |
1731 | ||
1732 | if (!stable_node_dup) { | |
1733 | /* | |
1734 | * If the stable_node is a chain and | |
1735 | * we got a payload match in memcmp | |
1736 | * but we cannot merge the scanned | |
1737 | * page in any of the existing | |
1738 | * stable_node dups because they're | |
1739 | * all full, we need to wait the | |
1740 | * scanned page to find itself a match | |
1741 | * in the unstable tree to create a | |
1742 | * brand new KSM page to add later to | |
1743 | * the dups of this stable_node. | |
1744 | */ | |
1745 | return NULL; | |
1746 | } | |
1747 | ||
c8d6553b HD |
1748 | /* |
1749 | * Lock and unlock the stable_node's page (which | |
1750 | * might already have been migrated) so that page | |
1751 | * migration is sure to notice its raised count. | |
1752 | * It would be more elegant to return stable_node | |
1753 | * than kpage, but that involves more changes. | |
1754 | */ | |
2cee57d1 YS |
1755 | tree_page = get_ksm_page(stable_node_dup, |
1756 | GET_KSM_PAGE_TRYLOCK); | |
1757 | ||
1758 | if (PTR_ERR(tree_page) == -EBUSY) | |
1759 | return ERR_PTR(-EBUSY); | |
1760 | ||
2c653d0e AA |
1761 | if (unlikely(!tree_page)) |
1762 | /* | |
1763 | * The tree may have been rebalanced, | |
1764 | * so re-evaluate parent and new. | |
1765 | */ | |
4146d2d6 | 1766 | goto again; |
2c653d0e AA |
1767 | unlock_page(tree_page); |
1768 | ||
1769 | if (get_kpfn_nid(stable_node_dup->kpfn) != | |
1770 | NUMA(stable_node_dup->nid)) { | |
1771 | put_page(tree_page); | |
1772 | goto replace; | |
1773 | } | |
1774 | return tree_page; | |
c8d6553b | 1775 | } |
31dbd01f IE |
1776 | } |
1777 | ||
4146d2d6 HD |
1778 | if (!page_node) |
1779 | return NULL; | |
1780 | ||
1781 | list_del(&page_node->list); | |
1782 | DO_NUMA(page_node->nid = nid); | |
1783 | rb_link_node(&page_node->node, parent, new); | |
ef53d16c | 1784 | rb_insert_color(&page_node->node, root); |
2c653d0e AA |
1785 | out: |
1786 | if (is_page_sharing_candidate(page_node)) { | |
1787 | get_page(page); | |
1788 | return page; | |
1789 | } else | |
1790 | return NULL; | |
4146d2d6 HD |
1791 | |
1792 | replace: | |
b4fecc67 AA |
1793 | /* |
1794 | * If stable_node was a chain and chain_prune collapsed it, | |
0ba1d0f7 AA |
1795 | * stable_node has been updated to be the new regular |
1796 | * stable_node. A collapse of the chain is indistinguishable | |
1797 | * from the case there was no chain in the stable | |
1798 | * rbtree. Otherwise stable_node is the chain and | |
1799 | * stable_node_dup is the dup to replace. | |
b4fecc67 | 1800 | */ |
0ba1d0f7 | 1801 | if (stable_node_dup == stable_node) { |
b4fecc67 AA |
1802 | VM_BUG_ON(is_stable_node_chain(stable_node_dup)); |
1803 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); | |
2c653d0e AA |
1804 | /* there is no chain */ |
1805 | if (page_node) { | |
1806 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1807 | list_del(&page_node->list); | |
1808 | DO_NUMA(page_node->nid = nid); | |
b4fecc67 AA |
1809 | rb_replace_node(&stable_node_dup->node, |
1810 | &page_node->node, | |
2c653d0e AA |
1811 | root); |
1812 | if (is_page_sharing_candidate(page_node)) | |
1813 | get_page(page); | |
1814 | else | |
1815 | page = NULL; | |
1816 | } else { | |
b4fecc67 | 1817 | rb_erase(&stable_node_dup->node, root); |
2c653d0e AA |
1818 | page = NULL; |
1819 | } | |
4146d2d6 | 1820 | } else { |
2c653d0e AA |
1821 | VM_BUG_ON(!is_stable_node_chain(stable_node)); |
1822 | __stable_node_dup_del(stable_node_dup); | |
1823 | if (page_node) { | |
1824 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1825 | list_del(&page_node->list); | |
1826 | DO_NUMA(page_node->nid = nid); | |
1827 | stable_node_chain_add_dup(page_node, stable_node); | |
1828 | if (is_page_sharing_candidate(page_node)) | |
1829 | get_page(page); | |
1830 | else | |
1831 | page = NULL; | |
1832 | } else { | |
1833 | page = NULL; | |
1834 | } | |
4146d2d6 | 1835 | } |
2c653d0e AA |
1836 | stable_node_dup->head = &migrate_nodes; |
1837 | list_add(&stable_node_dup->list, stable_node_dup->head); | |
4146d2d6 | 1838 | return page; |
2c653d0e AA |
1839 | |
1840 | chain_append: | |
1841 | /* stable_node_dup could be null if it reached the limit */ | |
1842 | if (!stable_node_dup) | |
1843 | stable_node_dup = stable_node_any; | |
b4fecc67 AA |
1844 | /* |
1845 | * If stable_node was a chain and chain_prune collapsed it, | |
0ba1d0f7 AA |
1846 | * stable_node has been updated to be the new regular |
1847 | * stable_node. A collapse of the chain is indistinguishable | |
1848 | * from the case there was no chain in the stable | |
1849 | * rbtree. Otherwise stable_node is the chain and | |
1850 | * stable_node_dup is the dup to replace. | |
b4fecc67 | 1851 | */ |
0ba1d0f7 | 1852 | if (stable_node_dup == stable_node) { |
b4fecc67 | 1853 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); |
2c653d0e AA |
1854 | /* chain is missing so create it */ |
1855 | stable_node = alloc_stable_node_chain(stable_node_dup, | |
1856 | root); | |
1857 | if (!stable_node) | |
1858 | return NULL; | |
1859 | } | |
1860 | /* | |
1861 | * Add this stable_node dup that was | |
1862 | * migrated to the stable_node chain | |
1863 | * of the current nid for this page | |
1864 | * content. | |
1865 | */ | |
b4fecc67 | 1866 | VM_BUG_ON(!is_stable_node_dup(stable_node_dup)); |
2c653d0e AA |
1867 | VM_BUG_ON(page_node->head != &migrate_nodes); |
1868 | list_del(&page_node->list); | |
1869 | DO_NUMA(page_node->nid = nid); | |
1870 | stable_node_chain_add_dup(page_node, stable_node); | |
1871 | goto out; | |
31dbd01f IE |
1872 | } |
1873 | ||
1874 | /* | |
e850dcf5 | 1875 | * stable_tree_insert - insert stable tree node pointing to new ksm page |
31dbd01f IE |
1876 | * into the stable tree. |
1877 | * | |
7b6ba2c7 HD |
1878 | * This function returns the stable tree node just allocated on success, |
1879 | * NULL otherwise. | |
31dbd01f | 1880 | */ |
21fbd591 | 1881 | static struct ksm_stable_node *stable_tree_insert(struct page *kpage) |
31dbd01f | 1882 | { |
90bd6fd3 PH |
1883 | int nid; |
1884 | unsigned long kpfn; | |
ef53d16c | 1885 | struct rb_root *root; |
90bd6fd3 | 1886 | struct rb_node **new; |
f2e5ff85 | 1887 | struct rb_node *parent; |
21fbd591 | 1888 | struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; |
2c653d0e | 1889 | bool need_chain = false; |
31dbd01f | 1890 | |
90bd6fd3 PH |
1891 | kpfn = page_to_pfn(kpage); |
1892 | nid = get_kpfn_nid(kpfn); | |
ef53d16c | 1893 | root = root_stable_tree + nid; |
f2e5ff85 AA |
1894 | again: |
1895 | parent = NULL; | |
ef53d16c | 1896 | new = &root->rb_node; |
90bd6fd3 | 1897 | |
31dbd01f | 1898 | while (*new) { |
4035c07a | 1899 | struct page *tree_page; |
31dbd01f IE |
1900 | int ret; |
1901 | ||
08beca44 | 1902 | cond_resched(); |
21fbd591 | 1903 | stable_node = rb_entry(*new, struct ksm_stable_node, node); |
2c653d0e | 1904 | stable_node_any = NULL; |
8dc5ffcd | 1905 | tree_page = chain(&stable_node_dup, stable_node, root); |
2c653d0e AA |
1906 | if (!stable_node_dup) { |
1907 | /* | |
1908 | * Either all stable_node dups were full in | |
1909 | * this stable_node chain, or this chain was | |
1910 | * empty and should be rb_erased. | |
1911 | */ | |
1912 | stable_node_any = stable_node_dup_any(stable_node, | |
1913 | root); | |
1914 | if (!stable_node_any) { | |
1915 | /* rb_erase just run */ | |
1916 | goto again; | |
1917 | } | |
1918 | /* | |
1919 | * Take any of the stable_node dups page of | |
1920 | * this stable_node chain to let the tree walk | |
1921 | * continue. All KSM pages belonging to the | |
1922 | * stable_node dups in a stable_node chain | |
1923 | * have the same content and they're | |
457aef94 | 1924 | * write protected at all times. Any will work |
2c653d0e AA |
1925 | * fine to continue the walk. |
1926 | */ | |
2cee57d1 YS |
1927 | tree_page = get_ksm_page(stable_node_any, |
1928 | GET_KSM_PAGE_NOLOCK); | |
2c653d0e AA |
1929 | } |
1930 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1931 | if (!tree_page) { |
1932 | /* | |
1933 | * If we walked over a stale stable_node, | |
1934 | * get_ksm_page() will call rb_erase() and it | |
1935 | * may rebalance the tree from under us. So | |
1936 | * restart the search from scratch. Returning | |
1937 | * NULL would be safe too, but we'd generate | |
1938 | * false negative insertions just because some | |
1939 | * stable_node was stale. | |
1940 | */ | |
1941 | goto again; | |
1942 | } | |
31dbd01f | 1943 | |
4035c07a HD |
1944 | ret = memcmp_pages(kpage, tree_page); |
1945 | put_page(tree_page); | |
31dbd01f IE |
1946 | |
1947 | parent = *new; | |
1948 | if (ret < 0) | |
1949 | new = &parent->rb_left; | |
1950 | else if (ret > 0) | |
1951 | new = &parent->rb_right; | |
1952 | else { | |
2c653d0e AA |
1953 | need_chain = true; |
1954 | break; | |
31dbd01f IE |
1955 | } |
1956 | } | |
1957 | ||
2c653d0e AA |
1958 | stable_node_dup = alloc_stable_node(); |
1959 | if (!stable_node_dup) | |
7b6ba2c7 | 1960 | return NULL; |
31dbd01f | 1961 | |
2c653d0e AA |
1962 | INIT_HLIST_HEAD(&stable_node_dup->hlist); |
1963 | stable_node_dup->kpfn = kpfn; | |
1964 | set_page_stable_node(kpage, stable_node_dup); | |
1965 | stable_node_dup->rmap_hlist_len = 0; | |
1966 | DO_NUMA(stable_node_dup->nid = nid); | |
1967 | if (!need_chain) { | |
1968 | rb_link_node(&stable_node_dup->node, parent, new); | |
1969 | rb_insert_color(&stable_node_dup->node, root); | |
1970 | } else { | |
1971 | if (!is_stable_node_chain(stable_node)) { | |
21fbd591 | 1972 | struct ksm_stable_node *orig = stable_node; |
2c653d0e AA |
1973 | /* chain is missing so create it */ |
1974 | stable_node = alloc_stable_node_chain(orig, root); | |
1975 | if (!stable_node) { | |
1976 | free_stable_node(stable_node_dup); | |
1977 | return NULL; | |
1978 | } | |
1979 | } | |
1980 | stable_node_chain_add_dup(stable_node_dup, stable_node); | |
1981 | } | |
08beca44 | 1982 | |
2c653d0e | 1983 | return stable_node_dup; |
31dbd01f IE |
1984 | } |
1985 | ||
1986 | /* | |
8dd3557a HD |
1987 | * unstable_tree_search_insert - search for identical page, |
1988 | * else insert rmap_item into the unstable tree. | |
31dbd01f IE |
1989 | * |
1990 | * This function searches for a page in the unstable tree identical to the | |
1991 | * page currently being scanned; and if no identical page is found in the | |
1992 | * tree, we insert rmap_item as a new object into the unstable tree. | |
1993 | * | |
1994 | * This function returns pointer to rmap_item found to be identical | |
1995 | * to the currently scanned page, NULL otherwise. | |
1996 | * | |
1997 | * This function does both searching and inserting, because they share | |
1998 | * the same walking algorithm in an rbtree. | |
1999 | */ | |
8dd3557a | 2000 | static |
21fbd591 | 2001 | struct ksm_rmap_item *unstable_tree_search_insert(struct ksm_rmap_item *rmap_item, |
8dd3557a HD |
2002 | struct page *page, |
2003 | struct page **tree_pagep) | |
31dbd01f | 2004 | { |
90bd6fd3 PH |
2005 | struct rb_node **new; |
2006 | struct rb_root *root; | |
31dbd01f | 2007 | struct rb_node *parent = NULL; |
90bd6fd3 PH |
2008 | int nid; |
2009 | ||
2010 | nid = get_kpfn_nid(page_to_pfn(page)); | |
ef53d16c | 2011 | root = root_unstable_tree + nid; |
90bd6fd3 | 2012 | new = &root->rb_node; |
31dbd01f IE |
2013 | |
2014 | while (*new) { | |
21fbd591 | 2015 | struct ksm_rmap_item *tree_rmap_item; |
8dd3557a | 2016 | struct page *tree_page; |
31dbd01f IE |
2017 | int ret; |
2018 | ||
d178f27f | 2019 | cond_resched(); |
21fbd591 | 2020 | tree_rmap_item = rb_entry(*new, struct ksm_rmap_item, node); |
8dd3557a | 2021 | tree_page = get_mergeable_page(tree_rmap_item); |
c8f95ed1 | 2022 | if (!tree_page) |
31dbd01f IE |
2023 | return NULL; |
2024 | ||
2025 | /* | |
8dd3557a | 2026 | * Don't substitute a ksm page for a forked page. |
31dbd01f | 2027 | */ |
8dd3557a HD |
2028 | if (page == tree_page) { |
2029 | put_page(tree_page); | |
31dbd01f IE |
2030 | return NULL; |
2031 | } | |
2032 | ||
8dd3557a | 2033 | ret = memcmp_pages(page, tree_page); |
31dbd01f IE |
2034 | |
2035 | parent = *new; | |
2036 | if (ret < 0) { | |
8dd3557a | 2037 | put_page(tree_page); |
31dbd01f IE |
2038 | new = &parent->rb_left; |
2039 | } else if (ret > 0) { | |
8dd3557a | 2040 | put_page(tree_page); |
31dbd01f | 2041 | new = &parent->rb_right; |
b599cbdf HD |
2042 | } else if (!ksm_merge_across_nodes && |
2043 | page_to_nid(tree_page) != nid) { | |
2044 | /* | |
2045 | * If tree_page has been migrated to another NUMA node, | |
2046 | * it will be flushed out and put in the right unstable | |
2047 | * tree next time: only merge with it when across_nodes. | |
2048 | */ | |
2049 | put_page(tree_page); | |
2050 | return NULL; | |
31dbd01f | 2051 | } else { |
8dd3557a | 2052 | *tree_pagep = tree_page; |
31dbd01f IE |
2053 | return tree_rmap_item; |
2054 | } | |
2055 | } | |
2056 | ||
7b6ba2c7 | 2057 | rmap_item->address |= UNSTABLE_FLAG; |
31dbd01f | 2058 | rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); |
e850dcf5 | 2059 | DO_NUMA(rmap_item->nid = nid); |
31dbd01f | 2060 | rb_link_node(&rmap_item->node, parent, new); |
90bd6fd3 | 2061 | rb_insert_color(&rmap_item->node, root); |
31dbd01f | 2062 | |
473b0ce4 | 2063 | ksm_pages_unshared++; |
31dbd01f IE |
2064 | return NULL; |
2065 | } | |
2066 | ||
2067 | /* | |
2068 | * stable_tree_append - add another rmap_item to the linked list of | |
2069 | * rmap_items hanging off a given node of the stable tree, all sharing | |
2070 | * the same ksm page. | |
2071 | */ | |
21fbd591 QZ |
2072 | static void stable_tree_append(struct ksm_rmap_item *rmap_item, |
2073 | struct ksm_stable_node *stable_node, | |
2c653d0e | 2074 | bool max_page_sharing_bypass) |
31dbd01f | 2075 | { |
2c653d0e AA |
2076 | /* |
2077 | * rmap won't find this mapping if we don't insert the | |
2078 | * rmap_item in the right stable_node | |
2079 | * duplicate. page_migration could break later if rmap breaks, | |
2080 | * so we can as well crash here. We really need to check for | |
2081 | * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check | |
457aef94 | 2082 | * for other negative values as an underflow if detected here |
2c653d0e AA |
2083 | * for the first time (and not when decreasing rmap_hlist_len) |
2084 | * would be sign of memory corruption in the stable_node. | |
2085 | */ | |
2086 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
2087 | ||
2088 | stable_node->rmap_hlist_len++; | |
2089 | if (!max_page_sharing_bypass) | |
2090 | /* possibly non fatal but unexpected overflow, only warn */ | |
2091 | WARN_ON_ONCE(stable_node->rmap_hlist_len > | |
2092 | ksm_max_page_sharing); | |
2093 | ||
7b6ba2c7 | 2094 | rmap_item->head = stable_node; |
31dbd01f | 2095 | rmap_item->address |= STABLE_FLAG; |
7b6ba2c7 | 2096 | hlist_add_head(&rmap_item->hlist, &stable_node->hlist); |
e178dfde | 2097 | |
7b6ba2c7 HD |
2098 | if (rmap_item->hlist.next) |
2099 | ksm_pages_sharing++; | |
2100 | else | |
2101 | ksm_pages_shared++; | |
76093853 | 2102 | |
2103 | rmap_item->mm->ksm_merging_pages++; | |
31dbd01f IE |
2104 | } |
2105 | ||
2106 | /* | |
81464e30 HD |
2107 | * cmp_and_merge_page - first see if page can be merged into the stable tree; |
2108 | * if not, compare checksum to previous and if it's the same, see if page can | |
2109 | * be inserted into the unstable tree, or merged with a page already there and | |
2110 | * both transferred to the stable tree. | |
31dbd01f IE |
2111 | * |
2112 | * @page: the page that we are searching identical page to. | |
2113 | * @rmap_item: the reverse mapping into the virtual address of this page | |
2114 | */ | |
21fbd591 | 2115 | static void cmp_and_merge_page(struct page *page, struct ksm_rmap_item *rmap_item) |
31dbd01f | 2116 | { |
4b22927f | 2117 | struct mm_struct *mm = rmap_item->mm; |
21fbd591 | 2118 | struct ksm_rmap_item *tree_rmap_item; |
8dd3557a | 2119 | struct page *tree_page = NULL; |
21fbd591 | 2120 | struct ksm_stable_node *stable_node; |
8dd3557a | 2121 | struct page *kpage; |
31dbd01f IE |
2122 | unsigned int checksum; |
2123 | int err; | |
2c653d0e | 2124 | bool max_page_sharing_bypass = false; |
31dbd01f | 2125 | |
4146d2d6 HD |
2126 | stable_node = page_stable_node(page); |
2127 | if (stable_node) { | |
2128 | if (stable_node->head != &migrate_nodes && | |
2c653d0e AA |
2129 | get_kpfn_nid(READ_ONCE(stable_node->kpfn)) != |
2130 | NUMA(stable_node->nid)) { | |
2131 | stable_node_dup_del(stable_node); | |
4146d2d6 HD |
2132 | stable_node->head = &migrate_nodes; |
2133 | list_add(&stable_node->list, stable_node->head); | |
2134 | } | |
2135 | if (stable_node->head != &migrate_nodes && | |
2136 | rmap_item->head == stable_node) | |
2137 | return; | |
2c653d0e AA |
2138 | /* |
2139 | * If it's a KSM fork, allow it to go over the sharing limit | |
2140 | * without warnings. | |
2141 | */ | |
2142 | if (!is_page_sharing_candidate(stable_node)) | |
2143 | max_page_sharing_bypass = true; | |
4146d2d6 | 2144 | } |
31dbd01f IE |
2145 | |
2146 | /* We first start with searching the page inside the stable tree */ | |
62b61f61 | 2147 | kpage = stable_tree_search(page); |
4146d2d6 HD |
2148 | if (kpage == page && rmap_item->head == stable_node) { |
2149 | put_page(kpage); | |
2150 | return; | |
2151 | } | |
2152 | ||
2153 | remove_rmap_item_from_tree(rmap_item); | |
2154 | ||
62b61f61 | 2155 | if (kpage) { |
2cee57d1 YS |
2156 | if (PTR_ERR(kpage) == -EBUSY) |
2157 | return; | |
2158 | ||
08beca44 | 2159 | err = try_to_merge_with_ksm_page(rmap_item, page, kpage); |
31dbd01f IE |
2160 | if (!err) { |
2161 | /* | |
2162 | * The page was successfully merged: | |
2163 | * add its rmap_item to the stable tree. | |
2164 | */ | |
5ad64688 | 2165 | lock_page(kpage); |
2c653d0e AA |
2166 | stable_tree_append(rmap_item, page_stable_node(kpage), |
2167 | max_page_sharing_bypass); | |
5ad64688 | 2168 | unlock_page(kpage); |
31dbd01f | 2169 | } |
8dd3557a | 2170 | put_page(kpage); |
31dbd01f IE |
2171 | return; |
2172 | } | |
2173 | ||
2174 | /* | |
4035c07a HD |
2175 | * If the hash value of the page has changed from the last time |
2176 | * we calculated it, this page is changing frequently: therefore we | |
2177 | * don't want to insert it in the unstable tree, and we don't want | |
2178 | * to waste our time searching for something identical to it there. | |
31dbd01f IE |
2179 | */ |
2180 | checksum = calc_checksum(page); | |
2181 | if (rmap_item->oldchecksum != checksum) { | |
2182 | rmap_item->oldchecksum = checksum; | |
2183 | return; | |
2184 | } | |
2185 | ||
e86c59b1 CI |
2186 | /* |
2187 | * Same checksum as an empty page. We attempt to merge it with the | |
2188 | * appropriate zero page if the user enabled this via sysfs. | |
2189 | */ | |
2190 | if (ksm_use_zero_pages && (checksum == zero_checksum)) { | |
2191 | struct vm_area_struct *vma; | |
2192 | ||
d8ed45c5 | 2193 | mmap_read_lock(mm); |
4b22927f | 2194 | vma = find_mergeable_vma(mm, rmap_item->address); |
56df70a6 MS |
2195 | if (vma) { |
2196 | err = try_to_merge_one_page(vma, page, | |
2197 | ZERO_PAGE(rmap_item->address)); | |
739100c8 SR |
2198 | trace_ksm_merge_one_page( |
2199 | page_to_pfn(ZERO_PAGE(rmap_item->address)), | |
2200 | rmap_item, mm, err); | |
56df70a6 MS |
2201 | } else { |
2202 | /* | |
2203 | * If the vma is out of date, we do not need to | |
2204 | * continue. | |
2205 | */ | |
2206 | err = 0; | |
2207 | } | |
d8ed45c5 | 2208 | mmap_read_unlock(mm); |
e86c59b1 CI |
2209 | /* |
2210 | * In case of failure, the page was not really empty, so we | |
2211 | * need to continue. Otherwise we're done. | |
2212 | */ | |
2213 | if (!err) | |
2214 | return; | |
2215 | } | |
8dd3557a HD |
2216 | tree_rmap_item = |
2217 | unstable_tree_search_insert(rmap_item, page, &tree_page); | |
31dbd01f | 2218 | if (tree_rmap_item) { |
77da2ba0 CI |
2219 | bool split; |
2220 | ||
8dd3557a HD |
2221 | kpage = try_to_merge_two_pages(rmap_item, page, |
2222 | tree_rmap_item, tree_page); | |
77da2ba0 CI |
2223 | /* |
2224 | * If both pages we tried to merge belong to the same compound | |
2225 | * page, then we actually ended up increasing the reference | |
2226 | * count of the same compound page twice, and split_huge_page | |
2227 | * failed. | |
2228 | * Here we set a flag if that happened, and we use it later to | |
2229 | * try split_huge_page again. Since we call put_page right | |
2230 | * afterwards, the reference count will be correct and | |
2231 | * split_huge_page should succeed. | |
2232 | */ | |
2233 | split = PageTransCompound(page) | |
2234 | && compound_head(page) == compound_head(tree_page); | |
8dd3557a | 2235 | put_page(tree_page); |
8dd3557a | 2236 | if (kpage) { |
bc56620b HD |
2237 | /* |
2238 | * The pages were successfully merged: insert new | |
2239 | * node in the stable tree and add both rmap_items. | |
2240 | */ | |
5ad64688 | 2241 | lock_page(kpage); |
7b6ba2c7 HD |
2242 | stable_node = stable_tree_insert(kpage); |
2243 | if (stable_node) { | |
2c653d0e AA |
2244 | stable_tree_append(tree_rmap_item, stable_node, |
2245 | false); | |
2246 | stable_tree_append(rmap_item, stable_node, | |
2247 | false); | |
7b6ba2c7 | 2248 | } |
5ad64688 | 2249 | unlock_page(kpage); |
7b6ba2c7 | 2250 | |
31dbd01f IE |
2251 | /* |
2252 | * If we fail to insert the page into the stable tree, | |
2253 | * we will have 2 virtual addresses that are pointing | |
2254 | * to a ksm page left outside the stable tree, | |
2255 | * in which case we need to break_cow on both. | |
2256 | */ | |
7b6ba2c7 | 2257 | if (!stable_node) { |
8dd3557a HD |
2258 | break_cow(tree_rmap_item); |
2259 | break_cow(rmap_item); | |
31dbd01f | 2260 | } |
77da2ba0 CI |
2261 | } else if (split) { |
2262 | /* | |
2263 | * We are here if we tried to merge two pages and | |
2264 | * failed because they both belonged to the same | |
2265 | * compound page. We will split the page now, but no | |
2266 | * merging will take place. | |
2267 | * We do not want to add the cost of a full lock; if | |
2268 | * the page is locked, it is better to skip it and | |
2269 | * perhaps try again later. | |
2270 | */ | |
2271 | if (!trylock_page(page)) | |
2272 | return; | |
2273 | split_huge_page(page); | |
2274 | unlock_page(page); | |
31dbd01f | 2275 | } |
31dbd01f IE |
2276 | } |
2277 | } | |
2278 | ||
21fbd591 QZ |
2279 | static struct ksm_rmap_item *get_next_rmap_item(struct ksm_mm_slot *mm_slot, |
2280 | struct ksm_rmap_item **rmap_list, | |
31dbd01f IE |
2281 | unsigned long addr) |
2282 | { | |
21fbd591 | 2283 | struct ksm_rmap_item *rmap_item; |
31dbd01f | 2284 | |
6514d511 HD |
2285 | while (*rmap_list) { |
2286 | rmap_item = *rmap_list; | |
93d17715 | 2287 | if ((rmap_item->address & PAGE_MASK) == addr) |
31dbd01f | 2288 | return rmap_item; |
31dbd01f IE |
2289 | if (rmap_item->address > addr) |
2290 | break; | |
6514d511 | 2291 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 2292 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
2293 | free_rmap_item(rmap_item); |
2294 | } | |
2295 | ||
2296 | rmap_item = alloc_rmap_item(); | |
2297 | if (rmap_item) { | |
2298 | /* It has already been zeroed */ | |
58730ab6 | 2299 | rmap_item->mm = mm_slot->slot.mm; |
cb4df4ca | 2300 | rmap_item->mm->ksm_rmap_items++; |
31dbd01f | 2301 | rmap_item->address = addr; |
6514d511 HD |
2302 | rmap_item->rmap_list = *rmap_list; |
2303 | *rmap_list = rmap_item; | |
31dbd01f IE |
2304 | } |
2305 | return rmap_item; | |
2306 | } | |
2307 | ||
21fbd591 | 2308 | static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page) |
31dbd01f IE |
2309 | { |
2310 | struct mm_struct *mm; | |
58730ab6 QZ |
2311 | struct ksm_mm_slot *mm_slot; |
2312 | struct mm_slot *slot; | |
31dbd01f | 2313 | struct vm_area_struct *vma; |
21fbd591 | 2314 | struct ksm_rmap_item *rmap_item; |
a5f18ba0 | 2315 | struct vma_iterator vmi; |
90bd6fd3 | 2316 | int nid; |
31dbd01f | 2317 | |
58730ab6 | 2318 | if (list_empty(&ksm_mm_head.slot.mm_node)) |
31dbd01f IE |
2319 | return NULL; |
2320 | ||
58730ab6 QZ |
2321 | mm_slot = ksm_scan.mm_slot; |
2322 | if (mm_slot == &ksm_mm_head) { | |
739100c8 SR |
2323 | trace_ksm_start_scan(ksm_scan.seqnr, ksm_rmap_items); |
2324 | ||
2919bfd0 | 2325 | /* |
1fec6890 MWO |
2326 | * A number of pages can hang around indefinitely in per-cpu |
2327 | * LRU cache, raised page count preventing write_protect_page | |
2919bfd0 HD |
2328 | * from merging them. Though it doesn't really matter much, |
2329 | * it is puzzling to see some stuck in pages_volatile until | |
2330 | * other activity jostles them out, and they also prevented | |
2331 | * LTP's KSM test from succeeding deterministically; so drain | |
2332 | * them here (here rather than on entry to ksm_do_scan(), | |
2333 | * so we don't IPI too often when pages_to_scan is set low). | |
2334 | */ | |
2335 | lru_add_drain_all(); | |
2336 | ||
4146d2d6 HD |
2337 | /* |
2338 | * Whereas stale stable_nodes on the stable_tree itself | |
2339 | * get pruned in the regular course of stable_tree_search(), | |
2340 | * those moved out to the migrate_nodes list can accumulate: | |
2341 | * so prune them once before each full scan. | |
2342 | */ | |
2343 | if (!ksm_merge_across_nodes) { | |
21fbd591 | 2344 | struct ksm_stable_node *stable_node, *next; |
4146d2d6 HD |
2345 | struct page *page; |
2346 | ||
03640418 GT |
2347 | list_for_each_entry_safe(stable_node, next, |
2348 | &migrate_nodes, list) { | |
2cee57d1 YS |
2349 | page = get_ksm_page(stable_node, |
2350 | GET_KSM_PAGE_NOLOCK); | |
4146d2d6 HD |
2351 | if (page) |
2352 | put_page(page); | |
2353 | cond_resched(); | |
2354 | } | |
2355 | } | |
2356 | ||
ef53d16c | 2357 | for (nid = 0; nid < ksm_nr_node_ids; nid++) |
90bd6fd3 | 2358 | root_unstable_tree[nid] = RB_ROOT; |
31dbd01f IE |
2359 | |
2360 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
2361 | slot = list_entry(mm_slot->slot.mm_node.next, |
2362 | struct mm_slot, mm_node); | |
2363 | mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
2364 | ksm_scan.mm_slot = mm_slot; | |
31dbd01f | 2365 | spin_unlock(&ksm_mmlist_lock); |
2b472611 HD |
2366 | /* |
2367 | * Although we tested list_empty() above, a racing __ksm_exit | |
2368 | * of the last mm on the list may have removed it since then. | |
2369 | */ | |
58730ab6 | 2370 | if (mm_slot == &ksm_mm_head) |
2b472611 | 2371 | return NULL; |
31dbd01f IE |
2372 | next_mm: |
2373 | ksm_scan.address = 0; | |
58730ab6 | 2374 | ksm_scan.rmap_list = &mm_slot->rmap_list; |
31dbd01f IE |
2375 | } |
2376 | ||
58730ab6 | 2377 | slot = &mm_slot->slot; |
31dbd01f | 2378 | mm = slot->mm; |
a5f18ba0 MWO |
2379 | vma_iter_init(&vmi, mm, ksm_scan.address); |
2380 | ||
d8ed45c5 | 2381 | mmap_read_lock(mm); |
9ba69294 | 2382 | if (ksm_test_exit(mm)) |
a5f18ba0 | 2383 | goto no_vmas; |
9ba69294 | 2384 | |
a5f18ba0 | 2385 | for_each_vma(vmi, vma) { |
31dbd01f IE |
2386 | if (!(vma->vm_flags & VM_MERGEABLE)) |
2387 | continue; | |
2388 | if (ksm_scan.address < vma->vm_start) | |
2389 | ksm_scan.address = vma->vm_start; | |
2390 | if (!vma->anon_vma) | |
2391 | ksm_scan.address = vma->vm_end; | |
2392 | ||
2393 | while (ksm_scan.address < vma->vm_end) { | |
9ba69294 HD |
2394 | if (ksm_test_exit(mm)) |
2395 | break; | |
31dbd01f | 2396 | *page = follow_page(vma, ksm_scan.address, FOLL_GET); |
f7091ed6 | 2397 | if (IS_ERR_OR_NULL(*page)) { |
21ae5b01 AA |
2398 | ksm_scan.address += PAGE_SIZE; |
2399 | cond_resched(); | |
2400 | continue; | |
2401 | } | |
f7091ed6 HW |
2402 | if (is_zone_device_page(*page)) |
2403 | goto next_page; | |
f765f540 | 2404 | if (PageAnon(*page)) { |
31dbd01f IE |
2405 | flush_anon_page(vma, *page, ksm_scan.address); |
2406 | flush_dcache_page(*page); | |
58730ab6 | 2407 | rmap_item = get_next_rmap_item(mm_slot, |
6514d511 | 2408 | ksm_scan.rmap_list, ksm_scan.address); |
31dbd01f | 2409 | if (rmap_item) { |
6514d511 HD |
2410 | ksm_scan.rmap_list = |
2411 | &rmap_item->rmap_list; | |
31dbd01f IE |
2412 | ksm_scan.address += PAGE_SIZE; |
2413 | } else | |
2414 | put_page(*page); | |
d8ed45c5 | 2415 | mmap_read_unlock(mm); |
31dbd01f IE |
2416 | return rmap_item; |
2417 | } | |
f7091ed6 | 2418 | next_page: |
21ae5b01 | 2419 | put_page(*page); |
31dbd01f IE |
2420 | ksm_scan.address += PAGE_SIZE; |
2421 | cond_resched(); | |
2422 | } | |
2423 | } | |
2424 | ||
9ba69294 | 2425 | if (ksm_test_exit(mm)) { |
a5f18ba0 | 2426 | no_vmas: |
9ba69294 | 2427 | ksm_scan.address = 0; |
58730ab6 | 2428 | ksm_scan.rmap_list = &mm_slot->rmap_list; |
9ba69294 | 2429 | } |
31dbd01f IE |
2430 | /* |
2431 | * Nuke all the rmap_items that are above this current rmap: | |
2432 | * because there were no VM_MERGEABLE vmas with such addresses. | |
2433 | */ | |
420be4ed | 2434 | remove_trailing_rmap_items(ksm_scan.rmap_list); |
31dbd01f IE |
2435 | |
2436 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
2437 | slot = list_entry(mm_slot->slot.mm_node.next, |
2438 | struct mm_slot, mm_node); | |
2439 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
cd551f97 HD |
2440 | if (ksm_scan.address == 0) { |
2441 | /* | |
c1e8d7c6 | 2442 | * We've completed a full scan of all vmas, holding mmap_lock |
cd551f97 HD |
2443 | * throughout, and found no VM_MERGEABLE: so do the same as |
2444 | * __ksm_exit does to remove this mm from all our lists now. | |
9ba69294 HD |
2445 | * This applies either when cleaning up after __ksm_exit |
2446 | * (but beware: we can reach here even before __ksm_exit), | |
2447 | * or when all VM_MERGEABLE areas have been unmapped (and | |
c1e8d7c6 | 2448 | * mmap_lock then protects against race with MADV_MERGEABLE). |
cd551f97 | 2449 | */ |
58730ab6 QZ |
2450 | hash_del(&mm_slot->slot.hash); |
2451 | list_del(&mm_slot->slot.mm_node); | |
9ba69294 HD |
2452 | spin_unlock(&ksm_mmlist_lock); |
2453 | ||
58730ab6 | 2454 | mm_slot_free(mm_slot_cache, mm_slot); |
cd551f97 | 2455 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
d7597f59 | 2456 | clear_bit(MMF_VM_MERGE_ANY, &mm->flags); |
d8ed45c5 | 2457 | mmap_read_unlock(mm); |
9ba69294 HD |
2458 | mmdrop(mm); |
2459 | } else { | |
d8ed45c5 | 2460 | mmap_read_unlock(mm); |
7496fea9 | 2461 | /* |
3e4e28c5 | 2462 | * mmap_read_unlock(mm) first because after |
7496fea9 ZC |
2463 | * spin_unlock(&ksm_mmlist_lock) run, the "mm" may |
2464 | * already have been freed under us by __ksm_exit() | |
2465 | * because the "mm_slot" is still hashed and | |
2466 | * ksm_scan.mm_slot doesn't point to it anymore. | |
2467 | */ | |
2468 | spin_unlock(&ksm_mmlist_lock); | |
cd551f97 | 2469 | } |
31dbd01f IE |
2470 | |
2471 | /* Repeat until we've completed scanning the whole list */ | |
58730ab6 QZ |
2472 | mm_slot = ksm_scan.mm_slot; |
2473 | if (mm_slot != &ksm_mm_head) | |
31dbd01f IE |
2474 | goto next_mm; |
2475 | ||
739100c8 | 2476 | trace_ksm_stop_scan(ksm_scan.seqnr, ksm_rmap_items); |
31dbd01f IE |
2477 | ksm_scan.seqnr++; |
2478 | return NULL; | |
2479 | } | |
2480 | ||
2481 | /** | |
2482 | * ksm_do_scan - the ksm scanner main worker function. | |
b7701a5f | 2483 | * @scan_npages: number of pages we want to scan before we return. |
31dbd01f IE |
2484 | */ |
2485 | static void ksm_do_scan(unsigned int scan_npages) | |
2486 | { | |
21fbd591 | 2487 | struct ksm_rmap_item *rmap_item; |
3f649ab7 | 2488 | struct page *page; |
b348b5fe | 2489 | unsigned int npages = scan_npages; |
31dbd01f | 2490 | |
b348b5fe | 2491 | while (npages-- && likely(!freezing(current))) { |
31dbd01f IE |
2492 | cond_resched(); |
2493 | rmap_item = scan_get_next_rmap_item(&page); | |
2494 | if (!rmap_item) | |
2495 | return; | |
4146d2d6 | 2496 | cmp_and_merge_page(page, rmap_item); |
31dbd01f IE |
2497 | put_page(page); |
2498 | } | |
b348b5fe SR |
2499 | |
2500 | ksm_pages_scanned += scan_npages - npages; | |
31dbd01f IE |
2501 | } |
2502 | ||
6e158384 HD |
2503 | static int ksmd_should_run(void) |
2504 | { | |
58730ab6 | 2505 | return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.slot.mm_node); |
6e158384 HD |
2506 | } |
2507 | ||
31dbd01f IE |
2508 | static int ksm_scan_thread(void *nothing) |
2509 | { | |
fcf9a0ef KT |
2510 | unsigned int sleep_ms; |
2511 | ||
878aee7d | 2512 | set_freezable(); |
339aa624 | 2513 | set_user_nice(current, 5); |
31dbd01f IE |
2514 | |
2515 | while (!kthread_should_stop()) { | |
6e158384 | 2516 | mutex_lock(&ksm_thread_mutex); |
ef4d43a8 | 2517 | wait_while_offlining(); |
6e158384 | 2518 | if (ksmd_should_run()) |
31dbd01f | 2519 | ksm_do_scan(ksm_thread_pages_to_scan); |
6e158384 HD |
2520 | mutex_unlock(&ksm_thread_mutex); |
2521 | ||
878aee7d AA |
2522 | try_to_freeze(); |
2523 | ||
6e158384 | 2524 | if (ksmd_should_run()) { |
fcf9a0ef KT |
2525 | sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs); |
2526 | wait_event_interruptible_timeout(ksm_iter_wait, | |
2527 | sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs), | |
2528 | msecs_to_jiffies(sleep_ms)); | |
31dbd01f | 2529 | } else { |
878aee7d | 2530 | wait_event_freezable(ksm_thread_wait, |
6e158384 | 2531 | ksmd_should_run() || kthread_should_stop()); |
31dbd01f IE |
2532 | } |
2533 | } | |
2534 | return 0; | |
2535 | } | |
2536 | ||
d7597f59 SR |
2537 | static void __ksm_add_vma(struct vm_area_struct *vma) |
2538 | { | |
2539 | unsigned long vm_flags = vma->vm_flags; | |
2540 | ||
2541 | if (vm_flags & VM_MERGEABLE) | |
2542 | return; | |
2543 | ||
2544 | if (vma_ksm_compatible(vma)) | |
2545 | vm_flags_set(vma, VM_MERGEABLE); | |
2546 | } | |
2547 | ||
24139c07 DH |
2548 | static int __ksm_del_vma(struct vm_area_struct *vma) |
2549 | { | |
2550 | int err; | |
2551 | ||
2552 | if (!(vma->vm_flags & VM_MERGEABLE)) | |
2553 | return 0; | |
2554 | ||
2555 | if (vma->anon_vma) { | |
49b06385 | 2556 | err = unmerge_ksm_pages(vma, vma->vm_start, vma->vm_end, true); |
24139c07 DH |
2557 | if (err) |
2558 | return err; | |
2559 | } | |
2560 | ||
2561 | vm_flags_clear(vma, VM_MERGEABLE); | |
2562 | return 0; | |
2563 | } | |
d7597f59 SR |
2564 | /** |
2565 | * ksm_add_vma - Mark vma as mergeable if compatible | |
2566 | * | |
2567 | * @vma: Pointer to vma | |
2568 | */ | |
2569 | void ksm_add_vma(struct vm_area_struct *vma) | |
2570 | { | |
2571 | struct mm_struct *mm = vma->vm_mm; | |
2572 | ||
2573 | if (test_bit(MMF_VM_MERGE_ANY, &mm->flags)) | |
2574 | __ksm_add_vma(vma); | |
2575 | } | |
2576 | ||
2577 | static void ksm_add_vmas(struct mm_struct *mm) | |
2578 | { | |
2579 | struct vm_area_struct *vma; | |
2580 | ||
2581 | VMA_ITERATOR(vmi, mm, 0); | |
2582 | for_each_vma(vmi, vma) | |
2583 | __ksm_add_vma(vma); | |
2584 | } | |
2585 | ||
24139c07 DH |
2586 | static int ksm_del_vmas(struct mm_struct *mm) |
2587 | { | |
2588 | struct vm_area_struct *vma; | |
2589 | int err; | |
2590 | ||
2591 | VMA_ITERATOR(vmi, mm, 0); | |
2592 | for_each_vma(vmi, vma) { | |
2593 | err = __ksm_del_vma(vma); | |
2594 | if (err) | |
2595 | return err; | |
2596 | } | |
2597 | return 0; | |
2598 | } | |
2599 | ||
d7597f59 SR |
2600 | /** |
2601 | * ksm_enable_merge_any - Add mm to mm ksm list and enable merging on all | |
2602 | * compatible VMA's | |
2603 | * | |
2604 | * @mm: Pointer to mm | |
2605 | * | |
2606 | * Returns 0 on success, otherwise error code | |
2607 | */ | |
2608 | int ksm_enable_merge_any(struct mm_struct *mm) | |
2609 | { | |
2610 | int err; | |
2611 | ||
2612 | if (test_bit(MMF_VM_MERGE_ANY, &mm->flags)) | |
2613 | return 0; | |
2614 | ||
2615 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { | |
2616 | err = __ksm_enter(mm); | |
2617 | if (err) | |
2618 | return err; | |
2619 | } | |
2620 | ||
2621 | set_bit(MMF_VM_MERGE_ANY, &mm->flags); | |
2622 | ksm_add_vmas(mm); | |
2623 | ||
2624 | return 0; | |
2625 | } | |
2626 | ||
24139c07 DH |
2627 | /** |
2628 | * ksm_disable_merge_any - Disable merging on all compatible VMA's of the mm, | |
2629 | * previously enabled via ksm_enable_merge_any(). | |
2630 | * | |
2631 | * Disabling merging implies unmerging any merged pages, like setting | |
2632 | * MADV_UNMERGEABLE would. If unmerging fails, the whole operation fails and | |
2633 | * merging on all compatible VMA's remains enabled. | |
2634 | * | |
2635 | * @mm: Pointer to mm | |
2636 | * | |
2637 | * Returns 0 on success, otherwise error code | |
2638 | */ | |
2639 | int ksm_disable_merge_any(struct mm_struct *mm) | |
2640 | { | |
2641 | int err; | |
2642 | ||
2643 | if (!test_bit(MMF_VM_MERGE_ANY, &mm->flags)) | |
2644 | return 0; | |
2645 | ||
2646 | err = ksm_del_vmas(mm); | |
2647 | if (err) { | |
2648 | ksm_add_vmas(mm); | |
2649 | return err; | |
2650 | } | |
2651 | ||
2652 | clear_bit(MMF_VM_MERGE_ANY, &mm->flags); | |
2653 | return 0; | |
2654 | } | |
2655 | ||
2c281f54 DH |
2656 | int ksm_disable(struct mm_struct *mm) |
2657 | { | |
2658 | mmap_assert_write_locked(mm); | |
2659 | ||
2660 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) | |
2661 | return 0; | |
2662 | if (test_bit(MMF_VM_MERGE_ANY, &mm->flags)) | |
2663 | return ksm_disable_merge_any(mm); | |
2664 | return ksm_del_vmas(mm); | |
2665 | } | |
2666 | ||
f8af4da3 HD |
2667 | int ksm_madvise(struct vm_area_struct *vma, unsigned long start, |
2668 | unsigned long end, int advice, unsigned long *vm_flags) | |
2669 | { | |
2670 | struct mm_struct *mm = vma->vm_mm; | |
d952b791 | 2671 | int err; |
f8af4da3 HD |
2672 | |
2673 | switch (advice) { | |
2674 | case MADV_MERGEABLE: | |
d7597f59 | 2675 | if (vma->vm_flags & VM_MERGEABLE) |
e1fb4a08 | 2676 | return 0; |
d7597f59 | 2677 | if (!vma_ksm_compatible(vma)) |
74a04967 | 2678 | return 0; |
cc2383ec | 2679 | |
d952b791 HD |
2680 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { |
2681 | err = __ksm_enter(mm); | |
2682 | if (err) | |
2683 | return err; | |
2684 | } | |
f8af4da3 HD |
2685 | |
2686 | *vm_flags |= VM_MERGEABLE; | |
2687 | break; | |
2688 | ||
2689 | case MADV_UNMERGEABLE: | |
2690 | if (!(*vm_flags & VM_MERGEABLE)) | |
2691 | return 0; /* just ignore the advice */ | |
2692 | ||
d952b791 | 2693 | if (vma->anon_vma) { |
49b06385 | 2694 | err = unmerge_ksm_pages(vma, start, end, true); |
d952b791 HD |
2695 | if (err) |
2696 | return err; | |
2697 | } | |
f8af4da3 HD |
2698 | |
2699 | *vm_flags &= ~VM_MERGEABLE; | |
2700 | break; | |
2701 | } | |
2702 | ||
2703 | return 0; | |
2704 | } | |
33cf1707 | 2705 | EXPORT_SYMBOL_GPL(ksm_madvise); |
f8af4da3 HD |
2706 | |
2707 | int __ksm_enter(struct mm_struct *mm) | |
2708 | { | |
21fbd591 | 2709 | struct ksm_mm_slot *mm_slot; |
58730ab6 | 2710 | struct mm_slot *slot; |
6e158384 HD |
2711 | int needs_wakeup; |
2712 | ||
58730ab6 | 2713 | mm_slot = mm_slot_alloc(mm_slot_cache); |
31dbd01f IE |
2714 | if (!mm_slot) |
2715 | return -ENOMEM; | |
2716 | ||
58730ab6 QZ |
2717 | slot = &mm_slot->slot; |
2718 | ||
6e158384 | 2719 | /* Check ksm_run too? Would need tighter locking */ |
58730ab6 | 2720 | needs_wakeup = list_empty(&ksm_mm_head.slot.mm_node); |
6e158384 | 2721 | |
31dbd01f | 2722 | spin_lock(&ksm_mmlist_lock); |
58730ab6 | 2723 | mm_slot_insert(mm_slots_hash, mm, slot); |
31dbd01f | 2724 | /* |
cbf86cfe HD |
2725 | * When KSM_RUN_MERGE (or KSM_RUN_STOP), |
2726 | * insert just behind the scanning cursor, to let the area settle | |
31dbd01f IE |
2727 | * down a little; when fork is followed by immediate exec, we don't |
2728 | * want ksmd to waste time setting up and tearing down an rmap_list. | |
cbf86cfe HD |
2729 | * |
2730 | * But when KSM_RUN_UNMERGE, it's important to insert ahead of its | |
2731 | * scanning cursor, otherwise KSM pages in newly forked mms will be | |
2732 | * missed: then we might as well insert at the end of the list. | |
31dbd01f | 2733 | */ |
cbf86cfe | 2734 | if (ksm_run & KSM_RUN_UNMERGE) |
58730ab6 | 2735 | list_add_tail(&slot->mm_node, &ksm_mm_head.slot.mm_node); |
cbf86cfe | 2736 | else |
58730ab6 | 2737 | list_add_tail(&slot->mm_node, &ksm_scan.mm_slot->slot.mm_node); |
31dbd01f IE |
2738 | spin_unlock(&ksm_mmlist_lock); |
2739 | ||
f8af4da3 | 2740 | set_bit(MMF_VM_MERGEABLE, &mm->flags); |
f1f10076 | 2741 | mmgrab(mm); |
6e158384 HD |
2742 | |
2743 | if (needs_wakeup) | |
2744 | wake_up_interruptible(&ksm_thread_wait); | |
2745 | ||
739100c8 | 2746 | trace_ksm_enter(mm); |
f8af4da3 HD |
2747 | return 0; |
2748 | } | |
2749 | ||
1c2fb7a4 | 2750 | void __ksm_exit(struct mm_struct *mm) |
f8af4da3 | 2751 | { |
21fbd591 | 2752 | struct ksm_mm_slot *mm_slot; |
58730ab6 | 2753 | struct mm_slot *slot; |
9ba69294 | 2754 | int easy_to_free = 0; |
cd551f97 | 2755 | |
31dbd01f | 2756 | /* |
9ba69294 HD |
2757 | * This process is exiting: if it's straightforward (as is the |
2758 | * case when ksmd was never running), free mm_slot immediately. | |
2759 | * But if it's at the cursor or has rmap_items linked to it, use | |
c1e8d7c6 | 2760 | * mmap_lock to synchronize with any break_cows before pagetables |
9ba69294 HD |
2761 | * are freed, and leave the mm_slot on the list for ksmd to free. |
2762 | * Beware: ksm may already have noticed it exiting and freed the slot. | |
31dbd01f | 2763 | */ |
9ba69294 | 2764 | |
cd551f97 | 2765 | spin_lock(&ksm_mmlist_lock); |
58730ab6 QZ |
2766 | slot = mm_slot_lookup(mm_slots_hash, mm); |
2767 | mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
9ba69294 | 2768 | if (mm_slot && ksm_scan.mm_slot != mm_slot) { |
6514d511 | 2769 | if (!mm_slot->rmap_list) { |
58730ab6 QZ |
2770 | hash_del(&slot->hash); |
2771 | list_del(&slot->mm_node); | |
9ba69294 HD |
2772 | easy_to_free = 1; |
2773 | } else { | |
58730ab6 QZ |
2774 | list_move(&slot->mm_node, |
2775 | &ksm_scan.mm_slot->slot.mm_node); | |
9ba69294 | 2776 | } |
cd551f97 | 2777 | } |
cd551f97 HD |
2778 | spin_unlock(&ksm_mmlist_lock); |
2779 | ||
9ba69294 | 2780 | if (easy_to_free) { |
58730ab6 | 2781 | mm_slot_free(mm_slot_cache, mm_slot); |
d7597f59 | 2782 | clear_bit(MMF_VM_MERGE_ANY, &mm->flags); |
9ba69294 HD |
2783 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
2784 | mmdrop(mm); | |
2785 | } else if (mm_slot) { | |
d8ed45c5 ML |
2786 | mmap_write_lock(mm); |
2787 | mmap_write_unlock(mm); | |
9ba69294 | 2788 | } |
739100c8 SR |
2789 | |
2790 | trace_ksm_exit(mm); | |
31dbd01f IE |
2791 | } |
2792 | ||
cbf86cfe | 2793 | struct page *ksm_might_need_to_copy(struct page *page, |
5ad64688 HD |
2794 | struct vm_area_struct *vma, unsigned long address) |
2795 | { | |
e05b3453 MWO |
2796 | struct folio *folio = page_folio(page); |
2797 | struct anon_vma *anon_vma = folio_anon_vma(folio); | |
5ad64688 HD |
2798 | struct page *new_page; |
2799 | ||
cbf86cfe HD |
2800 | if (PageKsm(page)) { |
2801 | if (page_stable_node(page) && | |
2802 | !(ksm_run & KSM_RUN_UNMERGE)) | |
2803 | return page; /* no need to copy it */ | |
2804 | } else if (!anon_vma) { | |
2805 | return page; /* no need to copy it */ | |
e1c63e11 NS |
2806 | } else if (page->index == linear_page_index(vma, address) && |
2807 | anon_vma->root == vma->anon_vma->root) { | |
cbf86cfe HD |
2808 | return page; /* still no need to copy it */ |
2809 | } | |
f985fc32 ML |
2810 | if (PageHWPoison(page)) |
2811 | return ERR_PTR(-EHWPOISON); | |
cbf86cfe HD |
2812 | if (!PageUptodate(page)) |
2813 | return page; /* let do_swap_page report the error */ | |
2814 | ||
5ad64688 | 2815 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
8f425e4e MWO |
2816 | if (new_page && |
2817 | mem_cgroup_charge(page_folio(new_page), vma->vm_mm, GFP_KERNEL)) { | |
62fdb163 HD |
2818 | put_page(new_page); |
2819 | new_page = NULL; | |
2820 | } | |
5ad64688 | 2821 | if (new_page) { |
6b970599 KW |
2822 | if (copy_mc_user_highpage(new_page, page, address, vma)) { |
2823 | put_page(new_page); | |
2824 | memory_failure_queue(page_to_pfn(page), 0); | |
2825 | return ERR_PTR(-EHWPOISON); | |
2826 | } | |
5ad64688 HD |
2827 | SetPageDirty(new_page); |
2828 | __SetPageUptodate(new_page); | |
48c935ad | 2829 | __SetPageLocked(new_page); |
4d45c3af YY |
2830 | #ifdef CONFIG_SWAP |
2831 | count_vm_event(KSM_SWPIN_COPY); | |
2832 | #endif | |
5ad64688 HD |
2833 | } |
2834 | ||
5ad64688 HD |
2835 | return new_page; |
2836 | } | |
2837 | ||
6d4675e6 | 2838 | void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc) |
e9995ef9 | 2839 | { |
21fbd591 QZ |
2840 | struct ksm_stable_node *stable_node; |
2841 | struct ksm_rmap_item *rmap_item; | |
e9995ef9 HD |
2842 | int search_new_forks = 0; |
2843 | ||
2f031c6f | 2844 | VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio); |
9f32624b JK |
2845 | |
2846 | /* | |
2847 | * Rely on the page lock to protect against concurrent modifications | |
2848 | * to that page's node of the stable tree. | |
2849 | */ | |
2f031c6f | 2850 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
e9995ef9 | 2851 | |
2f031c6f | 2852 | stable_node = folio_stable_node(folio); |
e9995ef9 | 2853 | if (!stable_node) |
1df631ae | 2854 | return; |
e9995ef9 | 2855 | again: |
b67bfe0d | 2856 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
e9995ef9 | 2857 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
5beb4930 | 2858 | struct anon_vma_chain *vmac; |
e9995ef9 HD |
2859 | struct vm_area_struct *vma; |
2860 | ||
ad12695f | 2861 | cond_resched(); |
6d4675e6 MK |
2862 | if (!anon_vma_trylock_read(anon_vma)) { |
2863 | if (rwc->try_lock) { | |
2864 | rwc->contended = true; | |
2865 | return; | |
2866 | } | |
2867 | anon_vma_lock_read(anon_vma); | |
2868 | } | |
bf181b9f ML |
2869 | anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, |
2870 | 0, ULONG_MAX) { | |
1105a2fc JH |
2871 | unsigned long addr; |
2872 | ||
ad12695f | 2873 | cond_resched(); |
5beb4930 | 2874 | vma = vmac->vma; |
1105a2fc JH |
2875 | |
2876 | /* Ignore the stable/unstable/sqnr flags */ | |
cd7fae26 | 2877 | addr = rmap_item->address & PAGE_MASK; |
1105a2fc JH |
2878 | |
2879 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
e9995ef9 HD |
2880 | continue; |
2881 | /* | |
2882 | * Initially we examine only the vma which covers this | |
2883 | * rmap_item; but later, if there is still work to do, | |
2884 | * we examine covering vmas in other mms: in case they | |
2885 | * were forked from the original since ksmd passed. | |
2886 | */ | |
2887 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
2888 | continue; | |
2889 | ||
0dd1c7bb JK |
2890 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
2891 | continue; | |
2892 | ||
2f031c6f | 2893 | if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) { |
b6b19f25 | 2894 | anon_vma_unlock_read(anon_vma); |
1df631ae | 2895 | return; |
e9995ef9 | 2896 | } |
2f031c6f | 2897 | if (rwc->done && rwc->done(folio)) { |
0dd1c7bb | 2898 | anon_vma_unlock_read(anon_vma); |
1df631ae | 2899 | return; |
0dd1c7bb | 2900 | } |
e9995ef9 | 2901 | } |
b6b19f25 | 2902 | anon_vma_unlock_read(anon_vma); |
e9995ef9 HD |
2903 | } |
2904 | if (!search_new_forks++) | |
2905 | goto again; | |
e9995ef9 HD |
2906 | } |
2907 | ||
4248d008 LX |
2908 | #ifdef CONFIG_MEMORY_FAILURE |
2909 | /* | |
2910 | * Collect processes when the error hit an ksm page. | |
2911 | */ | |
2912 | void collect_procs_ksm(struct page *page, struct list_head *to_kill, | |
2913 | int force_early) | |
2914 | { | |
2915 | struct ksm_stable_node *stable_node; | |
2916 | struct ksm_rmap_item *rmap_item; | |
2917 | struct folio *folio = page_folio(page); | |
2918 | struct vm_area_struct *vma; | |
2919 | struct task_struct *tsk; | |
2920 | ||
2921 | stable_node = folio_stable_node(folio); | |
2922 | if (!stable_node) | |
2923 | return; | |
2924 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { | |
2925 | struct anon_vma *av = rmap_item->anon_vma; | |
2926 | ||
2927 | anon_vma_lock_read(av); | |
d256d1cd | 2928 | rcu_read_lock(); |
4248d008 LX |
2929 | for_each_process(tsk) { |
2930 | struct anon_vma_chain *vmac; | |
2931 | unsigned long addr; | |
2932 | struct task_struct *t = | |
2933 | task_early_kill(tsk, force_early); | |
2934 | if (!t) | |
2935 | continue; | |
2936 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, 0, | |
2937 | ULONG_MAX) | |
2938 | { | |
2939 | vma = vmac->vma; | |
2940 | if (vma->vm_mm == t->mm) { | |
2941 | addr = rmap_item->address & PAGE_MASK; | |
2942 | add_to_kill_ksm(t, page, vma, to_kill, | |
2943 | addr); | |
2944 | } | |
2945 | } | |
2946 | } | |
d256d1cd | 2947 | rcu_read_unlock(); |
4248d008 LX |
2948 | anon_vma_unlock_read(av); |
2949 | } | |
2950 | } | |
2951 | #endif | |
2952 | ||
52629506 | 2953 | #ifdef CONFIG_MIGRATION |
19138349 | 2954 | void folio_migrate_ksm(struct folio *newfolio, struct folio *folio) |
e9995ef9 | 2955 | { |
21fbd591 | 2956 | struct ksm_stable_node *stable_node; |
e9995ef9 | 2957 | |
19138349 MWO |
2958 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
2959 | VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio); | |
2960 | VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio); | |
e9995ef9 | 2961 | |
19138349 | 2962 | stable_node = folio_stable_node(folio); |
e9995ef9 | 2963 | if (stable_node) { |
19138349 MWO |
2964 | VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio); |
2965 | stable_node->kpfn = folio_pfn(newfolio); | |
c8d6553b | 2966 | /* |
19138349 | 2967 | * newfolio->mapping was set in advance; now we need smp_wmb() |
c8d6553b | 2968 | * to make sure that the new stable_node->kpfn is visible |
19138349 MWO |
2969 | * to get_ksm_page() before it can see that folio->mapping |
2970 | * has gone stale (or that folio_test_swapcache has been cleared). | |
c8d6553b HD |
2971 | */ |
2972 | smp_wmb(); | |
19138349 | 2973 | set_page_stable_node(&folio->page, NULL); |
e9995ef9 HD |
2974 | } |
2975 | } | |
2976 | #endif /* CONFIG_MIGRATION */ | |
2977 | ||
62b61f61 | 2978 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 HD |
2979 | static void wait_while_offlining(void) |
2980 | { | |
2981 | while (ksm_run & KSM_RUN_OFFLINE) { | |
2982 | mutex_unlock(&ksm_thread_mutex); | |
2983 | wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), | |
74316201 | 2984 | TASK_UNINTERRUPTIBLE); |
ef4d43a8 HD |
2985 | mutex_lock(&ksm_thread_mutex); |
2986 | } | |
2987 | } | |
2988 | ||
21fbd591 | 2989 | static bool stable_node_dup_remove_range(struct ksm_stable_node *stable_node, |
2c653d0e AA |
2990 | unsigned long start_pfn, |
2991 | unsigned long end_pfn) | |
2992 | { | |
2993 | if (stable_node->kpfn >= start_pfn && | |
2994 | stable_node->kpfn < end_pfn) { | |
2995 | /* | |
2996 | * Don't get_ksm_page, page has already gone: | |
2997 | * which is why we keep kpfn instead of page* | |
2998 | */ | |
2999 | remove_node_from_stable_tree(stable_node); | |
3000 | return true; | |
3001 | } | |
3002 | return false; | |
3003 | } | |
3004 | ||
21fbd591 | 3005 | static bool stable_node_chain_remove_range(struct ksm_stable_node *stable_node, |
2c653d0e AA |
3006 | unsigned long start_pfn, |
3007 | unsigned long end_pfn, | |
3008 | struct rb_root *root) | |
3009 | { | |
21fbd591 | 3010 | struct ksm_stable_node *dup; |
2c653d0e AA |
3011 | struct hlist_node *hlist_safe; |
3012 | ||
3013 | if (!is_stable_node_chain(stable_node)) { | |
3014 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
3015 | return stable_node_dup_remove_range(stable_node, start_pfn, | |
3016 | end_pfn); | |
3017 | } | |
3018 | ||
3019 | hlist_for_each_entry_safe(dup, hlist_safe, | |
3020 | &stable_node->hlist, hlist_dup) { | |
3021 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
3022 | stable_node_dup_remove_range(dup, start_pfn, end_pfn); | |
3023 | } | |
3024 | if (hlist_empty(&stable_node->hlist)) { | |
3025 | free_stable_node_chain(stable_node, root); | |
3026 | return true; /* notify caller that tree was rebalanced */ | |
3027 | } else | |
3028 | return false; | |
3029 | } | |
3030 | ||
ee0ea59c HD |
3031 | static void ksm_check_stable_tree(unsigned long start_pfn, |
3032 | unsigned long end_pfn) | |
62b61f61 | 3033 | { |
21fbd591 | 3034 | struct ksm_stable_node *stable_node, *next; |
62b61f61 | 3035 | struct rb_node *node; |
90bd6fd3 | 3036 | int nid; |
62b61f61 | 3037 | |
ef53d16c HD |
3038 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
3039 | node = rb_first(root_stable_tree + nid); | |
ee0ea59c | 3040 | while (node) { |
21fbd591 | 3041 | stable_node = rb_entry(node, struct ksm_stable_node, node); |
2c653d0e AA |
3042 | if (stable_node_chain_remove_range(stable_node, |
3043 | start_pfn, end_pfn, | |
3044 | root_stable_tree + | |
3045 | nid)) | |
ef53d16c | 3046 | node = rb_first(root_stable_tree + nid); |
2c653d0e | 3047 | else |
ee0ea59c HD |
3048 | node = rb_next(node); |
3049 | cond_resched(); | |
90bd6fd3 | 3050 | } |
ee0ea59c | 3051 | } |
03640418 | 3052 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
3053 | if (stable_node->kpfn >= start_pfn && |
3054 | stable_node->kpfn < end_pfn) | |
3055 | remove_node_from_stable_tree(stable_node); | |
3056 | cond_resched(); | |
3057 | } | |
62b61f61 HD |
3058 | } |
3059 | ||
3060 | static int ksm_memory_callback(struct notifier_block *self, | |
3061 | unsigned long action, void *arg) | |
3062 | { | |
3063 | struct memory_notify *mn = arg; | |
62b61f61 HD |
3064 | |
3065 | switch (action) { | |
3066 | case MEM_GOING_OFFLINE: | |
3067 | /* | |
ef4d43a8 HD |
3068 | * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() |
3069 | * and remove_all_stable_nodes() while memory is going offline: | |
3070 | * it is unsafe for them to touch the stable tree at this time. | |
3071 | * But unmerge_ksm_pages(), rmap lookups and other entry points | |
3072 | * which do not need the ksm_thread_mutex are all safe. | |
62b61f61 | 3073 | */ |
ef4d43a8 HD |
3074 | mutex_lock(&ksm_thread_mutex); |
3075 | ksm_run |= KSM_RUN_OFFLINE; | |
3076 | mutex_unlock(&ksm_thread_mutex); | |
62b61f61 HD |
3077 | break; |
3078 | ||
3079 | case MEM_OFFLINE: | |
3080 | /* | |
3081 | * Most of the work is done by page migration; but there might | |
3082 | * be a few stable_nodes left over, still pointing to struct | |
ee0ea59c HD |
3083 | * pages which have been offlined: prune those from the tree, |
3084 | * otherwise get_ksm_page() might later try to access a | |
3085 | * non-existent struct page. | |
62b61f61 | 3086 | */ |
ee0ea59c HD |
3087 | ksm_check_stable_tree(mn->start_pfn, |
3088 | mn->start_pfn + mn->nr_pages); | |
e4a9bc58 | 3089 | fallthrough; |
62b61f61 | 3090 | case MEM_CANCEL_OFFLINE: |
ef4d43a8 HD |
3091 | mutex_lock(&ksm_thread_mutex); |
3092 | ksm_run &= ~KSM_RUN_OFFLINE; | |
62b61f61 | 3093 | mutex_unlock(&ksm_thread_mutex); |
ef4d43a8 HD |
3094 | |
3095 | smp_mb(); /* wake_up_bit advises this */ | |
3096 | wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); | |
62b61f61 HD |
3097 | break; |
3098 | } | |
3099 | return NOTIFY_OK; | |
3100 | } | |
ef4d43a8 HD |
3101 | #else |
3102 | static void wait_while_offlining(void) | |
3103 | { | |
3104 | } | |
62b61f61 HD |
3105 | #endif /* CONFIG_MEMORY_HOTREMOVE */ |
3106 | ||
d21077fb SR |
3107 | #ifdef CONFIG_PROC_FS |
3108 | long ksm_process_profit(struct mm_struct *mm) | |
3109 | { | |
1a8e8430 | 3110 | return (long)(mm->ksm_merging_pages + mm->ksm_zero_pages) * PAGE_SIZE - |
d21077fb SR |
3111 | mm->ksm_rmap_items * sizeof(struct ksm_rmap_item); |
3112 | } | |
3113 | #endif /* CONFIG_PROC_FS */ | |
3114 | ||
2ffd8679 HD |
3115 | #ifdef CONFIG_SYSFS |
3116 | /* | |
3117 | * This all compiles without CONFIG_SYSFS, but is a waste of space. | |
3118 | */ | |
3119 | ||
31dbd01f IE |
3120 | #define KSM_ATTR_RO(_name) \ |
3121 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
3122 | #define KSM_ATTR(_name) \ | |
1bad2e5c | 3123 | static struct kobj_attribute _name##_attr = __ATTR_RW(_name) |
31dbd01f IE |
3124 | |
3125 | static ssize_t sleep_millisecs_show(struct kobject *kobj, | |
3126 | struct kobj_attribute *attr, char *buf) | |
3127 | { | |
ae7a927d | 3128 | return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs); |
31dbd01f IE |
3129 | } |
3130 | ||
3131 | static ssize_t sleep_millisecs_store(struct kobject *kobj, | |
3132 | struct kobj_attribute *attr, | |
3133 | const char *buf, size_t count) | |
3134 | { | |
dfefd226 | 3135 | unsigned int msecs; |
31dbd01f IE |
3136 | int err; |
3137 | ||
dfefd226 AD |
3138 | err = kstrtouint(buf, 10, &msecs); |
3139 | if (err) | |
31dbd01f IE |
3140 | return -EINVAL; |
3141 | ||
3142 | ksm_thread_sleep_millisecs = msecs; | |
fcf9a0ef | 3143 | wake_up_interruptible(&ksm_iter_wait); |
31dbd01f IE |
3144 | |
3145 | return count; | |
3146 | } | |
3147 | KSM_ATTR(sleep_millisecs); | |
3148 | ||
3149 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
3150 | struct kobj_attribute *attr, char *buf) | |
3151 | { | |
ae7a927d | 3152 | return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan); |
31dbd01f IE |
3153 | } |
3154 | ||
3155 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
3156 | struct kobj_attribute *attr, | |
3157 | const char *buf, size_t count) | |
3158 | { | |
dfefd226 | 3159 | unsigned int nr_pages; |
31dbd01f | 3160 | int err; |
31dbd01f | 3161 | |
dfefd226 AD |
3162 | err = kstrtouint(buf, 10, &nr_pages); |
3163 | if (err) | |
31dbd01f IE |
3164 | return -EINVAL; |
3165 | ||
3166 | ksm_thread_pages_to_scan = nr_pages; | |
3167 | ||
3168 | return count; | |
3169 | } | |
3170 | KSM_ATTR(pages_to_scan); | |
3171 | ||
3172 | static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, | |
3173 | char *buf) | |
3174 | { | |
ae7a927d | 3175 | return sysfs_emit(buf, "%lu\n", ksm_run); |
31dbd01f IE |
3176 | } |
3177 | ||
3178 | static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, | |
3179 | const char *buf, size_t count) | |
3180 | { | |
dfefd226 | 3181 | unsigned int flags; |
31dbd01f | 3182 | int err; |
31dbd01f | 3183 | |
dfefd226 AD |
3184 | err = kstrtouint(buf, 10, &flags); |
3185 | if (err) | |
31dbd01f IE |
3186 | return -EINVAL; |
3187 | if (flags > KSM_RUN_UNMERGE) | |
3188 | return -EINVAL; | |
3189 | ||
3190 | /* | |
3191 | * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. | |
3192 | * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, | |
d0f209f6 HD |
3193 | * breaking COW to free the pages_shared (but leaves mm_slots |
3194 | * on the list for when ksmd may be set running again). | |
31dbd01f IE |
3195 | */ |
3196 | ||
3197 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 3198 | wait_while_offlining(); |
31dbd01f IE |
3199 | if (ksm_run != flags) { |
3200 | ksm_run = flags; | |
d952b791 | 3201 | if (flags & KSM_RUN_UNMERGE) { |
e1e12d2f | 3202 | set_current_oom_origin(); |
d952b791 | 3203 | err = unmerge_and_remove_all_rmap_items(); |
e1e12d2f | 3204 | clear_current_oom_origin(); |
d952b791 HD |
3205 | if (err) { |
3206 | ksm_run = KSM_RUN_STOP; | |
3207 | count = err; | |
3208 | } | |
3209 | } | |
31dbd01f IE |
3210 | } |
3211 | mutex_unlock(&ksm_thread_mutex); | |
3212 | ||
3213 | if (flags & KSM_RUN_MERGE) | |
3214 | wake_up_interruptible(&ksm_thread_wait); | |
3215 | ||
3216 | return count; | |
3217 | } | |
3218 | KSM_ATTR(run); | |
3219 | ||
90bd6fd3 PH |
3220 | #ifdef CONFIG_NUMA |
3221 | static ssize_t merge_across_nodes_show(struct kobject *kobj, | |
ae7a927d | 3222 | struct kobj_attribute *attr, char *buf) |
90bd6fd3 | 3223 | { |
ae7a927d | 3224 | return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes); |
90bd6fd3 PH |
3225 | } |
3226 | ||
3227 | static ssize_t merge_across_nodes_store(struct kobject *kobj, | |
3228 | struct kobj_attribute *attr, | |
3229 | const char *buf, size_t count) | |
3230 | { | |
3231 | int err; | |
3232 | unsigned long knob; | |
3233 | ||
3234 | err = kstrtoul(buf, 10, &knob); | |
3235 | if (err) | |
3236 | return err; | |
3237 | if (knob > 1) | |
3238 | return -EINVAL; | |
3239 | ||
3240 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 3241 | wait_while_offlining(); |
90bd6fd3 | 3242 | if (ksm_merge_across_nodes != knob) { |
cbf86cfe | 3243 | if (ksm_pages_shared || remove_all_stable_nodes()) |
90bd6fd3 | 3244 | err = -EBUSY; |
ef53d16c HD |
3245 | else if (root_stable_tree == one_stable_tree) { |
3246 | struct rb_root *buf; | |
3247 | /* | |
3248 | * This is the first time that we switch away from the | |
3249 | * default of merging across nodes: must now allocate | |
3250 | * a buffer to hold as many roots as may be needed. | |
3251 | * Allocate stable and unstable together: | |
3252 | * MAXSMP NODES_SHIFT 10 will use 16kB. | |
3253 | */ | |
bafe1e14 JP |
3254 | buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), |
3255 | GFP_KERNEL); | |
ef53d16c HD |
3256 | /* Let us assume that RB_ROOT is NULL is zero */ |
3257 | if (!buf) | |
3258 | err = -ENOMEM; | |
3259 | else { | |
3260 | root_stable_tree = buf; | |
3261 | root_unstable_tree = buf + nr_node_ids; | |
3262 | /* Stable tree is empty but not the unstable */ | |
3263 | root_unstable_tree[0] = one_unstable_tree[0]; | |
3264 | } | |
3265 | } | |
3266 | if (!err) { | |
90bd6fd3 | 3267 | ksm_merge_across_nodes = knob; |
ef53d16c HD |
3268 | ksm_nr_node_ids = knob ? 1 : nr_node_ids; |
3269 | } | |
90bd6fd3 PH |
3270 | } |
3271 | mutex_unlock(&ksm_thread_mutex); | |
3272 | ||
3273 | return err ? err : count; | |
3274 | } | |
3275 | KSM_ATTR(merge_across_nodes); | |
3276 | #endif | |
3277 | ||
e86c59b1 | 3278 | static ssize_t use_zero_pages_show(struct kobject *kobj, |
ae7a927d | 3279 | struct kobj_attribute *attr, char *buf) |
e86c59b1 | 3280 | { |
ae7a927d | 3281 | return sysfs_emit(buf, "%u\n", ksm_use_zero_pages); |
e86c59b1 CI |
3282 | } |
3283 | static ssize_t use_zero_pages_store(struct kobject *kobj, | |
3284 | struct kobj_attribute *attr, | |
3285 | const char *buf, size_t count) | |
3286 | { | |
3287 | int err; | |
3288 | bool value; | |
3289 | ||
3290 | err = kstrtobool(buf, &value); | |
3291 | if (err) | |
3292 | return -EINVAL; | |
3293 | ||
3294 | ksm_use_zero_pages = value; | |
3295 | ||
3296 | return count; | |
3297 | } | |
3298 | KSM_ATTR(use_zero_pages); | |
3299 | ||
2c653d0e AA |
3300 | static ssize_t max_page_sharing_show(struct kobject *kobj, |
3301 | struct kobj_attribute *attr, char *buf) | |
3302 | { | |
ae7a927d | 3303 | return sysfs_emit(buf, "%u\n", ksm_max_page_sharing); |
2c653d0e AA |
3304 | } |
3305 | ||
3306 | static ssize_t max_page_sharing_store(struct kobject *kobj, | |
3307 | struct kobj_attribute *attr, | |
3308 | const char *buf, size_t count) | |
3309 | { | |
3310 | int err; | |
3311 | int knob; | |
3312 | ||
3313 | err = kstrtoint(buf, 10, &knob); | |
3314 | if (err) | |
3315 | return err; | |
3316 | /* | |
3317 | * When a KSM page is created it is shared by 2 mappings. This | |
3318 | * being a signed comparison, it implicitly verifies it's not | |
3319 | * negative. | |
3320 | */ | |
3321 | if (knob < 2) | |
3322 | return -EINVAL; | |
3323 | ||
3324 | if (READ_ONCE(ksm_max_page_sharing) == knob) | |
3325 | return count; | |
3326 | ||
3327 | mutex_lock(&ksm_thread_mutex); | |
3328 | wait_while_offlining(); | |
3329 | if (ksm_max_page_sharing != knob) { | |
3330 | if (ksm_pages_shared || remove_all_stable_nodes()) | |
3331 | err = -EBUSY; | |
3332 | else | |
3333 | ksm_max_page_sharing = knob; | |
3334 | } | |
3335 | mutex_unlock(&ksm_thread_mutex); | |
3336 | ||
3337 | return err ? err : count; | |
3338 | } | |
3339 | KSM_ATTR(max_page_sharing); | |
3340 | ||
b348b5fe SR |
3341 | static ssize_t pages_scanned_show(struct kobject *kobj, |
3342 | struct kobj_attribute *attr, char *buf) | |
3343 | { | |
3344 | return sysfs_emit(buf, "%lu\n", ksm_pages_scanned); | |
3345 | } | |
3346 | KSM_ATTR_RO(pages_scanned); | |
3347 | ||
b4028260 HD |
3348 | static ssize_t pages_shared_show(struct kobject *kobj, |
3349 | struct kobj_attribute *attr, char *buf) | |
3350 | { | |
ae7a927d | 3351 | return sysfs_emit(buf, "%lu\n", ksm_pages_shared); |
b4028260 HD |
3352 | } |
3353 | KSM_ATTR_RO(pages_shared); | |
3354 | ||
3355 | static ssize_t pages_sharing_show(struct kobject *kobj, | |
3356 | struct kobj_attribute *attr, char *buf) | |
3357 | { | |
ae7a927d | 3358 | return sysfs_emit(buf, "%lu\n", ksm_pages_sharing); |
b4028260 HD |
3359 | } |
3360 | KSM_ATTR_RO(pages_sharing); | |
3361 | ||
473b0ce4 HD |
3362 | static ssize_t pages_unshared_show(struct kobject *kobj, |
3363 | struct kobj_attribute *attr, char *buf) | |
3364 | { | |
ae7a927d | 3365 | return sysfs_emit(buf, "%lu\n", ksm_pages_unshared); |
473b0ce4 HD |
3366 | } |
3367 | KSM_ATTR_RO(pages_unshared); | |
3368 | ||
3369 | static ssize_t pages_volatile_show(struct kobject *kobj, | |
3370 | struct kobj_attribute *attr, char *buf) | |
3371 | { | |
3372 | long ksm_pages_volatile; | |
3373 | ||
3374 | ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared | |
3375 | - ksm_pages_sharing - ksm_pages_unshared; | |
3376 | /* | |
3377 | * It was not worth any locking to calculate that statistic, | |
3378 | * but it might therefore sometimes be negative: conceal that. | |
3379 | */ | |
3380 | if (ksm_pages_volatile < 0) | |
3381 | ksm_pages_volatile = 0; | |
ae7a927d | 3382 | return sysfs_emit(buf, "%ld\n", ksm_pages_volatile); |
473b0ce4 HD |
3383 | } |
3384 | KSM_ATTR_RO(pages_volatile); | |
3385 | ||
e2942062 | 3386 | static ssize_t ksm_zero_pages_show(struct kobject *kobj, |
3387 | struct kobj_attribute *attr, char *buf) | |
3388 | { | |
3389 | return sysfs_emit(buf, "%ld\n", ksm_zero_pages); | |
3390 | } | |
3391 | KSM_ATTR_RO(ksm_zero_pages); | |
3392 | ||
d21077fb SR |
3393 | static ssize_t general_profit_show(struct kobject *kobj, |
3394 | struct kobj_attribute *attr, char *buf) | |
3395 | { | |
3396 | long general_profit; | |
3397 | ||
1a8e8430 | 3398 | general_profit = (ksm_pages_sharing + ksm_zero_pages) * PAGE_SIZE - |
d21077fb SR |
3399 | ksm_rmap_items * sizeof(struct ksm_rmap_item); |
3400 | ||
3401 | return sysfs_emit(buf, "%ld\n", general_profit); | |
3402 | } | |
3403 | KSM_ATTR_RO(general_profit); | |
3404 | ||
2c653d0e AA |
3405 | static ssize_t stable_node_dups_show(struct kobject *kobj, |
3406 | struct kobj_attribute *attr, char *buf) | |
3407 | { | |
ae7a927d | 3408 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups); |
2c653d0e AA |
3409 | } |
3410 | KSM_ATTR_RO(stable_node_dups); | |
3411 | ||
3412 | static ssize_t stable_node_chains_show(struct kobject *kobj, | |
3413 | struct kobj_attribute *attr, char *buf) | |
3414 | { | |
ae7a927d | 3415 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains); |
2c653d0e AA |
3416 | } |
3417 | KSM_ATTR_RO(stable_node_chains); | |
3418 | ||
3419 | static ssize_t | |
3420 | stable_node_chains_prune_millisecs_show(struct kobject *kobj, | |
3421 | struct kobj_attribute *attr, | |
3422 | char *buf) | |
3423 | { | |
ae7a927d | 3424 | return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs); |
2c653d0e AA |
3425 | } |
3426 | ||
3427 | static ssize_t | |
3428 | stable_node_chains_prune_millisecs_store(struct kobject *kobj, | |
3429 | struct kobj_attribute *attr, | |
3430 | const char *buf, size_t count) | |
3431 | { | |
584ff0df | 3432 | unsigned int msecs; |
2c653d0e AA |
3433 | int err; |
3434 | ||
584ff0df ZB |
3435 | err = kstrtouint(buf, 10, &msecs); |
3436 | if (err) | |
2c653d0e AA |
3437 | return -EINVAL; |
3438 | ||
3439 | ksm_stable_node_chains_prune_millisecs = msecs; | |
3440 | ||
3441 | return count; | |
3442 | } | |
3443 | KSM_ATTR(stable_node_chains_prune_millisecs); | |
3444 | ||
473b0ce4 HD |
3445 | static ssize_t full_scans_show(struct kobject *kobj, |
3446 | struct kobj_attribute *attr, char *buf) | |
3447 | { | |
ae7a927d | 3448 | return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr); |
473b0ce4 HD |
3449 | } |
3450 | KSM_ATTR_RO(full_scans); | |
3451 | ||
31dbd01f IE |
3452 | static struct attribute *ksm_attrs[] = { |
3453 | &sleep_millisecs_attr.attr, | |
3454 | &pages_to_scan_attr.attr, | |
3455 | &run_attr.attr, | |
b348b5fe | 3456 | &pages_scanned_attr.attr, |
b4028260 HD |
3457 | &pages_shared_attr.attr, |
3458 | &pages_sharing_attr.attr, | |
473b0ce4 HD |
3459 | &pages_unshared_attr.attr, |
3460 | &pages_volatile_attr.attr, | |
e2942062 | 3461 | &ksm_zero_pages_attr.attr, |
473b0ce4 | 3462 | &full_scans_attr.attr, |
90bd6fd3 PH |
3463 | #ifdef CONFIG_NUMA |
3464 | &merge_across_nodes_attr.attr, | |
3465 | #endif | |
2c653d0e AA |
3466 | &max_page_sharing_attr.attr, |
3467 | &stable_node_chains_attr.attr, | |
3468 | &stable_node_dups_attr.attr, | |
3469 | &stable_node_chains_prune_millisecs_attr.attr, | |
e86c59b1 | 3470 | &use_zero_pages_attr.attr, |
d21077fb | 3471 | &general_profit_attr.attr, |
31dbd01f IE |
3472 | NULL, |
3473 | }; | |
3474 | ||
f907c26a | 3475 | static const struct attribute_group ksm_attr_group = { |
31dbd01f IE |
3476 | .attrs = ksm_attrs, |
3477 | .name = "ksm", | |
3478 | }; | |
2ffd8679 | 3479 | #endif /* CONFIG_SYSFS */ |
31dbd01f IE |
3480 | |
3481 | static int __init ksm_init(void) | |
3482 | { | |
3483 | struct task_struct *ksm_thread; | |
3484 | int err; | |
3485 | ||
e86c59b1 CI |
3486 | /* The correct value depends on page size and endianness */ |
3487 | zero_checksum = calc_checksum(ZERO_PAGE(0)); | |
3488 | /* Default to false for backwards compatibility */ | |
3489 | ksm_use_zero_pages = false; | |
3490 | ||
31dbd01f IE |
3491 | err = ksm_slab_init(); |
3492 | if (err) | |
3493 | goto out; | |
3494 | ||
31dbd01f IE |
3495 | ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); |
3496 | if (IS_ERR(ksm_thread)) { | |
25acde31 | 3497 | pr_err("ksm: creating kthread failed\n"); |
31dbd01f | 3498 | err = PTR_ERR(ksm_thread); |
d9f8984c | 3499 | goto out_free; |
31dbd01f IE |
3500 | } |
3501 | ||
2ffd8679 | 3502 | #ifdef CONFIG_SYSFS |
31dbd01f IE |
3503 | err = sysfs_create_group(mm_kobj, &ksm_attr_group); |
3504 | if (err) { | |
25acde31 | 3505 | pr_err("ksm: register sysfs failed\n"); |
2ffd8679 | 3506 | kthread_stop(ksm_thread); |
d9f8984c | 3507 | goto out_free; |
31dbd01f | 3508 | } |
c73602ad HD |
3509 | #else |
3510 | ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ | |
3511 | ||
2ffd8679 | 3512 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 3513 | |
62b61f61 | 3514 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 | 3515 | /* There is no significance to this priority 100 */ |
1eeaa4fd | 3516 | hotplug_memory_notifier(ksm_memory_callback, KSM_CALLBACK_PRI); |
62b61f61 | 3517 | #endif |
31dbd01f IE |
3518 | return 0; |
3519 | ||
d9f8984c | 3520 | out_free: |
31dbd01f IE |
3521 | ksm_slab_free(); |
3522 | out: | |
3523 | return err; | |
f8af4da3 | 3524 | } |
a64fb3cd | 3525 | subsys_initcall(ksm_init); |