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 | |
e850dcf5 HD |
48 | #ifdef CONFIG_NUMA |
49 | #define NUMA(x) (x) | |
50 | #define DO_NUMA(x) do { (x); } while (0) | |
51 | #else | |
52 | #define NUMA(x) (0) | |
53 | #define DO_NUMA(x) do { } while (0) | |
54 | #endif | |
55 | ||
5a2ca3ef MR |
56 | /** |
57 | * DOC: Overview | |
58 | * | |
31dbd01f IE |
59 | * A few notes about the KSM scanning process, |
60 | * to make it easier to understand the data structures below: | |
61 | * | |
62 | * In order to reduce excessive scanning, KSM sorts the memory pages by their | |
63 | * contents into a data structure that holds pointers to the pages' locations. | |
64 | * | |
65 | * Since the contents of the pages may change at any moment, KSM cannot just | |
66 | * insert the pages into a normal sorted tree and expect it to find anything. | |
67 | * Therefore KSM uses two data structures - the stable and the unstable tree. | |
68 | * | |
69 | * The stable tree holds pointers to all the merged pages (ksm pages), sorted | |
70 | * by their contents. Because each such page is write-protected, searching on | |
71 | * this tree is fully assured to be working (except when pages are unmapped), | |
72 | * and therefore this tree is called the stable tree. | |
73 | * | |
5a2ca3ef MR |
74 | * The stable tree node includes information required for reverse |
75 | * mapping from a KSM page to virtual addresses that map this page. | |
76 | * | |
77 | * In order to avoid large latencies of the rmap walks on KSM pages, | |
78 | * KSM maintains two types of nodes in the stable tree: | |
79 | * | |
80 | * * the regular nodes that keep the reverse mapping structures in a | |
81 | * linked list | |
82 | * * the "chains" that link nodes ("dups") that represent the same | |
83 | * write protected memory content, but each "dup" corresponds to a | |
84 | * different KSM page copy of that content | |
85 | * | |
86 | * Internally, the regular nodes, "dups" and "chains" are represented | |
21fbd591 | 87 | * using the same struct ksm_stable_node structure. |
5a2ca3ef | 88 | * |
31dbd01f IE |
89 | * In addition to the stable tree, KSM uses a second data structure called the |
90 | * unstable tree: this tree holds pointers to pages which have been found to | |
91 | * be "unchanged for a period of time". The unstable tree sorts these pages | |
92 | * by their contents, but since they are not write-protected, KSM cannot rely | |
93 | * upon the unstable tree to work correctly - the unstable tree is liable to | |
94 | * be corrupted as its contents are modified, and so it is called unstable. | |
95 | * | |
96 | * KSM solves this problem by several techniques: | |
97 | * | |
98 | * 1) The unstable tree is flushed every time KSM completes scanning all | |
99 | * memory areas, and then the tree is rebuilt again from the beginning. | |
100 | * 2) KSM will only insert into the unstable tree, pages whose hash value | |
101 | * has not changed since the previous scan of all memory areas. | |
102 | * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the | |
103 | * colors of the nodes and not on their contents, assuring that even when | |
104 | * the tree gets "corrupted" it won't get out of balance, so scanning time | |
105 | * remains the same (also, searching and inserting nodes in an rbtree uses | |
106 | * the same algorithm, so we have no overhead when we flush and rebuild). | |
107 | * 4) KSM never flushes the stable tree, which means that even if it were to | |
108 | * take 10 attempts to find a page in the unstable tree, once it is found, | |
109 | * it is secured in the stable tree. (When we scan a new page, we first | |
110 | * compare it against the stable tree, and then against the unstable tree.) | |
8fdb3dbf HD |
111 | * |
112 | * If the merge_across_nodes tunable is unset, then KSM maintains multiple | |
113 | * stable trees and multiple unstable trees: one of each for each NUMA node. | |
31dbd01f IE |
114 | */ |
115 | ||
116 | /** | |
21fbd591 | 117 | * struct ksm_mm_slot - ksm information per mm that is being scanned |
58730ab6 | 118 | * @slot: hash lookup from mm to mm_slot |
6514d511 | 119 | * @rmap_list: head for this mm_slot's singly-linked list of rmap_items |
31dbd01f | 120 | */ |
21fbd591 | 121 | struct ksm_mm_slot { |
58730ab6 | 122 | struct mm_slot slot; |
21fbd591 | 123 | struct ksm_rmap_item *rmap_list; |
31dbd01f IE |
124 | }; |
125 | ||
126 | /** | |
127 | * struct ksm_scan - cursor for scanning | |
128 | * @mm_slot: the current mm_slot we are scanning | |
129 | * @address: the next address inside that to be scanned | |
6514d511 | 130 | * @rmap_list: link to the next rmap to be scanned in the rmap_list |
31dbd01f IE |
131 | * @seqnr: count of completed full scans (needed when removing unstable node) |
132 | * | |
133 | * There is only the one ksm_scan instance of this cursor structure. | |
134 | */ | |
135 | struct ksm_scan { | |
21fbd591 | 136 | struct ksm_mm_slot *mm_slot; |
31dbd01f | 137 | unsigned long address; |
21fbd591 | 138 | struct ksm_rmap_item **rmap_list; |
31dbd01f IE |
139 | unsigned long seqnr; |
140 | }; | |
141 | ||
7b6ba2c7 | 142 | /** |
21fbd591 | 143 | * struct ksm_stable_node - node of the stable rbtree |
7b6ba2c7 | 144 | * @node: rb node of this ksm page in the stable tree |
4146d2d6 | 145 | * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list |
2c653d0e | 146 | * @hlist_dup: linked into the stable_node->hlist with a stable_node chain |
4146d2d6 | 147 | * @list: linked into migrate_nodes, pending placement in the proper node tree |
7b6ba2c7 | 148 | * @hlist: hlist head of rmap_items using this ksm page |
4146d2d6 | 149 | * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) |
2c653d0e AA |
150 | * @chain_prune_time: time of the last full garbage collection |
151 | * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN | |
4146d2d6 | 152 | * @nid: NUMA node id of stable tree in which linked (may not match kpfn) |
7b6ba2c7 | 153 | */ |
21fbd591 | 154 | struct ksm_stable_node { |
4146d2d6 HD |
155 | union { |
156 | struct rb_node node; /* when node of stable tree */ | |
157 | struct { /* when listed for migration */ | |
158 | struct list_head *head; | |
2c653d0e AA |
159 | struct { |
160 | struct hlist_node hlist_dup; | |
161 | struct list_head list; | |
162 | }; | |
4146d2d6 HD |
163 | }; |
164 | }; | |
7b6ba2c7 | 165 | struct hlist_head hlist; |
2c653d0e AA |
166 | union { |
167 | unsigned long kpfn; | |
168 | unsigned long chain_prune_time; | |
169 | }; | |
170 | /* | |
171 | * STABLE_NODE_CHAIN can be any negative number in | |
172 | * rmap_hlist_len negative range, but better not -1 to be able | |
173 | * to reliably detect underflows. | |
174 | */ | |
175 | #define STABLE_NODE_CHAIN -1024 | |
176 | int rmap_hlist_len; | |
4146d2d6 HD |
177 | #ifdef CONFIG_NUMA |
178 | int nid; | |
179 | #endif | |
7b6ba2c7 HD |
180 | }; |
181 | ||
31dbd01f | 182 | /** |
21fbd591 | 183 | * struct ksm_rmap_item - reverse mapping item for virtual addresses |
6514d511 | 184 | * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list |
db114b83 | 185 | * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree |
bc56620b | 186 | * @nid: NUMA node id of unstable tree in which linked (may not match page) |
31dbd01f IE |
187 | * @mm: the memory structure this rmap_item is pointing into |
188 | * @address: the virtual address this rmap_item tracks (+ flags in low bits) | |
189 | * @oldchecksum: previous checksum of the page at that virtual address | |
7b6ba2c7 HD |
190 | * @node: rb node of this rmap_item in the unstable tree |
191 | * @head: pointer to stable_node heading this list in the stable tree | |
192 | * @hlist: link into hlist of rmap_items hanging off that stable_node | |
31dbd01f | 193 | */ |
21fbd591 QZ |
194 | struct ksm_rmap_item { |
195 | struct ksm_rmap_item *rmap_list; | |
bc56620b HD |
196 | union { |
197 | struct anon_vma *anon_vma; /* when stable */ | |
198 | #ifdef CONFIG_NUMA | |
199 | int nid; /* when node of unstable tree */ | |
200 | #endif | |
201 | }; | |
31dbd01f IE |
202 | struct mm_struct *mm; |
203 | unsigned long address; /* + low bits used for flags below */ | |
7b6ba2c7 | 204 | unsigned int oldchecksum; /* when unstable */ |
31dbd01f | 205 | union { |
7b6ba2c7 HD |
206 | struct rb_node node; /* when node of unstable tree */ |
207 | struct { /* when listed from stable tree */ | |
21fbd591 | 208 | struct ksm_stable_node *head; |
7b6ba2c7 HD |
209 | struct hlist_node hlist; |
210 | }; | |
31dbd01f IE |
211 | }; |
212 | }; | |
213 | ||
214 | #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ | |
7b6ba2c7 HD |
215 | #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ |
216 | #define STABLE_FLAG 0x200 /* is listed from the stable tree */ | |
31dbd01f IE |
217 | |
218 | /* The stable and unstable tree heads */ | |
ef53d16c HD |
219 | static struct rb_root one_stable_tree[1] = { RB_ROOT }; |
220 | static struct rb_root one_unstable_tree[1] = { RB_ROOT }; | |
221 | static struct rb_root *root_stable_tree = one_stable_tree; | |
222 | static struct rb_root *root_unstable_tree = one_unstable_tree; | |
31dbd01f | 223 | |
4146d2d6 HD |
224 | /* Recently migrated nodes of stable tree, pending proper placement */ |
225 | static LIST_HEAD(migrate_nodes); | |
2c653d0e | 226 | #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev) |
4146d2d6 | 227 | |
4ca3a69b SL |
228 | #define MM_SLOTS_HASH_BITS 10 |
229 | static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
31dbd01f | 230 | |
21fbd591 | 231 | static struct ksm_mm_slot ksm_mm_head = { |
58730ab6 | 232 | .slot.mm_node = LIST_HEAD_INIT(ksm_mm_head.slot.mm_node), |
31dbd01f IE |
233 | }; |
234 | static struct ksm_scan ksm_scan = { | |
235 | .mm_slot = &ksm_mm_head, | |
236 | }; | |
237 | ||
238 | static struct kmem_cache *rmap_item_cache; | |
7b6ba2c7 | 239 | static struct kmem_cache *stable_node_cache; |
31dbd01f IE |
240 | static struct kmem_cache *mm_slot_cache; |
241 | ||
242 | /* The number of nodes in the stable tree */ | |
b4028260 | 243 | static unsigned long ksm_pages_shared; |
31dbd01f | 244 | |
e178dfde | 245 | /* The number of page slots additionally sharing those nodes */ |
b4028260 | 246 | static unsigned long ksm_pages_sharing; |
31dbd01f | 247 | |
473b0ce4 HD |
248 | /* The number of nodes in the unstable tree */ |
249 | static unsigned long ksm_pages_unshared; | |
250 | ||
251 | /* The number of rmap_items in use: to calculate pages_volatile */ | |
252 | static unsigned long ksm_rmap_items; | |
253 | ||
2c653d0e AA |
254 | /* The number of stable_node chains */ |
255 | static unsigned long ksm_stable_node_chains; | |
256 | ||
257 | /* The number of stable_node dups linked to the stable_node chains */ | |
258 | static unsigned long ksm_stable_node_dups; | |
259 | ||
260 | /* Delay in pruning stale stable_node_dups in the stable_node_chains */ | |
584ff0df | 261 | static unsigned int ksm_stable_node_chains_prune_millisecs = 2000; |
2c653d0e AA |
262 | |
263 | /* Maximum number of page slots sharing a stable node */ | |
264 | static int ksm_max_page_sharing = 256; | |
265 | ||
31dbd01f | 266 | /* Number of pages ksmd should scan in one batch */ |
2c6854fd | 267 | static unsigned int ksm_thread_pages_to_scan = 100; |
31dbd01f IE |
268 | |
269 | /* Milliseconds ksmd should sleep between batches */ | |
2ffd8679 | 270 | static unsigned int ksm_thread_sleep_millisecs = 20; |
31dbd01f | 271 | |
e86c59b1 CI |
272 | /* Checksum of an empty (zeroed) page */ |
273 | static unsigned int zero_checksum __read_mostly; | |
274 | ||
275 | /* Whether to merge empty (zeroed) pages with actual zero pages */ | |
276 | static bool ksm_use_zero_pages __read_mostly; | |
277 | ||
e850dcf5 | 278 | #ifdef CONFIG_NUMA |
90bd6fd3 PH |
279 | /* Zeroed when merging across nodes is not allowed */ |
280 | static unsigned int ksm_merge_across_nodes = 1; | |
ef53d16c | 281 | static int ksm_nr_node_ids = 1; |
e850dcf5 HD |
282 | #else |
283 | #define ksm_merge_across_nodes 1U | |
ef53d16c | 284 | #define ksm_nr_node_ids 1 |
e850dcf5 | 285 | #endif |
90bd6fd3 | 286 | |
31dbd01f IE |
287 | #define KSM_RUN_STOP 0 |
288 | #define KSM_RUN_MERGE 1 | |
289 | #define KSM_RUN_UNMERGE 2 | |
ef4d43a8 HD |
290 | #define KSM_RUN_OFFLINE 4 |
291 | static unsigned long ksm_run = KSM_RUN_STOP; | |
292 | static void wait_while_offlining(void); | |
31dbd01f IE |
293 | |
294 | static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); | |
fcf9a0ef | 295 | static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait); |
31dbd01f IE |
296 | static DEFINE_MUTEX(ksm_thread_mutex); |
297 | static DEFINE_SPINLOCK(ksm_mmlist_lock); | |
298 | ||
21fbd591 | 299 | #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create(#__struct,\ |
31dbd01f IE |
300 | sizeof(struct __struct), __alignof__(struct __struct),\ |
301 | (__flags), NULL) | |
302 | ||
303 | static int __init ksm_slab_init(void) | |
304 | { | |
21fbd591 | 305 | rmap_item_cache = KSM_KMEM_CACHE(ksm_rmap_item, 0); |
31dbd01f IE |
306 | if (!rmap_item_cache) |
307 | goto out; | |
308 | ||
21fbd591 | 309 | stable_node_cache = KSM_KMEM_CACHE(ksm_stable_node, 0); |
7b6ba2c7 HD |
310 | if (!stable_node_cache) |
311 | goto out_free1; | |
312 | ||
21fbd591 | 313 | mm_slot_cache = KSM_KMEM_CACHE(ksm_mm_slot, 0); |
31dbd01f | 314 | if (!mm_slot_cache) |
7b6ba2c7 | 315 | goto out_free2; |
31dbd01f IE |
316 | |
317 | return 0; | |
318 | ||
7b6ba2c7 HD |
319 | out_free2: |
320 | kmem_cache_destroy(stable_node_cache); | |
321 | out_free1: | |
31dbd01f IE |
322 | kmem_cache_destroy(rmap_item_cache); |
323 | out: | |
324 | return -ENOMEM; | |
325 | } | |
326 | ||
327 | static void __init ksm_slab_free(void) | |
328 | { | |
329 | kmem_cache_destroy(mm_slot_cache); | |
7b6ba2c7 | 330 | kmem_cache_destroy(stable_node_cache); |
31dbd01f IE |
331 | kmem_cache_destroy(rmap_item_cache); |
332 | mm_slot_cache = NULL; | |
333 | } | |
334 | ||
21fbd591 | 335 | static __always_inline bool is_stable_node_chain(struct ksm_stable_node *chain) |
2c653d0e AA |
336 | { |
337 | return chain->rmap_hlist_len == STABLE_NODE_CHAIN; | |
338 | } | |
339 | ||
21fbd591 | 340 | static __always_inline bool is_stable_node_dup(struct ksm_stable_node *dup) |
2c653d0e AA |
341 | { |
342 | return dup->head == STABLE_NODE_DUP_HEAD; | |
343 | } | |
344 | ||
21fbd591 QZ |
345 | static inline void stable_node_chain_add_dup(struct ksm_stable_node *dup, |
346 | struct ksm_stable_node *chain) | |
2c653d0e AA |
347 | { |
348 | VM_BUG_ON(is_stable_node_dup(dup)); | |
349 | dup->head = STABLE_NODE_DUP_HEAD; | |
350 | VM_BUG_ON(!is_stable_node_chain(chain)); | |
351 | hlist_add_head(&dup->hlist_dup, &chain->hlist); | |
352 | ksm_stable_node_dups++; | |
353 | } | |
354 | ||
21fbd591 | 355 | static inline void __stable_node_dup_del(struct ksm_stable_node *dup) |
2c653d0e | 356 | { |
b4fecc67 | 357 | VM_BUG_ON(!is_stable_node_dup(dup)); |
2c653d0e AA |
358 | hlist_del(&dup->hlist_dup); |
359 | ksm_stable_node_dups--; | |
360 | } | |
361 | ||
21fbd591 | 362 | static inline void stable_node_dup_del(struct ksm_stable_node *dup) |
2c653d0e AA |
363 | { |
364 | VM_BUG_ON(is_stable_node_chain(dup)); | |
365 | if (is_stable_node_dup(dup)) | |
366 | __stable_node_dup_del(dup); | |
367 | else | |
368 | rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid)); | |
369 | #ifdef CONFIG_DEBUG_VM | |
370 | dup->head = NULL; | |
371 | #endif | |
372 | } | |
373 | ||
21fbd591 | 374 | static inline struct ksm_rmap_item *alloc_rmap_item(void) |
31dbd01f | 375 | { |
21fbd591 | 376 | struct ksm_rmap_item *rmap_item; |
473b0ce4 | 377 | |
5b398e41 | 378 | rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | |
379 | __GFP_NORETRY | __GFP_NOWARN); | |
473b0ce4 HD |
380 | if (rmap_item) |
381 | ksm_rmap_items++; | |
382 | return rmap_item; | |
31dbd01f IE |
383 | } |
384 | ||
21fbd591 | 385 | static inline void free_rmap_item(struct ksm_rmap_item *rmap_item) |
31dbd01f | 386 | { |
473b0ce4 | 387 | ksm_rmap_items--; |
cb4df4ca | 388 | rmap_item->mm->ksm_rmap_items--; |
31dbd01f IE |
389 | rmap_item->mm = NULL; /* debug safety */ |
390 | kmem_cache_free(rmap_item_cache, rmap_item); | |
391 | } | |
392 | ||
21fbd591 | 393 | static inline struct ksm_stable_node *alloc_stable_node(void) |
7b6ba2c7 | 394 | { |
6213055f | 395 | /* |
396 | * The allocation can take too long with GFP_KERNEL when memory is under | |
397 | * pressure, which may lead to hung task warnings. Adding __GFP_HIGH | |
398 | * grants access to memory reserves, helping to avoid this problem. | |
399 | */ | |
400 | return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); | |
7b6ba2c7 HD |
401 | } |
402 | ||
21fbd591 | 403 | static inline void free_stable_node(struct ksm_stable_node *stable_node) |
7b6ba2c7 | 404 | { |
2c653d0e AA |
405 | VM_BUG_ON(stable_node->rmap_hlist_len && |
406 | !is_stable_node_chain(stable_node)); | |
7b6ba2c7 HD |
407 | kmem_cache_free(stable_node_cache, stable_node); |
408 | } | |
409 | ||
a913e182 HD |
410 | /* |
411 | * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's | |
412 | * page tables after it has passed through ksm_exit() - which, if necessary, | |
c1e8d7c6 | 413 | * takes mmap_lock briefly to serialize against them. ksm_exit() does not set |
a913e182 HD |
414 | * a special flag: they can just back out as soon as mm_users goes to zero. |
415 | * ksm_test_exit() is used throughout to make this test for exit: in some | |
416 | * places for correctness, in some places just to avoid unnecessary work. | |
417 | */ | |
418 | static inline bool ksm_test_exit(struct mm_struct *mm) | |
419 | { | |
420 | return atomic_read(&mm->mm_users) == 0; | |
421 | } | |
422 | ||
d7c0e68d DH |
423 | static int break_ksm_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long next, |
424 | struct mm_walk *walk) | |
425 | { | |
426 | struct page *page = NULL; | |
427 | spinlock_t *ptl; | |
428 | pte_t *pte; | |
429 | int ret; | |
430 | ||
431 | if (pmd_leaf(*pmd) || !pmd_present(*pmd)) | |
432 | return 0; | |
433 | ||
434 | pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); | |
435 | if (pte_present(*pte)) { | |
436 | page = vm_normal_page(walk->vma, addr, *pte); | |
437 | } else if (!pte_none(*pte)) { | |
438 | swp_entry_t entry = pte_to_swp_entry(*pte); | |
439 | ||
440 | /* | |
441 | * As KSM pages remain KSM pages until freed, no need to wait | |
442 | * here for migration to end. | |
443 | */ | |
444 | if (is_migration_entry(entry)) | |
445 | page = pfn_swap_entry_to_page(entry); | |
446 | } | |
447 | ret = page && PageKsm(page); | |
448 | pte_unmap_unlock(pte, ptl); | |
449 | return ret; | |
450 | } | |
451 | ||
452 | static const struct mm_walk_ops break_ksm_ops = { | |
453 | .pmd_entry = break_ksm_pmd_entry, | |
454 | }; | |
455 | ||
31dbd01f | 456 | /* |
6cce3314 DH |
457 | * We use break_ksm to break COW on a ksm page by triggering unsharing, |
458 | * such that the ksm page will get replaced by an exclusive anonymous page. | |
31dbd01f | 459 | * |
6cce3314 | 460 | * We take great care only to touch a ksm page, in a VM_MERGEABLE vma, |
31dbd01f IE |
461 | * in case the application has unmapped and remapped mm,addr meanwhile. |
462 | * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP | |
bbcd53c9 | 463 | * mmap of /dev/mem, where we would not want to touch it. |
1b2ee126 | 464 | * |
6cce3314 | 465 | * FAULT_FLAG_REMOTE/FOLL_REMOTE are because we do this outside the context |
1b2ee126 DH |
466 | * of the process that owns 'vma'. We also do not want to enforce |
467 | * protection keys here anyway. | |
31dbd01f | 468 | */ |
d952b791 | 469 | static int break_ksm(struct vm_area_struct *vma, unsigned long addr) |
31dbd01f | 470 | { |
50a7ca3c | 471 | vm_fault_t ret = 0; |
31dbd01f IE |
472 | |
473 | do { | |
d7c0e68d | 474 | int ksm_page; |
58f595c6 | 475 | |
31dbd01f | 476 | cond_resched(); |
d7c0e68d DH |
477 | ksm_page = walk_page_range_vma(vma, addr, addr + 1, |
478 | &break_ksm_ops, NULL); | |
479 | if (WARN_ON_ONCE(ksm_page < 0)) | |
480 | return ksm_page; | |
58f595c6 DH |
481 | if (!ksm_page) |
482 | return 0; | |
483 | ret = handle_mm_fault(vma, addr, | |
6cce3314 | 484 | FAULT_FLAG_UNSHARE | FAULT_FLAG_REMOTE, |
58f595c6 DH |
485 | NULL); |
486 | } while (!(ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); | |
d952b791 | 487 | /* |
58f595c6 DH |
488 | * We must loop until we no longer find a KSM page because |
489 | * handle_mm_fault() may back out if there's any difficulty e.g. if | |
490 | * pte accessed bit gets updated concurrently. | |
d952b791 HD |
491 | * |
492 | * VM_FAULT_SIGBUS could occur if we race with truncation of the | |
493 | * backing file, which also invalidates anonymous pages: that's | |
494 | * okay, that truncation will have unmapped the PageKsm for us. | |
495 | * | |
496 | * VM_FAULT_OOM: at the time of writing (late July 2009), setting | |
497 | * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the | |
498 | * current task has TIF_MEMDIE set, and will be OOM killed on return | |
499 | * to user; and ksmd, having no mm, would never be chosen for that. | |
500 | * | |
501 | * But if the mm is in a limited mem_cgroup, then the fault may fail | |
502 | * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and | |
503 | * even ksmd can fail in this way - though it's usually breaking ksm | |
504 | * just to undo a merge it made a moment before, so unlikely to oom. | |
505 | * | |
506 | * That's a pity: we might therefore have more kernel pages allocated | |
507 | * than we're counting as nodes in the stable tree; but ksm_do_scan | |
508 | * will retry to break_cow on each pass, so should recover the page | |
509 | * in due course. The important thing is to not let VM_MERGEABLE | |
510 | * be cleared while any such pages might remain in the area. | |
511 | */ | |
512 | return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; | |
31dbd01f IE |
513 | } |
514 | ||
ef694222 BL |
515 | static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, |
516 | unsigned long addr) | |
517 | { | |
518 | struct vm_area_struct *vma; | |
519 | if (ksm_test_exit(mm)) | |
520 | return NULL; | |
ff69fb81 LH |
521 | vma = vma_lookup(mm, addr); |
522 | if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) | |
ef694222 BL |
523 | return NULL; |
524 | return vma; | |
525 | } | |
526 | ||
21fbd591 | 527 | static void break_cow(struct ksm_rmap_item *rmap_item) |
31dbd01f | 528 | { |
8dd3557a HD |
529 | struct mm_struct *mm = rmap_item->mm; |
530 | unsigned long addr = rmap_item->address; | |
31dbd01f IE |
531 | struct vm_area_struct *vma; |
532 | ||
4035c07a HD |
533 | /* |
534 | * It is not an accident that whenever we want to break COW | |
535 | * to undo, we also need to drop a reference to the anon_vma. | |
536 | */ | |
9e60109f | 537 | put_anon_vma(rmap_item->anon_vma); |
4035c07a | 538 | |
d8ed45c5 | 539 | mmap_read_lock(mm); |
ef694222 BL |
540 | vma = find_mergeable_vma(mm, addr); |
541 | if (vma) | |
542 | break_ksm(vma, addr); | |
d8ed45c5 | 543 | mmap_read_unlock(mm); |
31dbd01f IE |
544 | } |
545 | ||
21fbd591 | 546 | static struct page *get_mergeable_page(struct ksm_rmap_item *rmap_item) |
31dbd01f IE |
547 | { |
548 | struct mm_struct *mm = rmap_item->mm; | |
549 | unsigned long addr = rmap_item->address; | |
550 | struct vm_area_struct *vma; | |
551 | struct page *page; | |
552 | ||
d8ed45c5 | 553 | mmap_read_lock(mm); |
ef694222 BL |
554 | vma = find_mergeable_vma(mm, addr); |
555 | if (!vma) | |
31dbd01f IE |
556 | goto out; |
557 | ||
558 | page = follow_page(vma, addr, FOLL_GET); | |
f7091ed6 | 559 | if (IS_ERR_OR_NULL(page)) |
31dbd01f | 560 | goto out; |
f7091ed6 HW |
561 | if (is_zone_device_page(page)) |
562 | goto out_putpage; | |
f765f540 | 563 | if (PageAnon(page)) { |
31dbd01f IE |
564 | flush_anon_page(vma, page, addr); |
565 | flush_dcache_page(page); | |
566 | } else { | |
f7091ed6 | 567 | out_putpage: |
31dbd01f | 568 | put_page(page); |
c8f95ed1 AA |
569 | out: |
570 | page = NULL; | |
31dbd01f | 571 | } |
d8ed45c5 | 572 | mmap_read_unlock(mm); |
31dbd01f IE |
573 | return page; |
574 | } | |
575 | ||
90bd6fd3 PH |
576 | /* |
577 | * This helper is used for getting right index into array of tree roots. | |
578 | * When merge_across_nodes knob is set to 1, there are only two rb-trees for | |
579 | * stable and unstable pages from all nodes with roots in index 0. Otherwise, | |
580 | * every node has its own stable and unstable tree. | |
581 | */ | |
582 | static inline int get_kpfn_nid(unsigned long kpfn) | |
583 | { | |
d8fc16a8 | 584 | return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); |
90bd6fd3 PH |
585 | } |
586 | ||
21fbd591 | 587 | static struct ksm_stable_node *alloc_stable_node_chain(struct ksm_stable_node *dup, |
2c653d0e AA |
588 | struct rb_root *root) |
589 | { | |
21fbd591 | 590 | struct ksm_stable_node *chain = alloc_stable_node(); |
2c653d0e AA |
591 | VM_BUG_ON(is_stable_node_chain(dup)); |
592 | if (likely(chain)) { | |
593 | INIT_HLIST_HEAD(&chain->hlist); | |
594 | chain->chain_prune_time = jiffies; | |
595 | chain->rmap_hlist_len = STABLE_NODE_CHAIN; | |
596 | #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA) | |
98fa15f3 | 597 | chain->nid = NUMA_NO_NODE; /* debug */ |
2c653d0e AA |
598 | #endif |
599 | ksm_stable_node_chains++; | |
600 | ||
601 | /* | |
602 | * Put the stable node chain in the first dimension of | |
603 | * the stable tree and at the same time remove the old | |
604 | * stable node. | |
605 | */ | |
606 | rb_replace_node(&dup->node, &chain->node, root); | |
607 | ||
608 | /* | |
609 | * Move the old stable node to the second dimension | |
610 | * queued in the hlist_dup. The invariant is that all | |
611 | * dup stable_nodes in the chain->hlist point to pages | |
457aef94 | 612 | * that are write protected and have the exact same |
2c653d0e AA |
613 | * content. |
614 | */ | |
615 | stable_node_chain_add_dup(dup, chain); | |
616 | } | |
617 | return chain; | |
618 | } | |
619 | ||
21fbd591 | 620 | static inline void free_stable_node_chain(struct ksm_stable_node *chain, |
2c653d0e AA |
621 | struct rb_root *root) |
622 | { | |
623 | rb_erase(&chain->node, root); | |
624 | free_stable_node(chain); | |
625 | ksm_stable_node_chains--; | |
626 | } | |
627 | ||
21fbd591 | 628 | static void remove_node_from_stable_tree(struct ksm_stable_node *stable_node) |
4035c07a | 629 | { |
21fbd591 | 630 | struct ksm_rmap_item *rmap_item; |
4035c07a | 631 | |
2c653d0e AA |
632 | /* check it's not STABLE_NODE_CHAIN or negative */ |
633 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
634 | ||
b67bfe0d | 635 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
4035c07a HD |
636 | if (rmap_item->hlist.next) |
637 | ksm_pages_sharing--; | |
638 | else | |
639 | ksm_pages_shared--; | |
76093853 | 640 | |
641 | rmap_item->mm->ksm_merging_pages--; | |
642 | ||
2c653d0e AA |
643 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
644 | stable_node->rmap_hlist_len--; | |
9e60109f | 645 | put_anon_vma(rmap_item->anon_vma); |
4035c07a HD |
646 | rmap_item->address &= PAGE_MASK; |
647 | cond_resched(); | |
648 | } | |
649 | ||
2c653d0e AA |
650 | /* |
651 | * We need the second aligned pointer of the migrate_nodes | |
652 | * list_head to stay clear from the rb_parent_color union | |
653 | * (aligned and different than any node) and also different | |
654 | * from &migrate_nodes. This will verify that future list.h changes | |
815f0ddb | 655 | * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it. |
2c653d0e | 656 | */ |
2c653d0e AA |
657 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes); |
658 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1); | |
2c653d0e | 659 | |
4146d2d6 HD |
660 | if (stable_node->head == &migrate_nodes) |
661 | list_del(&stable_node->list); | |
662 | else | |
2c653d0e | 663 | stable_node_dup_del(stable_node); |
4035c07a HD |
664 | free_stable_node(stable_node); |
665 | } | |
666 | ||
2cee57d1 YS |
667 | enum get_ksm_page_flags { |
668 | GET_KSM_PAGE_NOLOCK, | |
669 | GET_KSM_PAGE_LOCK, | |
670 | GET_KSM_PAGE_TRYLOCK | |
671 | }; | |
672 | ||
4035c07a HD |
673 | /* |
674 | * get_ksm_page: checks if the page indicated by the stable node | |
675 | * is still its ksm page, despite having held no reference to it. | |
676 | * In which case we can trust the content of the page, and it | |
677 | * returns the gotten page; but if the page has now been zapped, | |
678 | * remove the stale node from the stable tree and return NULL. | |
c8d6553b | 679 | * But beware, the stable node's page might be being migrated. |
4035c07a HD |
680 | * |
681 | * You would expect the stable_node to hold a reference to the ksm page. | |
682 | * But if it increments the page's count, swapping out has to wait for | |
683 | * ksmd to come around again before it can free the page, which may take | |
684 | * seconds or even minutes: much too unresponsive. So instead we use a | |
685 | * "keyhole reference": access to the ksm page from the stable node peeps | |
686 | * out through its keyhole to see if that page still holds the right key, | |
687 | * pointing back to this stable node. This relies on freeing a PageAnon | |
688 | * page to reset its page->mapping to NULL, and relies on no other use of | |
689 | * a page to put something that might look like our key in page->mapping. | |
4035c07a HD |
690 | * is on its way to being freed; but it is an anomaly to bear in mind. |
691 | */ | |
21fbd591 | 692 | static struct page *get_ksm_page(struct ksm_stable_node *stable_node, |
2cee57d1 | 693 | enum get_ksm_page_flags flags) |
4035c07a HD |
694 | { |
695 | struct page *page; | |
696 | void *expected_mapping; | |
c8d6553b | 697 | unsigned long kpfn; |
4035c07a | 698 | |
bda807d4 MK |
699 | expected_mapping = (void *)((unsigned long)stable_node | |
700 | PAGE_MAPPING_KSM); | |
c8d6553b | 701 | again: |
08df4774 | 702 | kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */ |
c8d6553b | 703 | page = pfn_to_page(kpfn); |
4db0c3c2 | 704 | if (READ_ONCE(page->mapping) != expected_mapping) |
4035c07a | 705 | goto stale; |
c8d6553b HD |
706 | |
707 | /* | |
708 | * We cannot do anything with the page while its refcount is 0. | |
709 | * Usually 0 means free, or tail of a higher-order page: in which | |
710 | * case this node is no longer referenced, and should be freed; | |
1c4c3b99 | 711 | * however, it might mean that the page is under page_ref_freeze(). |
c8d6553b | 712 | * The __remove_mapping() case is easy, again the node is now stale; |
52d1e606 | 713 | * the same is in reuse_ksm_page() case; but if page is swapcache |
9800562f | 714 | * in folio_migrate_mapping(), it might still be our page, |
52d1e606 | 715 | * in which case it's essential to keep the node. |
c8d6553b HD |
716 | */ |
717 | while (!get_page_unless_zero(page)) { | |
718 | /* | |
719 | * Another check for page->mapping != expected_mapping would | |
720 | * work here too. We have chosen the !PageSwapCache test to | |
721 | * optimize the common case, when the page is or is about to | |
722 | * be freed: PageSwapCache is cleared (under spin_lock_irq) | |
1c4c3b99 | 723 | * in the ref_freeze section of __remove_mapping(); but Anon |
c8d6553b HD |
724 | * page->mapping reset to NULL later, in free_pages_prepare(). |
725 | */ | |
726 | if (!PageSwapCache(page)) | |
727 | goto stale; | |
728 | cpu_relax(); | |
729 | } | |
730 | ||
4db0c3c2 | 731 | if (READ_ONCE(page->mapping) != expected_mapping) { |
4035c07a HD |
732 | put_page(page); |
733 | goto stale; | |
734 | } | |
c8d6553b | 735 | |
2cee57d1 YS |
736 | if (flags == GET_KSM_PAGE_TRYLOCK) { |
737 | if (!trylock_page(page)) { | |
738 | put_page(page); | |
739 | return ERR_PTR(-EBUSY); | |
740 | } | |
741 | } else if (flags == GET_KSM_PAGE_LOCK) | |
8aafa6a4 | 742 | lock_page(page); |
2cee57d1 YS |
743 | |
744 | if (flags != GET_KSM_PAGE_NOLOCK) { | |
4db0c3c2 | 745 | if (READ_ONCE(page->mapping) != expected_mapping) { |
8aafa6a4 HD |
746 | unlock_page(page); |
747 | put_page(page); | |
748 | goto stale; | |
749 | } | |
750 | } | |
4035c07a | 751 | return page; |
c8d6553b | 752 | |
4035c07a | 753 | stale: |
c8d6553b HD |
754 | /* |
755 | * We come here from above when page->mapping or !PageSwapCache | |
756 | * suggests that the node is stale; but it might be under migration. | |
19138349 | 757 | * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(), |
c8d6553b HD |
758 | * before checking whether node->kpfn has been changed. |
759 | */ | |
760 | smp_rmb(); | |
4db0c3c2 | 761 | if (READ_ONCE(stable_node->kpfn) != kpfn) |
c8d6553b | 762 | goto again; |
4035c07a HD |
763 | remove_node_from_stable_tree(stable_node); |
764 | return NULL; | |
765 | } | |
766 | ||
31dbd01f IE |
767 | /* |
768 | * Removing rmap_item from stable or unstable tree. | |
769 | * This function will clean the information from the stable/unstable tree. | |
770 | */ | |
21fbd591 | 771 | static void remove_rmap_item_from_tree(struct ksm_rmap_item *rmap_item) |
31dbd01f | 772 | { |
7b6ba2c7 | 773 | if (rmap_item->address & STABLE_FLAG) { |
21fbd591 | 774 | struct ksm_stable_node *stable_node; |
5ad64688 | 775 | struct page *page; |
31dbd01f | 776 | |
7b6ba2c7 | 777 | stable_node = rmap_item->head; |
62862290 | 778 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
4035c07a HD |
779 | if (!page) |
780 | goto out; | |
5ad64688 | 781 | |
7b6ba2c7 | 782 | hlist_del(&rmap_item->hlist); |
62862290 | 783 | unlock_page(page); |
4035c07a | 784 | put_page(page); |
08beca44 | 785 | |
98666f8a | 786 | if (!hlist_empty(&stable_node->hlist)) |
4035c07a HD |
787 | ksm_pages_sharing--; |
788 | else | |
7b6ba2c7 | 789 | ksm_pages_shared--; |
76093853 | 790 | |
791 | rmap_item->mm->ksm_merging_pages--; | |
792 | ||
2c653d0e AA |
793 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
794 | stable_node->rmap_hlist_len--; | |
31dbd01f | 795 | |
9e60109f | 796 | put_anon_vma(rmap_item->anon_vma); |
c89a384e | 797 | rmap_item->head = NULL; |
93d17715 | 798 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 799 | |
7b6ba2c7 | 800 | } else if (rmap_item->address & UNSTABLE_FLAG) { |
31dbd01f IE |
801 | unsigned char age; |
802 | /* | |
9ba69294 | 803 | * Usually ksmd can and must skip the rb_erase, because |
31dbd01f | 804 | * root_unstable_tree was already reset to RB_ROOT. |
9ba69294 HD |
805 | * But be careful when an mm is exiting: do the rb_erase |
806 | * if this rmap_item was inserted by this scan, rather | |
807 | * than left over from before. | |
31dbd01f IE |
808 | */ |
809 | age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); | |
cd551f97 | 810 | BUG_ON(age > 1); |
31dbd01f | 811 | if (!age) |
90bd6fd3 | 812 | rb_erase(&rmap_item->node, |
ef53d16c | 813 | root_unstable_tree + NUMA(rmap_item->nid)); |
473b0ce4 | 814 | ksm_pages_unshared--; |
93d17715 | 815 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 816 | } |
4035c07a | 817 | out: |
31dbd01f IE |
818 | cond_resched(); /* we're called from many long loops */ |
819 | } | |
820 | ||
21fbd591 | 821 | static void remove_trailing_rmap_items(struct ksm_rmap_item **rmap_list) |
31dbd01f | 822 | { |
6514d511 | 823 | while (*rmap_list) { |
21fbd591 | 824 | struct ksm_rmap_item *rmap_item = *rmap_list; |
6514d511 | 825 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 826 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
827 | free_rmap_item(rmap_item); |
828 | } | |
829 | } | |
830 | ||
831 | /* | |
e850dcf5 | 832 | * Though it's very tempting to unmerge rmap_items from stable tree rather |
31dbd01f IE |
833 | * than check every pte of a given vma, the locking doesn't quite work for |
834 | * that - an rmap_item is assigned to the stable tree after inserting ksm | |
c1e8d7c6 | 835 | * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing |
31dbd01f IE |
836 | * rmap_items from parent to child at fork time (so as not to waste time |
837 | * if exit comes before the next scan reaches it). | |
81464e30 HD |
838 | * |
839 | * Similarly, although we'd like to remove rmap_items (so updating counts | |
840 | * and freeing memory) when unmerging an area, it's easier to leave that | |
841 | * to the next pass of ksmd - consider, for example, how ksmd might be | |
842 | * in cmp_and_merge_page on one of the rmap_items we would be removing. | |
31dbd01f | 843 | */ |
d952b791 HD |
844 | static int unmerge_ksm_pages(struct vm_area_struct *vma, |
845 | unsigned long start, unsigned long end) | |
31dbd01f IE |
846 | { |
847 | unsigned long addr; | |
d952b791 | 848 | int err = 0; |
31dbd01f | 849 | |
d952b791 | 850 | for (addr = start; addr < end && !err; addr += PAGE_SIZE) { |
9ba69294 HD |
851 | if (ksm_test_exit(vma->vm_mm)) |
852 | break; | |
d952b791 HD |
853 | if (signal_pending(current)) |
854 | err = -ERESTARTSYS; | |
855 | else | |
856 | err = break_ksm(vma, addr); | |
857 | } | |
858 | return err; | |
31dbd01f IE |
859 | } |
860 | ||
21fbd591 | 861 | static inline struct ksm_stable_node *folio_stable_node(struct folio *folio) |
19138349 MWO |
862 | { |
863 | return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL; | |
864 | } | |
865 | ||
21fbd591 | 866 | static inline struct ksm_stable_node *page_stable_node(struct page *page) |
88484826 | 867 | { |
19138349 | 868 | return folio_stable_node(page_folio(page)); |
88484826 MR |
869 | } |
870 | ||
871 | static inline void set_page_stable_node(struct page *page, | |
21fbd591 | 872 | struct ksm_stable_node *stable_node) |
88484826 | 873 | { |
6c287605 | 874 | VM_BUG_ON_PAGE(PageAnon(page) && PageAnonExclusive(page), page); |
88484826 MR |
875 | page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM); |
876 | } | |
877 | ||
2ffd8679 HD |
878 | #ifdef CONFIG_SYSFS |
879 | /* | |
880 | * Only called through the sysfs control interface: | |
881 | */ | |
21fbd591 | 882 | static int remove_stable_node(struct ksm_stable_node *stable_node) |
cbf86cfe HD |
883 | { |
884 | struct page *page; | |
885 | int err; | |
886 | ||
2cee57d1 | 887 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
cbf86cfe HD |
888 | if (!page) { |
889 | /* | |
890 | * get_ksm_page did remove_node_from_stable_tree itself. | |
891 | */ | |
892 | return 0; | |
893 | } | |
894 | ||
9a63236f AR |
895 | /* |
896 | * Page could be still mapped if this races with __mmput() running in | |
897 | * between ksm_exit() and exit_mmap(). Just refuse to let | |
898 | * merge_across_nodes/max_page_sharing be switched. | |
899 | */ | |
900 | err = -EBUSY; | |
901 | if (!page_mapped(page)) { | |
cbf86cfe | 902 | /* |
8fdb3dbf HD |
903 | * The stable node did not yet appear stale to get_ksm_page(), |
904 | * since that allows for an unmapped ksm page to be recognized | |
905 | * right up until it is freed; but the node is safe to remove. | |
cbf86cfe HD |
906 | * This page might be in a pagevec waiting to be freed, |
907 | * or it might be PageSwapCache (perhaps under writeback), | |
908 | * or it might have been removed from swapcache a moment ago. | |
909 | */ | |
910 | set_page_stable_node(page, NULL); | |
911 | remove_node_from_stable_tree(stable_node); | |
912 | err = 0; | |
913 | } | |
914 | ||
915 | unlock_page(page); | |
916 | put_page(page); | |
917 | return err; | |
918 | } | |
919 | ||
21fbd591 | 920 | static int remove_stable_node_chain(struct ksm_stable_node *stable_node, |
2c653d0e AA |
921 | struct rb_root *root) |
922 | { | |
21fbd591 | 923 | struct ksm_stable_node *dup; |
2c653d0e AA |
924 | struct hlist_node *hlist_safe; |
925 | ||
926 | if (!is_stable_node_chain(stable_node)) { | |
927 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
928 | if (remove_stable_node(stable_node)) | |
929 | return true; | |
930 | else | |
931 | return false; | |
932 | } | |
933 | ||
934 | hlist_for_each_entry_safe(dup, hlist_safe, | |
935 | &stable_node->hlist, hlist_dup) { | |
936 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
937 | if (remove_stable_node(dup)) | |
938 | return true; | |
939 | } | |
940 | BUG_ON(!hlist_empty(&stable_node->hlist)); | |
941 | free_stable_node_chain(stable_node, root); | |
942 | return false; | |
943 | } | |
944 | ||
cbf86cfe HD |
945 | static int remove_all_stable_nodes(void) |
946 | { | |
21fbd591 | 947 | struct ksm_stable_node *stable_node, *next; |
cbf86cfe HD |
948 | int nid; |
949 | int err = 0; | |
950 | ||
ef53d16c | 951 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
cbf86cfe HD |
952 | while (root_stable_tree[nid].rb_node) { |
953 | stable_node = rb_entry(root_stable_tree[nid].rb_node, | |
21fbd591 | 954 | struct ksm_stable_node, node); |
2c653d0e AA |
955 | if (remove_stable_node_chain(stable_node, |
956 | root_stable_tree + nid)) { | |
cbf86cfe HD |
957 | err = -EBUSY; |
958 | break; /* proceed to next nid */ | |
959 | } | |
960 | cond_resched(); | |
961 | } | |
962 | } | |
03640418 | 963 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
964 | if (remove_stable_node(stable_node)) |
965 | err = -EBUSY; | |
966 | cond_resched(); | |
967 | } | |
cbf86cfe HD |
968 | return err; |
969 | } | |
970 | ||
d952b791 | 971 | static int unmerge_and_remove_all_rmap_items(void) |
31dbd01f | 972 | { |
21fbd591 | 973 | struct ksm_mm_slot *mm_slot; |
58730ab6 | 974 | struct mm_slot *slot; |
31dbd01f IE |
975 | struct mm_struct *mm; |
976 | struct vm_area_struct *vma; | |
d952b791 HD |
977 | int err = 0; |
978 | ||
979 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
980 | slot = list_entry(ksm_mm_head.slot.mm_node.next, |
981 | struct mm_slot, mm_node); | |
982 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
d952b791 | 983 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f | 984 | |
a5f18ba0 MWO |
985 | for (mm_slot = ksm_scan.mm_slot; mm_slot != &ksm_mm_head; |
986 | mm_slot = ksm_scan.mm_slot) { | |
58730ab6 | 987 | VMA_ITERATOR(vmi, mm_slot->slot.mm, 0); |
a5f18ba0 | 988 | |
58730ab6 | 989 | mm = mm_slot->slot.mm; |
d8ed45c5 | 990 | mmap_read_lock(mm); |
a5f18ba0 | 991 | for_each_vma(vmi, vma) { |
9ba69294 HD |
992 | if (ksm_test_exit(mm)) |
993 | break; | |
31dbd01f IE |
994 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
995 | continue; | |
d952b791 HD |
996 | err = unmerge_ksm_pages(vma, |
997 | vma->vm_start, vma->vm_end); | |
9ba69294 HD |
998 | if (err) |
999 | goto error; | |
31dbd01f | 1000 | } |
9ba69294 | 1001 | |
420be4ed | 1002 | remove_trailing_rmap_items(&mm_slot->rmap_list); |
d8ed45c5 | 1003 | mmap_read_unlock(mm); |
d952b791 HD |
1004 | |
1005 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
1006 | slot = list_entry(mm_slot->slot.mm_node.next, |
1007 | struct mm_slot, mm_node); | |
1008 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
9ba69294 | 1009 | if (ksm_test_exit(mm)) { |
58730ab6 QZ |
1010 | hash_del(&mm_slot->slot.hash); |
1011 | list_del(&mm_slot->slot.mm_node); | |
9ba69294 HD |
1012 | spin_unlock(&ksm_mmlist_lock); |
1013 | ||
58730ab6 | 1014 | mm_slot_free(mm_slot_cache, mm_slot); |
9ba69294 | 1015 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
9ba69294 | 1016 | mmdrop(mm); |
7496fea9 | 1017 | } else |
9ba69294 | 1018 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f IE |
1019 | } |
1020 | ||
cbf86cfe HD |
1021 | /* Clean up stable nodes, but don't worry if some are still busy */ |
1022 | remove_all_stable_nodes(); | |
d952b791 | 1023 | ksm_scan.seqnr = 0; |
9ba69294 HD |
1024 | return 0; |
1025 | ||
1026 | error: | |
d8ed45c5 | 1027 | mmap_read_unlock(mm); |
31dbd01f | 1028 | spin_lock(&ksm_mmlist_lock); |
d952b791 | 1029 | ksm_scan.mm_slot = &ksm_mm_head; |
31dbd01f | 1030 | spin_unlock(&ksm_mmlist_lock); |
d952b791 | 1031 | return err; |
31dbd01f | 1032 | } |
2ffd8679 | 1033 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 1034 | |
31dbd01f IE |
1035 | static u32 calc_checksum(struct page *page) |
1036 | { | |
1037 | u32 checksum; | |
9b04c5fe | 1038 | void *addr = kmap_atomic(page); |
59e1a2f4 | 1039 | checksum = xxhash(addr, PAGE_SIZE, 0); |
9b04c5fe | 1040 | kunmap_atomic(addr); |
31dbd01f IE |
1041 | return checksum; |
1042 | } | |
1043 | ||
31dbd01f IE |
1044 | static int write_protect_page(struct vm_area_struct *vma, struct page *page, |
1045 | pte_t *orig_pte) | |
1046 | { | |
1047 | struct mm_struct *mm = vma->vm_mm; | |
eed05e54 | 1048 | DEFINE_PAGE_VMA_WALK(pvmw, page, vma, 0, 0); |
31dbd01f IE |
1049 | int swapped; |
1050 | int err = -EFAULT; | |
ac46d4f3 | 1051 | struct mmu_notifier_range range; |
6c287605 | 1052 | bool anon_exclusive; |
31dbd01f | 1053 | |
36eaff33 KS |
1054 | pvmw.address = page_address_in_vma(page, vma); |
1055 | if (pvmw.address == -EFAULT) | |
31dbd01f IE |
1056 | goto out; |
1057 | ||
29ad768c | 1058 | BUG_ON(PageTransCompound(page)); |
6bdb913f | 1059 | |
7269f999 | 1060 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, |
6f4f13e8 | 1061 | pvmw.address, |
ac46d4f3 JG |
1062 | pvmw.address + PAGE_SIZE); |
1063 | mmu_notifier_invalidate_range_start(&range); | |
6bdb913f | 1064 | |
36eaff33 | 1065 | if (!page_vma_mapped_walk(&pvmw)) |
6bdb913f | 1066 | goto out_mn; |
36eaff33 KS |
1067 | if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?")) |
1068 | goto out_unlock; | |
31dbd01f | 1069 | |
6c287605 | 1070 | anon_exclusive = PageAnonExclusive(page); |
595cd8f2 | 1071 | if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) || |
6c287605 | 1072 | anon_exclusive || mm_tlb_flush_pending(mm)) { |
31dbd01f IE |
1073 | pte_t entry; |
1074 | ||
1075 | swapped = PageSwapCache(page); | |
36eaff33 | 1076 | flush_cache_page(vma, pvmw.address, page_to_pfn(page)); |
31dbd01f | 1077 | /* |
25985edc | 1078 | * Ok this is tricky, when get_user_pages_fast() run it doesn't |
31dbd01f | 1079 | * take any lock, therefore the check that we are going to make |
f0953a1b | 1080 | * with the pagecount against the mapcount is racy and |
31dbd01f IE |
1081 | * O_DIRECT can happen right after the check. |
1082 | * So we clear the pte and flush the tlb before the check | |
1083 | * this assure us that no O_DIRECT can happen after the check | |
1084 | * or in the middle of the check. | |
0f10851e JG |
1085 | * |
1086 | * No need to notify as we are downgrading page table to read | |
1087 | * only not changing it to point to a new page. | |
1088 | * | |
ee65728e | 1089 | * See Documentation/mm/mmu_notifier.rst |
31dbd01f | 1090 | */ |
0f10851e | 1091 | entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte); |
31dbd01f IE |
1092 | /* |
1093 | * Check that no O_DIRECT or similar I/O is in progress on the | |
1094 | * page | |
1095 | */ | |
31e855ea | 1096 | if (page_mapcount(page) + 1 + swapped != page_count(page)) { |
36eaff33 | 1097 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f IE |
1098 | goto out_unlock; |
1099 | } | |
6c287605 | 1100 | |
088b8aa5 | 1101 | /* See page_try_share_anon_rmap(): clear PTE first. */ |
6c287605 DH |
1102 | if (anon_exclusive && page_try_share_anon_rmap(page)) { |
1103 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); | |
1104 | goto out_unlock; | |
1105 | } | |
1106 | ||
4e31635c HD |
1107 | if (pte_dirty(entry)) |
1108 | set_page_dirty(page); | |
6a56ccbc DH |
1109 | entry = pte_mkclean(entry); |
1110 | ||
1111 | if (pte_write(entry)) | |
1112 | entry = pte_wrprotect(entry); | |
595cd8f2 | 1113 | |
36eaff33 | 1114 | set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f | 1115 | } |
36eaff33 | 1116 | *orig_pte = *pvmw.pte; |
31dbd01f IE |
1117 | err = 0; |
1118 | ||
1119 | out_unlock: | |
36eaff33 | 1120 | page_vma_mapped_walk_done(&pvmw); |
6bdb913f | 1121 | out_mn: |
ac46d4f3 | 1122 | mmu_notifier_invalidate_range_end(&range); |
31dbd01f IE |
1123 | out: |
1124 | return err; | |
1125 | } | |
1126 | ||
1127 | /** | |
1128 | * replace_page - replace page in vma by new ksm page | |
8dd3557a HD |
1129 | * @vma: vma that holds the pte pointing to page |
1130 | * @page: the page we are replacing by kpage | |
1131 | * @kpage: the ksm page we replace page by | |
31dbd01f IE |
1132 | * @orig_pte: the original value of the pte |
1133 | * | |
1134 | * Returns 0 on success, -EFAULT on failure. | |
1135 | */ | |
8dd3557a HD |
1136 | static int replace_page(struct vm_area_struct *vma, struct page *page, |
1137 | struct page *kpage, pte_t orig_pte) | |
31dbd01f IE |
1138 | { |
1139 | struct mm_struct *mm = vma->vm_mm; | |
b4e6f66e | 1140 | struct folio *folio; |
31dbd01f | 1141 | pmd_t *pmd; |
50722804 | 1142 | pmd_t pmde; |
31dbd01f | 1143 | pte_t *ptep; |
e86c59b1 | 1144 | pte_t newpte; |
31dbd01f IE |
1145 | spinlock_t *ptl; |
1146 | unsigned long addr; | |
31dbd01f | 1147 | int err = -EFAULT; |
ac46d4f3 | 1148 | struct mmu_notifier_range range; |
31dbd01f | 1149 | |
8dd3557a | 1150 | addr = page_address_in_vma(page, vma); |
31dbd01f IE |
1151 | if (addr == -EFAULT) |
1152 | goto out; | |
1153 | ||
6219049a BL |
1154 | pmd = mm_find_pmd(mm, addr); |
1155 | if (!pmd) | |
31dbd01f | 1156 | goto out; |
50722804 ZK |
1157 | /* |
1158 | * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() | |
1159 | * without holding anon_vma lock for write. So when looking for a | |
1160 | * genuine pmde (in which to find pte), test present and !THP together. | |
1161 | */ | |
1162 | pmde = *pmd; | |
1163 | barrier(); | |
1164 | if (!pmd_present(pmde) || pmd_trans_huge(pmde)) | |
1165 | goto out; | |
31dbd01f | 1166 | |
7269f999 | 1167 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, addr, |
6f4f13e8 | 1168 | addr + PAGE_SIZE); |
ac46d4f3 | 1169 | mmu_notifier_invalidate_range_start(&range); |
6bdb913f | 1170 | |
31dbd01f IE |
1171 | ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); |
1172 | if (!pte_same(*ptep, orig_pte)) { | |
1173 | pte_unmap_unlock(ptep, ptl); | |
6bdb913f | 1174 | goto out_mn; |
31dbd01f | 1175 | } |
6c287605 DH |
1176 | VM_BUG_ON_PAGE(PageAnonExclusive(page), page); |
1177 | VM_BUG_ON_PAGE(PageAnon(kpage) && PageAnonExclusive(kpage), kpage); | |
31dbd01f | 1178 | |
e86c59b1 CI |
1179 | /* |
1180 | * No need to check ksm_use_zero_pages here: we can only have a | |
457aef94 | 1181 | * zero_page here if ksm_use_zero_pages was enabled already. |
e86c59b1 CI |
1182 | */ |
1183 | if (!is_zero_pfn(page_to_pfn(kpage))) { | |
1184 | get_page(kpage); | |
f1e2db12 | 1185 | page_add_anon_rmap(kpage, vma, addr, RMAP_NONE); |
e86c59b1 CI |
1186 | newpte = mk_pte(kpage, vma->vm_page_prot); |
1187 | } else { | |
1188 | newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage), | |
1189 | vma->vm_page_prot)); | |
a38c015f CI |
1190 | /* |
1191 | * We're replacing an anonymous page with a zero page, which is | |
1192 | * not anonymous. We need to do proper accounting otherwise we | |
1193 | * will get wrong values in /proc, and a BUG message in dmesg | |
1194 | * when tearing down the mm. | |
1195 | */ | |
1196 | dec_mm_counter(mm, MM_ANONPAGES); | |
e86c59b1 | 1197 | } |
31dbd01f IE |
1198 | |
1199 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
0f10851e JG |
1200 | /* |
1201 | * No need to notify as we are replacing a read only page with another | |
1202 | * read only page with the same content. | |
1203 | * | |
ee65728e | 1204 | * See Documentation/mm/mmu_notifier.rst |
0f10851e JG |
1205 | */ |
1206 | ptep_clear_flush(vma, addr, ptep); | |
e86c59b1 | 1207 | set_pte_at_notify(mm, addr, ptep, newpte); |
31dbd01f | 1208 | |
b4e6f66e | 1209 | folio = page_folio(page); |
cea86fe2 | 1210 | page_remove_rmap(page, vma, false); |
b4e6f66e MWO |
1211 | if (!folio_mapped(folio)) |
1212 | folio_free_swap(folio); | |
1213 | folio_put(folio); | |
31dbd01f IE |
1214 | |
1215 | pte_unmap_unlock(ptep, ptl); | |
1216 | err = 0; | |
6bdb913f | 1217 | out_mn: |
ac46d4f3 | 1218 | mmu_notifier_invalidate_range_end(&range); |
31dbd01f IE |
1219 | out: |
1220 | return err; | |
1221 | } | |
1222 | ||
1223 | /* | |
1224 | * try_to_merge_one_page - take two pages and merge them into one | |
8dd3557a HD |
1225 | * @vma: the vma that holds the pte pointing to page |
1226 | * @page: the PageAnon page that we want to replace with kpage | |
80e14822 HD |
1227 | * @kpage: the PageKsm page that we want to map instead of page, |
1228 | * or NULL the first time when we want to use page as kpage. | |
31dbd01f IE |
1229 | * |
1230 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
1231 | */ | |
1232 | static int try_to_merge_one_page(struct vm_area_struct *vma, | |
8dd3557a | 1233 | struct page *page, struct page *kpage) |
31dbd01f IE |
1234 | { |
1235 | pte_t orig_pte = __pte(0); | |
1236 | int err = -EFAULT; | |
1237 | ||
db114b83 HD |
1238 | if (page == kpage) /* ksm page forked */ |
1239 | return 0; | |
1240 | ||
8dd3557a | 1241 | if (!PageAnon(page)) |
31dbd01f IE |
1242 | goto out; |
1243 | ||
31dbd01f IE |
1244 | /* |
1245 | * We need the page lock to read a stable PageSwapCache in | |
1246 | * write_protect_page(). We use trylock_page() instead of | |
1247 | * lock_page() because we don't want to wait here - we | |
1248 | * prefer to continue scanning and merging different pages, | |
1249 | * then come back to this page when it is unlocked. | |
1250 | */ | |
8dd3557a | 1251 | if (!trylock_page(page)) |
31e855ea | 1252 | goto out; |
f765f540 KS |
1253 | |
1254 | if (PageTransCompound(page)) { | |
a7306c34 | 1255 | if (split_huge_page(page)) |
f765f540 KS |
1256 | goto out_unlock; |
1257 | } | |
1258 | ||
31dbd01f IE |
1259 | /* |
1260 | * If this anonymous page is mapped only here, its pte may need | |
1261 | * to be write-protected. If it's mapped elsewhere, all of its | |
1262 | * ptes are necessarily already write-protected. But in either | |
1263 | * case, we need to lock and check page_count is not raised. | |
1264 | */ | |
80e14822 HD |
1265 | if (write_protect_page(vma, page, &orig_pte) == 0) { |
1266 | if (!kpage) { | |
1267 | /* | |
1268 | * While we hold page lock, upgrade page from | |
1269 | * PageAnon+anon_vma to PageKsm+NULL stable_node: | |
1270 | * stable_tree_insert() will update stable_node. | |
1271 | */ | |
1272 | set_page_stable_node(page, NULL); | |
1273 | mark_page_accessed(page); | |
337ed7eb MK |
1274 | /* |
1275 | * Page reclaim just frees a clean page with no dirty | |
1276 | * ptes: make sure that the ksm page would be swapped. | |
1277 | */ | |
1278 | if (!PageDirty(page)) | |
1279 | SetPageDirty(page); | |
80e14822 HD |
1280 | err = 0; |
1281 | } else if (pages_identical(page, kpage)) | |
1282 | err = replace_page(vma, page, kpage, orig_pte); | |
1283 | } | |
31dbd01f | 1284 | |
f765f540 | 1285 | out_unlock: |
8dd3557a | 1286 | unlock_page(page); |
31dbd01f IE |
1287 | out: |
1288 | return err; | |
1289 | } | |
1290 | ||
81464e30 HD |
1291 | /* |
1292 | * try_to_merge_with_ksm_page - like try_to_merge_two_pages, | |
1293 | * but no new kernel page is allocated: kpage must already be a ksm page. | |
8dd3557a HD |
1294 | * |
1295 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
81464e30 | 1296 | */ |
21fbd591 | 1297 | static int try_to_merge_with_ksm_page(struct ksm_rmap_item *rmap_item, |
8dd3557a | 1298 | struct page *page, struct page *kpage) |
81464e30 | 1299 | { |
8dd3557a | 1300 | struct mm_struct *mm = rmap_item->mm; |
81464e30 HD |
1301 | struct vm_area_struct *vma; |
1302 | int err = -EFAULT; | |
1303 | ||
d8ed45c5 | 1304 | mmap_read_lock(mm); |
85c6e8dd AA |
1305 | vma = find_mergeable_vma(mm, rmap_item->address); |
1306 | if (!vma) | |
81464e30 HD |
1307 | goto out; |
1308 | ||
8dd3557a | 1309 | err = try_to_merge_one_page(vma, page, kpage); |
db114b83 HD |
1310 | if (err) |
1311 | goto out; | |
1312 | ||
bc56620b HD |
1313 | /* Unstable nid is in union with stable anon_vma: remove first */ |
1314 | remove_rmap_item_from_tree(rmap_item); | |
1315 | ||
c1e8d7c6 | 1316 | /* Must get reference to anon_vma while still holding mmap_lock */ |
9e60109f PZ |
1317 | rmap_item->anon_vma = vma->anon_vma; |
1318 | get_anon_vma(vma->anon_vma); | |
81464e30 | 1319 | out: |
d8ed45c5 | 1320 | mmap_read_unlock(mm); |
81464e30 HD |
1321 | return err; |
1322 | } | |
1323 | ||
31dbd01f IE |
1324 | /* |
1325 | * try_to_merge_two_pages - take two identical pages and prepare them | |
1326 | * to be merged into one page. | |
1327 | * | |
8dd3557a HD |
1328 | * This function returns the kpage if we successfully merged two identical |
1329 | * pages into one ksm page, NULL otherwise. | |
31dbd01f | 1330 | * |
80e14822 | 1331 | * Note that this function upgrades page to ksm page: if one of the pages |
31dbd01f IE |
1332 | * is already a ksm page, try_to_merge_with_ksm_page should be used. |
1333 | */ | |
21fbd591 | 1334 | static struct page *try_to_merge_two_pages(struct ksm_rmap_item *rmap_item, |
8dd3557a | 1335 | struct page *page, |
21fbd591 | 1336 | struct ksm_rmap_item *tree_rmap_item, |
8dd3557a | 1337 | struct page *tree_page) |
31dbd01f | 1338 | { |
80e14822 | 1339 | int err; |
31dbd01f | 1340 | |
80e14822 | 1341 | err = try_to_merge_with_ksm_page(rmap_item, page, NULL); |
31dbd01f | 1342 | if (!err) { |
8dd3557a | 1343 | err = try_to_merge_with_ksm_page(tree_rmap_item, |
80e14822 | 1344 | tree_page, page); |
31dbd01f | 1345 | /* |
81464e30 HD |
1346 | * If that fails, we have a ksm page with only one pte |
1347 | * pointing to it: so break it. | |
31dbd01f | 1348 | */ |
4035c07a | 1349 | if (err) |
8dd3557a | 1350 | break_cow(rmap_item); |
31dbd01f | 1351 | } |
80e14822 | 1352 | return err ? NULL : page; |
31dbd01f IE |
1353 | } |
1354 | ||
2c653d0e | 1355 | static __always_inline |
21fbd591 | 1356 | bool __is_page_sharing_candidate(struct ksm_stable_node *stable_node, int offset) |
2c653d0e AA |
1357 | { |
1358 | VM_BUG_ON(stable_node->rmap_hlist_len < 0); | |
1359 | /* | |
1360 | * Check that at least one mapping still exists, otherwise | |
1361 | * there's no much point to merge and share with this | |
1362 | * stable_node, as the underlying tree_page of the other | |
1363 | * sharer is going to be freed soon. | |
1364 | */ | |
1365 | return stable_node->rmap_hlist_len && | |
1366 | stable_node->rmap_hlist_len + offset < ksm_max_page_sharing; | |
1367 | } | |
1368 | ||
1369 | static __always_inline | |
21fbd591 | 1370 | bool is_page_sharing_candidate(struct ksm_stable_node *stable_node) |
2c653d0e AA |
1371 | { |
1372 | return __is_page_sharing_candidate(stable_node, 0); | |
1373 | } | |
1374 | ||
21fbd591 QZ |
1375 | static struct page *stable_node_dup(struct ksm_stable_node **_stable_node_dup, |
1376 | struct ksm_stable_node **_stable_node, | |
c01f0b54 CIK |
1377 | struct rb_root *root, |
1378 | bool prune_stale_stable_nodes) | |
2c653d0e | 1379 | { |
21fbd591 | 1380 | struct ksm_stable_node *dup, *found = NULL, *stable_node = *_stable_node; |
2c653d0e | 1381 | struct hlist_node *hlist_safe; |
8dc5ffcd | 1382 | struct page *_tree_page, *tree_page = NULL; |
2c653d0e AA |
1383 | int nr = 0; |
1384 | int found_rmap_hlist_len; | |
1385 | ||
1386 | if (!prune_stale_stable_nodes || | |
1387 | time_before(jiffies, stable_node->chain_prune_time + | |
1388 | msecs_to_jiffies( | |
1389 | ksm_stable_node_chains_prune_millisecs))) | |
1390 | prune_stale_stable_nodes = false; | |
1391 | else | |
1392 | stable_node->chain_prune_time = jiffies; | |
1393 | ||
1394 | hlist_for_each_entry_safe(dup, hlist_safe, | |
1395 | &stable_node->hlist, hlist_dup) { | |
1396 | cond_resched(); | |
1397 | /* | |
1398 | * We must walk all stable_node_dup to prune the stale | |
1399 | * stable nodes during lookup. | |
1400 | * | |
1401 | * get_ksm_page can drop the nodes from the | |
1402 | * stable_node->hlist if they point to freed pages | |
1403 | * (that's why we do a _safe walk). The "dup" | |
1404 | * stable_node parameter itself will be freed from | |
1405 | * under us if it returns NULL. | |
1406 | */ | |
2cee57d1 | 1407 | _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK); |
2c653d0e AA |
1408 | if (!_tree_page) |
1409 | continue; | |
1410 | nr += 1; | |
1411 | if (is_page_sharing_candidate(dup)) { | |
1412 | if (!found || | |
1413 | dup->rmap_hlist_len > found_rmap_hlist_len) { | |
1414 | if (found) | |
8dc5ffcd | 1415 | put_page(tree_page); |
2c653d0e AA |
1416 | found = dup; |
1417 | found_rmap_hlist_len = found->rmap_hlist_len; | |
8dc5ffcd | 1418 | tree_page = _tree_page; |
2c653d0e | 1419 | |
8dc5ffcd | 1420 | /* skip put_page for found dup */ |
2c653d0e AA |
1421 | if (!prune_stale_stable_nodes) |
1422 | break; | |
2c653d0e AA |
1423 | continue; |
1424 | } | |
1425 | } | |
1426 | put_page(_tree_page); | |
1427 | } | |
1428 | ||
80b18dfa AA |
1429 | if (found) { |
1430 | /* | |
1431 | * nr is counting all dups in the chain only if | |
1432 | * prune_stale_stable_nodes is true, otherwise we may | |
1433 | * break the loop at nr == 1 even if there are | |
1434 | * multiple entries. | |
1435 | */ | |
1436 | if (prune_stale_stable_nodes && nr == 1) { | |
2c653d0e AA |
1437 | /* |
1438 | * If there's not just one entry it would | |
1439 | * corrupt memory, better BUG_ON. In KSM | |
1440 | * context with no lock held it's not even | |
1441 | * fatal. | |
1442 | */ | |
1443 | BUG_ON(stable_node->hlist.first->next); | |
1444 | ||
1445 | /* | |
1446 | * There's just one entry and it is below the | |
1447 | * deduplication limit so drop the chain. | |
1448 | */ | |
1449 | rb_replace_node(&stable_node->node, &found->node, | |
1450 | root); | |
1451 | free_stable_node(stable_node); | |
1452 | ksm_stable_node_chains--; | |
1453 | ksm_stable_node_dups--; | |
b4fecc67 | 1454 | /* |
0ba1d0f7 AA |
1455 | * NOTE: the caller depends on the stable_node |
1456 | * to be equal to stable_node_dup if the chain | |
1457 | * was collapsed. | |
b4fecc67 | 1458 | */ |
0ba1d0f7 AA |
1459 | *_stable_node = found; |
1460 | /* | |
f0953a1b | 1461 | * Just for robustness, as stable_node is |
0ba1d0f7 AA |
1462 | * otherwise left as a stable pointer, the |
1463 | * compiler shall optimize it away at build | |
1464 | * time. | |
1465 | */ | |
1466 | stable_node = NULL; | |
80b18dfa AA |
1467 | } else if (stable_node->hlist.first != &found->hlist_dup && |
1468 | __is_page_sharing_candidate(found, 1)) { | |
2c653d0e | 1469 | /* |
80b18dfa AA |
1470 | * If the found stable_node dup can accept one |
1471 | * more future merge (in addition to the one | |
1472 | * that is underway) and is not at the head of | |
1473 | * the chain, put it there so next search will | |
1474 | * be quicker in the !prune_stale_stable_nodes | |
1475 | * case. | |
1476 | * | |
1477 | * NOTE: it would be inaccurate to use nr > 1 | |
1478 | * instead of checking the hlist.first pointer | |
1479 | * directly, because in the | |
1480 | * prune_stale_stable_nodes case "nr" isn't | |
1481 | * the position of the found dup in the chain, | |
1482 | * but the total number of dups in the chain. | |
2c653d0e AA |
1483 | */ |
1484 | hlist_del(&found->hlist_dup); | |
1485 | hlist_add_head(&found->hlist_dup, | |
1486 | &stable_node->hlist); | |
1487 | } | |
1488 | } | |
1489 | ||
8dc5ffcd AA |
1490 | *_stable_node_dup = found; |
1491 | return tree_page; | |
2c653d0e AA |
1492 | } |
1493 | ||
21fbd591 | 1494 | static struct ksm_stable_node *stable_node_dup_any(struct ksm_stable_node *stable_node, |
2c653d0e AA |
1495 | struct rb_root *root) |
1496 | { | |
1497 | if (!is_stable_node_chain(stable_node)) | |
1498 | return stable_node; | |
1499 | if (hlist_empty(&stable_node->hlist)) { | |
1500 | free_stable_node_chain(stable_node, root); | |
1501 | return NULL; | |
1502 | } | |
1503 | return hlist_entry(stable_node->hlist.first, | |
1504 | typeof(*stable_node), hlist_dup); | |
1505 | } | |
1506 | ||
8dc5ffcd AA |
1507 | /* |
1508 | * Like for get_ksm_page, this function can free the *_stable_node and | |
1509 | * *_stable_node_dup if the returned tree_page is NULL. | |
1510 | * | |
1511 | * It can also free and overwrite *_stable_node with the found | |
1512 | * stable_node_dup if the chain is collapsed (in which case | |
1513 | * *_stable_node will be equal to *_stable_node_dup like if the chain | |
1514 | * never existed). It's up to the caller to verify tree_page is not | |
1515 | * NULL before dereferencing *_stable_node or *_stable_node_dup. | |
1516 | * | |
1517 | * *_stable_node_dup is really a second output parameter of this | |
1518 | * function and will be overwritten in all cases, the caller doesn't | |
1519 | * need to initialize it. | |
1520 | */ | |
21fbd591 QZ |
1521 | static struct page *__stable_node_chain(struct ksm_stable_node **_stable_node_dup, |
1522 | struct ksm_stable_node **_stable_node, | |
8dc5ffcd AA |
1523 | struct rb_root *root, |
1524 | bool prune_stale_stable_nodes) | |
2c653d0e | 1525 | { |
21fbd591 | 1526 | struct ksm_stable_node *stable_node = *_stable_node; |
2c653d0e AA |
1527 | if (!is_stable_node_chain(stable_node)) { |
1528 | if (is_page_sharing_candidate(stable_node)) { | |
8dc5ffcd | 1529 | *_stable_node_dup = stable_node; |
2cee57d1 | 1530 | return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK); |
2c653d0e | 1531 | } |
8dc5ffcd AA |
1532 | /* |
1533 | * _stable_node_dup set to NULL means the stable_node | |
1534 | * reached the ksm_max_page_sharing limit. | |
1535 | */ | |
1536 | *_stable_node_dup = NULL; | |
2c653d0e AA |
1537 | return NULL; |
1538 | } | |
8dc5ffcd | 1539 | return stable_node_dup(_stable_node_dup, _stable_node, root, |
2c653d0e AA |
1540 | prune_stale_stable_nodes); |
1541 | } | |
1542 | ||
21fbd591 QZ |
1543 | static __always_inline struct page *chain_prune(struct ksm_stable_node **s_n_d, |
1544 | struct ksm_stable_node **s_n, | |
8dc5ffcd | 1545 | struct rb_root *root) |
2c653d0e | 1546 | { |
8dc5ffcd | 1547 | return __stable_node_chain(s_n_d, s_n, root, true); |
2c653d0e AA |
1548 | } |
1549 | ||
21fbd591 QZ |
1550 | static __always_inline struct page *chain(struct ksm_stable_node **s_n_d, |
1551 | struct ksm_stable_node *s_n, | |
8dc5ffcd | 1552 | struct rb_root *root) |
2c653d0e | 1553 | { |
21fbd591 | 1554 | struct ksm_stable_node *old_stable_node = s_n; |
8dc5ffcd AA |
1555 | struct page *tree_page; |
1556 | ||
1557 | tree_page = __stable_node_chain(s_n_d, &s_n, root, false); | |
1558 | /* not pruning dups so s_n cannot have changed */ | |
1559 | VM_BUG_ON(s_n != old_stable_node); | |
1560 | return tree_page; | |
2c653d0e AA |
1561 | } |
1562 | ||
31dbd01f | 1563 | /* |
8dd3557a | 1564 | * stable_tree_search - search for page inside the stable tree |
31dbd01f IE |
1565 | * |
1566 | * This function checks if there is a page inside the stable tree | |
1567 | * with identical content to the page that we are scanning right now. | |
1568 | * | |
7b6ba2c7 | 1569 | * This function returns the stable tree node of identical content if found, |
31dbd01f IE |
1570 | * NULL otherwise. |
1571 | */ | |
62b61f61 | 1572 | static struct page *stable_tree_search(struct page *page) |
31dbd01f | 1573 | { |
90bd6fd3 | 1574 | int nid; |
ef53d16c | 1575 | struct rb_root *root; |
4146d2d6 HD |
1576 | struct rb_node **new; |
1577 | struct rb_node *parent; | |
21fbd591 QZ |
1578 | struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; |
1579 | struct ksm_stable_node *page_node; | |
31dbd01f | 1580 | |
4146d2d6 HD |
1581 | page_node = page_stable_node(page); |
1582 | if (page_node && page_node->head != &migrate_nodes) { | |
1583 | /* ksm page forked */ | |
08beca44 | 1584 | get_page(page); |
62b61f61 | 1585 | return page; |
08beca44 HD |
1586 | } |
1587 | ||
90bd6fd3 | 1588 | nid = get_kpfn_nid(page_to_pfn(page)); |
ef53d16c | 1589 | root = root_stable_tree + nid; |
4146d2d6 | 1590 | again: |
ef53d16c | 1591 | new = &root->rb_node; |
4146d2d6 | 1592 | parent = NULL; |
90bd6fd3 | 1593 | |
4146d2d6 | 1594 | while (*new) { |
4035c07a | 1595 | struct page *tree_page; |
31dbd01f IE |
1596 | int ret; |
1597 | ||
08beca44 | 1598 | cond_resched(); |
21fbd591 | 1599 | stable_node = rb_entry(*new, struct ksm_stable_node, node); |
2c653d0e | 1600 | stable_node_any = NULL; |
8dc5ffcd | 1601 | tree_page = chain_prune(&stable_node_dup, &stable_node, root); |
b4fecc67 AA |
1602 | /* |
1603 | * NOTE: stable_node may have been freed by | |
1604 | * chain_prune() if the returned stable_node_dup is | |
1605 | * not NULL. stable_node_dup may have been inserted in | |
1606 | * the rbtree instead as a regular stable_node (in | |
1607 | * order to collapse the stable_node chain if a single | |
0ba1d0f7 | 1608 | * stable_node dup was found in it). In such case the |
3413b2c8 | 1609 | * stable_node is overwritten by the callee to point |
0ba1d0f7 AA |
1610 | * to the stable_node_dup that was collapsed in the |
1611 | * stable rbtree and stable_node will be equal to | |
1612 | * stable_node_dup like if the chain never existed. | |
b4fecc67 | 1613 | */ |
2c653d0e AA |
1614 | if (!stable_node_dup) { |
1615 | /* | |
1616 | * Either all stable_node dups were full in | |
1617 | * this stable_node chain, or this chain was | |
1618 | * empty and should be rb_erased. | |
1619 | */ | |
1620 | stable_node_any = stable_node_dup_any(stable_node, | |
1621 | root); | |
1622 | if (!stable_node_any) { | |
1623 | /* rb_erase just run */ | |
1624 | goto again; | |
1625 | } | |
1626 | /* | |
1627 | * Take any of the stable_node dups page of | |
1628 | * this stable_node chain to let the tree walk | |
1629 | * continue. All KSM pages belonging to the | |
1630 | * stable_node dups in a stable_node chain | |
1631 | * have the same content and they're | |
457aef94 | 1632 | * write protected at all times. Any will work |
2c653d0e AA |
1633 | * fine to continue the walk. |
1634 | */ | |
2cee57d1 YS |
1635 | tree_page = get_ksm_page(stable_node_any, |
1636 | GET_KSM_PAGE_NOLOCK); | |
2c653d0e AA |
1637 | } |
1638 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1639 | if (!tree_page) { |
1640 | /* | |
1641 | * If we walked over a stale stable_node, | |
1642 | * get_ksm_page() will call rb_erase() and it | |
1643 | * may rebalance the tree from under us. So | |
1644 | * restart the search from scratch. Returning | |
1645 | * NULL would be safe too, but we'd generate | |
1646 | * false negative insertions just because some | |
1647 | * stable_node was stale. | |
1648 | */ | |
1649 | goto again; | |
1650 | } | |
31dbd01f | 1651 | |
4035c07a | 1652 | ret = memcmp_pages(page, tree_page); |
c8d6553b | 1653 | put_page(tree_page); |
31dbd01f | 1654 | |
4146d2d6 | 1655 | parent = *new; |
c8d6553b | 1656 | if (ret < 0) |
4146d2d6 | 1657 | new = &parent->rb_left; |
c8d6553b | 1658 | else if (ret > 0) |
4146d2d6 | 1659 | new = &parent->rb_right; |
c8d6553b | 1660 | else { |
2c653d0e AA |
1661 | if (page_node) { |
1662 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1663 | /* | |
1664 | * Test if the migrated page should be merged | |
1665 | * into a stable node dup. If the mapcount is | |
1666 | * 1 we can migrate it with another KSM page | |
1667 | * without adding it to the chain. | |
1668 | */ | |
1669 | if (page_mapcount(page) > 1) | |
1670 | goto chain_append; | |
1671 | } | |
1672 | ||
1673 | if (!stable_node_dup) { | |
1674 | /* | |
1675 | * If the stable_node is a chain and | |
1676 | * we got a payload match in memcmp | |
1677 | * but we cannot merge the scanned | |
1678 | * page in any of the existing | |
1679 | * stable_node dups because they're | |
1680 | * all full, we need to wait the | |
1681 | * scanned page to find itself a match | |
1682 | * in the unstable tree to create a | |
1683 | * brand new KSM page to add later to | |
1684 | * the dups of this stable_node. | |
1685 | */ | |
1686 | return NULL; | |
1687 | } | |
1688 | ||
c8d6553b HD |
1689 | /* |
1690 | * Lock and unlock the stable_node's page (which | |
1691 | * might already have been migrated) so that page | |
1692 | * migration is sure to notice its raised count. | |
1693 | * It would be more elegant to return stable_node | |
1694 | * than kpage, but that involves more changes. | |
1695 | */ | |
2cee57d1 YS |
1696 | tree_page = get_ksm_page(stable_node_dup, |
1697 | GET_KSM_PAGE_TRYLOCK); | |
1698 | ||
1699 | if (PTR_ERR(tree_page) == -EBUSY) | |
1700 | return ERR_PTR(-EBUSY); | |
1701 | ||
2c653d0e AA |
1702 | if (unlikely(!tree_page)) |
1703 | /* | |
1704 | * The tree may have been rebalanced, | |
1705 | * so re-evaluate parent and new. | |
1706 | */ | |
4146d2d6 | 1707 | goto again; |
2c653d0e AA |
1708 | unlock_page(tree_page); |
1709 | ||
1710 | if (get_kpfn_nid(stable_node_dup->kpfn) != | |
1711 | NUMA(stable_node_dup->nid)) { | |
1712 | put_page(tree_page); | |
1713 | goto replace; | |
1714 | } | |
1715 | return tree_page; | |
c8d6553b | 1716 | } |
31dbd01f IE |
1717 | } |
1718 | ||
4146d2d6 HD |
1719 | if (!page_node) |
1720 | return NULL; | |
1721 | ||
1722 | list_del(&page_node->list); | |
1723 | DO_NUMA(page_node->nid = nid); | |
1724 | rb_link_node(&page_node->node, parent, new); | |
ef53d16c | 1725 | rb_insert_color(&page_node->node, root); |
2c653d0e AA |
1726 | out: |
1727 | if (is_page_sharing_candidate(page_node)) { | |
1728 | get_page(page); | |
1729 | return page; | |
1730 | } else | |
1731 | return NULL; | |
4146d2d6 HD |
1732 | |
1733 | replace: | |
b4fecc67 AA |
1734 | /* |
1735 | * If stable_node was a chain and chain_prune collapsed it, | |
0ba1d0f7 AA |
1736 | * stable_node has been updated to be the new regular |
1737 | * stable_node. A collapse of the chain is indistinguishable | |
1738 | * from the case there was no chain in the stable | |
1739 | * rbtree. Otherwise stable_node is the chain and | |
1740 | * stable_node_dup is the dup to replace. | |
b4fecc67 | 1741 | */ |
0ba1d0f7 | 1742 | if (stable_node_dup == stable_node) { |
b4fecc67 AA |
1743 | VM_BUG_ON(is_stable_node_chain(stable_node_dup)); |
1744 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); | |
2c653d0e AA |
1745 | /* there is no chain */ |
1746 | if (page_node) { | |
1747 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1748 | list_del(&page_node->list); | |
1749 | DO_NUMA(page_node->nid = nid); | |
b4fecc67 AA |
1750 | rb_replace_node(&stable_node_dup->node, |
1751 | &page_node->node, | |
2c653d0e AA |
1752 | root); |
1753 | if (is_page_sharing_candidate(page_node)) | |
1754 | get_page(page); | |
1755 | else | |
1756 | page = NULL; | |
1757 | } else { | |
b4fecc67 | 1758 | rb_erase(&stable_node_dup->node, root); |
2c653d0e AA |
1759 | page = NULL; |
1760 | } | |
4146d2d6 | 1761 | } else { |
2c653d0e AA |
1762 | VM_BUG_ON(!is_stable_node_chain(stable_node)); |
1763 | __stable_node_dup_del(stable_node_dup); | |
1764 | if (page_node) { | |
1765 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1766 | list_del(&page_node->list); | |
1767 | DO_NUMA(page_node->nid = nid); | |
1768 | stable_node_chain_add_dup(page_node, stable_node); | |
1769 | if (is_page_sharing_candidate(page_node)) | |
1770 | get_page(page); | |
1771 | else | |
1772 | page = NULL; | |
1773 | } else { | |
1774 | page = NULL; | |
1775 | } | |
4146d2d6 | 1776 | } |
2c653d0e AA |
1777 | stable_node_dup->head = &migrate_nodes; |
1778 | list_add(&stable_node_dup->list, stable_node_dup->head); | |
4146d2d6 | 1779 | return page; |
2c653d0e AA |
1780 | |
1781 | chain_append: | |
1782 | /* stable_node_dup could be null if it reached the limit */ | |
1783 | if (!stable_node_dup) | |
1784 | stable_node_dup = stable_node_any; | |
b4fecc67 AA |
1785 | /* |
1786 | * If stable_node was a chain and chain_prune collapsed it, | |
0ba1d0f7 AA |
1787 | * stable_node has been updated to be the new regular |
1788 | * stable_node. A collapse of the chain is indistinguishable | |
1789 | * from the case there was no chain in the stable | |
1790 | * rbtree. Otherwise stable_node is the chain and | |
1791 | * stable_node_dup is the dup to replace. | |
b4fecc67 | 1792 | */ |
0ba1d0f7 | 1793 | if (stable_node_dup == stable_node) { |
b4fecc67 | 1794 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); |
2c653d0e AA |
1795 | /* chain is missing so create it */ |
1796 | stable_node = alloc_stable_node_chain(stable_node_dup, | |
1797 | root); | |
1798 | if (!stable_node) | |
1799 | return NULL; | |
1800 | } | |
1801 | /* | |
1802 | * Add this stable_node dup that was | |
1803 | * migrated to the stable_node chain | |
1804 | * of the current nid for this page | |
1805 | * content. | |
1806 | */ | |
b4fecc67 | 1807 | VM_BUG_ON(!is_stable_node_dup(stable_node_dup)); |
2c653d0e AA |
1808 | VM_BUG_ON(page_node->head != &migrate_nodes); |
1809 | list_del(&page_node->list); | |
1810 | DO_NUMA(page_node->nid = nid); | |
1811 | stable_node_chain_add_dup(page_node, stable_node); | |
1812 | goto out; | |
31dbd01f IE |
1813 | } |
1814 | ||
1815 | /* | |
e850dcf5 | 1816 | * stable_tree_insert - insert stable tree node pointing to new ksm page |
31dbd01f IE |
1817 | * into the stable tree. |
1818 | * | |
7b6ba2c7 HD |
1819 | * This function returns the stable tree node just allocated on success, |
1820 | * NULL otherwise. | |
31dbd01f | 1821 | */ |
21fbd591 | 1822 | static struct ksm_stable_node *stable_tree_insert(struct page *kpage) |
31dbd01f | 1823 | { |
90bd6fd3 PH |
1824 | int nid; |
1825 | unsigned long kpfn; | |
ef53d16c | 1826 | struct rb_root *root; |
90bd6fd3 | 1827 | struct rb_node **new; |
f2e5ff85 | 1828 | struct rb_node *parent; |
21fbd591 | 1829 | struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; |
2c653d0e | 1830 | bool need_chain = false; |
31dbd01f | 1831 | |
90bd6fd3 PH |
1832 | kpfn = page_to_pfn(kpage); |
1833 | nid = get_kpfn_nid(kpfn); | |
ef53d16c | 1834 | root = root_stable_tree + nid; |
f2e5ff85 AA |
1835 | again: |
1836 | parent = NULL; | |
ef53d16c | 1837 | new = &root->rb_node; |
90bd6fd3 | 1838 | |
31dbd01f | 1839 | while (*new) { |
4035c07a | 1840 | struct page *tree_page; |
31dbd01f IE |
1841 | int ret; |
1842 | ||
08beca44 | 1843 | cond_resched(); |
21fbd591 | 1844 | stable_node = rb_entry(*new, struct ksm_stable_node, node); |
2c653d0e | 1845 | stable_node_any = NULL; |
8dc5ffcd | 1846 | tree_page = chain(&stable_node_dup, stable_node, root); |
2c653d0e AA |
1847 | if (!stable_node_dup) { |
1848 | /* | |
1849 | * Either all stable_node dups were full in | |
1850 | * this stable_node chain, or this chain was | |
1851 | * empty and should be rb_erased. | |
1852 | */ | |
1853 | stable_node_any = stable_node_dup_any(stable_node, | |
1854 | root); | |
1855 | if (!stable_node_any) { | |
1856 | /* rb_erase just run */ | |
1857 | goto again; | |
1858 | } | |
1859 | /* | |
1860 | * Take any of the stable_node dups page of | |
1861 | * this stable_node chain to let the tree walk | |
1862 | * continue. All KSM pages belonging to the | |
1863 | * stable_node dups in a stable_node chain | |
1864 | * have the same content and they're | |
457aef94 | 1865 | * write protected at all times. Any will work |
2c653d0e AA |
1866 | * fine to continue the walk. |
1867 | */ | |
2cee57d1 YS |
1868 | tree_page = get_ksm_page(stable_node_any, |
1869 | GET_KSM_PAGE_NOLOCK); | |
2c653d0e AA |
1870 | } |
1871 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1872 | if (!tree_page) { |
1873 | /* | |
1874 | * If we walked over a stale stable_node, | |
1875 | * get_ksm_page() will call rb_erase() and it | |
1876 | * may rebalance the tree from under us. So | |
1877 | * restart the search from scratch. Returning | |
1878 | * NULL would be safe too, but we'd generate | |
1879 | * false negative insertions just because some | |
1880 | * stable_node was stale. | |
1881 | */ | |
1882 | goto again; | |
1883 | } | |
31dbd01f | 1884 | |
4035c07a HD |
1885 | ret = memcmp_pages(kpage, tree_page); |
1886 | put_page(tree_page); | |
31dbd01f IE |
1887 | |
1888 | parent = *new; | |
1889 | if (ret < 0) | |
1890 | new = &parent->rb_left; | |
1891 | else if (ret > 0) | |
1892 | new = &parent->rb_right; | |
1893 | else { | |
2c653d0e AA |
1894 | need_chain = true; |
1895 | break; | |
31dbd01f IE |
1896 | } |
1897 | } | |
1898 | ||
2c653d0e AA |
1899 | stable_node_dup = alloc_stable_node(); |
1900 | if (!stable_node_dup) | |
7b6ba2c7 | 1901 | return NULL; |
31dbd01f | 1902 | |
2c653d0e AA |
1903 | INIT_HLIST_HEAD(&stable_node_dup->hlist); |
1904 | stable_node_dup->kpfn = kpfn; | |
1905 | set_page_stable_node(kpage, stable_node_dup); | |
1906 | stable_node_dup->rmap_hlist_len = 0; | |
1907 | DO_NUMA(stable_node_dup->nid = nid); | |
1908 | if (!need_chain) { | |
1909 | rb_link_node(&stable_node_dup->node, parent, new); | |
1910 | rb_insert_color(&stable_node_dup->node, root); | |
1911 | } else { | |
1912 | if (!is_stable_node_chain(stable_node)) { | |
21fbd591 | 1913 | struct ksm_stable_node *orig = stable_node; |
2c653d0e AA |
1914 | /* chain is missing so create it */ |
1915 | stable_node = alloc_stable_node_chain(orig, root); | |
1916 | if (!stable_node) { | |
1917 | free_stable_node(stable_node_dup); | |
1918 | return NULL; | |
1919 | } | |
1920 | } | |
1921 | stable_node_chain_add_dup(stable_node_dup, stable_node); | |
1922 | } | |
08beca44 | 1923 | |
2c653d0e | 1924 | return stable_node_dup; |
31dbd01f IE |
1925 | } |
1926 | ||
1927 | /* | |
8dd3557a HD |
1928 | * unstable_tree_search_insert - search for identical page, |
1929 | * else insert rmap_item into the unstable tree. | |
31dbd01f IE |
1930 | * |
1931 | * This function searches for a page in the unstable tree identical to the | |
1932 | * page currently being scanned; and if no identical page is found in the | |
1933 | * tree, we insert rmap_item as a new object into the unstable tree. | |
1934 | * | |
1935 | * This function returns pointer to rmap_item found to be identical | |
1936 | * to the currently scanned page, NULL otherwise. | |
1937 | * | |
1938 | * This function does both searching and inserting, because they share | |
1939 | * the same walking algorithm in an rbtree. | |
1940 | */ | |
8dd3557a | 1941 | static |
21fbd591 | 1942 | struct ksm_rmap_item *unstable_tree_search_insert(struct ksm_rmap_item *rmap_item, |
8dd3557a HD |
1943 | struct page *page, |
1944 | struct page **tree_pagep) | |
31dbd01f | 1945 | { |
90bd6fd3 PH |
1946 | struct rb_node **new; |
1947 | struct rb_root *root; | |
31dbd01f | 1948 | struct rb_node *parent = NULL; |
90bd6fd3 PH |
1949 | int nid; |
1950 | ||
1951 | nid = get_kpfn_nid(page_to_pfn(page)); | |
ef53d16c | 1952 | root = root_unstable_tree + nid; |
90bd6fd3 | 1953 | new = &root->rb_node; |
31dbd01f IE |
1954 | |
1955 | while (*new) { | |
21fbd591 | 1956 | struct ksm_rmap_item *tree_rmap_item; |
8dd3557a | 1957 | struct page *tree_page; |
31dbd01f IE |
1958 | int ret; |
1959 | ||
d178f27f | 1960 | cond_resched(); |
21fbd591 | 1961 | tree_rmap_item = rb_entry(*new, struct ksm_rmap_item, node); |
8dd3557a | 1962 | tree_page = get_mergeable_page(tree_rmap_item); |
c8f95ed1 | 1963 | if (!tree_page) |
31dbd01f IE |
1964 | return NULL; |
1965 | ||
1966 | /* | |
8dd3557a | 1967 | * Don't substitute a ksm page for a forked page. |
31dbd01f | 1968 | */ |
8dd3557a HD |
1969 | if (page == tree_page) { |
1970 | put_page(tree_page); | |
31dbd01f IE |
1971 | return NULL; |
1972 | } | |
1973 | ||
8dd3557a | 1974 | ret = memcmp_pages(page, tree_page); |
31dbd01f IE |
1975 | |
1976 | parent = *new; | |
1977 | if (ret < 0) { | |
8dd3557a | 1978 | put_page(tree_page); |
31dbd01f IE |
1979 | new = &parent->rb_left; |
1980 | } else if (ret > 0) { | |
8dd3557a | 1981 | put_page(tree_page); |
31dbd01f | 1982 | new = &parent->rb_right; |
b599cbdf HD |
1983 | } else if (!ksm_merge_across_nodes && |
1984 | page_to_nid(tree_page) != nid) { | |
1985 | /* | |
1986 | * If tree_page has been migrated to another NUMA node, | |
1987 | * it will be flushed out and put in the right unstable | |
1988 | * tree next time: only merge with it when across_nodes. | |
1989 | */ | |
1990 | put_page(tree_page); | |
1991 | return NULL; | |
31dbd01f | 1992 | } else { |
8dd3557a | 1993 | *tree_pagep = tree_page; |
31dbd01f IE |
1994 | return tree_rmap_item; |
1995 | } | |
1996 | } | |
1997 | ||
7b6ba2c7 | 1998 | rmap_item->address |= UNSTABLE_FLAG; |
31dbd01f | 1999 | rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); |
e850dcf5 | 2000 | DO_NUMA(rmap_item->nid = nid); |
31dbd01f | 2001 | rb_link_node(&rmap_item->node, parent, new); |
90bd6fd3 | 2002 | rb_insert_color(&rmap_item->node, root); |
31dbd01f | 2003 | |
473b0ce4 | 2004 | ksm_pages_unshared++; |
31dbd01f IE |
2005 | return NULL; |
2006 | } | |
2007 | ||
2008 | /* | |
2009 | * stable_tree_append - add another rmap_item to the linked list of | |
2010 | * rmap_items hanging off a given node of the stable tree, all sharing | |
2011 | * the same ksm page. | |
2012 | */ | |
21fbd591 QZ |
2013 | static void stable_tree_append(struct ksm_rmap_item *rmap_item, |
2014 | struct ksm_stable_node *stable_node, | |
2c653d0e | 2015 | bool max_page_sharing_bypass) |
31dbd01f | 2016 | { |
2c653d0e AA |
2017 | /* |
2018 | * rmap won't find this mapping if we don't insert the | |
2019 | * rmap_item in the right stable_node | |
2020 | * duplicate. page_migration could break later if rmap breaks, | |
2021 | * so we can as well crash here. We really need to check for | |
2022 | * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check | |
457aef94 | 2023 | * for other negative values as an underflow if detected here |
2c653d0e AA |
2024 | * for the first time (and not when decreasing rmap_hlist_len) |
2025 | * would be sign of memory corruption in the stable_node. | |
2026 | */ | |
2027 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
2028 | ||
2029 | stable_node->rmap_hlist_len++; | |
2030 | if (!max_page_sharing_bypass) | |
2031 | /* possibly non fatal but unexpected overflow, only warn */ | |
2032 | WARN_ON_ONCE(stable_node->rmap_hlist_len > | |
2033 | ksm_max_page_sharing); | |
2034 | ||
7b6ba2c7 | 2035 | rmap_item->head = stable_node; |
31dbd01f | 2036 | rmap_item->address |= STABLE_FLAG; |
7b6ba2c7 | 2037 | hlist_add_head(&rmap_item->hlist, &stable_node->hlist); |
e178dfde | 2038 | |
7b6ba2c7 HD |
2039 | if (rmap_item->hlist.next) |
2040 | ksm_pages_sharing++; | |
2041 | else | |
2042 | ksm_pages_shared++; | |
76093853 | 2043 | |
2044 | rmap_item->mm->ksm_merging_pages++; | |
31dbd01f IE |
2045 | } |
2046 | ||
2047 | /* | |
81464e30 HD |
2048 | * cmp_and_merge_page - first see if page can be merged into the stable tree; |
2049 | * if not, compare checksum to previous and if it's the same, see if page can | |
2050 | * be inserted into the unstable tree, or merged with a page already there and | |
2051 | * both transferred to the stable tree. | |
31dbd01f IE |
2052 | * |
2053 | * @page: the page that we are searching identical page to. | |
2054 | * @rmap_item: the reverse mapping into the virtual address of this page | |
2055 | */ | |
21fbd591 | 2056 | static void cmp_and_merge_page(struct page *page, struct ksm_rmap_item *rmap_item) |
31dbd01f | 2057 | { |
4b22927f | 2058 | struct mm_struct *mm = rmap_item->mm; |
21fbd591 | 2059 | struct ksm_rmap_item *tree_rmap_item; |
8dd3557a | 2060 | struct page *tree_page = NULL; |
21fbd591 | 2061 | struct ksm_stable_node *stable_node; |
8dd3557a | 2062 | struct page *kpage; |
31dbd01f IE |
2063 | unsigned int checksum; |
2064 | int err; | |
2c653d0e | 2065 | bool max_page_sharing_bypass = false; |
31dbd01f | 2066 | |
4146d2d6 HD |
2067 | stable_node = page_stable_node(page); |
2068 | if (stable_node) { | |
2069 | if (stable_node->head != &migrate_nodes && | |
2c653d0e AA |
2070 | get_kpfn_nid(READ_ONCE(stable_node->kpfn)) != |
2071 | NUMA(stable_node->nid)) { | |
2072 | stable_node_dup_del(stable_node); | |
4146d2d6 HD |
2073 | stable_node->head = &migrate_nodes; |
2074 | list_add(&stable_node->list, stable_node->head); | |
2075 | } | |
2076 | if (stable_node->head != &migrate_nodes && | |
2077 | rmap_item->head == stable_node) | |
2078 | return; | |
2c653d0e AA |
2079 | /* |
2080 | * If it's a KSM fork, allow it to go over the sharing limit | |
2081 | * without warnings. | |
2082 | */ | |
2083 | if (!is_page_sharing_candidate(stable_node)) | |
2084 | max_page_sharing_bypass = true; | |
4146d2d6 | 2085 | } |
31dbd01f IE |
2086 | |
2087 | /* We first start with searching the page inside the stable tree */ | |
62b61f61 | 2088 | kpage = stable_tree_search(page); |
4146d2d6 HD |
2089 | if (kpage == page && rmap_item->head == stable_node) { |
2090 | put_page(kpage); | |
2091 | return; | |
2092 | } | |
2093 | ||
2094 | remove_rmap_item_from_tree(rmap_item); | |
2095 | ||
62b61f61 | 2096 | if (kpage) { |
2cee57d1 YS |
2097 | if (PTR_ERR(kpage) == -EBUSY) |
2098 | return; | |
2099 | ||
08beca44 | 2100 | err = try_to_merge_with_ksm_page(rmap_item, page, kpage); |
31dbd01f IE |
2101 | if (!err) { |
2102 | /* | |
2103 | * The page was successfully merged: | |
2104 | * add its rmap_item to the stable tree. | |
2105 | */ | |
5ad64688 | 2106 | lock_page(kpage); |
2c653d0e AA |
2107 | stable_tree_append(rmap_item, page_stable_node(kpage), |
2108 | max_page_sharing_bypass); | |
5ad64688 | 2109 | unlock_page(kpage); |
31dbd01f | 2110 | } |
8dd3557a | 2111 | put_page(kpage); |
31dbd01f IE |
2112 | return; |
2113 | } | |
2114 | ||
2115 | /* | |
4035c07a HD |
2116 | * If the hash value of the page has changed from the last time |
2117 | * we calculated it, this page is changing frequently: therefore we | |
2118 | * don't want to insert it in the unstable tree, and we don't want | |
2119 | * to waste our time searching for something identical to it there. | |
31dbd01f IE |
2120 | */ |
2121 | checksum = calc_checksum(page); | |
2122 | if (rmap_item->oldchecksum != checksum) { | |
2123 | rmap_item->oldchecksum = checksum; | |
2124 | return; | |
2125 | } | |
2126 | ||
e86c59b1 CI |
2127 | /* |
2128 | * Same checksum as an empty page. We attempt to merge it with the | |
2129 | * appropriate zero page if the user enabled this via sysfs. | |
2130 | */ | |
2131 | if (ksm_use_zero_pages && (checksum == zero_checksum)) { | |
2132 | struct vm_area_struct *vma; | |
2133 | ||
d8ed45c5 | 2134 | mmap_read_lock(mm); |
4b22927f | 2135 | vma = find_mergeable_vma(mm, rmap_item->address); |
56df70a6 MS |
2136 | if (vma) { |
2137 | err = try_to_merge_one_page(vma, page, | |
2138 | ZERO_PAGE(rmap_item->address)); | |
2139 | } else { | |
2140 | /* | |
2141 | * If the vma is out of date, we do not need to | |
2142 | * continue. | |
2143 | */ | |
2144 | err = 0; | |
2145 | } | |
d8ed45c5 | 2146 | mmap_read_unlock(mm); |
e86c59b1 CI |
2147 | /* |
2148 | * In case of failure, the page was not really empty, so we | |
2149 | * need to continue. Otherwise we're done. | |
2150 | */ | |
2151 | if (!err) | |
2152 | return; | |
2153 | } | |
8dd3557a HD |
2154 | tree_rmap_item = |
2155 | unstable_tree_search_insert(rmap_item, page, &tree_page); | |
31dbd01f | 2156 | if (tree_rmap_item) { |
77da2ba0 CI |
2157 | bool split; |
2158 | ||
8dd3557a HD |
2159 | kpage = try_to_merge_two_pages(rmap_item, page, |
2160 | tree_rmap_item, tree_page); | |
77da2ba0 CI |
2161 | /* |
2162 | * If both pages we tried to merge belong to the same compound | |
2163 | * page, then we actually ended up increasing the reference | |
2164 | * count of the same compound page twice, and split_huge_page | |
2165 | * failed. | |
2166 | * Here we set a flag if that happened, and we use it later to | |
2167 | * try split_huge_page again. Since we call put_page right | |
2168 | * afterwards, the reference count will be correct and | |
2169 | * split_huge_page should succeed. | |
2170 | */ | |
2171 | split = PageTransCompound(page) | |
2172 | && compound_head(page) == compound_head(tree_page); | |
8dd3557a | 2173 | put_page(tree_page); |
8dd3557a | 2174 | if (kpage) { |
bc56620b HD |
2175 | /* |
2176 | * The pages were successfully merged: insert new | |
2177 | * node in the stable tree and add both rmap_items. | |
2178 | */ | |
5ad64688 | 2179 | lock_page(kpage); |
7b6ba2c7 HD |
2180 | stable_node = stable_tree_insert(kpage); |
2181 | if (stable_node) { | |
2c653d0e AA |
2182 | stable_tree_append(tree_rmap_item, stable_node, |
2183 | false); | |
2184 | stable_tree_append(rmap_item, stable_node, | |
2185 | false); | |
7b6ba2c7 | 2186 | } |
5ad64688 | 2187 | unlock_page(kpage); |
7b6ba2c7 | 2188 | |
31dbd01f IE |
2189 | /* |
2190 | * If we fail to insert the page into the stable tree, | |
2191 | * we will have 2 virtual addresses that are pointing | |
2192 | * to a ksm page left outside the stable tree, | |
2193 | * in which case we need to break_cow on both. | |
2194 | */ | |
7b6ba2c7 | 2195 | if (!stable_node) { |
8dd3557a HD |
2196 | break_cow(tree_rmap_item); |
2197 | break_cow(rmap_item); | |
31dbd01f | 2198 | } |
77da2ba0 CI |
2199 | } else if (split) { |
2200 | /* | |
2201 | * We are here if we tried to merge two pages and | |
2202 | * failed because they both belonged to the same | |
2203 | * compound page. We will split the page now, but no | |
2204 | * merging will take place. | |
2205 | * We do not want to add the cost of a full lock; if | |
2206 | * the page is locked, it is better to skip it and | |
2207 | * perhaps try again later. | |
2208 | */ | |
2209 | if (!trylock_page(page)) | |
2210 | return; | |
2211 | split_huge_page(page); | |
2212 | unlock_page(page); | |
31dbd01f | 2213 | } |
31dbd01f IE |
2214 | } |
2215 | } | |
2216 | ||
21fbd591 QZ |
2217 | static struct ksm_rmap_item *get_next_rmap_item(struct ksm_mm_slot *mm_slot, |
2218 | struct ksm_rmap_item **rmap_list, | |
31dbd01f IE |
2219 | unsigned long addr) |
2220 | { | |
21fbd591 | 2221 | struct ksm_rmap_item *rmap_item; |
31dbd01f | 2222 | |
6514d511 HD |
2223 | while (*rmap_list) { |
2224 | rmap_item = *rmap_list; | |
93d17715 | 2225 | if ((rmap_item->address & PAGE_MASK) == addr) |
31dbd01f | 2226 | return rmap_item; |
31dbd01f IE |
2227 | if (rmap_item->address > addr) |
2228 | break; | |
6514d511 | 2229 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 2230 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
2231 | free_rmap_item(rmap_item); |
2232 | } | |
2233 | ||
2234 | rmap_item = alloc_rmap_item(); | |
2235 | if (rmap_item) { | |
2236 | /* It has already been zeroed */ | |
58730ab6 | 2237 | rmap_item->mm = mm_slot->slot.mm; |
cb4df4ca | 2238 | rmap_item->mm->ksm_rmap_items++; |
31dbd01f | 2239 | rmap_item->address = addr; |
6514d511 HD |
2240 | rmap_item->rmap_list = *rmap_list; |
2241 | *rmap_list = rmap_item; | |
31dbd01f IE |
2242 | } |
2243 | return rmap_item; | |
2244 | } | |
2245 | ||
21fbd591 | 2246 | static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page) |
31dbd01f IE |
2247 | { |
2248 | struct mm_struct *mm; | |
58730ab6 QZ |
2249 | struct ksm_mm_slot *mm_slot; |
2250 | struct mm_slot *slot; | |
31dbd01f | 2251 | struct vm_area_struct *vma; |
21fbd591 | 2252 | struct ksm_rmap_item *rmap_item; |
a5f18ba0 | 2253 | struct vma_iterator vmi; |
90bd6fd3 | 2254 | int nid; |
31dbd01f | 2255 | |
58730ab6 | 2256 | if (list_empty(&ksm_mm_head.slot.mm_node)) |
31dbd01f IE |
2257 | return NULL; |
2258 | ||
58730ab6 QZ |
2259 | mm_slot = ksm_scan.mm_slot; |
2260 | if (mm_slot == &ksm_mm_head) { | |
2919bfd0 HD |
2261 | /* |
2262 | * A number of pages can hang around indefinitely on per-cpu | |
2263 | * pagevecs, raised page count preventing write_protect_page | |
2264 | * from merging them. Though it doesn't really matter much, | |
2265 | * it is puzzling to see some stuck in pages_volatile until | |
2266 | * other activity jostles them out, and they also prevented | |
2267 | * LTP's KSM test from succeeding deterministically; so drain | |
2268 | * them here (here rather than on entry to ksm_do_scan(), | |
2269 | * so we don't IPI too often when pages_to_scan is set low). | |
2270 | */ | |
2271 | lru_add_drain_all(); | |
2272 | ||
4146d2d6 HD |
2273 | /* |
2274 | * Whereas stale stable_nodes on the stable_tree itself | |
2275 | * get pruned in the regular course of stable_tree_search(), | |
2276 | * those moved out to the migrate_nodes list can accumulate: | |
2277 | * so prune them once before each full scan. | |
2278 | */ | |
2279 | if (!ksm_merge_across_nodes) { | |
21fbd591 | 2280 | struct ksm_stable_node *stable_node, *next; |
4146d2d6 HD |
2281 | struct page *page; |
2282 | ||
03640418 GT |
2283 | list_for_each_entry_safe(stable_node, next, |
2284 | &migrate_nodes, list) { | |
2cee57d1 YS |
2285 | page = get_ksm_page(stable_node, |
2286 | GET_KSM_PAGE_NOLOCK); | |
4146d2d6 HD |
2287 | if (page) |
2288 | put_page(page); | |
2289 | cond_resched(); | |
2290 | } | |
2291 | } | |
2292 | ||
ef53d16c | 2293 | for (nid = 0; nid < ksm_nr_node_ids; nid++) |
90bd6fd3 | 2294 | root_unstable_tree[nid] = RB_ROOT; |
31dbd01f IE |
2295 | |
2296 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
2297 | slot = list_entry(mm_slot->slot.mm_node.next, |
2298 | struct mm_slot, mm_node); | |
2299 | mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
2300 | ksm_scan.mm_slot = mm_slot; | |
31dbd01f | 2301 | spin_unlock(&ksm_mmlist_lock); |
2b472611 HD |
2302 | /* |
2303 | * Although we tested list_empty() above, a racing __ksm_exit | |
2304 | * of the last mm on the list may have removed it since then. | |
2305 | */ | |
58730ab6 | 2306 | if (mm_slot == &ksm_mm_head) |
2b472611 | 2307 | return NULL; |
31dbd01f IE |
2308 | next_mm: |
2309 | ksm_scan.address = 0; | |
58730ab6 | 2310 | ksm_scan.rmap_list = &mm_slot->rmap_list; |
31dbd01f IE |
2311 | } |
2312 | ||
58730ab6 | 2313 | slot = &mm_slot->slot; |
31dbd01f | 2314 | mm = slot->mm; |
a5f18ba0 MWO |
2315 | vma_iter_init(&vmi, mm, ksm_scan.address); |
2316 | ||
d8ed45c5 | 2317 | mmap_read_lock(mm); |
9ba69294 | 2318 | if (ksm_test_exit(mm)) |
a5f18ba0 | 2319 | goto no_vmas; |
9ba69294 | 2320 | |
a5f18ba0 | 2321 | for_each_vma(vmi, vma) { |
31dbd01f IE |
2322 | if (!(vma->vm_flags & VM_MERGEABLE)) |
2323 | continue; | |
2324 | if (ksm_scan.address < vma->vm_start) | |
2325 | ksm_scan.address = vma->vm_start; | |
2326 | if (!vma->anon_vma) | |
2327 | ksm_scan.address = vma->vm_end; | |
2328 | ||
2329 | while (ksm_scan.address < vma->vm_end) { | |
9ba69294 HD |
2330 | if (ksm_test_exit(mm)) |
2331 | break; | |
31dbd01f | 2332 | *page = follow_page(vma, ksm_scan.address, FOLL_GET); |
f7091ed6 | 2333 | if (IS_ERR_OR_NULL(*page)) { |
21ae5b01 AA |
2334 | ksm_scan.address += PAGE_SIZE; |
2335 | cond_resched(); | |
2336 | continue; | |
2337 | } | |
f7091ed6 HW |
2338 | if (is_zone_device_page(*page)) |
2339 | goto next_page; | |
f765f540 | 2340 | if (PageAnon(*page)) { |
31dbd01f IE |
2341 | flush_anon_page(vma, *page, ksm_scan.address); |
2342 | flush_dcache_page(*page); | |
58730ab6 | 2343 | rmap_item = get_next_rmap_item(mm_slot, |
6514d511 | 2344 | ksm_scan.rmap_list, ksm_scan.address); |
31dbd01f | 2345 | if (rmap_item) { |
6514d511 HD |
2346 | ksm_scan.rmap_list = |
2347 | &rmap_item->rmap_list; | |
31dbd01f IE |
2348 | ksm_scan.address += PAGE_SIZE; |
2349 | } else | |
2350 | put_page(*page); | |
d8ed45c5 | 2351 | mmap_read_unlock(mm); |
31dbd01f IE |
2352 | return rmap_item; |
2353 | } | |
f7091ed6 | 2354 | next_page: |
21ae5b01 | 2355 | put_page(*page); |
31dbd01f IE |
2356 | ksm_scan.address += PAGE_SIZE; |
2357 | cond_resched(); | |
2358 | } | |
2359 | } | |
2360 | ||
9ba69294 | 2361 | if (ksm_test_exit(mm)) { |
a5f18ba0 | 2362 | no_vmas: |
9ba69294 | 2363 | ksm_scan.address = 0; |
58730ab6 | 2364 | ksm_scan.rmap_list = &mm_slot->rmap_list; |
9ba69294 | 2365 | } |
31dbd01f IE |
2366 | /* |
2367 | * Nuke all the rmap_items that are above this current rmap: | |
2368 | * because there were no VM_MERGEABLE vmas with such addresses. | |
2369 | */ | |
420be4ed | 2370 | remove_trailing_rmap_items(ksm_scan.rmap_list); |
31dbd01f IE |
2371 | |
2372 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
2373 | slot = list_entry(mm_slot->slot.mm_node.next, |
2374 | struct mm_slot, mm_node); | |
2375 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
cd551f97 HD |
2376 | if (ksm_scan.address == 0) { |
2377 | /* | |
c1e8d7c6 | 2378 | * We've completed a full scan of all vmas, holding mmap_lock |
cd551f97 HD |
2379 | * throughout, and found no VM_MERGEABLE: so do the same as |
2380 | * __ksm_exit does to remove this mm from all our lists now. | |
9ba69294 HD |
2381 | * This applies either when cleaning up after __ksm_exit |
2382 | * (but beware: we can reach here even before __ksm_exit), | |
2383 | * or when all VM_MERGEABLE areas have been unmapped (and | |
c1e8d7c6 | 2384 | * mmap_lock then protects against race with MADV_MERGEABLE). |
cd551f97 | 2385 | */ |
58730ab6 QZ |
2386 | hash_del(&mm_slot->slot.hash); |
2387 | list_del(&mm_slot->slot.mm_node); | |
9ba69294 HD |
2388 | spin_unlock(&ksm_mmlist_lock); |
2389 | ||
58730ab6 | 2390 | mm_slot_free(mm_slot_cache, mm_slot); |
cd551f97 | 2391 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
d8ed45c5 | 2392 | mmap_read_unlock(mm); |
9ba69294 HD |
2393 | mmdrop(mm); |
2394 | } else { | |
d8ed45c5 | 2395 | mmap_read_unlock(mm); |
7496fea9 | 2396 | /* |
3e4e28c5 | 2397 | * mmap_read_unlock(mm) first because after |
7496fea9 ZC |
2398 | * spin_unlock(&ksm_mmlist_lock) run, the "mm" may |
2399 | * already have been freed under us by __ksm_exit() | |
2400 | * because the "mm_slot" is still hashed and | |
2401 | * ksm_scan.mm_slot doesn't point to it anymore. | |
2402 | */ | |
2403 | spin_unlock(&ksm_mmlist_lock); | |
cd551f97 | 2404 | } |
31dbd01f IE |
2405 | |
2406 | /* Repeat until we've completed scanning the whole list */ | |
58730ab6 QZ |
2407 | mm_slot = ksm_scan.mm_slot; |
2408 | if (mm_slot != &ksm_mm_head) | |
31dbd01f IE |
2409 | goto next_mm; |
2410 | ||
31dbd01f IE |
2411 | ksm_scan.seqnr++; |
2412 | return NULL; | |
2413 | } | |
2414 | ||
2415 | /** | |
2416 | * ksm_do_scan - the ksm scanner main worker function. | |
b7701a5f | 2417 | * @scan_npages: number of pages we want to scan before we return. |
31dbd01f IE |
2418 | */ |
2419 | static void ksm_do_scan(unsigned int scan_npages) | |
2420 | { | |
21fbd591 | 2421 | struct ksm_rmap_item *rmap_item; |
3f649ab7 | 2422 | struct page *page; |
31dbd01f | 2423 | |
878aee7d | 2424 | while (scan_npages-- && likely(!freezing(current))) { |
31dbd01f IE |
2425 | cond_resched(); |
2426 | rmap_item = scan_get_next_rmap_item(&page); | |
2427 | if (!rmap_item) | |
2428 | return; | |
4146d2d6 | 2429 | cmp_and_merge_page(page, rmap_item); |
31dbd01f IE |
2430 | put_page(page); |
2431 | } | |
2432 | } | |
2433 | ||
6e158384 HD |
2434 | static int ksmd_should_run(void) |
2435 | { | |
58730ab6 | 2436 | return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.slot.mm_node); |
6e158384 HD |
2437 | } |
2438 | ||
31dbd01f IE |
2439 | static int ksm_scan_thread(void *nothing) |
2440 | { | |
fcf9a0ef KT |
2441 | unsigned int sleep_ms; |
2442 | ||
878aee7d | 2443 | set_freezable(); |
339aa624 | 2444 | set_user_nice(current, 5); |
31dbd01f IE |
2445 | |
2446 | while (!kthread_should_stop()) { | |
6e158384 | 2447 | mutex_lock(&ksm_thread_mutex); |
ef4d43a8 | 2448 | wait_while_offlining(); |
6e158384 | 2449 | if (ksmd_should_run()) |
31dbd01f | 2450 | ksm_do_scan(ksm_thread_pages_to_scan); |
6e158384 HD |
2451 | mutex_unlock(&ksm_thread_mutex); |
2452 | ||
878aee7d AA |
2453 | try_to_freeze(); |
2454 | ||
6e158384 | 2455 | if (ksmd_should_run()) { |
fcf9a0ef KT |
2456 | sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs); |
2457 | wait_event_interruptible_timeout(ksm_iter_wait, | |
2458 | sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs), | |
2459 | msecs_to_jiffies(sleep_ms)); | |
31dbd01f | 2460 | } else { |
878aee7d | 2461 | wait_event_freezable(ksm_thread_wait, |
6e158384 | 2462 | ksmd_should_run() || kthread_should_stop()); |
31dbd01f IE |
2463 | } |
2464 | } | |
2465 | return 0; | |
2466 | } | |
2467 | ||
f8af4da3 HD |
2468 | int ksm_madvise(struct vm_area_struct *vma, unsigned long start, |
2469 | unsigned long end, int advice, unsigned long *vm_flags) | |
2470 | { | |
2471 | struct mm_struct *mm = vma->vm_mm; | |
d952b791 | 2472 | int err; |
f8af4da3 HD |
2473 | |
2474 | switch (advice) { | |
2475 | case MADV_MERGEABLE: | |
2476 | /* | |
2477 | * Be somewhat over-protective for now! | |
2478 | */ | |
2479 | if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE | | |
2480 | VM_PFNMAP | VM_IO | VM_DONTEXPAND | | |
0661a336 | 2481 | VM_HUGETLB | VM_MIXEDMAP)) |
f8af4da3 HD |
2482 | return 0; /* just ignore the advice */ |
2483 | ||
e1fb4a08 DJ |
2484 | if (vma_is_dax(vma)) |
2485 | return 0; | |
2486 | ||
12564485 SA |
2487 | #ifdef VM_SAO |
2488 | if (*vm_flags & VM_SAO) | |
2489 | return 0; | |
2490 | #endif | |
74a04967 KA |
2491 | #ifdef VM_SPARC_ADI |
2492 | if (*vm_flags & VM_SPARC_ADI) | |
2493 | return 0; | |
2494 | #endif | |
cc2383ec | 2495 | |
d952b791 HD |
2496 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { |
2497 | err = __ksm_enter(mm); | |
2498 | if (err) | |
2499 | return err; | |
2500 | } | |
f8af4da3 HD |
2501 | |
2502 | *vm_flags |= VM_MERGEABLE; | |
2503 | break; | |
2504 | ||
2505 | case MADV_UNMERGEABLE: | |
2506 | if (!(*vm_flags & VM_MERGEABLE)) | |
2507 | return 0; /* just ignore the advice */ | |
2508 | ||
d952b791 HD |
2509 | if (vma->anon_vma) { |
2510 | err = unmerge_ksm_pages(vma, start, end); | |
2511 | if (err) | |
2512 | return err; | |
2513 | } | |
f8af4da3 HD |
2514 | |
2515 | *vm_flags &= ~VM_MERGEABLE; | |
2516 | break; | |
2517 | } | |
2518 | ||
2519 | return 0; | |
2520 | } | |
33cf1707 | 2521 | EXPORT_SYMBOL_GPL(ksm_madvise); |
f8af4da3 HD |
2522 | |
2523 | int __ksm_enter(struct mm_struct *mm) | |
2524 | { | |
21fbd591 | 2525 | struct ksm_mm_slot *mm_slot; |
58730ab6 | 2526 | struct mm_slot *slot; |
6e158384 HD |
2527 | int needs_wakeup; |
2528 | ||
58730ab6 | 2529 | mm_slot = mm_slot_alloc(mm_slot_cache); |
31dbd01f IE |
2530 | if (!mm_slot) |
2531 | return -ENOMEM; | |
2532 | ||
58730ab6 QZ |
2533 | slot = &mm_slot->slot; |
2534 | ||
6e158384 | 2535 | /* Check ksm_run too? Would need tighter locking */ |
58730ab6 | 2536 | needs_wakeup = list_empty(&ksm_mm_head.slot.mm_node); |
6e158384 | 2537 | |
31dbd01f | 2538 | spin_lock(&ksm_mmlist_lock); |
58730ab6 | 2539 | mm_slot_insert(mm_slots_hash, mm, slot); |
31dbd01f | 2540 | /* |
cbf86cfe HD |
2541 | * When KSM_RUN_MERGE (or KSM_RUN_STOP), |
2542 | * insert just behind the scanning cursor, to let the area settle | |
31dbd01f IE |
2543 | * down a little; when fork is followed by immediate exec, we don't |
2544 | * want ksmd to waste time setting up and tearing down an rmap_list. | |
cbf86cfe HD |
2545 | * |
2546 | * But when KSM_RUN_UNMERGE, it's important to insert ahead of its | |
2547 | * scanning cursor, otherwise KSM pages in newly forked mms will be | |
2548 | * missed: then we might as well insert at the end of the list. | |
31dbd01f | 2549 | */ |
cbf86cfe | 2550 | if (ksm_run & KSM_RUN_UNMERGE) |
58730ab6 | 2551 | list_add_tail(&slot->mm_node, &ksm_mm_head.slot.mm_node); |
cbf86cfe | 2552 | else |
58730ab6 | 2553 | list_add_tail(&slot->mm_node, &ksm_scan.mm_slot->slot.mm_node); |
31dbd01f IE |
2554 | spin_unlock(&ksm_mmlist_lock); |
2555 | ||
f8af4da3 | 2556 | set_bit(MMF_VM_MERGEABLE, &mm->flags); |
f1f10076 | 2557 | mmgrab(mm); |
6e158384 HD |
2558 | |
2559 | if (needs_wakeup) | |
2560 | wake_up_interruptible(&ksm_thread_wait); | |
2561 | ||
f8af4da3 HD |
2562 | return 0; |
2563 | } | |
2564 | ||
1c2fb7a4 | 2565 | void __ksm_exit(struct mm_struct *mm) |
f8af4da3 | 2566 | { |
21fbd591 | 2567 | struct ksm_mm_slot *mm_slot; |
58730ab6 | 2568 | struct mm_slot *slot; |
9ba69294 | 2569 | int easy_to_free = 0; |
cd551f97 | 2570 | |
31dbd01f | 2571 | /* |
9ba69294 HD |
2572 | * This process is exiting: if it's straightforward (as is the |
2573 | * case when ksmd was never running), free mm_slot immediately. | |
2574 | * But if it's at the cursor or has rmap_items linked to it, use | |
c1e8d7c6 | 2575 | * mmap_lock to synchronize with any break_cows before pagetables |
9ba69294 HD |
2576 | * are freed, and leave the mm_slot on the list for ksmd to free. |
2577 | * Beware: ksm may already have noticed it exiting and freed the slot. | |
31dbd01f | 2578 | */ |
9ba69294 | 2579 | |
cd551f97 | 2580 | spin_lock(&ksm_mmlist_lock); |
58730ab6 QZ |
2581 | slot = mm_slot_lookup(mm_slots_hash, mm); |
2582 | mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
9ba69294 | 2583 | if (mm_slot && ksm_scan.mm_slot != mm_slot) { |
6514d511 | 2584 | if (!mm_slot->rmap_list) { |
58730ab6 QZ |
2585 | hash_del(&slot->hash); |
2586 | list_del(&slot->mm_node); | |
9ba69294 HD |
2587 | easy_to_free = 1; |
2588 | } else { | |
58730ab6 QZ |
2589 | list_move(&slot->mm_node, |
2590 | &ksm_scan.mm_slot->slot.mm_node); | |
9ba69294 | 2591 | } |
cd551f97 | 2592 | } |
cd551f97 HD |
2593 | spin_unlock(&ksm_mmlist_lock); |
2594 | ||
9ba69294 | 2595 | if (easy_to_free) { |
58730ab6 | 2596 | mm_slot_free(mm_slot_cache, mm_slot); |
9ba69294 HD |
2597 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
2598 | mmdrop(mm); | |
2599 | } else if (mm_slot) { | |
d8ed45c5 ML |
2600 | mmap_write_lock(mm); |
2601 | mmap_write_unlock(mm); | |
9ba69294 | 2602 | } |
31dbd01f IE |
2603 | } |
2604 | ||
cbf86cfe | 2605 | struct page *ksm_might_need_to_copy(struct page *page, |
5ad64688 HD |
2606 | struct vm_area_struct *vma, unsigned long address) |
2607 | { | |
e05b3453 MWO |
2608 | struct folio *folio = page_folio(page); |
2609 | struct anon_vma *anon_vma = folio_anon_vma(folio); | |
5ad64688 HD |
2610 | struct page *new_page; |
2611 | ||
cbf86cfe HD |
2612 | if (PageKsm(page)) { |
2613 | if (page_stable_node(page) && | |
2614 | !(ksm_run & KSM_RUN_UNMERGE)) | |
2615 | return page; /* no need to copy it */ | |
2616 | } else if (!anon_vma) { | |
2617 | return page; /* no need to copy it */ | |
e1c63e11 NS |
2618 | } else if (page->index == linear_page_index(vma, address) && |
2619 | anon_vma->root == vma->anon_vma->root) { | |
cbf86cfe HD |
2620 | return page; /* still no need to copy it */ |
2621 | } | |
2622 | if (!PageUptodate(page)) | |
2623 | return page; /* let do_swap_page report the error */ | |
2624 | ||
5ad64688 | 2625 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
8f425e4e MWO |
2626 | if (new_page && |
2627 | mem_cgroup_charge(page_folio(new_page), vma->vm_mm, GFP_KERNEL)) { | |
62fdb163 HD |
2628 | put_page(new_page); |
2629 | new_page = NULL; | |
2630 | } | |
5ad64688 HD |
2631 | if (new_page) { |
2632 | copy_user_highpage(new_page, page, address, vma); | |
2633 | ||
2634 | SetPageDirty(new_page); | |
2635 | __SetPageUptodate(new_page); | |
48c935ad | 2636 | __SetPageLocked(new_page); |
4d45c3af YY |
2637 | #ifdef CONFIG_SWAP |
2638 | count_vm_event(KSM_SWPIN_COPY); | |
2639 | #endif | |
5ad64688 HD |
2640 | } |
2641 | ||
5ad64688 HD |
2642 | return new_page; |
2643 | } | |
2644 | ||
6d4675e6 | 2645 | void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc) |
e9995ef9 | 2646 | { |
21fbd591 QZ |
2647 | struct ksm_stable_node *stable_node; |
2648 | struct ksm_rmap_item *rmap_item; | |
e9995ef9 HD |
2649 | int search_new_forks = 0; |
2650 | ||
2f031c6f | 2651 | VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio); |
9f32624b JK |
2652 | |
2653 | /* | |
2654 | * Rely on the page lock to protect against concurrent modifications | |
2655 | * to that page's node of the stable tree. | |
2656 | */ | |
2f031c6f | 2657 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
e9995ef9 | 2658 | |
2f031c6f | 2659 | stable_node = folio_stable_node(folio); |
e9995ef9 | 2660 | if (!stable_node) |
1df631ae | 2661 | return; |
e9995ef9 | 2662 | again: |
b67bfe0d | 2663 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
e9995ef9 | 2664 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
5beb4930 | 2665 | struct anon_vma_chain *vmac; |
e9995ef9 HD |
2666 | struct vm_area_struct *vma; |
2667 | ||
ad12695f | 2668 | cond_resched(); |
6d4675e6 MK |
2669 | if (!anon_vma_trylock_read(anon_vma)) { |
2670 | if (rwc->try_lock) { | |
2671 | rwc->contended = true; | |
2672 | return; | |
2673 | } | |
2674 | anon_vma_lock_read(anon_vma); | |
2675 | } | |
bf181b9f ML |
2676 | anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, |
2677 | 0, ULONG_MAX) { | |
1105a2fc JH |
2678 | unsigned long addr; |
2679 | ||
ad12695f | 2680 | cond_resched(); |
5beb4930 | 2681 | vma = vmac->vma; |
1105a2fc JH |
2682 | |
2683 | /* Ignore the stable/unstable/sqnr flags */ | |
cd7fae26 | 2684 | addr = rmap_item->address & PAGE_MASK; |
1105a2fc JH |
2685 | |
2686 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
e9995ef9 HD |
2687 | continue; |
2688 | /* | |
2689 | * Initially we examine only the vma which covers this | |
2690 | * rmap_item; but later, if there is still work to do, | |
2691 | * we examine covering vmas in other mms: in case they | |
2692 | * were forked from the original since ksmd passed. | |
2693 | */ | |
2694 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
2695 | continue; | |
2696 | ||
0dd1c7bb JK |
2697 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
2698 | continue; | |
2699 | ||
2f031c6f | 2700 | if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) { |
b6b19f25 | 2701 | anon_vma_unlock_read(anon_vma); |
1df631ae | 2702 | return; |
e9995ef9 | 2703 | } |
2f031c6f | 2704 | if (rwc->done && rwc->done(folio)) { |
0dd1c7bb | 2705 | anon_vma_unlock_read(anon_vma); |
1df631ae | 2706 | return; |
0dd1c7bb | 2707 | } |
e9995ef9 | 2708 | } |
b6b19f25 | 2709 | anon_vma_unlock_read(anon_vma); |
e9995ef9 HD |
2710 | } |
2711 | if (!search_new_forks++) | |
2712 | goto again; | |
e9995ef9 HD |
2713 | } |
2714 | ||
52629506 | 2715 | #ifdef CONFIG_MIGRATION |
19138349 | 2716 | void folio_migrate_ksm(struct folio *newfolio, struct folio *folio) |
e9995ef9 | 2717 | { |
21fbd591 | 2718 | struct ksm_stable_node *stable_node; |
e9995ef9 | 2719 | |
19138349 MWO |
2720 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
2721 | VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio); | |
2722 | VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio); | |
e9995ef9 | 2723 | |
19138349 | 2724 | stable_node = folio_stable_node(folio); |
e9995ef9 | 2725 | if (stable_node) { |
19138349 MWO |
2726 | VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio); |
2727 | stable_node->kpfn = folio_pfn(newfolio); | |
c8d6553b | 2728 | /* |
19138349 | 2729 | * newfolio->mapping was set in advance; now we need smp_wmb() |
c8d6553b | 2730 | * to make sure that the new stable_node->kpfn is visible |
19138349 MWO |
2731 | * to get_ksm_page() before it can see that folio->mapping |
2732 | * has gone stale (or that folio_test_swapcache has been cleared). | |
c8d6553b HD |
2733 | */ |
2734 | smp_wmb(); | |
19138349 | 2735 | set_page_stable_node(&folio->page, NULL); |
e9995ef9 HD |
2736 | } |
2737 | } | |
2738 | #endif /* CONFIG_MIGRATION */ | |
2739 | ||
62b61f61 | 2740 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 HD |
2741 | static void wait_while_offlining(void) |
2742 | { | |
2743 | while (ksm_run & KSM_RUN_OFFLINE) { | |
2744 | mutex_unlock(&ksm_thread_mutex); | |
2745 | wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), | |
74316201 | 2746 | TASK_UNINTERRUPTIBLE); |
ef4d43a8 HD |
2747 | mutex_lock(&ksm_thread_mutex); |
2748 | } | |
2749 | } | |
2750 | ||
21fbd591 | 2751 | static bool stable_node_dup_remove_range(struct ksm_stable_node *stable_node, |
2c653d0e AA |
2752 | unsigned long start_pfn, |
2753 | unsigned long end_pfn) | |
2754 | { | |
2755 | if (stable_node->kpfn >= start_pfn && | |
2756 | stable_node->kpfn < end_pfn) { | |
2757 | /* | |
2758 | * Don't get_ksm_page, page has already gone: | |
2759 | * which is why we keep kpfn instead of page* | |
2760 | */ | |
2761 | remove_node_from_stable_tree(stable_node); | |
2762 | return true; | |
2763 | } | |
2764 | return false; | |
2765 | } | |
2766 | ||
21fbd591 | 2767 | static bool stable_node_chain_remove_range(struct ksm_stable_node *stable_node, |
2c653d0e AA |
2768 | unsigned long start_pfn, |
2769 | unsigned long end_pfn, | |
2770 | struct rb_root *root) | |
2771 | { | |
21fbd591 | 2772 | struct ksm_stable_node *dup; |
2c653d0e AA |
2773 | struct hlist_node *hlist_safe; |
2774 | ||
2775 | if (!is_stable_node_chain(stable_node)) { | |
2776 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
2777 | return stable_node_dup_remove_range(stable_node, start_pfn, | |
2778 | end_pfn); | |
2779 | } | |
2780 | ||
2781 | hlist_for_each_entry_safe(dup, hlist_safe, | |
2782 | &stable_node->hlist, hlist_dup) { | |
2783 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
2784 | stable_node_dup_remove_range(dup, start_pfn, end_pfn); | |
2785 | } | |
2786 | if (hlist_empty(&stable_node->hlist)) { | |
2787 | free_stable_node_chain(stable_node, root); | |
2788 | return true; /* notify caller that tree was rebalanced */ | |
2789 | } else | |
2790 | return false; | |
2791 | } | |
2792 | ||
ee0ea59c HD |
2793 | static void ksm_check_stable_tree(unsigned long start_pfn, |
2794 | unsigned long end_pfn) | |
62b61f61 | 2795 | { |
21fbd591 | 2796 | struct ksm_stable_node *stable_node, *next; |
62b61f61 | 2797 | struct rb_node *node; |
90bd6fd3 | 2798 | int nid; |
62b61f61 | 2799 | |
ef53d16c HD |
2800 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
2801 | node = rb_first(root_stable_tree + nid); | |
ee0ea59c | 2802 | while (node) { |
21fbd591 | 2803 | stable_node = rb_entry(node, struct ksm_stable_node, node); |
2c653d0e AA |
2804 | if (stable_node_chain_remove_range(stable_node, |
2805 | start_pfn, end_pfn, | |
2806 | root_stable_tree + | |
2807 | nid)) | |
ef53d16c | 2808 | node = rb_first(root_stable_tree + nid); |
2c653d0e | 2809 | else |
ee0ea59c HD |
2810 | node = rb_next(node); |
2811 | cond_resched(); | |
90bd6fd3 | 2812 | } |
ee0ea59c | 2813 | } |
03640418 | 2814 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
2815 | if (stable_node->kpfn >= start_pfn && |
2816 | stable_node->kpfn < end_pfn) | |
2817 | remove_node_from_stable_tree(stable_node); | |
2818 | cond_resched(); | |
2819 | } | |
62b61f61 HD |
2820 | } |
2821 | ||
2822 | static int ksm_memory_callback(struct notifier_block *self, | |
2823 | unsigned long action, void *arg) | |
2824 | { | |
2825 | struct memory_notify *mn = arg; | |
62b61f61 HD |
2826 | |
2827 | switch (action) { | |
2828 | case MEM_GOING_OFFLINE: | |
2829 | /* | |
ef4d43a8 HD |
2830 | * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() |
2831 | * and remove_all_stable_nodes() while memory is going offline: | |
2832 | * it is unsafe for them to touch the stable tree at this time. | |
2833 | * But unmerge_ksm_pages(), rmap lookups and other entry points | |
2834 | * which do not need the ksm_thread_mutex are all safe. | |
62b61f61 | 2835 | */ |
ef4d43a8 HD |
2836 | mutex_lock(&ksm_thread_mutex); |
2837 | ksm_run |= KSM_RUN_OFFLINE; | |
2838 | mutex_unlock(&ksm_thread_mutex); | |
62b61f61 HD |
2839 | break; |
2840 | ||
2841 | case MEM_OFFLINE: | |
2842 | /* | |
2843 | * Most of the work is done by page migration; but there might | |
2844 | * be a few stable_nodes left over, still pointing to struct | |
ee0ea59c HD |
2845 | * pages which have been offlined: prune those from the tree, |
2846 | * otherwise get_ksm_page() might later try to access a | |
2847 | * non-existent struct page. | |
62b61f61 | 2848 | */ |
ee0ea59c HD |
2849 | ksm_check_stable_tree(mn->start_pfn, |
2850 | mn->start_pfn + mn->nr_pages); | |
e4a9bc58 | 2851 | fallthrough; |
62b61f61 | 2852 | case MEM_CANCEL_OFFLINE: |
ef4d43a8 HD |
2853 | mutex_lock(&ksm_thread_mutex); |
2854 | ksm_run &= ~KSM_RUN_OFFLINE; | |
62b61f61 | 2855 | mutex_unlock(&ksm_thread_mutex); |
ef4d43a8 HD |
2856 | |
2857 | smp_mb(); /* wake_up_bit advises this */ | |
2858 | wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); | |
62b61f61 HD |
2859 | break; |
2860 | } | |
2861 | return NOTIFY_OK; | |
2862 | } | |
ef4d43a8 HD |
2863 | #else |
2864 | static void wait_while_offlining(void) | |
2865 | { | |
2866 | } | |
62b61f61 HD |
2867 | #endif /* CONFIG_MEMORY_HOTREMOVE */ |
2868 | ||
2ffd8679 HD |
2869 | #ifdef CONFIG_SYSFS |
2870 | /* | |
2871 | * This all compiles without CONFIG_SYSFS, but is a waste of space. | |
2872 | */ | |
2873 | ||
31dbd01f IE |
2874 | #define KSM_ATTR_RO(_name) \ |
2875 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2876 | #define KSM_ATTR(_name) \ | |
1bad2e5c | 2877 | static struct kobj_attribute _name##_attr = __ATTR_RW(_name) |
31dbd01f IE |
2878 | |
2879 | static ssize_t sleep_millisecs_show(struct kobject *kobj, | |
2880 | struct kobj_attribute *attr, char *buf) | |
2881 | { | |
ae7a927d | 2882 | return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs); |
31dbd01f IE |
2883 | } |
2884 | ||
2885 | static ssize_t sleep_millisecs_store(struct kobject *kobj, | |
2886 | struct kobj_attribute *attr, | |
2887 | const char *buf, size_t count) | |
2888 | { | |
dfefd226 | 2889 | unsigned int msecs; |
31dbd01f IE |
2890 | int err; |
2891 | ||
dfefd226 AD |
2892 | err = kstrtouint(buf, 10, &msecs); |
2893 | if (err) | |
31dbd01f IE |
2894 | return -EINVAL; |
2895 | ||
2896 | ksm_thread_sleep_millisecs = msecs; | |
fcf9a0ef | 2897 | wake_up_interruptible(&ksm_iter_wait); |
31dbd01f IE |
2898 | |
2899 | return count; | |
2900 | } | |
2901 | KSM_ATTR(sleep_millisecs); | |
2902 | ||
2903 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
2904 | struct kobj_attribute *attr, char *buf) | |
2905 | { | |
ae7a927d | 2906 | return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan); |
31dbd01f IE |
2907 | } |
2908 | ||
2909 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
2910 | struct kobj_attribute *attr, | |
2911 | const char *buf, size_t count) | |
2912 | { | |
dfefd226 | 2913 | unsigned int nr_pages; |
31dbd01f | 2914 | int err; |
31dbd01f | 2915 | |
dfefd226 AD |
2916 | err = kstrtouint(buf, 10, &nr_pages); |
2917 | if (err) | |
31dbd01f IE |
2918 | return -EINVAL; |
2919 | ||
2920 | ksm_thread_pages_to_scan = nr_pages; | |
2921 | ||
2922 | return count; | |
2923 | } | |
2924 | KSM_ATTR(pages_to_scan); | |
2925 | ||
2926 | static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, | |
2927 | char *buf) | |
2928 | { | |
ae7a927d | 2929 | return sysfs_emit(buf, "%lu\n", ksm_run); |
31dbd01f IE |
2930 | } |
2931 | ||
2932 | static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, | |
2933 | const char *buf, size_t count) | |
2934 | { | |
dfefd226 | 2935 | unsigned int flags; |
31dbd01f | 2936 | int err; |
31dbd01f | 2937 | |
dfefd226 AD |
2938 | err = kstrtouint(buf, 10, &flags); |
2939 | if (err) | |
31dbd01f IE |
2940 | return -EINVAL; |
2941 | if (flags > KSM_RUN_UNMERGE) | |
2942 | return -EINVAL; | |
2943 | ||
2944 | /* | |
2945 | * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. | |
2946 | * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, | |
d0f209f6 HD |
2947 | * breaking COW to free the pages_shared (but leaves mm_slots |
2948 | * on the list for when ksmd may be set running again). | |
31dbd01f IE |
2949 | */ |
2950 | ||
2951 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 2952 | wait_while_offlining(); |
31dbd01f IE |
2953 | if (ksm_run != flags) { |
2954 | ksm_run = flags; | |
d952b791 | 2955 | if (flags & KSM_RUN_UNMERGE) { |
e1e12d2f | 2956 | set_current_oom_origin(); |
d952b791 | 2957 | err = unmerge_and_remove_all_rmap_items(); |
e1e12d2f | 2958 | clear_current_oom_origin(); |
d952b791 HD |
2959 | if (err) { |
2960 | ksm_run = KSM_RUN_STOP; | |
2961 | count = err; | |
2962 | } | |
2963 | } | |
31dbd01f IE |
2964 | } |
2965 | mutex_unlock(&ksm_thread_mutex); | |
2966 | ||
2967 | if (flags & KSM_RUN_MERGE) | |
2968 | wake_up_interruptible(&ksm_thread_wait); | |
2969 | ||
2970 | return count; | |
2971 | } | |
2972 | KSM_ATTR(run); | |
2973 | ||
90bd6fd3 PH |
2974 | #ifdef CONFIG_NUMA |
2975 | static ssize_t merge_across_nodes_show(struct kobject *kobj, | |
ae7a927d | 2976 | struct kobj_attribute *attr, char *buf) |
90bd6fd3 | 2977 | { |
ae7a927d | 2978 | return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes); |
90bd6fd3 PH |
2979 | } |
2980 | ||
2981 | static ssize_t merge_across_nodes_store(struct kobject *kobj, | |
2982 | struct kobj_attribute *attr, | |
2983 | const char *buf, size_t count) | |
2984 | { | |
2985 | int err; | |
2986 | unsigned long knob; | |
2987 | ||
2988 | err = kstrtoul(buf, 10, &knob); | |
2989 | if (err) | |
2990 | return err; | |
2991 | if (knob > 1) | |
2992 | return -EINVAL; | |
2993 | ||
2994 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 2995 | wait_while_offlining(); |
90bd6fd3 | 2996 | if (ksm_merge_across_nodes != knob) { |
cbf86cfe | 2997 | if (ksm_pages_shared || remove_all_stable_nodes()) |
90bd6fd3 | 2998 | err = -EBUSY; |
ef53d16c HD |
2999 | else if (root_stable_tree == one_stable_tree) { |
3000 | struct rb_root *buf; | |
3001 | /* | |
3002 | * This is the first time that we switch away from the | |
3003 | * default of merging across nodes: must now allocate | |
3004 | * a buffer to hold as many roots as may be needed. | |
3005 | * Allocate stable and unstable together: | |
3006 | * MAXSMP NODES_SHIFT 10 will use 16kB. | |
3007 | */ | |
bafe1e14 JP |
3008 | buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), |
3009 | GFP_KERNEL); | |
ef53d16c HD |
3010 | /* Let us assume that RB_ROOT is NULL is zero */ |
3011 | if (!buf) | |
3012 | err = -ENOMEM; | |
3013 | else { | |
3014 | root_stable_tree = buf; | |
3015 | root_unstable_tree = buf + nr_node_ids; | |
3016 | /* Stable tree is empty but not the unstable */ | |
3017 | root_unstable_tree[0] = one_unstable_tree[0]; | |
3018 | } | |
3019 | } | |
3020 | if (!err) { | |
90bd6fd3 | 3021 | ksm_merge_across_nodes = knob; |
ef53d16c HD |
3022 | ksm_nr_node_ids = knob ? 1 : nr_node_ids; |
3023 | } | |
90bd6fd3 PH |
3024 | } |
3025 | mutex_unlock(&ksm_thread_mutex); | |
3026 | ||
3027 | return err ? err : count; | |
3028 | } | |
3029 | KSM_ATTR(merge_across_nodes); | |
3030 | #endif | |
3031 | ||
e86c59b1 | 3032 | static ssize_t use_zero_pages_show(struct kobject *kobj, |
ae7a927d | 3033 | struct kobj_attribute *attr, char *buf) |
e86c59b1 | 3034 | { |
ae7a927d | 3035 | return sysfs_emit(buf, "%u\n", ksm_use_zero_pages); |
e86c59b1 CI |
3036 | } |
3037 | static ssize_t use_zero_pages_store(struct kobject *kobj, | |
3038 | struct kobj_attribute *attr, | |
3039 | const char *buf, size_t count) | |
3040 | { | |
3041 | int err; | |
3042 | bool value; | |
3043 | ||
3044 | err = kstrtobool(buf, &value); | |
3045 | if (err) | |
3046 | return -EINVAL; | |
3047 | ||
3048 | ksm_use_zero_pages = value; | |
3049 | ||
3050 | return count; | |
3051 | } | |
3052 | KSM_ATTR(use_zero_pages); | |
3053 | ||
2c653d0e AA |
3054 | static ssize_t max_page_sharing_show(struct kobject *kobj, |
3055 | struct kobj_attribute *attr, char *buf) | |
3056 | { | |
ae7a927d | 3057 | return sysfs_emit(buf, "%u\n", ksm_max_page_sharing); |
2c653d0e AA |
3058 | } |
3059 | ||
3060 | static ssize_t max_page_sharing_store(struct kobject *kobj, | |
3061 | struct kobj_attribute *attr, | |
3062 | const char *buf, size_t count) | |
3063 | { | |
3064 | int err; | |
3065 | int knob; | |
3066 | ||
3067 | err = kstrtoint(buf, 10, &knob); | |
3068 | if (err) | |
3069 | return err; | |
3070 | /* | |
3071 | * When a KSM page is created it is shared by 2 mappings. This | |
3072 | * being a signed comparison, it implicitly verifies it's not | |
3073 | * negative. | |
3074 | */ | |
3075 | if (knob < 2) | |
3076 | return -EINVAL; | |
3077 | ||
3078 | if (READ_ONCE(ksm_max_page_sharing) == knob) | |
3079 | return count; | |
3080 | ||
3081 | mutex_lock(&ksm_thread_mutex); | |
3082 | wait_while_offlining(); | |
3083 | if (ksm_max_page_sharing != knob) { | |
3084 | if (ksm_pages_shared || remove_all_stable_nodes()) | |
3085 | err = -EBUSY; | |
3086 | else | |
3087 | ksm_max_page_sharing = knob; | |
3088 | } | |
3089 | mutex_unlock(&ksm_thread_mutex); | |
3090 | ||
3091 | return err ? err : count; | |
3092 | } | |
3093 | KSM_ATTR(max_page_sharing); | |
3094 | ||
b4028260 HD |
3095 | static ssize_t pages_shared_show(struct kobject *kobj, |
3096 | struct kobj_attribute *attr, char *buf) | |
3097 | { | |
ae7a927d | 3098 | return sysfs_emit(buf, "%lu\n", ksm_pages_shared); |
b4028260 HD |
3099 | } |
3100 | KSM_ATTR_RO(pages_shared); | |
3101 | ||
3102 | static ssize_t pages_sharing_show(struct kobject *kobj, | |
3103 | struct kobj_attribute *attr, char *buf) | |
3104 | { | |
ae7a927d | 3105 | return sysfs_emit(buf, "%lu\n", ksm_pages_sharing); |
b4028260 HD |
3106 | } |
3107 | KSM_ATTR_RO(pages_sharing); | |
3108 | ||
473b0ce4 HD |
3109 | static ssize_t pages_unshared_show(struct kobject *kobj, |
3110 | struct kobj_attribute *attr, char *buf) | |
3111 | { | |
ae7a927d | 3112 | return sysfs_emit(buf, "%lu\n", ksm_pages_unshared); |
473b0ce4 HD |
3113 | } |
3114 | KSM_ATTR_RO(pages_unshared); | |
3115 | ||
3116 | static ssize_t pages_volatile_show(struct kobject *kobj, | |
3117 | struct kobj_attribute *attr, char *buf) | |
3118 | { | |
3119 | long ksm_pages_volatile; | |
3120 | ||
3121 | ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared | |
3122 | - ksm_pages_sharing - ksm_pages_unshared; | |
3123 | /* | |
3124 | * It was not worth any locking to calculate that statistic, | |
3125 | * but it might therefore sometimes be negative: conceal that. | |
3126 | */ | |
3127 | if (ksm_pages_volatile < 0) | |
3128 | ksm_pages_volatile = 0; | |
ae7a927d | 3129 | return sysfs_emit(buf, "%ld\n", ksm_pages_volatile); |
473b0ce4 HD |
3130 | } |
3131 | KSM_ATTR_RO(pages_volatile); | |
3132 | ||
2c653d0e AA |
3133 | static ssize_t stable_node_dups_show(struct kobject *kobj, |
3134 | struct kobj_attribute *attr, char *buf) | |
3135 | { | |
ae7a927d | 3136 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups); |
2c653d0e AA |
3137 | } |
3138 | KSM_ATTR_RO(stable_node_dups); | |
3139 | ||
3140 | static ssize_t stable_node_chains_show(struct kobject *kobj, | |
3141 | struct kobj_attribute *attr, char *buf) | |
3142 | { | |
ae7a927d | 3143 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains); |
2c653d0e AA |
3144 | } |
3145 | KSM_ATTR_RO(stable_node_chains); | |
3146 | ||
3147 | static ssize_t | |
3148 | stable_node_chains_prune_millisecs_show(struct kobject *kobj, | |
3149 | struct kobj_attribute *attr, | |
3150 | char *buf) | |
3151 | { | |
ae7a927d | 3152 | return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs); |
2c653d0e AA |
3153 | } |
3154 | ||
3155 | static ssize_t | |
3156 | stable_node_chains_prune_millisecs_store(struct kobject *kobj, | |
3157 | struct kobj_attribute *attr, | |
3158 | const char *buf, size_t count) | |
3159 | { | |
584ff0df | 3160 | unsigned int msecs; |
2c653d0e AA |
3161 | int err; |
3162 | ||
584ff0df ZB |
3163 | err = kstrtouint(buf, 10, &msecs); |
3164 | if (err) | |
2c653d0e AA |
3165 | return -EINVAL; |
3166 | ||
3167 | ksm_stable_node_chains_prune_millisecs = msecs; | |
3168 | ||
3169 | return count; | |
3170 | } | |
3171 | KSM_ATTR(stable_node_chains_prune_millisecs); | |
3172 | ||
473b0ce4 HD |
3173 | static ssize_t full_scans_show(struct kobject *kobj, |
3174 | struct kobj_attribute *attr, char *buf) | |
3175 | { | |
ae7a927d | 3176 | return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr); |
473b0ce4 HD |
3177 | } |
3178 | KSM_ATTR_RO(full_scans); | |
3179 | ||
31dbd01f IE |
3180 | static struct attribute *ksm_attrs[] = { |
3181 | &sleep_millisecs_attr.attr, | |
3182 | &pages_to_scan_attr.attr, | |
3183 | &run_attr.attr, | |
b4028260 HD |
3184 | &pages_shared_attr.attr, |
3185 | &pages_sharing_attr.attr, | |
473b0ce4 HD |
3186 | &pages_unshared_attr.attr, |
3187 | &pages_volatile_attr.attr, | |
3188 | &full_scans_attr.attr, | |
90bd6fd3 PH |
3189 | #ifdef CONFIG_NUMA |
3190 | &merge_across_nodes_attr.attr, | |
3191 | #endif | |
2c653d0e AA |
3192 | &max_page_sharing_attr.attr, |
3193 | &stable_node_chains_attr.attr, | |
3194 | &stable_node_dups_attr.attr, | |
3195 | &stable_node_chains_prune_millisecs_attr.attr, | |
e86c59b1 | 3196 | &use_zero_pages_attr.attr, |
31dbd01f IE |
3197 | NULL, |
3198 | }; | |
3199 | ||
f907c26a | 3200 | static const struct attribute_group ksm_attr_group = { |
31dbd01f IE |
3201 | .attrs = ksm_attrs, |
3202 | .name = "ksm", | |
3203 | }; | |
2ffd8679 | 3204 | #endif /* CONFIG_SYSFS */ |
31dbd01f IE |
3205 | |
3206 | static int __init ksm_init(void) | |
3207 | { | |
3208 | struct task_struct *ksm_thread; | |
3209 | int err; | |
3210 | ||
e86c59b1 CI |
3211 | /* The correct value depends on page size and endianness */ |
3212 | zero_checksum = calc_checksum(ZERO_PAGE(0)); | |
3213 | /* Default to false for backwards compatibility */ | |
3214 | ksm_use_zero_pages = false; | |
3215 | ||
31dbd01f IE |
3216 | err = ksm_slab_init(); |
3217 | if (err) | |
3218 | goto out; | |
3219 | ||
31dbd01f IE |
3220 | ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); |
3221 | if (IS_ERR(ksm_thread)) { | |
25acde31 | 3222 | pr_err("ksm: creating kthread failed\n"); |
31dbd01f | 3223 | err = PTR_ERR(ksm_thread); |
d9f8984c | 3224 | goto out_free; |
31dbd01f IE |
3225 | } |
3226 | ||
2ffd8679 | 3227 | #ifdef CONFIG_SYSFS |
31dbd01f IE |
3228 | err = sysfs_create_group(mm_kobj, &ksm_attr_group); |
3229 | if (err) { | |
25acde31 | 3230 | pr_err("ksm: register sysfs failed\n"); |
2ffd8679 | 3231 | kthread_stop(ksm_thread); |
d9f8984c | 3232 | goto out_free; |
31dbd01f | 3233 | } |
c73602ad HD |
3234 | #else |
3235 | ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ | |
3236 | ||
2ffd8679 | 3237 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 3238 | |
62b61f61 | 3239 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 | 3240 | /* There is no significance to this priority 100 */ |
1eeaa4fd | 3241 | hotplug_memory_notifier(ksm_memory_callback, KSM_CALLBACK_PRI); |
62b61f61 | 3242 | #endif |
31dbd01f IE |
3243 | return 0; |
3244 | ||
d9f8984c | 3245 | out_free: |
31dbd01f IE |
3246 | ksm_slab_free(); |
3247 | out: | |
3248 | return err; | |
f8af4da3 | 3249 | } |
a64fb3cd | 3250 | subsys_initcall(ksm_init); |