Commit | Line | Data |
---|---|---|
7a338472 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
f8af4da3 | 2 | /* |
31dbd01f IE |
3 | * Memory merging support. |
4 | * | |
5 | * This code enables dynamic sharing of identical pages found in different | |
6 | * memory areas, even if they are not shared by fork() | |
7 | * | |
36b2528d | 8 | * Copyright (C) 2008-2009 Red Hat, Inc. |
31dbd01f IE |
9 | * Authors: |
10 | * Izik Eidus | |
11 | * Andrea Arcangeli | |
12 | * Chris Wright | |
36b2528d | 13 | * Hugh Dickins |
f8af4da3 HD |
14 | */ |
15 | ||
16 | #include <linux/errno.h> | |
31dbd01f | 17 | #include <linux/mm.h> |
36090def | 18 | #include <linux/mm_inline.h> |
31dbd01f | 19 | #include <linux/fs.h> |
f8af4da3 | 20 | #include <linux/mman.h> |
31dbd01f | 21 | #include <linux/sched.h> |
6e84f315 | 22 | #include <linux/sched/mm.h> |
f7ccbae4 | 23 | #include <linux/sched/coredump.h> |
31dbd01f IE |
24 | #include <linux/rwsem.h> |
25 | #include <linux/pagemap.h> | |
26 | #include <linux/rmap.h> | |
27 | #include <linux/spinlock.h> | |
59e1a2f4 | 28 | #include <linux/xxhash.h> |
31dbd01f IE |
29 | #include <linux/delay.h> |
30 | #include <linux/kthread.h> | |
31 | #include <linux/wait.h> | |
32 | #include <linux/slab.h> | |
33 | #include <linux/rbtree.h> | |
62b61f61 | 34 | #include <linux/memory.h> |
31dbd01f | 35 | #include <linux/mmu_notifier.h> |
2c6854fd | 36 | #include <linux/swap.h> |
f8af4da3 | 37 | #include <linux/ksm.h> |
4ca3a69b | 38 | #include <linux/hashtable.h> |
878aee7d | 39 | #include <linux/freezer.h> |
72788c38 | 40 | #include <linux/oom.h> |
90bd6fd3 | 41 | #include <linux/numa.h> |
d7c0e68d | 42 | #include <linux/pagewalk.h> |
f8af4da3 | 43 | |
31dbd01f | 44 | #include <asm/tlbflush.h> |
73848b46 | 45 | #include "internal.h" |
58730ab6 | 46 | #include "mm_slot.h" |
31dbd01f | 47 | |
739100c8 SR |
48 | #define CREATE_TRACE_POINTS |
49 | #include <trace/events/ksm.h> | |
50 | ||
e850dcf5 HD |
51 | #ifdef CONFIG_NUMA |
52 | #define NUMA(x) (x) | |
53 | #define DO_NUMA(x) do { (x); } while (0) | |
54 | #else | |
55 | #define NUMA(x) (0) | |
56 | #define DO_NUMA(x) do { } while (0) | |
57 | #endif | |
58 | ||
5e924ff5 SR |
59 | typedef u8 rmap_age_t; |
60 | ||
5a2ca3ef MR |
61 | /** |
62 | * DOC: Overview | |
63 | * | |
31dbd01f IE |
64 | * A few notes about the KSM scanning process, |
65 | * to make it easier to understand the data structures below: | |
66 | * | |
67 | * In order to reduce excessive scanning, KSM sorts the memory pages by their | |
68 | * contents into a data structure that holds pointers to the pages' locations. | |
69 | * | |
70 | * Since the contents of the pages may change at any moment, KSM cannot just | |
71 | * insert the pages into a normal sorted tree and expect it to find anything. | |
72 | * Therefore KSM uses two data structures - the stable and the unstable tree. | |
73 | * | |
74 | * The stable tree holds pointers to all the merged pages (ksm pages), sorted | |
75 | * by their contents. Because each such page is write-protected, searching on | |
76 | * this tree is fully assured to be working (except when pages are unmapped), | |
77 | * and therefore this tree is called the stable tree. | |
78 | * | |
5a2ca3ef MR |
79 | * The stable tree node includes information required for reverse |
80 | * mapping from a KSM page to virtual addresses that map this page. | |
81 | * | |
82 | * In order to avoid large latencies of the rmap walks on KSM pages, | |
83 | * KSM maintains two types of nodes in the stable tree: | |
84 | * | |
85 | * * the regular nodes that keep the reverse mapping structures in a | |
86 | * linked list | |
87 | * * the "chains" that link nodes ("dups") that represent the same | |
88 | * write protected memory content, but each "dup" corresponds to a | |
89 | * different KSM page copy of that content | |
90 | * | |
91 | * Internally, the regular nodes, "dups" and "chains" are represented | |
21fbd591 | 92 | * using the same struct ksm_stable_node structure. |
5a2ca3ef | 93 | * |
31dbd01f IE |
94 | * In addition to the stable tree, KSM uses a second data structure called the |
95 | * unstable tree: this tree holds pointers to pages which have been found to | |
96 | * be "unchanged for a period of time". The unstable tree sorts these pages | |
97 | * by their contents, but since they are not write-protected, KSM cannot rely | |
98 | * upon the unstable tree to work correctly - the unstable tree is liable to | |
99 | * be corrupted as its contents are modified, and so it is called unstable. | |
100 | * | |
101 | * KSM solves this problem by several techniques: | |
102 | * | |
103 | * 1) The unstable tree is flushed every time KSM completes scanning all | |
104 | * memory areas, and then the tree is rebuilt again from the beginning. | |
105 | * 2) KSM will only insert into the unstable tree, pages whose hash value | |
106 | * has not changed since the previous scan of all memory areas. | |
107 | * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the | |
108 | * colors of the nodes and not on their contents, assuring that even when | |
109 | * the tree gets "corrupted" it won't get out of balance, so scanning time | |
110 | * remains the same (also, searching and inserting nodes in an rbtree uses | |
111 | * the same algorithm, so we have no overhead when we flush and rebuild). | |
112 | * 4) KSM never flushes the stable tree, which means that even if it were to | |
113 | * take 10 attempts to find a page in the unstable tree, once it is found, | |
114 | * it is secured in the stable tree. (When we scan a new page, we first | |
115 | * compare it against the stable tree, and then against the unstable tree.) | |
8fdb3dbf HD |
116 | * |
117 | * If the merge_across_nodes tunable is unset, then KSM maintains multiple | |
118 | * stable trees and multiple unstable trees: one of each for each NUMA node. | |
31dbd01f IE |
119 | */ |
120 | ||
121 | /** | |
21fbd591 | 122 | * struct ksm_mm_slot - ksm information per mm that is being scanned |
58730ab6 | 123 | * @slot: hash lookup from mm to mm_slot |
6514d511 | 124 | * @rmap_list: head for this mm_slot's singly-linked list of rmap_items |
31dbd01f | 125 | */ |
21fbd591 | 126 | struct ksm_mm_slot { |
58730ab6 | 127 | struct mm_slot slot; |
21fbd591 | 128 | struct ksm_rmap_item *rmap_list; |
31dbd01f IE |
129 | }; |
130 | ||
131 | /** | |
132 | * struct ksm_scan - cursor for scanning | |
133 | * @mm_slot: the current mm_slot we are scanning | |
134 | * @address: the next address inside that to be scanned | |
6514d511 | 135 | * @rmap_list: link to the next rmap to be scanned in the rmap_list |
31dbd01f IE |
136 | * @seqnr: count of completed full scans (needed when removing unstable node) |
137 | * | |
138 | * There is only the one ksm_scan instance of this cursor structure. | |
139 | */ | |
140 | struct ksm_scan { | |
21fbd591 | 141 | struct ksm_mm_slot *mm_slot; |
31dbd01f | 142 | unsigned long address; |
21fbd591 | 143 | struct ksm_rmap_item **rmap_list; |
31dbd01f IE |
144 | unsigned long seqnr; |
145 | }; | |
146 | ||
7b6ba2c7 | 147 | /** |
21fbd591 | 148 | * struct ksm_stable_node - node of the stable rbtree |
7b6ba2c7 | 149 | * @node: rb node of this ksm page in the stable tree |
4146d2d6 | 150 | * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list |
2c653d0e | 151 | * @hlist_dup: linked into the stable_node->hlist with a stable_node chain |
4146d2d6 | 152 | * @list: linked into migrate_nodes, pending placement in the proper node tree |
7b6ba2c7 | 153 | * @hlist: hlist head of rmap_items using this ksm page |
4146d2d6 | 154 | * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) |
2c653d0e AA |
155 | * @chain_prune_time: time of the last full garbage collection |
156 | * @rmap_hlist_len: number of rmap_item entries in hlist or STABLE_NODE_CHAIN | |
4146d2d6 | 157 | * @nid: NUMA node id of stable tree in which linked (may not match kpfn) |
7b6ba2c7 | 158 | */ |
21fbd591 | 159 | struct ksm_stable_node { |
4146d2d6 HD |
160 | union { |
161 | struct rb_node node; /* when node of stable tree */ | |
162 | struct { /* when listed for migration */ | |
163 | struct list_head *head; | |
2c653d0e AA |
164 | struct { |
165 | struct hlist_node hlist_dup; | |
166 | struct list_head list; | |
167 | }; | |
4146d2d6 HD |
168 | }; |
169 | }; | |
7b6ba2c7 | 170 | struct hlist_head hlist; |
2c653d0e AA |
171 | union { |
172 | unsigned long kpfn; | |
173 | unsigned long chain_prune_time; | |
174 | }; | |
175 | /* | |
176 | * STABLE_NODE_CHAIN can be any negative number in | |
177 | * rmap_hlist_len negative range, but better not -1 to be able | |
178 | * to reliably detect underflows. | |
179 | */ | |
180 | #define STABLE_NODE_CHAIN -1024 | |
181 | int rmap_hlist_len; | |
4146d2d6 HD |
182 | #ifdef CONFIG_NUMA |
183 | int nid; | |
184 | #endif | |
7b6ba2c7 HD |
185 | }; |
186 | ||
31dbd01f | 187 | /** |
21fbd591 | 188 | * struct ksm_rmap_item - reverse mapping item for virtual addresses |
6514d511 | 189 | * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list |
db114b83 | 190 | * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree |
bc56620b | 191 | * @nid: NUMA node id of unstable tree in which linked (may not match page) |
31dbd01f IE |
192 | * @mm: the memory structure this rmap_item is pointing into |
193 | * @address: the virtual address this rmap_item tracks (+ flags in low bits) | |
194 | * @oldchecksum: previous checksum of the page at that virtual address | |
7b6ba2c7 HD |
195 | * @node: rb node of this rmap_item in the unstable tree |
196 | * @head: pointer to stable_node heading this list in the stable tree | |
197 | * @hlist: link into hlist of rmap_items hanging off that stable_node | |
5e924ff5 SR |
198 | * @age: number of scan iterations since creation |
199 | * @remaining_skips: how many scans to skip | |
31dbd01f | 200 | */ |
21fbd591 QZ |
201 | struct ksm_rmap_item { |
202 | struct ksm_rmap_item *rmap_list; | |
bc56620b HD |
203 | union { |
204 | struct anon_vma *anon_vma; /* when stable */ | |
205 | #ifdef CONFIG_NUMA | |
206 | int nid; /* when node of unstable tree */ | |
207 | #endif | |
208 | }; | |
31dbd01f IE |
209 | struct mm_struct *mm; |
210 | unsigned long address; /* + low bits used for flags below */ | |
7b6ba2c7 | 211 | unsigned int oldchecksum; /* when unstable */ |
5e924ff5 SR |
212 | rmap_age_t age; |
213 | rmap_age_t remaining_skips; | |
31dbd01f | 214 | union { |
7b6ba2c7 HD |
215 | struct rb_node node; /* when node of unstable tree */ |
216 | struct { /* when listed from stable tree */ | |
21fbd591 | 217 | struct ksm_stable_node *head; |
7b6ba2c7 HD |
218 | struct hlist_node hlist; |
219 | }; | |
31dbd01f IE |
220 | }; |
221 | }; | |
222 | ||
223 | #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ | |
7b6ba2c7 HD |
224 | #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ |
225 | #define STABLE_FLAG 0x200 /* is listed from the stable tree */ | |
31dbd01f IE |
226 | |
227 | /* The stable and unstable tree heads */ | |
ef53d16c HD |
228 | static struct rb_root one_stable_tree[1] = { RB_ROOT }; |
229 | static struct rb_root one_unstable_tree[1] = { RB_ROOT }; | |
230 | static struct rb_root *root_stable_tree = one_stable_tree; | |
231 | static struct rb_root *root_unstable_tree = one_unstable_tree; | |
31dbd01f | 232 | |
4146d2d6 HD |
233 | /* Recently migrated nodes of stable tree, pending proper placement */ |
234 | static LIST_HEAD(migrate_nodes); | |
2c653d0e | 235 | #define STABLE_NODE_DUP_HEAD ((struct list_head *)&migrate_nodes.prev) |
4146d2d6 | 236 | |
4ca3a69b SL |
237 | #define MM_SLOTS_HASH_BITS 10 |
238 | static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
31dbd01f | 239 | |
21fbd591 | 240 | static struct ksm_mm_slot ksm_mm_head = { |
58730ab6 | 241 | .slot.mm_node = LIST_HEAD_INIT(ksm_mm_head.slot.mm_node), |
31dbd01f IE |
242 | }; |
243 | static struct ksm_scan ksm_scan = { | |
244 | .mm_slot = &ksm_mm_head, | |
245 | }; | |
246 | ||
247 | static struct kmem_cache *rmap_item_cache; | |
7b6ba2c7 | 248 | static struct kmem_cache *stable_node_cache; |
31dbd01f IE |
249 | static struct kmem_cache *mm_slot_cache; |
250 | ||
b348b5fe SR |
251 | /* The number of pages scanned */ |
252 | static unsigned long ksm_pages_scanned; | |
253 | ||
31dbd01f | 254 | /* The number of nodes in the stable tree */ |
b4028260 | 255 | static unsigned long ksm_pages_shared; |
31dbd01f | 256 | |
e178dfde | 257 | /* The number of page slots additionally sharing those nodes */ |
b4028260 | 258 | static unsigned long ksm_pages_sharing; |
31dbd01f | 259 | |
473b0ce4 HD |
260 | /* The number of nodes in the unstable tree */ |
261 | static unsigned long ksm_pages_unshared; | |
262 | ||
263 | /* The number of rmap_items in use: to calculate pages_volatile */ | |
264 | static unsigned long ksm_rmap_items; | |
265 | ||
2c653d0e AA |
266 | /* The number of stable_node chains */ |
267 | static unsigned long ksm_stable_node_chains; | |
268 | ||
269 | /* The number of stable_node dups linked to the stable_node chains */ | |
270 | static unsigned long ksm_stable_node_dups; | |
271 | ||
272 | /* Delay in pruning stale stable_node_dups in the stable_node_chains */ | |
584ff0df | 273 | static unsigned int ksm_stable_node_chains_prune_millisecs = 2000; |
2c653d0e AA |
274 | |
275 | /* Maximum number of page slots sharing a stable node */ | |
276 | static int ksm_max_page_sharing = 256; | |
277 | ||
31dbd01f | 278 | /* Number of pages ksmd should scan in one batch */ |
2c6854fd | 279 | static unsigned int ksm_thread_pages_to_scan = 100; |
31dbd01f IE |
280 | |
281 | /* Milliseconds ksmd should sleep between batches */ | |
2ffd8679 | 282 | static unsigned int ksm_thread_sleep_millisecs = 20; |
31dbd01f | 283 | |
e86c59b1 CI |
284 | /* Checksum of an empty (zeroed) page */ |
285 | static unsigned int zero_checksum __read_mostly; | |
286 | ||
287 | /* Whether to merge empty (zeroed) pages with actual zero pages */ | |
288 | static bool ksm_use_zero_pages __read_mostly; | |
289 | ||
5e924ff5 SR |
290 | /* Skip pages that couldn't be de-duplicated previously */ |
291 | /* Default to true at least temporarily, for testing */ | |
292 | static bool ksm_smart_scan = true; | |
293 | ||
e2942062 | 294 | /* The number of zero pages which is placed by KSM */ |
295 | unsigned long ksm_zero_pages; | |
296 | ||
e5a68991 SR |
297 | /* The number of pages that have been skipped due to "smart scanning" */ |
298 | static unsigned long ksm_pages_skipped; | |
299 | ||
e850dcf5 | 300 | #ifdef CONFIG_NUMA |
90bd6fd3 PH |
301 | /* Zeroed when merging across nodes is not allowed */ |
302 | static unsigned int ksm_merge_across_nodes = 1; | |
ef53d16c | 303 | static int ksm_nr_node_ids = 1; |
e850dcf5 HD |
304 | #else |
305 | #define ksm_merge_across_nodes 1U | |
ef53d16c | 306 | #define ksm_nr_node_ids 1 |
e850dcf5 | 307 | #endif |
90bd6fd3 | 308 | |
31dbd01f IE |
309 | #define KSM_RUN_STOP 0 |
310 | #define KSM_RUN_MERGE 1 | |
311 | #define KSM_RUN_UNMERGE 2 | |
ef4d43a8 HD |
312 | #define KSM_RUN_OFFLINE 4 |
313 | static unsigned long ksm_run = KSM_RUN_STOP; | |
314 | static void wait_while_offlining(void); | |
31dbd01f IE |
315 | |
316 | static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); | |
fcf9a0ef | 317 | static DECLARE_WAIT_QUEUE_HEAD(ksm_iter_wait); |
31dbd01f IE |
318 | static DEFINE_MUTEX(ksm_thread_mutex); |
319 | static DEFINE_SPINLOCK(ksm_mmlist_lock); | |
320 | ||
21fbd591 | 321 | #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create(#__struct,\ |
31dbd01f IE |
322 | sizeof(struct __struct), __alignof__(struct __struct),\ |
323 | (__flags), NULL) | |
324 | ||
325 | static int __init ksm_slab_init(void) | |
326 | { | |
21fbd591 | 327 | rmap_item_cache = KSM_KMEM_CACHE(ksm_rmap_item, 0); |
31dbd01f IE |
328 | if (!rmap_item_cache) |
329 | goto out; | |
330 | ||
21fbd591 | 331 | stable_node_cache = KSM_KMEM_CACHE(ksm_stable_node, 0); |
7b6ba2c7 HD |
332 | if (!stable_node_cache) |
333 | goto out_free1; | |
334 | ||
21fbd591 | 335 | mm_slot_cache = KSM_KMEM_CACHE(ksm_mm_slot, 0); |
31dbd01f | 336 | if (!mm_slot_cache) |
7b6ba2c7 | 337 | goto out_free2; |
31dbd01f IE |
338 | |
339 | return 0; | |
340 | ||
7b6ba2c7 HD |
341 | out_free2: |
342 | kmem_cache_destroy(stable_node_cache); | |
343 | out_free1: | |
31dbd01f IE |
344 | kmem_cache_destroy(rmap_item_cache); |
345 | out: | |
346 | return -ENOMEM; | |
347 | } | |
348 | ||
349 | static void __init ksm_slab_free(void) | |
350 | { | |
351 | kmem_cache_destroy(mm_slot_cache); | |
7b6ba2c7 | 352 | kmem_cache_destroy(stable_node_cache); |
31dbd01f IE |
353 | kmem_cache_destroy(rmap_item_cache); |
354 | mm_slot_cache = NULL; | |
355 | } | |
356 | ||
21fbd591 | 357 | static __always_inline bool is_stable_node_chain(struct ksm_stable_node *chain) |
2c653d0e AA |
358 | { |
359 | return chain->rmap_hlist_len == STABLE_NODE_CHAIN; | |
360 | } | |
361 | ||
21fbd591 | 362 | static __always_inline bool is_stable_node_dup(struct ksm_stable_node *dup) |
2c653d0e AA |
363 | { |
364 | return dup->head == STABLE_NODE_DUP_HEAD; | |
365 | } | |
366 | ||
21fbd591 QZ |
367 | static inline void stable_node_chain_add_dup(struct ksm_stable_node *dup, |
368 | struct ksm_stable_node *chain) | |
2c653d0e AA |
369 | { |
370 | VM_BUG_ON(is_stable_node_dup(dup)); | |
371 | dup->head = STABLE_NODE_DUP_HEAD; | |
372 | VM_BUG_ON(!is_stable_node_chain(chain)); | |
373 | hlist_add_head(&dup->hlist_dup, &chain->hlist); | |
374 | ksm_stable_node_dups++; | |
375 | } | |
376 | ||
21fbd591 | 377 | static inline void __stable_node_dup_del(struct ksm_stable_node *dup) |
2c653d0e | 378 | { |
b4fecc67 | 379 | VM_BUG_ON(!is_stable_node_dup(dup)); |
2c653d0e AA |
380 | hlist_del(&dup->hlist_dup); |
381 | ksm_stable_node_dups--; | |
382 | } | |
383 | ||
21fbd591 | 384 | static inline void stable_node_dup_del(struct ksm_stable_node *dup) |
2c653d0e AA |
385 | { |
386 | VM_BUG_ON(is_stable_node_chain(dup)); | |
387 | if (is_stable_node_dup(dup)) | |
388 | __stable_node_dup_del(dup); | |
389 | else | |
390 | rb_erase(&dup->node, root_stable_tree + NUMA(dup->nid)); | |
391 | #ifdef CONFIG_DEBUG_VM | |
392 | dup->head = NULL; | |
393 | #endif | |
394 | } | |
395 | ||
21fbd591 | 396 | static inline struct ksm_rmap_item *alloc_rmap_item(void) |
31dbd01f | 397 | { |
21fbd591 | 398 | struct ksm_rmap_item *rmap_item; |
473b0ce4 | 399 | |
5b398e41 | 400 | rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | |
401 | __GFP_NORETRY | __GFP_NOWARN); | |
473b0ce4 HD |
402 | if (rmap_item) |
403 | ksm_rmap_items++; | |
404 | return rmap_item; | |
31dbd01f IE |
405 | } |
406 | ||
21fbd591 | 407 | static inline void free_rmap_item(struct ksm_rmap_item *rmap_item) |
31dbd01f | 408 | { |
473b0ce4 | 409 | ksm_rmap_items--; |
cb4df4ca | 410 | rmap_item->mm->ksm_rmap_items--; |
31dbd01f IE |
411 | rmap_item->mm = NULL; /* debug safety */ |
412 | kmem_cache_free(rmap_item_cache, rmap_item); | |
413 | } | |
414 | ||
21fbd591 | 415 | static inline struct ksm_stable_node *alloc_stable_node(void) |
7b6ba2c7 | 416 | { |
6213055f | 417 | /* |
418 | * The allocation can take too long with GFP_KERNEL when memory is under | |
419 | * pressure, which may lead to hung task warnings. Adding __GFP_HIGH | |
420 | * grants access to memory reserves, helping to avoid this problem. | |
421 | */ | |
422 | return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); | |
7b6ba2c7 HD |
423 | } |
424 | ||
21fbd591 | 425 | static inline void free_stable_node(struct ksm_stable_node *stable_node) |
7b6ba2c7 | 426 | { |
2c653d0e AA |
427 | VM_BUG_ON(stable_node->rmap_hlist_len && |
428 | !is_stable_node_chain(stable_node)); | |
7b6ba2c7 HD |
429 | kmem_cache_free(stable_node_cache, stable_node); |
430 | } | |
431 | ||
a913e182 HD |
432 | /* |
433 | * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's | |
434 | * page tables after it has passed through ksm_exit() - which, if necessary, | |
c1e8d7c6 | 435 | * takes mmap_lock briefly to serialize against them. ksm_exit() does not set |
a913e182 HD |
436 | * a special flag: they can just back out as soon as mm_users goes to zero. |
437 | * ksm_test_exit() is used throughout to make this test for exit: in some | |
438 | * places for correctness, in some places just to avoid unnecessary work. | |
439 | */ | |
440 | static inline bool ksm_test_exit(struct mm_struct *mm) | |
441 | { | |
442 | return atomic_read(&mm->mm_users) == 0; | |
443 | } | |
444 | ||
d7c0e68d DH |
445 | static int break_ksm_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long next, |
446 | struct mm_walk *walk) | |
447 | { | |
448 | struct page *page = NULL; | |
449 | spinlock_t *ptl; | |
450 | pte_t *pte; | |
c33c7948 | 451 | pte_t ptent; |
d7c0e68d DH |
452 | int ret; |
453 | ||
d7c0e68d | 454 | pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); |
04dee9e8 HD |
455 | if (!pte) |
456 | return 0; | |
c33c7948 RR |
457 | ptent = ptep_get(pte); |
458 | if (pte_present(ptent)) { | |
459 | page = vm_normal_page(walk->vma, addr, ptent); | |
460 | } else if (!pte_none(ptent)) { | |
461 | swp_entry_t entry = pte_to_swp_entry(ptent); | |
d7c0e68d DH |
462 | |
463 | /* | |
464 | * As KSM pages remain KSM pages until freed, no need to wait | |
465 | * here for migration to end. | |
466 | */ | |
467 | if (is_migration_entry(entry)) | |
468 | page = pfn_swap_entry_to_page(entry); | |
469 | } | |
79271476 | 470 | /* return 1 if the page is an normal ksm page or KSM-placed zero page */ |
afccb080 | 471 | ret = (page && PageKsm(page)) || is_ksm_zero_pte(ptent); |
d7c0e68d DH |
472 | pte_unmap_unlock(pte, ptl); |
473 | return ret; | |
474 | } | |
475 | ||
476 | static const struct mm_walk_ops break_ksm_ops = { | |
477 | .pmd_entry = break_ksm_pmd_entry, | |
49b06385 SB |
478 | .walk_lock = PGWALK_RDLOCK, |
479 | }; | |
480 | ||
481 | static const struct mm_walk_ops break_ksm_lock_vma_ops = { | |
482 | .pmd_entry = break_ksm_pmd_entry, | |
483 | .walk_lock = PGWALK_WRLOCK, | |
d7c0e68d DH |
484 | }; |
485 | ||
31dbd01f | 486 | /* |
6cce3314 DH |
487 | * We use break_ksm to break COW on a ksm page by triggering unsharing, |
488 | * such that the ksm page will get replaced by an exclusive anonymous page. | |
31dbd01f | 489 | * |
6cce3314 | 490 | * We take great care only to touch a ksm page, in a VM_MERGEABLE vma, |
31dbd01f IE |
491 | * in case the application has unmapped and remapped mm,addr meanwhile. |
492 | * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP | |
bbcd53c9 | 493 | * mmap of /dev/mem, where we would not want to touch it. |
1b2ee126 | 494 | * |
6cce3314 | 495 | * FAULT_FLAG_REMOTE/FOLL_REMOTE are because we do this outside the context |
1b2ee126 DH |
496 | * of the process that owns 'vma'. We also do not want to enforce |
497 | * protection keys here anyway. | |
31dbd01f | 498 | */ |
49b06385 | 499 | static int break_ksm(struct vm_area_struct *vma, unsigned long addr, bool lock_vma) |
31dbd01f | 500 | { |
50a7ca3c | 501 | vm_fault_t ret = 0; |
49b06385 SB |
502 | const struct mm_walk_ops *ops = lock_vma ? |
503 | &break_ksm_lock_vma_ops : &break_ksm_ops; | |
31dbd01f IE |
504 | |
505 | do { | |
d7c0e68d | 506 | int ksm_page; |
58f595c6 | 507 | |
31dbd01f | 508 | cond_resched(); |
49b06385 | 509 | ksm_page = walk_page_range_vma(vma, addr, addr + 1, ops, NULL); |
d7c0e68d DH |
510 | if (WARN_ON_ONCE(ksm_page < 0)) |
511 | return ksm_page; | |
58f595c6 DH |
512 | if (!ksm_page) |
513 | return 0; | |
514 | ret = handle_mm_fault(vma, addr, | |
6cce3314 | 515 | FAULT_FLAG_UNSHARE | FAULT_FLAG_REMOTE, |
58f595c6 DH |
516 | NULL); |
517 | } while (!(ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); | |
d952b791 | 518 | /* |
58f595c6 DH |
519 | * We must loop until we no longer find a KSM page because |
520 | * handle_mm_fault() may back out if there's any difficulty e.g. if | |
521 | * pte accessed bit gets updated concurrently. | |
d952b791 HD |
522 | * |
523 | * VM_FAULT_SIGBUS could occur if we race with truncation of the | |
524 | * backing file, which also invalidates anonymous pages: that's | |
525 | * okay, that truncation will have unmapped the PageKsm for us. | |
526 | * | |
527 | * VM_FAULT_OOM: at the time of writing (late July 2009), setting | |
528 | * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the | |
529 | * current task has TIF_MEMDIE set, and will be OOM killed on return | |
530 | * to user; and ksmd, having no mm, would never be chosen for that. | |
531 | * | |
532 | * But if the mm is in a limited mem_cgroup, then the fault may fail | |
533 | * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and | |
534 | * even ksmd can fail in this way - though it's usually breaking ksm | |
535 | * just to undo a merge it made a moment before, so unlikely to oom. | |
536 | * | |
537 | * That's a pity: we might therefore have more kernel pages allocated | |
538 | * than we're counting as nodes in the stable tree; but ksm_do_scan | |
539 | * will retry to break_cow on each pass, so should recover the page | |
540 | * in due course. The important thing is to not let VM_MERGEABLE | |
541 | * be cleared while any such pages might remain in the area. | |
542 | */ | |
543 | return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; | |
31dbd01f IE |
544 | } |
545 | ||
d7597f59 SR |
546 | static bool vma_ksm_compatible(struct vm_area_struct *vma) |
547 | { | |
548 | if (vma->vm_flags & (VM_SHARED | VM_MAYSHARE | VM_PFNMAP | | |
549 | VM_IO | VM_DONTEXPAND | VM_HUGETLB | | |
550 | VM_MIXEDMAP)) | |
551 | return false; /* just ignore the advice */ | |
552 | ||
553 | if (vma_is_dax(vma)) | |
554 | return false; | |
555 | ||
556 | #ifdef VM_SAO | |
557 | if (vma->vm_flags & VM_SAO) | |
558 | return false; | |
559 | #endif | |
560 | #ifdef VM_SPARC_ADI | |
561 | if (vma->vm_flags & VM_SPARC_ADI) | |
562 | return false; | |
563 | #endif | |
564 | ||
565 | return true; | |
566 | } | |
567 | ||
ef694222 BL |
568 | static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, |
569 | unsigned long addr) | |
570 | { | |
571 | struct vm_area_struct *vma; | |
572 | if (ksm_test_exit(mm)) | |
573 | return NULL; | |
ff69fb81 LH |
574 | vma = vma_lookup(mm, addr); |
575 | if (!vma || !(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) | |
ef694222 BL |
576 | return NULL; |
577 | return vma; | |
578 | } | |
579 | ||
21fbd591 | 580 | static void break_cow(struct ksm_rmap_item *rmap_item) |
31dbd01f | 581 | { |
8dd3557a HD |
582 | struct mm_struct *mm = rmap_item->mm; |
583 | unsigned long addr = rmap_item->address; | |
31dbd01f IE |
584 | struct vm_area_struct *vma; |
585 | ||
4035c07a HD |
586 | /* |
587 | * It is not an accident that whenever we want to break COW | |
588 | * to undo, we also need to drop a reference to the anon_vma. | |
589 | */ | |
9e60109f | 590 | put_anon_vma(rmap_item->anon_vma); |
4035c07a | 591 | |
d8ed45c5 | 592 | mmap_read_lock(mm); |
ef694222 BL |
593 | vma = find_mergeable_vma(mm, addr); |
594 | if (vma) | |
49b06385 | 595 | break_ksm(vma, addr, false); |
d8ed45c5 | 596 | mmap_read_unlock(mm); |
31dbd01f IE |
597 | } |
598 | ||
21fbd591 | 599 | static struct page *get_mergeable_page(struct ksm_rmap_item *rmap_item) |
31dbd01f IE |
600 | { |
601 | struct mm_struct *mm = rmap_item->mm; | |
602 | unsigned long addr = rmap_item->address; | |
603 | struct vm_area_struct *vma; | |
604 | struct page *page; | |
605 | ||
d8ed45c5 | 606 | mmap_read_lock(mm); |
ef694222 BL |
607 | vma = find_mergeable_vma(mm, addr); |
608 | if (!vma) | |
31dbd01f IE |
609 | goto out; |
610 | ||
611 | page = follow_page(vma, addr, FOLL_GET); | |
f7091ed6 | 612 | if (IS_ERR_OR_NULL(page)) |
31dbd01f | 613 | goto out; |
f7091ed6 HW |
614 | if (is_zone_device_page(page)) |
615 | goto out_putpage; | |
f765f540 | 616 | if (PageAnon(page)) { |
31dbd01f IE |
617 | flush_anon_page(vma, page, addr); |
618 | flush_dcache_page(page); | |
619 | } else { | |
f7091ed6 | 620 | out_putpage: |
31dbd01f | 621 | put_page(page); |
c8f95ed1 AA |
622 | out: |
623 | page = NULL; | |
31dbd01f | 624 | } |
d8ed45c5 | 625 | mmap_read_unlock(mm); |
31dbd01f IE |
626 | return page; |
627 | } | |
628 | ||
90bd6fd3 PH |
629 | /* |
630 | * This helper is used for getting right index into array of tree roots. | |
631 | * When merge_across_nodes knob is set to 1, there are only two rb-trees for | |
632 | * stable and unstable pages from all nodes with roots in index 0. Otherwise, | |
633 | * every node has its own stable and unstable tree. | |
634 | */ | |
635 | static inline int get_kpfn_nid(unsigned long kpfn) | |
636 | { | |
d8fc16a8 | 637 | return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); |
90bd6fd3 PH |
638 | } |
639 | ||
21fbd591 | 640 | static struct ksm_stable_node *alloc_stable_node_chain(struct ksm_stable_node *dup, |
2c653d0e AA |
641 | struct rb_root *root) |
642 | { | |
21fbd591 | 643 | struct ksm_stable_node *chain = alloc_stable_node(); |
2c653d0e AA |
644 | VM_BUG_ON(is_stable_node_chain(dup)); |
645 | if (likely(chain)) { | |
646 | INIT_HLIST_HEAD(&chain->hlist); | |
647 | chain->chain_prune_time = jiffies; | |
648 | chain->rmap_hlist_len = STABLE_NODE_CHAIN; | |
649 | #if defined (CONFIG_DEBUG_VM) && defined(CONFIG_NUMA) | |
98fa15f3 | 650 | chain->nid = NUMA_NO_NODE; /* debug */ |
2c653d0e AA |
651 | #endif |
652 | ksm_stable_node_chains++; | |
653 | ||
654 | /* | |
655 | * Put the stable node chain in the first dimension of | |
656 | * the stable tree and at the same time remove the old | |
657 | * stable node. | |
658 | */ | |
659 | rb_replace_node(&dup->node, &chain->node, root); | |
660 | ||
661 | /* | |
662 | * Move the old stable node to the second dimension | |
663 | * queued in the hlist_dup. The invariant is that all | |
664 | * dup stable_nodes in the chain->hlist point to pages | |
457aef94 | 665 | * that are write protected and have the exact same |
2c653d0e AA |
666 | * content. |
667 | */ | |
668 | stable_node_chain_add_dup(dup, chain); | |
669 | } | |
670 | return chain; | |
671 | } | |
672 | ||
21fbd591 | 673 | static inline void free_stable_node_chain(struct ksm_stable_node *chain, |
2c653d0e AA |
674 | struct rb_root *root) |
675 | { | |
676 | rb_erase(&chain->node, root); | |
677 | free_stable_node(chain); | |
678 | ksm_stable_node_chains--; | |
679 | } | |
680 | ||
21fbd591 | 681 | static void remove_node_from_stable_tree(struct ksm_stable_node *stable_node) |
4035c07a | 682 | { |
21fbd591 | 683 | struct ksm_rmap_item *rmap_item; |
4035c07a | 684 | |
2c653d0e AA |
685 | /* check it's not STABLE_NODE_CHAIN or negative */ |
686 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
687 | ||
b67bfe0d | 688 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
739100c8 | 689 | if (rmap_item->hlist.next) { |
4035c07a | 690 | ksm_pages_sharing--; |
739100c8 SR |
691 | trace_ksm_remove_rmap_item(stable_node->kpfn, rmap_item, rmap_item->mm); |
692 | } else { | |
4035c07a | 693 | ksm_pages_shared--; |
739100c8 | 694 | } |
76093853 | 695 | |
696 | rmap_item->mm->ksm_merging_pages--; | |
697 | ||
2c653d0e AA |
698 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
699 | stable_node->rmap_hlist_len--; | |
9e60109f | 700 | put_anon_vma(rmap_item->anon_vma); |
4035c07a HD |
701 | rmap_item->address &= PAGE_MASK; |
702 | cond_resched(); | |
703 | } | |
704 | ||
2c653d0e AA |
705 | /* |
706 | * We need the second aligned pointer of the migrate_nodes | |
707 | * list_head to stay clear from the rb_parent_color union | |
708 | * (aligned and different than any node) and also different | |
709 | * from &migrate_nodes. This will verify that future list.h changes | |
815f0ddb | 710 | * don't break STABLE_NODE_DUP_HEAD. Only recent gcc can handle it. |
2c653d0e | 711 | */ |
2c653d0e AA |
712 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD <= &migrate_nodes); |
713 | BUILD_BUG_ON(STABLE_NODE_DUP_HEAD >= &migrate_nodes + 1); | |
2c653d0e | 714 | |
739100c8 | 715 | trace_ksm_remove_ksm_page(stable_node->kpfn); |
4146d2d6 HD |
716 | if (stable_node->head == &migrate_nodes) |
717 | list_del(&stable_node->list); | |
718 | else | |
2c653d0e | 719 | stable_node_dup_del(stable_node); |
4035c07a HD |
720 | free_stable_node(stable_node); |
721 | } | |
722 | ||
2cee57d1 YS |
723 | enum get_ksm_page_flags { |
724 | GET_KSM_PAGE_NOLOCK, | |
725 | GET_KSM_PAGE_LOCK, | |
726 | GET_KSM_PAGE_TRYLOCK | |
727 | }; | |
728 | ||
4035c07a HD |
729 | /* |
730 | * get_ksm_page: checks if the page indicated by the stable node | |
731 | * is still its ksm page, despite having held no reference to it. | |
732 | * In which case we can trust the content of the page, and it | |
733 | * returns the gotten page; but if the page has now been zapped, | |
734 | * remove the stale node from the stable tree and return NULL. | |
c8d6553b | 735 | * But beware, the stable node's page might be being migrated. |
4035c07a HD |
736 | * |
737 | * You would expect the stable_node to hold a reference to the ksm page. | |
738 | * But if it increments the page's count, swapping out has to wait for | |
739 | * ksmd to come around again before it can free the page, which may take | |
740 | * seconds or even minutes: much too unresponsive. So instead we use a | |
741 | * "keyhole reference": access to the ksm page from the stable node peeps | |
742 | * out through its keyhole to see if that page still holds the right key, | |
743 | * pointing back to this stable node. This relies on freeing a PageAnon | |
744 | * page to reset its page->mapping to NULL, and relies on no other use of | |
745 | * a page to put something that might look like our key in page->mapping. | |
4035c07a HD |
746 | * is on its way to being freed; but it is an anomaly to bear in mind. |
747 | */ | |
21fbd591 | 748 | static struct page *get_ksm_page(struct ksm_stable_node *stable_node, |
2cee57d1 | 749 | enum get_ksm_page_flags flags) |
4035c07a HD |
750 | { |
751 | struct page *page; | |
752 | void *expected_mapping; | |
c8d6553b | 753 | unsigned long kpfn; |
4035c07a | 754 | |
bda807d4 MK |
755 | expected_mapping = (void *)((unsigned long)stable_node | |
756 | PAGE_MAPPING_KSM); | |
c8d6553b | 757 | again: |
08df4774 | 758 | kpfn = READ_ONCE(stable_node->kpfn); /* Address dependency. */ |
c8d6553b | 759 | page = pfn_to_page(kpfn); |
4db0c3c2 | 760 | if (READ_ONCE(page->mapping) != expected_mapping) |
4035c07a | 761 | goto stale; |
c8d6553b HD |
762 | |
763 | /* | |
764 | * We cannot do anything with the page while its refcount is 0. | |
765 | * Usually 0 means free, or tail of a higher-order page: in which | |
766 | * case this node is no longer referenced, and should be freed; | |
1c4c3b99 | 767 | * however, it might mean that the page is under page_ref_freeze(). |
c8d6553b | 768 | * The __remove_mapping() case is easy, again the node is now stale; |
52d1e606 | 769 | * the same is in reuse_ksm_page() case; but if page is swapcache |
9800562f | 770 | * in folio_migrate_mapping(), it might still be our page, |
52d1e606 | 771 | * in which case it's essential to keep the node. |
c8d6553b HD |
772 | */ |
773 | while (!get_page_unless_zero(page)) { | |
774 | /* | |
775 | * Another check for page->mapping != expected_mapping would | |
776 | * work here too. We have chosen the !PageSwapCache test to | |
777 | * optimize the common case, when the page is or is about to | |
778 | * be freed: PageSwapCache is cleared (under spin_lock_irq) | |
1c4c3b99 | 779 | * in the ref_freeze section of __remove_mapping(); but Anon |
c8d6553b HD |
780 | * page->mapping reset to NULL later, in free_pages_prepare(). |
781 | */ | |
782 | if (!PageSwapCache(page)) | |
783 | goto stale; | |
784 | cpu_relax(); | |
785 | } | |
786 | ||
4db0c3c2 | 787 | if (READ_ONCE(page->mapping) != expected_mapping) { |
4035c07a HD |
788 | put_page(page); |
789 | goto stale; | |
790 | } | |
c8d6553b | 791 | |
2cee57d1 YS |
792 | if (flags == GET_KSM_PAGE_TRYLOCK) { |
793 | if (!trylock_page(page)) { | |
794 | put_page(page); | |
795 | return ERR_PTR(-EBUSY); | |
796 | } | |
797 | } else if (flags == GET_KSM_PAGE_LOCK) | |
8aafa6a4 | 798 | lock_page(page); |
2cee57d1 YS |
799 | |
800 | if (flags != GET_KSM_PAGE_NOLOCK) { | |
4db0c3c2 | 801 | if (READ_ONCE(page->mapping) != expected_mapping) { |
8aafa6a4 HD |
802 | unlock_page(page); |
803 | put_page(page); | |
804 | goto stale; | |
805 | } | |
806 | } | |
4035c07a | 807 | return page; |
c8d6553b | 808 | |
4035c07a | 809 | stale: |
c8d6553b HD |
810 | /* |
811 | * We come here from above when page->mapping or !PageSwapCache | |
812 | * suggests that the node is stale; but it might be under migration. | |
19138349 | 813 | * We need smp_rmb(), matching the smp_wmb() in folio_migrate_ksm(), |
c8d6553b HD |
814 | * before checking whether node->kpfn has been changed. |
815 | */ | |
816 | smp_rmb(); | |
4db0c3c2 | 817 | if (READ_ONCE(stable_node->kpfn) != kpfn) |
c8d6553b | 818 | goto again; |
4035c07a HD |
819 | remove_node_from_stable_tree(stable_node); |
820 | return NULL; | |
821 | } | |
822 | ||
31dbd01f IE |
823 | /* |
824 | * Removing rmap_item from stable or unstable tree. | |
825 | * This function will clean the information from the stable/unstable tree. | |
826 | */ | |
21fbd591 | 827 | static void remove_rmap_item_from_tree(struct ksm_rmap_item *rmap_item) |
31dbd01f | 828 | { |
7b6ba2c7 | 829 | if (rmap_item->address & STABLE_FLAG) { |
21fbd591 | 830 | struct ksm_stable_node *stable_node; |
5ad64688 | 831 | struct page *page; |
31dbd01f | 832 | |
7b6ba2c7 | 833 | stable_node = rmap_item->head; |
62862290 | 834 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
4035c07a HD |
835 | if (!page) |
836 | goto out; | |
5ad64688 | 837 | |
7b6ba2c7 | 838 | hlist_del(&rmap_item->hlist); |
62862290 | 839 | unlock_page(page); |
4035c07a | 840 | put_page(page); |
08beca44 | 841 | |
98666f8a | 842 | if (!hlist_empty(&stable_node->hlist)) |
4035c07a HD |
843 | ksm_pages_sharing--; |
844 | else | |
7b6ba2c7 | 845 | ksm_pages_shared--; |
76093853 | 846 | |
847 | rmap_item->mm->ksm_merging_pages--; | |
848 | ||
2c653d0e AA |
849 | VM_BUG_ON(stable_node->rmap_hlist_len <= 0); |
850 | stable_node->rmap_hlist_len--; | |
31dbd01f | 851 | |
9e60109f | 852 | put_anon_vma(rmap_item->anon_vma); |
c89a384e | 853 | rmap_item->head = NULL; |
93d17715 | 854 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 855 | |
7b6ba2c7 | 856 | } else if (rmap_item->address & UNSTABLE_FLAG) { |
31dbd01f IE |
857 | unsigned char age; |
858 | /* | |
9ba69294 | 859 | * Usually ksmd can and must skip the rb_erase, because |
31dbd01f | 860 | * root_unstable_tree was already reset to RB_ROOT. |
9ba69294 HD |
861 | * But be careful when an mm is exiting: do the rb_erase |
862 | * if this rmap_item was inserted by this scan, rather | |
863 | * than left over from before. | |
31dbd01f IE |
864 | */ |
865 | age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); | |
cd551f97 | 866 | BUG_ON(age > 1); |
31dbd01f | 867 | if (!age) |
90bd6fd3 | 868 | rb_erase(&rmap_item->node, |
ef53d16c | 869 | root_unstable_tree + NUMA(rmap_item->nid)); |
473b0ce4 | 870 | ksm_pages_unshared--; |
93d17715 | 871 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 872 | } |
4035c07a | 873 | out: |
31dbd01f IE |
874 | cond_resched(); /* we're called from many long loops */ |
875 | } | |
876 | ||
21fbd591 | 877 | static void remove_trailing_rmap_items(struct ksm_rmap_item **rmap_list) |
31dbd01f | 878 | { |
6514d511 | 879 | while (*rmap_list) { |
21fbd591 | 880 | struct ksm_rmap_item *rmap_item = *rmap_list; |
6514d511 | 881 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 882 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
883 | free_rmap_item(rmap_item); |
884 | } | |
885 | } | |
886 | ||
887 | /* | |
e850dcf5 | 888 | * Though it's very tempting to unmerge rmap_items from stable tree rather |
31dbd01f IE |
889 | * than check every pte of a given vma, the locking doesn't quite work for |
890 | * that - an rmap_item is assigned to the stable tree after inserting ksm | |
c1e8d7c6 | 891 | * page and upping mmap_lock. Nor does it fit with the way we skip dup'ing |
31dbd01f IE |
892 | * rmap_items from parent to child at fork time (so as not to waste time |
893 | * if exit comes before the next scan reaches it). | |
81464e30 HD |
894 | * |
895 | * Similarly, although we'd like to remove rmap_items (so updating counts | |
896 | * and freeing memory) when unmerging an area, it's easier to leave that | |
897 | * to the next pass of ksmd - consider, for example, how ksmd might be | |
898 | * in cmp_and_merge_page on one of the rmap_items we would be removing. | |
31dbd01f | 899 | */ |
d952b791 | 900 | static int unmerge_ksm_pages(struct vm_area_struct *vma, |
49b06385 | 901 | unsigned long start, unsigned long end, bool lock_vma) |
31dbd01f IE |
902 | { |
903 | unsigned long addr; | |
d952b791 | 904 | int err = 0; |
31dbd01f | 905 | |
d952b791 | 906 | for (addr = start; addr < end && !err; addr += PAGE_SIZE) { |
9ba69294 HD |
907 | if (ksm_test_exit(vma->vm_mm)) |
908 | break; | |
d952b791 HD |
909 | if (signal_pending(current)) |
910 | err = -ERESTARTSYS; | |
911 | else | |
49b06385 | 912 | err = break_ksm(vma, addr, lock_vma); |
d952b791 HD |
913 | } |
914 | return err; | |
31dbd01f IE |
915 | } |
916 | ||
21fbd591 | 917 | static inline struct ksm_stable_node *folio_stable_node(struct folio *folio) |
19138349 MWO |
918 | { |
919 | return folio_test_ksm(folio) ? folio_raw_mapping(folio) : NULL; | |
920 | } | |
921 | ||
21fbd591 | 922 | static inline struct ksm_stable_node *page_stable_node(struct page *page) |
88484826 | 923 | { |
19138349 | 924 | return folio_stable_node(page_folio(page)); |
88484826 MR |
925 | } |
926 | ||
927 | static inline void set_page_stable_node(struct page *page, | |
21fbd591 | 928 | struct ksm_stable_node *stable_node) |
88484826 | 929 | { |
6c287605 | 930 | VM_BUG_ON_PAGE(PageAnon(page) && PageAnonExclusive(page), page); |
88484826 MR |
931 | page->mapping = (void *)((unsigned long)stable_node | PAGE_MAPPING_KSM); |
932 | } | |
933 | ||
2ffd8679 HD |
934 | #ifdef CONFIG_SYSFS |
935 | /* | |
936 | * Only called through the sysfs control interface: | |
937 | */ | |
21fbd591 | 938 | static int remove_stable_node(struct ksm_stable_node *stable_node) |
cbf86cfe HD |
939 | { |
940 | struct page *page; | |
941 | int err; | |
942 | ||
2cee57d1 | 943 | page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK); |
cbf86cfe HD |
944 | if (!page) { |
945 | /* | |
946 | * get_ksm_page did remove_node_from_stable_tree itself. | |
947 | */ | |
948 | return 0; | |
949 | } | |
950 | ||
9a63236f AR |
951 | /* |
952 | * Page could be still mapped if this races with __mmput() running in | |
953 | * between ksm_exit() and exit_mmap(). Just refuse to let | |
954 | * merge_across_nodes/max_page_sharing be switched. | |
955 | */ | |
956 | err = -EBUSY; | |
957 | if (!page_mapped(page)) { | |
cbf86cfe | 958 | /* |
8fdb3dbf HD |
959 | * The stable node did not yet appear stale to get_ksm_page(), |
960 | * since that allows for an unmapped ksm page to be recognized | |
961 | * right up until it is freed; but the node is safe to remove. | |
1fec6890 | 962 | * This page might be in an LRU cache waiting to be freed, |
cbf86cfe HD |
963 | * or it might be PageSwapCache (perhaps under writeback), |
964 | * or it might have been removed from swapcache a moment ago. | |
965 | */ | |
966 | set_page_stable_node(page, NULL); | |
967 | remove_node_from_stable_tree(stable_node); | |
968 | err = 0; | |
969 | } | |
970 | ||
971 | unlock_page(page); | |
972 | put_page(page); | |
973 | return err; | |
974 | } | |
975 | ||
21fbd591 | 976 | static int remove_stable_node_chain(struct ksm_stable_node *stable_node, |
2c653d0e AA |
977 | struct rb_root *root) |
978 | { | |
21fbd591 | 979 | struct ksm_stable_node *dup; |
2c653d0e AA |
980 | struct hlist_node *hlist_safe; |
981 | ||
982 | if (!is_stable_node_chain(stable_node)) { | |
983 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
984 | if (remove_stable_node(stable_node)) | |
985 | return true; | |
986 | else | |
987 | return false; | |
988 | } | |
989 | ||
990 | hlist_for_each_entry_safe(dup, hlist_safe, | |
991 | &stable_node->hlist, hlist_dup) { | |
992 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
993 | if (remove_stable_node(dup)) | |
994 | return true; | |
995 | } | |
996 | BUG_ON(!hlist_empty(&stable_node->hlist)); | |
997 | free_stable_node_chain(stable_node, root); | |
998 | return false; | |
999 | } | |
1000 | ||
cbf86cfe HD |
1001 | static int remove_all_stable_nodes(void) |
1002 | { | |
21fbd591 | 1003 | struct ksm_stable_node *stable_node, *next; |
cbf86cfe HD |
1004 | int nid; |
1005 | int err = 0; | |
1006 | ||
ef53d16c | 1007 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
cbf86cfe HD |
1008 | while (root_stable_tree[nid].rb_node) { |
1009 | stable_node = rb_entry(root_stable_tree[nid].rb_node, | |
21fbd591 | 1010 | struct ksm_stable_node, node); |
2c653d0e AA |
1011 | if (remove_stable_node_chain(stable_node, |
1012 | root_stable_tree + nid)) { | |
cbf86cfe HD |
1013 | err = -EBUSY; |
1014 | break; /* proceed to next nid */ | |
1015 | } | |
1016 | cond_resched(); | |
1017 | } | |
1018 | } | |
03640418 | 1019 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
1020 | if (remove_stable_node(stable_node)) |
1021 | err = -EBUSY; | |
1022 | cond_resched(); | |
1023 | } | |
cbf86cfe HD |
1024 | return err; |
1025 | } | |
1026 | ||
d952b791 | 1027 | static int unmerge_and_remove_all_rmap_items(void) |
31dbd01f | 1028 | { |
21fbd591 | 1029 | struct ksm_mm_slot *mm_slot; |
58730ab6 | 1030 | struct mm_slot *slot; |
31dbd01f IE |
1031 | struct mm_struct *mm; |
1032 | struct vm_area_struct *vma; | |
d952b791 HD |
1033 | int err = 0; |
1034 | ||
1035 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
1036 | slot = list_entry(ksm_mm_head.slot.mm_node.next, |
1037 | struct mm_slot, mm_node); | |
1038 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
d952b791 | 1039 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f | 1040 | |
a5f18ba0 MWO |
1041 | for (mm_slot = ksm_scan.mm_slot; mm_slot != &ksm_mm_head; |
1042 | mm_slot = ksm_scan.mm_slot) { | |
58730ab6 | 1043 | VMA_ITERATOR(vmi, mm_slot->slot.mm, 0); |
a5f18ba0 | 1044 | |
58730ab6 | 1045 | mm = mm_slot->slot.mm; |
d8ed45c5 | 1046 | mmap_read_lock(mm); |
6db504ce LH |
1047 | |
1048 | /* | |
1049 | * Exit right away if mm is exiting to avoid lockdep issue in | |
1050 | * the maple tree | |
1051 | */ | |
1052 | if (ksm_test_exit(mm)) | |
1053 | goto mm_exiting; | |
1054 | ||
a5f18ba0 | 1055 | for_each_vma(vmi, vma) { |
31dbd01f IE |
1056 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
1057 | continue; | |
d952b791 | 1058 | err = unmerge_ksm_pages(vma, |
49b06385 | 1059 | vma->vm_start, vma->vm_end, false); |
9ba69294 HD |
1060 | if (err) |
1061 | goto error; | |
31dbd01f | 1062 | } |
9ba69294 | 1063 | |
6db504ce | 1064 | mm_exiting: |
420be4ed | 1065 | remove_trailing_rmap_items(&mm_slot->rmap_list); |
d8ed45c5 | 1066 | mmap_read_unlock(mm); |
d952b791 HD |
1067 | |
1068 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
1069 | slot = list_entry(mm_slot->slot.mm_node.next, |
1070 | struct mm_slot, mm_node); | |
1071 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
9ba69294 | 1072 | if (ksm_test_exit(mm)) { |
58730ab6 QZ |
1073 | hash_del(&mm_slot->slot.hash); |
1074 | list_del(&mm_slot->slot.mm_node); | |
9ba69294 HD |
1075 | spin_unlock(&ksm_mmlist_lock); |
1076 | ||
58730ab6 | 1077 | mm_slot_free(mm_slot_cache, mm_slot); |
9ba69294 | 1078 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
d7597f59 | 1079 | clear_bit(MMF_VM_MERGE_ANY, &mm->flags); |
9ba69294 | 1080 | mmdrop(mm); |
7496fea9 | 1081 | } else |
9ba69294 | 1082 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f IE |
1083 | } |
1084 | ||
cbf86cfe HD |
1085 | /* Clean up stable nodes, but don't worry if some are still busy */ |
1086 | remove_all_stable_nodes(); | |
d952b791 | 1087 | ksm_scan.seqnr = 0; |
9ba69294 HD |
1088 | return 0; |
1089 | ||
1090 | error: | |
d8ed45c5 | 1091 | mmap_read_unlock(mm); |
31dbd01f | 1092 | spin_lock(&ksm_mmlist_lock); |
d952b791 | 1093 | ksm_scan.mm_slot = &ksm_mm_head; |
31dbd01f | 1094 | spin_unlock(&ksm_mmlist_lock); |
d952b791 | 1095 | return err; |
31dbd01f | 1096 | } |
2ffd8679 | 1097 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 1098 | |
31dbd01f IE |
1099 | static u32 calc_checksum(struct page *page) |
1100 | { | |
1101 | u32 checksum; | |
9b04c5fe | 1102 | void *addr = kmap_atomic(page); |
59e1a2f4 | 1103 | checksum = xxhash(addr, PAGE_SIZE, 0); |
9b04c5fe | 1104 | kunmap_atomic(addr); |
31dbd01f IE |
1105 | return checksum; |
1106 | } | |
1107 | ||
31dbd01f IE |
1108 | static int write_protect_page(struct vm_area_struct *vma, struct page *page, |
1109 | pte_t *orig_pte) | |
1110 | { | |
1111 | struct mm_struct *mm = vma->vm_mm; | |
eed05e54 | 1112 | DEFINE_PAGE_VMA_WALK(pvmw, page, vma, 0, 0); |
31dbd01f IE |
1113 | int swapped; |
1114 | int err = -EFAULT; | |
ac46d4f3 | 1115 | struct mmu_notifier_range range; |
6c287605 | 1116 | bool anon_exclusive; |
c33c7948 | 1117 | pte_t entry; |
31dbd01f | 1118 | |
36eaff33 KS |
1119 | pvmw.address = page_address_in_vma(page, vma); |
1120 | if (pvmw.address == -EFAULT) | |
31dbd01f IE |
1121 | goto out; |
1122 | ||
29ad768c | 1123 | BUG_ON(PageTransCompound(page)); |
6bdb913f | 1124 | |
7d4a8be0 | 1125 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, pvmw.address, |
ac46d4f3 JG |
1126 | pvmw.address + PAGE_SIZE); |
1127 | mmu_notifier_invalidate_range_start(&range); | |
6bdb913f | 1128 | |
36eaff33 | 1129 | if (!page_vma_mapped_walk(&pvmw)) |
6bdb913f | 1130 | goto out_mn; |
36eaff33 KS |
1131 | if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?")) |
1132 | goto out_unlock; | |
31dbd01f | 1133 | |
6c287605 | 1134 | anon_exclusive = PageAnonExclusive(page); |
c33c7948 RR |
1135 | entry = ptep_get(pvmw.pte); |
1136 | if (pte_write(entry) || pte_dirty(entry) || | |
6c287605 | 1137 | anon_exclusive || mm_tlb_flush_pending(mm)) { |
31dbd01f | 1138 | swapped = PageSwapCache(page); |
36eaff33 | 1139 | flush_cache_page(vma, pvmw.address, page_to_pfn(page)); |
31dbd01f | 1140 | /* |
25985edc | 1141 | * Ok this is tricky, when get_user_pages_fast() run it doesn't |
31dbd01f | 1142 | * take any lock, therefore the check that we are going to make |
f0953a1b | 1143 | * with the pagecount against the mapcount is racy and |
31dbd01f IE |
1144 | * O_DIRECT can happen right after the check. |
1145 | * So we clear the pte and flush the tlb before the check | |
1146 | * this assure us that no O_DIRECT can happen after the check | |
1147 | * or in the middle of the check. | |
0f10851e JG |
1148 | * |
1149 | * No need to notify as we are downgrading page table to read | |
1150 | * only not changing it to point to a new page. | |
1151 | * | |
ee65728e | 1152 | * See Documentation/mm/mmu_notifier.rst |
31dbd01f | 1153 | */ |
0f10851e | 1154 | entry = ptep_clear_flush(vma, pvmw.address, pvmw.pte); |
31dbd01f IE |
1155 | /* |
1156 | * Check that no O_DIRECT or similar I/O is in progress on the | |
1157 | * page | |
1158 | */ | |
31e855ea | 1159 | if (page_mapcount(page) + 1 + swapped != page_count(page)) { |
36eaff33 | 1160 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f IE |
1161 | goto out_unlock; |
1162 | } | |
6c287605 | 1163 | |
088b8aa5 | 1164 | /* See page_try_share_anon_rmap(): clear PTE first. */ |
6c287605 DH |
1165 | if (anon_exclusive && page_try_share_anon_rmap(page)) { |
1166 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); | |
1167 | goto out_unlock; | |
1168 | } | |
1169 | ||
4e31635c HD |
1170 | if (pte_dirty(entry)) |
1171 | set_page_dirty(page); | |
6a56ccbc DH |
1172 | entry = pte_mkclean(entry); |
1173 | ||
1174 | if (pte_write(entry)) | |
1175 | entry = pte_wrprotect(entry); | |
595cd8f2 | 1176 | |
36eaff33 | 1177 | set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f | 1178 | } |
c33c7948 | 1179 | *orig_pte = entry; |
31dbd01f IE |
1180 | err = 0; |
1181 | ||
1182 | out_unlock: | |
36eaff33 | 1183 | page_vma_mapped_walk_done(&pvmw); |
6bdb913f | 1184 | out_mn: |
ac46d4f3 | 1185 | mmu_notifier_invalidate_range_end(&range); |
31dbd01f IE |
1186 | out: |
1187 | return err; | |
1188 | } | |
1189 | ||
1190 | /** | |
1191 | * replace_page - replace page in vma by new ksm page | |
8dd3557a HD |
1192 | * @vma: vma that holds the pte pointing to page |
1193 | * @page: the page we are replacing by kpage | |
1194 | * @kpage: the ksm page we replace page by | |
31dbd01f IE |
1195 | * @orig_pte: the original value of the pte |
1196 | * | |
1197 | * Returns 0 on success, -EFAULT on failure. | |
1198 | */ | |
8dd3557a HD |
1199 | static int replace_page(struct vm_area_struct *vma, struct page *page, |
1200 | struct page *kpage, pte_t orig_pte) | |
31dbd01f IE |
1201 | { |
1202 | struct mm_struct *mm = vma->vm_mm; | |
b4e6f66e | 1203 | struct folio *folio; |
31dbd01f | 1204 | pmd_t *pmd; |
50722804 | 1205 | pmd_t pmde; |
31dbd01f | 1206 | pte_t *ptep; |
e86c59b1 | 1207 | pte_t newpte; |
31dbd01f IE |
1208 | spinlock_t *ptl; |
1209 | unsigned long addr; | |
31dbd01f | 1210 | int err = -EFAULT; |
ac46d4f3 | 1211 | struct mmu_notifier_range range; |
31dbd01f | 1212 | |
8dd3557a | 1213 | addr = page_address_in_vma(page, vma); |
31dbd01f IE |
1214 | if (addr == -EFAULT) |
1215 | goto out; | |
1216 | ||
6219049a BL |
1217 | pmd = mm_find_pmd(mm, addr); |
1218 | if (!pmd) | |
31dbd01f | 1219 | goto out; |
50722804 ZK |
1220 | /* |
1221 | * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() | |
1222 | * without holding anon_vma lock for write. So when looking for a | |
1223 | * genuine pmde (in which to find pte), test present and !THP together. | |
1224 | */ | |
26e1a0c3 | 1225 | pmde = pmdp_get_lockless(pmd); |
50722804 ZK |
1226 | if (!pmd_present(pmde) || pmd_trans_huge(pmde)) |
1227 | goto out; | |
31dbd01f | 1228 | |
7d4a8be0 | 1229 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, addr, |
6f4f13e8 | 1230 | addr + PAGE_SIZE); |
ac46d4f3 | 1231 | mmu_notifier_invalidate_range_start(&range); |
6bdb913f | 1232 | |
31dbd01f | 1233 | ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); |
04dee9e8 HD |
1234 | if (!ptep) |
1235 | goto out_mn; | |
c33c7948 | 1236 | if (!pte_same(ptep_get(ptep), orig_pte)) { |
31dbd01f | 1237 | pte_unmap_unlock(ptep, ptl); |
6bdb913f | 1238 | goto out_mn; |
31dbd01f | 1239 | } |
6c287605 DH |
1240 | VM_BUG_ON_PAGE(PageAnonExclusive(page), page); |
1241 | VM_BUG_ON_PAGE(PageAnon(kpage) && PageAnonExclusive(kpage), kpage); | |
31dbd01f | 1242 | |
e86c59b1 CI |
1243 | /* |
1244 | * No need to check ksm_use_zero_pages here: we can only have a | |
457aef94 | 1245 | * zero_page here if ksm_use_zero_pages was enabled already. |
e86c59b1 CI |
1246 | */ |
1247 | if (!is_zero_pfn(page_to_pfn(kpage))) { | |
1248 | get_page(kpage); | |
f1e2db12 | 1249 | page_add_anon_rmap(kpage, vma, addr, RMAP_NONE); |
e86c59b1 CI |
1250 | newpte = mk_pte(kpage, vma->vm_page_prot); |
1251 | } else { | |
79271476 | 1252 | /* |
1253 | * Use pte_mkdirty to mark the zero page mapped by KSM, and then | |
1254 | * we can easily track all KSM-placed zero pages by checking if | |
1255 | * the dirty bit in zero page's PTE is set. | |
1256 | */ | |
1257 | newpte = pte_mkdirty(pte_mkspecial(pfn_pte(page_to_pfn(kpage), vma->vm_page_prot))); | |
e2942062 | 1258 | ksm_zero_pages++; |
6080d19f | 1259 | mm->ksm_zero_pages++; |
a38c015f CI |
1260 | /* |
1261 | * We're replacing an anonymous page with a zero page, which is | |
1262 | * not anonymous. We need to do proper accounting otherwise we | |
1263 | * will get wrong values in /proc, and a BUG message in dmesg | |
1264 | * when tearing down the mm. | |
1265 | */ | |
1266 | dec_mm_counter(mm, MM_ANONPAGES); | |
e86c59b1 | 1267 | } |
31dbd01f | 1268 | |
c33c7948 | 1269 | flush_cache_page(vma, addr, pte_pfn(ptep_get(ptep))); |
0f10851e JG |
1270 | /* |
1271 | * No need to notify as we are replacing a read only page with another | |
1272 | * read only page with the same content. | |
1273 | * | |
ee65728e | 1274 | * See Documentation/mm/mmu_notifier.rst |
0f10851e JG |
1275 | */ |
1276 | ptep_clear_flush(vma, addr, ptep); | |
e86c59b1 | 1277 | set_pte_at_notify(mm, addr, ptep, newpte); |
31dbd01f | 1278 | |
b4e6f66e | 1279 | folio = page_folio(page); |
cea86fe2 | 1280 | page_remove_rmap(page, vma, false); |
b4e6f66e MWO |
1281 | if (!folio_mapped(folio)) |
1282 | folio_free_swap(folio); | |
1283 | folio_put(folio); | |
31dbd01f IE |
1284 | |
1285 | pte_unmap_unlock(ptep, ptl); | |
1286 | err = 0; | |
6bdb913f | 1287 | out_mn: |
ac46d4f3 | 1288 | mmu_notifier_invalidate_range_end(&range); |
31dbd01f IE |
1289 | out: |
1290 | return err; | |
1291 | } | |
1292 | ||
1293 | /* | |
1294 | * try_to_merge_one_page - take two pages and merge them into one | |
8dd3557a HD |
1295 | * @vma: the vma that holds the pte pointing to page |
1296 | * @page: the PageAnon page that we want to replace with kpage | |
80e14822 HD |
1297 | * @kpage: the PageKsm page that we want to map instead of page, |
1298 | * or NULL the first time when we want to use page as kpage. | |
31dbd01f IE |
1299 | * |
1300 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
1301 | */ | |
1302 | static int try_to_merge_one_page(struct vm_area_struct *vma, | |
8dd3557a | 1303 | struct page *page, struct page *kpage) |
31dbd01f IE |
1304 | { |
1305 | pte_t orig_pte = __pte(0); | |
1306 | int err = -EFAULT; | |
1307 | ||
db114b83 HD |
1308 | if (page == kpage) /* ksm page forked */ |
1309 | return 0; | |
1310 | ||
8dd3557a | 1311 | if (!PageAnon(page)) |
31dbd01f IE |
1312 | goto out; |
1313 | ||
31dbd01f IE |
1314 | /* |
1315 | * We need the page lock to read a stable PageSwapCache in | |
1316 | * write_protect_page(). We use trylock_page() instead of | |
1317 | * lock_page() because we don't want to wait here - we | |
1318 | * prefer to continue scanning and merging different pages, | |
1319 | * then come back to this page when it is unlocked. | |
1320 | */ | |
8dd3557a | 1321 | if (!trylock_page(page)) |
31e855ea | 1322 | goto out; |
f765f540 KS |
1323 | |
1324 | if (PageTransCompound(page)) { | |
a7306c34 | 1325 | if (split_huge_page(page)) |
f765f540 KS |
1326 | goto out_unlock; |
1327 | } | |
1328 | ||
31dbd01f IE |
1329 | /* |
1330 | * If this anonymous page is mapped only here, its pte may need | |
1331 | * to be write-protected. If it's mapped elsewhere, all of its | |
1332 | * ptes are necessarily already write-protected. But in either | |
1333 | * case, we need to lock and check page_count is not raised. | |
1334 | */ | |
80e14822 HD |
1335 | if (write_protect_page(vma, page, &orig_pte) == 0) { |
1336 | if (!kpage) { | |
1337 | /* | |
1338 | * While we hold page lock, upgrade page from | |
1339 | * PageAnon+anon_vma to PageKsm+NULL stable_node: | |
1340 | * stable_tree_insert() will update stable_node. | |
1341 | */ | |
1342 | set_page_stable_node(page, NULL); | |
1343 | mark_page_accessed(page); | |
337ed7eb MK |
1344 | /* |
1345 | * Page reclaim just frees a clean page with no dirty | |
1346 | * ptes: make sure that the ksm page would be swapped. | |
1347 | */ | |
1348 | if (!PageDirty(page)) | |
1349 | SetPageDirty(page); | |
80e14822 HD |
1350 | err = 0; |
1351 | } else if (pages_identical(page, kpage)) | |
1352 | err = replace_page(vma, page, kpage, orig_pte); | |
1353 | } | |
31dbd01f | 1354 | |
f765f540 | 1355 | out_unlock: |
8dd3557a | 1356 | unlock_page(page); |
31dbd01f IE |
1357 | out: |
1358 | return err; | |
1359 | } | |
1360 | ||
81464e30 HD |
1361 | /* |
1362 | * try_to_merge_with_ksm_page - like try_to_merge_two_pages, | |
1363 | * but no new kernel page is allocated: kpage must already be a ksm page. | |
8dd3557a HD |
1364 | * |
1365 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
81464e30 | 1366 | */ |
21fbd591 | 1367 | static int try_to_merge_with_ksm_page(struct ksm_rmap_item *rmap_item, |
8dd3557a | 1368 | struct page *page, struct page *kpage) |
81464e30 | 1369 | { |
8dd3557a | 1370 | struct mm_struct *mm = rmap_item->mm; |
81464e30 HD |
1371 | struct vm_area_struct *vma; |
1372 | int err = -EFAULT; | |
1373 | ||
d8ed45c5 | 1374 | mmap_read_lock(mm); |
85c6e8dd AA |
1375 | vma = find_mergeable_vma(mm, rmap_item->address); |
1376 | if (!vma) | |
81464e30 HD |
1377 | goto out; |
1378 | ||
8dd3557a | 1379 | err = try_to_merge_one_page(vma, page, kpage); |
db114b83 HD |
1380 | if (err) |
1381 | goto out; | |
1382 | ||
bc56620b HD |
1383 | /* Unstable nid is in union with stable anon_vma: remove first */ |
1384 | remove_rmap_item_from_tree(rmap_item); | |
1385 | ||
c1e8d7c6 | 1386 | /* Must get reference to anon_vma while still holding mmap_lock */ |
9e60109f PZ |
1387 | rmap_item->anon_vma = vma->anon_vma; |
1388 | get_anon_vma(vma->anon_vma); | |
81464e30 | 1389 | out: |
d8ed45c5 | 1390 | mmap_read_unlock(mm); |
739100c8 SR |
1391 | trace_ksm_merge_with_ksm_page(kpage, page_to_pfn(kpage ? kpage : page), |
1392 | rmap_item, mm, err); | |
81464e30 HD |
1393 | return err; |
1394 | } | |
1395 | ||
31dbd01f IE |
1396 | /* |
1397 | * try_to_merge_two_pages - take two identical pages and prepare them | |
1398 | * to be merged into one page. | |
1399 | * | |
8dd3557a HD |
1400 | * This function returns the kpage if we successfully merged two identical |
1401 | * pages into one ksm page, NULL otherwise. | |
31dbd01f | 1402 | * |
80e14822 | 1403 | * Note that this function upgrades page to ksm page: if one of the pages |
31dbd01f IE |
1404 | * is already a ksm page, try_to_merge_with_ksm_page should be used. |
1405 | */ | |
21fbd591 | 1406 | static struct page *try_to_merge_two_pages(struct ksm_rmap_item *rmap_item, |
8dd3557a | 1407 | struct page *page, |
21fbd591 | 1408 | struct ksm_rmap_item *tree_rmap_item, |
8dd3557a | 1409 | struct page *tree_page) |
31dbd01f | 1410 | { |
80e14822 | 1411 | int err; |
31dbd01f | 1412 | |
80e14822 | 1413 | err = try_to_merge_with_ksm_page(rmap_item, page, NULL); |
31dbd01f | 1414 | if (!err) { |
8dd3557a | 1415 | err = try_to_merge_with_ksm_page(tree_rmap_item, |
80e14822 | 1416 | tree_page, page); |
31dbd01f | 1417 | /* |
81464e30 HD |
1418 | * If that fails, we have a ksm page with only one pte |
1419 | * pointing to it: so break it. | |
31dbd01f | 1420 | */ |
4035c07a | 1421 | if (err) |
8dd3557a | 1422 | break_cow(rmap_item); |
31dbd01f | 1423 | } |
80e14822 | 1424 | return err ? NULL : page; |
31dbd01f IE |
1425 | } |
1426 | ||
2c653d0e | 1427 | static __always_inline |
21fbd591 | 1428 | bool __is_page_sharing_candidate(struct ksm_stable_node *stable_node, int offset) |
2c653d0e AA |
1429 | { |
1430 | VM_BUG_ON(stable_node->rmap_hlist_len < 0); | |
1431 | /* | |
1432 | * Check that at least one mapping still exists, otherwise | |
1433 | * there's no much point to merge and share with this | |
1434 | * stable_node, as the underlying tree_page of the other | |
1435 | * sharer is going to be freed soon. | |
1436 | */ | |
1437 | return stable_node->rmap_hlist_len && | |
1438 | stable_node->rmap_hlist_len + offset < ksm_max_page_sharing; | |
1439 | } | |
1440 | ||
1441 | static __always_inline | |
21fbd591 | 1442 | bool is_page_sharing_candidate(struct ksm_stable_node *stable_node) |
2c653d0e AA |
1443 | { |
1444 | return __is_page_sharing_candidate(stable_node, 0); | |
1445 | } | |
1446 | ||
21fbd591 QZ |
1447 | static struct page *stable_node_dup(struct ksm_stable_node **_stable_node_dup, |
1448 | struct ksm_stable_node **_stable_node, | |
c01f0b54 CIK |
1449 | struct rb_root *root, |
1450 | bool prune_stale_stable_nodes) | |
2c653d0e | 1451 | { |
21fbd591 | 1452 | struct ksm_stable_node *dup, *found = NULL, *stable_node = *_stable_node; |
2c653d0e | 1453 | struct hlist_node *hlist_safe; |
8dc5ffcd | 1454 | struct page *_tree_page, *tree_page = NULL; |
2c653d0e AA |
1455 | int nr = 0; |
1456 | int found_rmap_hlist_len; | |
1457 | ||
1458 | if (!prune_stale_stable_nodes || | |
1459 | time_before(jiffies, stable_node->chain_prune_time + | |
1460 | msecs_to_jiffies( | |
1461 | ksm_stable_node_chains_prune_millisecs))) | |
1462 | prune_stale_stable_nodes = false; | |
1463 | else | |
1464 | stable_node->chain_prune_time = jiffies; | |
1465 | ||
1466 | hlist_for_each_entry_safe(dup, hlist_safe, | |
1467 | &stable_node->hlist, hlist_dup) { | |
1468 | cond_resched(); | |
1469 | /* | |
1470 | * We must walk all stable_node_dup to prune the stale | |
1471 | * stable nodes during lookup. | |
1472 | * | |
1473 | * get_ksm_page can drop the nodes from the | |
1474 | * stable_node->hlist if they point to freed pages | |
1475 | * (that's why we do a _safe walk). The "dup" | |
1476 | * stable_node parameter itself will be freed from | |
1477 | * under us if it returns NULL. | |
1478 | */ | |
2cee57d1 | 1479 | _tree_page = get_ksm_page(dup, GET_KSM_PAGE_NOLOCK); |
2c653d0e AA |
1480 | if (!_tree_page) |
1481 | continue; | |
1482 | nr += 1; | |
1483 | if (is_page_sharing_candidate(dup)) { | |
1484 | if (!found || | |
1485 | dup->rmap_hlist_len > found_rmap_hlist_len) { | |
1486 | if (found) | |
8dc5ffcd | 1487 | put_page(tree_page); |
2c653d0e AA |
1488 | found = dup; |
1489 | found_rmap_hlist_len = found->rmap_hlist_len; | |
8dc5ffcd | 1490 | tree_page = _tree_page; |
2c653d0e | 1491 | |
8dc5ffcd | 1492 | /* skip put_page for found dup */ |
2c653d0e AA |
1493 | if (!prune_stale_stable_nodes) |
1494 | break; | |
2c653d0e AA |
1495 | continue; |
1496 | } | |
1497 | } | |
1498 | put_page(_tree_page); | |
1499 | } | |
1500 | ||
80b18dfa AA |
1501 | if (found) { |
1502 | /* | |
1503 | * nr is counting all dups in the chain only if | |
1504 | * prune_stale_stable_nodes is true, otherwise we may | |
1505 | * break the loop at nr == 1 even if there are | |
1506 | * multiple entries. | |
1507 | */ | |
1508 | if (prune_stale_stable_nodes && nr == 1) { | |
2c653d0e AA |
1509 | /* |
1510 | * If there's not just one entry it would | |
1511 | * corrupt memory, better BUG_ON. In KSM | |
1512 | * context with no lock held it's not even | |
1513 | * fatal. | |
1514 | */ | |
1515 | BUG_ON(stable_node->hlist.first->next); | |
1516 | ||
1517 | /* | |
1518 | * There's just one entry and it is below the | |
1519 | * deduplication limit so drop the chain. | |
1520 | */ | |
1521 | rb_replace_node(&stable_node->node, &found->node, | |
1522 | root); | |
1523 | free_stable_node(stable_node); | |
1524 | ksm_stable_node_chains--; | |
1525 | ksm_stable_node_dups--; | |
b4fecc67 | 1526 | /* |
0ba1d0f7 AA |
1527 | * NOTE: the caller depends on the stable_node |
1528 | * to be equal to stable_node_dup if the chain | |
1529 | * was collapsed. | |
b4fecc67 | 1530 | */ |
0ba1d0f7 AA |
1531 | *_stable_node = found; |
1532 | /* | |
f0953a1b | 1533 | * Just for robustness, as stable_node is |
0ba1d0f7 AA |
1534 | * otherwise left as a stable pointer, the |
1535 | * compiler shall optimize it away at build | |
1536 | * time. | |
1537 | */ | |
1538 | stable_node = NULL; | |
80b18dfa AA |
1539 | } else if (stable_node->hlist.first != &found->hlist_dup && |
1540 | __is_page_sharing_candidate(found, 1)) { | |
2c653d0e | 1541 | /* |
80b18dfa AA |
1542 | * If the found stable_node dup can accept one |
1543 | * more future merge (in addition to the one | |
1544 | * that is underway) and is not at the head of | |
1545 | * the chain, put it there so next search will | |
1546 | * be quicker in the !prune_stale_stable_nodes | |
1547 | * case. | |
1548 | * | |
1549 | * NOTE: it would be inaccurate to use nr > 1 | |
1550 | * instead of checking the hlist.first pointer | |
1551 | * directly, because in the | |
1552 | * prune_stale_stable_nodes case "nr" isn't | |
1553 | * the position of the found dup in the chain, | |
1554 | * but the total number of dups in the chain. | |
2c653d0e AA |
1555 | */ |
1556 | hlist_del(&found->hlist_dup); | |
1557 | hlist_add_head(&found->hlist_dup, | |
1558 | &stable_node->hlist); | |
1559 | } | |
1560 | } | |
1561 | ||
8dc5ffcd AA |
1562 | *_stable_node_dup = found; |
1563 | return tree_page; | |
2c653d0e AA |
1564 | } |
1565 | ||
21fbd591 | 1566 | static struct ksm_stable_node *stable_node_dup_any(struct ksm_stable_node *stable_node, |
2c653d0e AA |
1567 | struct rb_root *root) |
1568 | { | |
1569 | if (!is_stable_node_chain(stable_node)) | |
1570 | return stable_node; | |
1571 | if (hlist_empty(&stable_node->hlist)) { | |
1572 | free_stable_node_chain(stable_node, root); | |
1573 | return NULL; | |
1574 | } | |
1575 | return hlist_entry(stable_node->hlist.first, | |
1576 | typeof(*stable_node), hlist_dup); | |
1577 | } | |
1578 | ||
8dc5ffcd AA |
1579 | /* |
1580 | * Like for get_ksm_page, this function can free the *_stable_node and | |
1581 | * *_stable_node_dup if the returned tree_page is NULL. | |
1582 | * | |
1583 | * It can also free and overwrite *_stable_node with the found | |
1584 | * stable_node_dup if the chain is collapsed (in which case | |
1585 | * *_stable_node will be equal to *_stable_node_dup like if the chain | |
1586 | * never existed). It's up to the caller to verify tree_page is not | |
1587 | * NULL before dereferencing *_stable_node or *_stable_node_dup. | |
1588 | * | |
1589 | * *_stable_node_dup is really a second output parameter of this | |
1590 | * function and will be overwritten in all cases, the caller doesn't | |
1591 | * need to initialize it. | |
1592 | */ | |
21fbd591 QZ |
1593 | static struct page *__stable_node_chain(struct ksm_stable_node **_stable_node_dup, |
1594 | struct ksm_stable_node **_stable_node, | |
8dc5ffcd AA |
1595 | struct rb_root *root, |
1596 | bool prune_stale_stable_nodes) | |
2c653d0e | 1597 | { |
21fbd591 | 1598 | struct ksm_stable_node *stable_node = *_stable_node; |
2c653d0e AA |
1599 | if (!is_stable_node_chain(stable_node)) { |
1600 | if (is_page_sharing_candidate(stable_node)) { | |
8dc5ffcd | 1601 | *_stable_node_dup = stable_node; |
2cee57d1 | 1602 | return get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK); |
2c653d0e | 1603 | } |
8dc5ffcd AA |
1604 | /* |
1605 | * _stable_node_dup set to NULL means the stable_node | |
1606 | * reached the ksm_max_page_sharing limit. | |
1607 | */ | |
1608 | *_stable_node_dup = NULL; | |
2c653d0e AA |
1609 | return NULL; |
1610 | } | |
8dc5ffcd | 1611 | return stable_node_dup(_stable_node_dup, _stable_node, root, |
2c653d0e AA |
1612 | prune_stale_stable_nodes); |
1613 | } | |
1614 | ||
21fbd591 QZ |
1615 | static __always_inline struct page *chain_prune(struct ksm_stable_node **s_n_d, |
1616 | struct ksm_stable_node **s_n, | |
8dc5ffcd | 1617 | struct rb_root *root) |
2c653d0e | 1618 | { |
8dc5ffcd | 1619 | return __stable_node_chain(s_n_d, s_n, root, true); |
2c653d0e AA |
1620 | } |
1621 | ||
21fbd591 QZ |
1622 | static __always_inline struct page *chain(struct ksm_stable_node **s_n_d, |
1623 | struct ksm_stable_node *s_n, | |
8dc5ffcd | 1624 | struct rb_root *root) |
2c653d0e | 1625 | { |
21fbd591 | 1626 | struct ksm_stable_node *old_stable_node = s_n; |
8dc5ffcd AA |
1627 | struct page *tree_page; |
1628 | ||
1629 | tree_page = __stable_node_chain(s_n_d, &s_n, root, false); | |
1630 | /* not pruning dups so s_n cannot have changed */ | |
1631 | VM_BUG_ON(s_n != old_stable_node); | |
1632 | return tree_page; | |
2c653d0e AA |
1633 | } |
1634 | ||
31dbd01f | 1635 | /* |
8dd3557a | 1636 | * stable_tree_search - search for page inside the stable tree |
31dbd01f IE |
1637 | * |
1638 | * This function checks if there is a page inside the stable tree | |
1639 | * with identical content to the page that we are scanning right now. | |
1640 | * | |
7b6ba2c7 | 1641 | * This function returns the stable tree node of identical content if found, |
31dbd01f IE |
1642 | * NULL otherwise. |
1643 | */ | |
62b61f61 | 1644 | static struct page *stable_tree_search(struct page *page) |
31dbd01f | 1645 | { |
90bd6fd3 | 1646 | int nid; |
ef53d16c | 1647 | struct rb_root *root; |
4146d2d6 HD |
1648 | struct rb_node **new; |
1649 | struct rb_node *parent; | |
21fbd591 QZ |
1650 | struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; |
1651 | struct ksm_stable_node *page_node; | |
31dbd01f | 1652 | |
4146d2d6 HD |
1653 | page_node = page_stable_node(page); |
1654 | if (page_node && page_node->head != &migrate_nodes) { | |
1655 | /* ksm page forked */ | |
08beca44 | 1656 | get_page(page); |
62b61f61 | 1657 | return page; |
08beca44 HD |
1658 | } |
1659 | ||
90bd6fd3 | 1660 | nid = get_kpfn_nid(page_to_pfn(page)); |
ef53d16c | 1661 | root = root_stable_tree + nid; |
4146d2d6 | 1662 | again: |
ef53d16c | 1663 | new = &root->rb_node; |
4146d2d6 | 1664 | parent = NULL; |
90bd6fd3 | 1665 | |
4146d2d6 | 1666 | while (*new) { |
4035c07a | 1667 | struct page *tree_page; |
31dbd01f IE |
1668 | int ret; |
1669 | ||
08beca44 | 1670 | cond_resched(); |
21fbd591 | 1671 | stable_node = rb_entry(*new, struct ksm_stable_node, node); |
2c653d0e | 1672 | stable_node_any = NULL; |
8dc5ffcd | 1673 | tree_page = chain_prune(&stable_node_dup, &stable_node, root); |
b4fecc67 AA |
1674 | /* |
1675 | * NOTE: stable_node may have been freed by | |
1676 | * chain_prune() if the returned stable_node_dup is | |
1677 | * not NULL. stable_node_dup may have been inserted in | |
1678 | * the rbtree instead as a regular stable_node (in | |
1679 | * order to collapse the stable_node chain if a single | |
0ba1d0f7 | 1680 | * stable_node dup was found in it). In such case the |
3413b2c8 | 1681 | * stable_node is overwritten by the callee to point |
0ba1d0f7 AA |
1682 | * to the stable_node_dup that was collapsed in the |
1683 | * stable rbtree and stable_node will be equal to | |
1684 | * stable_node_dup like if the chain never existed. | |
b4fecc67 | 1685 | */ |
2c653d0e AA |
1686 | if (!stable_node_dup) { |
1687 | /* | |
1688 | * Either all stable_node dups were full in | |
1689 | * this stable_node chain, or this chain was | |
1690 | * empty and should be rb_erased. | |
1691 | */ | |
1692 | stable_node_any = stable_node_dup_any(stable_node, | |
1693 | root); | |
1694 | if (!stable_node_any) { | |
1695 | /* rb_erase just run */ | |
1696 | goto again; | |
1697 | } | |
1698 | /* | |
1699 | * Take any of the stable_node dups page of | |
1700 | * this stable_node chain to let the tree walk | |
1701 | * continue. All KSM pages belonging to the | |
1702 | * stable_node dups in a stable_node chain | |
1703 | * have the same content and they're | |
457aef94 | 1704 | * write protected at all times. Any will work |
2c653d0e AA |
1705 | * fine to continue the walk. |
1706 | */ | |
2cee57d1 YS |
1707 | tree_page = get_ksm_page(stable_node_any, |
1708 | GET_KSM_PAGE_NOLOCK); | |
2c653d0e AA |
1709 | } |
1710 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1711 | if (!tree_page) { |
1712 | /* | |
1713 | * If we walked over a stale stable_node, | |
1714 | * get_ksm_page() will call rb_erase() and it | |
1715 | * may rebalance the tree from under us. So | |
1716 | * restart the search from scratch. Returning | |
1717 | * NULL would be safe too, but we'd generate | |
1718 | * false negative insertions just because some | |
1719 | * stable_node was stale. | |
1720 | */ | |
1721 | goto again; | |
1722 | } | |
31dbd01f | 1723 | |
4035c07a | 1724 | ret = memcmp_pages(page, tree_page); |
c8d6553b | 1725 | put_page(tree_page); |
31dbd01f | 1726 | |
4146d2d6 | 1727 | parent = *new; |
c8d6553b | 1728 | if (ret < 0) |
4146d2d6 | 1729 | new = &parent->rb_left; |
c8d6553b | 1730 | else if (ret > 0) |
4146d2d6 | 1731 | new = &parent->rb_right; |
c8d6553b | 1732 | else { |
2c653d0e AA |
1733 | if (page_node) { |
1734 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1735 | /* | |
1736 | * Test if the migrated page should be merged | |
1737 | * into a stable node dup. If the mapcount is | |
1738 | * 1 we can migrate it with another KSM page | |
1739 | * without adding it to the chain. | |
1740 | */ | |
1741 | if (page_mapcount(page) > 1) | |
1742 | goto chain_append; | |
1743 | } | |
1744 | ||
1745 | if (!stable_node_dup) { | |
1746 | /* | |
1747 | * If the stable_node is a chain and | |
1748 | * we got a payload match in memcmp | |
1749 | * but we cannot merge the scanned | |
1750 | * page in any of the existing | |
1751 | * stable_node dups because they're | |
1752 | * all full, we need to wait the | |
1753 | * scanned page to find itself a match | |
1754 | * in the unstable tree to create a | |
1755 | * brand new KSM page to add later to | |
1756 | * the dups of this stable_node. | |
1757 | */ | |
1758 | return NULL; | |
1759 | } | |
1760 | ||
c8d6553b HD |
1761 | /* |
1762 | * Lock and unlock the stable_node's page (which | |
1763 | * might already have been migrated) so that page | |
1764 | * migration is sure to notice its raised count. | |
1765 | * It would be more elegant to return stable_node | |
1766 | * than kpage, but that involves more changes. | |
1767 | */ | |
2cee57d1 YS |
1768 | tree_page = get_ksm_page(stable_node_dup, |
1769 | GET_KSM_PAGE_TRYLOCK); | |
1770 | ||
1771 | if (PTR_ERR(tree_page) == -EBUSY) | |
1772 | return ERR_PTR(-EBUSY); | |
1773 | ||
2c653d0e AA |
1774 | if (unlikely(!tree_page)) |
1775 | /* | |
1776 | * The tree may have been rebalanced, | |
1777 | * so re-evaluate parent and new. | |
1778 | */ | |
4146d2d6 | 1779 | goto again; |
2c653d0e AA |
1780 | unlock_page(tree_page); |
1781 | ||
1782 | if (get_kpfn_nid(stable_node_dup->kpfn) != | |
1783 | NUMA(stable_node_dup->nid)) { | |
1784 | put_page(tree_page); | |
1785 | goto replace; | |
1786 | } | |
1787 | return tree_page; | |
c8d6553b | 1788 | } |
31dbd01f IE |
1789 | } |
1790 | ||
4146d2d6 HD |
1791 | if (!page_node) |
1792 | return NULL; | |
1793 | ||
1794 | list_del(&page_node->list); | |
1795 | DO_NUMA(page_node->nid = nid); | |
1796 | rb_link_node(&page_node->node, parent, new); | |
ef53d16c | 1797 | rb_insert_color(&page_node->node, root); |
2c653d0e AA |
1798 | out: |
1799 | if (is_page_sharing_candidate(page_node)) { | |
1800 | get_page(page); | |
1801 | return page; | |
1802 | } else | |
1803 | return NULL; | |
4146d2d6 HD |
1804 | |
1805 | replace: | |
b4fecc67 AA |
1806 | /* |
1807 | * If stable_node was a chain and chain_prune collapsed it, | |
0ba1d0f7 AA |
1808 | * stable_node has been updated to be the new regular |
1809 | * stable_node. A collapse of the chain is indistinguishable | |
1810 | * from the case there was no chain in the stable | |
1811 | * rbtree. Otherwise stable_node is the chain and | |
1812 | * stable_node_dup is the dup to replace. | |
b4fecc67 | 1813 | */ |
0ba1d0f7 | 1814 | if (stable_node_dup == stable_node) { |
b4fecc67 AA |
1815 | VM_BUG_ON(is_stable_node_chain(stable_node_dup)); |
1816 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); | |
2c653d0e AA |
1817 | /* there is no chain */ |
1818 | if (page_node) { | |
1819 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1820 | list_del(&page_node->list); | |
1821 | DO_NUMA(page_node->nid = nid); | |
b4fecc67 AA |
1822 | rb_replace_node(&stable_node_dup->node, |
1823 | &page_node->node, | |
2c653d0e AA |
1824 | root); |
1825 | if (is_page_sharing_candidate(page_node)) | |
1826 | get_page(page); | |
1827 | else | |
1828 | page = NULL; | |
1829 | } else { | |
b4fecc67 | 1830 | rb_erase(&stable_node_dup->node, root); |
2c653d0e AA |
1831 | page = NULL; |
1832 | } | |
4146d2d6 | 1833 | } else { |
2c653d0e AA |
1834 | VM_BUG_ON(!is_stable_node_chain(stable_node)); |
1835 | __stable_node_dup_del(stable_node_dup); | |
1836 | if (page_node) { | |
1837 | VM_BUG_ON(page_node->head != &migrate_nodes); | |
1838 | list_del(&page_node->list); | |
1839 | DO_NUMA(page_node->nid = nid); | |
1840 | stable_node_chain_add_dup(page_node, stable_node); | |
1841 | if (is_page_sharing_candidate(page_node)) | |
1842 | get_page(page); | |
1843 | else | |
1844 | page = NULL; | |
1845 | } else { | |
1846 | page = NULL; | |
1847 | } | |
4146d2d6 | 1848 | } |
2c653d0e AA |
1849 | stable_node_dup->head = &migrate_nodes; |
1850 | list_add(&stable_node_dup->list, stable_node_dup->head); | |
4146d2d6 | 1851 | return page; |
2c653d0e AA |
1852 | |
1853 | chain_append: | |
1854 | /* stable_node_dup could be null if it reached the limit */ | |
1855 | if (!stable_node_dup) | |
1856 | stable_node_dup = stable_node_any; | |
b4fecc67 AA |
1857 | /* |
1858 | * If stable_node was a chain and chain_prune collapsed it, | |
0ba1d0f7 AA |
1859 | * stable_node has been updated to be the new regular |
1860 | * stable_node. A collapse of the chain is indistinguishable | |
1861 | * from the case there was no chain in the stable | |
1862 | * rbtree. Otherwise stable_node is the chain and | |
1863 | * stable_node_dup is the dup to replace. | |
b4fecc67 | 1864 | */ |
0ba1d0f7 | 1865 | if (stable_node_dup == stable_node) { |
b4fecc67 | 1866 | VM_BUG_ON(is_stable_node_dup(stable_node_dup)); |
2c653d0e AA |
1867 | /* chain is missing so create it */ |
1868 | stable_node = alloc_stable_node_chain(stable_node_dup, | |
1869 | root); | |
1870 | if (!stable_node) | |
1871 | return NULL; | |
1872 | } | |
1873 | /* | |
1874 | * Add this stable_node dup that was | |
1875 | * migrated to the stable_node chain | |
1876 | * of the current nid for this page | |
1877 | * content. | |
1878 | */ | |
b4fecc67 | 1879 | VM_BUG_ON(!is_stable_node_dup(stable_node_dup)); |
2c653d0e AA |
1880 | VM_BUG_ON(page_node->head != &migrate_nodes); |
1881 | list_del(&page_node->list); | |
1882 | DO_NUMA(page_node->nid = nid); | |
1883 | stable_node_chain_add_dup(page_node, stable_node); | |
1884 | goto out; | |
31dbd01f IE |
1885 | } |
1886 | ||
1887 | /* | |
e850dcf5 | 1888 | * stable_tree_insert - insert stable tree node pointing to new ksm page |
31dbd01f IE |
1889 | * into the stable tree. |
1890 | * | |
7b6ba2c7 HD |
1891 | * This function returns the stable tree node just allocated on success, |
1892 | * NULL otherwise. | |
31dbd01f | 1893 | */ |
21fbd591 | 1894 | static struct ksm_stable_node *stable_tree_insert(struct page *kpage) |
31dbd01f | 1895 | { |
90bd6fd3 PH |
1896 | int nid; |
1897 | unsigned long kpfn; | |
ef53d16c | 1898 | struct rb_root *root; |
90bd6fd3 | 1899 | struct rb_node **new; |
f2e5ff85 | 1900 | struct rb_node *parent; |
21fbd591 | 1901 | struct ksm_stable_node *stable_node, *stable_node_dup, *stable_node_any; |
2c653d0e | 1902 | bool need_chain = false; |
31dbd01f | 1903 | |
90bd6fd3 PH |
1904 | kpfn = page_to_pfn(kpage); |
1905 | nid = get_kpfn_nid(kpfn); | |
ef53d16c | 1906 | root = root_stable_tree + nid; |
f2e5ff85 AA |
1907 | again: |
1908 | parent = NULL; | |
ef53d16c | 1909 | new = &root->rb_node; |
90bd6fd3 | 1910 | |
31dbd01f | 1911 | while (*new) { |
4035c07a | 1912 | struct page *tree_page; |
31dbd01f IE |
1913 | int ret; |
1914 | ||
08beca44 | 1915 | cond_resched(); |
21fbd591 | 1916 | stable_node = rb_entry(*new, struct ksm_stable_node, node); |
2c653d0e | 1917 | stable_node_any = NULL; |
8dc5ffcd | 1918 | tree_page = chain(&stable_node_dup, stable_node, root); |
2c653d0e AA |
1919 | if (!stable_node_dup) { |
1920 | /* | |
1921 | * Either all stable_node dups were full in | |
1922 | * this stable_node chain, or this chain was | |
1923 | * empty and should be rb_erased. | |
1924 | */ | |
1925 | stable_node_any = stable_node_dup_any(stable_node, | |
1926 | root); | |
1927 | if (!stable_node_any) { | |
1928 | /* rb_erase just run */ | |
1929 | goto again; | |
1930 | } | |
1931 | /* | |
1932 | * Take any of the stable_node dups page of | |
1933 | * this stable_node chain to let the tree walk | |
1934 | * continue. All KSM pages belonging to the | |
1935 | * stable_node dups in a stable_node chain | |
1936 | * have the same content and they're | |
457aef94 | 1937 | * write protected at all times. Any will work |
2c653d0e AA |
1938 | * fine to continue the walk. |
1939 | */ | |
2cee57d1 YS |
1940 | tree_page = get_ksm_page(stable_node_any, |
1941 | GET_KSM_PAGE_NOLOCK); | |
2c653d0e AA |
1942 | } |
1943 | VM_BUG_ON(!stable_node_dup ^ !!stable_node_any); | |
f2e5ff85 AA |
1944 | if (!tree_page) { |
1945 | /* | |
1946 | * If we walked over a stale stable_node, | |
1947 | * get_ksm_page() will call rb_erase() and it | |
1948 | * may rebalance the tree from under us. So | |
1949 | * restart the search from scratch. Returning | |
1950 | * NULL would be safe too, but we'd generate | |
1951 | * false negative insertions just because some | |
1952 | * stable_node was stale. | |
1953 | */ | |
1954 | goto again; | |
1955 | } | |
31dbd01f | 1956 | |
4035c07a HD |
1957 | ret = memcmp_pages(kpage, tree_page); |
1958 | put_page(tree_page); | |
31dbd01f IE |
1959 | |
1960 | parent = *new; | |
1961 | if (ret < 0) | |
1962 | new = &parent->rb_left; | |
1963 | else if (ret > 0) | |
1964 | new = &parent->rb_right; | |
1965 | else { | |
2c653d0e AA |
1966 | need_chain = true; |
1967 | break; | |
31dbd01f IE |
1968 | } |
1969 | } | |
1970 | ||
2c653d0e AA |
1971 | stable_node_dup = alloc_stable_node(); |
1972 | if (!stable_node_dup) | |
7b6ba2c7 | 1973 | return NULL; |
31dbd01f | 1974 | |
2c653d0e AA |
1975 | INIT_HLIST_HEAD(&stable_node_dup->hlist); |
1976 | stable_node_dup->kpfn = kpfn; | |
1977 | set_page_stable_node(kpage, stable_node_dup); | |
1978 | stable_node_dup->rmap_hlist_len = 0; | |
1979 | DO_NUMA(stable_node_dup->nid = nid); | |
1980 | if (!need_chain) { | |
1981 | rb_link_node(&stable_node_dup->node, parent, new); | |
1982 | rb_insert_color(&stable_node_dup->node, root); | |
1983 | } else { | |
1984 | if (!is_stable_node_chain(stable_node)) { | |
21fbd591 | 1985 | struct ksm_stable_node *orig = stable_node; |
2c653d0e AA |
1986 | /* chain is missing so create it */ |
1987 | stable_node = alloc_stable_node_chain(orig, root); | |
1988 | if (!stable_node) { | |
1989 | free_stable_node(stable_node_dup); | |
1990 | return NULL; | |
1991 | } | |
1992 | } | |
1993 | stable_node_chain_add_dup(stable_node_dup, stable_node); | |
1994 | } | |
08beca44 | 1995 | |
2c653d0e | 1996 | return stable_node_dup; |
31dbd01f IE |
1997 | } |
1998 | ||
1999 | /* | |
8dd3557a HD |
2000 | * unstable_tree_search_insert - search for identical page, |
2001 | * else insert rmap_item into the unstable tree. | |
31dbd01f IE |
2002 | * |
2003 | * This function searches for a page in the unstable tree identical to the | |
2004 | * page currently being scanned; and if no identical page is found in the | |
2005 | * tree, we insert rmap_item as a new object into the unstable tree. | |
2006 | * | |
2007 | * This function returns pointer to rmap_item found to be identical | |
2008 | * to the currently scanned page, NULL otherwise. | |
2009 | * | |
2010 | * This function does both searching and inserting, because they share | |
2011 | * the same walking algorithm in an rbtree. | |
2012 | */ | |
8dd3557a | 2013 | static |
21fbd591 | 2014 | struct ksm_rmap_item *unstable_tree_search_insert(struct ksm_rmap_item *rmap_item, |
8dd3557a HD |
2015 | struct page *page, |
2016 | struct page **tree_pagep) | |
31dbd01f | 2017 | { |
90bd6fd3 PH |
2018 | struct rb_node **new; |
2019 | struct rb_root *root; | |
31dbd01f | 2020 | struct rb_node *parent = NULL; |
90bd6fd3 PH |
2021 | int nid; |
2022 | ||
2023 | nid = get_kpfn_nid(page_to_pfn(page)); | |
ef53d16c | 2024 | root = root_unstable_tree + nid; |
90bd6fd3 | 2025 | new = &root->rb_node; |
31dbd01f IE |
2026 | |
2027 | while (*new) { | |
21fbd591 | 2028 | struct ksm_rmap_item *tree_rmap_item; |
8dd3557a | 2029 | struct page *tree_page; |
31dbd01f IE |
2030 | int ret; |
2031 | ||
d178f27f | 2032 | cond_resched(); |
21fbd591 | 2033 | tree_rmap_item = rb_entry(*new, struct ksm_rmap_item, node); |
8dd3557a | 2034 | tree_page = get_mergeable_page(tree_rmap_item); |
c8f95ed1 | 2035 | if (!tree_page) |
31dbd01f IE |
2036 | return NULL; |
2037 | ||
2038 | /* | |
8dd3557a | 2039 | * Don't substitute a ksm page for a forked page. |
31dbd01f | 2040 | */ |
8dd3557a HD |
2041 | if (page == tree_page) { |
2042 | put_page(tree_page); | |
31dbd01f IE |
2043 | return NULL; |
2044 | } | |
2045 | ||
8dd3557a | 2046 | ret = memcmp_pages(page, tree_page); |
31dbd01f IE |
2047 | |
2048 | parent = *new; | |
2049 | if (ret < 0) { | |
8dd3557a | 2050 | put_page(tree_page); |
31dbd01f IE |
2051 | new = &parent->rb_left; |
2052 | } else if (ret > 0) { | |
8dd3557a | 2053 | put_page(tree_page); |
31dbd01f | 2054 | new = &parent->rb_right; |
b599cbdf HD |
2055 | } else if (!ksm_merge_across_nodes && |
2056 | page_to_nid(tree_page) != nid) { | |
2057 | /* | |
2058 | * If tree_page has been migrated to another NUMA node, | |
2059 | * it will be flushed out and put in the right unstable | |
2060 | * tree next time: only merge with it when across_nodes. | |
2061 | */ | |
2062 | put_page(tree_page); | |
2063 | return NULL; | |
31dbd01f | 2064 | } else { |
8dd3557a | 2065 | *tree_pagep = tree_page; |
31dbd01f IE |
2066 | return tree_rmap_item; |
2067 | } | |
2068 | } | |
2069 | ||
7b6ba2c7 | 2070 | rmap_item->address |= UNSTABLE_FLAG; |
31dbd01f | 2071 | rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); |
e850dcf5 | 2072 | DO_NUMA(rmap_item->nid = nid); |
31dbd01f | 2073 | rb_link_node(&rmap_item->node, parent, new); |
90bd6fd3 | 2074 | rb_insert_color(&rmap_item->node, root); |
31dbd01f | 2075 | |
473b0ce4 | 2076 | ksm_pages_unshared++; |
31dbd01f IE |
2077 | return NULL; |
2078 | } | |
2079 | ||
2080 | /* | |
2081 | * stable_tree_append - add another rmap_item to the linked list of | |
2082 | * rmap_items hanging off a given node of the stable tree, all sharing | |
2083 | * the same ksm page. | |
2084 | */ | |
21fbd591 QZ |
2085 | static void stable_tree_append(struct ksm_rmap_item *rmap_item, |
2086 | struct ksm_stable_node *stable_node, | |
2c653d0e | 2087 | bool max_page_sharing_bypass) |
31dbd01f | 2088 | { |
2c653d0e AA |
2089 | /* |
2090 | * rmap won't find this mapping if we don't insert the | |
2091 | * rmap_item in the right stable_node | |
2092 | * duplicate. page_migration could break later if rmap breaks, | |
2093 | * so we can as well crash here. We really need to check for | |
2094 | * rmap_hlist_len == STABLE_NODE_CHAIN, but we can as well check | |
457aef94 | 2095 | * for other negative values as an underflow if detected here |
2c653d0e AA |
2096 | * for the first time (and not when decreasing rmap_hlist_len) |
2097 | * would be sign of memory corruption in the stable_node. | |
2098 | */ | |
2099 | BUG_ON(stable_node->rmap_hlist_len < 0); | |
2100 | ||
2101 | stable_node->rmap_hlist_len++; | |
2102 | if (!max_page_sharing_bypass) | |
2103 | /* possibly non fatal but unexpected overflow, only warn */ | |
2104 | WARN_ON_ONCE(stable_node->rmap_hlist_len > | |
2105 | ksm_max_page_sharing); | |
2106 | ||
7b6ba2c7 | 2107 | rmap_item->head = stable_node; |
31dbd01f | 2108 | rmap_item->address |= STABLE_FLAG; |
7b6ba2c7 | 2109 | hlist_add_head(&rmap_item->hlist, &stable_node->hlist); |
e178dfde | 2110 | |
7b6ba2c7 HD |
2111 | if (rmap_item->hlist.next) |
2112 | ksm_pages_sharing++; | |
2113 | else | |
2114 | ksm_pages_shared++; | |
76093853 | 2115 | |
2116 | rmap_item->mm->ksm_merging_pages++; | |
31dbd01f IE |
2117 | } |
2118 | ||
2119 | /* | |
81464e30 HD |
2120 | * cmp_and_merge_page - first see if page can be merged into the stable tree; |
2121 | * if not, compare checksum to previous and if it's the same, see if page can | |
2122 | * be inserted into the unstable tree, or merged with a page already there and | |
2123 | * both transferred to the stable tree. | |
31dbd01f IE |
2124 | * |
2125 | * @page: the page that we are searching identical page to. | |
2126 | * @rmap_item: the reverse mapping into the virtual address of this page | |
2127 | */ | |
21fbd591 | 2128 | static void cmp_and_merge_page(struct page *page, struct ksm_rmap_item *rmap_item) |
31dbd01f | 2129 | { |
4b22927f | 2130 | struct mm_struct *mm = rmap_item->mm; |
21fbd591 | 2131 | struct ksm_rmap_item *tree_rmap_item; |
8dd3557a | 2132 | struct page *tree_page = NULL; |
21fbd591 | 2133 | struct ksm_stable_node *stable_node; |
8dd3557a | 2134 | struct page *kpage; |
31dbd01f IE |
2135 | unsigned int checksum; |
2136 | int err; | |
2c653d0e | 2137 | bool max_page_sharing_bypass = false; |
31dbd01f | 2138 | |
4146d2d6 HD |
2139 | stable_node = page_stable_node(page); |
2140 | if (stable_node) { | |
2141 | if (stable_node->head != &migrate_nodes && | |
2c653d0e AA |
2142 | get_kpfn_nid(READ_ONCE(stable_node->kpfn)) != |
2143 | NUMA(stable_node->nid)) { | |
2144 | stable_node_dup_del(stable_node); | |
4146d2d6 HD |
2145 | stable_node->head = &migrate_nodes; |
2146 | list_add(&stable_node->list, stable_node->head); | |
2147 | } | |
2148 | if (stable_node->head != &migrate_nodes && | |
2149 | rmap_item->head == stable_node) | |
2150 | return; | |
2c653d0e AA |
2151 | /* |
2152 | * If it's a KSM fork, allow it to go over the sharing limit | |
2153 | * without warnings. | |
2154 | */ | |
2155 | if (!is_page_sharing_candidate(stable_node)) | |
2156 | max_page_sharing_bypass = true; | |
4146d2d6 | 2157 | } |
31dbd01f IE |
2158 | |
2159 | /* We first start with searching the page inside the stable tree */ | |
62b61f61 | 2160 | kpage = stable_tree_search(page); |
4146d2d6 HD |
2161 | if (kpage == page && rmap_item->head == stable_node) { |
2162 | put_page(kpage); | |
2163 | return; | |
2164 | } | |
2165 | ||
2166 | remove_rmap_item_from_tree(rmap_item); | |
2167 | ||
62b61f61 | 2168 | if (kpage) { |
2cee57d1 YS |
2169 | if (PTR_ERR(kpage) == -EBUSY) |
2170 | return; | |
2171 | ||
08beca44 | 2172 | err = try_to_merge_with_ksm_page(rmap_item, page, kpage); |
31dbd01f IE |
2173 | if (!err) { |
2174 | /* | |
2175 | * The page was successfully merged: | |
2176 | * add its rmap_item to the stable tree. | |
2177 | */ | |
5ad64688 | 2178 | lock_page(kpage); |
2c653d0e AA |
2179 | stable_tree_append(rmap_item, page_stable_node(kpage), |
2180 | max_page_sharing_bypass); | |
5ad64688 | 2181 | unlock_page(kpage); |
31dbd01f | 2182 | } |
8dd3557a | 2183 | put_page(kpage); |
31dbd01f IE |
2184 | return; |
2185 | } | |
2186 | ||
2187 | /* | |
4035c07a HD |
2188 | * If the hash value of the page has changed from the last time |
2189 | * we calculated it, this page is changing frequently: therefore we | |
2190 | * don't want to insert it in the unstable tree, and we don't want | |
2191 | * to waste our time searching for something identical to it there. | |
31dbd01f IE |
2192 | */ |
2193 | checksum = calc_checksum(page); | |
2194 | if (rmap_item->oldchecksum != checksum) { | |
2195 | rmap_item->oldchecksum = checksum; | |
2196 | return; | |
2197 | } | |
2198 | ||
e86c59b1 CI |
2199 | /* |
2200 | * Same checksum as an empty page. We attempt to merge it with the | |
2201 | * appropriate zero page if the user enabled this via sysfs. | |
2202 | */ | |
2203 | if (ksm_use_zero_pages && (checksum == zero_checksum)) { | |
2204 | struct vm_area_struct *vma; | |
2205 | ||
d8ed45c5 | 2206 | mmap_read_lock(mm); |
4b22927f | 2207 | vma = find_mergeable_vma(mm, rmap_item->address); |
56df70a6 MS |
2208 | if (vma) { |
2209 | err = try_to_merge_one_page(vma, page, | |
2210 | ZERO_PAGE(rmap_item->address)); | |
739100c8 SR |
2211 | trace_ksm_merge_one_page( |
2212 | page_to_pfn(ZERO_PAGE(rmap_item->address)), | |
2213 | rmap_item, mm, err); | |
56df70a6 MS |
2214 | } else { |
2215 | /* | |
2216 | * If the vma is out of date, we do not need to | |
2217 | * continue. | |
2218 | */ | |
2219 | err = 0; | |
2220 | } | |
d8ed45c5 | 2221 | mmap_read_unlock(mm); |
e86c59b1 CI |
2222 | /* |
2223 | * In case of failure, the page was not really empty, so we | |
2224 | * need to continue. Otherwise we're done. | |
2225 | */ | |
2226 | if (!err) | |
2227 | return; | |
2228 | } | |
8dd3557a HD |
2229 | tree_rmap_item = |
2230 | unstable_tree_search_insert(rmap_item, page, &tree_page); | |
31dbd01f | 2231 | if (tree_rmap_item) { |
77da2ba0 CI |
2232 | bool split; |
2233 | ||
8dd3557a HD |
2234 | kpage = try_to_merge_two_pages(rmap_item, page, |
2235 | tree_rmap_item, tree_page); | |
77da2ba0 CI |
2236 | /* |
2237 | * If both pages we tried to merge belong to the same compound | |
2238 | * page, then we actually ended up increasing the reference | |
2239 | * count of the same compound page twice, and split_huge_page | |
2240 | * failed. | |
2241 | * Here we set a flag if that happened, and we use it later to | |
2242 | * try split_huge_page again. Since we call put_page right | |
2243 | * afterwards, the reference count will be correct and | |
2244 | * split_huge_page should succeed. | |
2245 | */ | |
2246 | split = PageTransCompound(page) | |
2247 | && compound_head(page) == compound_head(tree_page); | |
8dd3557a | 2248 | put_page(tree_page); |
8dd3557a | 2249 | if (kpage) { |
bc56620b HD |
2250 | /* |
2251 | * The pages were successfully merged: insert new | |
2252 | * node in the stable tree and add both rmap_items. | |
2253 | */ | |
5ad64688 | 2254 | lock_page(kpage); |
7b6ba2c7 HD |
2255 | stable_node = stable_tree_insert(kpage); |
2256 | if (stable_node) { | |
2c653d0e AA |
2257 | stable_tree_append(tree_rmap_item, stable_node, |
2258 | false); | |
2259 | stable_tree_append(rmap_item, stable_node, | |
2260 | false); | |
7b6ba2c7 | 2261 | } |
5ad64688 | 2262 | unlock_page(kpage); |
7b6ba2c7 | 2263 | |
31dbd01f IE |
2264 | /* |
2265 | * If we fail to insert the page into the stable tree, | |
2266 | * we will have 2 virtual addresses that are pointing | |
2267 | * to a ksm page left outside the stable tree, | |
2268 | * in which case we need to break_cow on both. | |
2269 | */ | |
7b6ba2c7 | 2270 | if (!stable_node) { |
8dd3557a HD |
2271 | break_cow(tree_rmap_item); |
2272 | break_cow(rmap_item); | |
31dbd01f | 2273 | } |
77da2ba0 CI |
2274 | } else if (split) { |
2275 | /* | |
2276 | * We are here if we tried to merge two pages and | |
2277 | * failed because they both belonged to the same | |
2278 | * compound page. We will split the page now, but no | |
2279 | * merging will take place. | |
2280 | * We do not want to add the cost of a full lock; if | |
2281 | * the page is locked, it is better to skip it and | |
2282 | * perhaps try again later. | |
2283 | */ | |
2284 | if (!trylock_page(page)) | |
2285 | return; | |
2286 | split_huge_page(page); | |
2287 | unlock_page(page); | |
31dbd01f | 2288 | } |
31dbd01f IE |
2289 | } |
2290 | } | |
2291 | ||
21fbd591 QZ |
2292 | static struct ksm_rmap_item *get_next_rmap_item(struct ksm_mm_slot *mm_slot, |
2293 | struct ksm_rmap_item **rmap_list, | |
31dbd01f IE |
2294 | unsigned long addr) |
2295 | { | |
21fbd591 | 2296 | struct ksm_rmap_item *rmap_item; |
31dbd01f | 2297 | |
6514d511 HD |
2298 | while (*rmap_list) { |
2299 | rmap_item = *rmap_list; | |
93d17715 | 2300 | if ((rmap_item->address & PAGE_MASK) == addr) |
31dbd01f | 2301 | return rmap_item; |
31dbd01f IE |
2302 | if (rmap_item->address > addr) |
2303 | break; | |
6514d511 | 2304 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 2305 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
2306 | free_rmap_item(rmap_item); |
2307 | } | |
2308 | ||
2309 | rmap_item = alloc_rmap_item(); | |
2310 | if (rmap_item) { | |
2311 | /* It has already been zeroed */ | |
58730ab6 | 2312 | rmap_item->mm = mm_slot->slot.mm; |
cb4df4ca | 2313 | rmap_item->mm->ksm_rmap_items++; |
31dbd01f | 2314 | rmap_item->address = addr; |
6514d511 HD |
2315 | rmap_item->rmap_list = *rmap_list; |
2316 | *rmap_list = rmap_item; | |
31dbd01f IE |
2317 | } |
2318 | return rmap_item; | |
2319 | } | |
2320 | ||
5e924ff5 SR |
2321 | /* |
2322 | * Calculate skip age for the ksm page age. The age determines how often | |
2323 | * de-duplicating has already been tried unsuccessfully. If the age is | |
2324 | * smaller, the scanning of this page is skipped for less scans. | |
2325 | * | |
2326 | * @age: rmap_item age of page | |
2327 | */ | |
2328 | static unsigned int skip_age(rmap_age_t age) | |
2329 | { | |
2330 | if (age <= 3) | |
2331 | return 1; | |
2332 | if (age <= 5) | |
2333 | return 2; | |
2334 | if (age <= 8) | |
2335 | return 4; | |
2336 | ||
2337 | return 8; | |
2338 | } | |
2339 | ||
2340 | /* | |
2341 | * Determines if a page should be skipped for the current scan. | |
2342 | * | |
2343 | * @page: page to check | |
2344 | * @rmap_item: associated rmap_item of page | |
2345 | */ | |
2346 | static bool should_skip_rmap_item(struct page *page, | |
2347 | struct ksm_rmap_item *rmap_item) | |
2348 | { | |
2349 | rmap_age_t age; | |
2350 | ||
2351 | if (!ksm_smart_scan) | |
2352 | return false; | |
2353 | ||
2354 | /* | |
2355 | * Never skip pages that are already KSM; pages cmp_and_merge_page() | |
2356 | * will essentially ignore them, but we still have to process them | |
2357 | * properly. | |
2358 | */ | |
2359 | if (PageKsm(page)) | |
2360 | return false; | |
2361 | ||
2362 | age = rmap_item->age; | |
2363 | if (age != U8_MAX) | |
2364 | rmap_item->age++; | |
2365 | ||
2366 | /* | |
2367 | * Smaller ages are not skipped, they need to get a chance to go | |
2368 | * through the different phases of the KSM merging. | |
2369 | */ | |
2370 | if (age < 3) | |
2371 | return false; | |
2372 | ||
2373 | /* | |
2374 | * Are we still allowed to skip? If not, then don't skip it | |
2375 | * and determine how much more often we are allowed to skip next. | |
2376 | */ | |
2377 | if (!rmap_item->remaining_skips) { | |
2378 | rmap_item->remaining_skips = skip_age(age); | |
2379 | return false; | |
2380 | } | |
2381 | ||
2382 | /* Skip this page */ | |
e5a68991 | 2383 | ksm_pages_skipped++; |
5e924ff5 SR |
2384 | rmap_item->remaining_skips--; |
2385 | remove_rmap_item_from_tree(rmap_item); | |
2386 | return true; | |
2387 | } | |
2388 | ||
21fbd591 | 2389 | static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page) |
31dbd01f IE |
2390 | { |
2391 | struct mm_struct *mm; | |
58730ab6 QZ |
2392 | struct ksm_mm_slot *mm_slot; |
2393 | struct mm_slot *slot; | |
31dbd01f | 2394 | struct vm_area_struct *vma; |
21fbd591 | 2395 | struct ksm_rmap_item *rmap_item; |
a5f18ba0 | 2396 | struct vma_iterator vmi; |
90bd6fd3 | 2397 | int nid; |
31dbd01f | 2398 | |
58730ab6 | 2399 | if (list_empty(&ksm_mm_head.slot.mm_node)) |
31dbd01f IE |
2400 | return NULL; |
2401 | ||
58730ab6 QZ |
2402 | mm_slot = ksm_scan.mm_slot; |
2403 | if (mm_slot == &ksm_mm_head) { | |
739100c8 SR |
2404 | trace_ksm_start_scan(ksm_scan.seqnr, ksm_rmap_items); |
2405 | ||
2919bfd0 | 2406 | /* |
1fec6890 MWO |
2407 | * A number of pages can hang around indefinitely in per-cpu |
2408 | * LRU cache, raised page count preventing write_protect_page | |
2919bfd0 HD |
2409 | * from merging them. Though it doesn't really matter much, |
2410 | * it is puzzling to see some stuck in pages_volatile until | |
2411 | * other activity jostles them out, and they also prevented | |
2412 | * LTP's KSM test from succeeding deterministically; so drain | |
2413 | * them here (here rather than on entry to ksm_do_scan(), | |
2414 | * so we don't IPI too often when pages_to_scan is set low). | |
2415 | */ | |
2416 | lru_add_drain_all(); | |
2417 | ||
4146d2d6 HD |
2418 | /* |
2419 | * Whereas stale stable_nodes on the stable_tree itself | |
2420 | * get pruned in the regular course of stable_tree_search(), | |
2421 | * those moved out to the migrate_nodes list can accumulate: | |
2422 | * so prune them once before each full scan. | |
2423 | */ | |
2424 | if (!ksm_merge_across_nodes) { | |
21fbd591 | 2425 | struct ksm_stable_node *stable_node, *next; |
4146d2d6 HD |
2426 | struct page *page; |
2427 | ||
03640418 GT |
2428 | list_for_each_entry_safe(stable_node, next, |
2429 | &migrate_nodes, list) { | |
2cee57d1 YS |
2430 | page = get_ksm_page(stable_node, |
2431 | GET_KSM_PAGE_NOLOCK); | |
4146d2d6 HD |
2432 | if (page) |
2433 | put_page(page); | |
2434 | cond_resched(); | |
2435 | } | |
2436 | } | |
2437 | ||
ef53d16c | 2438 | for (nid = 0; nid < ksm_nr_node_ids; nid++) |
90bd6fd3 | 2439 | root_unstable_tree[nid] = RB_ROOT; |
31dbd01f IE |
2440 | |
2441 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
2442 | slot = list_entry(mm_slot->slot.mm_node.next, |
2443 | struct mm_slot, mm_node); | |
2444 | mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
2445 | ksm_scan.mm_slot = mm_slot; | |
31dbd01f | 2446 | spin_unlock(&ksm_mmlist_lock); |
2b472611 HD |
2447 | /* |
2448 | * Although we tested list_empty() above, a racing __ksm_exit | |
2449 | * of the last mm on the list may have removed it since then. | |
2450 | */ | |
58730ab6 | 2451 | if (mm_slot == &ksm_mm_head) |
2b472611 | 2452 | return NULL; |
31dbd01f IE |
2453 | next_mm: |
2454 | ksm_scan.address = 0; | |
58730ab6 | 2455 | ksm_scan.rmap_list = &mm_slot->rmap_list; |
31dbd01f IE |
2456 | } |
2457 | ||
58730ab6 | 2458 | slot = &mm_slot->slot; |
31dbd01f | 2459 | mm = slot->mm; |
a5f18ba0 MWO |
2460 | vma_iter_init(&vmi, mm, ksm_scan.address); |
2461 | ||
d8ed45c5 | 2462 | mmap_read_lock(mm); |
9ba69294 | 2463 | if (ksm_test_exit(mm)) |
a5f18ba0 | 2464 | goto no_vmas; |
9ba69294 | 2465 | |
a5f18ba0 | 2466 | for_each_vma(vmi, vma) { |
31dbd01f IE |
2467 | if (!(vma->vm_flags & VM_MERGEABLE)) |
2468 | continue; | |
2469 | if (ksm_scan.address < vma->vm_start) | |
2470 | ksm_scan.address = vma->vm_start; | |
2471 | if (!vma->anon_vma) | |
2472 | ksm_scan.address = vma->vm_end; | |
2473 | ||
2474 | while (ksm_scan.address < vma->vm_end) { | |
9ba69294 HD |
2475 | if (ksm_test_exit(mm)) |
2476 | break; | |
31dbd01f | 2477 | *page = follow_page(vma, ksm_scan.address, FOLL_GET); |
f7091ed6 | 2478 | if (IS_ERR_OR_NULL(*page)) { |
21ae5b01 AA |
2479 | ksm_scan.address += PAGE_SIZE; |
2480 | cond_resched(); | |
2481 | continue; | |
2482 | } | |
f7091ed6 HW |
2483 | if (is_zone_device_page(*page)) |
2484 | goto next_page; | |
f765f540 | 2485 | if (PageAnon(*page)) { |
31dbd01f IE |
2486 | flush_anon_page(vma, *page, ksm_scan.address); |
2487 | flush_dcache_page(*page); | |
58730ab6 | 2488 | rmap_item = get_next_rmap_item(mm_slot, |
6514d511 | 2489 | ksm_scan.rmap_list, ksm_scan.address); |
31dbd01f | 2490 | if (rmap_item) { |
6514d511 HD |
2491 | ksm_scan.rmap_list = |
2492 | &rmap_item->rmap_list; | |
5e924ff5 SR |
2493 | |
2494 | if (should_skip_rmap_item(*page, rmap_item)) | |
2495 | goto next_page; | |
2496 | ||
31dbd01f IE |
2497 | ksm_scan.address += PAGE_SIZE; |
2498 | } else | |
2499 | put_page(*page); | |
d8ed45c5 | 2500 | mmap_read_unlock(mm); |
31dbd01f IE |
2501 | return rmap_item; |
2502 | } | |
f7091ed6 | 2503 | next_page: |
21ae5b01 | 2504 | put_page(*page); |
31dbd01f IE |
2505 | ksm_scan.address += PAGE_SIZE; |
2506 | cond_resched(); | |
2507 | } | |
2508 | } | |
2509 | ||
9ba69294 | 2510 | if (ksm_test_exit(mm)) { |
a5f18ba0 | 2511 | no_vmas: |
9ba69294 | 2512 | ksm_scan.address = 0; |
58730ab6 | 2513 | ksm_scan.rmap_list = &mm_slot->rmap_list; |
9ba69294 | 2514 | } |
31dbd01f IE |
2515 | /* |
2516 | * Nuke all the rmap_items that are above this current rmap: | |
2517 | * because there were no VM_MERGEABLE vmas with such addresses. | |
2518 | */ | |
420be4ed | 2519 | remove_trailing_rmap_items(ksm_scan.rmap_list); |
31dbd01f IE |
2520 | |
2521 | spin_lock(&ksm_mmlist_lock); | |
58730ab6 QZ |
2522 | slot = list_entry(mm_slot->slot.mm_node.next, |
2523 | struct mm_slot, mm_node); | |
2524 | ksm_scan.mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
cd551f97 HD |
2525 | if (ksm_scan.address == 0) { |
2526 | /* | |
c1e8d7c6 | 2527 | * We've completed a full scan of all vmas, holding mmap_lock |
cd551f97 HD |
2528 | * throughout, and found no VM_MERGEABLE: so do the same as |
2529 | * __ksm_exit does to remove this mm from all our lists now. | |
9ba69294 HD |
2530 | * This applies either when cleaning up after __ksm_exit |
2531 | * (but beware: we can reach here even before __ksm_exit), | |
2532 | * or when all VM_MERGEABLE areas have been unmapped (and | |
c1e8d7c6 | 2533 | * mmap_lock then protects against race with MADV_MERGEABLE). |
cd551f97 | 2534 | */ |
58730ab6 QZ |
2535 | hash_del(&mm_slot->slot.hash); |
2536 | list_del(&mm_slot->slot.mm_node); | |
9ba69294 HD |
2537 | spin_unlock(&ksm_mmlist_lock); |
2538 | ||
58730ab6 | 2539 | mm_slot_free(mm_slot_cache, mm_slot); |
cd551f97 | 2540 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
d7597f59 | 2541 | clear_bit(MMF_VM_MERGE_ANY, &mm->flags); |
d8ed45c5 | 2542 | mmap_read_unlock(mm); |
9ba69294 HD |
2543 | mmdrop(mm); |
2544 | } else { | |
d8ed45c5 | 2545 | mmap_read_unlock(mm); |
7496fea9 | 2546 | /* |
3e4e28c5 | 2547 | * mmap_read_unlock(mm) first because after |
7496fea9 ZC |
2548 | * spin_unlock(&ksm_mmlist_lock) run, the "mm" may |
2549 | * already have been freed under us by __ksm_exit() | |
2550 | * because the "mm_slot" is still hashed and | |
2551 | * ksm_scan.mm_slot doesn't point to it anymore. | |
2552 | */ | |
2553 | spin_unlock(&ksm_mmlist_lock); | |
cd551f97 | 2554 | } |
31dbd01f IE |
2555 | |
2556 | /* Repeat until we've completed scanning the whole list */ | |
58730ab6 QZ |
2557 | mm_slot = ksm_scan.mm_slot; |
2558 | if (mm_slot != &ksm_mm_head) | |
31dbd01f IE |
2559 | goto next_mm; |
2560 | ||
739100c8 | 2561 | trace_ksm_stop_scan(ksm_scan.seqnr, ksm_rmap_items); |
31dbd01f IE |
2562 | ksm_scan.seqnr++; |
2563 | return NULL; | |
2564 | } | |
2565 | ||
2566 | /** | |
2567 | * ksm_do_scan - the ksm scanner main worker function. | |
b7701a5f | 2568 | * @scan_npages: number of pages we want to scan before we return. |
31dbd01f IE |
2569 | */ |
2570 | static void ksm_do_scan(unsigned int scan_npages) | |
2571 | { | |
21fbd591 | 2572 | struct ksm_rmap_item *rmap_item; |
3f649ab7 | 2573 | struct page *page; |
b348b5fe | 2574 | unsigned int npages = scan_npages; |
31dbd01f | 2575 | |
b348b5fe | 2576 | while (npages-- && likely(!freezing(current))) { |
31dbd01f IE |
2577 | cond_resched(); |
2578 | rmap_item = scan_get_next_rmap_item(&page); | |
2579 | if (!rmap_item) | |
2580 | return; | |
4146d2d6 | 2581 | cmp_and_merge_page(page, rmap_item); |
31dbd01f IE |
2582 | put_page(page); |
2583 | } | |
b348b5fe SR |
2584 | |
2585 | ksm_pages_scanned += scan_npages - npages; | |
31dbd01f IE |
2586 | } |
2587 | ||
6e158384 HD |
2588 | static int ksmd_should_run(void) |
2589 | { | |
58730ab6 | 2590 | return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.slot.mm_node); |
6e158384 HD |
2591 | } |
2592 | ||
31dbd01f IE |
2593 | static int ksm_scan_thread(void *nothing) |
2594 | { | |
fcf9a0ef KT |
2595 | unsigned int sleep_ms; |
2596 | ||
878aee7d | 2597 | set_freezable(); |
339aa624 | 2598 | set_user_nice(current, 5); |
31dbd01f IE |
2599 | |
2600 | while (!kthread_should_stop()) { | |
6e158384 | 2601 | mutex_lock(&ksm_thread_mutex); |
ef4d43a8 | 2602 | wait_while_offlining(); |
6e158384 | 2603 | if (ksmd_should_run()) |
31dbd01f | 2604 | ksm_do_scan(ksm_thread_pages_to_scan); |
6e158384 HD |
2605 | mutex_unlock(&ksm_thread_mutex); |
2606 | ||
878aee7d AA |
2607 | try_to_freeze(); |
2608 | ||
6e158384 | 2609 | if (ksmd_should_run()) { |
fcf9a0ef KT |
2610 | sleep_ms = READ_ONCE(ksm_thread_sleep_millisecs); |
2611 | wait_event_interruptible_timeout(ksm_iter_wait, | |
2612 | sleep_ms != READ_ONCE(ksm_thread_sleep_millisecs), | |
2613 | msecs_to_jiffies(sleep_ms)); | |
31dbd01f | 2614 | } else { |
878aee7d | 2615 | wait_event_freezable(ksm_thread_wait, |
6e158384 | 2616 | ksmd_should_run() || kthread_should_stop()); |
31dbd01f IE |
2617 | } |
2618 | } | |
2619 | return 0; | |
2620 | } | |
2621 | ||
d7597f59 SR |
2622 | static void __ksm_add_vma(struct vm_area_struct *vma) |
2623 | { | |
2624 | unsigned long vm_flags = vma->vm_flags; | |
2625 | ||
2626 | if (vm_flags & VM_MERGEABLE) | |
2627 | return; | |
2628 | ||
2629 | if (vma_ksm_compatible(vma)) | |
2630 | vm_flags_set(vma, VM_MERGEABLE); | |
2631 | } | |
2632 | ||
24139c07 DH |
2633 | static int __ksm_del_vma(struct vm_area_struct *vma) |
2634 | { | |
2635 | int err; | |
2636 | ||
2637 | if (!(vma->vm_flags & VM_MERGEABLE)) | |
2638 | return 0; | |
2639 | ||
2640 | if (vma->anon_vma) { | |
49b06385 | 2641 | err = unmerge_ksm_pages(vma, vma->vm_start, vma->vm_end, true); |
24139c07 DH |
2642 | if (err) |
2643 | return err; | |
2644 | } | |
2645 | ||
2646 | vm_flags_clear(vma, VM_MERGEABLE); | |
2647 | return 0; | |
2648 | } | |
d7597f59 SR |
2649 | /** |
2650 | * ksm_add_vma - Mark vma as mergeable if compatible | |
2651 | * | |
2652 | * @vma: Pointer to vma | |
2653 | */ | |
2654 | void ksm_add_vma(struct vm_area_struct *vma) | |
2655 | { | |
2656 | struct mm_struct *mm = vma->vm_mm; | |
2657 | ||
2658 | if (test_bit(MMF_VM_MERGE_ANY, &mm->flags)) | |
2659 | __ksm_add_vma(vma); | |
2660 | } | |
2661 | ||
2662 | static void ksm_add_vmas(struct mm_struct *mm) | |
2663 | { | |
2664 | struct vm_area_struct *vma; | |
2665 | ||
2666 | VMA_ITERATOR(vmi, mm, 0); | |
2667 | for_each_vma(vmi, vma) | |
2668 | __ksm_add_vma(vma); | |
2669 | } | |
2670 | ||
24139c07 DH |
2671 | static int ksm_del_vmas(struct mm_struct *mm) |
2672 | { | |
2673 | struct vm_area_struct *vma; | |
2674 | int err; | |
2675 | ||
2676 | VMA_ITERATOR(vmi, mm, 0); | |
2677 | for_each_vma(vmi, vma) { | |
2678 | err = __ksm_del_vma(vma); | |
2679 | if (err) | |
2680 | return err; | |
2681 | } | |
2682 | return 0; | |
2683 | } | |
2684 | ||
d7597f59 SR |
2685 | /** |
2686 | * ksm_enable_merge_any - Add mm to mm ksm list and enable merging on all | |
2687 | * compatible VMA's | |
2688 | * | |
2689 | * @mm: Pointer to mm | |
2690 | * | |
2691 | * Returns 0 on success, otherwise error code | |
2692 | */ | |
2693 | int ksm_enable_merge_any(struct mm_struct *mm) | |
2694 | { | |
2695 | int err; | |
2696 | ||
2697 | if (test_bit(MMF_VM_MERGE_ANY, &mm->flags)) | |
2698 | return 0; | |
2699 | ||
2700 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { | |
2701 | err = __ksm_enter(mm); | |
2702 | if (err) | |
2703 | return err; | |
2704 | } | |
2705 | ||
2706 | set_bit(MMF_VM_MERGE_ANY, &mm->flags); | |
2707 | ksm_add_vmas(mm); | |
2708 | ||
2709 | return 0; | |
2710 | } | |
2711 | ||
24139c07 DH |
2712 | /** |
2713 | * ksm_disable_merge_any - Disable merging on all compatible VMA's of the mm, | |
2714 | * previously enabled via ksm_enable_merge_any(). | |
2715 | * | |
2716 | * Disabling merging implies unmerging any merged pages, like setting | |
2717 | * MADV_UNMERGEABLE would. If unmerging fails, the whole operation fails and | |
2718 | * merging on all compatible VMA's remains enabled. | |
2719 | * | |
2720 | * @mm: Pointer to mm | |
2721 | * | |
2722 | * Returns 0 on success, otherwise error code | |
2723 | */ | |
2724 | int ksm_disable_merge_any(struct mm_struct *mm) | |
2725 | { | |
2726 | int err; | |
2727 | ||
2728 | if (!test_bit(MMF_VM_MERGE_ANY, &mm->flags)) | |
2729 | return 0; | |
2730 | ||
2731 | err = ksm_del_vmas(mm); | |
2732 | if (err) { | |
2733 | ksm_add_vmas(mm); | |
2734 | return err; | |
2735 | } | |
2736 | ||
2737 | clear_bit(MMF_VM_MERGE_ANY, &mm->flags); | |
2738 | return 0; | |
2739 | } | |
2740 | ||
2c281f54 DH |
2741 | int ksm_disable(struct mm_struct *mm) |
2742 | { | |
2743 | mmap_assert_write_locked(mm); | |
2744 | ||
2745 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) | |
2746 | return 0; | |
2747 | if (test_bit(MMF_VM_MERGE_ANY, &mm->flags)) | |
2748 | return ksm_disable_merge_any(mm); | |
2749 | return ksm_del_vmas(mm); | |
2750 | } | |
2751 | ||
f8af4da3 HD |
2752 | int ksm_madvise(struct vm_area_struct *vma, unsigned long start, |
2753 | unsigned long end, int advice, unsigned long *vm_flags) | |
2754 | { | |
2755 | struct mm_struct *mm = vma->vm_mm; | |
d952b791 | 2756 | int err; |
f8af4da3 HD |
2757 | |
2758 | switch (advice) { | |
2759 | case MADV_MERGEABLE: | |
d7597f59 | 2760 | if (vma->vm_flags & VM_MERGEABLE) |
e1fb4a08 | 2761 | return 0; |
d7597f59 | 2762 | if (!vma_ksm_compatible(vma)) |
74a04967 | 2763 | return 0; |
cc2383ec | 2764 | |
d952b791 HD |
2765 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { |
2766 | err = __ksm_enter(mm); | |
2767 | if (err) | |
2768 | return err; | |
2769 | } | |
f8af4da3 HD |
2770 | |
2771 | *vm_flags |= VM_MERGEABLE; | |
2772 | break; | |
2773 | ||
2774 | case MADV_UNMERGEABLE: | |
2775 | if (!(*vm_flags & VM_MERGEABLE)) | |
2776 | return 0; /* just ignore the advice */ | |
2777 | ||
d952b791 | 2778 | if (vma->anon_vma) { |
49b06385 | 2779 | err = unmerge_ksm_pages(vma, start, end, true); |
d952b791 HD |
2780 | if (err) |
2781 | return err; | |
2782 | } | |
f8af4da3 HD |
2783 | |
2784 | *vm_flags &= ~VM_MERGEABLE; | |
2785 | break; | |
2786 | } | |
2787 | ||
2788 | return 0; | |
2789 | } | |
33cf1707 | 2790 | EXPORT_SYMBOL_GPL(ksm_madvise); |
f8af4da3 HD |
2791 | |
2792 | int __ksm_enter(struct mm_struct *mm) | |
2793 | { | |
21fbd591 | 2794 | struct ksm_mm_slot *mm_slot; |
58730ab6 | 2795 | struct mm_slot *slot; |
6e158384 HD |
2796 | int needs_wakeup; |
2797 | ||
58730ab6 | 2798 | mm_slot = mm_slot_alloc(mm_slot_cache); |
31dbd01f IE |
2799 | if (!mm_slot) |
2800 | return -ENOMEM; | |
2801 | ||
58730ab6 QZ |
2802 | slot = &mm_slot->slot; |
2803 | ||
6e158384 | 2804 | /* Check ksm_run too? Would need tighter locking */ |
58730ab6 | 2805 | needs_wakeup = list_empty(&ksm_mm_head.slot.mm_node); |
6e158384 | 2806 | |
31dbd01f | 2807 | spin_lock(&ksm_mmlist_lock); |
58730ab6 | 2808 | mm_slot_insert(mm_slots_hash, mm, slot); |
31dbd01f | 2809 | /* |
cbf86cfe HD |
2810 | * When KSM_RUN_MERGE (or KSM_RUN_STOP), |
2811 | * insert just behind the scanning cursor, to let the area settle | |
31dbd01f IE |
2812 | * down a little; when fork is followed by immediate exec, we don't |
2813 | * want ksmd to waste time setting up and tearing down an rmap_list. | |
cbf86cfe HD |
2814 | * |
2815 | * But when KSM_RUN_UNMERGE, it's important to insert ahead of its | |
2816 | * scanning cursor, otherwise KSM pages in newly forked mms will be | |
2817 | * missed: then we might as well insert at the end of the list. | |
31dbd01f | 2818 | */ |
cbf86cfe | 2819 | if (ksm_run & KSM_RUN_UNMERGE) |
58730ab6 | 2820 | list_add_tail(&slot->mm_node, &ksm_mm_head.slot.mm_node); |
cbf86cfe | 2821 | else |
58730ab6 | 2822 | list_add_tail(&slot->mm_node, &ksm_scan.mm_slot->slot.mm_node); |
31dbd01f IE |
2823 | spin_unlock(&ksm_mmlist_lock); |
2824 | ||
f8af4da3 | 2825 | set_bit(MMF_VM_MERGEABLE, &mm->flags); |
f1f10076 | 2826 | mmgrab(mm); |
6e158384 HD |
2827 | |
2828 | if (needs_wakeup) | |
2829 | wake_up_interruptible(&ksm_thread_wait); | |
2830 | ||
739100c8 | 2831 | trace_ksm_enter(mm); |
f8af4da3 HD |
2832 | return 0; |
2833 | } | |
2834 | ||
1c2fb7a4 | 2835 | void __ksm_exit(struct mm_struct *mm) |
f8af4da3 | 2836 | { |
21fbd591 | 2837 | struct ksm_mm_slot *mm_slot; |
58730ab6 | 2838 | struct mm_slot *slot; |
9ba69294 | 2839 | int easy_to_free = 0; |
cd551f97 | 2840 | |
31dbd01f | 2841 | /* |
9ba69294 HD |
2842 | * This process is exiting: if it's straightforward (as is the |
2843 | * case when ksmd was never running), free mm_slot immediately. | |
2844 | * But if it's at the cursor or has rmap_items linked to it, use | |
c1e8d7c6 | 2845 | * mmap_lock to synchronize with any break_cows before pagetables |
9ba69294 HD |
2846 | * are freed, and leave the mm_slot on the list for ksmd to free. |
2847 | * Beware: ksm may already have noticed it exiting and freed the slot. | |
31dbd01f | 2848 | */ |
9ba69294 | 2849 | |
cd551f97 | 2850 | spin_lock(&ksm_mmlist_lock); |
58730ab6 QZ |
2851 | slot = mm_slot_lookup(mm_slots_hash, mm); |
2852 | mm_slot = mm_slot_entry(slot, struct ksm_mm_slot, slot); | |
9ba69294 | 2853 | if (mm_slot && ksm_scan.mm_slot != mm_slot) { |
6514d511 | 2854 | if (!mm_slot->rmap_list) { |
58730ab6 QZ |
2855 | hash_del(&slot->hash); |
2856 | list_del(&slot->mm_node); | |
9ba69294 HD |
2857 | easy_to_free = 1; |
2858 | } else { | |
58730ab6 QZ |
2859 | list_move(&slot->mm_node, |
2860 | &ksm_scan.mm_slot->slot.mm_node); | |
9ba69294 | 2861 | } |
cd551f97 | 2862 | } |
cd551f97 HD |
2863 | spin_unlock(&ksm_mmlist_lock); |
2864 | ||
9ba69294 | 2865 | if (easy_to_free) { |
58730ab6 | 2866 | mm_slot_free(mm_slot_cache, mm_slot); |
d7597f59 | 2867 | clear_bit(MMF_VM_MERGE_ANY, &mm->flags); |
9ba69294 HD |
2868 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); |
2869 | mmdrop(mm); | |
2870 | } else if (mm_slot) { | |
d8ed45c5 ML |
2871 | mmap_write_lock(mm); |
2872 | mmap_write_unlock(mm); | |
9ba69294 | 2873 | } |
739100c8 SR |
2874 | |
2875 | trace_ksm_exit(mm); | |
31dbd01f IE |
2876 | } |
2877 | ||
cbf86cfe | 2878 | struct page *ksm_might_need_to_copy(struct page *page, |
5ad64688 HD |
2879 | struct vm_area_struct *vma, unsigned long address) |
2880 | { | |
e05b3453 MWO |
2881 | struct folio *folio = page_folio(page); |
2882 | struct anon_vma *anon_vma = folio_anon_vma(folio); | |
5ad64688 HD |
2883 | struct page *new_page; |
2884 | ||
cbf86cfe HD |
2885 | if (PageKsm(page)) { |
2886 | if (page_stable_node(page) && | |
2887 | !(ksm_run & KSM_RUN_UNMERGE)) | |
2888 | return page; /* no need to copy it */ | |
2889 | } else if (!anon_vma) { | |
2890 | return page; /* no need to copy it */ | |
e1c63e11 NS |
2891 | } else if (page->index == linear_page_index(vma, address) && |
2892 | anon_vma->root == vma->anon_vma->root) { | |
cbf86cfe HD |
2893 | return page; /* still no need to copy it */ |
2894 | } | |
f985fc32 ML |
2895 | if (PageHWPoison(page)) |
2896 | return ERR_PTR(-EHWPOISON); | |
cbf86cfe HD |
2897 | if (!PageUptodate(page)) |
2898 | return page; /* let do_swap_page report the error */ | |
2899 | ||
5ad64688 | 2900 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
8f425e4e MWO |
2901 | if (new_page && |
2902 | mem_cgroup_charge(page_folio(new_page), vma->vm_mm, GFP_KERNEL)) { | |
62fdb163 HD |
2903 | put_page(new_page); |
2904 | new_page = NULL; | |
2905 | } | |
5ad64688 | 2906 | if (new_page) { |
6b970599 KW |
2907 | if (copy_mc_user_highpage(new_page, page, address, vma)) { |
2908 | put_page(new_page); | |
2909 | memory_failure_queue(page_to_pfn(page), 0); | |
2910 | return ERR_PTR(-EHWPOISON); | |
2911 | } | |
5ad64688 HD |
2912 | SetPageDirty(new_page); |
2913 | __SetPageUptodate(new_page); | |
48c935ad | 2914 | __SetPageLocked(new_page); |
4d45c3af YY |
2915 | #ifdef CONFIG_SWAP |
2916 | count_vm_event(KSM_SWPIN_COPY); | |
2917 | #endif | |
5ad64688 HD |
2918 | } |
2919 | ||
5ad64688 HD |
2920 | return new_page; |
2921 | } | |
2922 | ||
6d4675e6 | 2923 | void rmap_walk_ksm(struct folio *folio, struct rmap_walk_control *rwc) |
e9995ef9 | 2924 | { |
21fbd591 QZ |
2925 | struct ksm_stable_node *stable_node; |
2926 | struct ksm_rmap_item *rmap_item; | |
e9995ef9 HD |
2927 | int search_new_forks = 0; |
2928 | ||
2f031c6f | 2929 | VM_BUG_ON_FOLIO(!folio_test_ksm(folio), folio); |
9f32624b JK |
2930 | |
2931 | /* | |
2932 | * Rely on the page lock to protect against concurrent modifications | |
2933 | * to that page's node of the stable tree. | |
2934 | */ | |
2f031c6f | 2935 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
e9995ef9 | 2936 | |
2f031c6f | 2937 | stable_node = folio_stable_node(folio); |
e9995ef9 | 2938 | if (!stable_node) |
1df631ae | 2939 | return; |
e9995ef9 | 2940 | again: |
b67bfe0d | 2941 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
e9995ef9 | 2942 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
5beb4930 | 2943 | struct anon_vma_chain *vmac; |
e9995ef9 HD |
2944 | struct vm_area_struct *vma; |
2945 | ||
ad12695f | 2946 | cond_resched(); |
6d4675e6 MK |
2947 | if (!anon_vma_trylock_read(anon_vma)) { |
2948 | if (rwc->try_lock) { | |
2949 | rwc->contended = true; | |
2950 | return; | |
2951 | } | |
2952 | anon_vma_lock_read(anon_vma); | |
2953 | } | |
bf181b9f ML |
2954 | anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, |
2955 | 0, ULONG_MAX) { | |
1105a2fc JH |
2956 | unsigned long addr; |
2957 | ||
ad12695f | 2958 | cond_resched(); |
5beb4930 | 2959 | vma = vmac->vma; |
1105a2fc JH |
2960 | |
2961 | /* Ignore the stable/unstable/sqnr flags */ | |
cd7fae26 | 2962 | addr = rmap_item->address & PAGE_MASK; |
1105a2fc JH |
2963 | |
2964 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
e9995ef9 HD |
2965 | continue; |
2966 | /* | |
2967 | * Initially we examine only the vma which covers this | |
2968 | * rmap_item; but later, if there is still work to do, | |
2969 | * we examine covering vmas in other mms: in case they | |
2970 | * were forked from the original since ksmd passed. | |
2971 | */ | |
2972 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
2973 | continue; | |
2974 | ||
0dd1c7bb JK |
2975 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
2976 | continue; | |
2977 | ||
2f031c6f | 2978 | if (!rwc->rmap_one(folio, vma, addr, rwc->arg)) { |
b6b19f25 | 2979 | anon_vma_unlock_read(anon_vma); |
1df631ae | 2980 | return; |
e9995ef9 | 2981 | } |
2f031c6f | 2982 | if (rwc->done && rwc->done(folio)) { |
0dd1c7bb | 2983 | anon_vma_unlock_read(anon_vma); |
1df631ae | 2984 | return; |
0dd1c7bb | 2985 | } |
e9995ef9 | 2986 | } |
b6b19f25 | 2987 | anon_vma_unlock_read(anon_vma); |
e9995ef9 HD |
2988 | } |
2989 | if (!search_new_forks++) | |
2990 | goto again; | |
e9995ef9 HD |
2991 | } |
2992 | ||
4248d008 LX |
2993 | #ifdef CONFIG_MEMORY_FAILURE |
2994 | /* | |
2995 | * Collect processes when the error hit an ksm page. | |
2996 | */ | |
2997 | void collect_procs_ksm(struct page *page, struct list_head *to_kill, | |
2998 | int force_early) | |
2999 | { | |
3000 | struct ksm_stable_node *stable_node; | |
3001 | struct ksm_rmap_item *rmap_item; | |
3002 | struct folio *folio = page_folio(page); | |
3003 | struct vm_area_struct *vma; | |
3004 | struct task_struct *tsk; | |
3005 | ||
3006 | stable_node = folio_stable_node(folio); | |
3007 | if (!stable_node) | |
3008 | return; | |
3009 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { | |
3010 | struct anon_vma *av = rmap_item->anon_vma; | |
3011 | ||
3012 | anon_vma_lock_read(av); | |
d256d1cd | 3013 | rcu_read_lock(); |
4248d008 LX |
3014 | for_each_process(tsk) { |
3015 | struct anon_vma_chain *vmac; | |
3016 | unsigned long addr; | |
3017 | struct task_struct *t = | |
3018 | task_early_kill(tsk, force_early); | |
3019 | if (!t) | |
3020 | continue; | |
3021 | anon_vma_interval_tree_foreach(vmac, &av->rb_root, 0, | |
3022 | ULONG_MAX) | |
3023 | { | |
3024 | vma = vmac->vma; | |
3025 | if (vma->vm_mm == t->mm) { | |
3026 | addr = rmap_item->address & PAGE_MASK; | |
3027 | add_to_kill_ksm(t, page, vma, to_kill, | |
3028 | addr); | |
3029 | } | |
3030 | } | |
3031 | } | |
d256d1cd | 3032 | rcu_read_unlock(); |
4248d008 LX |
3033 | anon_vma_unlock_read(av); |
3034 | } | |
3035 | } | |
3036 | #endif | |
3037 | ||
52629506 | 3038 | #ifdef CONFIG_MIGRATION |
19138349 | 3039 | void folio_migrate_ksm(struct folio *newfolio, struct folio *folio) |
e9995ef9 | 3040 | { |
21fbd591 | 3041 | struct ksm_stable_node *stable_node; |
e9995ef9 | 3042 | |
19138349 MWO |
3043 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
3044 | VM_BUG_ON_FOLIO(!folio_test_locked(newfolio), newfolio); | |
3045 | VM_BUG_ON_FOLIO(newfolio->mapping != folio->mapping, newfolio); | |
e9995ef9 | 3046 | |
19138349 | 3047 | stable_node = folio_stable_node(folio); |
e9995ef9 | 3048 | if (stable_node) { |
19138349 MWO |
3049 | VM_BUG_ON_FOLIO(stable_node->kpfn != folio_pfn(folio), folio); |
3050 | stable_node->kpfn = folio_pfn(newfolio); | |
c8d6553b | 3051 | /* |
19138349 | 3052 | * newfolio->mapping was set in advance; now we need smp_wmb() |
c8d6553b | 3053 | * to make sure that the new stable_node->kpfn is visible |
19138349 MWO |
3054 | * to get_ksm_page() before it can see that folio->mapping |
3055 | * has gone stale (or that folio_test_swapcache has been cleared). | |
c8d6553b HD |
3056 | */ |
3057 | smp_wmb(); | |
19138349 | 3058 | set_page_stable_node(&folio->page, NULL); |
e9995ef9 HD |
3059 | } |
3060 | } | |
3061 | #endif /* CONFIG_MIGRATION */ | |
3062 | ||
62b61f61 | 3063 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 HD |
3064 | static void wait_while_offlining(void) |
3065 | { | |
3066 | while (ksm_run & KSM_RUN_OFFLINE) { | |
3067 | mutex_unlock(&ksm_thread_mutex); | |
3068 | wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), | |
74316201 | 3069 | TASK_UNINTERRUPTIBLE); |
ef4d43a8 HD |
3070 | mutex_lock(&ksm_thread_mutex); |
3071 | } | |
3072 | } | |
3073 | ||
21fbd591 | 3074 | static bool stable_node_dup_remove_range(struct ksm_stable_node *stable_node, |
2c653d0e AA |
3075 | unsigned long start_pfn, |
3076 | unsigned long end_pfn) | |
3077 | { | |
3078 | if (stable_node->kpfn >= start_pfn && | |
3079 | stable_node->kpfn < end_pfn) { | |
3080 | /* | |
3081 | * Don't get_ksm_page, page has already gone: | |
3082 | * which is why we keep kpfn instead of page* | |
3083 | */ | |
3084 | remove_node_from_stable_tree(stable_node); | |
3085 | return true; | |
3086 | } | |
3087 | return false; | |
3088 | } | |
3089 | ||
21fbd591 | 3090 | static bool stable_node_chain_remove_range(struct ksm_stable_node *stable_node, |
2c653d0e AA |
3091 | unsigned long start_pfn, |
3092 | unsigned long end_pfn, | |
3093 | struct rb_root *root) | |
3094 | { | |
21fbd591 | 3095 | struct ksm_stable_node *dup; |
2c653d0e AA |
3096 | struct hlist_node *hlist_safe; |
3097 | ||
3098 | if (!is_stable_node_chain(stable_node)) { | |
3099 | VM_BUG_ON(is_stable_node_dup(stable_node)); | |
3100 | return stable_node_dup_remove_range(stable_node, start_pfn, | |
3101 | end_pfn); | |
3102 | } | |
3103 | ||
3104 | hlist_for_each_entry_safe(dup, hlist_safe, | |
3105 | &stable_node->hlist, hlist_dup) { | |
3106 | VM_BUG_ON(!is_stable_node_dup(dup)); | |
3107 | stable_node_dup_remove_range(dup, start_pfn, end_pfn); | |
3108 | } | |
3109 | if (hlist_empty(&stable_node->hlist)) { | |
3110 | free_stable_node_chain(stable_node, root); | |
3111 | return true; /* notify caller that tree was rebalanced */ | |
3112 | } else | |
3113 | return false; | |
3114 | } | |
3115 | ||
ee0ea59c HD |
3116 | static void ksm_check_stable_tree(unsigned long start_pfn, |
3117 | unsigned long end_pfn) | |
62b61f61 | 3118 | { |
21fbd591 | 3119 | struct ksm_stable_node *stable_node, *next; |
62b61f61 | 3120 | struct rb_node *node; |
90bd6fd3 | 3121 | int nid; |
62b61f61 | 3122 | |
ef53d16c HD |
3123 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
3124 | node = rb_first(root_stable_tree + nid); | |
ee0ea59c | 3125 | while (node) { |
21fbd591 | 3126 | stable_node = rb_entry(node, struct ksm_stable_node, node); |
2c653d0e AA |
3127 | if (stable_node_chain_remove_range(stable_node, |
3128 | start_pfn, end_pfn, | |
3129 | root_stable_tree + | |
3130 | nid)) | |
ef53d16c | 3131 | node = rb_first(root_stable_tree + nid); |
2c653d0e | 3132 | else |
ee0ea59c HD |
3133 | node = rb_next(node); |
3134 | cond_resched(); | |
90bd6fd3 | 3135 | } |
ee0ea59c | 3136 | } |
03640418 | 3137 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
3138 | if (stable_node->kpfn >= start_pfn && |
3139 | stable_node->kpfn < end_pfn) | |
3140 | remove_node_from_stable_tree(stable_node); | |
3141 | cond_resched(); | |
3142 | } | |
62b61f61 HD |
3143 | } |
3144 | ||
3145 | static int ksm_memory_callback(struct notifier_block *self, | |
3146 | unsigned long action, void *arg) | |
3147 | { | |
3148 | struct memory_notify *mn = arg; | |
62b61f61 HD |
3149 | |
3150 | switch (action) { | |
3151 | case MEM_GOING_OFFLINE: | |
3152 | /* | |
ef4d43a8 HD |
3153 | * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() |
3154 | * and remove_all_stable_nodes() while memory is going offline: | |
3155 | * it is unsafe for them to touch the stable tree at this time. | |
3156 | * But unmerge_ksm_pages(), rmap lookups and other entry points | |
3157 | * which do not need the ksm_thread_mutex are all safe. | |
62b61f61 | 3158 | */ |
ef4d43a8 HD |
3159 | mutex_lock(&ksm_thread_mutex); |
3160 | ksm_run |= KSM_RUN_OFFLINE; | |
3161 | mutex_unlock(&ksm_thread_mutex); | |
62b61f61 HD |
3162 | break; |
3163 | ||
3164 | case MEM_OFFLINE: | |
3165 | /* | |
3166 | * Most of the work is done by page migration; but there might | |
3167 | * be a few stable_nodes left over, still pointing to struct | |
ee0ea59c HD |
3168 | * pages which have been offlined: prune those from the tree, |
3169 | * otherwise get_ksm_page() might later try to access a | |
3170 | * non-existent struct page. | |
62b61f61 | 3171 | */ |
ee0ea59c HD |
3172 | ksm_check_stable_tree(mn->start_pfn, |
3173 | mn->start_pfn + mn->nr_pages); | |
e4a9bc58 | 3174 | fallthrough; |
62b61f61 | 3175 | case MEM_CANCEL_OFFLINE: |
ef4d43a8 HD |
3176 | mutex_lock(&ksm_thread_mutex); |
3177 | ksm_run &= ~KSM_RUN_OFFLINE; | |
62b61f61 | 3178 | mutex_unlock(&ksm_thread_mutex); |
ef4d43a8 HD |
3179 | |
3180 | smp_mb(); /* wake_up_bit advises this */ | |
3181 | wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); | |
62b61f61 HD |
3182 | break; |
3183 | } | |
3184 | return NOTIFY_OK; | |
3185 | } | |
ef4d43a8 HD |
3186 | #else |
3187 | static void wait_while_offlining(void) | |
3188 | { | |
3189 | } | |
62b61f61 HD |
3190 | #endif /* CONFIG_MEMORY_HOTREMOVE */ |
3191 | ||
d21077fb SR |
3192 | #ifdef CONFIG_PROC_FS |
3193 | long ksm_process_profit(struct mm_struct *mm) | |
3194 | { | |
1a8e8430 | 3195 | return (long)(mm->ksm_merging_pages + mm->ksm_zero_pages) * PAGE_SIZE - |
d21077fb SR |
3196 | mm->ksm_rmap_items * sizeof(struct ksm_rmap_item); |
3197 | } | |
3198 | #endif /* CONFIG_PROC_FS */ | |
3199 | ||
2ffd8679 HD |
3200 | #ifdef CONFIG_SYSFS |
3201 | /* | |
3202 | * This all compiles without CONFIG_SYSFS, but is a waste of space. | |
3203 | */ | |
3204 | ||
31dbd01f IE |
3205 | #define KSM_ATTR_RO(_name) \ |
3206 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
3207 | #define KSM_ATTR(_name) \ | |
1bad2e5c | 3208 | static struct kobj_attribute _name##_attr = __ATTR_RW(_name) |
31dbd01f IE |
3209 | |
3210 | static ssize_t sleep_millisecs_show(struct kobject *kobj, | |
3211 | struct kobj_attribute *attr, char *buf) | |
3212 | { | |
ae7a927d | 3213 | return sysfs_emit(buf, "%u\n", ksm_thread_sleep_millisecs); |
31dbd01f IE |
3214 | } |
3215 | ||
3216 | static ssize_t sleep_millisecs_store(struct kobject *kobj, | |
3217 | struct kobj_attribute *attr, | |
3218 | const char *buf, size_t count) | |
3219 | { | |
dfefd226 | 3220 | unsigned int msecs; |
31dbd01f IE |
3221 | int err; |
3222 | ||
dfefd226 AD |
3223 | err = kstrtouint(buf, 10, &msecs); |
3224 | if (err) | |
31dbd01f IE |
3225 | return -EINVAL; |
3226 | ||
3227 | ksm_thread_sleep_millisecs = msecs; | |
fcf9a0ef | 3228 | wake_up_interruptible(&ksm_iter_wait); |
31dbd01f IE |
3229 | |
3230 | return count; | |
3231 | } | |
3232 | KSM_ATTR(sleep_millisecs); | |
3233 | ||
3234 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
3235 | struct kobj_attribute *attr, char *buf) | |
3236 | { | |
ae7a927d | 3237 | return sysfs_emit(buf, "%u\n", ksm_thread_pages_to_scan); |
31dbd01f IE |
3238 | } |
3239 | ||
3240 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
3241 | struct kobj_attribute *attr, | |
3242 | const char *buf, size_t count) | |
3243 | { | |
dfefd226 | 3244 | unsigned int nr_pages; |
31dbd01f | 3245 | int err; |
31dbd01f | 3246 | |
dfefd226 AD |
3247 | err = kstrtouint(buf, 10, &nr_pages); |
3248 | if (err) | |
31dbd01f IE |
3249 | return -EINVAL; |
3250 | ||
3251 | ksm_thread_pages_to_scan = nr_pages; | |
3252 | ||
3253 | return count; | |
3254 | } | |
3255 | KSM_ATTR(pages_to_scan); | |
3256 | ||
3257 | static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, | |
3258 | char *buf) | |
3259 | { | |
ae7a927d | 3260 | return sysfs_emit(buf, "%lu\n", ksm_run); |
31dbd01f IE |
3261 | } |
3262 | ||
3263 | static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, | |
3264 | const char *buf, size_t count) | |
3265 | { | |
dfefd226 | 3266 | unsigned int flags; |
31dbd01f | 3267 | int err; |
31dbd01f | 3268 | |
dfefd226 AD |
3269 | err = kstrtouint(buf, 10, &flags); |
3270 | if (err) | |
31dbd01f IE |
3271 | return -EINVAL; |
3272 | if (flags > KSM_RUN_UNMERGE) | |
3273 | return -EINVAL; | |
3274 | ||
3275 | /* | |
3276 | * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. | |
3277 | * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, | |
d0f209f6 HD |
3278 | * breaking COW to free the pages_shared (but leaves mm_slots |
3279 | * on the list for when ksmd may be set running again). | |
31dbd01f IE |
3280 | */ |
3281 | ||
3282 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 3283 | wait_while_offlining(); |
31dbd01f IE |
3284 | if (ksm_run != flags) { |
3285 | ksm_run = flags; | |
d952b791 | 3286 | if (flags & KSM_RUN_UNMERGE) { |
e1e12d2f | 3287 | set_current_oom_origin(); |
d952b791 | 3288 | err = unmerge_and_remove_all_rmap_items(); |
e1e12d2f | 3289 | clear_current_oom_origin(); |
d952b791 HD |
3290 | if (err) { |
3291 | ksm_run = KSM_RUN_STOP; | |
3292 | count = err; | |
3293 | } | |
3294 | } | |
31dbd01f IE |
3295 | } |
3296 | mutex_unlock(&ksm_thread_mutex); | |
3297 | ||
3298 | if (flags & KSM_RUN_MERGE) | |
3299 | wake_up_interruptible(&ksm_thread_wait); | |
3300 | ||
3301 | return count; | |
3302 | } | |
3303 | KSM_ATTR(run); | |
3304 | ||
90bd6fd3 PH |
3305 | #ifdef CONFIG_NUMA |
3306 | static ssize_t merge_across_nodes_show(struct kobject *kobj, | |
ae7a927d | 3307 | struct kobj_attribute *attr, char *buf) |
90bd6fd3 | 3308 | { |
ae7a927d | 3309 | return sysfs_emit(buf, "%u\n", ksm_merge_across_nodes); |
90bd6fd3 PH |
3310 | } |
3311 | ||
3312 | static ssize_t merge_across_nodes_store(struct kobject *kobj, | |
3313 | struct kobj_attribute *attr, | |
3314 | const char *buf, size_t count) | |
3315 | { | |
3316 | int err; | |
3317 | unsigned long knob; | |
3318 | ||
3319 | err = kstrtoul(buf, 10, &knob); | |
3320 | if (err) | |
3321 | return err; | |
3322 | if (knob > 1) | |
3323 | return -EINVAL; | |
3324 | ||
3325 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 3326 | wait_while_offlining(); |
90bd6fd3 | 3327 | if (ksm_merge_across_nodes != knob) { |
cbf86cfe | 3328 | if (ksm_pages_shared || remove_all_stable_nodes()) |
90bd6fd3 | 3329 | err = -EBUSY; |
ef53d16c HD |
3330 | else if (root_stable_tree == one_stable_tree) { |
3331 | struct rb_root *buf; | |
3332 | /* | |
3333 | * This is the first time that we switch away from the | |
3334 | * default of merging across nodes: must now allocate | |
3335 | * a buffer to hold as many roots as may be needed. | |
3336 | * Allocate stable and unstable together: | |
3337 | * MAXSMP NODES_SHIFT 10 will use 16kB. | |
3338 | */ | |
bafe1e14 JP |
3339 | buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), |
3340 | GFP_KERNEL); | |
ef53d16c HD |
3341 | /* Let us assume that RB_ROOT is NULL is zero */ |
3342 | if (!buf) | |
3343 | err = -ENOMEM; | |
3344 | else { | |
3345 | root_stable_tree = buf; | |
3346 | root_unstable_tree = buf + nr_node_ids; | |
3347 | /* Stable tree is empty but not the unstable */ | |
3348 | root_unstable_tree[0] = one_unstable_tree[0]; | |
3349 | } | |
3350 | } | |
3351 | if (!err) { | |
90bd6fd3 | 3352 | ksm_merge_across_nodes = knob; |
ef53d16c HD |
3353 | ksm_nr_node_ids = knob ? 1 : nr_node_ids; |
3354 | } | |
90bd6fd3 PH |
3355 | } |
3356 | mutex_unlock(&ksm_thread_mutex); | |
3357 | ||
3358 | return err ? err : count; | |
3359 | } | |
3360 | KSM_ATTR(merge_across_nodes); | |
3361 | #endif | |
3362 | ||
e86c59b1 | 3363 | static ssize_t use_zero_pages_show(struct kobject *kobj, |
ae7a927d | 3364 | struct kobj_attribute *attr, char *buf) |
e86c59b1 | 3365 | { |
ae7a927d | 3366 | return sysfs_emit(buf, "%u\n", ksm_use_zero_pages); |
e86c59b1 CI |
3367 | } |
3368 | static ssize_t use_zero_pages_store(struct kobject *kobj, | |
3369 | struct kobj_attribute *attr, | |
3370 | const char *buf, size_t count) | |
3371 | { | |
3372 | int err; | |
3373 | bool value; | |
3374 | ||
3375 | err = kstrtobool(buf, &value); | |
3376 | if (err) | |
3377 | return -EINVAL; | |
3378 | ||
3379 | ksm_use_zero_pages = value; | |
3380 | ||
3381 | return count; | |
3382 | } | |
3383 | KSM_ATTR(use_zero_pages); | |
3384 | ||
2c653d0e AA |
3385 | static ssize_t max_page_sharing_show(struct kobject *kobj, |
3386 | struct kobj_attribute *attr, char *buf) | |
3387 | { | |
ae7a927d | 3388 | return sysfs_emit(buf, "%u\n", ksm_max_page_sharing); |
2c653d0e AA |
3389 | } |
3390 | ||
3391 | static ssize_t max_page_sharing_store(struct kobject *kobj, | |
3392 | struct kobj_attribute *attr, | |
3393 | const char *buf, size_t count) | |
3394 | { | |
3395 | int err; | |
3396 | int knob; | |
3397 | ||
3398 | err = kstrtoint(buf, 10, &knob); | |
3399 | if (err) | |
3400 | return err; | |
3401 | /* | |
3402 | * When a KSM page is created it is shared by 2 mappings. This | |
3403 | * being a signed comparison, it implicitly verifies it's not | |
3404 | * negative. | |
3405 | */ | |
3406 | if (knob < 2) | |
3407 | return -EINVAL; | |
3408 | ||
3409 | if (READ_ONCE(ksm_max_page_sharing) == knob) | |
3410 | return count; | |
3411 | ||
3412 | mutex_lock(&ksm_thread_mutex); | |
3413 | wait_while_offlining(); | |
3414 | if (ksm_max_page_sharing != knob) { | |
3415 | if (ksm_pages_shared || remove_all_stable_nodes()) | |
3416 | err = -EBUSY; | |
3417 | else | |
3418 | ksm_max_page_sharing = knob; | |
3419 | } | |
3420 | mutex_unlock(&ksm_thread_mutex); | |
3421 | ||
3422 | return err ? err : count; | |
3423 | } | |
3424 | KSM_ATTR(max_page_sharing); | |
3425 | ||
b348b5fe SR |
3426 | static ssize_t pages_scanned_show(struct kobject *kobj, |
3427 | struct kobj_attribute *attr, char *buf) | |
3428 | { | |
3429 | return sysfs_emit(buf, "%lu\n", ksm_pages_scanned); | |
3430 | } | |
3431 | KSM_ATTR_RO(pages_scanned); | |
3432 | ||
b4028260 HD |
3433 | static ssize_t pages_shared_show(struct kobject *kobj, |
3434 | struct kobj_attribute *attr, char *buf) | |
3435 | { | |
ae7a927d | 3436 | return sysfs_emit(buf, "%lu\n", ksm_pages_shared); |
b4028260 HD |
3437 | } |
3438 | KSM_ATTR_RO(pages_shared); | |
3439 | ||
3440 | static ssize_t pages_sharing_show(struct kobject *kobj, | |
3441 | struct kobj_attribute *attr, char *buf) | |
3442 | { | |
ae7a927d | 3443 | return sysfs_emit(buf, "%lu\n", ksm_pages_sharing); |
b4028260 HD |
3444 | } |
3445 | KSM_ATTR_RO(pages_sharing); | |
3446 | ||
473b0ce4 HD |
3447 | static ssize_t pages_unshared_show(struct kobject *kobj, |
3448 | struct kobj_attribute *attr, char *buf) | |
3449 | { | |
ae7a927d | 3450 | return sysfs_emit(buf, "%lu\n", ksm_pages_unshared); |
473b0ce4 HD |
3451 | } |
3452 | KSM_ATTR_RO(pages_unshared); | |
3453 | ||
3454 | static ssize_t pages_volatile_show(struct kobject *kobj, | |
3455 | struct kobj_attribute *attr, char *buf) | |
3456 | { | |
3457 | long ksm_pages_volatile; | |
3458 | ||
3459 | ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared | |
3460 | - ksm_pages_sharing - ksm_pages_unshared; | |
3461 | /* | |
3462 | * It was not worth any locking to calculate that statistic, | |
3463 | * but it might therefore sometimes be negative: conceal that. | |
3464 | */ | |
3465 | if (ksm_pages_volatile < 0) | |
3466 | ksm_pages_volatile = 0; | |
ae7a927d | 3467 | return sysfs_emit(buf, "%ld\n", ksm_pages_volatile); |
473b0ce4 HD |
3468 | } |
3469 | KSM_ATTR_RO(pages_volatile); | |
3470 | ||
e5a68991 SR |
3471 | static ssize_t pages_skipped_show(struct kobject *kobj, |
3472 | struct kobj_attribute *attr, char *buf) | |
3473 | { | |
3474 | return sysfs_emit(buf, "%lu\n", ksm_pages_skipped); | |
3475 | } | |
3476 | KSM_ATTR_RO(pages_skipped); | |
3477 | ||
e2942062 | 3478 | static ssize_t ksm_zero_pages_show(struct kobject *kobj, |
3479 | struct kobj_attribute *attr, char *buf) | |
3480 | { | |
3481 | return sysfs_emit(buf, "%ld\n", ksm_zero_pages); | |
3482 | } | |
3483 | KSM_ATTR_RO(ksm_zero_pages); | |
3484 | ||
d21077fb SR |
3485 | static ssize_t general_profit_show(struct kobject *kobj, |
3486 | struct kobj_attribute *attr, char *buf) | |
3487 | { | |
3488 | long general_profit; | |
3489 | ||
1a8e8430 | 3490 | general_profit = (ksm_pages_sharing + ksm_zero_pages) * PAGE_SIZE - |
d21077fb SR |
3491 | ksm_rmap_items * sizeof(struct ksm_rmap_item); |
3492 | ||
3493 | return sysfs_emit(buf, "%ld\n", general_profit); | |
3494 | } | |
3495 | KSM_ATTR_RO(general_profit); | |
3496 | ||
2c653d0e AA |
3497 | static ssize_t stable_node_dups_show(struct kobject *kobj, |
3498 | struct kobj_attribute *attr, char *buf) | |
3499 | { | |
ae7a927d | 3500 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_dups); |
2c653d0e AA |
3501 | } |
3502 | KSM_ATTR_RO(stable_node_dups); | |
3503 | ||
3504 | static ssize_t stable_node_chains_show(struct kobject *kobj, | |
3505 | struct kobj_attribute *attr, char *buf) | |
3506 | { | |
ae7a927d | 3507 | return sysfs_emit(buf, "%lu\n", ksm_stable_node_chains); |
2c653d0e AA |
3508 | } |
3509 | KSM_ATTR_RO(stable_node_chains); | |
3510 | ||
3511 | static ssize_t | |
3512 | stable_node_chains_prune_millisecs_show(struct kobject *kobj, | |
3513 | struct kobj_attribute *attr, | |
3514 | char *buf) | |
3515 | { | |
ae7a927d | 3516 | return sysfs_emit(buf, "%u\n", ksm_stable_node_chains_prune_millisecs); |
2c653d0e AA |
3517 | } |
3518 | ||
3519 | static ssize_t | |
3520 | stable_node_chains_prune_millisecs_store(struct kobject *kobj, | |
3521 | struct kobj_attribute *attr, | |
3522 | const char *buf, size_t count) | |
3523 | { | |
584ff0df | 3524 | unsigned int msecs; |
2c653d0e AA |
3525 | int err; |
3526 | ||
584ff0df ZB |
3527 | err = kstrtouint(buf, 10, &msecs); |
3528 | if (err) | |
2c653d0e AA |
3529 | return -EINVAL; |
3530 | ||
3531 | ksm_stable_node_chains_prune_millisecs = msecs; | |
3532 | ||
3533 | return count; | |
3534 | } | |
3535 | KSM_ATTR(stable_node_chains_prune_millisecs); | |
3536 | ||
473b0ce4 HD |
3537 | static ssize_t full_scans_show(struct kobject *kobj, |
3538 | struct kobj_attribute *attr, char *buf) | |
3539 | { | |
ae7a927d | 3540 | return sysfs_emit(buf, "%lu\n", ksm_scan.seqnr); |
473b0ce4 HD |
3541 | } |
3542 | KSM_ATTR_RO(full_scans); | |
3543 | ||
5e924ff5 SR |
3544 | static ssize_t smart_scan_show(struct kobject *kobj, |
3545 | struct kobj_attribute *attr, char *buf) | |
3546 | { | |
3547 | return sysfs_emit(buf, "%u\n", ksm_smart_scan); | |
3548 | } | |
3549 | ||
3550 | static ssize_t smart_scan_store(struct kobject *kobj, | |
3551 | struct kobj_attribute *attr, | |
3552 | const char *buf, size_t count) | |
3553 | { | |
3554 | int err; | |
3555 | bool value; | |
3556 | ||
3557 | err = kstrtobool(buf, &value); | |
3558 | if (err) | |
3559 | return -EINVAL; | |
3560 | ||
3561 | ksm_smart_scan = value; | |
3562 | return count; | |
3563 | } | |
3564 | KSM_ATTR(smart_scan); | |
3565 | ||
31dbd01f IE |
3566 | static struct attribute *ksm_attrs[] = { |
3567 | &sleep_millisecs_attr.attr, | |
3568 | &pages_to_scan_attr.attr, | |
3569 | &run_attr.attr, | |
b348b5fe | 3570 | &pages_scanned_attr.attr, |
b4028260 HD |
3571 | &pages_shared_attr.attr, |
3572 | &pages_sharing_attr.attr, | |
473b0ce4 HD |
3573 | &pages_unshared_attr.attr, |
3574 | &pages_volatile_attr.attr, | |
e5a68991 | 3575 | &pages_skipped_attr.attr, |
e2942062 | 3576 | &ksm_zero_pages_attr.attr, |
473b0ce4 | 3577 | &full_scans_attr.attr, |
90bd6fd3 PH |
3578 | #ifdef CONFIG_NUMA |
3579 | &merge_across_nodes_attr.attr, | |
3580 | #endif | |
2c653d0e AA |
3581 | &max_page_sharing_attr.attr, |
3582 | &stable_node_chains_attr.attr, | |
3583 | &stable_node_dups_attr.attr, | |
3584 | &stable_node_chains_prune_millisecs_attr.attr, | |
e86c59b1 | 3585 | &use_zero_pages_attr.attr, |
d21077fb | 3586 | &general_profit_attr.attr, |
5e924ff5 | 3587 | &smart_scan_attr.attr, |
31dbd01f IE |
3588 | NULL, |
3589 | }; | |
3590 | ||
f907c26a | 3591 | static const struct attribute_group ksm_attr_group = { |
31dbd01f IE |
3592 | .attrs = ksm_attrs, |
3593 | .name = "ksm", | |
3594 | }; | |
2ffd8679 | 3595 | #endif /* CONFIG_SYSFS */ |
31dbd01f IE |
3596 | |
3597 | static int __init ksm_init(void) | |
3598 | { | |
3599 | struct task_struct *ksm_thread; | |
3600 | int err; | |
3601 | ||
e86c59b1 CI |
3602 | /* The correct value depends on page size and endianness */ |
3603 | zero_checksum = calc_checksum(ZERO_PAGE(0)); | |
3604 | /* Default to false for backwards compatibility */ | |
3605 | ksm_use_zero_pages = false; | |
3606 | ||
31dbd01f IE |
3607 | err = ksm_slab_init(); |
3608 | if (err) | |
3609 | goto out; | |
3610 | ||
31dbd01f IE |
3611 | ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); |
3612 | if (IS_ERR(ksm_thread)) { | |
25acde31 | 3613 | pr_err("ksm: creating kthread failed\n"); |
31dbd01f | 3614 | err = PTR_ERR(ksm_thread); |
d9f8984c | 3615 | goto out_free; |
31dbd01f IE |
3616 | } |
3617 | ||
2ffd8679 | 3618 | #ifdef CONFIG_SYSFS |
31dbd01f IE |
3619 | err = sysfs_create_group(mm_kobj, &ksm_attr_group); |
3620 | if (err) { | |
25acde31 | 3621 | pr_err("ksm: register sysfs failed\n"); |
2ffd8679 | 3622 | kthread_stop(ksm_thread); |
d9f8984c | 3623 | goto out_free; |
31dbd01f | 3624 | } |
c73602ad HD |
3625 | #else |
3626 | ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ | |
3627 | ||
2ffd8679 | 3628 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 3629 | |
62b61f61 | 3630 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 | 3631 | /* There is no significance to this priority 100 */ |
1eeaa4fd | 3632 | hotplug_memory_notifier(ksm_memory_callback, KSM_CALLBACK_PRI); |
62b61f61 | 3633 | #endif |
31dbd01f IE |
3634 | return 0; |
3635 | ||
d9f8984c | 3636 | out_free: |
31dbd01f IE |
3637 | ksm_slab_free(); |
3638 | out: | |
3639 | return err; | |
f8af4da3 | 3640 | } |
a64fb3cd | 3641 | subsys_initcall(ksm_init); |