]>
Commit | Line | Data |
---|---|---|
f8af4da3 | 1 | /* |
31dbd01f IE |
2 | * Memory merging support. |
3 | * | |
4 | * This code enables dynamic sharing of identical pages found in different | |
5 | * memory areas, even if they are not shared by fork() | |
6 | * | |
36b2528d | 7 | * Copyright (C) 2008-2009 Red Hat, Inc. |
31dbd01f IE |
8 | * Authors: |
9 | * Izik Eidus | |
10 | * Andrea Arcangeli | |
11 | * Chris Wright | |
36b2528d | 12 | * Hugh Dickins |
31dbd01f IE |
13 | * |
14 | * This work is licensed under the terms of the GNU GPL, version 2. | |
f8af4da3 HD |
15 | */ |
16 | ||
17 | #include <linux/errno.h> | |
31dbd01f IE |
18 | #include <linux/mm.h> |
19 | #include <linux/fs.h> | |
f8af4da3 | 20 | #include <linux/mman.h> |
31dbd01f IE |
21 | #include <linux/sched.h> |
22 | #include <linux/rwsem.h> | |
23 | #include <linux/pagemap.h> | |
24 | #include <linux/rmap.h> | |
25 | #include <linux/spinlock.h> | |
26 | #include <linux/jhash.h> | |
27 | #include <linux/delay.h> | |
28 | #include <linux/kthread.h> | |
29 | #include <linux/wait.h> | |
30 | #include <linux/slab.h> | |
31 | #include <linux/rbtree.h> | |
62b61f61 | 32 | #include <linux/memory.h> |
31dbd01f | 33 | #include <linux/mmu_notifier.h> |
2c6854fd | 34 | #include <linux/swap.h> |
f8af4da3 | 35 | #include <linux/ksm.h> |
4ca3a69b | 36 | #include <linux/hashtable.h> |
878aee7d | 37 | #include <linux/freezer.h> |
72788c38 | 38 | #include <linux/oom.h> |
90bd6fd3 | 39 | #include <linux/numa.h> |
f8af4da3 | 40 | |
31dbd01f | 41 | #include <asm/tlbflush.h> |
73848b46 | 42 | #include "internal.h" |
31dbd01f | 43 | |
e850dcf5 HD |
44 | #ifdef CONFIG_NUMA |
45 | #define NUMA(x) (x) | |
46 | #define DO_NUMA(x) do { (x); } while (0) | |
47 | #else | |
48 | #define NUMA(x) (0) | |
49 | #define DO_NUMA(x) do { } while (0) | |
50 | #endif | |
51 | ||
31dbd01f IE |
52 | /* |
53 | * A few notes about the KSM scanning process, | |
54 | * to make it easier to understand the data structures below: | |
55 | * | |
56 | * In order to reduce excessive scanning, KSM sorts the memory pages by their | |
57 | * contents into a data structure that holds pointers to the pages' locations. | |
58 | * | |
59 | * Since the contents of the pages may change at any moment, KSM cannot just | |
60 | * insert the pages into a normal sorted tree and expect it to find anything. | |
61 | * Therefore KSM uses two data structures - the stable and the unstable tree. | |
62 | * | |
63 | * The stable tree holds pointers to all the merged pages (ksm pages), sorted | |
64 | * by their contents. Because each such page is write-protected, searching on | |
65 | * this tree is fully assured to be working (except when pages are unmapped), | |
66 | * and therefore this tree is called the stable tree. | |
67 | * | |
68 | * In addition to the stable tree, KSM uses a second data structure called the | |
69 | * unstable tree: this tree holds pointers to pages which have been found to | |
70 | * be "unchanged for a period of time". The unstable tree sorts these pages | |
71 | * by their contents, but since they are not write-protected, KSM cannot rely | |
72 | * upon the unstable tree to work correctly - the unstable tree is liable to | |
73 | * be corrupted as its contents are modified, and so it is called unstable. | |
74 | * | |
75 | * KSM solves this problem by several techniques: | |
76 | * | |
77 | * 1) The unstable tree is flushed every time KSM completes scanning all | |
78 | * memory areas, and then the tree is rebuilt again from the beginning. | |
79 | * 2) KSM will only insert into the unstable tree, pages whose hash value | |
80 | * has not changed since the previous scan of all memory areas. | |
81 | * 3) The unstable tree is a RedBlack Tree - so its balancing is based on the | |
82 | * colors of the nodes and not on their contents, assuring that even when | |
83 | * the tree gets "corrupted" it won't get out of balance, so scanning time | |
84 | * remains the same (also, searching and inserting nodes in an rbtree uses | |
85 | * the same algorithm, so we have no overhead when we flush and rebuild). | |
86 | * 4) KSM never flushes the stable tree, which means that even if it were to | |
87 | * take 10 attempts to find a page in the unstable tree, once it is found, | |
88 | * it is secured in the stable tree. (When we scan a new page, we first | |
89 | * compare it against the stable tree, and then against the unstable tree.) | |
8fdb3dbf HD |
90 | * |
91 | * If the merge_across_nodes tunable is unset, then KSM maintains multiple | |
92 | * stable trees and multiple unstable trees: one of each for each NUMA node. | |
31dbd01f IE |
93 | */ |
94 | ||
95 | /** | |
96 | * struct mm_slot - ksm information per mm that is being scanned | |
97 | * @link: link to the mm_slots hash list | |
98 | * @mm_list: link into the mm_slots list, rooted in ksm_mm_head | |
6514d511 | 99 | * @rmap_list: head for this mm_slot's singly-linked list of rmap_items |
31dbd01f IE |
100 | * @mm: the mm that this information is valid for |
101 | */ | |
102 | struct mm_slot { | |
103 | struct hlist_node link; | |
104 | struct list_head mm_list; | |
6514d511 | 105 | struct rmap_item *rmap_list; |
31dbd01f IE |
106 | struct mm_struct *mm; |
107 | }; | |
108 | ||
109 | /** | |
110 | * struct ksm_scan - cursor for scanning | |
111 | * @mm_slot: the current mm_slot we are scanning | |
112 | * @address: the next address inside that to be scanned | |
6514d511 | 113 | * @rmap_list: link to the next rmap to be scanned in the rmap_list |
31dbd01f IE |
114 | * @seqnr: count of completed full scans (needed when removing unstable node) |
115 | * | |
116 | * There is only the one ksm_scan instance of this cursor structure. | |
117 | */ | |
118 | struct ksm_scan { | |
119 | struct mm_slot *mm_slot; | |
120 | unsigned long address; | |
6514d511 | 121 | struct rmap_item **rmap_list; |
31dbd01f IE |
122 | unsigned long seqnr; |
123 | }; | |
124 | ||
7b6ba2c7 HD |
125 | /** |
126 | * struct stable_node - node of the stable rbtree | |
127 | * @node: rb node of this ksm page in the stable tree | |
4146d2d6 HD |
128 | * @head: (overlaying parent) &migrate_nodes indicates temporarily on that list |
129 | * @list: linked into migrate_nodes, pending placement in the proper node tree | |
7b6ba2c7 | 130 | * @hlist: hlist head of rmap_items using this ksm page |
4146d2d6 HD |
131 | * @kpfn: page frame number of this ksm page (perhaps temporarily on wrong nid) |
132 | * @nid: NUMA node id of stable tree in which linked (may not match kpfn) | |
7b6ba2c7 HD |
133 | */ |
134 | struct stable_node { | |
4146d2d6 HD |
135 | union { |
136 | struct rb_node node; /* when node of stable tree */ | |
137 | struct { /* when listed for migration */ | |
138 | struct list_head *head; | |
139 | struct list_head list; | |
140 | }; | |
141 | }; | |
7b6ba2c7 | 142 | struct hlist_head hlist; |
62b61f61 | 143 | unsigned long kpfn; |
4146d2d6 HD |
144 | #ifdef CONFIG_NUMA |
145 | int nid; | |
146 | #endif | |
7b6ba2c7 HD |
147 | }; |
148 | ||
31dbd01f IE |
149 | /** |
150 | * struct rmap_item - reverse mapping item for virtual addresses | |
6514d511 | 151 | * @rmap_list: next rmap_item in mm_slot's singly-linked rmap_list |
db114b83 | 152 | * @anon_vma: pointer to anon_vma for this mm,address, when in stable tree |
bc56620b | 153 | * @nid: NUMA node id of unstable tree in which linked (may not match page) |
31dbd01f IE |
154 | * @mm: the memory structure this rmap_item is pointing into |
155 | * @address: the virtual address this rmap_item tracks (+ flags in low bits) | |
156 | * @oldchecksum: previous checksum of the page at that virtual address | |
7b6ba2c7 HD |
157 | * @node: rb node of this rmap_item in the unstable tree |
158 | * @head: pointer to stable_node heading this list in the stable tree | |
159 | * @hlist: link into hlist of rmap_items hanging off that stable_node | |
31dbd01f IE |
160 | */ |
161 | struct rmap_item { | |
6514d511 | 162 | struct rmap_item *rmap_list; |
bc56620b HD |
163 | union { |
164 | struct anon_vma *anon_vma; /* when stable */ | |
165 | #ifdef CONFIG_NUMA | |
166 | int nid; /* when node of unstable tree */ | |
167 | #endif | |
168 | }; | |
31dbd01f IE |
169 | struct mm_struct *mm; |
170 | unsigned long address; /* + low bits used for flags below */ | |
7b6ba2c7 | 171 | unsigned int oldchecksum; /* when unstable */ |
31dbd01f | 172 | union { |
7b6ba2c7 HD |
173 | struct rb_node node; /* when node of unstable tree */ |
174 | struct { /* when listed from stable tree */ | |
175 | struct stable_node *head; | |
176 | struct hlist_node hlist; | |
177 | }; | |
31dbd01f IE |
178 | }; |
179 | }; | |
180 | ||
181 | #define SEQNR_MASK 0x0ff /* low bits of unstable tree seqnr */ | |
7b6ba2c7 HD |
182 | #define UNSTABLE_FLAG 0x100 /* is a node of the unstable tree */ |
183 | #define STABLE_FLAG 0x200 /* is listed from the stable tree */ | |
31dbd01f IE |
184 | |
185 | /* The stable and unstable tree heads */ | |
ef53d16c HD |
186 | static struct rb_root one_stable_tree[1] = { RB_ROOT }; |
187 | static struct rb_root one_unstable_tree[1] = { RB_ROOT }; | |
188 | static struct rb_root *root_stable_tree = one_stable_tree; | |
189 | static struct rb_root *root_unstable_tree = one_unstable_tree; | |
31dbd01f | 190 | |
4146d2d6 HD |
191 | /* Recently migrated nodes of stable tree, pending proper placement */ |
192 | static LIST_HEAD(migrate_nodes); | |
193 | ||
4ca3a69b SL |
194 | #define MM_SLOTS_HASH_BITS 10 |
195 | static DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
31dbd01f IE |
196 | |
197 | static struct mm_slot ksm_mm_head = { | |
198 | .mm_list = LIST_HEAD_INIT(ksm_mm_head.mm_list), | |
199 | }; | |
200 | static struct ksm_scan ksm_scan = { | |
201 | .mm_slot = &ksm_mm_head, | |
202 | }; | |
203 | ||
204 | static struct kmem_cache *rmap_item_cache; | |
7b6ba2c7 | 205 | static struct kmem_cache *stable_node_cache; |
31dbd01f IE |
206 | static struct kmem_cache *mm_slot_cache; |
207 | ||
208 | /* The number of nodes in the stable tree */ | |
b4028260 | 209 | static unsigned long ksm_pages_shared; |
31dbd01f | 210 | |
e178dfde | 211 | /* The number of page slots additionally sharing those nodes */ |
b4028260 | 212 | static unsigned long ksm_pages_sharing; |
31dbd01f | 213 | |
473b0ce4 HD |
214 | /* The number of nodes in the unstable tree */ |
215 | static unsigned long ksm_pages_unshared; | |
216 | ||
217 | /* The number of rmap_items in use: to calculate pages_volatile */ | |
218 | static unsigned long ksm_rmap_items; | |
219 | ||
31dbd01f | 220 | /* Number of pages ksmd should scan in one batch */ |
2c6854fd | 221 | static unsigned int ksm_thread_pages_to_scan = 100; |
31dbd01f IE |
222 | |
223 | /* Milliseconds ksmd should sleep between batches */ | |
2ffd8679 | 224 | static unsigned int ksm_thread_sleep_millisecs = 20; |
31dbd01f | 225 | |
e86c59b1 CI |
226 | /* Checksum of an empty (zeroed) page */ |
227 | static unsigned int zero_checksum __read_mostly; | |
228 | ||
229 | /* Whether to merge empty (zeroed) pages with actual zero pages */ | |
230 | static bool ksm_use_zero_pages __read_mostly; | |
231 | ||
e850dcf5 | 232 | #ifdef CONFIG_NUMA |
90bd6fd3 PH |
233 | /* Zeroed when merging across nodes is not allowed */ |
234 | static unsigned int ksm_merge_across_nodes = 1; | |
ef53d16c | 235 | static int ksm_nr_node_ids = 1; |
e850dcf5 HD |
236 | #else |
237 | #define ksm_merge_across_nodes 1U | |
ef53d16c | 238 | #define ksm_nr_node_ids 1 |
e850dcf5 | 239 | #endif |
90bd6fd3 | 240 | |
31dbd01f IE |
241 | #define KSM_RUN_STOP 0 |
242 | #define KSM_RUN_MERGE 1 | |
243 | #define KSM_RUN_UNMERGE 2 | |
ef4d43a8 HD |
244 | #define KSM_RUN_OFFLINE 4 |
245 | static unsigned long ksm_run = KSM_RUN_STOP; | |
246 | static void wait_while_offlining(void); | |
31dbd01f IE |
247 | |
248 | static DECLARE_WAIT_QUEUE_HEAD(ksm_thread_wait); | |
249 | static DEFINE_MUTEX(ksm_thread_mutex); | |
250 | static DEFINE_SPINLOCK(ksm_mmlist_lock); | |
251 | ||
252 | #define KSM_KMEM_CACHE(__struct, __flags) kmem_cache_create("ksm_"#__struct,\ | |
253 | sizeof(struct __struct), __alignof__(struct __struct),\ | |
254 | (__flags), NULL) | |
255 | ||
256 | static int __init ksm_slab_init(void) | |
257 | { | |
258 | rmap_item_cache = KSM_KMEM_CACHE(rmap_item, 0); | |
259 | if (!rmap_item_cache) | |
260 | goto out; | |
261 | ||
7b6ba2c7 HD |
262 | stable_node_cache = KSM_KMEM_CACHE(stable_node, 0); |
263 | if (!stable_node_cache) | |
264 | goto out_free1; | |
265 | ||
31dbd01f IE |
266 | mm_slot_cache = KSM_KMEM_CACHE(mm_slot, 0); |
267 | if (!mm_slot_cache) | |
7b6ba2c7 | 268 | goto out_free2; |
31dbd01f IE |
269 | |
270 | return 0; | |
271 | ||
7b6ba2c7 HD |
272 | out_free2: |
273 | kmem_cache_destroy(stable_node_cache); | |
274 | out_free1: | |
31dbd01f IE |
275 | kmem_cache_destroy(rmap_item_cache); |
276 | out: | |
277 | return -ENOMEM; | |
278 | } | |
279 | ||
280 | static void __init ksm_slab_free(void) | |
281 | { | |
282 | kmem_cache_destroy(mm_slot_cache); | |
7b6ba2c7 | 283 | kmem_cache_destroy(stable_node_cache); |
31dbd01f IE |
284 | kmem_cache_destroy(rmap_item_cache); |
285 | mm_slot_cache = NULL; | |
286 | } | |
287 | ||
288 | static inline struct rmap_item *alloc_rmap_item(void) | |
289 | { | |
473b0ce4 HD |
290 | struct rmap_item *rmap_item; |
291 | ||
5b398e41 | 292 | rmap_item = kmem_cache_zalloc(rmap_item_cache, GFP_KERNEL | |
293 | __GFP_NORETRY | __GFP_NOWARN); | |
473b0ce4 HD |
294 | if (rmap_item) |
295 | ksm_rmap_items++; | |
296 | return rmap_item; | |
31dbd01f IE |
297 | } |
298 | ||
299 | static inline void free_rmap_item(struct rmap_item *rmap_item) | |
300 | { | |
473b0ce4 | 301 | ksm_rmap_items--; |
31dbd01f IE |
302 | rmap_item->mm = NULL; /* debug safety */ |
303 | kmem_cache_free(rmap_item_cache, rmap_item); | |
304 | } | |
305 | ||
7b6ba2c7 HD |
306 | static inline struct stable_node *alloc_stable_node(void) |
307 | { | |
6213055f | 308 | /* |
309 | * The allocation can take too long with GFP_KERNEL when memory is under | |
310 | * pressure, which may lead to hung task warnings. Adding __GFP_HIGH | |
311 | * grants access to memory reserves, helping to avoid this problem. | |
312 | */ | |
313 | return kmem_cache_alloc(stable_node_cache, GFP_KERNEL | __GFP_HIGH); | |
7b6ba2c7 HD |
314 | } |
315 | ||
316 | static inline void free_stable_node(struct stable_node *stable_node) | |
317 | { | |
318 | kmem_cache_free(stable_node_cache, stable_node); | |
319 | } | |
320 | ||
31dbd01f IE |
321 | static inline struct mm_slot *alloc_mm_slot(void) |
322 | { | |
323 | if (!mm_slot_cache) /* initialization failed */ | |
324 | return NULL; | |
325 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
326 | } | |
327 | ||
328 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
329 | { | |
330 | kmem_cache_free(mm_slot_cache, mm_slot); | |
331 | } | |
332 | ||
31dbd01f IE |
333 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) |
334 | { | |
4ca3a69b SL |
335 | struct mm_slot *slot; |
336 | ||
b67bfe0d | 337 | hash_for_each_possible(mm_slots_hash, slot, link, (unsigned long)mm) |
4ca3a69b SL |
338 | if (slot->mm == mm) |
339 | return slot; | |
31dbd01f | 340 | |
31dbd01f IE |
341 | return NULL; |
342 | } | |
343 | ||
344 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
345 | struct mm_slot *mm_slot) | |
346 | { | |
31dbd01f | 347 | mm_slot->mm = mm; |
4ca3a69b | 348 | hash_add(mm_slots_hash, &mm_slot->link, (unsigned long)mm); |
31dbd01f IE |
349 | } |
350 | ||
a913e182 HD |
351 | /* |
352 | * ksmd, and unmerge_and_remove_all_rmap_items(), must not touch an mm's | |
353 | * page tables after it has passed through ksm_exit() - which, if necessary, | |
354 | * takes mmap_sem briefly to serialize against them. ksm_exit() does not set | |
355 | * a special flag: they can just back out as soon as mm_users goes to zero. | |
356 | * ksm_test_exit() is used throughout to make this test for exit: in some | |
357 | * places for correctness, in some places just to avoid unnecessary work. | |
358 | */ | |
359 | static inline bool ksm_test_exit(struct mm_struct *mm) | |
360 | { | |
361 | return atomic_read(&mm->mm_users) == 0; | |
362 | } | |
363 | ||
31dbd01f IE |
364 | /* |
365 | * We use break_ksm to break COW on a ksm page: it's a stripped down | |
366 | * | |
d4edcf0d | 367 | * if (get_user_pages(addr, 1, 1, 1, &page, NULL) == 1) |
31dbd01f IE |
368 | * put_page(page); |
369 | * | |
370 | * but taking great care only to touch a ksm page, in a VM_MERGEABLE vma, | |
371 | * in case the application has unmapped and remapped mm,addr meanwhile. | |
372 | * Could a ksm page appear anywhere else? Actually yes, in a VM_PFNMAP | |
373 | * mmap of /dev/mem or /dev/kmem, where we would not want to touch it. | |
1b2ee126 DH |
374 | * |
375 | * FAULT_FLAG/FOLL_REMOTE are because we do this outside the context | |
376 | * of the process that owns 'vma'. We also do not want to enforce | |
377 | * protection keys here anyway. | |
31dbd01f | 378 | */ |
d952b791 | 379 | static int break_ksm(struct vm_area_struct *vma, unsigned long addr) |
31dbd01f IE |
380 | { |
381 | struct page *page; | |
d952b791 | 382 | int ret = 0; |
31dbd01f IE |
383 | |
384 | do { | |
385 | cond_resched(); | |
1b2ee126 DH |
386 | page = follow_page(vma, addr, |
387 | FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE); | |
22eccdd7 | 388 | if (IS_ERR_OR_NULL(page)) |
31dbd01f IE |
389 | break; |
390 | if (PageKsm(page)) | |
dcddffd4 KS |
391 | ret = handle_mm_fault(vma, addr, |
392 | FAULT_FLAG_WRITE | FAULT_FLAG_REMOTE); | |
31dbd01f IE |
393 | else |
394 | ret = VM_FAULT_WRITE; | |
395 | put_page(page); | |
33692f27 | 396 | } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | VM_FAULT_OOM))); |
d952b791 HD |
397 | /* |
398 | * We must loop because handle_mm_fault() may back out if there's | |
399 | * any difficulty e.g. if pte accessed bit gets updated concurrently. | |
400 | * | |
401 | * VM_FAULT_WRITE is what we have been hoping for: it indicates that | |
402 | * COW has been broken, even if the vma does not permit VM_WRITE; | |
403 | * but note that a concurrent fault might break PageKsm for us. | |
404 | * | |
405 | * VM_FAULT_SIGBUS could occur if we race with truncation of the | |
406 | * backing file, which also invalidates anonymous pages: that's | |
407 | * okay, that truncation will have unmapped the PageKsm for us. | |
408 | * | |
409 | * VM_FAULT_OOM: at the time of writing (late July 2009), setting | |
410 | * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the | |
411 | * current task has TIF_MEMDIE set, and will be OOM killed on return | |
412 | * to user; and ksmd, having no mm, would never be chosen for that. | |
413 | * | |
414 | * But if the mm is in a limited mem_cgroup, then the fault may fail | |
415 | * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and | |
416 | * even ksmd can fail in this way - though it's usually breaking ksm | |
417 | * just to undo a merge it made a moment before, so unlikely to oom. | |
418 | * | |
419 | * That's a pity: we might therefore have more kernel pages allocated | |
420 | * than we're counting as nodes in the stable tree; but ksm_do_scan | |
421 | * will retry to break_cow on each pass, so should recover the page | |
422 | * in due course. The important thing is to not let VM_MERGEABLE | |
423 | * be cleared while any such pages might remain in the area. | |
424 | */ | |
425 | return (ret & VM_FAULT_OOM) ? -ENOMEM : 0; | |
31dbd01f IE |
426 | } |
427 | ||
ef694222 BL |
428 | static struct vm_area_struct *find_mergeable_vma(struct mm_struct *mm, |
429 | unsigned long addr) | |
430 | { | |
431 | struct vm_area_struct *vma; | |
432 | if (ksm_test_exit(mm)) | |
433 | return NULL; | |
434 | vma = find_vma(mm, addr); | |
435 | if (!vma || vma->vm_start > addr) | |
436 | return NULL; | |
437 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) | |
438 | return NULL; | |
439 | return vma; | |
440 | } | |
441 | ||
8dd3557a | 442 | static void break_cow(struct rmap_item *rmap_item) |
31dbd01f | 443 | { |
8dd3557a HD |
444 | struct mm_struct *mm = rmap_item->mm; |
445 | unsigned long addr = rmap_item->address; | |
31dbd01f IE |
446 | struct vm_area_struct *vma; |
447 | ||
4035c07a HD |
448 | /* |
449 | * It is not an accident that whenever we want to break COW | |
450 | * to undo, we also need to drop a reference to the anon_vma. | |
451 | */ | |
9e60109f | 452 | put_anon_vma(rmap_item->anon_vma); |
4035c07a | 453 | |
81464e30 | 454 | down_read(&mm->mmap_sem); |
ef694222 BL |
455 | vma = find_mergeable_vma(mm, addr); |
456 | if (vma) | |
457 | break_ksm(vma, addr); | |
31dbd01f IE |
458 | up_read(&mm->mmap_sem); |
459 | } | |
460 | ||
461 | static struct page *get_mergeable_page(struct rmap_item *rmap_item) | |
462 | { | |
463 | struct mm_struct *mm = rmap_item->mm; | |
464 | unsigned long addr = rmap_item->address; | |
465 | struct vm_area_struct *vma; | |
466 | struct page *page; | |
467 | ||
468 | down_read(&mm->mmap_sem); | |
ef694222 BL |
469 | vma = find_mergeable_vma(mm, addr); |
470 | if (!vma) | |
31dbd01f IE |
471 | goto out; |
472 | ||
473 | page = follow_page(vma, addr, FOLL_GET); | |
22eccdd7 | 474 | if (IS_ERR_OR_NULL(page)) |
31dbd01f | 475 | goto out; |
f765f540 | 476 | if (PageAnon(page)) { |
31dbd01f IE |
477 | flush_anon_page(vma, page, addr); |
478 | flush_dcache_page(page); | |
479 | } else { | |
480 | put_page(page); | |
c8f95ed1 AA |
481 | out: |
482 | page = NULL; | |
31dbd01f IE |
483 | } |
484 | up_read(&mm->mmap_sem); | |
485 | return page; | |
486 | } | |
487 | ||
90bd6fd3 PH |
488 | /* |
489 | * This helper is used for getting right index into array of tree roots. | |
490 | * When merge_across_nodes knob is set to 1, there are only two rb-trees for | |
491 | * stable and unstable pages from all nodes with roots in index 0. Otherwise, | |
492 | * every node has its own stable and unstable tree. | |
493 | */ | |
494 | static inline int get_kpfn_nid(unsigned long kpfn) | |
495 | { | |
d8fc16a8 | 496 | return ksm_merge_across_nodes ? 0 : NUMA(pfn_to_nid(kpfn)); |
90bd6fd3 PH |
497 | } |
498 | ||
4035c07a HD |
499 | static void remove_node_from_stable_tree(struct stable_node *stable_node) |
500 | { | |
501 | struct rmap_item *rmap_item; | |
4035c07a | 502 | |
b67bfe0d | 503 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
4035c07a HD |
504 | if (rmap_item->hlist.next) |
505 | ksm_pages_sharing--; | |
506 | else | |
507 | ksm_pages_shared--; | |
9e60109f | 508 | put_anon_vma(rmap_item->anon_vma); |
4035c07a HD |
509 | rmap_item->address &= PAGE_MASK; |
510 | cond_resched(); | |
511 | } | |
512 | ||
4146d2d6 HD |
513 | if (stable_node->head == &migrate_nodes) |
514 | list_del(&stable_node->list); | |
515 | else | |
516 | rb_erase(&stable_node->node, | |
ef53d16c | 517 | root_stable_tree + NUMA(stable_node->nid)); |
4035c07a HD |
518 | free_stable_node(stable_node); |
519 | } | |
520 | ||
521 | /* | |
522 | * get_ksm_page: checks if the page indicated by the stable node | |
523 | * is still its ksm page, despite having held no reference to it. | |
524 | * In which case we can trust the content of the page, and it | |
525 | * returns the gotten page; but if the page has now been zapped, | |
526 | * remove the stale node from the stable tree and return NULL. | |
c8d6553b | 527 | * But beware, the stable node's page might be being migrated. |
4035c07a HD |
528 | * |
529 | * You would expect the stable_node to hold a reference to the ksm page. | |
530 | * But if it increments the page's count, swapping out has to wait for | |
531 | * ksmd to come around again before it can free the page, which may take | |
532 | * seconds or even minutes: much too unresponsive. So instead we use a | |
533 | * "keyhole reference": access to the ksm page from the stable node peeps | |
534 | * out through its keyhole to see if that page still holds the right key, | |
535 | * pointing back to this stable node. This relies on freeing a PageAnon | |
536 | * page to reset its page->mapping to NULL, and relies on no other use of | |
537 | * a page to put something that might look like our key in page->mapping. | |
4035c07a HD |
538 | * is on its way to being freed; but it is an anomaly to bear in mind. |
539 | */ | |
8fdb3dbf | 540 | static struct page *get_ksm_page(struct stable_node *stable_node, bool lock_it) |
4035c07a HD |
541 | { |
542 | struct page *page; | |
543 | void *expected_mapping; | |
c8d6553b | 544 | unsigned long kpfn; |
4035c07a | 545 | |
bda807d4 MK |
546 | expected_mapping = (void *)((unsigned long)stable_node | |
547 | PAGE_MAPPING_KSM); | |
c8d6553b | 548 | again: |
4db0c3c2 | 549 | kpfn = READ_ONCE(stable_node->kpfn); |
c8d6553b HD |
550 | page = pfn_to_page(kpfn); |
551 | ||
552 | /* | |
553 | * page is computed from kpfn, so on most architectures reading | |
554 | * page->mapping is naturally ordered after reading node->kpfn, | |
555 | * but on Alpha we need to be more careful. | |
556 | */ | |
557 | smp_read_barrier_depends(); | |
4db0c3c2 | 558 | if (READ_ONCE(page->mapping) != expected_mapping) |
4035c07a | 559 | goto stale; |
c8d6553b HD |
560 | |
561 | /* | |
562 | * We cannot do anything with the page while its refcount is 0. | |
563 | * Usually 0 means free, or tail of a higher-order page: in which | |
564 | * case this node is no longer referenced, and should be freed; | |
565 | * however, it might mean that the page is under page_freeze_refs(). | |
566 | * The __remove_mapping() case is easy, again the node is now stale; | |
567 | * but if page is swapcache in migrate_page_move_mapping(), it might | |
568 | * still be our page, in which case it's essential to keep the node. | |
569 | */ | |
570 | while (!get_page_unless_zero(page)) { | |
571 | /* | |
572 | * Another check for page->mapping != expected_mapping would | |
573 | * work here too. We have chosen the !PageSwapCache test to | |
574 | * optimize the common case, when the page is or is about to | |
575 | * be freed: PageSwapCache is cleared (under spin_lock_irq) | |
576 | * in the freeze_refs section of __remove_mapping(); but Anon | |
577 | * page->mapping reset to NULL later, in free_pages_prepare(). | |
578 | */ | |
579 | if (!PageSwapCache(page)) | |
580 | goto stale; | |
581 | cpu_relax(); | |
582 | } | |
583 | ||
4db0c3c2 | 584 | if (READ_ONCE(page->mapping) != expected_mapping) { |
4035c07a HD |
585 | put_page(page); |
586 | goto stale; | |
587 | } | |
c8d6553b | 588 | |
8fdb3dbf | 589 | if (lock_it) { |
8aafa6a4 | 590 | lock_page(page); |
4db0c3c2 | 591 | if (READ_ONCE(page->mapping) != expected_mapping) { |
8aafa6a4 HD |
592 | unlock_page(page); |
593 | put_page(page); | |
594 | goto stale; | |
595 | } | |
596 | } | |
4035c07a | 597 | return page; |
c8d6553b | 598 | |
4035c07a | 599 | stale: |
c8d6553b HD |
600 | /* |
601 | * We come here from above when page->mapping or !PageSwapCache | |
602 | * suggests that the node is stale; but it might be under migration. | |
603 | * We need smp_rmb(), matching the smp_wmb() in ksm_migrate_page(), | |
604 | * before checking whether node->kpfn has been changed. | |
605 | */ | |
606 | smp_rmb(); | |
4db0c3c2 | 607 | if (READ_ONCE(stable_node->kpfn) != kpfn) |
c8d6553b | 608 | goto again; |
4035c07a HD |
609 | remove_node_from_stable_tree(stable_node); |
610 | return NULL; | |
611 | } | |
612 | ||
31dbd01f IE |
613 | /* |
614 | * Removing rmap_item from stable or unstable tree. | |
615 | * This function will clean the information from the stable/unstable tree. | |
616 | */ | |
617 | static void remove_rmap_item_from_tree(struct rmap_item *rmap_item) | |
618 | { | |
7b6ba2c7 HD |
619 | if (rmap_item->address & STABLE_FLAG) { |
620 | struct stable_node *stable_node; | |
5ad64688 | 621 | struct page *page; |
31dbd01f | 622 | |
7b6ba2c7 | 623 | stable_node = rmap_item->head; |
8aafa6a4 | 624 | page = get_ksm_page(stable_node, true); |
4035c07a HD |
625 | if (!page) |
626 | goto out; | |
5ad64688 | 627 | |
7b6ba2c7 | 628 | hlist_del(&rmap_item->hlist); |
4035c07a HD |
629 | unlock_page(page); |
630 | put_page(page); | |
08beca44 | 631 | |
98666f8a | 632 | if (!hlist_empty(&stable_node->hlist)) |
4035c07a HD |
633 | ksm_pages_sharing--; |
634 | else | |
7b6ba2c7 | 635 | ksm_pages_shared--; |
31dbd01f | 636 | |
9e60109f | 637 | put_anon_vma(rmap_item->anon_vma); |
93d17715 | 638 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 639 | |
7b6ba2c7 | 640 | } else if (rmap_item->address & UNSTABLE_FLAG) { |
31dbd01f IE |
641 | unsigned char age; |
642 | /* | |
9ba69294 | 643 | * Usually ksmd can and must skip the rb_erase, because |
31dbd01f | 644 | * root_unstable_tree was already reset to RB_ROOT. |
9ba69294 HD |
645 | * But be careful when an mm is exiting: do the rb_erase |
646 | * if this rmap_item was inserted by this scan, rather | |
647 | * than left over from before. | |
31dbd01f IE |
648 | */ |
649 | age = (unsigned char)(ksm_scan.seqnr - rmap_item->address); | |
cd551f97 | 650 | BUG_ON(age > 1); |
31dbd01f | 651 | if (!age) |
90bd6fd3 | 652 | rb_erase(&rmap_item->node, |
ef53d16c | 653 | root_unstable_tree + NUMA(rmap_item->nid)); |
473b0ce4 | 654 | ksm_pages_unshared--; |
93d17715 | 655 | rmap_item->address &= PAGE_MASK; |
31dbd01f | 656 | } |
4035c07a | 657 | out: |
31dbd01f IE |
658 | cond_resched(); /* we're called from many long loops */ |
659 | } | |
660 | ||
31dbd01f | 661 | static void remove_trailing_rmap_items(struct mm_slot *mm_slot, |
6514d511 | 662 | struct rmap_item **rmap_list) |
31dbd01f | 663 | { |
6514d511 HD |
664 | while (*rmap_list) { |
665 | struct rmap_item *rmap_item = *rmap_list; | |
666 | *rmap_list = rmap_item->rmap_list; | |
31dbd01f | 667 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
668 | free_rmap_item(rmap_item); |
669 | } | |
670 | } | |
671 | ||
672 | /* | |
e850dcf5 | 673 | * Though it's very tempting to unmerge rmap_items from stable tree rather |
31dbd01f IE |
674 | * than check every pte of a given vma, the locking doesn't quite work for |
675 | * that - an rmap_item is assigned to the stable tree after inserting ksm | |
676 | * page and upping mmap_sem. Nor does it fit with the way we skip dup'ing | |
677 | * rmap_items from parent to child at fork time (so as not to waste time | |
678 | * if exit comes before the next scan reaches it). | |
81464e30 HD |
679 | * |
680 | * Similarly, although we'd like to remove rmap_items (so updating counts | |
681 | * and freeing memory) when unmerging an area, it's easier to leave that | |
682 | * to the next pass of ksmd - consider, for example, how ksmd might be | |
683 | * in cmp_and_merge_page on one of the rmap_items we would be removing. | |
31dbd01f | 684 | */ |
d952b791 HD |
685 | static int unmerge_ksm_pages(struct vm_area_struct *vma, |
686 | unsigned long start, unsigned long end) | |
31dbd01f IE |
687 | { |
688 | unsigned long addr; | |
d952b791 | 689 | int err = 0; |
31dbd01f | 690 | |
d952b791 | 691 | for (addr = start; addr < end && !err; addr += PAGE_SIZE) { |
9ba69294 HD |
692 | if (ksm_test_exit(vma->vm_mm)) |
693 | break; | |
d952b791 HD |
694 | if (signal_pending(current)) |
695 | err = -ERESTARTSYS; | |
696 | else | |
697 | err = break_ksm(vma, addr); | |
698 | } | |
699 | return err; | |
31dbd01f IE |
700 | } |
701 | ||
2ffd8679 HD |
702 | #ifdef CONFIG_SYSFS |
703 | /* | |
704 | * Only called through the sysfs control interface: | |
705 | */ | |
cbf86cfe HD |
706 | static int remove_stable_node(struct stable_node *stable_node) |
707 | { | |
708 | struct page *page; | |
709 | int err; | |
710 | ||
711 | page = get_ksm_page(stable_node, true); | |
712 | if (!page) { | |
713 | /* | |
714 | * get_ksm_page did remove_node_from_stable_tree itself. | |
715 | */ | |
716 | return 0; | |
717 | } | |
718 | ||
8fdb3dbf HD |
719 | if (WARN_ON_ONCE(page_mapped(page))) { |
720 | /* | |
721 | * This should not happen: but if it does, just refuse to let | |
722 | * merge_across_nodes be switched - there is no need to panic. | |
723 | */ | |
cbf86cfe | 724 | err = -EBUSY; |
8fdb3dbf | 725 | } else { |
cbf86cfe | 726 | /* |
8fdb3dbf HD |
727 | * The stable node did not yet appear stale to get_ksm_page(), |
728 | * since that allows for an unmapped ksm page to be recognized | |
729 | * right up until it is freed; but the node is safe to remove. | |
cbf86cfe HD |
730 | * This page might be in a pagevec waiting to be freed, |
731 | * or it might be PageSwapCache (perhaps under writeback), | |
732 | * or it might have been removed from swapcache a moment ago. | |
733 | */ | |
734 | set_page_stable_node(page, NULL); | |
735 | remove_node_from_stable_tree(stable_node); | |
736 | err = 0; | |
737 | } | |
738 | ||
739 | unlock_page(page); | |
740 | put_page(page); | |
741 | return err; | |
742 | } | |
743 | ||
744 | static int remove_all_stable_nodes(void) | |
745 | { | |
03640418 | 746 | struct stable_node *stable_node, *next; |
cbf86cfe HD |
747 | int nid; |
748 | int err = 0; | |
749 | ||
ef53d16c | 750 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
cbf86cfe HD |
751 | while (root_stable_tree[nid].rb_node) { |
752 | stable_node = rb_entry(root_stable_tree[nid].rb_node, | |
753 | struct stable_node, node); | |
754 | if (remove_stable_node(stable_node)) { | |
755 | err = -EBUSY; | |
756 | break; /* proceed to next nid */ | |
757 | } | |
758 | cond_resched(); | |
759 | } | |
760 | } | |
03640418 | 761 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
762 | if (remove_stable_node(stable_node)) |
763 | err = -EBUSY; | |
764 | cond_resched(); | |
765 | } | |
cbf86cfe HD |
766 | return err; |
767 | } | |
768 | ||
d952b791 | 769 | static int unmerge_and_remove_all_rmap_items(void) |
31dbd01f IE |
770 | { |
771 | struct mm_slot *mm_slot; | |
772 | struct mm_struct *mm; | |
773 | struct vm_area_struct *vma; | |
d952b791 HD |
774 | int err = 0; |
775 | ||
776 | spin_lock(&ksm_mmlist_lock); | |
9ba69294 | 777 | ksm_scan.mm_slot = list_entry(ksm_mm_head.mm_list.next, |
d952b791 HD |
778 | struct mm_slot, mm_list); |
779 | spin_unlock(&ksm_mmlist_lock); | |
31dbd01f | 780 | |
9ba69294 HD |
781 | for (mm_slot = ksm_scan.mm_slot; |
782 | mm_slot != &ksm_mm_head; mm_slot = ksm_scan.mm_slot) { | |
31dbd01f IE |
783 | mm = mm_slot->mm; |
784 | down_read(&mm->mmap_sem); | |
785 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
9ba69294 HD |
786 | if (ksm_test_exit(mm)) |
787 | break; | |
31dbd01f IE |
788 | if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma) |
789 | continue; | |
d952b791 HD |
790 | err = unmerge_ksm_pages(vma, |
791 | vma->vm_start, vma->vm_end); | |
9ba69294 HD |
792 | if (err) |
793 | goto error; | |
31dbd01f | 794 | } |
9ba69294 | 795 | |
6514d511 | 796 | remove_trailing_rmap_items(mm_slot, &mm_slot->rmap_list); |
7496fea9 | 797 | up_read(&mm->mmap_sem); |
d952b791 HD |
798 | |
799 | spin_lock(&ksm_mmlist_lock); | |
9ba69294 | 800 | ksm_scan.mm_slot = list_entry(mm_slot->mm_list.next, |
d952b791 | 801 | struct mm_slot, mm_list); |
9ba69294 | 802 | if (ksm_test_exit(mm)) { |
4ca3a69b | 803 | hash_del(&mm_slot->link); |
9ba69294 HD |
804 | list_del(&mm_slot->mm_list); |
805 | spin_unlock(&ksm_mmlist_lock); | |
806 | ||
807 | free_mm_slot(mm_slot); | |
808 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
9ba69294 | 809 | mmdrop(mm); |
7496fea9 | 810 | } else |
9ba69294 | 811 | spin_unlock(&ksm_mmlist_lock); |
31dbd01f IE |
812 | } |
813 | ||
cbf86cfe HD |
814 | /* Clean up stable nodes, but don't worry if some are still busy */ |
815 | remove_all_stable_nodes(); | |
d952b791 | 816 | ksm_scan.seqnr = 0; |
9ba69294 HD |
817 | return 0; |
818 | ||
819 | error: | |
820 | up_read(&mm->mmap_sem); | |
31dbd01f | 821 | spin_lock(&ksm_mmlist_lock); |
d952b791 | 822 | ksm_scan.mm_slot = &ksm_mm_head; |
31dbd01f | 823 | spin_unlock(&ksm_mmlist_lock); |
d952b791 | 824 | return err; |
31dbd01f | 825 | } |
2ffd8679 | 826 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 827 | |
31dbd01f IE |
828 | static u32 calc_checksum(struct page *page) |
829 | { | |
830 | u32 checksum; | |
9b04c5fe | 831 | void *addr = kmap_atomic(page); |
31dbd01f | 832 | checksum = jhash2(addr, PAGE_SIZE / 4, 17); |
9b04c5fe | 833 | kunmap_atomic(addr); |
31dbd01f IE |
834 | return checksum; |
835 | } | |
836 | ||
837 | static int memcmp_pages(struct page *page1, struct page *page2) | |
838 | { | |
839 | char *addr1, *addr2; | |
840 | int ret; | |
841 | ||
9b04c5fe CW |
842 | addr1 = kmap_atomic(page1); |
843 | addr2 = kmap_atomic(page2); | |
31dbd01f | 844 | ret = memcmp(addr1, addr2, PAGE_SIZE); |
9b04c5fe CW |
845 | kunmap_atomic(addr2); |
846 | kunmap_atomic(addr1); | |
31dbd01f IE |
847 | return ret; |
848 | } | |
849 | ||
850 | static inline int pages_identical(struct page *page1, struct page *page2) | |
851 | { | |
852 | return !memcmp_pages(page1, page2); | |
853 | } | |
854 | ||
855 | static int write_protect_page(struct vm_area_struct *vma, struct page *page, | |
856 | pte_t *orig_pte) | |
857 | { | |
858 | struct mm_struct *mm = vma->vm_mm; | |
36eaff33 KS |
859 | struct page_vma_mapped_walk pvmw = { |
860 | .page = page, | |
861 | .vma = vma, | |
862 | }; | |
31dbd01f IE |
863 | int swapped; |
864 | int err = -EFAULT; | |
6bdb913f HE |
865 | unsigned long mmun_start; /* For mmu_notifiers */ |
866 | unsigned long mmun_end; /* For mmu_notifiers */ | |
31dbd01f | 867 | |
36eaff33 KS |
868 | pvmw.address = page_address_in_vma(page, vma); |
869 | if (pvmw.address == -EFAULT) | |
31dbd01f IE |
870 | goto out; |
871 | ||
29ad768c | 872 | BUG_ON(PageTransCompound(page)); |
6bdb913f | 873 | |
36eaff33 KS |
874 | mmun_start = pvmw.address; |
875 | mmun_end = pvmw.address + PAGE_SIZE; | |
6bdb913f HE |
876 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
877 | ||
36eaff33 | 878 | if (!page_vma_mapped_walk(&pvmw)) |
6bdb913f | 879 | goto out_mn; |
36eaff33 KS |
880 | if (WARN_ONCE(!pvmw.pte, "Unexpected PMD mapping?")) |
881 | goto out_unlock; | |
31dbd01f | 882 | |
595cd8f2 AK |
883 | if (pte_write(*pvmw.pte) || pte_dirty(*pvmw.pte) || |
884 | (pte_protnone(*pvmw.pte) && pte_savedwrite(*pvmw.pte))) { | |
31dbd01f IE |
885 | pte_t entry; |
886 | ||
887 | swapped = PageSwapCache(page); | |
36eaff33 | 888 | flush_cache_page(vma, pvmw.address, page_to_pfn(page)); |
31dbd01f | 889 | /* |
25985edc | 890 | * Ok this is tricky, when get_user_pages_fast() run it doesn't |
31dbd01f IE |
891 | * take any lock, therefore the check that we are going to make |
892 | * with the pagecount against the mapcount is racey and | |
893 | * O_DIRECT can happen right after the check. | |
894 | * So we clear the pte and flush the tlb before the check | |
895 | * this assure us that no O_DIRECT can happen after the check | |
896 | * or in the middle of the check. | |
897 | */ | |
36eaff33 | 898 | entry = ptep_clear_flush_notify(vma, pvmw.address, pvmw.pte); |
31dbd01f IE |
899 | /* |
900 | * Check that no O_DIRECT or similar I/O is in progress on the | |
901 | * page | |
902 | */ | |
31e855ea | 903 | if (page_mapcount(page) + 1 + swapped != page_count(page)) { |
36eaff33 | 904 | set_pte_at(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f IE |
905 | goto out_unlock; |
906 | } | |
4e31635c HD |
907 | if (pte_dirty(entry)) |
908 | set_page_dirty(page); | |
595cd8f2 AK |
909 | |
910 | if (pte_protnone(entry)) | |
911 | entry = pte_mkclean(pte_clear_savedwrite(entry)); | |
912 | else | |
913 | entry = pte_mkclean(pte_wrprotect(entry)); | |
36eaff33 | 914 | set_pte_at_notify(mm, pvmw.address, pvmw.pte, entry); |
31dbd01f | 915 | } |
36eaff33 | 916 | *orig_pte = *pvmw.pte; |
31dbd01f IE |
917 | err = 0; |
918 | ||
919 | out_unlock: | |
36eaff33 | 920 | page_vma_mapped_walk_done(&pvmw); |
6bdb913f HE |
921 | out_mn: |
922 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
31dbd01f IE |
923 | out: |
924 | return err; | |
925 | } | |
926 | ||
927 | /** | |
928 | * replace_page - replace page in vma by new ksm page | |
8dd3557a HD |
929 | * @vma: vma that holds the pte pointing to page |
930 | * @page: the page we are replacing by kpage | |
931 | * @kpage: the ksm page we replace page by | |
31dbd01f IE |
932 | * @orig_pte: the original value of the pte |
933 | * | |
934 | * Returns 0 on success, -EFAULT on failure. | |
935 | */ | |
8dd3557a HD |
936 | static int replace_page(struct vm_area_struct *vma, struct page *page, |
937 | struct page *kpage, pte_t orig_pte) | |
31dbd01f IE |
938 | { |
939 | struct mm_struct *mm = vma->vm_mm; | |
31dbd01f IE |
940 | pmd_t *pmd; |
941 | pte_t *ptep; | |
e86c59b1 | 942 | pte_t newpte; |
31dbd01f IE |
943 | spinlock_t *ptl; |
944 | unsigned long addr; | |
31dbd01f | 945 | int err = -EFAULT; |
6bdb913f HE |
946 | unsigned long mmun_start; /* For mmu_notifiers */ |
947 | unsigned long mmun_end; /* For mmu_notifiers */ | |
31dbd01f | 948 | |
8dd3557a | 949 | addr = page_address_in_vma(page, vma); |
31dbd01f IE |
950 | if (addr == -EFAULT) |
951 | goto out; | |
952 | ||
6219049a BL |
953 | pmd = mm_find_pmd(mm, addr); |
954 | if (!pmd) | |
31dbd01f | 955 | goto out; |
31dbd01f | 956 | |
6bdb913f HE |
957 | mmun_start = addr; |
958 | mmun_end = addr + PAGE_SIZE; | |
959 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
960 | ||
31dbd01f IE |
961 | ptep = pte_offset_map_lock(mm, pmd, addr, &ptl); |
962 | if (!pte_same(*ptep, orig_pte)) { | |
963 | pte_unmap_unlock(ptep, ptl); | |
6bdb913f | 964 | goto out_mn; |
31dbd01f IE |
965 | } |
966 | ||
e86c59b1 CI |
967 | /* |
968 | * No need to check ksm_use_zero_pages here: we can only have a | |
969 | * zero_page here if ksm_use_zero_pages was enabled alreaady. | |
970 | */ | |
971 | if (!is_zero_pfn(page_to_pfn(kpage))) { | |
972 | get_page(kpage); | |
973 | page_add_anon_rmap(kpage, vma, addr, false); | |
974 | newpte = mk_pte(kpage, vma->vm_page_prot); | |
975 | } else { | |
976 | newpte = pte_mkspecial(pfn_pte(page_to_pfn(kpage), | |
977 | vma->vm_page_prot)); | |
978 | } | |
31dbd01f IE |
979 | |
980 | flush_cache_page(vma, addr, pte_pfn(*ptep)); | |
34ee645e | 981 | ptep_clear_flush_notify(vma, addr, ptep); |
e86c59b1 | 982 | set_pte_at_notify(mm, addr, ptep, newpte); |
31dbd01f | 983 | |
d281ee61 | 984 | page_remove_rmap(page, false); |
ae52a2ad HD |
985 | if (!page_mapped(page)) |
986 | try_to_free_swap(page); | |
8dd3557a | 987 | put_page(page); |
31dbd01f IE |
988 | |
989 | pte_unmap_unlock(ptep, ptl); | |
990 | err = 0; | |
6bdb913f HE |
991 | out_mn: |
992 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
31dbd01f IE |
993 | out: |
994 | return err; | |
995 | } | |
996 | ||
997 | /* | |
998 | * try_to_merge_one_page - take two pages and merge them into one | |
8dd3557a HD |
999 | * @vma: the vma that holds the pte pointing to page |
1000 | * @page: the PageAnon page that we want to replace with kpage | |
80e14822 HD |
1001 | * @kpage: the PageKsm page that we want to map instead of page, |
1002 | * or NULL the first time when we want to use page as kpage. | |
31dbd01f IE |
1003 | * |
1004 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
1005 | */ | |
1006 | static int try_to_merge_one_page(struct vm_area_struct *vma, | |
8dd3557a | 1007 | struct page *page, struct page *kpage) |
31dbd01f IE |
1008 | { |
1009 | pte_t orig_pte = __pte(0); | |
1010 | int err = -EFAULT; | |
1011 | ||
db114b83 HD |
1012 | if (page == kpage) /* ksm page forked */ |
1013 | return 0; | |
1014 | ||
8dd3557a | 1015 | if (!PageAnon(page)) |
31dbd01f IE |
1016 | goto out; |
1017 | ||
31dbd01f IE |
1018 | /* |
1019 | * We need the page lock to read a stable PageSwapCache in | |
1020 | * write_protect_page(). We use trylock_page() instead of | |
1021 | * lock_page() because we don't want to wait here - we | |
1022 | * prefer to continue scanning and merging different pages, | |
1023 | * then come back to this page when it is unlocked. | |
1024 | */ | |
8dd3557a | 1025 | if (!trylock_page(page)) |
31e855ea | 1026 | goto out; |
f765f540 KS |
1027 | |
1028 | if (PageTransCompound(page)) { | |
1029 | err = split_huge_page(page); | |
1030 | if (err) | |
1031 | goto out_unlock; | |
1032 | } | |
1033 | ||
31dbd01f IE |
1034 | /* |
1035 | * If this anonymous page is mapped only here, its pte may need | |
1036 | * to be write-protected. If it's mapped elsewhere, all of its | |
1037 | * ptes are necessarily already write-protected. But in either | |
1038 | * case, we need to lock and check page_count is not raised. | |
1039 | */ | |
80e14822 HD |
1040 | if (write_protect_page(vma, page, &orig_pte) == 0) { |
1041 | if (!kpage) { | |
1042 | /* | |
1043 | * While we hold page lock, upgrade page from | |
1044 | * PageAnon+anon_vma to PageKsm+NULL stable_node: | |
1045 | * stable_tree_insert() will update stable_node. | |
1046 | */ | |
1047 | set_page_stable_node(page, NULL); | |
1048 | mark_page_accessed(page); | |
337ed7eb MK |
1049 | /* |
1050 | * Page reclaim just frees a clean page with no dirty | |
1051 | * ptes: make sure that the ksm page would be swapped. | |
1052 | */ | |
1053 | if (!PageDirty(page)) | |
1054 | SetPageDirty(page); | |
80e14822 HD |
1055 | err = 0; |
1056 | } else if (pages_identical(page, kpage)) | |
1057 | err = replace_page(vma, page, kpage, orig_pte); | |
1058 | } | |
31dbd01f | 1059 | |
80e14822 | 1060 | if ((vma->vm_flags & VM_LOCKED) && kpage && !err) { |
73848b46 | 1061 | munlock_vma_page(page); |
5ad64688 HD |
1062 | if (!PageMlocked(kpage)) { |
1063 | unlock_page(page); | |
5ad64688 HD |
1064 | lock_page(kpage); |
1065 | mlock_vma_page(kpage); | |
1066 | page = kpage; /* for final unlock */ | |
1067 | } | |
1068 | } | |
73848b46 | 1069 | |
f765f540 | 1070 | out_unlock: |
8dd3557a | 1071 | unlock_page(page); |
31dbd01f IE |
1072 | out: |
1073 | return err; | |
1074 | } | |
1075 | ||
81464e30 HD |
1076 | /* |
1077 | * try_to_merge_with_ksm_page - like try_to_merge_two_pages, | |
1078 | * but no new kernel page is allocated: kpage must already be a ksm page. | |
8dd3557a HD |
1079 | * |
1080 | * This function returns 0 if the pages were merged, -EFAULT otherwise. | |
81464e30 | 1081 | */ |
8dd3557a HD |
1082 | static int try_to_merge_with_ksm_page(struct rmap_item *rmap_item, |
1083 | struct page *page, struct page *kpage) | |
81464e30 | 1084 | { |
8dd3557a | 1085 | struct mm_struct *mm = rmap_item->mm; |
81464e30 HD |
1086 | struct vm_area_struct *vma; |
1087 | int err = -EFAULT; | |
1088 | ||
8dd3557a | 1089 | down_read(&mm->mmap_sem); |
85c6e8dd AA |
1090 | vma = find_mergeable_vma(mm, rmap_item->address); |
1091 | if (!vma) | |
81464e30 HD |
1092 | goto out; |
1093 | ||
8dd3557a | 1094 | err = try_to_merge_one_page(vma, page, kpage); |
db114b83 HD |
1095 | if (err) |
1096 | goto out; | |
1097 | ||
bc56620b HD |
1098 | /* Unstable nid is in union with stable anon_vma: remove first */ |
1099 | remove_rmap_item_from_tree(rmap_item); | |
1100 | ||
db114b83 | 1101 | /* Must get reference to anon_vma while still holding mmap_sem */ |
9e60109f PZ |
1102 | rmap_item->anon_vma = vma->anon_vma; |
1103 | get_anon_vma(vma->anon_vma); | |
81464e30 | 1104 | out: |
8dd3557a | 1105 | up_read(&mm->mmap_sem); |
81464e30 HD |
1106 | return err; |
1107 | } | |
1108 | ||
31dbd01f IE |
1109 | /* |
1110 | * try_to_merge_two_pages - take two identical pages and prepare them | |
1111 | * to be merged into one page. | |
1112 | * | |
8dd3557a HD |
1113 | * This function returns the kpage if we successfully merged two identical |
1114 | * pages into one ksm page, NULL otherwise. | |
31dbd01f | 1115 | * |
80e14822 | 1116 | * Note that this function upgrades page to ksm page: if one of the pages |
31dbd01f IE |
1117 | * is already a ksm page, try_to_merge_with_ksm_page should be used. |
1118 | */ | |
8dd3557a HD |
1119 | static struct page *try_to_merge_two_pages(struct rmap_item *rmap_item, |
1120 | struct page *page, | |
1121 | struct rmap_item *tree_rmap_item, | |
1122 | struct page *tree_page) | |
31dbd01f | 1123 | { |
80e14822 | 1124 | int err; |
31dbd01f | 1125 | |
80e14822 | 1126 | err = try_to_merge_with_ksm_page(rmap_item, page, NULL); |
31dbd01f | 1127 | if (!err) { |
8dd3557a | 1128 | err = try_to_merge_with_ksm_page(tree_rmap_item, |
80e14822 | 1129 | tree_page, page); |
31dbd01f | 1130 | /* |
81464e30 HD |
1131 | * If that fails, we have a ksm page with only one pte |
1132 | * pointing to it: so break it. | |
31dbd01f | 1133 | */ |
4035c07a | 1134 | if (err) |
8dd3557a | 1135 | break_cow(rmap_item); |
31dbd01f | 1136 | } |
80e14822 | 1137 | return err ? NULL : page; |
31dbd01f IE |
1138 | } |
1139 | ||
31dbd01f | 1140 | /* |
8dd3557a | 1141 | * stable_tree_search - search for page inside the stable tree |
31dbd01f IE |
1142 | * |
1143 | * This function checks if there is a page inside the stable tree | |
1144 | * with identical content to the page that we are scanning right now. | |
1145 | * | |
7b6ba2c7 | 1146 | * This function returns the stable tree node of identical content if found, |
31dbd01f IE |
1147 | * NULL otherwise. |
1148 | */ | |
62b61f61 | 1149 | static struct page *stable_tree_search(struct page *page) |
31dbd01f | 1150 | { |
90bd6fd3 | 1151 | int nid; |
ef53d16c | 1152 | struct rb_root *root; |
4146d2d6 HD |
1153 | struct rb_node **new; |
1154 | struct rb_node *parent; | |
1155 | struct stable_node *stable_node; | |
1156 | struct stable_node *page_node; | |
31dbd01f | 1157 | |
4146d2d6 HD |
1158 | page_node = page_stable_node(page); |
1159 | if (page_node && page_node->head != &migrate_nodes) { | |
1160 | /* ksm page forked */ | |
08beca44 | 1161 | get_page(page); |
62b61f61 | 1162 | return page; |
08beca44 HD |
1163 | } |
1164 | ||
90bd6fd3 | 1165 | nid = get_kpfn_nid(page_to_pfn(page)); |
ef53d16c | 1166 | root = root_stable_tree + nid; |
4146d2d6 | 1167 | again: |
ef53d16c | 1168 | new = &root->rb_node; |
4146d2d6 | 1169 | parent = NULL; |
90bd6fd3 | 1170 | |
4146d2d6 | 1171 | while (*new) { |
4035c07a | 1172 | struct page *tree_page; |
31dbd01f IE |
1173 | int ret; |
1174 | ||
08beca44 | 1175 | cond_resched(); |
4146d2d6 | 1176 | stable_node = rb_entry(*new, struct stable_node, node); |
8aafa6a4 | 1177 | tree_page = get_ksm_page(stable_node, false); |
f2e5ff85 AA |
1178 | if (!tree_page) { |
1179 | /* | |
1180 | * If we walked over a stale stable_node, | |
1181 | * get_ksm_page() will call rb_erase() and it | |
1182 | * may rebalance the tree from under us. So | |
1183 | * restart the search from scratch. Returning | |
1184 | * NULL would be safe too, but we'd generate | |
1185 | * false negative insertions just because some | |
1186 | * stable_node was stale. | |
1187 | */ | |
1188 | goto again; | |
1189 | } | |
31dbd01f | 1190 | |
4035c07a | 1191 | ret = memcmp_pages(page, tree_page); |
c8d6553b | 1192 | put_page(tree_page); |
31dbd01f | 1193 | |
4146d2d6 | 1194 | parent = *new; |
c8d6553b | 1195 | if (ret < 0) |
4146d2d6 | 1196 | new = &parent->rb_left; |
c8d6553b | 1197 | else if (ret > 0) |
4146d2d6 | 1198 | new = &parent->rb_right; |
c8d6553b HD |
1199 | else { |
1200 | /* | |
1201 | * Lock and unlock the stable_node's page (which | |
1202 | * might already have been migrated) so that page | |
1203 | * migration is sure to notice its raised count. | |
1204 | * It would be more elegant to return stable_node | |
1205 | * than kpage, but that involves more changes. | |
1206 | */ | |
1207 | tree_page = get_ksm_page(stable_node, true); | |
4146d2d6 | 1208 | if (tree_page) { |
c8d6553b | 1209 | unlock_page(tree_page); |
4146d2d6 HD |
1210 | if (get_kpfn_nid(stable_node->kpfn) != |
1211 | NUMA(stable_node->nid)) { | |
1212 | put_page(tree_page); | |
1213 | goto replace; | |
1214 | } | |
1215 | return tree_page; | |
1216 | } | |
1217 | /* | |
1218 | * There is now a place for page_node, but the tree may | |
1219 | * have been rebalanced, so re-evaluate parent and new. | |
1220 | */ | |
1221 | if (page_node) | |
1222 | goto again; | |
1223 | return NULL; | |
c8d6553b | 1224 | } |
31dbd01f IE |
1225 | } |
1226 | ||
4146d2d6 HD |
1227 | if (!page_node) |
1228 | return NULL; | |
1229 | ||
1230 | list_del(&page_node->list); | |
1231 | DO_NUMA(page_node->nid = nid); | |
1232 | rb_link_node(&page_node->node, parent, new); | |
ef53d16c | 1233 | rb_insert_color(&page_node->node, root); |
4146d2d6 HD |
1234 | get_page(page); |
1235 | return page; | |
1236 | ||
1237 | replace: | |
1238 | if (page_node) { | |
1239 | list_del(&page_node->list); | |
1240 | DO_NUMA(page_node->nid = nid); | |
ef53d16c | 1241 | rb_replace_node(&stable_node->node, &page_node->node, root); |
4146d2d6 HD |
1242 | get_page(page); |
1243 | } else { | |
ef53d16c | 1244 | rb_erase(&stable_node->node, root); |
4146d2d6 HD |
1245 | page = NULL; |
1246 | } | |
1247 | stable_node->head = &migrate_nodes; | |
1248 | list_add(&stable_node->list, stable_node->head); | |
1249 | return page; | |
31dbd01f IE |
1250 | } |
1251 | ||
1252 | /* | |
e850dcf5 | 1253 | * stable_tree_insert - insert stable tree node pointing to new ksm page |
31dbd01f IE |
1254 | * into the stable tree. |
1255 | * | |
7b6ba2c7 HD |
1256 | * This function returns the stable tree node just allocated on success, |
1257 | * NULL otherwise. | |
31dbd01f | 1258 | */ |
7b6ba2c7 | 1259 | static struct stable_node *stable_tree_insert(struct page *kpage) |
31dbd01f | 1260 | { |
90bd6fd3 PH |
1261 | int nid; |
1262 | unsigned long kpfn; | |
ef53d16c | 1263 | struct rb_root *root; |
90bd6fd3 | 1264 | struct rb_node **new; |
f2e5ff85 | 1265 | struct rb_node *parent; |
7b6ba2c7 | 1266 | struct stable_node *stable_node; |
31dbd01f | 1267 | |
90bd6fd3 PH |
1268 | kpfn = page_to_pfn(kpage); |
1269 | nid = get_kpfn_nid(kpfn); | |
ef53d16c | 1270 | root = root_stable_tree + nid; |
f2e5ff85 AA |
1271 | again: |
1272 | parent = NULL; | |
ef53d16c | 1273 | new = &root->rb_node; |
90bd6fd3 | 1274 | |
31dbd01f | 1275 | while (*new) { |
4035c07a | 1276 | struct page *tree_page; |
31dbd01f IE |
1277 | int ret; |
1278 | ||
08beca44 | 1279 | cond_resched(); |
7b6ba2c7 | 1280 | stable_node = rb_entry(*new, struct stable_node, node); |
8aafa6a4 | 1281 | tree_page = get_ksm_page(stable_node, false); |
f2e5ff85 AA |
1282 | if (!tree_page) { |
1283 | /* | |
1284 | * If we walked over a stale stable_node, | |
1285 | * get_ksm_page() will call rb_erase() and it | |
1286 | * may rebalance the tree from under us. So | |
1287 | * restart the search from scratch. Returning | |
1288 | * NULL would be safe too, but we'd generate | |
1289 | * false negative insertions just because some | |
1290 | * stable_node was stale. | |
1291 | */ | |
1292 | goto again; | |
1293 | } | |
31dbd01f | 1294 | |
4035c07a HD |
1295 | ret = memcmp_pages(kpage, tree_page); |
1296 | put_page(tree_page); | |
31dbd01f IE |
1297 | |
1298 | parent = *new; | |
1299 | if (ret < 0) | |
1300 | new = &parent->rb_left; | |
1301 | else if (ret > 0) | |
1302 | new = &parent->rb_right; | |
1303 | else { | |
1304 | /* | |
1305 | * It is not a bug that stable_tree_search() didn't | |
1306 | * find this node: because at that time our page was | |
1307 | * not yet write-protected, so may have changed since. | |
1308 | */ | |
1309 | return NULL; | |
1310 | } | |
1311 | } | |
1312 | ||
7b6ba2c7 HD |
1313 | stable_node = alloc_stable_node(); |
1314 | if (!stable_node) | |
1315 | return NULL; | |
31dbd01f | 1316 | |
7b6ba2c7 | 1317 | INIT_HLIST_HEAD(&stable_node->hlist); |
90bd6fd3 | 1318 | stable_node->kpfn = kpfn; |
08beca44 | 1319 | set_page_stable_node(kpage, stable_node); |
4146d2d6 | 1320 | DO_NUMA(stable_node->nid = nid); |
e850dcf5 | 1321 | rb_link_node(&stable_node->node, parent, new); |
ef53d16c | 1322 | rb_insert_color(&stable_node->node, root); |
08beca44 | 1323 | |
7b6ba2c7 | 1324 | return stable_node; |
31dbd01f IE |
1325 | } |
1326 | ||
1327 | /* | |
8dd3557a HD |
1328 | * unstable_tree_search_insert - search for identical page, |
1329 | * else insert rmap_item into the unstable tree. | |
31dbd01f IE |
1330 | * |
1331 | * This function searches for a page in the unstable tree identical to the | |
1332 | * page currently being scanned; and if no identical page is found in the | |
1333 | * tree, we insert rmap_item as a new object into the unstable tree. | |
1334 | * | |
1335 | * This function returns pointer to rmap_item found to be identical | |
1336 | * to the currently scanned page, NULL otherwise. | |
1337 | * | |
1338 | * This function does both searching and inserting, because they share | |
1339 | * the same walking algorithm in an rbtree. | |
1340 | */ | |
8dd3557a HD |
1341 | static |
1342 | struct rmap_item *unstable_tree_search_insert(struct rmap_item *rmap_item, | |
1343 | struct page *page, | |
1344 | struct page **tree_pagep) | |
31dbd01f | 1345 | { |
90bd6fd3 PH |
1346 | struct rb_node **new; |
1347 | struct rb_root *root; | |
31dbd01f | 1348 | struct rb_node *parent = NULL; |
90bd6fd3 PH |
1349 | int nid; |
1350 | ||
1351 | nid = get_kpfn_nid(page_to_pfn(page)); | |
ef53d16c | 1352 | root = root_unstable_tree + nid; |
90bd6fd3 | 1353 | new = &root->rb_node; |
31dbd01f IE |
1354 | |
1355 | while (*new) { | |
1356 | struct rmap_item *tree_rmap_item; | |
8dd3557a | 1357 | struct page *tree_page; |
31dbd01f IE |
1358 | int ret; |
1359 | ||
d178f27f | 1360 | cond_resched(); |
31dbd01f | 1361 | tree_rmap_item = rb_entry(*new, struct rmap_item, node); |
8dd3557a | 1362 | tree_page = get_mergeable_page(tree_rmap_item); |
c8f95ed1 | 1363 | if (!tree_page) |
31dbd01f IE |
1364 | return NULL; |
1365 | ||
1366 | /* | |
8dd3557a | 1367 | * Don't substitute a ksm page for a forked page. |
31dbd01f | 1368 | */ |
8dd3557a HD |
1369 | if (page == tree_page) { |
1370 | put_page(tree_page); | |
31dbd01f IE |
1371 | return NULL; |
1372 | } | |
1373 | ||
8dd3557a | 1374 | ret = memcmp_pages(page, tree_page); |
31dbd01f IE |
1375 | |
1376 | parent = *new; | |
1377 | if (ret < 0) { | |
8dd3557a | 1378 | put_page(tree_page); |
31dbd01f IE |
1379 | new = &parent->rb_left; |
1380 | } else if (ret > 0) { | |
8dd3557a | 1381 | put_page(tree_page); |
31dbd01f | 1382 | new = &parent->rb_right; |
b599cbdf HD |
1383 | } else if (!ksm_merge_across_nodes && |
1384 | page_to_nid(tree_page) != nid) { | |
1385 | /* | |
1386 | * If tree_page has been migrated to another NUMA node, | |
1387 | * it will be flushed out and put in the right unstable | |
1388 | * tree next time: only merge with it when across_nodes. | |
1389 | */ | |
1390 | put_page(tree_page); | |
1391 | return NULL; | |
31dbd01f | 1392 | } else { |
8dd3557a | 1393 | *tree_pagep = tree_page; |
31dbd01f IE |
1394 | return tree_rmap_item; |
1395 | } | |
1396 | } | |
1397 | ||
7b6ba2c7 | 1398 | rmap_item->address |= UNSTABLE_FLAG; |
31dbd01f | 1399 | rmap_item->address |= (ksm_scan.seqnr & SEQNR_MASK); |
e850dcf5 | 1400 | DO_NUMA(rmap_item->nid = nid); |
31dbd01f | 1401 | rb_link_node(&rmap_item->node, parent, new); |
90bd6fd3 | 1402 | rb_insert_color(&rmap_item->node, root); |
31dbd01f | 1403 | |
473b0ce4 | 1404 | ksm_pages_unshared++; |
31dbd01f IE |
1405 | return NULL; |
1406 | } | |
1407 | ||
1408 | /* | |
1409 | * stable_tree_append - add another rmap_item to the linked list of | |
1410 | * rmap_items hanging off a given node of the stable tree, all sharing | |
1411 | * the same ksm page. | |
1412 | */ | |
1413 | static void stable_tree_append(struct rmap_item *rmap_item, | |
7b6ba2c7 | 1414 | struct stable_node *stable_node) |
31dbd01f | 1415 | { |
7b6ba2c7 | 1416 | rmap_item->head = stable_node; |
31dbd01f | 1417 | rmap_item->address |= STABLE_FLAG; |
7b6ba2c7 | 1418 | hlist_add_head(&rmap_item->hlist, &stable_node->hlist); |
e178dfde | 1419 | |
7b6ba2c7 HD |
1420 | if (rmap_item->hlist.next) |
1421 | ksm_pages_sharing++; | |
1422 | else | |
1423 | ksm_pages_shared++; | |
31dbd01f IE |
1424 | } |
1425 | ||
1426 | /* | |
81464e30 HD |
1427 | * cmp_and_merge_page - first see if page can be merged into the stable tree; |
1428 | * if not, compare checksum to previous and if it's the same, see if page can | |
1429 | * be inserted into the unstable tree, or merged with a page already there and | |
1430 | * both transferred to the stable tree. | |
31dbd01f IE |
1431 | * |
1432 | * @page: the page that we are searching identical page to. | |
1433 | * @rmap_item: the reverse mapping into the virtual address of this page | |
1434 | */ | |
1435 | static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item) | |
1436 | { | |
31dbd01f | 1437 | struct rmap_item *tree_rmap_item; |
8dd3557a | 1438 | struct page *tree_page = NULL; |
7b6ba2c7 | 1439 | struct stable_node *stable_node; |
8dd3557a | 1440 | struct page *kpage; |
31dbd01f IE |
1441 | unsigned int checksum; |
1442 | int err; | |
1443 | ||
4146d2d6 HD |
1444 | stable_node = page_stable_node(page); |
1445 | if (stable_node) { | |
1446 | if (stable_node->head != &migrate_nodes && | |
1447 | get_kpfn_nid(stable_node->kpfn) != NUMA(stable_node->nid)) { | |
1448 | rb_erase(&stable_node->node, | |
ef53d16c | 1449 | root_stable_tree + NUMA(stable_node->nid)); |
4146d2d6 HD |
1450 | stable_node->head = &migrate_nodes; |
1451 | list_add(&stable_node->list, stable_node->head); | |
1452 | } | |
1453 | if (stable_node->head != &migrate_nodes && | |
1454 | rmap_item->head == stable_node) | |
1455 | return; | |
1456 | } | |
31dbd01f IE |
1457 | |
1458 | /* We first start with searching the page inside the stable tree */ | |
62b61f61 | 1459 | kpage = stable_tree_search(page); |
4146d2d6 HD |
1460 | if (kpage == page && rmap_item->head == stable_node) { |
1461 | put_page(kpage); | |
1462 | return; | |
1463 | } | |
1464 | ||
1465 | remove_rmap_item_from_tree(rmap_item); | |
1466 | ||
62b61f61 | 1467 | if (kpage) { |
08beca44 | 1468 | err = try_to_merge_with_ksm_page(rmap_item, page, kpage); |
31dbd01f IE |
1469 | if (!err) { |
1470 | /* | |
1471 | * The page was successfully merged: | |
1472 | * add its rmap_item to the stable tree. | |
1473 | */ | |
5ad64688 | 1474 | lock_page(kpage); |
62b61f61 | 1475 | stable_tree_append(rmap_item, page_stable_node(kpage)); |
5ad64688 | 1476 | unlock_page(kpage); |
31dbd01f | 1477 | } |
8dd3557a | 1478 | put_page(kpage); |
31dbd01f IE |
1479 | return; |
1480 | } | |
1481 | ||
1482 | /* | |
4035c07a HD |
1483 | * If the hash value of the page has changed from the last time |
1484 | * we calculated it, this page is changing frequently: therefore we | |
1485 | * don't want to insert it in the unstable tree, and we don't want | |
1486 | * to waste our time searching for something identical to it there. | |
31dbd01f IE |
1487 | */ |
1488 | checksum = calc_checksum(page); | |
1489 | if (rmap_item->oldchecksum != checksum) { | |
1490 | rmap_item->oldchecksum = checksum; | |
1491 | return; | |
1492 | } | |
1493 | ||
e86c59b1 CI |
1494 | /* |
1495 | * Same checksum as an empty page. We attempt to merge it with the | |
1496 | * appropriate zero page if the user enabled this via sysfs. | |
1497 | */ | |
1498 | if (ksm_use_zero_pages && (checksum == zero_checksum)) { | |
1499 | struct vm_area_struct *vma; | |
1500 | ||
1501 | vma = find_mergeable_vma(rmap_item->mm, rmap_item->address); | |
1502 | err = try_to_merge_one_page(vma, page, | |
1503 | ZERO_PAGE(rmap_item->address)); | |
1504 | /* | |
1505 | * In case of failure, the page was not really empty, so we | |
1506 | * need to continue. Otherwise we're done. | |
1507 | */ | |
1508 | if (!err) | |
1509 | return; | |
1510 | } | |
8dd3557a HD |
1511 | tree_rmap_item = |
1512 | unstable_tree_search_insert(rmap_item, page, &tree_page); | |
31dbd01f | 1513 | if (tree_rmap_item) { |
8dd3557a HD |
1514 | kpage = try_to_merge_two_pages(rmap_item, page, |
1515 | tree_rmap_item, tree_page); | |
1516 | put_page(tree_page); | |
8dd3557a | 1517 | if (kpage) { |
bc56620b HD |
1518 | /* |
1519 | * The pages were successfully merged: insert new | |
1520 | * node in the stable tree and add both rmap_items. | |
1521 | */ | |
5ad64688 | 1522 | lock_page(kpage); |
7b6ba2c7 HD |
1523 | stable_node = stable_tree_insert(kpage); |
1524 | if (stable_node) { | |
1525 | stable_tree_append(tree_rmap_item, stable_node); | |
1526 | stable_tree_append(rmap_item, stable_node); | |
1527 | } | |
5ad64688 | 1528 | unlock_page(kpage); |
7b6ba2c7 | 1529 | |
31dbd01f IE |
1530 | /* |
1531 | * If we fail to insert the page into the stable tree, | |
1532 | * we will have 2 virtual addresses that are pointing | |
1533 | * to a ksm page left outside the stable tree, | |
1534 | * in which case we need to break_cow on both. | |
1535 | */ | |
7b6ba2c7 | 1536 | if (!stable_node) { |
8dd3557a HD |
1537 | break_cow(tree_rmap_item); |
1538 | break_cow(rmap_item); | |
31dbd01f IE |
1539 | } |
1540 | } | |
31dbd01f IE |
1541 | } |
1542 | } | |
1543 | ||
1544 | static struct rmap_item *get_next_rmap_item(struct mm_slot *mm_slot, | |
6514d511 | 1545 | struct rmap_item **rmap_list, |
31dbd01f IE |
1546 | unsigned long addr) |
1547 | { | |
1548 | struct rmap_item *rmap_item; | |
1549 | ||
6514d511 HD |
1550 | while (*rmap_list) { |
1551 | rmap_item = *rmap_list; | |
93d17715 | 1552 | if ((rmap_item->address & PAGE_MASK) == addr) |
31dbd01f | 1553 | return rmap_item; |
31dbd01f IE |
1554 | if (rmap_item->address > addr) |
1555 | break; | |
6514d511 | 1556 | *rmap_list = rmap_item->rmap_list; |
31dbd01f | 1557 | remove_rmap_item_from_tree(rmap_item); |
31dbd01f IE |
1558 | free_rmap_item(rmap_item); |
1559 | } | |
1560 | ||
1561 | rmap_item = alloc_rmap_item(); | |
1562 | if (rmap_item) { | |
1563 | /* It has already been zeroed */ | |
1564 | rmap_item->mm = mm_slot->mm; | |
1565 | rmap_item->address = addr; | |
6514d511 HD |
1566 | rmap_item->rmap_list = *rmap_list; |
1567 | *rmap_list = rmap_item; | |
31dbd01f IE |
1568 | } |
1569 | return rmap_item; | |
1570 | } | |
1571 | ||
1572 | static struct rmap_item *scan_get_next_rmap_item(struct page **page) | |
1573 | { | |
1574 | struct mm_struct *mm; | |
1575 | struct mm_slot *slot; | |
1576 | struct vm_area_struct *vma; | |
1577 | struct rmap_item *rmap_item; | |
90bd6fd3 | 1578 | int nid; |
31dbd01f IE |
1579 | |
1580 | if (list_empty(&ksm_mm_head.mm_list)) | |
1581 | return NULL; | |
1582 | ||
1583 | slot = ksm_scan.mm_slot; | |
1584 | if (slot == &ksm_mm_head) { | |
2919bfd0 HD |
1585 | /* |
1586 | * A number of pages can hang around indefinitely on per-cpu | |
1587 | * pagevecs, raised page count preventing write_protect_page | |
1588 | * from merging them. Though it doesn't really matter much, | |
1589 | * it is puzzling to see some stuck in pages_volatile until | |
1590 | * other activity jostles them out, and they also prevented | |
1591 | * LTP's KSM test from succeeding deterministically; so drain | |
1592 | * them here (here rather than on entry to ksm_do_scan(), | |
1593 | * so we don't IPI too often when pages_to_scan is set low). | |
1594 | */ | |
1595 | lru_add_drain_all(); | |
1596 | ||
4146d2d6 HD |
1597 | /* |
1598 | * Whereas stale stable_nodes on the stable_tree itself | |
1599 | * get pruned in the regular course of stable_tree_search(), | |
1600 | * those moved out to the migrate_nodes list can accumulate: | |
1601 | * so prune them once before each full scan. | |
1602 | */ | |
1603 | if (!ksm_merge_across_nodes) { | |
03640418 | 1604 | struct stable_node *stable_node, *next; |
4146d2d6 HD |
1605 | struct page *page; |
1606 | ||
03640418 GT |
1607 | list_for_each_entry_safe(stable_node, next, |
1608 | &migrate_nodes, list) { | |
4146d2d6 HD |
1609 | page = get_ksm_page(stable_node, false); |
1610 | if (page) | |
1611 | put_page(page); | |
1612 | cond_resched(); | |
1613 | } | |
1614 | } | |
1615 | ||
ef53d16c | 1616 | for (nid = 0; nid < ksm_nr_node_ids; nid++) |
90bd6fd3 | 1617 | root_unstable_tree[nid] = RB_ROOT; |
31dbd01f IE |
1618 | |
1619 | spin_lock(&ksm_mmlist_lock); | |
1620 | slot = list_entry(slot->mm_list.next, struct mm_slot, mm_list); | |
1621 | ksm_scan.mm_slot = slot; | |
1622 | spin_unlock(&ksm_mmlist_lock); | |
2b472611 HD |
1623 | /* |
1624 | * Although we tested list_empty() above, a racing __ksm_exit | |
1625 | * of the last mm on the list may have removed it since then. | |
1626 | */ | |
1627 | if (slot == &ksm_mm_head) | |
1628 | return NULL; | |
31dbd01f IE |
1629 | next_mm: |
1630 | ksm_scan.address = 0; | |
6514d511 | 1631 | ksm_scan.rmap_list = &slot->rmap_list; |
31dbd01f IE |
1632 | } |
1633 | ||
1634 | mm = slot->mm; | |
1635 | down_read(&mm->mmap_sem); | |
9ba69294 HD |
1636 | if (ksm_test_exit(mm)) |
1637 | vma = NULL; | |
1638 | else | |
1639 | vma = find_vma(mm, ksm_scan.address); | |
1640 | ||
1641 | for (; vma; vma = vma->vm_next) { | |
31dbd01f IE |
1642 | if (!(vma->vm_flags & VM_MERGEABLE)) |
1643 | continue; | |
1644 | if (ksm_scan.address < vma->vm_start) | |
1645 | ksm_scan.address = vma->vm_start; | |
1646 | if (!vma->anon_vma) | |
1647 | ksm_scan.address = vma->vm_end; | |
1648 | ||
1649 | while (ksm_scan.address < vma->vm_end) { | |
9ba69294 HD |
1650 | if (ksm_test_exit(mm)) |
1651 | break; | |
31dbd01f | 1652 | *page = follow_page(vma, ksm_scan.address, FOLL_GET); |
21ae5b01 AA |
1653 | if (IS_ERR_OR_NULL(*page)) { |
1654 | ksm_scan.address += PAGE_SIZE; | |
1655 | cond_resched(); | |
1656 | continue; | |
1657 | } | |
f765f540 | 1658 | if (PageAnon(*page)) { |
31dbd01f IE |
1659 | flush_anon_page(vma, *page, ksm_scan.address); |
1660 | flush_dcache_page(*page); | |
1661 | rmap_item = get_next_rmap_item(slot, | |
6514d511 | 1662 | ksm_scan.rmap_list, ksm_scan.address); |
31dbd01f | 1663 | if (rmap_item) { |
6514d511 HD |
1664 | ksm_scan.rmap_list = |
1665 | &rmap_item->rmap_list; | |
31dbd01f IE |
1666 | ksm_scan.address += PAGE_SIZE; |
1667 | } else | |
1668 | put_page(*page); | |
1669 | up_read(&mm->mmap_sem); | |
1670 | return rmap_item; | |
1671 | } | |
21ae5b01 | 1672 | put_page(*page); |
31dbd01f IE |
1673 | ksm_scan.address += PAGE_SIZE; |
1674 | cond_resched(); | |
1675 | } | |
1676 | } | |
1677 | ||
9ba69294 HD |
1678 | if (ksm_test_exit(mm)) { |
1679 | ksm_scan.address = 0; | |
6514d511 | 1680 | ksm_scan.rmap_list = &slot->rmap_list; |
9ba69294 | 1681 | } |
31dbd01f IE |
1682 | /* |
1683 | * Nuke all the rmap_items that are above this current rmap: | |
1684 | * because there were no VM_MERGEABLE vmas with such addresses. | |
1685 | */ | |
6514d511 | 1686 | remove_trailing_rmap_items(slot, ksm_scan.rmap_list); |
31dbd01f IE |
1687 | |
1688 | spin_lock(&ksm_mmlist_lock); | |
cd551f97 HD |
1689 | ksm_scan.mm_slot = list_entry(slot->mm_list.next, |
1690 | struct mm_slot, mm_list); | |
1691 | if (ksm_scan.address == 0) { | |
1692 | /* | |
1693 | * We've completed a full scan of all vmas, holding mmap_sem | |
1694 | * throughout, and found no VM_MERGEABLE: so do the same as | |
1695 | * __ksm_exit does to remove this mm from all our lists now. | |
9ba69294 HD |
1696 | * This applies either when cleaning up after __ksm_exit |
1697 | * (but beware: we can reach here even before __ksm_exit), | |
1698 | * or when all VM_MERGEABLE areas have been unmapped (and | |
1699 | * mmap_sem then protects against race with MADV_MERGEABLE). | |
cd551f97 | 1700 | */ |
4ca3a69b | 1701 | hash_del(&slot->link); |
cd551f97 | 1702 | list_del(&slot->mm_list); |
9ba69294 HD |
1703 | spin_unlock(&ksm_mmlist_lock); |
1704 | ||
cd551f97 HD |
1705 | free_mm_slot(slot); |
1706 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
9ba69294 HD |
1707 | up_read(&mm->mmap_sem); |
1708 | mmdrop(mm); | |
1709 | } else { | |
9ba69294 | 1710 | up_read(&mm->mmap_sem); |
7496fea9 ZC |
1711 | /* |
1712 | * up_read(&mm->mmap_sem) first because after | |
1713 | * spin_unlock(&ksm_mmlist_lock) run, the "mm" may | |
1714 | * already have been freed under us by __ksm_exit() | |
1715 | * because the "mm_slot" is still hashed and | |
1716 | * ksm_scan.mm_slot doesn't point to it anymore. | |
1717 | */ | |
1718 | spin_unlock(&ksm_mmlist_lock); | |
cd551f97 | 1719 | } |
31dbd01f IE |
1720 | |
1721 | /* Repeat until we've completed scanning the whole list */ | |
cd551f97 | 1722 | slot = ksm_scan.mm_slot; |
31dbd01f IE |
1723 | if (slot != &ksm_mm_head) |
1724 | goto next_mm; | |
1725 | ||
31dbd01f IE |
1726 | ksm_scan.seqnr++; |
1727 | return NULL; | |
1728 | } | |
1729 | ||
1730 | /** | |
1731 | * ksm_do_scan - the ksm scanner main worker function. | |
1732 | * @scan_npages - number of pages we want to scan before we return. | |
1733 | */ | |
1734 | static void ksm_do_scan(unsigned int scan_npages) | |
1735 | { | |
1736 | struct rmap_item *rmap_item; | |
22eccdd7 | 1737 | struct page *uninitialized_var(page); |
31dbd01f | 1738 | |
878aee7d | 1739 | while (scan_npages-- && likely(!freezing(current))) { |
31dbd01f IE |
1740 | cond_resched(); |
1741 | rmap_item = scan_get_next_rmap_item(&page); | |
1742 | if (!rmap_item) | |
1743 | return; | |
4146d2d6 | 1744 | cmp_and_merge_page(page, rmap_item); |
31dbd01f IE |
1745 | put_page(page); |
1746 | } | |
1747 | } | |
1748 | ||
6e158384 HD |
1749 | static int ksmd_should_run(void) |
1750 | { | |
1751 | return (ksm_run & KSM_RUN_MERGE) && !list_empty(&ksm_mm_head.mm_list); | |
1752 | } | |
1753 | ||
31dbd01f IE |
1754 | static int ksm_scan_thread(void *nothing) |
1755 | { | |
878aee7d | 1756 | set_freezable(); |
339aa624 | 1757 | set_user_nice(current, 5); |
31dbd01f IE |
1758 | |
1759 | while (!kthread_should_stop()) { | |
6e158384 | 1760 | mutex_lock(&ksm_thread_mutex); |
ef4d43a8 | 1761 | wait_while_offlining(); |
6e158384 | 1762 | if (ksmd_should_run()) |
31dbd01f | 1763 | ksm_do_scan(ksm_thread_pages_to_scan); |
6e158384 HD |
1764 | mutex_unlock(&ksm_thread_mutex); |
1765 | ||
878aee7d AA |
1766 | try_to_freeze(); |
1767 | ||
6e158384 | 1768 | if (ksmd_should_run()) { |
31dbd01f IE |
1769 | schedule_timeout_interruptible( |
1770 | msecs_to_jiffies(ksm_thread_sleep_millisecs)); | |
1771 | } else { | |
878aee7d | 1772 | wait_event_freezable(ksm_thread_wait, |
6e158384 | 1773 | ksmd_should_run() || kthread_should_stop()); |
31dbd01f IE |
1774 | } |
1775 | } | |
1776 | return 0; | |
1777 | } | |
1778 | ||
f8af4da3 HD |
1779 | int ksm_madvise(struct vm_area_struct *vma, unsigned long start, |
1780 | unsigned long end, int advice, unsigned long *vm_flags) | |
1781 | { | |
1782 | struct mm_struct *mm = vma->vm_mm; | |
d952b791 | 1783 | int err; |
f8af4da3 HD |
1784 | |
1785 | switch (advice) { | |
1786 | case MADV_MERGEABLE: | |
1787 | /* | |
1788 | * Be somewhat over-protective for now! | |
1789 | */ | |
1790 | if (*vm_flags & (VM_MERGEABLE | VM_SHARED | VM_MAYSHARE | | |
1791 | VM_PFNMAP | VM_IO | VM_DONTEXPAND | | |
0661a336 | 1792 | VM_HUGETLB | VM_MIXEDMAP)) |
f8af4da3 HD |
1793 | return 0; /* just ignore the advice */ |
1794 | ||
cc2383ec KK |
1795 | #ifdef VM_SAO |
1796 | if (*vm_flags & VM_SAO) | |
1797 | return 0; | |
1798 | #endif | |
1799 | ||
d952b791 HD |
1800 | if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) { |
1801 | err = __ksm_enter(mm); | |
1802 | if (err) | |
1803 | return err; | |
1804 | } | |
f8af4da3 HD |
1805 | |
1806 | *vm_flags |= VM_MERGEABLE; | |
1807 | break; | |
1808 | ||
1809 | case MADV_UNMERGEABLE: | |
1810 | if (!(*vm_flags & VM_MERGEABLE)) | |
1811 | return 0; /* just ignore the advice */ | |
1812 | ||
d952b791 HD |
1813 | if (vma->anon_vma) { |
1814 | err = unmerge_ksm_pages(vma, start, end); | |
1815 | if (err) | |
1816 | return err; | |
1817 | } | |
f8af4da3 HD |
1818 | |
1819 | *vm_flags &= ~VM_MERGEABLE; | |
1820 | break; | |
1821 | } | |
1822 | ||
1823 | return 0; | |
1824 | } | |
1825 | ||
1826 | int __ksm_enter(struct mm_struct *mm) | |
1827 | { | |
6e158384 HD |
1828 | struct mm_slot *mm_slot; |
1829 | int needs_wakeup; | |
1830 | ||
1831 | mm_slot = alloc_mm_slot(); | |
31dbd01f IE |
1832 | if (!mm_slot) |
1833 | return -ENOMEM; | |
1834 | ||
6e158384 HD |
1835 | /* Check ksm_run too? Would need tighter locking */ |
1836 | needs_wakeup = list_empty(&ksm_mm_head.mm_list); | |
1837 | ||
31dbd01f IE |
1838 | spin_lock(&ksm_mmlist_lock); |
1839 | insert_to_mm_slots_hash(mm, mm_slot); | |
1840 | /* | |
cbf86cfe HD |
1841 | * When KSM_RUN_MERGE (or KSM_RUN_STOP), |
1842 | * insert just behind the scanning cursor, to let the area settle | |
31dbd01f IE |
1843 | * down a little; when fork is followed by immediate exec, we don't |
1844 | * want ksmd to waste time setting up and tearing down an rmap_list. | |
cbf86cfe HD |
1845 | * |
1846 | * But when KSM_RUN_UNMERGE, it's important to insert ahead of its | |
1847 | * scanning cursor, otherwise KSM pages in newly forked mms will be | |
1848 | * missed: then we might as well insert at the end of the list. | |
31dbd01f | 1849 | */ |
cbf86cfe HD |
1850 | if (ksm_run & KSM_RUN_UNMERGE) |
1851 | list_add_tail(&mm_slot->mm_list, &ksm_mm_head.mm_list); | |
1852 | else | |
1853 | list_add_tail(&mm_slot->mm_list, &ksm_scan.mm_slot->mm_list); | |
31dbd01f IE |
1854 | spin_unlock(&ksm_mmlist_lock); |
1855 | ||
f8af4da3 | 1856 | set_bit(MMF_VM_MERGEABLE, &mm->flags); |
9ba69294 | 1857 | atomic_inc(&mm->mm_count); |
6e158384 HD |
1858 | |
1859 | if (needs_wakeup) | |
1860 | wake_up_interruptible(&ksm_thread_wait); | |
1861 | ||
f8af4da3 HD |
1862 | return 0; |
1863 | } | |
1864 | ||
1c2fb7a4 | 1865 | void __ksm_exit(struct mm_struct *mm) |
f8af4da3 | 1866 | { |
cd551f97 | 1867 | struct mm_slot *mm_slot; |
9ba69294 | 1868 | int easy_to_free = 0; |
cd551f97 | 1869 | |
31dbd01f | 1870 | /* |
9ba69294 HD |
1871 | * This process is exiting: if it's straightforward (as is the |
1872 | * case when ksmd was never running), free mm_slot immediately. | |
1873 | * But if it's at the cursor or has rmap_items linked to it, use | |
1874 | * mmap_sem to synchronize with any break_cows before pagetables | |
1875 | * are freed, and leave the mm_slot on the list for ksmd to free. | |
1876 | * Beware: ksm may already have noticed it exiting and freed the slot. | |
31dbd01f | 1877 | */ |
9ba69294 | 1878 | |
cd551f97 HD |
1879 | spin_lock(&ksm_mmlist_lock); |
1880 | mm_slot = get_mm_slot(mm); | |
9ba69294 | 1881 | if (mm_slot && ksm_scan.mm_slot != mm_slot) { |
6514d511 | 1882 | if (!mm_slot->rmap_list) { |
4ca3a69b | 1883 | hash_del(&mm_slot->link); |
9ba69294 HD |
1884 | list_del(&mm_slot->mm_list); |
1885 | easy_to_free = 1; | |
1886 | } else { | |
1887 | list_move(&mm_slot->mm_list, | |
1888 | &ksm_scan.mm_slot->mm_list); | |
1889 | } | |
cd551f97 | 1890 | } |
cd551f97 HD |
1891 | spin_unlock(&ksm_mmlist_lock); |
1892 | ||
9ba69294 HD |
1893 | if (easy_to_free) { |
1894 | free_mm_slot(mm_slot); | |
1895 | clear_bit(MMF_VM_MERGEABLE, &mm->flags); | |
1896 | mmdrop(mm); | |
1897 | } else if (mm_slot) { | |
9ba69294 HD |
1898 | down_write(&mm->mmap_sem); |
1899 | up_write(&mm->mmap_sem); | |
9ba69294 | 1900 | } |
31dbd01f IE |
1901 | } |
1902 | ||
cbf86cfe | 1903 | struct page *ksm_might_need_to_copy(struct page *page, |
5ad64688 HD |
1904 | struct vm_area_struct *vma, unsigned long address) |
1905 | { | |
cbf86cfe | 1906 | struct anon_vma *anon_vma = page_anon_vma(page); |
5ad64688 HD |
1907 | struct page *new_page; |
1908 | ||
cbf86cfe HD |
1909 | if (PageKsm(page)) { |
1910 | if (page_stable_node(page) && | |
1911 | !(ksm_run & KSM_RUN_UNMERGE)) | |
1912 | return page; /* no need to copy it */ | |
1913 | } else if (!anon_vma) { | |
1914 | return page; /* no need to copy it */ | |
1915 | } else if (anon_vma->root == vma->anon_vma->root && | |
1916 | page->index == linear_page_index(vma, address)) { | |
1917 | return page; /* still no need to copy it */ | |
1918 | } | |
1919 | if (!PageUptodate(page)) | |
1920 | return page; /* let do_swap_page report the error */ | |
1921 | ||
5ad64688 HD |
1922 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
1923 | if (new_page) { | |
1924 | copy_user_highpage(new_page, page, address, vma); | |
1925 | ||
1926 | SetPageDirty(new_page); | |
1927 | __SetPageUptodate(new_page); | |
48c935ad | 1928 | __SetPageLocked(new_page); |
5ad64688 HD |
1929 | } |
1930 | ||
5ad64688 HD |
1931 | return new_page; |
1932 | } | |
1933 | ||
051ac83a | 1934 | int rmap_walk_ksm(struct page *page, struct rmap_walk_control *rwc) |
e9995ef9 HD |
1935 | { |
1936 | struct stable_node *stable_node; | |
e9995ef9 HD |
1937 | struct rmap_item *rmap_item; |
1938 | int ret = SWAP_AGAIN; | |
1939 | int search_new_forks = 0; | |
1940 | ||
309381fe | 1941 | VM_BUG_ON_PAGE(!PageKsm(page), page); |
9f32624b JK |
1942 | |
1943 | /* | |
1944 | * Rely on the page lock to protect against concurrent modifications | |
1945 | * to that page's node of the stable tree. | |
1946 | */ | |
309381fe | 1947 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
e9995ef9 HD |
1948 | |
1949 | stable_node = page_stable_node(page); | |
1950 | if (!stable_node) | |
1951 | return ret; | |
1952 | again: | |
b67bfe0d | 1953 | hlist_for_each_entry(rmap_item, &stable_node->hlist, hlist) { |
e9995ef9 | 1954 | struct anon_vma *anon_vma = rmap_item->anon_vma; |
5beb4930 | 1955 | struct anon_vma_chain *vmac; |
e9995ef9 HD |
1956 | struct vm_area_struct *vma; |
1957 | ||
ad12695f | 1958 | cond_resched(); |
b6b19f25 | 1959 | anon_vma_lock_read(anon_vma); |
bf181b9f ML |
1960 | anon_vma_interval_tree_foreach(vmac, &anon_vma->rb_root, |
1961 | 0, ULONG_MAX) { | |
ad12695f | 1962 | cond_resched(); |
5beb4930 | 1963 | vma = vmac->vma; |
e9995ef9 HD |
1964 | if (rmap_item->address < vma->vm_start || |
1965 | rmap_item->address >= vma->vm_end) | |
1966 | continue; | |
1967 | /* | |
1968 | * Initially we examine only the vma which covers this | |
1969 | * rmap_item; but later, if there is still work to do, | |
1970 | * we examine covering vmas in other mms: in case they | |
1971 | * were forked from the original since ksmd passed. | |
1972 | */ | |
1973 | if ((rmap_item->mm == vma->vm_mm) == search_new_forks) | |
1974 | continue; | |
1975 | ||
0dd1c7bb JK |
1976 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
1977 | continue; | |
1978 | ||
051ac83a JK |
1979 | ret = rwc->rmap_one(page, vma, |
1980 | rmap_item->address, rwc->arg); | |
e9995ef9 | 1981 | if (ret != SWAP_AGAIN) { |
b6b19f25 | 1982 | anon_vma_unlock_read(anon_vma); |
e9995ef9 HD |
1983 | goto out; |
1984 | } | |
0dd1c7bb JK |
1985 | if (rwc->done && rwc->done(page)) { |
1986 | anon_vma_unlock_read(anon_vma); | |
1987 | goto out; | |
1988 | } | |
e9995ef9 | 1989 | } |
b6b19f25 | 1990 | anon_vma_unlock_read(anon_vma); |
e9995ef9 HD |
1991 | } |
1992 | if (!search_new_forks++) | |
1993 | goto again; | |
1994 | out: | |
1995 | return ret; | |
1996 | } | |
1997 | ||
52629506 | 1998 | #ifdef CONFIG_MIGRATION |
e9995ef9 HD |
1999 | void ksm_migrate_page(struct page *newpage, struct page *oldpage) |
2000 | { | |
2001 | struct stable_node *stable_node; | |
2002 | ||
309381fe SL |
2003 | VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage); |
2004 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
2005 | VM_BUG_ON_PAGE(newpage->mapping != oldpage->mapping, newpage); | |
e9995ef9 HD |
2006 | |
2007 | stable_node = page_stable_node(newpage); | |
2008 | if (stable_node) { | |
309381fe | 2009 | VM_BUG_ON_PAGE(stable_node->kpfn != page_to_pfn(oldpage), oldpage); |
62b61f61 | 2010 | stable_node->kpfn = page_to_pfn(newpage); |
c8d6553b HD |
2011 | /* |
2012 | * newpage->mapping was set in advance; now we need smp_wmb() | |
2013 | * to make sure that the new stable_node->kpfn is visible | |
2014 | * to get_ksm_page() before it can see that oldpage->mapping | |
2015 | * has gone stale (or that PageSwapCache has been cleared). | |
2016 | */ | |
2017 | smp_wmb(); | |
2018 | set_page_stable_node(oldpage, NULL); | |
e9995ef9 HD |
2019 | } |
2020 | } | |
2021 | #endif /* CONFIG_MIGRATION */ | |
2022 | ||
62b61f61 | 2023 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 HD |
2024 | static void wait_while_offlining(void) |
2025 | { | |
2026 | while (ksm_run & KSM_RUN_OFFLINE) { | |
2027 | mutex_unlock(&ksm_thread_mutex); | |
2028 | wait_on_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE), | |
74316201 | 2029 | TASK_UNINTERRUPTIBLE); |
ef4d43a8 HD |
2030 | mutex_lock(&ksm_thread_mutex); |
2031 | } | |
2032 | } | |
2033 | ||
ee0ea59c HD |
2034 | static void ksm_check_stable_tree(unsigned long start_pfn, |
2035 | unsigned long end_pfn) | |
62b61f61 | 2036 | { |
03640418 | 2037 | struct stable_node *stable_node, *next; |
62b61f61 | 2038 | struct rb_node *node; |
90bd6fd3 | 2039 | int nid; |
62b61f61 | 2040 | |
ef53d16c HD |
2041 | for (nid = 0; nid < ksm_nr_node_ids; nid++) { |
2042 | node = rb_first(root_stable_tree + nid); | |
ee0ea59c | 2043 | while (node) { |
90bd6fd3 PH |
2044 | stable_node = rb_entry(node, struct stable_node, node); |
2045 | if (stable_node->kpfn >= start_pfn && | |
ee0ea59c HD |
2046 | stable_node->kpfn < end_pfn) { |
2047 | /* | |
2048 | * Don't get_ksm_page, page has already gone: | |
2049 | * which is why we keep kpfn instead of page* | |
2050 | */ | |
2051 | remove_node_from_stable_tree(stable_node); | |
ef53d16c | 2052 | node = rb_first(root_stable_tree + nid); |
ee0ea59c HD |
2053 | } else |
2054 | node = rb_next(node); | |
2055 | cond_resched(); | |
90bd6fd3 | 2056 | } |
ee0ea59c | 2057 | } |
03640418 | 2058 | list_for_each_entry_safe(stable_node, next, &migrate_nodes, list) { |
4146d2d6 HD |
2059 | if (stable_node->kpfn >= start_pfn && |
2060 | stable_node->kpfn < end_pfn) | |
2061 | remove_node_from_stable_tree(stable_node); | |
2062 | cond_resched(); | |
2063 | } | |
62b61f61 HD |
2064 | } |
2065 | ||
2066 | static int ksm_memory_callback(struct notifier_block *self, | |
2067 | unsigned long action, void *arg) | |
2068 | { | |
2069 | struct memory_notify *mn = arg; | |
62b61f61 HD |
2070 | |
2071 | switch (action) { | |
2072 | case MEM_GOING_OFFLINE: | |
2073 | /* | |
ef4d43a8 HD |
2074 | * Prevent ksm_do_scan(), unmerge_and_remove_all_rmap_items() |
2075 | * and remove_all_stable_nodes() while memory is going offline: | |
2076 | * it is unsafe for them to touch the stable tree at this time. | |
2077 | * But unmerge_ksm_pages(), rmap lookups and other entry points | |
2078 | * which do not need the ksm_thread_mutex are all safe. | |
62b61f61 | 2079 | */ |
ef4d43a8 HD |
2080 | mutex_lock(&ksm_thread_mutex); |
2081 | ksm_run |= KSM_RUN_OFFLINE; | |
2082 | mutex_unlock(&ksm_thread_mutex); | |
62b61f61 HD |
2083 | break; |
2084 | ||
2085 | case MEM_OFFLINE: | |
2086 | /* | |
2087 | * Most of the work is done by page migration; but there might | |
2088 | * be a few stable_nodes left over, still pointing to struct | |
ee0ea59c HD |
2089 | * pages which have been offlined: prune those from the tree, |
2090 | * otherwise get_ksm_page() might later try to access a | |
2091 | * non-existent struct page. | |
62b61f61 | 2092 | */ |
ee0ea59c HD |
2093 | ksm_check_stable_tree(mn->start_pfn, |
2094 | mn->start_pfn + mn->nr_pages); | |
62b61f61 HD |
2095 | /* fallthrough */ |
2096 | ||
2097 | case MEM_CANCEL_OFFLINE: | |
ef4d43a8 HD |
2098 | mutex_lock(&ksm_thread_mutex); |
2099 | ksm_run &= ~KSM_RUN_OFFLINE; | |
62b61f61 | 2100 | mutex_unlock(&ksm_thread_mutex); |
ef4d43a8 HD |
2101 | |
2102 | smp_mb(); /* wake_up_bit advises this */ | |
2103 | wake_up_bit(&ksm_run, ilog2(KSM_RUN_OFFLINE)); | |
62b61f61 HD |
2104 | break; |
2105 | } | |
2106 | return NOTIFY_OK; | |
2107 | } | |
ef4d43a8 HD |
2108 | #else |
2109 | static void wait_while_offlining(void) | |
2110 | { | |
2111 | } | |
62b61f61 HD |
2112 | #endif /* CONFIG_MEMORY_HOTREMOVE */ |
2113 | ||
2ffd8679 HD |
2114 | #ifdef CONFIG_SYSFS |
2115 | /* | |
2116 | * This all compiles without CONFIG_SYSFS, but is a waste of space. | |
2117 | */ | |
2118 | ||
31dbd01f IE |
2119 | #define KSM_ATTR_RO(_name) \ |
2120 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2121 | #define KSM_ATTR(_name) \ | |
2122 | static struct kobj_attribute _name##_attr = \ | |
2123 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2124 | ||
2125 | static ssize_t sleep_millisecs_show(struct kobject *kobj, | |
2126 | struct kobj_attribute *attr, char *buf) | |
2127 | { | |
2128 | return sprintf(buf, "%u\n", ksm_thread_sleep_millisecs); | |
2129 | } | |
2130 | ||
2131 | static ssize_t sleep_millisecs_store(struct kobject *kobj, | |
2132 | struct kobj_attribute *attr, | |
2133 | const char *buf, size_t count) | |
2134 | { | |
2135 | unsigned long msecs; | |
2136 | int err; | |
2137 | ||
3dbb95f7 | 2138 | err = kstrtoul(buf, 10, &msecs); |
31dbd01f IE |
2139 | if (err || msecs > UINT_MAX) |
2140 | return -EINVAL; | |
2141 | ||
2142 | ksm_thread_sleep_millisecs = msecs; | |
2143 | ||
2144 | return count; | |
2145 | } | |
2146 | KSM_ATTR(sleep_millisecs); | |
2147 | ||
2148 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
2149 | struct kobj_attribute *attr, char *buf) | |
2150 | { | |
2151 | return sprintf(buf, "%u\n", ksm_thread_pages_to_scan); | |
2152 | } | |
2153 | ||
2154 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
2155 | struct kobj_attribute *attr, | |
2156 | const char *buf, size_t count) | |
2157 | { | |
2158 | int err; | |
2159 | unsigned long nr_pages; | |
2160 | ||
3dbb95f7 | 2161 | err = kstrtoul(buf, 10, &nr_pages); |
31dbd01f IE |
2162 | if (err || nr_pages > UINT_MAX) |
2163 | return -EINVAL; | |
2164 | ||
2165 | ksm_thread_pages_to_scan = nr_pages; | |
2166 | ||
2167 | return count; | |
2168 | } | |
2169 | KSM_ATTR(pages_to_scan); | |
2170 | ||
2171 | static ssize_t run_show(struct kobject *kobj, struct kobj_attribute *attr, | |
2172 | char *buf) | |
2173 | { | |
ef4d43a8 | 2174 | return sprintf(buf, "%lu\n", ksm_run); |
31dbd01f IE |
2175 | } |
2176 | ||
2177 | static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr, | |
2178 | const char *buf, size_t count) | |
2179 | { | |
2180 | int err; | |
2181 | unsigned long flags; | |
2182 | ||
3dbb95f7 | 2183 | err = kstrtoul(buf, 10, &flags); |
31dbd01f IE |
2184 | if (err || flags > UINT_MAX) |
2185 | return -EINVAL; | |
2186 | if (flags > KSM_RUN_UNMERGE) | |
2187 | return -EINVAL; | |
2188 | ||
2189 | /* | |
2190 | * KSM_RUN_MERGE sets ksmd running, and 0 stops it running. | |
2191 | * KSM_RUN_UNMERGE stops it running and unmerges all rmap_items, | |
d0f209f6 HD |
2192 | * breaking COW to free the pages_shared (but leaves mm_slots |
2193 | * on the list for when ksmd may be set running again). | |
31dbd01f IE |
2194 | */ |
2195 | ||
2196 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 2197 | wait_while_offlining(); |
31dbd01f IE |
2198 | if (ksm_run != flags) { |
2199 | ksm_run = flags; | |
d952b791 | 2200 | if (flags & KSM_RUN_UNMERGE) { |
e1e12d2f | 2201 | set_current_oom_origin(); |
d952b791 | 2202 | err = unmerge_and_remove_all_rmap_items(); |
e1e12d2f | 2203 | clear_current_oom_origin(); |
d952b791 HD |
2204 | if (err) { |
2205 | ksm_run = KSM_RUN_STOP; | |
2206 | count = err; | |
2207 | } | |
2208 | } | |
31dbd01f IE |
2209 | } |
2210 | mutex_unlock(&ksm_thread_mutex); | |
2211 | ||
2212 | if (flags & KSM_RUN_MERGE) | |
2213 | wake_up_interruptible(&ksm_thread_wait); | |
2214 | ||
2215 | return count; | |
2216 | } | |
2217 | KSM_ATTR(run); | |
2218 | ||
90bd6fd3 PH |
2219 | #ifdef CONFIG_NUMA |
2220 | static ssize_t merge_across_nodes_show(struct kobject *kobj, | |
2221 | struct kobj_attribute *attr, char *buf) | |
2222 | { | |
2223 | return sprintf(buf, "%u\n", ksm_merge_across_nodes); | |
2224 | } | |
2225 | ||
2226 | static ssize_t merge_across_nodes_store(struct kobject *kobj, | |
2227 | struct kobj_attribute *attr, | |
2228 | const char *buf, size_t count) | |
2229 | { | |
2230 | int err; | |
2231 | unsigned long knob; | |
2232 | ||
2233 | err = kstrtoul(buf, 10, &knob); | |
2234 | if (err) | |
2235 | return err; | |
2236 | if (knob > 1) | |
2237 | return -EINVAL; | |
2238 | ||
2239 | mutex_lock(&ksm_thread_mutex); | |
ef4d43a8 | 2240 | wait_while_offlining(); |
90bd6fd3 | 2241 | if (ksm_merge_across_nodes != knob) { |
cbf86cfe | 2242 | if (ksm_pages_shared || remove_all_stable_nodes()) |
90bd6fd3 | 2243 | err = -EBUSY; |
ef53d16c HD |
2244 | else if (root_stable_tree == one_stable_tree) { |
2245 | struct rb_root *buf; | |
2246 | /* | |
2247 | * This is the first time that we switch away from the | |
2248 | * default of merging across nodes: must now allocate | |
2249 | * a buffer to hold as many roots as may be needed. | |
2250 | * Allocate stable and unstable together: | |
2251 | * MAXSMP NODES_SHIFT 10 will use 16kB. | |
2252 | */ | |
bafe1e14 JP |
2253 | buf = kcalloc(nr_node_ids + nr_node_ids, sizeof(*buf), |
2254 | GFP_KERNEL); | |
ef53d16c HD |
2255 | /* Let us assume that RB_ROOT is NULL is zero */ |
2256 | if (!buf) | |
2257 | err = -ENOMEM; | |
2258 | else { | |
2259 | root_stable_tree = buf; | |
2260 | root_unstable_tree = buf + nr_node_ids; | |
2261 | /* Stable tree is empty but not the unstable */ | |
2262 | root_unstable_tree[0] = one_unstable_tree[0]; | |
2263 | } | |
2264 | } | |
2265 | if (!err) { | |
90bd6fd3 | 2266 | ksm_merge_across_nodes = knob; |
ef53d16c HD |
2267 | ksm_nr_node_ids = knob ? 1 : nr_node_ids; |
2268 | } | |
90bd6fd3 PH |
2269 | } |
2270 | mutex_unlock(&ksm_thread_mutex); | |
2271 | ||
2272 | return err ? err : count; | |
2273 | } | |
2274 | KSM_ATTR(merge_across_nodes); | |
2275 | #endif | |
2276 | ||
e86c59b1 CI |
2277 | static ssize_t use_zero_pages_show(struct kobject *kobj, |
2278 | struct kobj_attribute *attr, char *buf) | |
2279 | { | |
2280 | return sprintf(buf, "%u\n", ksm_use_zero_pages); | |
2281 | } | |
2282 | static ssize_t use_zero_pages_store(struct kobject *kobj, | |
2283 | struct kobj_attribute *attr, | |
2284 | const char *buf, size_t count) | |
2285 | { | |
2286 | int err; | |
2287 | bool value; | |
2288 | ||
2289 | err = kstrtobool(buf, &value); | |
2290 | if (err) | |
2291 | return -EINVAL; | |
2292 | ||
2293 | ksm_use_zero_pages = value; | |
2294 | ||
2295 | return count; | |
2296 | } | |
2297 | KSM_ATTR(use_zero_pages); | |
2298 | ||
b4028260 HD |
2299 | static ssize_t pages_shared_show(struct kobject *kobj, |
2300 | struct kobj_attribute *attr, char *buf) | |
2301 | { | |
2302 | return sprintf(buf, "%lu\n", ksm_pages_shared); | |
2303 | } | |
2304 | KSM_ATTR_RO(pages_shared); | |
2305 | ||
2306 | static ssize_t pages_sharing_show(struct kobject *kobj, | |
2307 | struct kobj_attribute *attr, char *buf) | |
2308 | { | |
e178dfde | 2309 | return sprintf(buf, "%lu\n", ksm_pages_sharing); |
b4028260 HD |
2310 | } |
2311 | KSM_ATTR_RO(pages_sharing); | |
2312 | ||
473b0ce4 HD |
2313 | static ssize_t pages_unshared_show(struct kobject *kobj, |
2314 | struct kobj_attribute *attr, char *buf) | |
2315 | { | |
2316 | return sprintf(buf, "%lu\n", ksm_pages_unshared); | |
2317 | } | |
2318 | KSM_ATTR_RO(pages_unshared); | |
2319 | ||
2320 | static ssize_t pages_volatile_show(struct kobject *kobj, | |
2321 | struct kobj_attribute *attr, char *buf) | |
2322 | { | |
2323 | long ksm_pages_volatile; | |
2324 | ||
2325 | ksm_pages_volatile = ksm_rmap_items - ksm_pages_shared | |
2326 | - ksm_pages_sharing - ksm_pages_unshared; | |
2327 | /* | |
2328 | * It was not worth any locking to calculate that statistic, | |
2329 | * but it might therefore sometimes be negative: conceal that. | |
2330 | */ | |
2331 | if (ksm_pages_volatile < 0) | |
2332 | ksm_pages_volatile = 0; | |
2333 | return sprintf(buf, "%ld\n", ksm_pages_volatile); | |
2334 | } | |
2335 | KSM_ATTR_RO(pages_volatile); | |
2336 | ||
2337 | static ssize_t full_scans_show(struct kobject *kobj, | |
2338 | struct kobj_attribute *attr, char *buf) | |
2339 | { | |
2340 | return sprintf(buf, "%lu\n", ksm_scan.seqnr); | |
2341 | } | |
2342 | KSM_ATTR_RO(full_scans); | |
2343 | ||
31dbd01f IE |
2344 | static struct attribute *ksm_attrs[] = { |
2345 | &sleep_millisecs_attr.attr, | |
2346 | &pages_to_scan_attr.attr, | |
2347 | &run_attr.attr, | |
b4028260 HD |
2348 | &pages_shared_attr.attr, |
2349 | &pages_sharing_attr.attr, | |
473b0ce4 HD |
2350 | &pages_unshared_attr.attr, |
2351 | &pages_volatile_attr.attr, | |
2352 | &full_scans_attr.attr, | |
90bd6fd3 PH |
2353 | #ifdef CONFIG_NUMA |
2354 | &merge_across_nodes_attr.attr, | |
2355 | #endif | |
e86c59b1 | 2356 | &use_zero_pages_attr.attr, |
31dbd01f IE |
2357 | NULL, |
2358 | }; | |
2359 | ||
2360 | static struct attribute_group ksm_attr_group = { | |
2361 | .attrs = ksm_attrs, | |
2362 | .name = "ksm", | |
2363 | }; | |
2ffd8679 | 2364 | #endif /* CONFIG_SYSFS */ |
31dbd01f IE |
2365 | |
2366 | static int __init ksm_init(void) | |
2367 | { | |
2368 | struct task_struct *ksm_thread; | |
2369 | int err; | |
2370 | ||
e86c59b1 CI |
2371 | /* The correct value depends on page size and endianness */ |
2372 | zero_checksum = calc_checksum(ZERO_PAGE(0)); | |
2373 | /* Default to false for backwards compatibility */ | |
2374 | ksm_use_zero_pages = false; | |
2375 | ||
31dbd01f IE |
2376 | err = ksm_slab_init(); |
2377 | if (err) | |
2378 | goto out; | |
2379 | ||
31dbd01f IE |
2380 | ksm_thread = kthread_run(ksm_scan_thread, NULL, "ksmd"); |
2381 | if (IS_ERR(ksm_thread)) { | |
25acde31 | 2382 | pr_err("ksm: creating kthread failed\n"); |
31dbd01f | 2383 | err = PTR_ERR(ksm_thread); |
d9f8984c | 2384 | goto out_free; |
31dbd01f IE |
2385 | } |
2386 | ||
2ffd8679 | 2387 | #ifdef CONFIG_SYSFS |
31dbd01f IE |
2388 | err = sysfs_create_group(mm_kobj, &ksm_attr_group); |
2389 | if (err) { | |
25acde31 | 2390 | pr_err("ksm: register sysfs failed\n"); |
2ffd8679 | 2391 | kthread_stop(ksm_thread); |
d9f8984c | 2392 | goto out_free; |
31dbd01f | 2393 | } |
c73602ad HD |
2394 | #else |
2395 | ksm_run = KSM_RUN_MERGE; /* no way for user to start it */ | |
2396 | ||
2ffd8679 | 2397 | #endif /* CONFIG_SYSFS */ |
31dbd01f | 2398 | |
62b61f61 | 2399 | #ifdef CONFIG_MEMORY_HOTREMOVE |
ef4d43a8 | 2400 | /* There is no significance to this priority 100 */ |
62b61f61 HD |
2401 | hotplug_memory_notifier(ksm_memory_callback, 100); |
2402 | #endif | |
31dbd01f IE |
2403 | return 0; |
2404 | ||
d9f8984c | 2405 | out_free: |
31dbd01f IE |
2406 | ksm_slab_free(); |
2407 | out: | |
2408 | return err; | |
f8af4da3 | 2409 | } |
a64fb3cd | 2410 | subsys_initcall(ksm_init); |