Commit | Line | Data |
---|---|---|
7a338472 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
29f233cf DM |
2 | /* |
3 | * Frontswap frontend | |
4 | * | |
5 | * This code provides the generic "frontend" layer to call a matching | |
6 | * "backend" driver implementation of frontswap. See | |
ad56b738 | 7 | * Documentation/vm/frontswap.rst for more information. |
29f233cf DM |
8 | * |
9 | * Copyright (C) 2009-2012 Oracle Corp. All rights reserved. | |
10 | * Author: Dan Magenheimer | |
29f233cf DM |
11 | */ |
12 | ||
29f233cf DM |
13 | #include <linux/mman.h> |
14 | #include <linux/swap.h> | |
15 | #include <linux/swapops.h> | |
29f233cf | 16 | #include <linux/security.h> |
29f233cf | 17 | #include <linux/module.h> |
29f233cf DM |
18 | #include <linux/debugfs.h> |
19 | #include <linux/frontswap.h> | |
20 | #include <linux/swapfile.h> | |
21 | ||
8ea1d2a1 VB |
22 | DEFINE_STATIC_KEY_FALSE(frontswap_enabled_key); |
23 | ||
29f233cf | 24 | /* |
d1dc6f1b DS |
25 | * frontswap_ops are added by frontswap_register_ops, and provide the |
26 | * frontswap "backend" implementation functions. Multiple implementations | |
27 | * may be registered, but implementations can never deregister. This | |
28 | * is a simple singly-linked list of all registered implementations. | |
29f233cf | 29 | */ |
1e01c968 | 30 | static struct frontswap_ops *frontswap_ops __read_mostly; |
29f233cf | 31 | |
d1dc6f1b DS |
32 | #define for_each_frontswap_ops(ops) \ |
33 | for ((ops) = frontswap_ops; (ops); (ops) = (ops)->next) | |
34 | ||
29f233cf DM |
35 | #ifdef CONFIG_DEBUG_FS |
36 | /* | |
37 | * Counters available via /sys/kernel/debug/frontswap (if debugfs is | |
38 | * properly configured). These are for information only so are not protected | |
39 | * against increment races. | |
40 | */ | |
165c8aed KRW |
41 | static u64 frontswap_loads; |
42 | static u64 frontswap_succ_stores; | |
43 | static u64 frontswap_failed_stores; | |
29f233cf DM |
44 | static u64 frontswap_invalidates; |
45 | ||
68d68ff6 ZD |
46 | static inline void inc_frontswap_loads(void) |
47 | { | |
96bdd2bc | 48 | data_race(frontswap_loads++); |
29f233cf | 49 | } |
68d68ff6 ZD |
50 | static inline void inc_frontswap_succ_stores(void) |
51 | { | |
96bdd2bc | 52 | data_race(frontswap_succ_stores++); |
29f233cf | 53 | } |
68d68ff6 ZD |
54 | static inline void inc_frontswap_failed_stores(void) |
55 | { | |
96bdd2bc | 56 | data_race(frontswap_failed_stores++); |
29f233cf | 57 | } |
68d68ff6 ZD |
58 | static inline void inc_frontswap_invalidates(void) |
59 | { | |
96bdd2bc | 60 | data_race(frontswap_invalidates++); |
29f233cf DM |
61 | } |
62 | #else | |
165c8aed KRW |
63 | static inline void inc_frontswap_loads(void) { } |
64 | static inline void inc_frontswap_succ_stores(void) { } | |
65 | static inline void inc_frontswap_failed_stores(void) { } | |
29f233cf DM |
66 | static inline void inc_frontswap_invalidates(void) { } |
67 | #endif | |
905cd0e1 DM |
68 | |
69 | /* | |
70 | * Due to the asynchronous nature of the backends loading potentially | |
71 | * _after_ the swap system has been activated, we have chokepoints | |
72 | * on all frontswap functions to not call the backend until the backend | |
73 | * has registered. | |
74 | * | |
905cd0e1 DM |
75 | * This would not guards us against the user deciding to call swapoff right as |
76 | * we are calling the backend to initialize (so swapon is in action). | |
404f3ecf | 77 | * Fortunately for us, the swapon_mutex has been taken by the callee so we are |
905cd0e1 DM |
78 | * OK. The other scenario where calls to frontswap_store (called via |
79 | * swap_writepage) is racing with frontswap_invalidate_area (called via | |
80 | * swapoff) is again guarded by the swap subsystem. | |
81 | * | |
82 | * While no backend is registered all calls to frontswap_[store|load| | |
83 | * invalidate_area|invalidate_page] are ignored or fail. | |
84 | * | |
85 | * The time between the backend being registered and the swap file system | |
86 | * calling the backend (via the frontswap_* functions) is indeterminate as | |
1e01c968 | 87 | * frontswap_ops is not atomic_t (or a value guarded by a spinlock). |
905cd0e1 DM |
88 | * That is OK as we are comfortable missing some of these calls to the newly |
89 | * registered backend. | |
90 | * | |
91 | * Obviously the opposite (unloading the backend) must be done after all | |
92 | * the frontswap_[store|load|invalidate_area|invalidate_page] start | |
d1dc6f1b DS |
93 | * ignoring or failing the requests. However, there is currently no way |
94 | * to unload a backend once it is registered. | |
905cd0e1 | 95 | */ |
905cd0e1 | 96 | |
29f233cf | 97 | /* |
d1dc6f1b | 98 | * Register operations for frontswap |
29f233cf | 99 | */ |
d1dc6f1b | 100 | void frontswap_register_ops(struct frontswap_ops *ops) |
29f233cf | 101 | { |
d1dc6f1b DS |
102 | DECLARE_BITMAP(a, MAX_SWAPFILES); |
103 | DECLARE_BITMAP(b, MAX_SWAPFILES); | |
104 | struct swap_info_struct *si; | |
105 | unsigned int i; | |
106 | ||
107 | bitmap_zero(a, MAX_SWAPFILES); | |
108 | bitmap_zero(b, MAX_SWAPFILES); | |
109 | ||
110 | spin_lock(&swap_lock); | |
111 | plist_for_each_entry(si, &swap_active_head, list) { | |
112 | if (!WARN_ON(!si->frontswap_map)) | |
3795f46b | 113 | __set_bit(si->type, a); |
d1dc6f1b DS |
114 | } |
115 | spin_unlock(&swap_lock); | |
116 | ||
117 | /* the new ops needs to know the currently active swap devices */ | |
118 | for_each_set_bit(i, a, MAX_SWAPFILES) | |
119 | ops->init(i); | |
120 | ||
121 | /* | |
122 | * Setting frontswap_ops must happen after the ops->init() calls | |
123 | * above; cmpxchg implies smp_mb() which will ensure the init is | |
124 | * complete at this point. | |
125 | */ | |
126 | do { | |
127 | ops->next = frontswap_ops; | |
128 | } while (cmpxchg(&frontswap_ops, ops->next, ops) != ops->next); | |
129 | ||
8ea1d2a1 VB |
130 | static_branch_inc(&frontswap_enabled_key); |
131 | ||
d1dc6f1b DS |
132 | spin_lock(&swap_lock); |
133 | plist_for_each_entry(si, &swap_active_head, list) { | |
134 | if (si->frontswap_map) | |
3795f46b | 135 | __set_bit(si->type, b); |
905cd0e1 | 136 | } |
d1dc6f1b DS |
137 | spin_unlock(&swap_lock); |
138 | ||
905cd0e1 | 139 | /* |
d1dc6f1b DS |
140 | * On the very unlikely chance that a swap device was added or |
141 | * removed between setting the "a" list bits and the ops init | |
142 | * calls, we re-check and do init or invalidate for any changed | |
143 | * bits. | |
905cd0e1 | 144 | */ |
d1dc6f1b DS |
145 | if (unlikely(!bitmap_equal(a, b, MAX_SWAPFILES))) { |
146 | for (i = 0; i < MAX_SWAPFILES; i++) { | |
147 | if (!test_bit(i, a) && test_bit(i, b)) | |
148 | ops->init(i); | |
149 | else if (test_bit(i, a) && !test_bit(i, b)) | |
150 | ops->invalidate_area(i); | |
151 | } | |
152 | } | |
29f233cf | 153 | } |
29f233cf | 154 | |
29f233cf DM |
155 | /* |
156 | * Called when a swap device is swapon'd. | |
157 | */ | |
1cf53c89 | 158 | void frontswap_init(unsigned type, unsigned long *map) |
29f233cf DM |
159 | { |
160 | struct swap_info_struct *sis = swap_info[type]; | |
d1dc6f1b | 161 | struct frontswap_ops *ops; |
29f233cf | 162 | |
8ea1d2a1 | 163 | VM_BUG_ON(sis == NULL); |
4f89849d MK |
164 | |
165 | /* | |
166 | * p->frontswap is a bitmap that we MUST have to figure out which page | |
167 | * has gone in frontswap. Without it there is no point of continuing. | |
168 | */ | |
169 | if (WARN_ON(!map)) | |
170 | return; | |
171 | /* | |
172 | * Irregardless of whether the frontswap backend has been loaded | |
173 | * before this function or it will be later, we _MUST_ have the | |
174 | * p->frontswap set to something valid to work properly. | |
175 | */ | |
176 | frontswap_map_set(sis, map); | |
d1dc6f1b DS |
177 | |
178 | for_each_frontswap_ops(ops) | |
179 | ops->init(type); | |
29f233cf | 180 | } |
29f233cf | 181 | |
bd9cd521 | 182 | static bool __frontswap_test(struct swap_info_struct *sis, |
f066ea23 BL |
183 | pgoff_t offset) |
184 | { | |
d1dc6f1b DS |
185 | if (sis->frontswap_map) |
186 | return test_bit(offset, sis->frontswap_map); | |
187 | return false; | |
f066ea23 | 188 | } |
f066ea23 | 189 | |
d1dc6f1b DS |
190 | static inline void __frontswap_set(struct swap_info_struct *sis, |
191 | pgoff_t offset) | |
192 | { | |
193 | set_bit(offset, sis->frontswap_map); | |
194 | atomic_inc(&sis->frontswap_pages); | |
195 | } | |
196 | ||
f066ea23 | 197 | static inline void __frontswap_clear(struct swap_info_struct *sis, |
d1dc6f1b | 198 | pgoff_t offset) |
611edfed | 199 | { |
f066ea23 | 200 | clear_bit(offset, sis->frontswap_map); |
611edfed SL |
201 | atomic_dec(&sis->frontswap_pages); |
202 | } | |
203 | ||
29f233cf | 204 | /* |
165c8aed | 205 | * "Store" data from a page to frontswap and associate it with the page's |
29f233cf DM |
206 | * swaptype and offset. Page must be locked and in the swap cache. |
207 | * If frontswap already contains a page with matching swaptype and | |
1d00015e | 208 | * offset, the frontswap implementation may either overwrite the data and |
29f233cf DM |
209 | * return success or invalidate the page from frontswap and return failure. |
210 | */ | |
165c8aed | 211 | int __frontswap_store(struct page *page) |
29f233cf | 212 | { |
d1dc6f1b | 213 | int ret = -1; |
29f233cf DM |
214 | swp_entry_t entry = { .val = page_private(page), }; |
215 | int type = swp_type(entry); | |
216 | struct swap_info_struct *sis = swap_info[type]; | |
217 | pgoff_t offset = swp_offset(entry); | |
d1dc6f1b | 218 | struct frontswap_ops *ops; |
29f233cf | 219 | |
8ea1d2a1 VB |
220 | VM_BUG_ON(!frontswap_ops); |
221 | VM_BUG_ON(!PageLocked(page)); | |
222 | VM_BUG_ON(sis == NULL); | |
d1dc6f1b DS |
223 | |
224 | /* | |
225 | * If a dup, we must remove the old page first; we can't leave the | |
226 | * old page no matter if the store of the new page succeeds or fails, | |
227 | * and we can't rely on the new page replacing the old page as we may | |
228 | * not store to the same implementation that contains the old page. | |
229 | */ | |
230 | if (__frontswap_test(sis, offset)) { | |
231 | __frontswap_clear(sis, offset); | |
232 | for_each_frontswap_ops(ops) | |
233 | ops->invalidate_page(type, offset); | |
234 | } | |
235 | ||
236 | /* Try to store in each implementation, until one succeeds. */ | |
237 | for_each_frontswap_ops(ops) { | |
238 | ret = ops->store(type, offset, page); | |
239 | if (!ret) /* successful store */ | |
240 | break; | |
241 | } | |
29f233cf | 242 | if (ret == 0) { |
d1dc6f1b | 243 | __frontswap_set(sis, offset); |
165c8aed | 244 | inc_frontswap_succ_stores(); |
d9674dda | 245 | } else { |
165c8aed | 246 | inc_frontswap_failed_stores(); |
4bb3e31e | 247 | } |
3d6035f1 | 248 | |
29f233cf DM |
249 | return ret; |
250 | } | |
29f233cf DM |
251 | |
252 | /* | |
253 | * "Get" data from frontswap associated with swaptype and offset that were | |
254 | * specified when the data was put to frontswap and use it to fill the | |
255 | * specified page with data. Page must be locked and in the swap cache. | |
256 | */ | |
165c8aed | 257 | int __frontswap_load(struct page *page) |
29f233cf DM |
258 | { |
259 | int ret = -1; | |
260 | swp_entry_t entry = { .val = page_private(page), }; | |
261 | int type = swp_type(entry); | |
262 | struct swap_info_struct *sis = swap_info[type]; | |
263 | pgoff_t offset = swp_offset(entry); | |
d1dc6f1b DS |
264 | struct frontswap_ops *ops; |
265 | ||
8ea1d2a1 VB |
266 | VM_BUG_ON(!frontswap_ops); |
267 | VM_BUG_ON(!PageLocked(page)); | |
268 | VM_BUG_ON(sis == NULL); | |
29f233cf | 269 | |
d1dc6f1b DS |
270 | if (!__frontswap_test(sis, offset)) |
271 | return -1; | |
272 | ||
273 | /* Try loading from each implementation, until one succeeds. */ | |
274 | for_each_frontswap_ops(ops) { | |
275 | ret = ops->load(type, offset, page); | |
276 | if (!ret) /* successful load */ | |
277 | break; | |
278 | } | |
71024cb4 | 279 | if (ret == 0) |
165c8aed | 280 | inc_frontswap_loads(); |
29f233cf DM |
281 | return ret; |
282 | } | |
29f233cf DM |
283 | |
284 | /* | |
285 | * Invalidate any data from frontswap associated with the specified swaptype | |
286 | * and offset so that a subsequent "get" will fail. | |
287 | */ | |
288 | void __frontswap_invalidate_page(unsigned type, pgoff_t offset) | |
289 | { | |
290 | struct swap_info_struct *sis = swap_info[type]; | |
d1dc6f1b DS |
291 | struct frontswap_ops *ops; |
292 | ||
8ea1d2a1 VB |
293 | VM_BUG_ON(!frontswap_ops); |
294 | VM_BUG_ON(sis == NULL); | |
29f233cf | 295 | |
d1dc6f1b DS |
296 | if (!__frontswap_test(sis, offset)) |
297 | return; | |
298 | ||
299 | for_each_frontswap_ops(ops) | |
300 | ops->invalidate_page(type, offset); | |
301 | __frontswap_clear(sis, offset); | |
302 | inc_frontswap_invalidates(); | |
29f233cf | 303 | } |
29f233cf DM |
304 | |
305 | /* | |
306 | * Invalidate all data from frontswap associated with all offsets for the | |
307 | * specified swaptype. | |
308 | */ | |
309 | void __frontswap_invalidate_area(unsigned type) | |
310 | { | |
311 | struct swap_info_struct *sis = swap_info[type]; | |
d1dc6f1b | 312 | struct frontswap_ops *ops; |
29f233cf | 313 | |
8ea1d2a1 VB |
314 | VM_BUG_ON(!frontswap_ops); |
315 | VM_BUG_ON(sis == NULL); | |
d1dc6f1b | 316 | |
d1dc6f1b DS |
317 | if (sis->frontswap_map == NULL) |
318 | return; | |
319 | ||
320 | for_each_frontswap_ops(ops) | |
321 | ops->invalidate_area(type); | |
322 | atomic_set(&sis->frontswap_pages, 0); | |
323 | bitmap_zero(sis->frontswap_map, sis->max); | |
29f233cf | 324 | } |
29f233cf | 325 | |
29f233cf DM |
326 | static int __init init_frontswap(void) |
327 | { | |
328 | #ifdef CONFIG_DEBUG_FS | |
329 | struct dentry *root = debugfs_create_dir("frontswap", NULL); | |
330 | if (root == NULL) | |
331 | return -ENXIO; | |
0825a6f9 JP |
332 | debugfs_create_u64("loads", 0444, root, &frontswap_loads); |
333 | debugfs_create_u64("succ_stores", 0444, root, &frontswap_succ_stores); | |
334 | debugfs_create_u64("failed_stores", 0444, root, | |
335 | &frontswap_failed_stores); | |
336 | debugfs_create_u64("invalidates", 0444, root, &frontswap_invalidates); | |
29f233cf DM |
337 | #endif |
338 | return 0; | |
339 | } | |
340 | ||
341 | module_init(init_frontswap); |