1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Frontswap frontend
4  *
5  * This code provides the generic "frontend" layer to call a matching
6  * "backend" driver implementation of frontswap.  See
7  * Documentation/mm/frontswap.rst for more information.
8  *
9  * Copyright (C) 2009-2012 Oracle Corp.  All rights reserved.
10  * Author: Dan Magenheimer
11  */
12 
13 #include <linux/mman.h>
14 #include <linux/swap.h>
15 #include <linux/swapops.h>
16 #include <linux/security.h>
17 #include <linux/module.h>
18 #include <linux/debugfs.h>
19 #include <linux/frontswap.h>
20 #include <linux/swapfile.h>
21 
22 DEFINE_STATIC_KEY_FALSE(frontswap_enabled_key);
23 
24 /*
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.
29  */
30 static const struct frontswap_ops *frontswap_ops __read_mostly;
31 
32 #ifdef CONFIG_DEBUG_FS
33 /*
34  * Counters available via /sys/kernel/debug/frontswap (if debugfs is
35  * properly configured).  These are for information only so are not protected
36  * against increment races.
37  */
38 static u64 frontswap_loads;
39 static u64 frontswap_succ_stores;
40 static u64 frontswap_failed_stores;
41 static u64 frontswap_invalidates;
42 
inc_frontswap_loads(void)43 static inline void inc_frontswap_loads(void)
44 {
45 	data_race(frontswap_loads++);
46 }
inc_frontswap_succ_stores(void)47 static inline void inc_frontswap_succ_stores(void)
48 {
49 	data_race(frontswap_succ_stores++);
50 }
inc_frontswap_failed_stores(void)51 static inline void inc_frontswap_failed_stores(void)
52 {
53 	data_race(frontswap_failed_stores++);
54 }
inc_frontswap_invalidates(void)55 static inline void inc_frontswap_invalidates(void)
56 {
57 	data_race(frontswap_invalidates++);
58 }
59 #else
inc_frontswap_loads(void)60 static inline void inc_frontswap_loads(void) { }
inc_frontswap_succ_stores(void)61 static inline void inc_frontswap_succ_stores(void) { }
inc_frontswap_failed_stores(void)62 static inline void inc_frontswap_failed_stores(void) { }
inc_frontswap_invalidates(void)63 static inline void inc_frontswap_invalidates(void) { }
64 #endif
65 
66 /*
67  * Due to the asynchronous nature of the backends loading potentially
68  * _after_ the swap system has been activated, we have chokepoints
69  * on all frontswap functions to not call the backend until the backend
70  * has registered.
71  *
72  * This would not guards us against the user deciding to call swapoff right as
73  * we are calling the backend to initialize (so swapon is in action).
74  * Fortunately for us, the swapon_mutex has been taken by the callee so we are
75  * OK. The other scenario where calls to frontswap_store (called via
76  * swap_writepage) is racing with frontswap_invalidate_area (called via
77  * swapoff) is again guarded by the swap subsystem.
78  *
79  * While no backend is registered all calls to frontswap_[store|load|
80  * invalidate_area|invalidate_page] are ignored or fail.
81  *
82  * The time between the backend being registered and the swap file system
83  * calling the backend (via the frontswap_* functions) is indeterminate as
84  * frontswap_ops is not atomic_t (or a value guarded by a spinlock).
85  * That is OK as we are comfortable missing some of these calls to the newly
86  * registered backend.
87  *
88  * Obviously the opposite (unloading the backend) must be done after all
89  * the frontswap_[store|load|invalidate_area|invalidate_page] start
90  * ignoring or failing the requests.  However, there is currently no way
91  * to unload a backend once it is registered.
92  */
93 
94 /*
95  * Register operations for frontswap
96  */
frontswap_register_ops(const struct frontswap_ops * ops)97 int frontswap_register_ops(const struct frontswap_ops *ops)
98 {
99 	if (frontswap_ops)
100 		return -EINVAL;
101 
102 	frontswap_ops = ops;
103 	static_branch_inc(&frontswap_enabled_key);
104 	return 0;
105 }
106 
107 /*
108  * Called when a swap device is swapon'd.
109  */
frontswap_init(unsigned type,unsigned long * map)110 void frontswap_init(unsigned type, unsigned long *map)
111 {
112 	struct swap_info_struct *sis = swap_info[type];
113 
114 	VM_BUG_ON(sis == NULL);
115 
116 	/*
117 	 * p->frontswap is a bitmap that we MUST have to figure out which page
118 	 * has gone in frontswap. Without it there is no point of continuing.
119 	 */
120 	if (WARN_ON(!map))
121 		return;
122 	/*
123 	 * Irregardless of whether the frontswap backend has been loaded
124 	 * before this function or it will be later, we _MUST_ have the
125 	 * p->frontswap set to something valid to work properly.
126 	 */
127 	frontswap_map_set(sis, map);
128 
129 	if (!frontswap_enabled())
130 		return;
131 	frontswap_ops->init(type);
132 }
133 
__frontswap_test(struct swap_info_struct * sis,pgoff_t offset)134 static bool __frontswap_test(struct swap_info_struct *sis,
135 				pgoff_t offset)
136 {
137 	if (sis->frontswap_map)
138 		return test_bit(offset, sis->frontswap_map);
139 	return false;
140 }
141 
__frontswap_set(struct swap_info_struct * sis,pgoff_t offset)142 static inline void __frontswap_set(struct swap_info_struct *sis,
143 				   pgoff_t offset)
144 {
145 	set_bit(offset, sis->frontswap_map);
146 	atomic_inc(&sis->frontswap_pages);
147 }
148 
__frontswap_clear(struct swap_info_struct * sis,pgoff_t offset)149 static inline void __frontswap_clear(struct swap_info_struct *sis,
150 				     pgoff_t offset)
151 {
152 	clear_bit(offset, sis->frontswap_map);
153 	atomic_dec(&sis->frontswap_pages);
154 }
155 
156 /*
157  * "Store" data from a page to frontswap and associate it with the page's
158  * swaptype and offset.  Page must be locked and in the swap cache.
159  * If frontswap already contains a page with matching swaptype and
160  * offset, the frontswap implementation may either overwrite the data and
161  * return success or invalidate the page from frontswap and return failure.
162  */
__frontswap_store(struct page * page)163 int __frontswap_store(struct page *page)
164 {
165 	int ret = -1;
166 	swp_entry_t entry = { .val = page_private(page), };
167 	int type = swp_type(entry);
168 	struct swap_info_struct *sis = swap_info[type];
169 	pgoff_t offset = swp_offset(entry);
170 
171 	VM_BUG_ON(!frontswap_ops);
172 	VM_BUG_ON(!PageLocked(page));
173 	VM_BUG_ON(sis == NULL);
174 
175 	/*
176 	 * If a dup, we must remove the old page first; we can't leave the
177 	 * old page no matter if the store of the new page succeeds or fails,
178 	 * and we can't rely on the new page replacing the old page as we may
179 	 * not store to the same implementation that contains the old page.
180 	 */
181 	if (__frontswap_test(sis, offset)) {
182 		__frontswap_clear(sis, offset);
183 		frontswap_ops->invalidate_page(type, offset);
184 	}
185 
186 	ret = frontswap_ops->store(type, offset, page);
187 	if (ret == 0) {
188 		__frontswap_set(sis, offset);
189 		inc_frontswap_succ_stores();
190 	} else {
191 		inc_frontswap_failed_stores();
192 	}
193 
194 	return ret;
195 }
196 
197 /*
198  * "Get" data from frontswap associated with swaptype and offset that were
199  * specified when the data was put to frontswap and use it to fill the
200  * specified page with data. Page must be locked and in the swap cache.
201  */
__frontswap_load(struct page * page)202 int __frontswap_load(struct page *page)
203 {
204 	int ret = -1;
205 	swp_entry_t entry = { .val = page_private(page), };
206 	int type = swp_type(entry);
207 	struct swap_info_struct *sis = swap_info[type];
208 	pgoff_t offset = swp_offset(entry);
209 
210 	VM_BUG_ON(!frontswap_ops);
211 	VM_BUG_ON(!PageLocked(page));
212 	VM_BUG_ON(sis == NULL);
213 
214 	if (!__frontswap_test(sis, offset))
215 		return -1;
216 
217 	/* Try loading from each implementation, until one succeeds. */
218 	ret = frontswap_ops->load(type, offset, page);
219 	if (ret == 0)
220 		inc_frontswap_loads();
221 	return ret;
222 }
223 
224 /*
225  * Invalidate any data from frontswap associated with the specified swaptype
226  * and offset so that a subsequent "get" will fail.
227  */
__frontswap_invalidate_page(unsigned type,pgoff_t offset)228 void __frontswap_invalidate_page(unsigned type, pgoff_t offset)
229 {
230 	struct swap_info_struct *sis = swap_info[type];
231 
232 	VM_BUG_ON(!frontswap_ops);
233 	VM_BUG_ON(sis == NULL);
234 
235 	if (!__frontswap_test(sis, offset))
236 		return;
237 
238 	frontswap_ops->invalidate_page(type, offset);
239 	__frontswap_clear(sis, offset);
240 	inc_frontswap_invalidates();
241 }
242 
243 /*
244  * Invalidate all data from frontswap associated with all offsets for the
245  * specified swaptype.
246  */
__frontswap_invalidate_area(unsigned type)247 void __frontswap_invalidate_area(unsigned type)
248 {
249 	struct swap_info_struct *sis = swap_info[type];
250 
251 	VM_BUG_ON(!frontswap_ops);
252 	VM_BUG_ON(sis == NULL);
253 
254 	if (sis->frontswap_map == NULL)
255 		return;
256 
257 	frontswap_ops->invalidate_area(type);
258 	atomic_set(&sis->frontswap_pages, 0);
259 	bitmap_zero(sis->frontswap_map, sis->max);
260 }
261 
init_frontswap(void)262 static int __init init_frontswap(void)
263 {
264 #ifdef CONFIG_DEBUG_FS
265 	struct dentry *root = debugfs_create_dir("frontswap", NULL);
266 	if (root == NULL)
267 		return -ENXIO;
268 	debugfs_create_u64("loads", 0444, root, &frontswap_loads);
269 	debugfs_create_u64("succ_stores", 0444, root, &frontswap_succ_stores);
270 	debugfs_create_u64("failed_stores", 0444, root,
271 			   &frontswap_failed_stores);
272 	debugfs_create_u64("invalidates", 0444, root, &frontswap_invalidates);
273 #endif
274 	return 0;
275 }
276 
277 module_init(init_frontswap);
278