1 // SPDX-License-Identifier: GPL-2.0-only
5 * This code provides the generic "frontend" layer to call a matching
6 * "backend" driver implementation of frontswap. See
7 * Documentation/vm/frontswap.rst for more information.
9 * Copyright (C) 2009-2012 Oracle Corp. All rights reserved.
10 * Author: Dan Magenheimer
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>
22 DEFINE_STATIC_KEY_FALSE(frontswap_enabled_key);
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.
30 static struct frontswap_ops *frontswap_ops __read_mostly;
32 #define for_each_frontswap_ops(ops) \
33 for ((ops) = frontswap_ops; (ops); (ops) = (ops)->next)
36 * If enabled, frontswap_store will return failure even on success. As
37 * a result, the swap subsystem will always write the page to swap, in
38 * effect converting frontswap into a writethrough cache. In this mode,
39 * there is no direct reduction in swap writes, but a frontswap backend
40 * can unilaterally "reclaim" any pages in use with no data loss, thus
41 * providing increases control over maximum memory usage due to frontswap.
43 static bool frontswap_writethrough_enabled __read_mostly;
46 * If enabled, the underlying tmem implementation is capable of doing
47 * exclusive gets, so frontswap_load, on a successful tmem_get must
48 * mark the page as no longer in frontswap AND mark it dirty.
50 static bool frontswap_tmem_exclusive_gets_enabled __read_mostly;
52 #ifdef CONFIG_DEBUG_FS
54 * Counters available via /sys/kernel/debug/frontswap (if debugfs is
55 * properly configured). These are for information only so are not protected
56 * against increment races.
58 static u64 frontswap_loads;
59 static u64 frontswap_succ_stores;
60 static u64 frontswap_failed_stores;
61 static u64 frontswap_invalidates;
63 static inline void inc_frontswap_loads(void)
65 data_race(frontswap_loads++);
67 static inline void inc_frontswap_succ_stores(void)
69 data_race(frontswap_succ_stores++);
71 static inline void inc_frontswap_failed_stores(void)
73 data_race(frontswap_failed_stores++);
75 static inline void inc_frontswap_invalidates(void)
77 data_race(frontswap_invalidates++);
80 static inline void inc_frontswap_loads(void) { }
81 static inline void inc_frontswap_succ_stores(void) { }
82 static inline void inc_frontswap_failed_stores(void) { }
83 static inline void inc_frontswap_invalidates(void) { }
87 * Due to the asynchronous nature of the backends loading potentially
88 * _after_ the swap system has been activated, we have chokepoints
89 * on all frontswap functions to not call the backend until the backend
92 * This would not guards us against the user deciding to call swapoff right as
93 * we are calling the backend to initialize (so swapon is in action).
94 * Fortunately for us, the swapon_mutex has been taken by the callee so we are
95 * OK. The other scenario where calls to frontswap_store (called via
96 * swap_writepage) is racing with frontswap_invalidate_area (called via
97 * swapoff) is again guarded by the swap subsystem.
99 * While no backend is registered all calls to frontswap_[store|load|
100 * invalidate_area|invalidate_page] are ignored or fail.
102 * The time between the backend being registered and the swap file system
103 * calling the backend (via the frontswap_* functions) is indeterminate as
104 * frontswap_ops is not atomic_t (or a value guarded by a spinlock).
105 * That is OK as we are comfortable missing some of these calls to the newly
106 * registered backend.
108 * Obviously the opposite (unloading the backend) must be done after all
109 * the frontswap_[store|load|invalidate_area|invalidate_page] start
110 * ignoring or failing the requests. However, there is currently no way
111 * to unload a backend once it is registered.
115 * Register operations for frontswap
117 void frontswap_register_ops(struct frontswap_ops *ops)
119 DECLARE_BITMAP(a, MAX_SWAPFILES);
120 DECLARE_BITMAP(b, MAX_SWAPFILES);
121 struct swap_info_struct *si;
124 bitmap_zero(a, MAX_SWAPFILES);
125 bitmap_zero(b, MAX_SWAPFILES);
127 spin_lock(&swap_lock);
128 plist_for_each_entry(si, &swap_active_head, list) {
129 if (!WARN_ON(!si->frontswap_map))
130 set_bit(si->type, a);
132 spin_unlock(&swap_lock);
134 /* the new ops needs to know the currently active swap devices */
135 for_each_set_bit(i, a, MAX_SWAPFILES)
139 * Setting frontswap_ops must happen after the ops->init() calls
140 * above; cmpxchg implies smp_mb() which will ensure the init is
141 * complete at this point.
144 ops->next = frontswap_ops;
145 } while (cmpxchg(&frontswap_ops, ops->next, ops) != ops->next);
147 static_branch_inc(&frontswap_enabled_key);
149 spin_lock(&swap_lock);
150 plist_for_each_entry(si, &swap_active_head, list) {
151 if (si->frontswap_map)
152 set_bit(si->type, b);
154 spin_unlock(&swap_lock);
157 * On the very unlikely chance that a swap device was added or
158 * removed between setting the "a" list bits and the ops init
159 * calls, we re-check and do init or invalidate for any changed
162 if (unlikely(!bitmap_equal(a, b, MAX_SWAPFILES))) {
163 for (i = 0; i < MAX_SWAPFILES; i++) {
164 if (!test_bit(i, a) && test_bit(i, b))
166 else if (test_bit(i, a) && !test_bit(i, b))
167 ops->invalidate_area(i);
171 EXPORT_SYMBOL(frontswap_register_ops);
174 * Enable/disable frontswap writethrough (see above).
176 void frontswap_writethrough(bool enable)
178 frontswap_writethrough_enabled = enable;
180 EXPORT_SYMBOL(frontswap_writethrough);
183 * Enable/disable frontswap exclusive gets (see above).
185 void frontswap_tmem_exclusive_gets(bool enable)
187 frontswap_tmem_exclusive_gets_enabled = enable;
189 EXPORT_SYMBOL(frontswap_tmem_exclusive_gets);
192 * Called when a swap device is swapon'd.
194 void __frontswap_init(unsigned type, unsigned long *map)
196 struct swap_info_struct *sis = swap_info[type];
197 struct frontswap_ops *ops;
199 VM_BUG_ON(sis == NULL);
202 * p->frontswap is a bitmap that we MUST have to figure out which page
203 * has gone in frontswap. Without it there is no point of continuing.
208 * Irregardless of whether the frontswap backend has been loaded
209 * before this function or it will be later, we _MUST_ have the
210 * p->frontswap set to something valid to work properly.
212 frontswap_map_set(sis, map);
214 for_each_frontswap_ops(ops)
217 EXPORT_SYMBOL(__frontswap_init);
219 bool __frontswap_test(struct swap_info_struct *sis,
222 if (sis->frontswap_map)
223 return test_bit(offset, sis->frontswap_map);
226 EXPORT_SYMBOL(__frontswap_test);
228 static inline void __frontswap_set(struct swap_info_struct *sis,
231 set_bit(offset, sis->frontswap_map);
232 atomic_inc(&sis->frontswap_pages);
235 static inline void __frontswap_clear(struct swap_info_struct *sis,
238 clear_bit(offset, sis->frontswap_map);
239 atomic_dec(&sis->frontswap_pages);
243 * "Store" data from a page to frontswap and associate it with the page's
244 * swaptype and offset. Page must be locked and in the swap cache.
245 * If frontswap already contains a page with matching swaptype and
246 * offset, the frontswap implementation may either overwrite the data and
247 * return success or invalidate the page from frontswap and return failure.
249 int __frontswap_store(struct page *page)
252 swp_entry_t entry = { .val = page_private(page), };
253 int type = swp_type(entry);
254 struct swap_info_struct *sis = swap_info[type];
255 pgoff_t offset = swp_offset(entry);
256 struct frontswap_ops *ops;
258 VM_BUG_ON(!frontswap_ops);
259 VM_BUG_ON(!PageLocked(page));
260 VM_BUG_ON(sis == NULL);
263 * If a dup, we must remove the old page first; we can't leave the
264 * old page no matter if the store of the new page succeeds or fails,
265 * and we can't rely on the new page replacing the old page as we may
266 * not store to the same implementation that contains the old page.
268 if (__frontswap_test(sis, offset)) {
269 __frontswap_clear(sis, offset);
270 for_each_frontswap_ops(ops)
271 ops->invalidate_page(type, offset);
274 /* Try to store in each implementation, until one succeeds. */
275 for_each_frontswap_ops(ops) {
276 ret = ops->store(type, offset, page);
277 if (!ret) /* successful store */
281 __frontswap_set(sis, offset);
282 inc_frontswap_succ_stores();
284 inc_frontswap_failed_stores();
286 if (frontswap_writethrough_enabled)
287 /* report failure so swap also writes to swap device */
291 EXPORT_SYMBOL(__frontswap_store);
294 * "Get" data from frontswap associated with swaptype and offset that were
295 * specified when the data was put to frontswap and use it to fill the
296 * specified page with data. Page must be locked and in the swap cache.
298 int __frontswap_load(struct page *page)
301 swp_entry_t entry = { .val = page_private(page), };
302 int type = swp_type(entry);
303 struct swap_info_struct *sis = swap_info[type];
304 pgoff_t offset = swp_offset(entry);
305 struct frontswap_ops *ops;
307 VM_BUG_ON(!frontswap_ops);
308 VM_BUG_ON(!PageLocked(page));
309 VM_BUG_ON(sis == NULL);
311 if (!__frontswap_test(sis, offset))
314 /* Try loading from each implementation, until one succeeds. */
315 for_each_frontswap_ops(ops) {
316 ret = ops->load(type, offset, page);
317 if (!ret) /* successful load */
321 inc_frontswap_loads();
322 if (frontswap_tmem_exclusive_gets_enabled) {
324 __frontswap_clear(sis, offset);
329 EXPORT_SYMBOL(__frontswap_load);
332 * Invalidate any data from frontswap associated with the specified swaptype
333 * and offset so that a subsequent "get" will fail.
335 void __frontswap_invalidate_page(unsigned type, pgoff_t offset)
337 struct swap_info_struct *sis = swap_info[type];
338 struct frontswap_ops *ops;
340 VM_BUG_ON(!frontswap_ops);
341 VM_BUG_ON(sis == NULL);
343 if (!__frontswap_test(sis, offset))
346 for_each_frontswap_ops(ops)
347 ops->invalidate_page(type, offset);
348 __frontswap_clear(sis, offset);
349 inc_frontswap_invalidates();
351 EXPORT_SYMBOL(__frontswap_invalidate_page);
354 * Invalidate all data from frontswap associated with all offsets for the
355 * specified swaptype.
357 void __frontswap_invalidate_area(unsigned type)
359 struct swap_info_struct *sis = swap_info[type];
360 struct frontswap_ops *ops;
362 VM_BUG_ON(!frontswap_ops);
363 VM_BUG_ON(sis == NULL);
365 if (sis->frontswap_map == NULL)
368 for_each_frontswap_ops(ops)
369 ops->invalidate_area(type);
370 atomic_set(&sis->frontswap_pages, 0);
371 bitmap_zero(sis->frontswap_map, sis->max);
373 EXPORT_SYMBOL(__frontswap_invalidate_area);
375 static unsigned long __frontswap_curr_pages(void)
377 unsigned long totalpages = 0;
378 struct swap_info_struct *si = NULL;
380 assert_spin_locked(&swap_lock);
381 plist_for_each_entry(si, &swap_active_head, list)
382 totalpages += atomic_read(&si->frontswap_pages);
386 static int __frontswap_unuse_pages(unsigned long total, unsigned long *unused,
390 struct swap_info_struct *si = NULL;
391 int si_frontswap_pages;
392 unsigned long total_pages_to_unuse = total;
393 unsigned long pages = 0, pages_to_unuse = 0;
395 assert_spin_locked(&swap_lock);
396 plist_for_each_entry(si, &swap_active_head, list) {
397 si_frontswap_pages = atomic_read(&si->frontswap_pages);
398 if (total_pages_to_unuse < si_frontswap_pages) {
399 pages = pages_to_unuse = total_pages_to_unuse;
401 pages = si_frontswap_pages;
402 pages_to_unuse = 0; /* unuse all */
404 /* ensure there is enough RAM to fetch pages from frontswap */
405 if (security_vm_enough_memory_mm(current->mm, pages)) {
409 vm_unacct_memory(pages);
410 *unused = pages_to_unuse;
420 * Used to check if it's necessary and feasible to unuse pages.
421 * Return 1 when nothing to do, 0 when need to shrink pages,
422 * error code when there is an error.
424 static int __frontswap_shrink(unsigned long target_pages,
425 unsigned long *pages_to_unuse,
428 unsigned long total_pages = 0, total_pages_to_unuse;
430 assert_spin_locked(&swap_lock);
432 total_pages = __frontswap_curr_pages();
433 if (total_pages <= target_pages) {
438 total_pages_to_unuse = total_pages - target_pages;
439 return __frontswap_unuse_pages(total_pages_to_unuse, pages_to_unuse, type);
443 * Frontswap, like a true swap device, may unnecessarily retain pages
444 * under certain circumstances; "shrink" frontswap is essentially a
445 * "partial swapoff" and works by calling try_to_unuse to attempt to
446 * unuse enough frontswap pages to attempt to -- subject to memory
447 * constraints -- reduce the number of pages in frontswap to the
448 * number given in the parameter target_pages.
450 void frontswap_shrink(unsigned long target_pages)
452 unsigned long pages_to_unuse = 0;
456 * we don't want to hold swap_lock while doing a very
457 * lengthy try_to_unuse, but swap_list may change
458 * so restart scan from swap_active_head each time
460 spin_lock(&swap_lock);
461 ret = __frontswap_shrink(target_pages, &pages_to_unuse, &type);
462 spin_unlock(&swap_lock);
464 try_to_unuse(type, true, pages_to_unuse);
467 EXPORT_SYMBOL(frontswap_shrink);
470 * Count and return the number of frontswap pages across all
471 * swap devices. This is exported so that backend drivers can
472 * determine current usage without reading debugfs.
474 unsigned long frontswap_curr_pages(void)
476 unsigned long totalpages = 0;
478 spin_lock(&swap_lock);
479 totalpages = __frontswap_curr_pages();
480 spin_unlock(&swap_lock);
484 EXPORT_SYMBOL(frontswap_curr_pages);
486 static int __init init_frontswap(void)
488 #ifdef CONFIG_DEBUG_FS
489 struct dentry *root = debugfs_create_dir("frontswap", NULL);
492 debugfs_create_u64("loads", 0444, root, &frontswap_loads);
493 debugfs_create_u64("succ_stores", 0444, root, &frontswap_succ_stores);
494 debugfs_create_u64("failed_stores", 0444, root,
495 &frontswap_failed_stores);
496 debugfs_create_u64("invalidates", 0444, root, &frontswap_invalidates);
501 module_init(init_frontswap);