1/*
2 * Frontswap frontend
3 *
4 * This code provides the generic "frontend" layer to call a matching
5 * "backend" driver implementation of frontswap.  See
6 * Documentation/vm/frontswap.txt for more information.
7 *
8 * Copyright (C) 2009-2012 Oracle Corp.  All rights reserved.
9 * Author: Dan Magenheimer
10 *
11 * This work is licensed under the terms of the GNU GPL, version 2.
12 */
13
14#include <linux/mman.h>
15#include <linux/swap.h>
16#include <linux/swapops.h>
17#include <linux/security.h>
18#include <linux/module.h>
19#include <linux/debugfs.h>
20#include <linux/frontswap.h>
21#include <linux/swapfile.h>
22
23/*
24 * frontswap_ops is set by frontswap_register_ops to contain the pointers
25 * to the frontswap "backend" implementation functions.
26 */
27static struct frontswap_ops *frontswap_ops __read_mostly;
28
29/*
30 * If enabled, frontswap_store will return failure even on success.  As
31 * a result, the swap subsystem will always write the page to swap, in
32 * effect converting frontswap into a writethrough cache.  In this mode,
33 * there is no direct reduction in swap writes, but a frontswap backend
34 * can unilaterally "reclaim" any pages in use with no data loss, thus
35 * providing increases control over maximum memory usage due to frontswap.
36 */
37static bool frontswap_writethrough_enabled __read_mostly;
38
39/*
40 * If enabled, the underlying tmem implementation is capable of doing
41 * exclusive gets, so frontswap_load, on a successful tmem_get must
42 * mark the page as no longer in frontswap AND mark it dirty.
43 */
44static bool frontswap_tmem_exclusive_gets_enabled __read_mostly;
45
46#ifdef CONFIG_DEBUG_FS
47/*
48 * Counters available via /sys/kernel/debug/frontswap (if debugfs is
49 * properly configured).  These are for information only so are not protected
50 * against increment races.
51 */
52static u64 frontswap_loads;
53static u64 frontswap_succ_stores;
54static u64 frontswap_failed_stores;
55static u64 frontswap_invalidates;
56
57static inline void inc_frontswap_loads(void) {
58	frontswap_loads++;
59}
60static inline void inc_frontswap_succ_stores(void) {
61	frontswap_succ_stores++;
62}
63static inline void inc_frontswap_failed_stores(void) {
64	frontswap_failed_stores++;
65}
66static inline void inc_frontswap_invalidates(void) {
67	frontswap_invalidates++;
68}
69#else
70static inline void inc_frontswap_loads(void) { }
71static inline void inc_frontswap_succ_stores(void) { }
72static inline void inc_frontswap_failed_stores(void) { }
73static inline void inc_frontswap_invalidates(void) { }
74#endif
75
76/*
77 * Due to the asynchronous nature of the backends loading potentially
78 * _after_ the swap system has been activated, we have chokepoints
79 * on all frontswap functions to not call the backend until the backend
80 * has registered.
81 *
82 * Specifically when no backend is registered (nobody called
83 * frontswap_register_ops) all calls to frontswap_init (which is done via
84 * swapon -> enable_swap_info -> frontswap_init) are registered and remembered
85 * (via the setting of need_init bitmap) but fail to create tmem_pools. When a
86 * backend registers with frontswap at some later point the previous
87 * calls to frontswap_init are executed (by iterating over the need_init
88 * bitmap) to create tmem_pools and set the respective poolids. All of that is
89 * guarded by us using atomic bit operations on the 'need_init' bitmap.
90 *
91 * This would not guards us against the user deciding to call swapoff right as
92 * we are calling the backend to initialize (so swapon is in action).
93 * Fortunatly for us, the swapon_mutex has been taked by the callee so we are
94 * OK. The other scenario where calls to frontswap_store (called via
95 * swap_writepage) is racing with frontswap_invalidate_area (called via
96 * swapoff) is again guarded by the swap subsystem.
97 *
98 * While no backend is registered all calls to frontswap_[store|load|
99 * invalidate_area|invalidate_page] are ignored or fail.
100 *
101 * The time between the backend being registered and the swap file system
102 * calling the backend (via the frontswap_* functions) is indeterminate as
103 * frontswap_ops is not atomic_t (or a value guarded by a spinlock).
104 * That is OK as we are comfortable missing some of these calls to the newly
105 * registered backend.
106 *
107 * Obviously the opposite (unloading the backend) must be done after all
108 * the frontswap_[store|load|invalidate_area|invalidate_page] start
109 * ignorning or failing the requests - at which point frontswap_ops
110 * would have to be made in some fashion atomic.
111 */
112static DECLARE_BITMAP(need_init, MAX_SWAPFILES);
113
114/*
115 * Register operations for frontswap, returning previous thus allowing
116 * detection of multiple backends and possible nesting.
117 */
118struct frontswap_ops *frontswap_register_ops(struct frontswap_ops *ops)
119{
120	struct frontswap_ops *old = frontswap_ops;
121	int i;
122
123	for (i = 0; i < MAX_SWAPFILES; i++) {
124		if (test_and_clear_bit(i, need_init)) {
125			struct swap_info_struct *sis = swap_info[i];
126			/* __frontswap_init _should_ have set it! */
127			if (!sis->frontswap_map)
128				return ERR_PTR(-EINVAL);
129			ops->init(i);
130		}
131	}
132	/*
133	 * We MUST have frontswap_ops set _after_ the frontswap_init's
134	 * have been called. Otherwise __frontswap_store might fail. Hence
135	 * the barrier to make sure compiler does not re-order us.
136	 */
137	barrier();
138	frontswap_ops = ops;
139	return old;
140}
141EXPORT_SYMBOL(frontswap_register_ops);
142
143/*
144 * Enable/disable frontswap writethrough (see above).
145 */
146void frontswap_writethrough(bool enable)
147{
148	frontswap_writethrough_enabled = enable;
149}
150EXPORT_SYMBOL(frontswap_writethrough);
151
152/*
153 * Enable/disable frontswap exclusive gets (see above).
154 */
155void frontswap_tmem_exclusive_gets(bool enable)
156{
157	frontswap_tmem_exclusive_gets_enabled = enable;
158}
159EXPORT_SYMBOL(frontswap_tmem_exclusive_gets);
160
161/*
162 * Called when a swap device is swapon'd.
163 */
164void __frontswap_init(unsigned type, unsigned long *map)
165{
166	struct swap_info_struct *sis = swap_info[type];
167
168	BUG_ON(sis == NULL);
169
170	/*
171	 * p->frontswap is a bitmap that we MUST have to figure out which page
172	 * has gone in frontswap. Without it there is no point of continuing.
173	 */
174	if (WARN_ON(!map))
175		return;
176	/*
177	 * Irregardless of whether the frontswap backend has been loaded
178	 * before this function or it will be later, we _MUST_ have the
179	 * p->frontswap set to something valid to work properly.
180	 */
181	frontswap_map_set(sis, map);
182	if (frontswap_ops)
183		frontswap_ops->init(type);
184	else {
185		BUG_ON(type >= MAX_SWAPFILES);
186		set_bit(type, need_init);
187	}
188}
189EXPORT_SYMBOL(__frontswap_init);
190
191bool __frontswap_test(struct swap_info_struct *sis,
192				pgoff_t offset)
193{
194	bool ret = false;
195
196	if (frontswap_ops && sis->frontswap_map)
197		ret = test_bit(offset, sis->frontswap_map);
198	return ret;
199}
200EXPORT_SYMBOL(__frontswap_test);
201
202static inline void __frontswap_clear(struct swap_info_struct *sis,
203				pgoff_t offset)
204{
205	clear_bit(offset, sis->frontswap_map);
206	atomic_dec(&sis->frontswap_pages);
207}
208
209/*
210 * "Store" data from a page to frontswap and associate it with the page's
211 * swaptype and offset.  Page must be locked and in the swap cache.
212 * If frontswap already contains a page with matching swaptype and
213 * offset, the frontswap implementation may either overwrite the data and
214 * return success or invalidate the page from frontswap and return failure.
215 */
216int __frontswap_store(struct page *page)
217{
218	int ret = -1, dup = 0;
219	swp_entry_t entry = { .val = page_private(page), };
220	int type = swp_type(entry);
221	struct swap_info_struct *sis = swap_info[type];
222	pgoff_t offset = swp_offset(entry);
223
224	/*
225	 * Return if no backend registed.
226	 * Don't need to inc frontswap_failed_stores here.
227	 */
228	if (!frontswap_ops)
229		return ret;
230
231	BUG_ON(!PageLocked(page));
232	BUG_ON(sis == NULL);
233	if (__frontswap_test(sis, offset))
234		dup = 1;
235	ret = frontswap_ops->store(type, offset, page);
236	if (ret == 0) {
237		set_bit(offset, sis->frontswap_map);
238		inc_frontswap_succ_stores();
239		if (!dup)
240			atomic_inc(&sis->frontswap_pages);
241	} else {
242		/*
243		  failed dup always results in automatic invalidate of
244		  the (older) page from frontswap
245		 */
246		inc_frontswap_failed_stores();
247		if (dup) {
248			__frontswap_clear(sis, offset);
249			frontswap_ops->invalidate_page(type, offset);
250		}
251	}
252	if (frontswap_writethrough_enabled)
253		/* report failure so swap also writes to swap device */
254		ret = -1;
255	return ret;
256}
257EXPORT_SYMBOL(__frontswap_store);
258
259/*
260 * "Get" data from frontswap associated with swaptype and offset that were
261 * specified when the data was put to frontswap and use it to fill the
262 * specified page with data. Page must be locked and in the swap cache.
263 */
264int __frontswap_load(struct page *page)
265{
266	int ret = -1;
267	swp_entry_t entry = { .val = page_private(page), };
268	int type = swp_type(entry);
269	struct swap_info_struct *sis = swap_info[type];
270	pgoff_t offset = swp_offset(entry);
271
272	BUG_ON(!PageLocked(page));
273	BUG_ON(sis == NULL);
274	/*
275	 * __frontswap_test() will check whether there is backend registered
276	 */
277	if (__frontswap_test(sis, offset))
278		ret = frontswap_ops->load(type, offset, page);
279	if (ret == 0) {
280		inc_frontswap_loads();
281		if (frontswap_tmem_exclusive_gets_enabled) {
282			SetPageDirty(page);
283			__frontswap_clear(sis, offset);
284		}
285	}
286	return ret;
287}
288EXPORT_SYMBOL(__frontswap_load);
289
290/*
291 * Invalidate any data from frontswap associated with the specified swaptype
292 * and offset so that a subsequent "get" will fail.
293 */
294void __frontswap_invalidate_page(unsigned type, pgoff_t offset)
295{
296	struct swap_info_struct *sis = swap_info[type];
297
298	BUG_ON(sis == NULL);
299	/*
300	 * __frontswap_test() will check whether there is backend registered
301	 */
302	if (__frontswap_test(sis, offset)) {
303		frontswap_ops->invalidate_page(type, offset);
304		__frontswap_clear(sis, offset);
305		inc_frontswap_invalidates();
306	}
307}
308EXPORT_SYMBOL(__frontswap_invalidate_page);
309
310/*
311 * Invalidate all data from frontswap associated with all offsets for the
312 * specified swaptype.
313 */
314void __frontswap_invalidate_area(unsigned type)
315{
316	struct swap_info_struct *sis = swap_info[type];
317
318	if (frontswap_ops) {
319		BUG_ON(sis == NULL);
320		if (sis->frontswap_map == NULL)
321			return;
322		frontswap_ops->invalidate_area(type);
323		atomic_set(&sis->frontswap_pages, 0);
324		bitmap_zero(sis->frontswap_map, sis->max);
325	}
326	clear_bit(type, need_init);
327}
328EXPORT_SYMBOL(__frontswap_invalidate_area);
329
330static unsigned long __frontswap_curr_pages(void)
331{
332	unsigned long totalpages = 0;
333	struct swap_info_struct *si = NULL;
334
335	assert_spin_locked(&swap_lock);
336	plist_for_each_entry(si, &swap_active_head, list)
337		totalpages += atomic_read(&si->frontswap_pages);
338	return totalpages;
339}
340
341static int __frontswap_unuse_pages(unsigned long total, unsigned long *unused,
342					int *swapid)
343{
344	int ret = -EINVAL;
345	struct swap_info_struct *si = NULL;
346	int si_frontswap_pages;
347	unsigned long total_pages_to_unuse = total;
348	unsigned long pages = 0, pages_to_unuse = 0;
349
350	assert_spin_locked(&swap_lock);
351	plist_for_each_entry(si, &swap_active_head, list) {
352		si_frontswap_pages = atomic_read(&si->frontswap_pages);
353		if (total_pages_to_unuse < si_frontswap_pages) {
354			pages = pages_to_unuse = total_pages_to_unuse;
355		} else {
356			pages = si_frontswap_pages;
357			pages_to_unuse = 0; /* unuse all */
358		}
359		/* ensure there is enough RAM to fetch pages from frontswap */
360		if (security_vm_enough_memory_mm(current->mm, pages)) {
361			ret = -ENOMEM;
362			continue;
363		}
364		vm_unacct_memory(pages);
365		*unused = pages_to_unuse;
366		*swapid = si->type;
367		ret = 0;
368		break;
369	}
370
371	return ret;
372}
373
374/*
375 * Used to check if it's necessory and feasible to unuse pages.
376 * Return 1 when nothing to do, 0 when need to shink pages,
377 * error code when there is an error.
378 */
379static int __frontswap_shrink(unsigned long target_pages,
380				unsigned long *pages_to_unuse,
381				int *type)
382{
383	unsigned long total_pages = 0, total_pages_to_unuse;
384
385	assert_spin_locked(&swap_lock);
386
387	total_pages = __frontswap_curr_pages();
388	if (total_pages <= target_pages) {
389		/* Nothing to do */
390		*pages_to_unuse = 0;
391		return 1;
392	}
393	total_pages_to_unuse = total_pages - target_pages;
394	return __frontswap_unuse_pages(total_pages_to_unuse, pages_to_unuse, type);
395}
396
397/*
398 * Frontswap, like a true swap device, may unnecessarily retain pages
399 * under certain circumstances; "shrink" frontswap is essentially a
400 * "partial swapoff" and works by calling try_to_unuse to attempt to
401 * unuse enough frontswap pages to attempt to -- subject to memory
402 * constraints -- reduce the number of pages in frontswap to the
403 * number given in the parameter target_pages.
404 */
405void frontswap_shrink(unsigned long target_pages)
406{
407	unsigned long pages_to_unuse = 0;
408	int uninitialized_var(type), ret;
409
410	/*
411	 * we don't want to hold swap_lock while doing a very
412	 * lengthy try_to_unuse, but swap_list may change
413	 * so restart scan from swap_active_head each time
414	 */
415	spin_lock(&swap_lock);
416	ret = __frontswap_shrink(target_pages, &pages_to_unuse, &type);
417	spin_unlock(&swap_lock);
418	if (ret == 0)
419		try_to_unuse(type, true, pages_to_unuse);
420	return;
421}
422EXPORT_SYMBOL(frontswap_shrink);
423
424/*
425 * Count and return the number of frontswap pages across all
426 * swap devices.  This is exported so that backend drivers can
427 * determine current usage without reading debugfs.
428 */
429unsigned long frontswap_curr_pages(void)
430{
431	unsigned long totalpages = 0;
432
433	spin_lock(&swap_lock);
434	totalpages = __frontswap_curr_pages();
435	spin_unlock(&swap_lock);
436
437	return totalpages;
438}
439EXPORT_SYMBOL(frontswap_curr_pages);
440
441static int __init init_frontswap(void)
442{
443#ifdef CONFIG_DEBUG_FS
444	struct dentry *root = debugfs_create_dir("frontswap", NULL);
445	if (root == NULL)
446		return -ENXIO;
447	debugfs_create_u64("loads", S_IRUGO, root, &frontswap_loads);
448	debugfs_create_u64("succ_stores", S_IRUGO, root, &frontswap_succ_stores);
449	debugfs_create_u64("failed_stores", S_IRUGO, root,
450				&frontswap_failed_stores);
451	debugfs_create_u64("invalidates", S_IRUGO,
452				root, &frontswap_invalidates);
453#endif
454	return 0;
455}
456
457module_init(init_frontswap);
458