xref: /openbmc/linux/mm/swap_slots.c (revision 239480ab)
1 /*
2  * Manage cache of swap slots to be used for and returned from
3  * swap.
4  *
5  * Copyright(c) 2016 Intel Corporation.
6  *
7  * Author: Tim Chen <tim.c.chen@linux.intel.com>
8  *
9  * We allocate the swap slots from the global pool and put
10  * it into local per cpu caches.  This has the advantage
11  * of no needing to acquire the swap_info lock every time
12  * we need a new slot.
13  *
14  * There is also opportunity to simply return the slot
15  * to local caches without needing to acquire swap_info
16  * lock.  We do not reuse the returned slots directly but
17  * move them back to the global pool in a batch.  This
18  * allows the slots to coaellesce and reduce fragmentation.
19  *
20  * The swap entry allocated is marked with SWAP_HAS_CACHE
21  * flag in map_count that prevents it from being allocated
22  * again from the global pool.
23  *
24  * The swap slots cache is protected by a mutex instead of
25  * a spin lock as when we search for slots with scan_swap_map,
26  * we can possibly sleep.
27  */
28 
29 #include <linux/swap_slots.h>
30 #include <linux/cpu.h>
31 #include <linux/cpumask.h>
32 #include <linux/vmalloc.h>
33 #include <linux/mutex.h>
34 #include <linux/mm.h>
35 
36 #ifdef CONFIG_SWAP
37 
38 static DEFINE_PER_CPU(struct swap_slots_cache, swp_slots);
39 static bool	swap_slot_cache_active;
40 bool	swap_slot_cache_enabled;
41 static bool	swap_slot_cache_initialized;
42 DEFINE_MUTEX(swap_slots_cache_mutex);
43 /* Serialize swap slots cache enable/disable operations */
44 DEFINE_MUTEX(swap_slots_cache_enable_mutex);
45 
46 static void __drain_swap_slots_cache(unsigned int type);
47 static void deactivate_swap_slots_cache(void);
48 static void reactivate_swap_slots_cache(void);
49 
50 #define use_swap_slot_cache (swap_slot_cache_active && \
51 		swap_slot_cache_enabled && swap_slot_cache_initialized)
52 #define SLOTS_CACHE 0x1
53 #define SLOTS_CACHE_RET 0x2
54 
55 static void deactivate_swap_slots_cache(void)
56 {
57 	mutex_lock(&swap_slots_cache_mutex);
58 	swap_slot_cache_active = false;
59 	__drain_swap_slots_cache(SLOTS_CACHE|SLOTS_CACHE_RET);
60 	mutex_unlock(&swap_slots_cache_mutex);
61 }
62 
63 static void reactivate_swap_slots_cache(void)
64 {
65 	mutex_lock(&swap_slots_cache_mutex);
66 	swap_slot_cache_active = true;
67 	mutex_unlock(&swap_slots_cache_mutex);
68 }
69 
70 /* Must not be called with cpu hot plug lock */
71 void disable_swap_slots_cache_lock(void)
72 {
73 	mutex_lock(&swap_slots_cache_enable_mutex);
74 	swap_slot_cache_enabled = false;
75 	if (swap_slot_cache_initialized) {
76 		/* serialize with cpu hotplug operations */
77 		get_online_cpus();
78 		__drain_swap_slots_cache(SLOTS_CACHE|SLOTS_CACHE_RET);
79 		put_online_cpus();
80 	}
81 }
82 
83 static void __reenable_swap_slots_cache(void)
84 {
85 	swap_slot_cache_enabled = has_usable_swap();
86 }
87 
88 void reenable_swap_slots_cache_unlock(void)
89 {
90 	__reenable_swap_slots_cache();
91 	mutex_unlock(&swap_slots_cache_enable_mutex);
92 }
93 
94 static bool check_cache_active(void)
95 {
96 	long pages;
97 
98 	if (!swap_slot_cache_enabled || !swap_slot_cache_initialized)
99 		return false;
100 
101 	pages = get_nr_swap_pages();
102 	if (!swap_slot_cache_active) {
103 		if (pages > num_online_cpus() *
104 		    THRESHOLD_ACTIVATE_SWAP_SLOTS_CACHE)
105 			reactivate_swap_slots_cache();
106 		goto out;
107 	}
108 
109 	/* if global pool of slot caches too low, deactivate cache */
110 	if (pages < num_online_cpus() * THRESHOLD_DEACTIVATE_SWAP_SLOTS_CACHE)
111 		deactivate_swap_slots_cache();
112 out:
113 	return swap_slot_cache_active;
114 }
115 
116 static int alloc_swap_slot_cache(unsigned int cpu)
117 {
118 	struct swap_slots_cache *cache;
119 	swp_entry_t *slots, *slots_ret;
120 
121 	/*
122 	 * Do allocation outside swap_slots_cache_mutex
123 	 * as kvzalloc could trigger reclaim and get_swap_page,
124 	 * which can lock swap_slots_cache_mutex.
125 	 */
126 	slots = kvzalloc(sizeof(swp_entry_t) * SWAP_SLOTS_CACHE_SIZE,
127 			 GFP_KERNEL);
128 	if (!slots)
129 		return -ENOMEM;
130 
131 	slots_ret = kvzalloc(sizeof(swp_entry_t) * SWAP_SLOTS_CACHE_SIZE,
132 			     GFP_KERNEL);
133 	if (!slots_ret) {
134 		kvfree(slots);
135 		return -ENOMEM;
136 	}
137 
138 	mutex_lock(&swap_slots_cache_mutex);
139 	cache = &per_cpu(swp_slots, cpu);
140 	if (cache->slots || cache->slots_ret)
141 		/* cache already allocated */
142 		goto out;
143 	if (!cache->lock_initialized) {
144 		mutex_init(&cache->alloc_lock);
145 		spin_lock_init(&cache->free_lock);
146 		cache->lock_initialized = true;
147 	}
148 	cache->nr = 0;
149 	cache->cur = 0;
150 	cache->n_ret = 0;
151 	cache->slots = slots;
152 	slots = NULL;
153 	cache->slots_ret = slots_ret;
154 	slots_ret = NULL;
155 out:
156 	mutex_unlock(&swap_slots_cache_mutex);
157 	if (slots)
158 		kvfree(slots);
159 	if (slots_ret)
160 		kvfree(slots_ret);
161 	return 0;
162 }
163 
164 static void drain_slots_cache_cpu(unsigned int cpu, unsigned int type,
165 				  bool free_slots)
166 {
167 	struct swap_slots_cache *cache;
168 	swp_entry_t *slots = NULL;
169 
170 	cache = &per_cpu(swp_slots, cpu);
171 	if ((type & SLOTS_CACHE) && cache->slots) {
172 		mutex_lock(&cache->alloc_lock);
173 		swapcache_free_entries(cache->slots + cache->cur, cache->nr);
174 		cache->cur = 0;
175 		cache->nr = 0;
176 		if (free_slots && cache->slots) {
177 			kvfree(cache->slots);
178 			cache->slots = NULL;
179 		}
180 		mutex_unlock(&cache->alloc_lock);
181 	}
182 	if ((type & SLOTS_CACHE_RET) && cache->slots_ret) {
183 		spin_lock_irq(&cache->free_lock);
184 		swapcache_free_entries(cache->slots_ret, cache->n_ret);
185 		cache->n_ret = 0;
186 		if (free_slots && cache->slots_ret) {
187 			slots = cache->slots_ret;
188 			cache->slots_ret = NULL;
189 		}
190 		spin_unlock_irq(&cache->free_lock);
191 		if (slots)
192 			kvfree(slots);
193 	}
194 }
195 
196 static void __drain_swap_slots_cache(unsigned int type)
197 {
198 	unsigned int cpu;
199 
200 	/*
201 	 * This function is called during
202 	 *	1) swapoff, when we have to make sure no
203 	 *	   left over slots are in cache when we remove
204 	 *	   a swap device;
205 	 *      2) disabling of swap slot cache, when we run low
206 	 *	   on swap slots when allocating memory and need
207 	 *	   to return swap slots to global pool.
208 	 *
209 	 * We cannot acquire cpu hot plug lock here as
210 	 * this function can be invoked in the cpu
211 	 * hot plug path:
212 	 * cpu_up -> lock cpu_hotplug -> cpu hotplug state callback
213 	 *   -> memory allocation -> direct reclaim -> get_swap_page
214 	 *   -> drain_swap_slots_cache
215 	 *
216 	 * Hence the loop over current online cpu below could miss cpu that
217 	 * is being brought online but not yet marked as online.
218 	 * That is okay as we do not schedule and run anything on a
219 	 * cpu before it has been marked online. Hence, we will not
220 	 * fill any swap slots in slots cache of such cpu.
221 	 * There are no slots on such cpu that need to be drained.
222 	 */
223 	for_each_online_cpu(cpu)
224 		drain_slots_cache_cpu(cpu, type, false);
225 }
226 
227 static int free_slot_cache(unsigned int cpu)
228 {
229 	mutex_lock(&swap_slots_cache_mutex);
230 	drain_slots_cache_cpu(cpu, SLOTS_CACHE | SLOTS_CACHE_RET, true);
231 	mutex_unlock(&swap_slots_cache_mutex);
232 	return 0;
233 }
234 
235 int enable_swap_slots_cache(void)
236 {
237 	int ret = 0;
238 
239 	mutex_lock(&swap_slots_cache_enable_mutex);
240 	if (swap_slot_cache_initialized) {
241 		__reenable_swap_slots_cache();
242 		goto out_unlock;
243 	}
244 
245 	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "swap_slots_cache",
246 				alloc_swap_slot_cache, free_slot_cache);
247 	if (WARN_ONCE(ret < 0, "Cache allocation failed (%s), operating "
248 			       "without swap slots cache.\n", __func__))
249 		goto out_unlock;
250 
251 	swap_slot_cache_initialized = true;
252 	__reenable_swap_slots_cache();
253 out_unlock:
254 	mutex_unlock(&swap_slots_cache_enable_mutex);
255 	return 0;
256 }
257 
258 /* called with swap slot cache's alloc lock held */
259 static int refill_swap_slots_cache(struct swap_slots_cache *cache)
260 {
261 	if (!use_swap_slot_cache || cache->nr)
262 		return 0;
263 
264 	cache->cur = 0;
265 	if (swap_slot_cache_active)
266 		cache->nr = get_swap_pages(SWAP_SLOTS_CACHE_SIZE, cache->slots);
267 
268 	return cache->nr;
269 }
270 
271 int free_swap_slot(swp_entry_t entry)
272 {
273 	struct swap_slots_cache *cache;
274 
275 	cache = &get_cpu_var(swp_slots);
276 	if (use_swap_slot_cache && cache->slots_ret) {
277 		spin_lock_irq(&cache->free_lock);
278 		/* Swap slots cache may be deactivated before acquiring lock */
279 		if (!use_swap_slot_cache) {
280 			spin_unlock_irq(&cache->free_lock);
281 			goto direct_free;
282 		}
283 		if (cache->n_ret >= SWAP_SLOTS_CACHE_SIZE) {
284 			/*
285 			 * Return slots to global pool.
286 			 * The current swap_map value is SWAP_HAS_CACHE.
287 			 * Set it to 0 to indicate it is available for
288 			 * allocation in global pool
289 			 */
290 			swapcache_free_entries(cache->slots_ret, cache->n_ret);
291 			cache->n_ret = 0;
292 		}
293 		cache->slots_ret[cache->n_ret++] = entry;
294 		spin_unlock_irq(&cache->free_lock);
295 	} else {
296 direct_free:
297 		swapcache_free_entries(&entry, 1);
298 	}
299 	put_cpu_var(swp_slots);
300 
301 	return 0;
302 }
303 
304 swp_entry_t get_swap_page(void)
305 {
306 	swp_entry_t entry, *pentry;
307 	struct swap_slots_cache *cache;
308 
309 	/*
310 	 * Preemption is allowed here, because we may sleep
311 	 * in refill_swap_slots_cache().  But it is safe, because
312 	 * accesses to the per-CPU data structure are protected by the
313 	 * mutex cache->alloc_lock.
314 	 *
315 	 * The alloc path here does not touch cache->slots_ret
316 	 * so cache->free_lock is not taken.
317 	 */
318 	cache = raw_cpu_ptr(&swp_slots);
319 
320 	entry.val = 0;
321 	if (check_cache_active()) {
322 		mutex_lock(&cache->alloc_lock);
323 		if (cache->slots) {
324 repeat:
325 			if (cache->nr) {
326 				pentry = &cache->slots[cache->cur++];
327 				entry = *pentry;
328 				pentry->val = 0;
329 				cache->nr--;
330 			} else {
331 				if (refill_swap_slots_cache(cache))
332 					goto repeat;
333 			}
334 		}
335 		mutex_unlock(&cache->alloc_lock);
336 		if (entry.val)
337 			return entry;
338 	}
339 
340 	get_swap_pages(1, &entry);
341 
342 	return entry;
343 }
344 
345 #endif /* CONFIG_SWAP */
346