Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
67afa38e TC |
2 | /* |
3 | * Manage cache of swap slots to be used for and returned from | |
4 | * swap. | |
5 | * | |
6 | * Copyright(c) 2016 Intel Corporation. | |
7 | * | |
8 | * Author: Tim Chen <tim.c.chen@linux.intel.com> | |
9 | * | |
10 | * We allocate the swap slots from the global pool and put | |
11 | * it into local per cpu caches. This has the advantage | |
12 | * of no needing to acquire the swap_info lock every time | |
13 | * we need a new slot. | |
14 | * | |
15 | * There is also opportunity to simply return the slot | |
16 | * to local caches without needing to acquire swap_info | |
17 | * lock. We do not reuse the returned slots directly but | |
18 | * move them back to the global pool in a batch. This | |
19 | * allows the slots to coaellesce and reduce fragmentation. | |
20 | * | |
21 | * The swap entry allocated is marked with SWAP_HAS_CACHE | |
22 | * flag in map_count that prevents it from being allocated | |
23 | * again from the global pool. | |
24 | * | |
25 | * The swap slots cache is protected by a mutex instead of | |
26 | * a spin lock as when we search for slots with scan_swap_map, | |
27 | * we can possibly sleep. | |
28 | */ | |
29 | ||
30 | #include <linux/swap_slots.h> | |
31 | #include <linux/cpu.h> | |
32 | #include <linux/cpumask.h> | |
33 | #include <linux/vmalloc.h> | |
34 | #include <linux/mutex.h> | |
54f180d3 | 35 | #include <linux/mm.h> |
67afa38e | 36 | |
67afa38e TC |
37 | static DEFINE_PER_CPU(struct swap_slots_cache, swp_slots); |
38 | static bool swap_slot_cache_active; | |
ba81f838 | 39 | bool swap_slot_cache_enabled; |
67afa38e | 40 | static bool swap_slot_cache_initialized; |
31f21da1 | 41 | static DEFINE_MUTEX(swap_slots_cache_mutex); |
67afa38e | 42 | /* Serialize swap slots cache enable/disable operations */ |
31f21da1 | 43 | static DEFINE_MUTEX(swap_slots_cache_enable_mutex); |
67afa38e TC |
44 | |
45 | static void __drain_swap_slots_cache(unsigned int type); | |
46 | static void deactivate_swap_slots_cache(void); | |
47 | static void reactivate_swap_slots_cache(void); | |
48 | ||
49 | #define use_swap_slot_cache (swap_slot_cache_active && \ | |
50 | swap_slot_cache_enabled && swap_slot_cache_initialized) | |
51 | #define SLOTS_CACHE 0x1 | |
52 | #define SLOTS_CACHE_RET 0x2 | |
53 | ||
54 | static void deactivate_swap_slots_cache(void) | |
55 | { | |
56 | mutex_lock(&swap_slots_cache_mutex); | |
57 | swap_slot_cache_active = false; | |
58 | __drain_swap_slots_cache(SLOTS_CACHE|SLOTS_CACHE_RET); | |
59 | mutex_unlock(&swap_slots_cache_mutex); | |
60 | } | |
61 | ||
62 | static void reactivate_swap_slots_cache(void) | |
63 | { | |
64 | mutex_lock(&swap_slots_cache_mutex); | |
65 | swap_slot_cache_active = true; | |
66 | mutex_unlock(&swap_slots_cache_mutex); | |
67 | } | |
68 | ||
69 | /* Must not be called with cpu hot plug lock */ | |
70 | void disable_swap_slots_cache_lock(void) | |
71 | { | |
72 | mutex_lock(&swap_slots_cache_enable_mutex); | |
73 | swap_slot_cache_enabled = false; | |
74 | if (swap_slot_cache_initialized) { | |
75 | /* serialize with cpu hotplug operations */ | |
76 | get_online_cpus(); | |
77 | __drain_swap_slots_cache(SLOTS_CACHE|SLOTS_CACHE_RET); | |
78 | put_online_cpus(); | |
79 | } | |
80 | } | |
81 | ||
82 | static void __reenable_swap_slots_cache(void) | |
83 | { | |
84 | swap_slot_cache_enabled = has_usable_swap(); | |
85 | } | |
86 | ||
87 | void reenable_swap_slots_cache_unlock(void) | |
88 | { | |
89 | __reenable_swap_slots_cache(); | |
90 | mutex_unlock(&swap_slots_cache_enable_mutex); | |
91 | } | |
92 | ||
93 | static bool check_cache_active(void) | |
94 | { | |
95 | long pages; | |
96 | ||
97 | if (!swap_slot_cache_enabled || !swap_slot_cache_initialized) | |
98 | return false; | |
99 | ||
100 | pages = get_nr_swap_pages(); | |
101 | if (!swap_slot_cache_active) { | |
102 | if (pages > num_online_cpus() * | |
103 | THRESHOLD_ACTIVATE_SWAP_SLOTS_CACHE) | |
104 | reactivate_swap_slots_cache(); | |
105 | goto out; | |
106 | } | |
107 | ||
108 | /* if global pool of slot caches too low, deactivate cache */ | |
109 | if (pages < num_online_cpus() * THRESHOLD_DEACTIVATE_SWAP_SLOTS_CACHE) | |
110 | deactivate_swap_slots_cache(); | |
111 | out: | |
112 | return swap_slot_cache_active; | |
113 | } | |
114 | ||
115 | static int alloc_swap_slot_cache(unsigned int cpu) | |
116 | { | |
117 | struct swap_slots_cache *cache; | |
118 | swp_entry_t *slots, *slots_ret; | |
119 | ||
120 | /* | |
121 | * Do allocation outside swap_slots_cache_mutex | |
54f180d3 | 122 | * as kvzalloc could trigger reclaim and get_swap_page, |
67afa38e TC |
123 | * which can lock swap_slots_cache_mutex. |
124 | */ | |
778e1cdd | 125 | slots = kvcalloc(SWAP_SLOTS_CACHE_SIZE, sizeof(swp_entry_t), |
54f180d3 | 126 | GFP_KERNEL); |
67afa38e TC |
127 | if (!slots) |
128 | return -ENOMEM; | |
129 | ||
778e1cdd | 130 | slots_ret = kvcalloc(SWAP_SLOTS_CACHE_SIZE, sizeof(swp_entry_t), |
54f180d3 | 131 | GFP_KERNEL); |
67afa38e | 132 | if (!slots_ret) { |
54f180d3 | 133 | kvfree(slots); |
67afa38e TC |
134 | return -ENOMEM; |
135 | } | |
136 | ||
137 | mutex_lock(&swap_slots_cache_mutex); | |
138 | cache = &per_cpu(swp_slots, cpu); | |
139 | if (cache->slots || cache->slots_ret) | |
140 | /* cache already allocated */ | |
141 | goto out; | |
142 | if (!cache->lock_initialized) { | |
143 | mutex_init(&cache->alloc_lock); | |
144 | spin_lock_init(&cache->free_lock); | |
145 | cache->lock_initialized = true; | |
146 | } | |
147 | cache->nr = 0; | |
148 | cache->cur = 0; | |
149 | cache->n_ret = 0; | |
a2e16731 TC |
150 | /* |
151 | * We initialized alloc_lock and free_lock earlier. We use | |
152 | * !cache->slots or !cache->slots_ret to know if it is safe to acquire | |
153 | * the corresponding lock and use the cache. Memory barrier below | |
154 | * ensures the assumption. | |
155 | */ | |
156 | mb(); | |
67afa38e TC |
157 | cache->slots = slots; |
158 | slots = NULL; | |
159 | cache->slots_ret = slots_ret; | |
160 | slots_ret = NULL; | |
161 | out: | |
162 | mutex_unlock(&swap_slots_cache_mutex); | |
163 | if (slots) | |
54f180d3 | 164 | kvfree(slots); |
67afa38e | 165 | if (slots_ret) |
54f180d3 | 166 | kvfree(slots_ret); |
67afa38e TC |
167 | return 0; |
168 | } | |
169 | ||
170 | static void drain_slots_cache_cpu(unsigned int cpu, unsigned int type, | |
171 | bool free_slots) | |
172 | { | |
173 | struct swap_slots_cache *cache; | |
174 | swp_entry_t *slots = NULL; | |
175 | ||
176 | cache = &per_cpu(swp_slots, cpu); | |
177 | if ((type & SLOTS_CACHE) && cache->slots) { | |
178 | mutex_lock(&cache->alloc_lock); | |
179 | swapcache_free_entries(cache->slots + cache->cur, cache->nr); | |
180 | cache->cur = 0; | |
181 | cache->nr = 0; | |
182 | if (free_slots && cache->slots) { | |
54f180d3 | 183 | kvfree(cache->slots); |
67afa38e TC |
184 | cache->slots = NULL; |
185 | } | |
186 | mutex_unlock(&cache->alloc_lock); | |
187 | } | |
188 | if ((type & SLOTS_CACHE_RET) && cache->slots_ret) { | |
189 | spin_lock_irq(&cache->free_lock); | |
190 | swapcache_free_entries(cache->slots_ret, cache->n_ret); | |
191 | cache->n_ret = 0; | |
192 | if (free_slots && cache->slots_ret) { | |
193 | slots = cache->slots_ret; | |
194 | cache->slots_ret = NULL; | |
195 | } | |
196 | spin_unlock_irq(&cache->free_lock); | |
197 | if (slots) | |
54f180d3 | 198 | kvfree(slots); |
67afa38e TC |
199 | } |
200 | } | |
201 | ||
202 | static void __drain_swap_slots_cache(unsigned int type) | |
203 | { | |
204 | unsigned int cpu; | |
205 | ||
206 | /* | |
207 | * This function is called during | |
208 | * 1) swapoff, when we have to make sure no | |
209 | * left over slots are in cache when we remove | |
210 | * a swap device; | |
211 | * 2) disabling of swap slot cache, when we run low | |
212 | * on swap slots when allocating memory and need | |
213 | * to return swap slots to global pool. | |
214 | * | |
215 | * We cannot acquire cpu hot plug lock here as | |
216 | * this function can be invoked in the cpu | |
217 | * hot plug path: | |
218 | * cpu_up -> lock cpu_hotplug -> cpu hotplug state callback | |
219 | * -> memory allocation -> direct reclaim -> get_swap_page | |
220 | * -> drain_swap_slots_cache | |
221 | * | |
222 | * Hence the loop over current online cpu below could miss cpu that | |
223 | * is being brought online but not yet marked as online. | |
224 | * That is okay as we do not schedule and run anything on a | |
225 | * cpu before it has been marked online. Hence, we will not | |
226 | * fill any swap slots in slots cache of such cpu. | |
227 | * There are no slots on such cpu that need to be drained. | |
228 | */ | |
229 | for_each_online_cpu(cpu) | |
230 | drain_slots_cache_cpu(cpu, type, false); | |
231 | } | |
232 | ||
233 | static int free_slot_cache(unsigned int cpu) | |
234 | { | |
235 | mutex_lock(&swap_slots_cache_mutex); | |
236 | drain_slots_cache_cpu(cpu, SLOTS_CACHE | SLOTS_CACHE_RET, true); | |
237 | mutex_unlock(&swap_slots_cache_mutex); | |
238 | return 0; | |
239 | } | |
240 | ||
241 | int enable_swap_slots_cache(void) | |
242 | { | |
243 | int ret = 0; | |
244 | ||
245 | mutex_lock(&swap_slots_cache_enable_mutex); | |
246 | if (swap_slot_cache_initialized) { | |
247 | __reenable_swap_slots_cache(); | |
248 | goto out_unlock; | |
249 | } | |
250 | ||
251 | ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "swap_slots_cache", | |
252 | alloc_swap_slot_cache, free_slot_cache); | |
9b7a8143 TC |
253 | if (WARN_ONCE(ret < 0, "Cache allocation failed (%s), operating " |
254 | "without swap slots cache.\n", __func__)) | |
67afa38e | 255 | goto out_unlock; |
9b7a8143 | 256 | |
67afa38e TC |
257 | swap_slot_cache_initialized = true; |
258 | __reenable_swap_slots_cache(); | |
259 | out_unlock: | |
260 | mutex_unlock(&swap_slots_cache_enable_mutex); | |
261 | return 0; | |
262 | } | |
263 | ||
264 | /* called with swap slot cache's alloc lock held */ | |
265 | static int refill_swap_slots_cache(struct swap_slots_cache *cache) | |
266 | { | |
267 | if (!use_swap_slot_cache || cache->nr) | |
268 | return 0; | |
269 | ||
270 | cache->cur = 0; | |
271 | if (swap_slot_cache_active) | |
5d5e8f19 HY |
272 | cache->nr = get_swap_pages(SWAP_SLOTS_CACHE_SIZE, |
273 | cache->slots, 1); | |
67afa38e TC |
274 | |
275 | return cache->nr; | |
276 | } | |
277 | ||
278 | int free_swap_slot(swp_entry_t entry) | |
279 | { | |
280 | struct swap_slots_cache *cache; | |
281 | ||
f07e0f84 | 282 | cache = raw_cpu_ptr(&swp_slots); |
a2e16731 | 283 | if (likely(use_swap_slot_cache && cache->slots_ret)) { |
67afa38e TC |
284 | spin_lock_irq(&cache->free_lock); |
285 | /* Swap slots cache may be deactivated before acquiring lock */ | |
f07e0f84 | 286 | if (!use_swap_slot_cache || !cache->slots_ret) { |
67afa38e TC |
287 | spin_unlock_irq(&cache->free_lock); |
288 | goto direct_free; | |
289 | } | |
290 | if (cache->n_ret >= SWAP_SLOTS_CACHE_SIZE) { | |
291 | /* | |
292 | * Return slots to global pool. | |
293 | * The current swap_map value is SWAP_HAS_CACHE. | |
294 | * Set it to 0 to indicate it is available for | |
295 | * allocation in global pool | |
296 | */ | |
297 | swapcache_free_entries(cache->slots_ret, cache->n_ret); | |
298 | cache->n_ret = 0; | |
299 | } | |
300 | cache->slots_ret[cache->n_ret++] = entry; | |
301 | spin_unlock_irq(&cache->free_lock); | |
302 | } else { | |
303 | direct_free: | |
304 | swapcache_free_entries(&entry, 1); | |
305 | } | |
67afa38e TC |
306 | |
307 | return 0; | |
308 | } | |
309 | ||
38d8b4e6 | 310 | swp_entry_t get_swap_page(struct page *page) |
67afa38e TC |
311 | { |
312 | swp_entry_t entry, *pentry; | |
313 | struct swap_slots_cache *cache; | |
314 | ||
38d8b4e6 HY |
315 | entry.val = 0; |
316 | ||
317 | if (PageTransHuge(page)) { | |
318 | if (IS_ENABLED(CONFIG_THP_SWAP)) | |
5d5e8f19 | 319 | get_swap_pages(1, &entry, HPAGE_PMD_NR); |
bb98f2c5 | 320 | goto out; |
38d8b4e6 HY |
321 | } |
322 | ||
67afa38e TC |
323 | /* |
324 | * Preemption is allowed here, because we may sleep | |
325 | * in refill_swap_slots_cache(). But it is safe, because | |
326 | * accesses to the per-CPU data structure are protected by the | |
327 | * mutex cache->alloc_lock. | |
328 | * | |
329 | * The alloc path here does not touch cache->slots_ret | |
330 | * so cache->free_lock is not taken. | |
331 | */ | |
332 | cache = raw_cpu_ptr(&swp_slots); | |
333 | ||
a2e16731 | 334 | if (likely(check_cache_active() && cache->slots)) { |
67afa38e TC |
335 | mutex_lock(&cache->alloc_lock); |
336 | if (cache->slots) { | |
337 | repeat: | |
338 | if (cache->nr) { | |
339 | pentry = &cache->slots[cache->cur++]; | |
340 | entry = *pentry; | |
341 | pentry->val = 0; | |
342 | cache->nr--; | |
343 | } else { | |
344 | if (refill_swap_slots_cache(cache)) | |
345 | goto repeat; | |
346 | } | |
347 | } | |
348 | mutex_unlock(&cache->alloc_lock); | |
349 | if (entry.val) | |
bb98f2c5 | 350 | goto out; |
67afa38e TC |
351 | } |
352 | ||
5d5e8f19 | 353 | get_swap_pages(1, &entry, 1); |
bb98f2c5 TH |
354 | out: |
355 | if (mem_cgroup_try_charge_swap(page, entry)) { | |
356 | put_swap_page(page, entry); | |
357 | entry.val = 0; | |
358 | } | |
67afa38e TC |
359 | return entry; |
360 | } |