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 | |
f0953a1b | 19 | * allows the slots to coalesce and reduce fragmentation. |
67afa38e TC |
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); | |
67afa38e | 46 | |
e0f3ebba | 47 | #define use_swap_slot_cache (swap_slot_cache_active && swap_slot_cache_enabled) |
67afa38e TC |
48 | #define SLOTS_CACHE 0x1 |
49 | #define SLOTS_CACHE_RET 0x2 | |
50 | ||
51 | static void deactivate_swap_slots_cache(void) | |
52 | { | |
53 | mutex_lock(&swap_slots_cache_mutex); | |
54 | swap_slot_cache_active = false; | |
55 | __drain_swap_slots_cache(SLOTS_CACHE|SLOTS_CACHE_RET); | |
56 | mutex_unlock(&swap_slots_cache_mutex); | |
57 | } | |
58 | ||
59 | static void reactivate_swap_slots_cache(void) | |
60 | { | |
61 | mutex_lock(&swap_slots_cache_mutex); | |
62 | swap_slot_cache_active = true; | |
63 | mutex_unlock(&swap_slots_cache_mutex); | |
64 | } | |
65 | ||
66 | /* Must not be called with cpu hot plug lock */ | |
67 | void disable_swap_slots_cache_lock(void) | |
68 | { | |
69 | mutex_lock(&swap_slots_cache_enable_mutex); | |
70 | swap_slot_cache_enabled = false; | |
71 | if (swap_slot_cache_initialized) { | |
72 | /* serialize with cpu hotplug operations */ | |
7625eccd | 73 | cpus_read_lock(); |
67afa38e | 74 | __drain_swap_slots_cache(SLOTS_CACHE|SLOTS_CACHE_RET); |
7625eccd | 75 | cpus_read_unlock(); |
67afa38e TC |
76 | } |
77 | } | |
78 | ||
79 | static void __reenable_swap_slots_cache(void) | |
80 | { | |
81 | swap_slot_cache_enabled = has_usable_swap(); | |
82 | } | |
83 | ||
84 | void reenable_swap_slots_cache_unlock(void) | |
85 | { | |
86 | __reenable_swap_slots_cache(); | |
87 | mutex_unlock(&swap_slots_cache_enable_mutex); | |
88 | } | |
89 | ||
90 | static bool check_cache_active(void) | |
91 | { | |
92 | long pages; | |
93 | ||
e0f3ebba | 94 | if (!swap_slot_cache_enabled) |
67afa38e TC |
95 | return false; |
96 | ||
97 | pages = get_nr_swap_pages(); | |
98 | if (!swap_slot_cache_active) { | |
99 | if (pages > num_online_cpus() * | |
100 | THRESHOLD_ACTIVATE_SWAP_SLOTS_CACHE) | |
101 | reactivate_swap_slots_cache(); | |
102 | goto out; | |
103 | } | |
104 | ||
105 | /* if global pool of slot caches too low, deactivate cache */ | |
106 | if (pages < num_online_cpus() * THRESHOLD_DEACTIVATE_SWAP_SLOTS_CACHE) | |
107 | deactivate_swap_slots_cache(); | |
108 | out: | |
109 | return swap_slot_cache_active; | |
110 | } | |
111 | ||
112 | static int alloc_swap_slot_cache(unsigned int cpu) | |
113 | { | |
114 | struct swap_slots_cache *cache; | |
115 | swp_entry_t *slots, *slots_ret; | |
116 | ||
117 | /* | |
118 | * Do allocation outside swap_slots_cache_mutex | |
54f180d3 | 119 | * as kvzalloc could trigger reclaim and get_swap_page, |
67afa38e TC |
120 | * which can lock swap_slots_cache_mutex. |
121 | */ | |
778e1cdd | 122 | slots = kvcalloc(SWAP_SLOTS_CACHE_SIZE, sizeof(swp_entry_t), |
54f180d3 | 123 | GFP_KERNEL); |
67afa38e TC |
124 | if (!slots) |
125 | return -ENOMEM; | |
126 | ||
778e1cdd | 127 | slots_ret = kvcalloc(SWAP_SLOTS_CACHE_SIZE, sizeof(swp_entry_t), |
54f180d3 | 128 | GFP_KERNEL); |
67afa38e | 129 | if (!slots_ret) { |
54f180d3 | 130 | kvfree(slots); |
67afa38e TC |
131 | return -ENOMEM; |
132 | } | |
133 | ||
134 | mutex_lock(&swap_slots_cache_mutex); | |
135 | cache = &per_cpu(swp_slots, cpu); | |
f90eae2a | 136 | if (cache->slots || cache->slots_ret) { |
67afa38e | 137 | /* cache already allocated */ |
f90eae2a ZL |
138 | mutex_unlock(&swap_slots_cache_mutex); |
139 | ||
140 | kvfree(slots); | |
141 | kvfree(slots_ret); | |
142 | ||
143 | return 0; | |
144 | } | |
145 | ||
67afa38e TC |
146 | if (!cache->lock_initialized) { |
147 | mutex_init(&cache->alloc_lock); | |
148 | spin_lock_init(&cache->free_lock); | |
149 | cache->lock_initialized = true; | |
150 | } | |
151 | cache->nr = 0; | |
152 | cache->cur = 0; | |
153 | cache->n_ret = 0; | |
a2e16731 TC |
154 | /* |
155 | * We initialized alloc_lock and free_lock earlier. We use | |
156 | * !cache->slots or !cache->slots_ret to know if it is safe to acquire | |
157 | * the corresponding lock and use the cache. Memory barrier below | |
158 | * ensures the assumption. | |
159 | */ | |
160 | mb(); | |
67afa38e | 161 | cache->slots = slots; |
67afa38e | 162 | cache->slots_ret = slots_ret; |
67afa38e | 163 | mutex_unlock(&swap_slots_cache_mutex); |
67afa38e TC |
164 | return 0; |
165 | } | |
166 | ||
167 | static void drain_slots_cache_cpu(unsigned int cpu, unsigned int type, | |
168 | bool free_slots) | |
169 | { | |
170 | struct swap_slots_cache *cache; | |
171 | swp_entry_t *slots = NULL; | |
172 | ||
173 | cache = &per_cpu(swp_slots, cpu); | |
174 | if ((type & SLOTS_CACHE) && cache->slots) { | |
175 | mutex_lock(&cache->alloc_lock); | |
176 | swapcache_free_entries(cache->slots + cache->cur, cache->nr); | |
177 | cache->cur = 0; | |
178 | cache->nr = 0; | |
179 | if (free_slots && cache->slots) { | |
54f180d3 | 180 | kvfree(cache->slots); |
67afa38e TC |
181 | cache->slots = NULL; |
182 | } | |
183 | mutex_unlock(&cache->alloc_lock); | |
184 | } | |
185 | if ((type & SLOTS_CACHE_RET) && cache->slots_ret) { | |
186 | spin_lock_irq(&cache->free_lock); | |
187 | swapcache_free_entries(cache->slots_ret, cache->n_ret); | |
188 | cache->n_ret = 0; | |
189 | if (free_slots && cache->slots_ret) { | |
190 | slots = cache->slots_ret; | |
191 | cache->slots_ret = NULL; | |
192 | } | |
193 | spin_unlock_irq(&cache->free_lock); | |
191a7221 | 194 | kvfree(slots); |
67afa38e TC |
195 | } |
196 | } | |
197 | ||
198 | static void __drain_swap_slots_cache(unsigned int type) | |
199 | { | |
200 | unsigned int cpu; | |
201 | ||
202 | /* | |
203 | * This function is called during | |
204 | * 1) swapoff, when we have to make sure no | |
205 | * left over slots are in cache when we remove | |
206 | * a swap device; | |
207 | * 2) disabling of swap slot cache, when we run low | |
208 | * on swap slots when allocating memory and need | |
209 | * to return swap slots to global pool. | |
210 | * | |
211 | * We cannot acquire cpu hot plug lock here as | |
212 | * this function can be invoked in the cpu | |
213 | * hot plug path: | |
214 | * cpu_up -> lock cpu_hotplug -> cpu hotplug state callback | |
215 | * -> memory allocation -> direct reclaim -> get_swap_page | |
216 | * -> drain_swap_slots_cache | |
217 | * | |
218 | * Hence the loop over current online cpu below could miss cpu that | |
219 | * is being brought online but not yet marked as online. | |
220 | * That is okay as we do not schedule and run anything on a | |
221 | * cpu before it has been marked online. Hence, we will not | |
222 | * fill any swap slots in slots cache of such cpu. | |
223 | * There are no slots on such cpu that need to be drained. | |
224 | */ | |
225 | for_each_online_cpu(cpu) | |
226 | drain_slots_cache_cpu(cpu, type, false); | |
227 | } | |
228 | ||
229 | static int free_slot_cache(unsigned int cpu) | |
230 | { | |
231 | mutex_lock(&swap_slots_cache_mutex); | |
232 | drain_slots_cache_cpu(cpu, SLOTS_CACHE | SLOTS_CACHE_RET, true); | |
233 | mutex_unlock(&swap_slots_cache_mutex); | |
234 | return 0; | |
235 | } | |
236 | ||
f3bc52cb | 237 | void enable_swap_slots_cache(void) |
67afa38e | 238 | { |
67afa38e | 239 | mutex_lock(&swap_slots_cache_enable_mutex); |
d69a9575 ZL |
240 | if (!swap_slot_cache_initialized) { |
241 | int ret; | |
67afa38e | 242 | |
d69a9575 ZL |
243 | ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "swap_slots_cache", |
244 | alloc_swap_slot_cache, free_slot_cache); | |
245 | if (WARN_ONCE(ret < 0, "Cache allocation failed (%s), operating " | |
246 | "without swap slots cache.\n", __func__)) | |
247 | goto out_unlock; | |
248 | ||
249 | swap_slot_cache_initialized = true; | |
250 | } | |
9b7a8143 | 251 | |
67afa38e TC |
252 | __reenable_swap_slots_cache(); |
253 | out_unlock: | |
254 | mutex_unlock(&swap_slots_cache_enable_mutex); | |
67afa38e TC |
255 | } |
256 | ||
257 | /* called with swap slot cache's alloc lock held */ | |
258 | static int refill_swap_slots_cache(struct swap_slots_cache *cache) | |
259 | { | |
260 | if (!use_swap_slot_cache || cache->nr) | |
261 | return 0; | |
262 | ||
263 | cache->cur = 0; | |
264 | if (swap_slot_cache_active) | |
5d5e8f19 HY |
265 | cache->nr = get_swap_pages(SWAP_SLOTS_CACHE_SIZE, |
266 | cache->slots, 1); | |
67afa38e TC |
267 | |
268 | return cache->nr; | |
269 | } | |
270 | ||
271 | int free_swap_slot(swp_entry_t entry) | |
272 | { | |
273 | struct swap_slots_cache *cache; | |
274 | ||
f07e0f84 | 275 | cache = raw_cpu_ptr(&swp_slots); |
a2e16731 | 276 | if (likely(use_swap_slot_cache && cache->slots_ret)) { |
67afa38e TC |
277 | spin_lock_irq(&cache->free_lock); |
278 | /* Swap slots cache may be deactivated before acquiring lock */ | |
f07e0f84 | 279 | if (!use_swap_slot_cache || !cache->slots_ret) { |
67afa38e TC |
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 | } | |
67afa38e TC |
299 | |
300 | return 0; | |
301 | } | |
302 | ||
38d8b4e6 | 303 | swp_entry_t get_swap_page(struct page *page) |
67afa38e | 304 | { |
2406b76f | 305 | swp_entry_t entry; |
67afa38e TC |
306 | struct swap_slots_cache *cache; |
307 | ||
38d8b4e6 HY |
308 | entry.val = 0; |
309 | ||
310 | if (PageTransHuge(page)) { | |
311 | if (IS_ENABLED(CONFIG_THP_SWAP)) | |
5d5e8f19 | 312 | get_swap_pages(1, &entry, HPAGE_PMD_NR); |
bb98f2c5 | 313 | goto out; |
38d8b4e6 HY |
314 | } |
315 | ||
67afa38e TC |
316 | /* |
317 | * Preemption is allowed here, because we may sleep | |
318 | * in refill_swap_slots_cache(). But it is safe, because | |
319 | * accesses to the per-CPU data structure are protected by the | |
320 | * mutex cache->alloc_lock. | |
321 | * | |
322 | * The alloc path here does not touch cache->slots_ret | |
323 | * so cache->free_lock is not taken. | |
324 | */ | |
325 | cache = raw_cpu_ptr(&swp_slots); | |
326 | ||
a2e16731 | 327 | if (likely(check_cache_active() && cache->slots)) { |
67afa38e TC |
328 | mutex_lock(&cache->alloc_lock); |
329 | if (cache->slots) { | |
330 | repeat: | |
331 | if (cache->nr) { | |
2406b76f WY |
332 | entry = cache->slots[cache->cur]; |
333 | cache->slots[cache->cur++].val = 0; | |
67afa38e | 334 | cache->nr--; |
2406b76f WY |
335 | } else if (refill_swap_slots_cache(cache)) { |
336 | goto repeat; | |
67afa38e TC |
337 | } |
338 | } | |
339 | mutex_unlock(&cache->alloc_lock); | |
340 | if (entry.val) | |
bb98f2c5 | 341 | goto out; |
67afa38e TC |
342 | } |
343 | ||
5d5e8f19 | 344 | get_swap_pages(1, &entry, 1); |
bb98f2c5 TH |
345 | out: |
346 | if (mem_cgroup_try_charge_swap(page, entry)) { | |
347 | put_swap_page(page, entry); | |
348 | entry.val = 0; | |
349 | } | |
67afa38e TC |
350 | return entry; |
351 | } |