]>
Commit | Line | Data |
---|---|---|
798ab48e KO |
1 | /* |
2 | * Percpu IDA library | |
3 | * | |
4 | * Copyright (C) 2013 Datera, Inc. Kent Overstreet | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or | |
7 | * modify it under the terms of the GNU General Public License as | |
8 | * published by the Free Software Foundation; either version 2, or (at | |
9 | * your option) any later version. | |
10 | * | |
11 | * This program is distributed in the hope that it will be useful, but | |
12 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
14 | * General Public License for more details. | |
15 | */ | |
16 | ||
fd771233 | 17 | #include <linux/mm.h> |
798ab48e KO |
18 | #include <linux/bitmap.h> |
19 | #include <linux/bitops.h> | |
20 | #include <linux/bug.h> | |
21 | #include <linux/err.h> | |
22 | #include <linux/export.h> | |
798ab48e KO |
23 | #include <linux/init.h> |
24 | #include <linux/kernel.h> | |
25 | #include <linux/percpu.h> | |
174cd4b1 | 26 | #include <linux/sched/signal.h> |
798ab48e KO |
27 | #include <linux/string.h> |
28 | #include <linux/spinlock.h> | |
29 | #include <linux/percpu_ida.h> | |
30 | ||
798ab48e KO |
31 | struct percpu_ida_cpu { |
32 | /* | |
33 | * Even though this is percpu, we need a lock for tag stealing by remote | |
34 | * CPUs: | |
35 | */ | |
36 | spinlock_t lock; | |
37 | ||
38 | /* nr_free/freelist form a stack of free IDs */ | |
39 | unsigned nr_free; | |
40 | unsigned freelist[]; | |
41 | }; | |
42 | ||
43 | static inline void move_tags(unsigned *dst, unsigned *dst_nr, | |
44 | unsigned *src, unsigned *src_nr, | |
45 | unsigned nr) | |
46 | { | |
47 | *src_nr -= nr; | |
48 | memcpy(dst + *dst_nr, src + *src_nr, sizeof(unsigned) * nr); | |
49 | *dst_nr += nr; | |
50 | } | |
51 | ||
52 | /* | |
53 | * Try to steal tags from a remote cpu's percpu freelist. | |
54 | * | |
d835502f | 55 | * We first check how many percpu freelists have tags |
798ab48e KO |
56 | * |
57 | * Then we iterate through the cpus until we find some tags - we don't attempt | |
58 | * to find the "best" cpu to steal from, to keep cacheline bouncing to a | |
59 | * minimum. | |
60 | */ | |
61 | static inline void steal_tags(struct percpu_ida *pool, | |
62 | struct percpu_ida_cpu *tags) | |
63 | { | |
64 | unsigned cpus_have_tags, cpu = pool->cpu_last_stolen; | |
65 | struct percpu_ida_cpu *remote; | |
66 | ||
67 | for (cpus_have_tags = cpumask_weight(&pool->cpus_have_tags); | |
d835502f | 68 | cpus_have_tags; cpus_have_tags--) { |
798ab48e KO |
69 | cpu = cpumask_next(cpu, &pool->cpus_have_tags); |
70 | ||
71 | if (cpu >= nr_cpu_ids) { | |
72 | cpu = cpumask_first(&pool->cpus_have_tags); | |
73 | if (cpu >= nr_cpu_ids) | |
74 | BUG(); | |
75 | } | |
76 | ||
77 | pool->cpu_last_stolen = cpu; | |
78 | remote = per_cpu_ptr(pool->tag_cpu, cpu); | |
79 | ||
80 | cpumask_clear_cpu(cpu, &pool->cpus_have_tags); | |
81 | ||
82 | if (remote == tags) | |
83 | continue; | |
84 | ||
85 | spin_lock(&remote->lock); | |
86 | ||
87 | if (remote->nr_free) { | |
88 | memcpy(tags->freelist, | |
89 | remote->freelist, | |
90 | sizeof(unsigned) * remote->nr_free); | |
91 | ||
92 | tags->nr_free = remote->nr_free; | |
93 | remote->nr_free = 0; | |
94 | } | |
95 | ||
96 | spin_unlock(&remote->lock); | |
97 | ||
98 | if (tags->nr_free) | |
99 | break; | |
100 | } | |
101 | } | |
102 | ||
103 | /* | |
104 | * Pop up to IDA_PCPU_BATCH_MOVE IDs off the global freelist, and push them onto | |
105 | * our percpu freelist: | |
106 | */ | |
107 | static inline void alloc_global_tags(struct percpu_ida *pool, | |
108 | struct percpu_ida_cpu *tags) | |
109 | { | |
110 | move_tags(tags->freelist, &tags->nr_free, | |
111 | pool->freelist, &pool->nr_free, | |
e26b53d0 | 112 | min(pool->nr_free, pool->percpu_batch_size)); |
798ab48e KO |
113 | } |
114 | ||
6b378382 | 115 | static inline unsigned alloc_local_tag(struct percpu_ida_cpu *tags) |
798ab48e KO |
116 | { |
117 | int tag = -ENOSPC; | |
118 | ||
119 | spin_lock(&tags->lock); | |
120 | if (tags->nr_free) | |
121 | tag = tags->freelist[--tags->nr_free]; | |
122 | spin_unlock(&tags->lock); | |
123 | ||
124 | return tag; | |
125 | } | |
126 | ||
127 | /** | |
128 | * percpu_ida_alloc - allocate a tag | |
129 | * @pool: pool to allocate from | |
6f6b5d1e | 130 | * @state: task state for prepare_to_wait |
798ab48e KO |
131 | * |
132 | * Returns a tag - an integer in the range [0..nr_tags) (passed to | |
133 | * tag_pool_init()), or otherwise -ENOSPC on allocation failure. | |
134 | * | |
135 | * Safe to be called from interrupt context (assuming it isn't passed | |
555b270e | 136 | * TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, of course). |
798ab48e KO |
137 | * |
138 | * @gfp indicates whether or not to wait until a free id is available (it's not | |
71baba4b | 139 | * used for internal memory allocations); thus if passed __GFP_RECLAIM we may sleep |
798ab48e KO |
140 | * however long it takes until another thread frees an id (same semantics as a |
141 | * mempool). | |
142 | * | |
555b270e | 143 | * Will not fail if passed TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE. |
798ab48e | 144 | */ |
6f6b5d1e | 145 | int percpu_ida_alloc(struct percpu_ida *pool, int state) |
798ab48e KO |
146 | { |
147 | DEFINE_WAIT(wait); | |
148 | struct percpu_ida_cpu *tags; | |
149 | unsigned long flags; | |
150 | int tag; | |
151 | ||
152 | local_irq_save(flags); | |
153 | tags = this_cpu_ptr(pool->tag_cpu); | |
154 | ||
155 | /* Fastpath */ | |
6b378382 | 156 | tag = alloc_local_tag(tags); |
798ab48e KO |
157 | if (likely(tag >= 0)) { |
158 | local_irq_restore(flags); | |
159 | return tag; | |
160 | } | |
161 | ||
162 | while (1) { | |
163 | spin_lock(&pool->lock); | |
164 | ||
165 | /* | |
166 | * prepare_to_wait() must come before steal_tags(), in case | |
167 | * percpu_ida_free() on another cpu flips a bit in | |
168 | * cpus_have_tags | |
169 | * | |
170 | * global lock held and irqs disabled, don't need percpu lock | |
171 | */ | |
6f6b5d1e KO |
172 | if (state != TASK_RUNNING) |
173 | prepare_to_wait(&pool->wait, &wait, state); | |
798ab48e KO |
174 | |
175 | if (!tags->nr_free) | |
176 | alloc_global_tags(pool, tags); | |
177 | if (!tags->nr_free) | |
178 | steal_tags(pool, tags); | |
179 | ||
180 | if (tags->nr_free) { | |
181 | tag = tags->freelist[--tags->nr_free]; | |
182 | if (tags->nr_free) | |
183 | cpumask_set_cpu(smp_processor_id(), | |
184 | &pool->cpus_have_tags); | |
185 | } | |
186 | ||
187 | spin_unlock(&pool->lock); | |
188 | local_irq_restore(flags); | |
189 | ||
6f6b5d1e | 190 | if (tag >= 0 || state == TASK_RUNNING) |
798ab48e KO |
191 | break; |
192 | ||
555b270e NB |
193 | if (signal_pending_state(state, current)) { |
194 | tag = -ERESTARTSYS; | |
195 | break; | |
196 | } | |
197 | ||
798ab48e KO |
198 | schedule(); |
199 | ||
200 | local_irq_save(flags); | |
201 | tags = this_cpu_ptr(pool->tag_cpu); | |
202 | } | |
6f6b5d1e KO |
203 | if (state != TASK_RUNNING) |
204 | finish_wait(&pool->wait, &wait); | |
798ab48e | 205 | |
798ab48e KO |
206 | return tag; |
207 | } | |
208 | EXPORT_SYMBOL_GPL(percpu_ida_alloc); | |
209 | ||
210 | /** | |
211 | * percpu_ida_free - free a tag | |
212 | * @pool: pool @tag was allocated from | |
213 | * @tag: a tag previously allocated with percpu_ida_alloc() | |
214 | * | |
215 | * Safe to be called from interrupt context. | |
216 | */ | |
217 | void percpu_ida_free(struct percpu_ida *pool, unsigned tag) | |
218 | { | |
219 | struct percpu_ida_cpu *tags; | |
220 | unsigned long flags; | |
221 | unsigned nr_free; | |
222 | ||
223 | BUG_ON(tag >= pool->nr_tags); | |
224 | ||
225 | local_irq_save(flags); | |
226 | tags = this_cpu_ptr(pool->tag_cpu); | |
227 | ||
228 | spin_lock(&tags->lock); | |
229 | tags->freelist[tags->nr_free++] = tag; | |
230 | ||
231 | nr_free = tags->nr_free; | |
232 | spin_unlock(&tags->lock); | |
233 | ||
234 | if (nr_free == 1) { | |
235 | cpumask_set_cpu(smp_processor_id(), | |
236 | &pool->cpus_have_tags); | |
237 | wake_up(&pool->wait); | |
238 | } | |
239 | ||
e26b53d0 | 240 | if (nr_free == pool->percpu_max_size) { |
798ab48e KO |
241 | spin_lock(&pool->lock); |
242 | ||
243 | /* | |
244 | * Global lock held and irqs disabled, don't need percpu | |
245 | * lock | |
246 | */ | |
e26b53d0 | 247 | if (tags->nr_free == pool->percpu_max_size) { |
798ab48e KO |
248 | move_tags(pool->freelist, &pool->nr_free, |
249 | tags->freelist, &tags->nr_free, | |
e26b53d0 | 250 | pool->percpu_batch_size); |
798ab48e KO |
251 | |
252 | wake_up(&pool->wait); | |
253 | } | |
254 | spin_unlock(&pool->lock); | |
255 | } | |
256 | ||
257 | local_irq_restore(flags); | |
258 | } | |
259 | EXPORT_SYMBOL_GPL(percpu_ida_free); | |
260 | ||
261 | /** | |
262 | * percpu_ida_destroy - release a tag pool's resources | |
263 | * @pool: pool to free | |
264 | * | |
265 | * Frees the resources allocated by percpu_ida_init(). | |
266 | */ | |
267 | void percpu_ida_destroy(struct percpu_ida *pool) | |
268 | { | |
269 | free_percpu(pool->tag_cpu); | |
270 | free_pages((unsigned long) pool->freelist, | |
271 | get_order(pool->nr_tags * sizeof(unsigned))); | |
272 | } | |
273 | EXPORT_SYMBOL_GPL(percpu_ida_destroy); | |
274 | ||
275 | /** | |
276 | * percpu_ida_init - initialize a percpu tag pool | |
277 | * @pool: pool to initialize | |
278 | * @nr_tags: number of tags that will be available for allocation | |
279 | * | |
280 | * Initializes @pool so that it can be used to allocate tags - integers in the | |
281 | * range [0, nr_tags). Typically, they'll be used by driver code to refer to a | |
282 | * preallocated array of tag structures. | |
283 | * | |
284 | * Allocation is percpu, but sharding is limited by nr_tags - for best | |
285 | * performance, the workload should not span more cpus than nr_tags / 128. | |
286 | */ | |
e26b53d0 SL |
287 | int __percpu_ida_init(struct percpu_ida *pool, unsigned long nr_tags, |
288 | unsigned long max_size, unsigned long batch_size) | |
798ab48e KO |
289 | { |
290 | unsigned i, cpu, order; | |
291 | ||
292 | memset(pool, 0, sizeof(*pool)); | |
293 | ||
294 | init_waitqueue_head(&pool->wait); | |
295 | spin_lock_init(&pool->lock); | |
296 | pool->nr_tags = nr_tags; | |
e26b53d0 SL |
297 | pool->percpu_max_size = max_size; |
298 | pool->percpu_batch_size = batch_size; | |
798ab48e KO |
299 | |
300 | /* Guard against overflow */ | |
301 | if (nr_tags > (unsigned) INT_MAX + 1) { | |
302 | pr_err("percpu_ida_init(): nr_tags too large\n"); | |
303 | return -EINVAL; | |
304 | } | |
305 | ||
306 | order = get_order(nr_tags * sizeof(unsigned)); | |
307 | pool->freelist = (void *) __get_free_pages(GFP_KERNEL, order); | |
308 | if (!pool->freelist) | |
309 | return -ENOMEM; | |
310 | ||
311 | for (i = 0; i < nr_tags; i++) | |
312 | pool->freelist[i] = i; | |
313 | ||
314 | pool->nr_free = nr_tags; | |
315 | ||
316 | pool->tag_cpu = __alloc_percpu(sizeof(struct percpu_ida_cpu) + | |
e26b53d0 | 317 | pool->percpu_max_size * sizeof(unsigned), |
798ab48e KO |
318 | sizeof(unsigned)); |
319 | if (!pool->tag_cpu) | |
320 | goto err; | |
321 | ||
322 | for_each_possible_cpu(cpu) | |
323 | spin_lock_init(&per_cpu_ptr(pool->tag_cpu, cpu)->lock); | |
324 | ||
325 | return 0; | |
326 | err: | |
327 | percpu_ida_destroy(pool); | |
328 | return -ENOMEM; | |
329 | } | |
e26b53d0 | 330 | EXPORT_SYMBOL_GPL(__percpu_ida_init); |
7fc2ba17 SL |
331 | |
332 | /** | |
333 | * percpu_ida_for_each_free - iterate free ids of a pool | |
334 | * @pool: pool to iterate | |
335 | * @fn: interate callback function | |
336 | * @data: parameter for @fn | |
337 | * | |
338 | * Note, this doesn't guarantee to iterate all free ids restrictly. Some free | |
339 | * ids might be missed, some might be iterated duplicated, and some might | |
340 | * be iterated and not free soon. | |
341 | */ | |
342 | int percpu_ida_for_each_free(struct percpu_ida *pool, percpu_ida_cb fn, | |
343 | void *data) | |
344 | { | |
345 | unsigned long flags; | |
346 | struct percpu_ida_cpu *remote; | |
347 | unsigned cpu, i, err = 0; | |
348 | ||
349 | local_irq_save(flags); | |
350 | for_each_possible_cpu(cpu) { | |
351 | remote = per_cpu_ptr(pool->tag_cpu, cpu); | |
352 | spin_lock(&remote->lock); | |
353 | for (i = 0; i < remote->nr_free; i++) { | |
354 | err = fn(remote->freelist[i], data); | |
355 | if (err) | |
356 | break; | |
357 | } | |
358 | spin_unlock(&remote->lock); | |
359 | if (err) | |
360 | goto out; | |
361 | } | |
362 | ||
363 | spin_lock(&pool->lock); | |
364 | for (i = 0; i < pool->nr_free; i++) { | |
365 | err = fn(pool->freelist[i], data); | |
366 | if (err) | |
367 | break; | |
368 | } | |
369 | spin_unlock(&pool->lock); | |
370 | out: | |
371 | local_irq_restore(flags); | |
372 | return err; | |
373 | } | |
374 | EXPORT_SYMBOL_GPL(percpu_ida_for_each_free); | |
1dddc01a SL |
375 | |
376 | /** | |
377 | * percpu_ida_free_tags - return free tags number of a specific cpu or global pool | |
378 | * @pool: pool related | |
379 | * @cpu: specific cpu or global pool if @cpu == nr_cpu_ids | |
380 | * | |
381 | * Note: this just returns a snapshot of free tags number. | |
382 | */ | |
383 | unsigned percpu_ida_free_tags(struct percpu_ida *pool, int cpu) | |
384 | { | |
385 | struct percpu_ida_cpu *remote; | |
386 | if (cpu == nr_cpu_ids) | |
387 | return pool->nr_free; | |
388 | remote = per_cpu_ptr(pool->tag_cpu, cpu); | |
389 | return remote->nr_free; | |
390 | } | |
391 | EXPORT_SYMBOL_GPL(percpu_ida_free_tags); |