4 * Copyright (C) 2013 Datera, Inc. Kent Overstreet
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.
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.
17 #include <linux/bitmap.h>
18 #include <linux/bitops.h>
19 #include <linux/bug.h>
20 #include <linux/err.h>
21 #include <linux/export.h>
22 #include <linux/hardirq.h>
23 #include <linux/idr.h>
24 #include <linux/init.h>
25 #include <linux/kernel.h>
26 #include <linux/percpu.h>
27 #include <linux/sched.h>
28 #include <linux/slab.h>
29 #include <linux/string.h>
30 #include <linux/spinlock.h>
31 #include <linux/percpu_ida.h>
33 struct percpu_ida_cpu
{
35 * Even though this is percpu, we need a lock for tag stealing by remote
40 /* nr_free/freelist form a stack of free IDs */
45 static inline void move_tags(unsigned *dst
, unsigned *dst_nr
,
46 unsigned *src
, unsigned *src_nr
,
50 memcpy(dst
+ *dst_nr
, src
+ *src_nr
, sizeof(unsigned) * nr
);
55 * Try to steal tags from a remote cpu's percpu freelist.
57 * We first check how many percpu freelists have tags - we don't steal tags
58 * unless enough percpu freelists have tags on them that it's possible more than
59 * half the total tags could be stuck on remote percpu freelists.
61 * Then we iterate through the cpus until we find some tags - we don't attempt
62 * to find the "best" cpu to steal from, to keep cacheline bouncing to a
65 static inline void steal_tags(struct percpu_ida
*pool
,
66 struct percpu_ida_cpu
*tags
)
68 unsigned cpus_have_tags
, cpu
= pool
->cpu_last_stolen
;
69 struct percpu_ida_cpu
*remote
;
71 for (cpus_have_tags
= cpumask_weight(&pool
->cpus_have_tags
);
72 cpus_have_tags
* pool
->percpu_max_size
> pool
->nr_tags
/ 2;
74 cpu
= cpumask_next(cpu
, &pool
->cpus_have_tags
);
76 if (cpu
>= nr_cpu_ids
) {
77 cpu
= cpumask_first(&pool
->cpus_have_tags
);
78 if (cpu
>= nr_cpu_ids
)
82 pool
->cpu_last_stolen
= cpu
;
83 remote
= per_cpu_ptr(pool
->tag_cpu
, cpu
);
85 cpumask_clear_cpu(cpu
, &pool
->cpus_have_tags
);
90 spin_lock(&remote
->lock
);
92 if (remote
->nr_free
) {
93 memcpy(tags
->freelist
,
95 sizeof(unsigned) * remote
->nr_free
);
97 tags
->nr_free
= remote
->nr_free
;
101 spin_unlock(&remote
->lock
);
109 * Pop up to IDA_PCPU_BATCH_MOVE IDs off the global freelist, and push them onto
110 * our percpu freelist:
112 static inline void alloc_global_tags(struct percpu_ida
*pool
,
113 struct percpu_ida_cpu
*tags
)
115 move_tags(tags
->freelist
, &tags
->nr_free
,
116 pool
->freelist
, &pool
->nr_free
,
117 min(pool
->nr_free
, pool
->percpu_batch_size
));
120 static inline unsigned alloc_local_tag(struct percpu_ida
*pool
,
121 struct percpu_ida_cpu
*tags
)
125 spin_lock(&tags
->lock
);
127 tag
= tags
->freelist
[--tags
->nr_free
];
128 spin_unlock(&tags
->lock
);
134 * percpu_ida_alloc - allocate a tag
135 * @pool: pool to allocate from
138 * Returns a tag - an integer in the range [0..nr_tags) (passed to
139 * tag_pool_init()), or otherwise -ENOSPC on allocation failure.
141 * Safe to be called from interrupt context (assuming it isn't passed
142 * __GFP_WAIT, of course).
144 * @gfp indicates whether or not to wait until a free id is available (it's not
145 * used for internal memory allocations); thus if passed __GFP_WAIT we may sleep
146 * however long it takes until another thread frees an id (same semantics as a
149 * Will not fail if passed __GFP_WAIT.
151 int percpu_ida_alloc(struct percpu_ida
*pool
, gfp_t gfp
)
154 struct percpu_ida_cpu
*tags
;
158 local_irq_save(flags
);
159 tags
= this_cpu_ptr(pool
->tag_cpu
);
162 tag
= alloc_local_tag(pool
, tags
);
163 if (likely(tag
>= 0)) {
164 local_irq_restore(flags
);
169 spin_lock(&pool
->lock
);
172 * prepare_to_wait() must come before steal_tags(), in case
173 * percpu_ida_free() on another cpu flips a bit in
176 * global lock held and irqs disabled, don't need percpu lock
178 prepare_to_wait(&pool
->wait
, &wait
, TASK_UNINTERRUPTIBLE
);
181 alloc_global_tags(pool
, tags
);
183 steal_tags(pool
, tags
);
186 tag
= tags
->freelist
[--tags
->nr_free
];
188 cpumask_set_cpu(smp_processor_id(),
189 &pool
->cpus_have_tags
);
192 spin_unlock(&pool
->lock
);
193 local_irq_restore(flags
);
195 if (tag
>= 0 || !(gfp
& __GFP_WAIT
))
200 local_irq_save(flags
);
201 tags
= this_cpu_ptr(pool
->tag_cpu
);
204 finish_wait(&pool
->wait
, &wait
);
207 EXPORT_SYMBOL_GPL(percpu_ida_alloc
);
210 * percpu_ida_free - free a tag
211 * @pool: pool @tag was allocated from
212 * @tag: a tag previously allocated with percpu_ida_alloc()
214 * Safe to be called from interrupt context.
216 void percpu_ida_free(struct percpu_ida
*pool
, unsigned tag
)
218 struct percpu_ida_cpu
*tags
;
222 BUG_ON(tag
>= pool
->nr_tags
);
224 local_irq_save(flags
);
225 tags
= this_cpu_ptr(pool
->tag_cpu
);
227 spin_lock(&tags
->lock
);
228 tags
->freelist
[tags
->nr_free
++] = tag
;
230 nr_free
= tags
->nr_free
;
231 spin_unlock(&tags
->lock
);
234 cpumask_set_cpu(smp_processor_id(),
235 &pool
->cpus_have_tags
);
236 wake_up(&pool
->wait
);
239 if (nr_free
== pool
->percpu_max_size
) {
240 spin_lock(&pool
->lock
);
243 * Global lock held and irqs disabled, don't need percpu
246 if (tags
->nr_free
== pool
->percpu_max_size
) {
247 move_tags(pool
->freelist
, &pool
->nr_free
,
248 tags
->freelist
, &tags
->nr_free
,
249 pool
->percpu_batch_size
);
251 wake_up(&pool
->wait
);
253 spin_unlock(&pool
->lock
);
256 local_irq_restore(flags
);
258 EXPORT_SYMBOL_GPL(percpu_ida_free
);
261 * percpu_ida_destroy - release a tag pool's resources
262 * @pool: pool to free
264 * Frees the resources allocated by percpu_ida_init().
266 void percpu_ida_destroy(struct percpu_ida
*pool
)
268 free_percpu(pool
->tag_cpu
);
269 free_pages((unsigned long) pool
->freelist
,
270 get_order(pool
->nr_tags
* sizeof(unsigned)));
272 EXPORT_SYMBOL_GPL(percpu_ida_destroy
);
275 * percpu_ida_init - initialize a percpu tag pool
276 * @pool: pool to initialize
277 * @nr_tags: number of tags that will be available for allocation
279 * Initializes @pool so that it can be used to allocate tags - integers in the
280 * range [0, nr_tags). Typically, they'll be used by driver code to refer to a
281 * preallocated array of tag structures.
283 * Allocation is percpu, but sharding is limited by nr_tags - for best
284 * performance, the workload should not span more cpus than nr_tags / 128.
286 int __percpu_ida_init(struct percpu_ida
*pool
, unsigned long nr_tags
,
287 unsigned long max_size
, unsigned long batch_size
)
289 unsigned i
, cpu
, order
;
291 memset(pool
, 0, sizeof(*pool
));
293 init_waitqueue_head(&pool
->wait
);
294 spin_lock_init(&pool
->lock
);
295 pool
->nr_tags
= nr_tags
;
296 pool
->percpu_max_size
= max_size
;
297 pool
->percpu_batch_size
= batch_size
;
299 /* Guard against overflow */
300 if (nr_tags
> (unsigned) INT_MAX
+ 1) {
301 pr_err("percpu_ida_init(): nr_tags too large\n");
305 order
= get_order(nr_tags
* sizeof(unsigned));
306 pool
->freelist
= (void *) __get_free_pages(GFP_KERNEL
, order
);
310 for (i
= 0; i
< nr_tags
; i
++)
311 pool
->freelist
[i
] = i
;
313 pool
->nr_free
= nr_tags
;
315 pool
->tag_cpu
= __alloc_percpu(sizeof(struct percpu_ida_cpu
) +
316 pool
->percpu_max_size
* sizeof(unsigned),
321 for_each_possible_cpu(cpu
)
322 spin_lock_init(&per_cpu_ptr(pool
->tag_cpu
, cpu
)->lock
);
326 percpu_ida_destroy(pool
);
329 EXPORT_SYMBOL_GPL(__percpu_ida_init
);
332 * percpu_ida_for_each_free - iterate free ids of a pool
333 * @pool: pool to iterate
334 * @fn: interate callback function
335 * @data: parameter for @fn
337 * Note, this doesn't guarantee to iterate all free ids restrictly. Some free
338 * ids might be missed, some might be iterated duplicated, and some might
339 * be iterated and not free soon.
341 int percpu_ida_for_each_free(struct percpu_ida
*pool
, percpu_ida_cb fn
,
345 struct percpu_ida_cpu
*remote
;
346 unsigned cpu
, i
, err
= 0;
348 local_irq_save(flags
);
349 for_each_possible_cpu(cpu
) {
350 remote
= per_cpu_ptr(pool
->tag_cpu
, cpu
);
351 spin_lock(&remote
->lock
);
352 for (i
= 0; i
< remote
->nr_free
; i
++) {
353 err
= fn(remote
->freelist
[i
], data
);
357 spin_unlock(&remote
->lock
);
362 spin_lock(&pool
->lock
);
363 for (i
= 0; i
< pool
->nr_free
; i
++) {
364 err
= fn(pool
->freelist
[i
], data
);
368 spin_unlock(&pool
->lock
);
370 local_irq_restore(flags
);
373 EXPORT_SYMBOL_GPL(percpu_ida_for_each_free
);
376 * percpu_ida_free_tags - return free tags number of a specific cpu or global pool
377 * @pool: pool related
378 * @cpu: specific cpu or global pool if @cpu == nr_cpu_ids
380 * Note: this just returns a snapshot of free tags number.
382 unsigned percpu_ida_free_tags(struct percpu_ida
*pool
, int cpu
)
384 struct percpu_ida_cpu
*remote
;
385 if (cpu
== nr_cpu_ids
)
386 return pool
->nr_free
;
387 remote
= per_cpu_ptr(pool
->tag_cpu
, cpu
);
388 return remote
->nr_free
;
390 EXPORT_SYMBOL_GPL(percpu_ida_free_tags
);
This page took 0.05258 seconds and 5 git commands to generate.