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/init.h>
23 #include <linux/kernel.h>
24 #include <linux/percpu.h>
25 #include <linux/sched.h>
26 #include <linux/string.h>
27 #include <linux/spinlock.h>
28 #include <linux/percpu_ida.h>
29 #include <linux/locallock.h>
31 static DEFINE_LOCAL_IRQ_LOCK(irq_off_lock);
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
59 * Then we iterate through the cpus until we find some tags - we don't attempt
60 * to find the "best" cpu to steal from, to keep cacheline bouncing to a
63 static inline void steal_tags(struct percpu_ida *pool,
64 struct percpu_ida_cpu *tags)
66 unsigned cpus_have_tags, cpu = pool->cpu_last_stolen;
67 struct percpu_ida_cpu *remote;
69 for (cpus_have_tags = cpumask_weight(&pool->cpus_have_tags);
70 cpus_have_tags; cpus_have_tags--) {
71 cpu = cpumask_next(cpu, &pool->cpus_have_tags);
73 if (cpu >= nr_cpu_ids) {
74 cpu = cpumask_first(&pool->cpus_have_tags);
75 if (cpu >= nr_cpu_ids)
79 pool->cpu_last_stolen = cpu;
80 remote = per_cpu_ptr(pool->tag_cpu, cpu);
82 cpumask_clear_cpu(cpu, &pool->cpus_have_tags);
87 spin_lock(&remote->lock);
89 if (remote->nr_free) {
90 memcpy(tags->freelist,
92 sizeof(unsigned) * remote->nr_free);
94 tags->nr_free = remote->nr_free;
98 spin_unlock(&remote->lock);
106 * Pop up to IDA_PCPU_BATCH_MOVE IDs off the global freelist, and push them onto
107 * our percpu freelist:
109 static inline void alloc_global_tags(struct percpu_ida *pool,
110 struct percpu_ida_cpu *tags)
112 move_tags(tags->freelist, &tags->nr_free,
113 pool->freelist, &pool->nr_free,
114 min(pool->nr_free, pool->percpu_batch_size));
117 static inline unsigned alloc_local_tag(struct percpu_ida_cpu *tags)
121 spin_lock(&tags->lock);
123 tag = tags->freelist[--tags->nr_free];
124 spin_unlock(&tags->lock);
130 * percpu_ida_alloc - allocate a tag
131 * @pool: pool to allocate from
132 * @state: task state for prepare_to_wait
134 * Returns a tag - an integer in the range [0..nr_tags) (passed to
135 * tag_pool_init()), or otherwise -ENOSPC on allocation failure.
137 * Safe to be called from interrupt context (assuming it isn't passed
138 * TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE, of course).
140 * @gfp indicates whether or not to wait until a free id is available (it's not
141 * used for internal memory allocations); thus if passed __GFP_RECLAIM we may sleep
142 * however long it takes until another thread frees an id (same semantics as a
145 * Will not fail if passed TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE.
147 int percpu_ida_alloc(struct percpu_ida *pool, int state)
150 struct percpu_ida_cpu *tags;
154 local_lock_irqsave(irq_off_lock, flags);
155 tags = this_cpu_ptr(pool->tag_cpu);
158 tag = alloc_local_tag(tags);
159 if (likely(tag >= 0)) {
160 local_unlock_irqrestore(irq_off_lock, flags);
165 spin_lock(&pool->lock);
168 * prepare_to_wait() must come before steal_tags(), in case
169 * percpu_ida_free() on another cpu flips a bit in
172 * global lock held and irqs disabled, don't need percpu lock
174 if (state != TASK_RUNNING)
175 prepare_to_wait(&pool->wait, &wait, state);
178 alloc_global_tags(pool, tags);
181 steal_tags(pool, tags);
184 tag = tags->freelist[--tags->nr_free];
186 cpumask_set_cpu(smp_processor_id(),
187 &pool->cpus_have_tags);
190 spin_unlock(&pool->lock);
191 local_unlock_irqrestore(irq_off_lock, flags);
193 if (tag >= 0 || state == TASK_RUNNING)
196 if (signal_pending_state(state, current)) {
203 local_lock_irqsave(irq_off_lock, flags);
204 tags = this_cpu_ptr(pool->tag_cpu);
206 if (state != TASK_RUNNING)
207 finish_wait(&pool->wait, &wait);
211 EXPORT_SYMBOL_GPL(percpu_ida_alloc);
214 * percpu_ida_free - free a tag
215 * @pool: pool @tag was allocated from
216 * @tag: a tag previously allocated with percpu_ida_alloc()
218 * Safe to be called from interrupt context.
220 void percpu_ida_free(struct percpu_ida *pool, unsigned tag)
222 struct percpu_ida_cpu *tags;
226 BUG_ON(tag >= pool->nr_tags);
228 local_lock_irqsave(irq_off_lock, flags);
229 tags = this_cpu_ptr(pool->tag_cpu);
231 spin_lock(&tags->lock);
232 tags->freelist[tags->nr_free++] = tag;
234 nr_free = tags->nr_free;
235 spin_unlock(&tags->lock);
238 cpumask_set_cpu(smp_processor_id(),
239 &pool->cpus_have_tags);
240 wake_up(&pool->wait);
243 if (nr_free == pool->percpu_max_size) {
244 spin_lock(&pool->lock);
247 * Global lock held and irqs disabled, don't need percpu
250 if (tags->nr_free == pool->percpu_max_size) {
251 move_tags(pool->freelist, &pool->nr_free,
252 tags->freelist, &tags->nr_free,
253 pool->percpu_batch_size);
255 wake_up(&pool->wait);
257 spin_unlock(&pool->lock);
260 local_unlock_irqrestore(irq_off_lock, flags);
262 EXPORT_SYMBOL_GPL(percpu_ida_free);
265 * percpu_ida_destroy - release a tag pool's resources
266 * @pool: pool to free
268 * Frees the resources allocated by percpu_ida_init().
270 void percpu_ida_destroy(struct percpu_ida *pool)
272 free_percpu(pool->tag_cpu);
273 free_pages((unsigned long) pool->freelist,
274 get_order(pool->nr_tags * sizeof(unsigned)));
276 EXPORT_SYMBOL_GPL(percpu_ida_destroy);
279 * percpu_ida_init - initialize a percpu tag pool
280 * @pool: pool to initialize
281 * @nr_tags: number of tags that will be available for allocation
283 * Initializes @pool so that it can be used to allocate tags - integers in the
284 * range [0, nr_tags). Typically, they'll be used by driver code to refer to a
285 * preallocated array of tag structures.
287 * Allocation is percpu, but sharding is limited by nr_tags - for best
288 * performance, the workload should not span more cpus than nr_tags / 128.
290 int __percpu_ida_init(struct percpu_ida *pool, unsigned long nr_tags,
291 unsigned long max_size, unsigned long batch_size)
293 unsigned i, cpu, order;
295 memset(pool, 0, sizeof(*pool));
297 init_waitqueue_head(&pool->wait);
298 spin_lock_init(&pool->lock);
299 pool->nr_tags = nr_tags;
300 pool->percpu_max_size = max_size;
301 pool->percpu_batch_size = batch_size;
303 /* Guard against overflow */
304 if (nr_tags > (unsigned) INT_MAX + 1) {
305 pr_err("percpu_ida_init(): nr_tags too large\n");
309 order = get_order(nr_tags * sizeof(unsigned));
310 pool->freelist = (void *) __get_free_pages(GFP_KERNEL, order);
314 for (i = 0; i < nr_tags; i++)
315 pool->freelist[i] = i;
317 pool->nr_free = nr_tags;
319 pool->tag_cpu = __alloc_percpu(sizeof(struct percpu_ida_cpu) +
320 pool->percpu_max_size * sizeof(unsigned),
325 for_each_possible_cpu(cpu)
326 spin_lock_init(&per_cpu_ptr(pool->tag_cpu, cpu)->lock);
330 percpu_ida_destroy(pool);
333 EXPORT_SYMBOL_GPL(__percpu_ida_init);
336 * percpu_ida_for_each_free - iterate free ids of a pool
337 * @pool: pool to iterate
338 * @fn: interate callback function
339 * @data: parameter for @fn
341 * Note, this doesn't guarantee to iterate all free ids restrictly. Some free
342 * ids might be missed, some might be iterated duplicated, and some might
343 * be iterated and not free soon.
345 int percpu_ida_for_each_free(struct percpu_ida *pool, percpu_ida_cb fn,
349 struct percpu_ida_cpu *remote;
350 unsigned cpu, i, err = 0;
352 local_lock_irqsave(irq_off_lock, flags);
353 for_each_possible_cpu(cpu) {
354 remote = per_cpu_ptr(pool->tag_cpu, cpu);
355 spin_lock(&remote->lock);
356 for (i = 0; i < remote->nr_free; i++) {
357 err = fn(remote->freelist[i], data);
361 spin_unlock(&remote->lock);
366 spin_lock(&pool->lock);
367 for (i = 0; i < pool->nr_free; i++) {
368 err = fn(pool->freelist[i], data);
372 spin_unlock(&pool->lock);
374 local_unlock_irqrestore(irq_off_lock, flags);
377 EXPORT_SYMBOL_GPL(percpu_ida_for_each_free);
380 * percpu_ida_free_tags - return free tags number of a specific cpu or global pool
381 * @pool: pool related
382 * @cpu: specific cpu or global pool if @cpu == nr_cpu_ids
384 * Note: this just returns a snapshot of free tags number.
386 unsigned percpu_ida_free_tags(struct percpu_ida *pool, int cpu)
388 struct percpu_ida_cpu *remote;
389 if (cpu == nr_cpu_ids)
390 return pool->nr_free;
391 remote = per_cpu_ptr(pool->tag_cpu, cpu);
392 return remote->nr_free;
394 EXPORT_SYMBOL_GPL(percpu_ida_free_tags);