2 * Copyright (C) 2016 Facebook
3 * Copyright (C) 2013-2014 Jens Axboe
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public
7 * License v2 as published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program. If not, see <https://www.gnu.org/licenses/>.
18 #include <linux/sched.h>
19 #include <linux/random.h>
20 #include <linux/sbitmap.h>
21 #include <linux/seq_file.h>
23 int sbitmap_init_node(struct sbitmap
*sb
, unsigned int depth
, int shift
,
24 gfp_t flags
, int node
)
26 unsigned int bits_per_word
;
30 shift
= ilog2(BITS_PER_LONG
);
32 * If the bitmap is small, shrink the number of bits per word so
33 * we spread over a few cachelines, at least. If less than 4
34 * bits, just forget about it, it's not going to work optimally
38 while ((4U << shift
) > depth
)
42 bits_per_word
= 1U << shift
;
43 if (bits_per_word
> BITS_PER_LONG
)
48 sb
->map_nr
= DIV_ROUND_UP(sb
->depth
, bits_per_word
);
55 sb
->map
= kzalloc_node(sb
->map_nr
* sizeof(*sb
->map
), flags
, node
);
59 for (i
= 0; i
< sb
->map_nr
; i
++) {
60 sb
->map
[i
].depth
= min(depth
, bits_per_word
);
61 depth
-= sb
->map
[i
].depth
;
65 EXPORT_SYMBOL_GPL(sbitmap_init_node
);
67 void sbitmap_resize(struct sbitmap
*sb
, unsigned int depth
)
69 unsigned int bits_per_word
= 1U << sb
->shift
;
73 sb
->map_nr
= DIV_ROUND_UP(sb
->depth
, bits_per_word
);
75 for (i
= 0; i
< sb
->map_nr
; i
++) {
76 sb
->map
[i
].depth
= min(depth
, bits_per_word
);
77 depth
-= sb
->map
[i
].depth
;
80 EXPORT_SYMBOL_GPL(sbitmap_resize
);
82 static int __sbitmap_get_word(struct sbitmap_word
*word
, unsigned int hint
,
85 unsigned int orig_hint
= hint
;
89 nr
= find_next_zero_bit(&word
->word
, word
->depth
, hint
);
90 if (unlikely(nr
>= word
->depth
)) {
92 * We started with an offset, and we didn't reset the
93 * offset to 0 in a failure case, so start from 0 to
96 if (orig_hint
&& hint
&& wrap
) {
103 if (!test_and_set_bit(nr
, &word
->word
))
107 if (hint
>= word
->depth
- 1)
114 int sbitmap_get(struct sbitmap
*sb
, unsigned int alloc_hint
, bool round_robin
)
116 unsigned int i
, index
;
119 index
= SB_NR_TO_INDEX(sb
, alloc_hint
);
121 for (i
= 0; i
< sb
->map_nr
; i
++) {
122 nr
= __sbitmap_get_word(&sb
->map
[index
],
123 SB_NR_TO_BIT(sb
, alloc_hint
),
126 nr
+= index
<< sb
->shift
;
130 /* Jump to next index. */
132 alloc_hint
= index
<< sb
->shift
;
134 if (index
>= sb
->map_nr
) {
142 EXPORT_SYMBOL_GPL(sbitmap_get
);
144 bool sbitmap_any_bit_set(const struct sbitmap
*sb
)
148 for (i
= 0; i
< sb
->map_nr
; i
++) {
154 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set
);
156 bool sbitmap_any_bit_clear(const struct sbitmap
*sb
)
160 for (i
= 0; i
< sb
->map_nr
; i
++) {
161 const struct sbitmap_word
*word
= &sb
->map
[i
];
164 ret
= find_first_zero_bit(&word
->word
, word
->depth
);
165 if (ret
< word
->depth
)
170 EXPORT_SYMBOL_GPL(sbitmap_any_bit_clear
);
172 unsigned int sbitmap_weight(const struct sbitmap
*sb
)
174 unsigned int i
, weight
= 0;
176 for (i
= 0; i
< sb
->map_nr
; i
++) {
177 const struct sbitmap_word
*word
= &sb
->map
[i
];
179 weight
+= bitmap_weight(&word
->word
, word
->depth
);
183 EXPORT_SYMBOL_GPL(sbitmap_weight
);
185 void sbitmap_show(struct sbitmap
*sb
, struct seq_file
*m
)
187 seq_printf(m
, "depth=%u\n", sb
->depth
);
188 seq_printf(m
, "busy=%u\n", sbitmap_weight(sb
));
189 seq_printf(m
, "bits_per_word=%u\n", 1U << sb
->shift
);
190 seq_printf(m
, "map_nr=%u\n", sb
->map_nr
);
192 EXPORT_SYMBOL_GPL(sbitmap_show
);
194 static inline void emit_byte(struct seq_file
*m
, unsigned int offset
, u8 byte
)
196 if ((offset
& 0xf) == 0) {
199 seq_printf(m
, "%08x:", offset
);
201 if ((offset
& 0x1) == 0)
203 seq_printf(m
, "%02x", byte
);
206 void sbitmap_bitmap_show(struct sbitmap
*sb
, struct seq_file
*m
)
209 unsigned int byte_bits
= 0;
210 unsigned int offset
= 0;
213 for (i
= 0; i
< sb
->map_nr
; i
++) {
214 unsigned long word
= READ_ONCE(sb
->map
[i
].word
);
215 unsigned int word_bits
= READ_ONCE(sb
->map
[i
].depth
);
217 while (word_bits
> 0) {
218 unsigned int bits
= min(8 - byte_bits
, word_bits
);
220 byte
|= (word
& (BIT(bits
) - 1)) << byte_bits
;
222 if (byte_bits
== 8) {
223 emit_byte(m
, offset
, byte
);
233 emit_byte(m
, offset
, byte
);
239 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show
);
241 static unsigned int sbq_calc_wake_batch(unsigned int depth
)
243 unsigned int wake_batch
;
246 * For each batch, we wake up one queue. We need to make sure that our
247 * batch size is small enough that the full depth of the bitmap is
248 * enough to wake up all of the queues.
250 wake_batch
= SBQ_WAKE_BATCH
;
251 if (wake_batch
> depth
/ SBQ_WAIT_QUEUES
)
252 wake_batch
= max(1U, depth
/ SBQ_WAIT_QUEUES
);
257 int sbitmap_queue_init_node(struct sbitmap_queue
*sbq
, unsigned int depth
,
258 int shift
, bool round_robin
, gfp_t flags
, int node
)
263 ret
= sbitmap_init_node(&sbq
->sb
, depth
, shift
, flags
, node
);
267 sbq
->alloc_hint
= alloc_percpu_gfp(unsigned int, flags
);
268 if (!sbq
->alloc_hint
) {
269 sbitmap_free(&sbq
->sb
);
273 if (depth
&& !round_robin
) {
274 for_each_possible_cpu(i
)
275 *per_cpu_ptr(sbq
->alloc_hint
, i
) = prandom_u32() % depth
;
278 sbq
->wake_batch
= sbq_calc_wake_batch(depth
);
279 atomic_set(&sbq
->wake_index
, 0);
281 sbq
->ws
= kzalloc_node(SBQ_WAIT_QUEUES
* sizeof(*sbq
->ws
), flags
, node
);
283 free_percpu(sbq
->alloc_hint
);
284 sbitmap_free(&sbq
->sb
);
288 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
289 init_waitqueue_head(&sbq
->ws
[i
].wait
);
290 atomic_set(&sbq
->ws
[i
].wait_cnt
, sbq
->wake_batch
);
293 sbq
->round_robin
= round_robin
;
296 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node
);
298 void sbitmap_queue_resize(struct sbitmap_queue
*sbq
, unsigned int depth
)
300 unsigned int wake_batch
= sbq_calc_wake_batch(depth
);
303 if (sbq
->wake_batch
!= wake_batch
) {
304 WRITE_ONCE(sbq
->wake_batch
, wake_batch
);
306 * Pairs with the memory barrier in sbq_wake_up() to ensure that
307 * the batch size is updated before the wait counts.
309 smp_mb__before_atomic();
310 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++)
311 atomic_set(&sbq
->ws
[i
].wait_cnt
, 1);
313 sbitmap_resize(&sbq
->sb
, depth
);
315 EXPORT_SYMBOL_GPL(sbitmap_queue_resize
);
317 int __sbitmap_queue_get(struct sbitmap_queue
*sbq
)
319 unsigned int hint
, depth
;
322 hint
= this_cpu_read(*sbq
->alloc_hint
);
323 depth
= READ_ONCE(sbq
->sb
.depth
);
324 if (unlikely(hint
>= depth
)) {
325 hint
= depth
? prandom_u32() % depth
: 0;
326 this_cpu_write(*sbq
->alloc_hint
, hint
);
328 nr
= sbitmap_get(&sbq
->sb
, hint
, sbq
->round_robin
);
331 /* If the map is full, a hint won't do us much good. */
332 this_cpu_write(*sbq
->alloc_hint
, 0);
333 } else if (nr
== hint
|| unlikely(sbq
->round_robin
)) {
334 /* Only update the hint if we used it. */
336 if (hint
>= depth
- 1)
338 this_cpu_write(*sbq
->alloc_hint
, hint
);
343 EXPORT_SYMBOL_GPL(__sbitmap_queue_get
);
345 static struct sbq_wait_state
*sbq_wake_ptr(struct sbitmap_queue
*sbq
)
349 wake_index
= atomic_read(&sbq
->wake_index
);
350 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
351 struct sbq_wait_state
*ws
= &sbq
->ws
[wake_index
];
353 if (waitqueue_active(&ws
->wait
)) {
354 int o
= atomic_read(&sbq
->wake_index
);
357 atomic_cmpxchg(&sbq
->wake_index
, o
, wake_index
);
361 wake_index
= sbq_index_inc(wake_index
);
367 static void sbq_wake_up(struct sbitmap_queue
*sbq
)
369 struct sbq_wait_state
*ws
;
370 unsigned int wake_batch
;
374 * Pairs with the memory barrier in set_current_state() to ensure the
375 * proper ordering of clear_bit()/waitqueue_active() in the waker and
376 * test_and_set_bit()/prepare_to_wait()/finish_wait() in the waiter. See
377 * the comment on waitqueue_active(). This is __after_atomic because we
378 * just did clear_bit() in the caller.
380 smp_mb__after_atomic();
382 ws
= sbq_wake_ptr(sbq
);
386 wait_cnt
= atomic_dec_return(&ws
->wait_cnt
);
388 wake_batch
= READ_ONCE(sbq
->wake_batch
);
390 * Pairs with the memory barrier in sbitmap_queue_resize() to
391 * ensure that we see the batch size update before the wait
394 smp_mb__before_atomic();
396 * If there are concurrent callers to sbq_wake_up(), the last
397 * one to decrement the wait count below zero will bump it back
398 * up. If there is a concurrent resize, the count reset will
399 * either cause the cmpxchg to fail or overwrite after the
402 atomic_cmpxchg(&ws
->wait_cnt
, wait_cnt
, wait_cnt
+ wake_batch
);
403 sbq_index_atomic_inc(&sbq
->wake_index
);
408 void sbitmap_queue_clear(struct sbitmap_queue
*sbq
, unsigned int nr
,
411 sbitmap_clear_bit(&sbq
->sb
, nr
);
413 if (likely(!sbq
->round_robin
&& nr
< sbq
->sb
.depth
))
414 *per_cpu_ptr(sbq
->alloc_hint
, cpu
) = nr
;
416 EXPORT_SYMBOL_GPL(sbitmap_queue_clear
);
418 void sbitmap_queue_wake_all(struct sbitmap_queue
*sbq
)
423 * Pairs with the memory barrier in set_current_state() like in
427 wake_index
= atomic_read(&sbq
->wake_index
);
428 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
429 struct sbq_wait_state
*ws
= &sbq
->ws
[wake_index
];
431 if (waitqueue_active(&ws
->wait
))
434 wake_index
= sbq_index_inc(wake_index
);
437 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all
);
439 void sbitmap_queue_show(struct sbitmap_queue
*sbq
, struct seq_file
*m
)
444 sbitmap_show(&sbq
->sb
, m
);
446 seq_puts(m
, "alloc_hint={");
448 for_each_possible_cpu(i
) {
452 seq_printf(m
, "%u", *per_cpu_ptr(sbq
->alloc_hint
, i
));
456 seq_printf(m
, "wake_batch=%u\n", sbq
->wake_batch
);
457 seq_printf(m
, "wake_index=%d\n", atomic_read(&sbq
->wake_index
));
459 seq_puts(m
, "ws={\n");
460 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
461 struct sbq_wait_state
*ws
= &sbq
->ws
[i
];
463 seq_printf(m
, "\t{.wait_cnt=%d, .wait=%s},\n",
464 atomic_read(&ws
->wait_cnt
),
465 waitqueue_active(&ws
->wait
) ? "active" : "inactive");
469 seq_printf(m
, "round_robin=%d\n", sbq
->round_robin
);
471 EXPORT_SYMBOL_GPL(sbitmap_queue_show
);