1 /* SPDX-License-Identifier: GPL-2.0-only */
3 * Fast and scalable bitmaps.
5 * Copyright (C) 2016 Facebook
6 * Copyright (C) 2013-2014 Jens Axboe
9 #ifndef __LINUX_SCALE_BITMAP_H
10 #define __LINUX_SCALE_BITMAP_H
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
18 * struct sbitmap_word - Word in a &struct sbitmap.
22 * @depth: Number of bits being used in @word/@cleared
27 * @word: word holding free bits
29 unsigned long word ____cacheline_aligned_in_smp
;
32 * @cleared: word holding cleared bits
34 unsigned long cleared ____cacheline_aligned_in_smp
;
35 } ____cacheline_aligned_in_smp
;
38 * struct sbitmap - Scalable bitmap.
40 * A &struct sbitmap is spread over multiple cachelines to avoid ping-pong. This
41 * trades off higher memory usage for better scalability.
45 * @depth: Number of bits used in the whole bitmap.
50 * @shift: log2(number of bits used per word)
55 * @map_nr: Number of words (cachelines) being used for the bitmap.
60 * @round_robin: Allocate bits in strict round-robin order.
65 * @map: Allocated bitmap.
67 struct sbitmap_word
*map
;
70 * @alloc_hint: Cache of last successfully allocated or freed bit.
72 * This is per-cpu, which allows multiple users to stick to different
73 * cachelines until the map is exhausted.
75 unsigned int __percpu
*alloc_hint
;
78 #define SBQ_WAIT_QUEUES 8
79 #define SBQ_WAKE_BATCH 8
82 * struct sbq_wait_state - Wait queue in a &struct sbitmap_queue.
84 struct sbq_wait_state
{
86 * @wait_cnt: Number of frees remaining before we wake up.
93 wait_queue_head_t wait
;
94 } ____cacheline_aligned_in_smp
;
97 * struct sbitmap_queue - Scalable bitmap with the added ability to wait on free
100 * A &struct sbitmap_queue uses multiple wait queues and rolling wakeups to
101 * avoid contention on the wait queue spinlock. This ensures that we don't hit a
102 * scalability wall when we run out of free bits and have to start putting tasks
105 struct sbitmap_queue
{
107 * @sb: Scalable bitmap.
112 * @wake_batch: Number of bits which must be freed before we wake up any
115 unsigned int wake_batch
;
118 * @wake_index: Next wait queue in @ws to wake up.
125 struct sbq_wait_state
*ws
;
128 * @ws_active: count of currently active ws waitqueues
133 * @min_shallow_depth: The minimum shallow depth which may be passed to
134 * sbitmap_queue_get_shallow() or __sbitmap_queue_get_shallow().
136 unsigned int min_shallow_depth
;
140 * sbitmap_init_node() - Initialize a &struct sbitmap on a specific memory node.
141 * @sb: Bitmap to initialize.
142 * @depth: Number of bits to allocate.
143 * @shift: Use 2^@shift bits per word in the bitmap; if a negative number if
144 * given, a good default is chosen.
145 * @flags: Allocation flags.
146 * @node: Memory node to allocate on.
147 * @round_robin: If true, be stricter about allocation order; always allocate
148 * starting from the last allocated bit. This is less efficient
149 * than the default behavior (false).
150 * @alloc_hint: If true, apply percpu hint for where to start searching for
153 * Return: Zero on success or negative errno on failure.
155 int sbitmap_init_node(struct sbitmap
*sb
, unsigned int depth
, int shift
,
156 gfp_t flags
, int node
, bool round_robin
, bool alloc_hint
);
159 * sbitmap_free() - Free memory used by a &struct sbitmap.
160 * @sb: Bitmap to free.
162 static inline void sbitmap_free(struct sbitmap
*sb
)
164 free_percpu(sb
->alloc_hint
);
170 * sbitmap_resize() - Resize a &struct sbitmap.
171 * @sb: Bitmap to resize.
172 * @depth: New number of bits to resize to.
174 * Doesn't reallocate anything. It's up to the caller to ensure that the new
175 * depth doesn't exceed the depth that the sb was initialized with.
177 void sbitmap_resize(struct sbitmap
*sb
, unsigned int depth
);
180 * sbitmap_get() - Try to allocate a free bit from a &struct sbitmap.
181 * @sb: Bitmap to allocate from.
183 * This operation provides acquire barrier semantics if it succeeds.
185 * Return: Non-negative allocated bit number if successful, -1 otherwise.
187 int sbitmap_get(struct sbitmap
*sb
);
190 * sbitmap_get_shallow() - Try to allocate a free bit from a &struct sbitmap,
191 * limiting the depth used from each word.
192 * @sb: Bitmap to allocate from.
193 * @shallow_depth: The maximum number of bits to allocate from a single word.
195 * This rather specific operation allows for having multiple users with
196 * different allocation limits. E.g., there can be a high-priority class that
197 * uses sbitmap_get() and a low-priority class that uses sbitmap_get_shallow()
198 * with a @shallow_depth of (1 << (@sb->shift - 1)). Then, the low-priority
199 * class can only allocate half of the total bits in the bitmap, preventing it
200 * from starving out the high-priority class.
202 * Return: Non-negative allocated bit number if successful, -1 otherwise.
204 int sbitmap_get_shallow(struct sbitmap
*sb
, unsigned long shallow_depth
);
207 * sbitmap_any_bit_set() - Check for a set bit in a &struct sbitmap.
208 * @sb: Bitmap to check.
210 * Return: true if any bit in the bitmap is set, false otherwise.
212 bool sbitmap_any_bit_set(const struct sbitmap
*sb
);
214 #define SB_NR_TO_INDEX(sb, bitnr) ((bitnr) >> (sb)->shift)
215 #define SB_NR_TO_BIT(sb, bitnr) ((bitnr) & ((1U << (sb)->shift) - 1U))
217 typedef bool (*sb_for_each_fn
)(struct sbitmap
*, unsigned int, void *);
220 * __sbitmap_for_each_set() - Iterate over each set bit in a &struct sbitmap.
221 * @start: Where to start the iteration.
222 * @sb: Bitmap to iterate over.
223 * @fn: Callback. Should return true to continue or false to break early.
224 * @data: Pointer to pass to callback.
226 * This is inline even though it's non-trivial so that the function calls to the
227 * callback will hopefully get optimized away.
229 static inline void __sbitmap_for_each_set(struct sbitmap
*sb
,
231 sb_for_each_fn fn
, void *data
)
235 unsigned int scanned
= 0;
237 if (start
>= sb
->depth
)
239 index
= SB_NR_TO_INDEX(sb
, start
);
240 nr
= SB_NR_TO_BIT(sb
, start
);
242 while (scanned
< sb
->depth
) {
244 unsigned int depth
= min_t(unsigned int,
245 sb
->map
[index
].depth
- nr
,
246 sb
->depth
- scanned
);
249 word
= sb
->map
[index
].word
& ~sb
->map
[index
].cleared
;
254 * On the first iteration of the outer loop, we need to add the
255 * bit offset back to the size of the word for find_next_bit().
256 * On all other iterations, nr is zero, so this is a noop.
260 nr
= find_next_bit(&word
, depth
, nr
);
263 if (!fn(sb
, (index
<< sb
->shift
) + nr
, data
))
270 if (++index
>= sb
->map_nr
)
276 * sbitmap_for_each_set() - Iterate over each set bit in a &struct sbitmap.
277 * @sb: Bitmap to iterate over.
278 * @fn: Callback. Should return true to continue or false to break early.
279 * @data: Pointer to pass to callback.
281 static inline void sbitmap_for_each_set(struct sbitmap
*sb
, sb_for_each_fn fn
,
284 __sbitmap_for_each_set(sb
, 0, fn
, data
);
287 static inline unsigned long *__sbitmap_word(struct sbitmap
*sb
,
290 return &sb
->map
[SB_NR_TO_INDEX(sb
, bitnr
)].word
;
293 /* Helpers equivalent to the operations in asm/bitops.h and linux/bitmap.h */
295 static inline void sbitmap_set_bit(struct sbitmap
*sb
, unsigned int bitnr
)
297 set_bit(SB_NR_TO_BIT(sb
, bitnr
), __sbitmap_word(sb
, bitnr
));
300 static inline void sbitmap_clear_bit(struct sbitmap
*sb
, unsigned int bitnr
)
302 clear_bit(SB_NR_TO_BIT(sb
, bitnr
), __sbitmap_word(sb
, bitnr
));
306 * This one is special, since it doesn't actually clear the bit, rather it
307 * sets the corresponding bit in the ->cleared mask instead. Paired with
308 * the caller doing sbitmap_deferred_clear() if a given index is full, which
309 * will clear the previously freed entries in the corresponding ->word.
311 static inline void sbitmap_deferred_clear_bit(struct sbitmap
*sb
, unsigned int bitnr
)
313 unsigned long *addr
= &sb
->map
[SB_NR_TO_INDEX(sb
, bitnr
)].cleared
;
315 set_bit(SB_NR_TO_BIT(sb
, bitnr
), addr
);
319 * Pair of sbitmap_get, and this one applies both cleared bit and
322 static inline void sbitmap_put(struct sbitmap
*sb
, unsigned int bitnr
)
324 sbitmap_deferred_clear_bit(sb
, bitnr
);
326 if (likely(sb
->alloc_hint
&& !sb
->round_robin
&& bitnr
< sb
->depth
))
327 *raw_cpu_ptr(sb
->alloc_hint
) = bitnr
;
330 static inline int sbitmap_test_bit(struct sbitmap
*sb
, unsigned int bitnr
)
332 return test_bit(SB_NR_TO_BIT(sb
, bitnr
), __sbitmap_word(sb
, bitnr
));
335 static inline int sbitmap_calculate_shift(unsigned int depth
)
337 int shift
= ilog2(BITS_PER_LONG
);
340 * If the bitmap is small, shrink the number of bits per word so
341 * we spread over a few cachelines, at least. If less than 4
342 * bits, just forget about it, it's not going to work optimally
346 while ((4U << shift
) > depth
)
354 * sbitmap_show() - Dump &struct sbitmap information to a &struct seq_file.
355 * @sb: Bitmap to show.
356 * @m: struct seq_file to write to.
358 * This is intended for debugging. The format may change at any time.
360 void sbitmap_show(struct sbitmap
*sb
, struct seq_file
*m
);
364 * sbitmap_weight() - Return how many set and not cleared bits in a &struct
366 * @sb: Bitmap to check.
368 * Return: How many set and not cleared bits set
370 unsigned int sbitmap_weight(const struct sbitmap
*sb
);
373 * sbitmap_bitmap_show() - Write a hex dump of a &struct sbitmap to a &struct
375 * @sb: Bitmap to show.
376 * @m: struct seq_file to write to.
378 * This is intended for debugging. The output isn't guaranteed to be internally
381 void sbitmap_bitmap_show(struct sbitmap
*sb
, struct seq_file
*m
);
384 * sbitmap_queue_init_node() - Initialize a &struct sbitmap_queue on a specific
386 * @sbq: Bitmap queue to initialize.
387 * @depth: See sbitmap_init_node().
388 * @shift: See sbitmap_init_node().
389 * @round_robin: See sbitmap_get().
390 * @flags: Allocation flags.
391 * @node: Memory node to allocate on.
393 * Return: Zero on success or negative errno on failure.
395 int sbitmap_queue_init_node(struct sbitmap_queue
*sbq
, unsigned int depth
,
396 int shift
, bool round_robin
, gfp_t flags
, int node
);
399 * sbitmap_queue_free() - Free memory used by a &struct sbitmap_queue.
401 * @sbq: Bitmap queue to free.
403 static inline void sbitmap_queue_free(struct sbitmap_queue
*sbq
)
406 sbitmap_free(&sbq
->sb
);
410 * sbitmap_queue_resize() - Resize a &struct sbitmap_queue.
411 * @sbq: Bitmap queue to resize.
412 * @depth: New number of bits to resize to.
414 * Like sbitmap_resize(), this doesn't reallocate anything. It has to do
415 * some extra work on the &struct sbitmap_queue, so it's not safe to just
416 * resize the underlying &struct sbitmap.
418 void sbitmap_queue_resize(struct sbitmap_queue
*sbq
, unsigned int depth
);
421 * __sbitmap_queue_get() - Try to allocate a free bit from a &struct
422 * sbitmap_queue with preemption already disabled.
423 * @sbq: Bitmap queue to allocate from.
425 * Return: Non-negative allocated bit number if successful, -1 otherwise.
427 int __sbitmap_queue_get(struct sbitmap_queue
*sbq
);
430 * __sbitmap_queue_get_shallow() - Try to allocate a free bit from a &struct
431 * sbitmap_queue, limiting the depth used from each word, with preemption
433 * @sbq: Bitmap queue to allocate from.
434 * @shallow_depth: The maximum number of bits to allocate from a single word.
435 * See sbitmap_get_shallow().
437 * If you call this, make sure to call sbitmap_queue_min_shallow_depth() after
440 * Return: Non-negative allocated bit number if successful, -1 otherwise.
442 int __sbitmap_queue_get_shallow(struct sbitmap_queue
*sbq
,
443 unsigned int shallow_depth
);
446 * sbitmap_queue_get() - Try to allocate a free bit from a &struct
448 * @sbq: Bitmap queue to allocate from.
449 * @cpu: Output parameter; will contain the CPU we ran on (e.g., to be passed to
450 * sbitmap_queue_clear()).
452 * Return: Non-negative allocated bit number if successful, -1 otherwise.
454 static inline int sbitmap_queue_get(struct sbitmap_queue
*sbq
,
460 nr
= __sbitmap_queue_get(sbq
);
466 * sbitmap_queue_get_shallow() - Try to allocate a free bit from a &struct
467 * sbitmap_queue, limiting the depth used from each word.
468 * @sbq: Bitmap queue to allocate from.
469 * @cpu: Output parameter; will contain the CPU we ran on (e.g., to be passed to
470 * sbitmap_queue_clear()).
471 * @shallow_depth: The maximum number of bits to allocate from a single word.
472 * See sbitmap_get_shallow().
474 * If you call this, make sure to call sbitmap_queue_min_shallow_depth() after
477 * Return: Non-negative allocated bit number if successful, -1 otherwise.
479 static inline int sbitmap_queue_get_shallow(struct sbitmap_queue
*sbq
,
481 unsigned int shallow_depth
)
486 nr
= __sbitmap_queue_get_shallow(sbq
, shallow_depth
);
492 * sbitmap_queue_min_shallow_depth() - Inform a &struct sbitmap_queue of the
493 * minimum shallow depth that will be used.
494 * @sbq: Bitmap queue in question.
495 * @min_shallow_depth: The minimum shallow depth that will be passed to
496 * sbitmap_queue_get_shallow() or __sbitmap_queue_get_shallow().
498 * sbitmap_queue_clear() batches wakeups as an optimization. The batch size
499 * depends on the depth of the bitmap. Since the shallow allocation functions
500 * effectively operate with a different depth, the shallow depth must be taken
501 * into account when calculating the batch size. This function must be called
502 * with the minimum shallow depth that will be used. Failure to do so can result
505 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue
*sbq
,
506 unsigned int min_shallow_depth
);
509 * sbitmap_queue_clear() - Free an allocated bit and wake up waiters on a
510 * &struct sbitmap_queue.
511 * @sbq: Bitmap to free from.
512 * @nr: Bit number to free.
513 * @cpu: CPU the bit was allocated on.
515 void sbitmap_queue_clear(struct sbitmap_queue
*sbq
, unsigned int nr
,
518 static inline int sbq_index_inc(int index
)
520 return (index
+ 1) & (SBQ_WAIT_QUEUES
- 1);
523 static inline void sbq_index_atomic_inc(atomic_t
*index
)
525 int old
= atomic_read(index
);
526 int new = sbq_index_inc(old
);
527 atomic_cmpxchg(index
, old
, new);
531 * sbq_wait_ptr() - Get the next wait queue to use for a &struct
533 * @sbq: Bitmap queue to wait on.
534 * @wait_index: A counter per "user" of @sbq.
536 static inline struct sbq_wait_state
*sbq_wait_ptr(struct sbitmap_queue
*sbq
,
537 atomic_t
*wait_index
)
539 struct sbq_wait_state
*ws
;
541 ws
= &sbq
->ws
[atomic_read(wait_index
)];
542 sbq_index_atomic_inc(wait_index
);
547 * sbitmap_queue_wake_all() - Wake up everything waiting on a &struct
549 * @sbq: Bitmap queue to wake up.
551 void sbitmap_queue_wake_all(struct sbitmap_queue
*sbq
);
554 * sbitmap_queue_wake_up() - Wake up some of waiters in one waitqueue
555 * on a &struct sbitmap_queue.
556 * @sbq: Bitmap queue to wake up.
558 void sbitmap_queue_wake_up(struct sbitmap_queue
*sbq
);
561 * sbitmap_queue_show() - Dump &struct sbitmap_queue information to a &struct
563 * @sbq: Bitmap queue to show.
564 * @m: struct seq_file to write to.
566 * This is intended for debugging. The format may change at any time.
568 void sbitmap_queue_show(struct sbitmap_queue
*sbq
, struct seq_file
*m
);
571 struct sbitmap_queue
*sbq
; /* if set, sbq_wait is accounted */
572 struct wait_queue_entry wait
;
575 #define DEFINE_SBQ_WAIT(name) \
576 struct sbq_wait name = { \
579 .private = current, \
580 .func = autoremove_wake_function, \
581 .entry = LIST_HEAD_INIT((name).wait.entry), \
586 * Wrapper around prepare_to_wait_exclusive(), which maintains some extra
589 void sbitmap_prepare_to_wait(struct sbitmap_queue
*sbq
,
590 struct sbq_wait_state
*ws
,
591 struct sbq_wait
*sbq_wait
, int state
);
594 * Must be paired with sbitmap_prepare_to_wait().
596 void sbitmap_finish_wait(struct sbitmap_queue
*sbq
, struct sbq_wait_state
*ws
,
597 struct sbq_wait
*sbq_wait
);
600 * Wrapper around add_wait_queue(), which maintains some extra internal state
602 void sbitmap_add_wait_queue(struct sbitmap_queue
*sbq
,
603 struct sbq_wait_state
*ws
,
604 struct sbq_wait
*sbq_wait
);
607 * Must be paired with sbitmap_add_wait_queue()
609 void sbitmap_del_wait_queue(struct sbq_wait
*sbq_wait
);
611 #endif /* __LINUX_SCALE_BITMAP_H */