blob: eff4e42c425a4b150e13cf1aefbab66f22f6ca28 [file] [log] [blame]
Thomas Gleixner0fc479b2019-05-29 16:57:42 -07001// SPDX-License-Identifier: GPL-2.0-only
Omar Sandoval88459642016-09-17 08:38:44 -06002/*
3 * Copyright (C) 2016 Facebook
4 * Copyright (C) 2013-2014 Jens Axboe
Omar Sandoval88459642016-09-17 08:38:44 -06005 */
6
Ingo Molnaraf8601a2017-02-03 09:57:00 +01007#include <linux/sched.h>
Omar Sandoval98d95412016-09-17 01:28:25 -07008#include <linux/random.h>
Omar Sandoval88459642016-09-17 08:38:44 -06009#include <linux/sbitmap.h>
Omar Sandoval24af1ccf2017-01-25 14:32:13 -080010#include <linux/seq_file.h>
Omar Sandoval88459642016-09-17 08:38:44 -060011
Ming Leic548e622021-01-22 10:33:08 +080012static int init_alloc_hint(struct sbitmap *sb, gfp_t flags)
Ming Leibf2c4282021-01-22 10:33:07 +080013{
Ming Leic548e622021-01-22 10:33:08 +080014 unsigned depth = sb->depth;
Ming Leibf2c4282021-01-22 10:33:07 +080015
Ming Leic548e622021-01-22 10:33:08 +080016 sb->alloc_hint = alloc_percpu_gfp(unsigned int, flags);
17 if (!sb->alloc_hint)
Ming Leibf2c4282021-01-22 10:33:07 +080018 return -ENOMEM;
19
Ming Leic548e622021-01-22 10:33:08 +080020 if (depth && !sb->round_robin) {
Ming Leibf2c4282021-01-22 10:33:07 +080021 int i;
22
23 for_each_possible_cpu(i)
Jason A. Donenfeld8032bf12022-10-09 20:44:02 -060024 *per_cpu_ptr(sb->alloc_hint, i) = get_random_u32_below(depth);
Ming Leibf2c4282021-01-22 10:33:07 +080025 }
Ming Leibf2c4282021-01-22 10:33:07 +080026 return 0;
27}
28
Ming Leic548e622021-01-22 10:33:08 +080029static inline unsigned update_alloc_hint_before_get(struct sbitmap *sb,
Ming Leibf2c4282021-01-22 10:33:07 +080030 unsigned int depth)
31{
32 unsigned hint;
33
Ming Leic548e622021-01-22 10:33:08 +080034 hint = this_cpu_read(*sb->alloc_hint);
Ming Leibf2c4282021-01-22 10:33:07 +080035 if (unlikely(hint >= depth)) {
Jason A. Donenfeld8032bf12022-10-09 20:44:02 -060036 hint = depth ? get_random_u32_below(depth) : 0;
Ming Leic548e622021-01-22 10:33:08 +080037 this_cpu_write(*sb->alloc_hint, hint);
Ming Leibf2c4282021-01-22 10:33:07 +080038 }
39
40 return hint;
41}
42
Ming Leic548e622021-01-22 10:33:08 +080043static inline void update_alloc_hint_after_get(struct sbitmap *sb,
Ming Leibf2c4282021-01-22 10:33:07 +080044 unsigned int depth,
45 unsigned int hint,
46 unsigned int nr)
47{
48 if (nr == -1) {
49 /* If the map is full, a hint won't do us much good. */
Ming Leic548e622021-01-22 10:33:08 +080050 this_cpu_write(*sb->alloc_hint, 0);
51 } else if (nr == hint || unlikely(sb->round_robin)) {
Ming Leibf2c4282021-01-22 10:33:07 +080052 /* Only update the hint if we used it. */
53 hint = nr + 1;
54 if (hint >= depth - 1)
55 hint = 0;
Ming Leic548e622021-01-22 10:33:08 +080056 this_cpu_write(*sb->alloc_hint, hint);
Ming Leibf2c4282021-01-22 10:33:07 +080057 }
58}
59
Jens Axboeb2dbff12018-12-11 18:39:41 -070060/*
61 * See if we have deferred clears that we can batch move
62 */
Pavel Begunkovb78beea2020-11-22 15:35:45 +000063static inline bool sbitmap_deferred_clear(struct sbitmap_word *map)
Jens Axboeb2dbff12018-12-11 18:39:41 -070064{
Pavel Begunkovc3250c82020-11-22 15:35:47 +000065 unsigned long mask;
Jens Axboeb2dbff12018-12-11 18:39:41 -070066
Pavel Begunkov661d4f52020-11-22 15:35:46 +000067 if (!READ_ONCE(map->cleared))
68 return false;
Jens Axboeb2dbff12018-12-11 18:39:41 -070069
70 /*
71 * First get a stable cleared mask, setting the old mask to 0.
72 */
Pavel Begunkovb78beea2020-11-22 15:35:45 +000073 mask = xchg(&map->cleared, 0);
Jens Axboeb2dbff12018-12-11 18:39:41 -070074
75 /*
76 * Now clear the masked bits in our free word
77 */
Pavel Begunkovc3250c82020-11-22 15:35:47 +000078 atomic_long_andnot(mask, (atomic_long_t *)&map->word);
79 BUILD_BUG_ON(sizeof(atomic_long_t) != sizeof(map->word));
Pavel Begunkov661d4f52020-11-22 15:35:46 +000080 return true;
Jens Axboeb2dbff12018-12-11 18:39:41 -070081}
82
Omar Sandoval88459642016-09-17 08:38:44 -060083int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
Ming Leic548e622021-01-22 10:33:08 +080084 gfp_t flags, int node, bool round_robin,
85 bool alloc_hint)
Omar Sandoval88459642016-09-17 08:38:44 -060086{
87 unsigned int bits_per_word;
Omar Sandoval88459642016-09-17 08:38:44 -060088
Ming Lei2d13b1e2021-01-22 10:33:10 +080089 if (shift < 0)
90 shift = sbitmap_calculate_shift(depth);
91
Omar Sandoval88459642016-09-17 08:38:44 -060092 bits_per_word = 1U << shift;
93 if (bits_per_word > BITS_PER_LONG)
94 return -EINVAL;
95
96 sb->shift = shift;
97 sb->depth = depth;
98 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
Ming Leiefe1f3a2021-01-22 10:33:06 +080099 sb->round_robin = round_robin;
Omar Sandoval88459642016-09-17 08:38:44 -0600100
101 if (depth == 0) {
102 sb->map = NULL;
103 return 0;
104 }
105
Ming Leic548e622021-01-22 10:33:08 +0800106 if (alloc_hint) {
107 if (init_alloc_hint(sb, flags))
108 return -ENOMEM;
109 } else {
110 sb->alloc_hint = NULL;
111 }
112
Ming Lei863a66c2022-03-16 09:27:08 +0800113 sb->map = kvzalloc_node(sb->map_nr * sizeof(*sb->map), flags, node);
Ming Leic548e622021-01-22 10:33:08 +0800114 if (!sb->map) {
115 free_percpu(sb->alloc_hint);
Omar Sandoval88459642016-09-17 08:38:44 -0600116 return -ENOMEM;
Ming Leic548e622021-01-22 10:33:08 +0800117 }
Omar Sandoval88459642016-09-17 08:38:44 -0600118
Omar Sandoval88459642016-09-17 08:38:44 -0600119 return 0;
120}
121EXPORT_SYMBOL_GPL(sbitmap_init_node);
122
123void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
124{
125 unsigned int bits_per_word = 1U << sb->shift;
126 unsigned int i;
127
Jens Axboeb2dbff12018-12-11 18:39:41 -0700128 for (i = 0; i < sb->map_nr; i++)
Pavel Begunkovb78beea2020-11-22 15:35:45 +0000129 sbitmap_deferred_clear(&sb->map[i]);
Jens Axboeb2dbff12018-12-11 18:39:41 -0700130
Omar Sandoval88459642016-09-17 08:38:44 -0600131 sb->depth = depth;
132 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
Omar Sandoval88459642016-09-17 08:38:44 -0600133}
134EXPORT_SYMBOL_GPL(sbitmap_resize);
135
Omar Sandovalc05e6672017-04-14 00:59:58 -0700136static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
137 unsigned int hint, bool wrap)
Omar Sandoval88459642016-09-17 08:38:44 -0600138{
Omar Sandoval88459642016-09-17 08:38:44 -0600139 int nr;
140
Pavel Begunkov0eff1f12020-11-22 15:35:48 +0000141 /* don't wrap if starting from 0 */
142 wrap = wrap && hint;
143
Omar Sandoval88459642016-09-17 08:38:44 -0600144 while (1) {
Omar Sandovalc05e6672017-04-14 00:59:58 -0700145 nr = find_next_zero_bit(word, depth, hint);
146 if (unlikely(nr >= depth)) {
Omar Sandoval88459642016-09-17 08:38:44 -0600147 /*
148 * We started with an offset, and we didn't reset the
149 * offset to 0 in a failure case, so start from 0 to
150 * exhaust the map.
151 */
Pavel Begunkov0eff1f12020-11-22 15:35:48 +0000152 if (hint && wrap) {
153 hint = 0;
Omar Sandoval88459642016-09-17 08:38:44 -0600154 continue;
155 }
156 return -1;
157 }
158
Omar Sandoval4ace53f2018-02-27 16:56:43 -0800159 if (!test_and_set_bit_lock(nr, word))
Omar Sandoval88459642016-09-17 08:38:44 -0600160 break;
161
162 hint = nr + 1;
Omar Sandovalc05e6672017-04-14 00:59:58 -0700163 if (hint >= depth - 1)
Omar Sandoval88459642016-09-17 08:38:44 -0600164 hint = 0;
165 }
166
167 return nr;
168}
169
Kemeng Shi08470a92023-01-17 04:50:57 +0800170static int sbitmap_find_bit_in_word(struct sbitmap_word *map,
171 unsigned int depth,
172 unsigned int alloc_hint,
173 bool wrap)
Jens Axboeea86ea22018-11-30 13:18:06 -0700174{
175 int nr;
176
177 do {
Kemeng Shi08470a92023-01-17 04:50:57 +0800178 nr = __sbitmap_get_word(&map->word, depth,
179 alloc_hint, wrap);
Jens Axboeea86ea22018-11-30 13:18:06 -0700180 if (nr != -1)
181 break;
Pavel Begunkovb78beea2020-11-22 15:35:45 +0000182 if (!sbitmap_deferred_clear(map))
Jens Axboeea86ea22018-11-30 13:18:06 -0700183 break;
184 } while (1);
185
186 return nr;
187}
188
Kemeng Shi678418c2023-01-17 04:50:58 +0800189static int sbitmap_find_bit(struct sbitmap *sb,
190 unsigned int depth,
191 unsigned int index,
192 unsigned int alloc_hint,
193 bool wrap)
194{
195 unsigned int i;
196 int nr = -1;
197
198 for (i = 0; i < sb->map_nr; i++) {
199 nr = sbitmap_find_bit_in_word(&sb->map[index],
200 min_t(unsigned int,
201 __map_depth(sb, index),
202 depth),
203 alloc_hint, wrap);
204
205 if (nr != -1) {
206 nr += index << sb->shift;
207 break;
208 }
209
210 /* Jump to next index. */
211 alloc_hint = 0;
212 if (++index >= sb->map_nr)
213 index = 0;
214 }
215
216 return nr;
217}
218
Ming Leic548e622021-01-22 10:33:08 +0800219static int __sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint)
Omar Sandoval88459642016-09-17 08:38:44 -0600220{
Kemeng Shi678418c2023-01-17 04:50:58 +0800221 unsigned int index;
Omar Sandoval88459642016-09-17 08:38:44 -0600222
223 index = SB_NR_TO_INDEX(sb, alloc_hint);
224
Jens Axboe27fae422018-11-29 12:35:16 -0700225 /*
226 * Unless we're doing round robin tag allocation, just use the
227 * alloc_hint to find the right word index. No point in looping
228 * twice in find_next_zero_bit() for that case.
229 */
Ming Leiefe1f3a2021-01-22 10:33:06 +0800230 if (sb->round_robin)
Jens Axboe27fae422018-11-29 12:35:16 -0700231 alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
232 else
233 alloc_hint = 0;
234
Kemeng Shi678418c2023-01-17 04:50:58 +0800235 return sbitmap_find_bit(sb, UINT_MAX, index, alloc_hint,
236 !sb->round_robin);
Omar Sandoval88459642016-09-17 08:38:44 -0600237}
Ming Leic548e622021-01-22 10:33:08 +0800238
239int sbitmap_get(struct sbitmap *sb)
240{
241 int nr;
242 unsigned int hint, depth;
243
244 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
245 return -1;
246
247 depth = READ_ONCE(sb->depth);
248 hint = update_alloc_hint_before_get(sb, depth);
249 nr = __sbitmap_get(sb, hint);
250 update_alloc_hint_after_get(sb, depth, hint, nr);
251
252 return nr;
253}
Omar Sandoval88459642016-09-17 08:38:44 -0600254EXPORT_SYMBOL_GPL(sbitmap_get);
255
Ming Leic548e622021-01-22 10:33:08 +0800256static int __sbitmap_get_shallow(struct sbitmap *sb,
257 unsigned int alloc_hint,
258 unsigned long shallow_depth)
Omar Sandovalc05e6672017-04-14 00:59:58 -0700259{
Kemeng Shi678418c2023-01-17 04:50:58 +0800260 unsigned int index;
Omar Sandovalc05e6672017-04-14 00:59:58 -0700261
262 index = SB_NR_TO_INDEX(sb, alloc_hint);
Kemeng Shif1591a82023-01-17 04:50:55 +0800263 alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
Omar Sandovalc05e6672017-04-14 00:59:58 -0700264
Kemeng Shi678418c2023-01-17 04:50:58 +0800265 return sbitmap_find_bit(sb, shallow_depth, index, alloc_hint, true);
Omar Sandovalc05e6672017-04-14 00:59:58 -0700266}
Ming Leic548e622021-01-22 10:33:08 +0800267
268int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth)
269{
270 int nr;
271 unsigned int hint, depth;
272
273 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
274 return -1;
275
276 depth = READ_ONCE(sb->depth);
277 hint = update_alloc_hint_before_get(sb, depth);
278 nr = __sbitmap_get_shallow(sb, hint, shallow_depth);
279 update_alloc_hint_after_get(sb, depth, hint, nr);
280
281 return nr;
282}
Omar Sandovalc05e6672017-04-14 00:59:58 -0700283EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
284
Omar Sandoval88459642016-09-17 08:38:44 -0600285bool sbitmap_any_bit_set(const struct sbitmap *sb)
286{
287 unsigned int i;
288
289 for (i = 0; i < sb->map_nr; i++) {
Jens Axboeb2dbff12018-12-11 18:39:41 -0700290 if (sb->map[i].word & ~sb->map[i].cleared)
Omar Sandoval88459642016-09-17 08:38:44 -0600291 return true;
292 }
293 return false;
294}
295EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
296
Jens Axboeea86ea22018-11-30 13:18:06 -0700297static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
Omar Sandoval88459642016-09-17 08:38:44 -0600298{
Colin Ian King60658e02016-09-19 14:34:08 +0100299 unsigned int i, weight = 0;
Omar Sandoval88459642016-09-17 08:38:44 -0600300
301 for (i = 0; i < sb->map_nr; i++) {
302 const struct sbitmap_word *word = &sb->map[i];
Ming Lei3301bc52022-01-10 15:29:45 +0800303 unsigned int word_depth = __map_depth(sb, i);
Omar Sandoval88459642016-09-17 08:38:44 -0600304
Jens Axboeea86ea22018-11-30 13:18:06 -0700305 if (set)
Ming Lei3301bc52022-01-10 15:29:45 +0800306 weight += bitmap_weight(&word->word, word_depth);
Jens Axboeea86ea22018-11-30 13:18:06 -0700307 else
Ming Lei3301bc52022-01-10 15:29:45 +0800308 weight += bitmap_weight(&word->cleared, word_depth);
Omar Sandoval88459642016-09-17 08:38:44 -0600309 }
310 return weight;
311}
Jens Axboeea86ea22018-11-30 13:18:06 -0700312
Jens Axboeea86ea22018-11-30 13:18:06 -0700313static unsigned int sbitmap_cleared(const struct sbitmap *sb)
314{
315 return __sbitmap_weight(sb, false);
316}
Omar Sandoval88459642016-09-17 08:38:44 -0600317
Ming Leicbb99502021-01-22 10:33:09 +0800318unsigned int sbitmap_weight(const struct sbitmap *sb)
319{
320 return __sbitmap_weight(sb, true) - sbitmap_cleared(sb);
321}
322EXPORT_SYMBOL_GPL(sbitmap_weight);
323
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800324void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
325{
326 seq_printf(m, "depth=%u\n", sb->depth);
Ming Leicbb99502021-01-22 10:33:09 +0800327 seq_printf(m, "busy=%u\n", sbitmap_weight(sb));
Jens Axboeea86ea22018-11-30 13:18:06 -0700328 seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800329 seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
330 seq_printf(m, "map_nr=%u\n", sb->map_nr);
331}
332EXPORT_SYMBOL_GPL(sbitmap_show);
333
334static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
335{
336 if ((offset & 0xf) == 0) {
337 if (offset != 0)
338 seq_putc(m, '\n');
339 seq_printf(m, "%08x:", offset);
340 }
341 if ((offset & 0x1) == 0)
342 seq_putc(m, ' ');
343 seq_printf(m, "%02x", byte);
344}
345
346void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
347{
348 u8 byte = 0;
349 unsigned int byte_bits = 0;
350 unsigned int offset = 0;
351 int i;
352
353 for (i = 0; i < sb->map_nr; i++) {
354 unsigned long word = READ_ONCE(sb->map[i].word);
John Garry6bf0eb52020-07-01 16:06:25 +0800355 unsigned long cleared = READ_ONCE(sb->map[i].cleared);
Ming Lei3301bc52022-01-10 15:29:45 +0800356 unsigned int word_bits = __map_depth(sb, i);
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800357
John Garry6bf0eb52020-07-01 16:06:25 +0800358 word &= ~cleared;
359
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800360 while (word_bits > 0) {
361 unsigned int bits = min(8 - byte_bits, word_bits);
362
363 byte |= (word & (BIT(bits) - 1)) << byte_bits;
364 byte_bits += bits;
365 if (byte_bits == 8) {
366 emit_byte(m, offset, byte);
367 byte = 0;
368 byte_bits = 0;
369 offset++;
370 }
371 word >>= bits;
372 word_bits -= bits;
373 }
374 }
375 if (byte_bits) {
376 emit_byte(m, offset, byte);
377 offset++;
378 }
379 if (offset)
380 seq_putc(m, '\n');
381}
382EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
383
Omar Sandovala3275532018-05-09 17:16:31 -0700384static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
385 unsigned int depth)
Omar Sandoval88459642016-09-17 08:38:44 -0600386{
387 unsigned int wake_batch;
Omar Sandovala3275532018-05-09 17:16:31 -0700388 unsigned int shallow_depth;
Omar Sandoval88459642016-09-17 08:38:44 -0600389
390 /*
391 * For each batch, we wake up one queue. We need to make sure that our
Omar Sandovala3275532018-05-09 17:16:31 -0700392 * batch size is small enough that the full depth of the bitmap,
393 * potentially limited by a shallow depth, is enough to wake up all of
394 * the queues.
395 *
396 * Each full word of the bitmap has bits_per_word bits, and there might
397 * be a partial word. There are depth / bits_per_word full words and
398 * depth % bits_per_word bits left over. In bitwise arithmetic:
399 *
400 * bits_per_word = 1 << shift
401 * depth / bits_per_word = depth >> shift
402 * depth % bits_per_word = depth & ((1 << shift) - 1)
403 *
404 * Each word can be limited to sbq->min_shallow_depth bits.
Omar Sandoval88459642016-09-17 08:38:44 -0600405 */
Omar Sandovala3275532018-05-09 17:16:31 -0700406 shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
407 depth = ((depth >> sbq->sb.shift) * shallow_depth +
408 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
409 wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
410 SBQ_WAKE_BATCH);
Omar Sandoval88459642016-09-17 08:38:44 -0600411
412 return wake_batch;
413}
414
415int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
Omar Sandovalf4a644d2016-09-17 01:28:24 -0700416 int shift, bool round_robin, gfp_t flags, int node)
Omar Sandoval88459642016-09-17 08:38:44 -0600417{
418 int ret;
419 int i;
420
Ming Leiefe1f3a2021-01-22 10:33:06 +0800421 ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node,
Ming Leic548e622021-01-22 10:33:08 +0800422 round_robin, true);
Omar Sandoval88459642016-09-17 08:38:44 -0600423 if (ret)
424 return ret;
425
Omar Sandovala3275532018-05-09 17:16:31 -0700426 sbq->min_shallow_depth = UINT_MAX;
427 sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
Omar Sandoval88459642016-09-17 08:38:44 -0600428 atomic_set(&sbq->wake_index, 0);
Jens Axboe5d2ee712018-11-29 17:36:41 -0700429 atomic_set(&sbq->ws_active, 0);
Gabriel Krisman Bertazi4f8126b2022-11-05 19:10:55 -0400430 atomic_set(&sbq->completion_cnt, 0);
431 atomic_set(&sbq->wakeup_cnt, 0);
Omar Sandoval88459642016-09-17 08:38:44 -0600432
Omar Sandoval48e28162016-09-17 01:28:22 -0700433 sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
Omar Sandoval88459642016-09-17 08:38:44 -0600434 if (!sbq->ws) {
435 sbitmap_free(&sbq->sb);
436 return -ENOMEM;
437 }
438
Gabriel Krisman Bertazi4f8126b2022-11-05 19:10:55 -0400439 for (i = 0; i < SBQ_WAIT_QUEUES; i++)
Omar Sandoval88459642016-09-17 08:38:44 -0600440 init_waitqueue_head(&sbq->ws[i].wait);
Omar Sandovalf4a644d2016-09-17 01:28:24 -0700441
Omar Sandoval88459642016-09-17 08:38:44 -0600442 return 0;
443}
444EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
445
Laibin Qiu180dccb2022-01-13 10:55:36 +0800446static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
447 unsigned int depth)
448{
449 unsigned int wake_batch;
450
451 wake_batch = sbq_calc_wake_batch(sbq, depth);
Gabriel Krisman Bertazi4f8126b2022-11-05 19:10:55 -0400452 if (sbq->wake_batch != wake_batch)
453 WRITE_ONCE(sbq->wake_batch, wake_batch);
Laibin Qiu180dccb2022-01-13 10:55:36 +0800454}
455
456void sbitmap_queue_recalculate_wake_batch(struct sbitmap_queue *sbq,
457 unsigned int users)
458{
459 unsigned int wake_batch;
Laibin Qiu108254102022-01-27 18:00:47 +0800460 unsigned int depth = (sbq->sb.depth + users - 1) / users;
Laibin Qiu180dccb2022-01-13 10:55:36 +0800461
Laibin Qiu108254102022-01-27 18:00:47 +0800462 wake_batch = clamp_val(depth / SBQ_WAIT_QUEUES,
Kemeng Shib5fcf782023-01-17 04:50:59 +0800463 1, SBQ_WAKE_BATCH);
Gabriel Krisman Bertazi4f8126b2022-11-05 19:10:55 -0400464
465 WRITE_ONCE(sbq->wake_batch, wake_batch);
Laibin Qiu180dccb2022-01-13 10:55:36 +0800466}
467EXPORT_SYMBOL_GPL(sbitmap_queue_recalculate_wake_batch);
468
Omar Sandovala3275532018-05-09 17:16:31 -0700469void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
470{
471 sbitmap_queue_update_wake_batch(sbq, depth);
Omar Sandoval88459642016-09-17 08:38:44 -0600472 sbitmap_resize(&sbq->sb, depth);
473}
474EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
475
Omar Sandovalf4a644d2016-09-17 01:28:24 -0700476int __sbitmap_queue_get(struct sbitmap_queue *sbq)
Omar Sandoval40aabb62016-09-17 01:28:23 -0700477{
Ming Leic548e622021-01-22 10:33:08 +0800478 return sbitmap_get(&sbq->sb);
Omar Sandoval40aabb62016-09-17 01:28:23 -0700479}
480EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
481
Jens Axboe9672b0d2021-10-09 13:02:23 -0600482unsigned long __sbitmap_queue_get_batch(struct sbitmap_queue *sbq, int nr_tags,
483 unsigned int *offset)
484{
485 struct sbitmap *sb = &sbq->sb;
486 unsigned int hint, depth;
487 unsigned long index, nr;
488 int i;
489
490 if (unlikely(sb->round_robin))
491 return 0;
492
493 depth = READ_ONCE(sb->depth);
494 hint = update_alloc_hint_before_get(sb, depth);
495
496 index = SB_NR_TO_INDEX(sb, hint);
497
498 for (i = 0; i < sb->map_nr; i++) {
499 struct sbitmap_word *map = &sb->map[index];
500 unsigned long get_mask;
Ming Lei3301bc52022-01-10 15:29:45 +0800501 unsigned int map_depth = __map_depth(sb, index);
Jens Axboe9672b0d2021-10-09 13:02:23 -0600502
503 sbitmap_deferred_clear(map);
Ming Lei3301bc52022-01-10 15:29:45 +0800504 if (map->word == (1UL << (map_depth - 1)) - 1)
wuchifbb564a2022-06-05 22:58:35 +0800505 goto next;
Jens Axboe9672b0d2021-10-09 13:02:23 -0600506
Ming Lei3301bc52022-01-10 15:29:45 +0800507 nr = find_first_zero_bit(&map->word, map_depth);
508 if (nr + nr_tags <= map_depth) {
Jens Axboe9672b0d2021-10-09 13:02:23 -0600509 atomic_long_t *ptr = (atomic_long_t *) &map->word;
Uros Bizjakc35227d2022-09-08 17:12:00 +0200510 unsigned long val;
Jens Axboe9672b0d2021-10-09 13:02:23 -0600511
Liu Songddbfc342022-08-26 11:14:13 +0800512 get_mask = ((1UL << nr_tags) - 1) << nr;
Uros Bizjakc35227d2022-09-08 17:12:00 +0200513 val = READ_ONCE(map->word);
Kemeng Shi903e86f2023-01-17 04:50:56 +0800514 while (!atomic_long_try_cmpxchg(ptr, &val,
515 get_mask | val))
516 ;
Uros Bizjakc35227d2022-09-08 17:12:00 +0200517 get_mask = (get_mask & ~val) >> nr;
Jens Axboe9672b0d2021-10-09 13:02:23 -0600518 if (get_mask) {
519 *offset = nr + (index << sb->shift);
520 update_alloc_hint_after_get(sb, depth, hint,
Liu Songddbfc342022-08-26 11:14:13 +0800521 *offset + nr_tags - 1);
Jens Axboe9672b0d2021-10-09 13:02:23 -0600522 return get_mask;
523 }
524 }
wuchifbb564a2022-06-05 22:58:35 +0800525next:
Jens Axboe9672b0d2021-10-09 13:02:23 -0600526 /* Jump to next index. */
527 if (++index >= sb->map_nr)
528 index = 0;
529 }
530
531 return 0;
532}
533
John Garry3f607292022-02-08 20:07:04 +0800534int sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
535 unsigned int shallow_depth)
Omar Sandovalc05e6672017-04-14 00:59:58 -0700536{
Omar Sandoval61445b562018-05-09 17:29:24 -0700537 WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth);
538
Ming Leic548e622021-01-22 10:33:08 +0800539 return sbitmap_get_shallow(&sbq->sb, shallow_depth);
Omar Sandovalc05e6672017-04-14 00:59:58 -0700540}
John Garry3f607292022-02-08 20:07:04 +0800541EXPORT_SYMBOL_GPL(sbitmap_queue_get_shallow);
Omar Sandovalc05e6672017-04-14 00:59:58 -0700542
Omar Sandovala3275532018-05-09 17:16:31 -0700543void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
544 unsigned int min_shallow_depth)
545{
546 sbq->min_shallow_depth = min_shallow_depth;
547 sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth);
548}
549EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth);
550
Gabriel Krisman Bertazi26edb302022-11-15 17:45:53 -0500551static void __sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)
Omar Sandoval88459642016-09-17 08:38:44 -0600552{
553 int i, wake_index;
554
Jens Axboe5d2ee712018-11-29 17:36:41 -0700555 if (!atomic_read(&sbq->ws_active))
Gabriel Krisman Bertazi26edb302022-11-15 17:45:53 -0500556 return;
Jens Axboe5d2ee712018-11-29 17:36:41 -0700557
Omar Sandoval88459642016-09-17 08:38:44 -0600558 wake_index = atomic_read(&sbq->wake_index);
559 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
560 struct sbq_wait_state *ws = &sbq->ws[wake_index];
561
Gabriel Krisman Bertazi976570b2022-11-15 17:45:51 -0500562 /*
563 * Advance the index before checking the current queue.
564 * It improves fairness, by ensuring the queue doesn't
565 * need to be fully emptied before trying to wake up
566 * from the next one.
567 */
Omar Sandoval88459642016-09-17 08:38:44 -0600568 wake_index = sbq_index_inc(wake_index);
Gabriel Krisman Bertazi976570b2022-11-15 17:45:51 -0500569
Gabriel Krisman Bertazi26edb302022-11-15 17:45:53 -0500570 /*
571 * It is sufficient to wake up at least one waiter to
572 * guarantee forward progress.
573 */
574 if (waitqueue_active(&ws->wait) &&
575 wake_up_nr(&ws->wait, nr))
576 break;
Omar Sandoval88459642016-09-17 08:38:44 -0600577 }
578
Gabriel Krisman Bertazi26edb302022-11-15 17:45:53 -0500579 if (wake_index != atomic_read(&sbq->wake_index))
580 atomic_set(&sbq->wake_index, wake_index);
Jens Axboec854ab572018-05-14 12:17:31 -0600581}
582
Keith Busch4acb8342022-09-09 11:40:22 -0700583void sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)
Jens Axboec854ab572018-05-14 12:17:31 -0600584{
Gabriel Krisman Bertazi4f8126b2022-11-05 19:10:55 -0400585 unsigned int wake_batch = READ_ONCE(sbq->wake_batch);
Gabriel Krisman Bertazi4f8126b2022-11-05 19:10:55 -0400586 unsigned int wakeups;
587
588 if (!atomic_read(&sbq->ws_active))
589 return;
590
591 atomic_add(nr, &sbq->completion_cnt);
592 wakeups = atomic_read(&sbq->wakeup_cnt);
593
594 do {
595 if (atomic_read(&sbq->completion_cnt) - wakeups < wake_batch)
596 return;
Gabriel Krisman Bertazi4f8126b2022-11-05 19:10:55 -0400597 } while (!atomic_try_cmpxchg(&sbq->wakeup_cnt,
598 &wakeups, wakeups + wake_batch));
599
Gabriel Krisman Bertazi26edb302022-11-15 17:45:53 -0500600 __sbitmap_queue_wake_up(sbq, wake_batch);
Omar Sandoval88459642016-09-17 08:38:44 -0600601}
Jens Axboebce1b562022-09-04 06:39:25 -0600602EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
Omar Sandoval88459642016-09-17 08:38:44 -0600603
Jens Axboe1aec5e42021-10-08 05:44:23 -0600604static inline void sbitmap_update_cpu_hint(struct sbitmap *sb, int cpu, int tag)
605{
606 if (likely(!sb->round_robin && tag < sb->depth))
Jens Axboe9f8b93a2021-10-25 10:45:01 -0600607 data_race(*per_cpu_ptr(sb->alloc_hint, cpu) = tag);
Jens Axboe1aec5e42021-10-08 05:44:23 -0600608}
609
610void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
611 int *tags, int nr_tags)
612{
613 struct sbitmap *sb = &sbq->sb;
614 unsigned long *addr = NULL;
615 unsigned long mask = 0;
616 int i;
617
618 smp_mb__before_atomic();
619 for (i = 0; i < nr_tags; i++) {
620 const int tag = tags[i] - offset;
621 unsigned long *this_addr;
622
623 /* since we're clearing a batch, skip the deferred map */
624 this_addr = &sb->map[SB_NR_TO_INDEX(sb, tag)].word;
625 if (!addr) {
626 addr = this_addr;
627 } else if (addr != this_addr) {
628 atomic_long_andnot(mask, (atomic_long_t *) addr);
629 mask = 0;
630 addr = this_addr;
631 }
632 mask |= (1UL << SB_NR_TO_BIT(sb, tag));
633 }
634
635 if (mask)
636 atomic_long_andnot(mask, (atomic_long_t *) addr);
637
638 smp_mb__after_atomic();
Keith Busch4acb8342022-09-09 11:40:22 -0700639 sbitmap_queue_wake_up(sbq, nr_tags);
Jens Axboe1aec5e42021-10-08 05:44:23 -0600640 sbitmap_update_cpu_hint(&sbq->sb, raw_smp_processor_id(),
641 tags[nr_tags - 1] - offset);
642}
643
Omar Sandoval40aabb62016-09-17 01:28:23 -0700644void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
Omar Sandovalf4a644d2016-09-17 01:28:24 -0700645 unsigned int cpu)
Omar Sandoval88459642016-09-17 08:38:44 -0600646{
Ming Leie6d1fa52019-03-22 09:13:51 +0800647 /*
648 * Once the clear bit is set, the bit may be allocated out.
649 *
Zhen Lei9dbbc3b2021-07-07 18:07:31 -0700650 * Orders READ/WRITE on the associated instance(such as request
Ming Leie6d1fa52019-03-22 09:13:51 +0800651 * of blk_mq) by this bit for avoiding race with re-allocation,
652 * and its pair is the memory barrier implied in __sbitmap_get_word.
653 *
654 * One invariant is that the clear bit has to be zero when the bit
655 * is in use.
656 */
657 smp_mb__before_atomic();
Jens Axboeea86ea22018-11-30 13:18:06 -0700658 sbitmap_deferred_clear_bit(&sbq->sb, nr);
659
Ming Leie6fc4642018-05-24 11:00:39 -0600660 /*
661 * Pairs with the memory barrier in set_current_state() to ensure the
662 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
663 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
664 * waiter. See the comment on waitqueue_active().
665 */
666 smp_mb__after_atomic();
Keith Busch4acb8342022-09-09 11:40:22 -0700667 sbitmap_queue_wake_up(sbq, 1);
Jens Axboe1aec5e42021-10-08 05:44:23 -0600668 sbitmap_update_cpu_hint(&sbq->sb, cpu, nr);
Omar Sandoval88459642016-09-17 08:38:44 -0600669}
670EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
671
672void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
673{
674 int i, wake_index;
675
676 /*
Omar Sandovalf66227d2017-01-18 11:55:21 -0800677 * Pairs with the memory barrier in set_current_state() like in
Ming Leie6fc4642018-05-24 11:00:39 -0600678 * sbitmap_queue_wake_up().
Omar Sandoval88459642016-09-17 08:38:44 -0600679 */
680 smp_mb();
681 wake_index = atomic_read(&sbq->wake_index);
682 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
683 struct sbq_wait_state *ws = &sbq->ws[wake_index];
684
685 if (waitqueue_active(&ws->wait))
686 wake_up(&ws->wait);
687
688 wake_index = sbq_index_inc(wake_index);
689 }
690}
691EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800692
693void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
694{
695 bool first;
696 int i;
697
698 sbitmap_show(&sbq->sb, m);
699
700 seq_puts(m, "alloc_hint={");
701 first = true;
702 for_each_possible_cpu(i) {
703 if (!first)
704 seq_puts(m, ", ");
705 first = false;
Ming Leic548e622021-01-22 10:33:08 +0800706 seq_printf(m, "%u", *per_cpu_ptr(sbq->sb.alloc_hint, i));
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800707 }
708 seq_puts(m, "}\n");
709
710 seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
711 seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
Jens Axboe5d2ee712018-11-29 17:36:41 -0700712 seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800713
714 seq_puts(m, "ws={\n");
715 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
716 struct sbq_wait_state *ws = &sbq->ws[i];
Gabriel Krisman Bertazi4f8126b2022-11-05 19:10:55 -0400717 seq_printf(m, "\t{.wait=%s},\n",
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800718 waitqueue_active(&ws->wait) ? "active" : "inactive");
719 }
720 seq_puts(m, "}\n");
721
Ming Leiefe1f3a2021-01-22 10:33:06 +0800722 seq_printf(m, "round_robin=%d\n", sbq->sb.round_robin);
Omar Sandovala3275532018-05-09 17:16:31 -0700723 seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800724}
725EXPORT_SYMBOL_GPL(sbitmap_queue_show);
Jens Axboe5d2ee712018-11-29 17:36:41 -0700726
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700727void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
728 struct sbq_wait_state *ws,
729 struct sbq_wait *sbq_wait)
730{
731 if (!sbq_wait->sbq) {
732 sbq_wait->sbq = sbq;
733 atomic_inc(&sbq->ws_active);
David Jefferydf034c92019-12-17 11:00:24 -0500734 add_wait_queue(&ws->wait, &sbq_wait->wait);
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700735 }
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700736}
737EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
738
739void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
740{
741 list_del_init(&sbq_wait->wait.entry);
742 if (sbq_wait->sbq) {
743 atomic_dec(&sbq_wait->sbq->ws_active);
744 sbq_wait->sbq = NULL;
745 }
746}
747EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
748
Jens Axboe5d2ee712018-11-29 17:36:41 -0700749void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
750 struct sbq_wait_state *ws,
751 struct sbq_wait *sbq_wait, int state)
752{
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700753 if (!sbq_wait->sbq) {
Jens Axboe5d2ee712018-11-29 17:36:41 -0700754 atomic_inc(&sbq->ws_active);
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700755 sbq_wait->sbq = sbq;
Jens Axboe5d2ee712018-11-29 17:36:41 -0700756 }
757 prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
758}
759EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
760
761void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
762 struct sbq_wait *sbq_wait)
763{
764 finish_wait(&ws->wait, &sbq_wait->wait);
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700765 if (sbq_wait->sbq) {
Jens Axboe5d2ee712018-11-29 17:36:41 -0700766 atomic_dec(&sbq->ws_active);
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700767 sbq_wait->sbq = NULL;
Jens Axboe5d2ee712018-11-29 17:36:41 -0700768 }
769}
770EXPORT_SYMBOL_GPL(sbitmap_finish_wait);