blob: 6220fa67fb7e6cd2c383efe28b4b9cced6d8e70a [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)
Ming Leic548e622021-01-22 10:33:08 +080024 *per_cpu_ptr(sb->alloc_hint, i) = prandom_u32() % 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)) {
36 hint = depth ? prandom_u32() % 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;
88 unsigned int i;
89
Ming Lei2d13b1e2021-01-22 10:33:10 +080090 if (shift < 0)
91 shift = sbitmap_calculate_shift(depth);
92
Omar Sandoval88459642016-09-17 08:38:44 -060093 bits_per_word = 1U << shift;
94 if (bits_per_word > BITS_PER_LONG)
95 return -EINVAL;
96
97 sb->shift = shift;
98 sb->depth = depth;
99 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
Ming Leiefe1f3a2021-01-22 10:33:06 +0800100 sb->round_robin = round_robin;
Omar Sandoval88459642016-09-17 08:38:44 -0600101
102 if (depth == 0) {
103 sb->map = NULL;
104 return 0;
105 }
106
Ming Leic548e622021-01-22 10:33:08 +0800107 if (alloc_hint) {
108 if (init_alloc_hint(sb, flags))
109 return -ENOMEM;
110 } else {
111 sb->alloc_hint = NULL;
112 }
113
Kees Cook590b5b72018-06-12 14:04:20 -0700114 sb->map = kcalloc_node(sb->map_nr, sizeof(*sb->map), flags, node);
Ming Leic548e622021-01-22 10:33:08 +0800115 if (!sb->map) {
116 free_percpu(sb->alloc_hint);
Omar Sandoval88459642016-09-17 08:38:44 -0600117 return -ENOMEM;
Ming Leic548e622021-01-22 10:33:08 +0800118 }
Omar Sandoval88459642016-09-17 08:38:44 -0600119
120 for (i = 0; i < sb->map_nr; i++) {
121 sb->map[i].depth = min(depth, bits_per_word);
122 depth -= sb->map[i].depth;
123 }
124 return 0;
125}
126EXPORT_SYMBOL_GPL(sbitmap_init_node);
127
128void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
129{
130 unsigned int bits_per_word = 1U << sb->shift;
131 unsigned int i;
132
Jens Axboeb2dbff12018-12-11 18:39:41 -0700133 for (i = 0; i < sb->map_nr; i++)
Pavel Begunkovb78beea2020-11-22 15:35:45 +0000134 sbitmap_deferred_clear(&sb->map[i]);
Jens Axboeb2dbff12018-12-11 18:39:41 -0700135
Omar Sandoval88459642016-09-17 08:38:44 -0600136 sb->depth = depth;
137 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
138
139 for (i = 0; i < sb->map_nr; i++) {
140 sb->map[i].depth = min(depth, bits_per_word);
141 depth -= sb->map[i].depth;
142 }
143}
144EXPORT_SYMBOL_GPL(sbitmap_resize);
145
Omar Sandovalc05e6672017-04-14 00:59:58 -0700146static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
147 unsigned int hint, bool wrap)
Omar Sandoval88459642016-09-17 08:38:44 -0600148{
Omar Sandoval88459642016-09-17 08:38:44 -0600149 int nr;
150
Pavel Begunkov0eff1f12020-11-22 15:35:48 +0000151 /* don't wrap if starting from 0 */
152 wrap = wrap && hint;
153
Omar Sandoval88459642016-09-17 08:38:44 -0600154 while (1) {
Omar Sandovalc05e6672017-04-14 00:59:58 -0700155 nr = find_next_zero_bit(word, depth, hint);
156 if (unlikely(nr >= depth)) {
Omar Sandoval88459642016-09-17 08:38:44 -0600157 /*
158 * We started with an offset, and we didn't reset the
159 * offset to 0 in a failure case, so start from 0 to
160 * exhaust the map.
161 */
Pavel Begunkov0eff1f12020-11-22 15:35:48 +0000162 if (hint && wrap) {
163 hint = 0;
Omar Sandoval88459642016-09-17 08:38:44 -0600164 continue;
165 }
166 return -1;
167 }
168
Omar Sandoval4ace53f2018-02-27 16:56:43 -0800169 if (!test_and_set_bit_lock(nr, word))
Omar Sandoval88459642016-09-17 08:38:44 -0600170 break;
171
172 hint = nr + 1;
Omar Sandovalc05e6672017-04-14 00:59:58 -0700173 if (hint >= depth - 1)
Omar Sandoval88459642016-09-17 08:38:44 -0600174 hint = 0;
175 }
176
177 return nr;
178}
179
Jens Axboeea86ea22018-11-30 13:18:06 -0700180static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index,
Ming Leiefe1f3a2021-01-22 10:33:06 +0800181 unsigned int alloc_hint)
Jens Axboeea86ea22018-11-30 13:18:06 -0700182{
Pavel Begunkovb78beea2020-11-22 15:35:45 +0000183 struct sbitmap_word *map = &sb->map[index];
Jens Axboeea86ea22018-11-30 13:18:06 -0700184 int nr;
185
186 do {
Pavel Begunkovb78beea2020-11-22 15:35:45 +0000187 nr = __sbitmap_get_word(&map->word, map->depth, alloc_hint,
Ming Leiefe1f3a2021-01-22 10:33:06 +0800188 !sb->round_robin);
Jens Axboeea86ea22018-11-30 13:18:06 -0700189 if (nr != -1)
190 break;
Pavel Begunkovb78beea2020-11-22 15:35:45 +0000191 if (!sbitmap_deferred_clear(map))
Jens Axboeea86ea22018-11-30 13:18:06 -0700192 break;
193 } while (1);
194
195 return nr;
196}
197
Ming Leic548e622021-01-22 10:33:08 +0800198static int __sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint)
Omar Sandoval88459642016-09-17 08:38:44 -0600199{
200 unsigned int i, index;
201 int nr = -1;
202
203 index = SB_NR_TO_INDEX(sb, alloc_hint);
204
Jens Axboe27fae422018-11-29 12:35:16 -0700205 /*
206 * Unless we're doing round robin tag allocation, just use the
207 * alloc_hint to find the right word index. No point in looping
208 * twice in find_next_zero_bit() for that case.
209 */
Ming Leiefe1f3a2021-01-22 10:33:06 +0800210 if (sb->round_robin)
Jens Axboe27fae422018-11-29 12:35:16 -0700211 alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
212 else
213 alloc_hint = 0;
214
Omar Sandoval88459642016-09-17 08:38:44 -0600215 for (i = 0; i < sb->map_nr; i++) {
Ming Leiefe1f3a2021-01-22 10:33:06 +0800216 nr = sbitmap_find_bit_in_index(sb, index, alloc_hint);
Omar Sandoval88459642016-09-17 08:38:44 -0600217 if (nr != -1) {
218 nr += index << sb->shift;
219 break;
220 }
221
222 /* Jump to next index. */
Jens Axboe27fae422018-11-29 12:35:16 -0700223 alloc_hint = 0;
224 if (++index >= sb->map_nr)
Omar Sandoval88459642016-09-17 08:38:44 -0600225 index = 0;
Omar Sandoval88459642016-09-17 08:38:44 -0600226 }
227
228 return nr;
229}
Ming Leic548e622021-01-22 10:33:08 +0800230
231int sbitmap_get(struct sbitmap *sb)
232{
233 int nr;
234 unsigned int hint, depth;
235
236 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
237 return -1;
238
239 depth = READ_ONCE(sb->depth);
240 hint = update_alloc_hint_before_get(sb, depth);
241 nr = __sbitmap_get(sb, hint);
242 update_alloc_hint_after_get(sb, depth, hint, nr);
243
244 return nr;
245}
Omar Sandoval88459642016-09-17 08:38:44 -0600246EXPORT_SYMBOL_GPL(sbitmap_get);
247
Ming Leic548e622021-01-22 10:33:08 +0800248static int __sbitmap_get_shallow(struct sbitmap *sb,
249 unsigned int alloc_hint,
250 unsigned long shallow_depth)
Omar Sandovalc05e6672017-04-14 00:59:58 -0700251{
252 unsigned int i, index;
253 int nr = -1;
254
255 index = SB_NR_TO_INDEX(sb, alloc_hint);
256
257 for (i = 0; i < sb->map_nr; i++) {
Jens Axboeb2dbff12018-12-11 18:39:41 -0700258again:
Omar Sandovalc05e6672017-04-14 00:59:58 -0700259 nr = __sbitmap_get_word(&sb->map[index].word,
260 min(sb->map[index].depth, shallow_depth),
261 SB_NR_TO_BIT(sb, alloc_hint), true);
262 if (nr != -1) {
263 nr += index << sb->shift;
264 break;
265 }
266
Pavel Begunkovb78beea2020-11-22 15:35:45 +0000267 if (sbitmap_deferred_clear(&sb->map[index]))
Jens Axboeb2dbff12018-12-11 18:39:41 -0700268 goto again;
269
Omar Sandovalc05e6672017-04-14 00:59:58 -0700270 /* Jump to next index. */
271 index++;
272 alloc_hint = index << sb->shift;
273
274 if (index >= sb->map_nr) {
275 index = 0;
276 alloc_hint = 0;
277 }
278 }
279
280 return nr;
281}
Ming Leic548e622021-01-22 10:33:08 +0800282
283int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth)
284{
285 int nr;
286 unsigned int hint, depth;
287
288 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
289 return -1;
290
291 depth = READ_ONCE(sb->depth);
292 hint = update_alloc_hint_before_get(sb, depth);
293 nr = __sbitmap_get_shallow(sb, hint, shallow_depth);
294 update_alloc_hint_after_get(sb, depth, hint, nr);
295
296 return nr;
297}
Omar Sandovalc05e6672017-04-14 00:59:58 -0700298EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
299
Omar Sandoval88459642016-09-17 08:38:44 -0600300bool sbitmap_any_bit_set(const struct sbitmap *sb)
301{
302 unsigned int i;
303
304 for (i = 0; i < sb->map_nr; i++) {
Jens Axboeb2dbff12018-12-11 18:39:41 -0700305 if (sb->map[i].word & ~sb->map[i].cleared)
Omar Sandoval88459642016-09-17 08:38:44 -0600306 return true;
307 }
308 return false;
309}
310EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
311
Jens Axboeea86ea22018-11-30 13:18:06 -0700312static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
Omar Sandoval88459642016-09-17 08:38:44 -0600313{
Colin Ian King60658e02016-09-19 14:34:08 +0100314 unsigned int i, weight = 0;
Omar Sandoval88459642016-09-17 08:38:44 -0600315
316 for (i = 0; i < sb->map_nr; i++) {
317 const struct sbitmap_word *word = &sb->map[i];
318
Jens Axboeea86ea22018-11-30 13:18:06 -0700319 if (set)
320 weight += bitmap_weight(&word->word, word->depth);
321 else
322 weight += bitmap_weight(&word->cleared, word->depth);
Omar Sandoval88459642016-09-17 08:38:44 -0600323 }
324 return weight;
325}
Jens Axboeea86ea22018-11-30 13:18:06 -0700326
Jens Axboeea86ea22018-11-30 13:18:06 -0700327static unsigned int sbitmap_cleared(const struct sbitmap *sb)
328{
329 return __sbitmap_weight(sb, false);
330}
Omar Sandoval88459642016-09-17 08:38:44 -0600331
Ming Leicbb99502021-01-22 10:33:09 +0800332unsigned int sbitmap_weight(const struct sbitmap *sb)
333{
334 return __sbitmap_weight(sb, true) - sbitmap_cleared(sb);
335}
336EXPORT_SYMBOL_GPL(sbitmap_weight);
337
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800338void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
339{
340 seq_printf(m, "depth=%u\n", sb->depth);
Ming Leicbb99502021-01-22 10:33:09 +0800341 seq_printf(m, "busy=%u\n", sbitmap_weight(sb));
Jens Axboeea86ea22018-11-30 13:18:06 -0700342 seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800343 seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
344 seq_printf(m, "map_nr=%u\n", sb->map_nr);
345}
346EXPORT_SYMBOL_GPL(sbitmap_show);
347
348static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
349{
350 if ((offset & 0xf) == 0) {
351 if (offset != 0)
352 seq_putc(m, '\n');
353 seq_printf(m, "%08x:", offset);
354 }
355 if ((offset & 0x1) == 0)
356 seq_putc(m, ' ');
357 seq_printf(m, "%02x", byte);
358}
359
360void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
361{
362 u8 byte = 0;
363 unsigned int byte_bits = 0;
364 unsigned int offset = 0;
365 int i;
366
367 for (i = 0; i < sb->map_nr; i++) {
368 unsigned long word = READ_ONCE(sb->map[i].word);
John Garry6bf0eb52020-07-01 16:06:25 +0800369 unsigned long cleared = READ_ONCE(sb->map[i].cleared);
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800370 unsigned int word_bits = READ_ONCE(sb->map[i].depth);
371
John Garry6bf0eb52020-07-01 16:06:25 +0800372 word &= ~cleared;
373
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800374 while (word_bits > 0) {
375 unsigned int bits = min(8 - byte_bits, word_bits);
376
377 byte |= (word & (BIT(bits) - 1)) << byte_bits;
378 byte_bits += bits;
379 if (byte_bits == 8) {
380 emit_byte(m, offset, byte);
381 byte = 0;
382 byte_bits = 0;
383 offset++;
384 }
385 word >>= bits;
386 word_bits -= bits;
387 }
388 }
389 if (byte_bits) {
390 emit_byte(m, offset, byte);
391 offset++;
392 }
393 if (offset)
394 seq_putc(m, '\n');
395}
396EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
397
Omar Sandovala3275532018-05-09 17:16:31 -0700398static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
399 unsigned int depth)
Omar Sandoval88459642016-09-17 08:38:44 -0600400{
401 unsigned int wake_batch;
Omar Sandovala3275532018-05-09 17:16:31 -0700402 unsigned int shallow_depth;
Omar Sandoval88459642016-09-17 08:38:44 -0600403
404 /*
405 * For each batch, we wake up one queue. We need to make sure that our
Omar Sandovala3275532018-05-09 17:16:31 -0700406 * batch size is small enough that the full depth of the bitmap,
407 * potentially limited by a shallow depth, is enough to wake up all of
408 * the queues.
409 *
410 * Each full word of the bitmap has bits_per_word bits, and there might
411 * be a partial word. There are depth / bits_per_word full words and
412 * depth % bits_per_word bits left over. In bitwise arithmetic:
413 *
414 * bits_per_word = 1 << shift
415 * depth / bits_per_word = depth >> shift
416 * depth % bits_per_word = depth & ((1 << shift) - 1)
417 *
418 * Each word can be limited to sbq->min_shallow_depth bits.
Omar Sandoval88459642016-09-17 08:38:44 -0600419 */
Omar Sandovala3275532018-05-09 17:16:31 -0700420 shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
421 depth = ((depth >> sbq->sb.shift) * shallow_depth +
422 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
423 wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
424 SBQ_WAKE_BATCH);
Omar Sandoval88459642016-09-17 08:38:44 -0600425
426 return wake_batch;
427}
428
429int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
Omar Sandovalf4a644d2016-09-17 01:28:24 -0700430 int shift, bool round_robin, gfp_t flags, int node)
Omar Sandoval88459642016-09-17 08:38:44 -0600431{
432 int ret;
433 int i;
434
Ming Leiefe1f3a2021-01-22 10:33:06 +0800435 ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node,
Ming Leic548e622021-01-22 10:33:08 +0800436 round_robin, true);
Omar Sandoval88459642016-09-17 08:38:44 -0600437 if (ret)
438 return ret;
439
Omar Sandovala3275532018-05-09 17:16:31 -0700440 sbq->min_shallow_depth = UINT_MAX;
441 sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
Omar Sandoval88459642016-09-17 08:38:44 -0600442 atomic_set(&sbq->wake_index, 0);
Jens Axboe5d2ee712018-11-29 17:36:41 -0700443 atomic_set(&sbq->ws_active, 0);
Omar Sandoval88459642016-09-17 08:38:44 -0600444
Omar Sandoval48e28162016-09-17 01:28:22 -0700445 sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
Omar Sandoval88459642016-09-17 08:38:44 -0600446 if (!sbq->ws) {
447 sbitmap_free(&sbq->sb);
448 return -ENOMEM;
449 }
450
451 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
452 init_waitqueue_head(&sbq->ws[i].wait);
453 atomic_set(&sbq->ws[i].wait_cnt, sbq->wake_batch);
454 }
Omar Sandovalf4a644d2016-09-17 01:28:24 -0700455
Omar Sandoval88459642016-09-17 08:38:44 -0600456 return 0;
457}
458EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
459
Laibin Qiu180dccb2022-01-13 10:55:36 +0800460static inline void __sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
461 unsigned int wake_batch)
Omar Sandoval88459642016-09-17 08:38:44 -0600462{
Omar Sandoval6c0ca7a2017-01-18 11:55:22 -0800463 int i;
464
465 if (sbq->wake_batch != wake_batch) {
466 WRITE_ONCE(sbq->wake_batch, wake_batch);
467 /*
Ming Leie6fc4642018-05-24 11:00:39 -0600468 * Pairs with the memory barrier in sbitmap_queue_wake_up()
469 * to ensure that the batch size is updated before the wait
470 * counts.
Omar Sandoval6c0ca7a2017-01-18 11:55:22 -0800471 */
Andrea Parria0934fd2019-05-20 19:23:57 +0200472 smp_mb();
Omar Sandoval6c0ca7a2017-01-18 11:55:22 -0800473 for (i = 0; i < SBQ_WAIT_QUEUES; i++)
474 atomic_set(&sbq->ws[i].wait_cnt, 1);
475 }
Omar Sandovala3275532018-05-09 17:16:31 -0700476}
477
Laibin Qiu180dccb2022-01-13 10:55:36 +0800478static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
479 unsigned int depth)
480{
481 unsigned int wake_batch;
482
483 wake_batch = sbq_calc_wake_batch(sbq, depth);
484 __sbitmap_queue_update_wake_batch(sbq, wake_batch);
485}
486
487void sbitmap_queue_recalculate_wake_batch(struct sbitmap_queue *sbq,
488 unsigned int users)
489{
490 unsigned int wake_batch;
491
492 wake_batch = clamp_val((sbq->sb.depth + users - 1) /
493 users, 4, SBQ_WAKE_BATCH);
494 __sbitmap_queue_update_wake_batch(sbq, wake_batch);
495}
496EXPORT_SYMBOL_GPL(sbitmap_queue_recalculate_wake_batch);
497
Omar Sandovala3275532018-05-09 17:16:31 -0700498void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
499{
500 sbitmap_queue_update_wake_batch(sbq, depth);
Omar Sandoval88459642016-09-17 08:38:44 -0600501 sbitmap_resize(&sbq->sb, depth);
502}
503EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
504
Omar Sandovalf4a644d2016-09-17 01:28:24 -0700505int __sbitmap_queue_get(struct sbitmap_queue *sbq)
Omar Sandoval40aabb62016-09-17 01:28:23 -0700506{
Ming Leic548e622021-01-22 10:33:08 +0800507 return sbitmap_get(&sbq->sb);
Omar Sandoval40aabb62016-09-17 01:28:23 -0700508}
509EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
510
Jens Axboe9672b0d2021-10-09 13:02:23 -0600511unsigned long __sbitmap_queue_get_batch(struct sbitmap_queue *sbq, int nr_tags,
512 unsigned int *offset)
513{
514 struct sbitmap *sb = &sbq->sb;
515 unsigned int hint, depth;
516 unsigned long index, nr;
517 int i;
518
519 if (unlikely(sb->round_robin))
520 return 0;
521
522 depth = READ_ONCE(sb->depth);
523 hint = update_alloc_hint_before_get(sb, depth);
524
525 index = SB_NR_TO_INDEX(sb, hint);
526
527 for (i = 0; i < sb->map_nr; i++) {
528 struct sbitmap_word *map = &sb->map[index];
529 unsigned long get_mask;
530
531 sbitmap_deferred_clear(map);
532 if (map->word == (1UL << (map->depth - 1)) - 1)
533 continue;
534
535 nr = find_first_zero_bit(&map->word, map->depth);
536 if (nr + nr_tags <= map->depth) {
537 atomic_long_t *ptr = (atomic_long_t *) &map->word;
538 int map_tags = min_t(int, nr_tags, map->depth);
539 unsigned long val, ret;
540
541 get_mask = ((1UL << map_tags) - 1) << nr;
542 do {
543 val = READ_ONCE(map->word);
544 ret = atomic_long_cmpxchg(ptr, val, get_mask | val);
545 } while (ret != val);
546 get_mask = (get_mask & ~ret) >> nr;
547 if (get_mask) {
548 *offset = nr + (index << sb->shift);
549 update_alloc_hint_after_get(sb, depth, hint,
550 *offset + map_tags - 1);
551 return get_mask;
552 }
553 }
554 /* Jump to next index. */
555 if (++index >= sb->map_nr)
556 index = 0;
557 }
558
559 return 0;
560}
561
Omar Sandovalc05e6672017-04-14 00:59:58 -0700562int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
563 unsigned int shallow_depth)
564{
Omar Sandoval61445b562018-05-09 17:29:24 -0700565 WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth);
566
Ming Leic548e622021-01-22 10:33:08 +0800567 return sbitmap_get_shallow(&sbq->sb, shallow_depth);
Omar Sandovalc05e6672017-04-14 00:59:58 -0700568}
569EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow);
570
Omar Sandovala3275532018-05-09 17:16:31 -0700571void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
572 unsigned int min_shallow_depth)
573{
574 sbq->min_shallow_depth = min_shallow_depth;
575 sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth);
576}
577EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth);
578
Omar Sandoval88459642016-09-17 08:38:44 -0600579static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)
580{
581 int i, wake_index;
582
Jens Axboe5d2ee712018-11-29 17:36:41 -0700583 if (!atomic_read(&sbq->ws_active))
584 return NULL;
585
Omar Sandoval88459642016-09-17 08:38:44 -0600586 wake_index = atomic_read(&sbq->wake_index);
587 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
588 struct sbq_wait_state *ws = &sbq->ws[wake_index];
589
590 if (waitqueue_active(&ws->wait)) {
Pavel Begunkov41723282019-05-23 18:39:16 +0300591 if (wake_index != atomic_read(&sbq->wake_index))
592 atomic_set(&sbq->wake_index, wake_index);
Omar Sandoval88459642016-09-17 08:38:44 -0600593 return ws;
594 }
595
596 wake_index = sbq_index_inc(wake_index);
597 }
598
599 return NULL;
600}
601
Jens Axboec854ab52018-05-14 12:17:31 -0600602static bool __sbq_wake_up(struct sbitmap_queue *sbq)
Omar Sandoval88459642016-09-17 08:38:44 -0600603{
604 struct sbq_wait_state *ws;
Omar Sandoval6c0ca7a2017-01-18 11:55:22 -0800605 unsigned int wake_batch;
Omar Sandoval88459642016-09-17 08:38:44 -0600606 int wait_cnt;
607
Omar Sandoval88459642016-09-17 08:38:44 -0600608 ws = sbq_wake_ptr(sbq);
609 if (!ws)
Jens Axboec854ab52018-05-14 12:17:31 -0600610 return false;
Omar Sandoval88459642016-09-17 08:38:44 -0600611
612 wait_cnt = atomic_dec_return(&ws->wait_cnt);
Omar Sandoval6c0ca7a2017-01-18 11:55:22 -0800613 if (wait_cnt <= 0) {
Jens Axboec854ab52018-05-14 12:17:31 -0600614 int ret;
615
Omar Sandoval6c0ca7a2017-01-18 11:55:22 -0800616 wake_batch = READ_ONCE(sbq->wake_batch);
Jens Axboec854ab52018-05-14 12:17:31 -0600617
Omar Sandoval6c0ca7a2017-01-18 11:55:22 -0800618 /*
619 * Pairs with the memory barrier in sbitmap_queue_resize() to
620 * ensure that we see the batch size update before the wait
621 * count is reset.
622 */
623 smp_mb__before_atomic();
Jens Axboec854ab52018-05-14 12:17:31 -0600624
Omar Sandoval6c0ca7a2017-01-18 11:55:22 -0800625 /*
Jens Axboec854ab52018-05-14 12:17:31 -0600626 * For concurrent callers of this, the one that failed the
627 * atomic_cmpxhcg() race should call this function again
628 * to wakeup a new batch on a different 'ws'.
Omar Sandoval6c0ca7a2017-01-18 11:55:22 -0800629 */
Jens Axboec854ab52018-05-14 12:17:31 -0600630 ret = atomic_cmpxchg(&ws->wait_cnt, wait_cnt, wake_batch);
631 if (ret == wait_cnt) {
632 sbq_index_atomic_inc(&sbq->wake_index);
633 wake_up_nr(&ws->wait, wake_batch);
634 return false;
635 }
636
637 return true;
Omar Sandoval88459642016-09-17 08:38:44 -0600638 }
Jens Axboec854ab52018-05-14 12:17:31 -0600639
640 return false;
641}
642
Ming Leie6fc4642018-05-24 11:00:39 -0600643void sbitmap_queue_wake_up(struct sbitmap_queue *sbq)
Jens Axboec854ab52018-05-14 12:17:31 -0600644{
645 while (__sbq_wake_up(sbq))
646 ;
Omar Sandoval88459642016-09-17 08:38:44 -0600647}
Ming Leie6fc4642018-05-24 11:00:39 -0600648EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
Omar Sandoval88459642016-09-17 08:38:44 -0600649
Jens Axboe1aec5e42021-10-08 05:44:23 -0600650static inline void sbitmap_update_cpu_hint(struct sbitmap *sb, int cpu, int tag)
651{
652 if (likely(!sb->round_robin && tag < sb->depth))
Jens Axboe9f8b93a2021-10-25 10:45:01 -0600653 data_race(*per_cpu_ptr(sb->alloc_hint, cpu) = tag);
Jens Axboe1aec5e42021-10-08 05:44:23 -0600654}
655
656void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
657 int *tags, int nr_tags)
658{
659 struct sbitmap *sb = &sbq->sb;
660 unsigned long *addr = NULL;
661 unsigned long mask = 0;
662 int i;
663
664 smp_mb__before_atomic();
665 for (i = 0; i < nr_tags; i++) {
666 const int tag = tags[i] - offset;
667 unsigned long *this_addr;
668
669 /* since we're clearing a batch, skip the deferred map */
670 this_addr = &sb->map[SB_NR_TO_INDEX(sb, tag)].word;
671 if (!addr) {
672 addr = this_addr;
673 } else if (addr != this_addr) {
674 atomic_long_andnot(mask, (atomic_long_t *) addr);
675 mask = 0;
676 addr = this_addr;
677 }
678 mask |= (1UL << SB_NR_TO_BIT(sb, tag));
679 }
680
681 if (mask)
682 atomic_long_andnot(mask, (atomic_long_t *) addr);
683
684 smp_mb__after_atomic();
685 sbitmap_queue_wake_up(sbq);
686 sbitmap_update_cpu_hint(&sbq->sb, raw_smp_processor_id(),
687 tags[nr_tags - 1] - offset);
688}
689
Omar Sandoval40aabb62016-09-17 01:28:23 -0700690void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
Omar Sandovalf4a644d2016-09-17 01:28:24 -0700691 unsigned int cpu)
Omar Sandoval88459642016-09-17 08:38:44 -0600692{
Ming Leie6d1fa52019-03-22 09:13:51 +0800693 /*
694 * Once the clear bit is set, the bit may be allocated out.
695 *
Zhen Lei9dbbc3b2021-07-07 18:07:31 -0700696 * Orders READ/WRITE on the associated instance(such as request
Ming Leie6d1fa52019-03-22 09:13:51 +0800697 * of blk_mq) by this bit for avoiding race with re-allocation,
698 * and its pair is the memory barrier implied in __sbitmap_get_word.
699 *
700 * One invariant is that the clear bit has to be zero when the bit
701 * is in use.
702 */
703 smp_mb__before_atomic();
Jens Axboeea86ea22018-11-30 13:18:06 -0700704 sbitmap_deferred_clear_bit(&sbq->sb, nr);
705
Ming Leie6fc4642018-05-24 11:00:39 -0600706 /*
707 * Pairs with the memory barrier in set_current_state() to ensure the
708 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
709 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
710 * waiter. See the comment on waitqueue_active().
711 */
712 smp_mb__after_atomic();
713 sbitmap_queue_wake_up(sbq);
Jens Axboe1aec5e42021-10-08 05:44:23 -0600714 sbitmap_update_cpu_hint(&sbq->sb, cpu, nr);
Omar Sandoval88459642016-09-17 08:38:44 -0600715}
716EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
717
718void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
719{
720 int i, wake_index;
721
722 /*
Omar Sandovalf66227d2017-01-18 11:55:21 -0800723 * Pairs with the memory barrier in set_current_state() like in
Ming Leie6fc4642018-05-24 11:00:39 -0600724 * sbitmap_queue_wake_up().
Omar Sandoval88459642016-09-17 08:38:44 -0600725 */
726 smp_mb();
727 wake_index = atomic_read(&sbq->wake_index);
728 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
729 struct sbq_wait_state *ws = &sbq->ws[wake_index];
730
731 if (waitqueue_active(&ws->wait))
732 wake_up(&ws->wait);
733
734 wake_index = sbq_index_inc(wake_index);
735 }
736}
737EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800738
739void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
740{
741 bool first;
742 int i;
743
744 sbitmap_show(&sbq->sb, m);
745
746 seq_puts(m, "alloc_hint={");
747 first = true;
748 for_each_possible_cpu(i) {
749 if (!first)
750 seq_puts(m, ", ");
751 first = false;
Ming Leic548e622021-01-22 10:33:08 +0800752 seq_printf(m, "%u", *per_cpu_ptr(sbq->sb.alloc_hint, i));
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800753 }
754 seq_puts(m, "}\n");
755
756 seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
757 seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
Jens Axboe5d2ee712018-11-29 17:36:41 -0700758 seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800759
760 seq_puts(m, "ws={\n");
761 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
762 struct sbq_wait_state *ws = &sbq->ws[i];
763
764 seq_printf(m, "\t{.wait_cnt=%d, .wait=%s},\n",
765 atomic_read(&ws->wait_cnt),
766 waitqueue_active(&ws->wait) ? "active" : "inactive");
767 }
768 seq_puts(m, "}\n");
769
Ming Leiefe1f3a2021-01-22 10:33:06 +0800770 seq_printf(m, "round_robin=%d\n", sbq->sb.round_robin);
Omar Sandovala3275532018-05-09 17:16:31 -0700771 seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
Omar Sandoval24af1ccf2017-01-25 14:32:13 -0800772}
773EXPORT_SYMBOL_GPL(sbitmap_queue_show);
Jens Axboe5d2ee712018-11-29 17:36:41 -0700774
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700775void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
776 struct sbq_wait_state *ws,
777 struct sbq_wait *sbq_wait)
778{
779 if (!sbq_wait->sbq) {
780 sbq_wait->sbq = sbq;
781 atomic_inc(&sbq->ws_active);
David Jefferydf034c92019-12-17 11:00:24 -0500782 add_wait_queue(&ws->wait, &sbq_wait->wait);
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700783 }
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700784}
785EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
786
787void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
788{
789 list_del_init(&sbq_wait->wait.entry);
790 if (sbq_wait->sbq) {
791 atomic_dec(&sbq_wait->sbq->ws_active);
792 sbq_wait->sbq = NULL;
793 }
794}
795EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
796
Jens Axboe5d2ee712018-11-29 17:36:41 -0700797void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
798 struct sbq_wait_state *ws,
799 struct sbq_wait *sbq_wait, int state)
800{
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700801 if (!sbq_wait->sbq) {
Jens Axboe5d2ee712018-11-29 17:36:41 -0700802 atomic_inc(&sbq->ws_active);
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700803 sbq_wait->sbq = sbq;
Jens Axboe5d2ee712018-11-29 17:36:41 -0700804 }
805 prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
806}
807EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
808
809void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
810 struct sbq_wait *sbq_wait)
811{
812 finish_wait(&ws->wait, &sbq_wait->wait);
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700813 if (sbq_wait->sbq) {
Jens Axboe5d2ee712018-11-29 17:36:41 -0700814 atomic_dec(&sbq->ws_active);
Jens Axboe9f6b7ef2018-12-20 08:49:00 -0700815 sbq_wait->sbq = NULL;
Jens Axboe5d2ee712018-11-29 17:36:41 -0700816 }
817}
818EXPORT_SYMBOL_GPL(sbitmap_finish_wait);