Loading...
1/*
2 * Tag allocation using scalable bitmaps. Uses active queue tracking to support
3 * fairer distribution of tags between multiple submitters when a shared tag map
4 * is used.
5 *
6 * Copyright (C) 2013-2014 Jens Axboe
7 */
8#include <linux/kernel.h>
9#include <linux/module.h>
10
11#include <linux/blk-mq.h>
12#include "blk.h"
13#include "blk-mq.h"
14#include "blk-mq-tag.h"
15
16bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
17{
18 if (!tags)
19 return true;
20
21 return sbitmap_any_bit_clear(&tags->bitmap_tags.sb);
22}
23
24/*
25 * If a previously inactive queue goes active, bump the active user count.
26 */
27bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
28{
29 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
30 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
31 atomic_inc(&hctx->tags->active_queues);
32
33 return true;
34}
35
36/*
37 * Wakeup all potentially sleeping on tags
38 */
39void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
40{
41 sbitmap_queue_wake_all(&tags->bitmap_tags);
42 if (include_reserve)
43 sbitmap_queue_wake_all(&tags->breserved_tags);
44}
45
46/*
47 * If a previously busy queue goes inactive, potential waiters could now
48 * be allowed to queue. Wake them up and check.
49 */
50void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
51{
52 struct blk_mq_tags *tags = hctx->tags;
53
54 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
55 return;
56
57 atomic_dec(&tags->active_queues);
58
59 blk_mq_tag_wakeup_all(tags, false);
60}
61
62/*
63 * For shared tag users, we track the number of currently active users
64 * and attempt to provide a fair share of the tag depth for each of them.
65 */
66static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
67 struct sbitmap_queue *bt)
68{
69 unsigned int depth, users;
70
71 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
72 return true;
73 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
74 return true;
75
76 /*
77 * Don't try dividing an ant
78 */
79 if (bt->sb.depth == 1)
80 return true;
81
82 users = atomic_read(&hctx->tags->active_queues);
83 if (!users)
84 return true;
85
86 /*
87 * Allow at least some tags
88 */
89 depth = max((bt->sb.depth + users - 1) / users, 4U);
90 return atomic_read(&hctx->nr_active) < depth;
91}
92
93static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
94 struct sbitmap_queue *bt)
95{
96 if (!(data->flags & BLK_MQ_REQ_INTERNAL) &&
97 !hctx_may_queue(data->hctx, bt))
98 return -1;
99 if (data->shallow_depth)
100 return __sbitmap_queue_get_shallow(bt, data->shallow_depth);
101 else
102 return __sbitmap_queue_get(bt);
103}
104
105unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
106{
107 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
108 struct sbitmap_queue *bt;
109 struct sbq_wait_state *ws;
110 DEFINE_WAIT(wait);
111 unsigned int tag_offset;
112 bool drop_ctx;
113 int tag;
114
115 if (data->flags & BLK_MQ_REQ_RESERVED) {
116 if (unlikely(!tags->nr_reserved_tags)) {
117 WARN_ON_ONCE(1);
118 return BLK_MQ_TAG_FAIL;
119 }
120 bt = &tags->breserved_tags;
121 tag_offset = 0;
122 } else {
123 bt = &tags->bitmap_tags;
124 tag_offset = tags->nr_reserved_tags;
125 }
126
127 tag = __blk_mq_get_tag(data, bt);
128 if (tag != -1)
129 goto found_tag;
130
131 if (data->flags & BLK_MQ_REQ_NOWAIT)
132 return BLK_MQ_TAG_FAIL;
133
134 ws = bt_wait_ptr(bt, data->hctx);
135 drop_ctx = data->ctx == NULL;
136 do {
137 /*
138 * We're out of tags on this hardware queue, kick any
139 * pending IO submits before going to sleep waiting for
140 * some to complete.
141 */
142 blk_mq_run_hw_queue(data->hctx, false);
143
144 /*
145 * Retry tag allocation after running the hardware queue,
146 * as running the queue may also have found completions.
147 */
148 tag = __blk_mq_get_tag(data, bt);
149 if (tag != -1)
150 break;
151
152 prepare_to_wait_exclusive(&ws->wait, &wait,
153 TASK_UNINTERRUPTIBLE);
154
155 tag = __blk_mq_get_tag(data, bt);
156 if (tag != -1)
157 break;
158
159 if (data->ctx)
160 blk_mq_put_ctx(data->ctx);
161
162 io_schedule();
163
164 data->ctx = blk_mq_get_ctx(data->q);
165 data->hctx = blk_mq_map_queue(data->q, data->ctx->cpu);
166 tags = blk_mq_tags_from_data(data);
167 if (data->flags & BLK_MQ_REQ_RESERVED)
168 bt = &tags->breserved_tags;
169 else
170 bt = &tags->bitmap_tags;
171
172 finish_wait(&ws->wait, &wait);
173 ws = bt_wait_ptr(bt, data->hctx);
174 } while (1);
175
176 if (drop_ctx && data->ctx)
177 blk_mq_put_ctx(data->ctx);
178
179 finish_wait(&ws->wait, &wait);
180
181found_tag:
182 return tag + tag_offset;
183}
184
185void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags,
186 struct blk_mq_ctx *ctx, unsigned int tag)
187{
188 if (!blk_mq_tag_is_reserved(tags, tag)) {
189 const int real_tag = tag - tags->nr_reserved_tags;
190
191 BUG_ON(real_tag >= tags->nr_tags);
192 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
193 } else {
194 BUG_ON(tag >= tags->nr_reserved_tags);
195 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
196 }
197}
198
199struct bt_iter_data {
200 struct blk_mq_hw_ctx *hctx;
201 busy_iter_fn *fn;
202 void *data;
203 bool reserved;
204};
205
206static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
207{
208 struct bt_iter_data *iter_data = data;
209 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
210 struct blk_mq_tags *tags = hctx->tags;
211 bool reserved = iter_data->reserved;
212 struct request *rq;
213
214 if (!reserved)
215 bitnr += tags->nr_reserved_tags;
216 rq = tags->rqs[bitnr];
217
218 /*
219 * We can hit rq == NULL here, because the tagging functions
220 * test and set the bit before assining ->rqs[].
221 */
222 if (rq && rq->q == hctx->queue)
223 iter_data->fn(hctx, rq, iter_data->data, reserved);
224 return true;
225}
226
227static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
228 busy_iter_fn *fn, void *data, bool reserved)
229{
230 struct bt_iter_data iter_data = {
231 .hctx = hctx,
232 .fn = fn,
233 .data = data,
234 .reserved = reserved,
235 };
236
237 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
238}
239
240struct bt_tags_iter_data {
241 struct blk_mq_tags *tags;
242 busy_tag_iter_fn *fn;
243 void *data;
244 bool reserved;
245};
246
247static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
248{
249 struct bt_tags_iter_data *iter_data = data;
250 struct blk_mq_tags *tags = iter_data->tags;
251 bool reserved = iter_data->reserved;
252 struct request *rq;
253
254 if (!reserved)
255 bitnr += tags->nr_reserved_tags;
256
257 /*
258 * We can hit rq == NULL here, because the tagging functions
259 * test and set the bit before assining ->rqs[].
260 */
261 rq = tags->rqs[bitnr];
262 if (rq)
263 iter_data->fn(rq, iter_data->data, reserved);
264
265 return true;
266}
267
268static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
269 busy_tag_iter_fn *fn, void *data, bool reserved)
270{
271 struct bt_tags_iter_data iter_data = {
272 .tags = tags,
273 .fn = fn,
274 .data = data,
275 .reserved = reserved,
276 };
277
278 if (tags->rqs)
279 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
280}
281
282static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags,
283 busy_tag_iter_fn *fn, void *priv)
284{
285 if (tags->nr_reserved_tags)
286 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true);
287 bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false);
288}
289
290void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
291 busy_tag_iter_fn *fn, void *priv)
292{
293 int i;
294
295 for (i = 0; i < tagset->nr_hw_queues; i++) {
296 if (tagset->tags && tagset->tags[i])
297 blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
298 }
299}
300EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
301
302int blk_mq_tagset_iter(struct blk_mq_tag_set *set, void *data,
303 int (fn)(void *, struct request *))
304{
305 int i, j, ret = 0;
306
307 if (WARN_ON_ONCE(!fn))
308 goto out;
309
310 for (i = 0; i < set->nr_hw_queues; i++) {
311 struct blk_mq_tags *tags = set->tags[i];
312
313 if (!tags)
314 continue;
315
316 for (j = 0; j < tags->nr_tags; j++) {
317 if (!tags->static_rqs[j])
318 continue;
319
320 ret = fn(data, tags->static_rqs[j]);
321 if (ret)
322 goto out;
323 }
324 }
325
326out:
327 return ret;
328}
329EXPORT_SYMBOL_GPL(blk_mq_tagset_iter);
330
331void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
332 void *priv)
333{
334 struct blk_mq_hw_ctx *hctx;
335 int i;
336
337
338 queue_for_each_hw_ctx(q, hctx, i) {
339 struct blk_mq_tags *tags = hctx->tags;
340
341 /*
342 * If not software queues are currently mapped to this
343 * hardware queue, there's nothing to check
344 */
345 if (!blk_mq_hw_queue_mapped(hctx))
346 continue;
347
348 if (tags->nr_reserved_tags)
349 bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
350 bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
351 }
352
353}
354
355static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
356 bool round_robin, int node)
357{
358 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
359 node);
360}
361
362static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
363 int node, int alloc_policy)
364{
365 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
366 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
367
368 if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
369 goto free_tags;
370 if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
371 node))
372 goto free_bitmap_tags;
373
374 return tags;
375free_bitmap_tags:
376 sbitmap_queue_free(&tags->bitmap_tags);
377free_tags:
378 kfree(tags);
379 return NULL;
380}
381
382struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
383 unsigned int reserved_tags,
384 int node, int alloc_policy)
385{
386 struct blk_mq_tags *tags;
387
388 if (total_tags > BLK_MQ_TAG_MAX) {
389 pr_err("blk-mq: tag depth too large\n");
390 return NULL;
391 }
392
393 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
394 if (!tags)
395 return NULL;
396
397 tags->nr_tags = total_tags;
398 tags->nr_reserved_tags = reserved_tags;
399
400 return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
401}
402
403void blk_mq_free_tags(struct blk_mq_tags *tags)
404{
405 sbitmap_queue_free(&tags->bitmap_tags);
406 sbitmap_queue_free(&tags->breserved_tags);
407 kfree(tags);
408}
409
410int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
411 struct blk_mq_tags **tagsptr, unsigned int tdepth,
412 bool can_grow)
413{
414 struct blk_mq_tags *tags = *tagsptr;
415
416 if (tdepth <= tags->nr_reserved_tags)
417 return -EINVAL;
418
419 tdepth -= tags->nr_reserved_tags;
420
421 /*
422 * If we are allowed to grow beyond the original size, allocate
423 * a new set of tags before freeing the old one.
424 */
425 if (tdepth > tags->nr_tags) {
426 struct blk_mq_tag_set *set = hctx->queue->tag_set;
427 struct blk_mq_tags *new;
428 bool ret;
429
430 if (!can_grow)
431 return -EINVAL;
432
433 /*
434 * We need some sort of upper limit, set it high enough that
435 * no valid use cases should require more.
436 */
437 if (tdepth > 16 * BLKDEV_MAX_RQ)
438 return -EINVAL;
439
440 new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth, 0);
441 if (!new)
442 return -ENOMEM;
443 ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth);
444 if (ret) {
445 blk_mq_free_rq_map(new);
446 return -ENOMEM;
447 }
448
449 blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
450 blk_mq_free_rq_map(*tagsptr);
451 *tagsptr = new;
452 } else {
453 /*
454 * Don't need (or can't) update reserved tags here, they
455 * remain static and should never need resizing.
456 */
457 sbitmap_queue_resize(&tags->bitmap_tags, tdepth);
458 }
459
460 return 0;
461}
462
463/**
464 * blk_mq_unique_tag() - return a tag that is unique queue-wide
465 * @rq: request for which to compute a unique tag
466 *
467 * The tag field in struct request is unique per hardware queue but not over
468 * all hardware queues. Hence this function that returns a tag with the
469 * hardware context index in the upper bits and the per hardware queue tag in
470 * the lower bits.
471 *
472 * Note: When called for a request that is queued on a non-multiqueue request
473 * queue, the hardware context index is set to zero.
474 */
475u32 blk_mq_unique_tag(struct request *rq)
476{
477 struct request_queue *q = rq->q;
478 struct blk_mq_hw_ctx *hctx;
479 int hwq = 0;
480
481 if (q->mq_ops) {
482 hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
483 hwq = hctx->queue_num;
484 }
485
486 return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
487 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
488}
489EXPORT_SYMBOL(blk_mq_unique_tag);
1/*
2 * Fast and scalable bitmap tagging variant. Uses sparser bitmaps spread
3 * over multiple cachelines to avoid ping-pong between multiple submitters
4 * or submitter and completer. Uses rolling wakeups to avoid falling of
5 * the scaling cliff when we run out of tags and have to start putting
6 * submitters to sleep.
7 *
8 * Uses active queue tracking to support fairer distribution of tags
9 * between multiple submitters when a shared tag map is used.
10 *
11 * Copyright (C) 2013-2014 Jens Axboe
12 */
13#include <linux/kernel.h>
14#include <linux/module.h>
15#include <linux/random.h>
16
17#include <linux/blk-mq.h>
18#include "blk.h"
19#include "blk-mq.h"
20#include "blk-mq-tag.h"
21
22static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt)
23{
24 int i;
25
26 for (i = 0; i < bt->map_nr; i++) {
27 struct blk_align_bitmap *bm = &bt->map[i];
28 int ret;
29
30 ret = find_first_zero_bit(&bm->word, bm->depth);
31 if (ret < bm->depth)
32 return true;
33 }
34
35 return false;
36}
37
38bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
39{
40 if (!tags)
41 return true;
42
43 return bt_has_free_tags(&tags->bitmap_tags);
44}
45
46static inline int bt_index_inc(int index)
47{
48 return (index + 1) & (BT_WAIT_QUEUES - 1);
49}
50
51static inline void bt_index_atomic_inc(atomic_t *index)
52{
53 int old = atomic_read(index);
54 int new = bt_index_inc(old);
55 atomic_cmpxchg(index, old, new);
56}
57
58/*
59 * If a previously inactive queue goes active, bump the active user count.
60 */
61bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
62{
63 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
64 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
65 atomic_inc(&hctx->tags->active_queues);
66
67 return true;
68}
69
70/*
71 * Wakeup all potentially sleeping on tags
72 */
73void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
74{
75 struct blk_mq_bitmap_tags *bt;
76 int i, wake_index;
77
78 /*
79 * Make sure all changes prior to this are visible from other CPUs.
80 */
81 smp_mb();
82 bt = &tags->bitmap_tags;
83 wake_index = atomic_read(&bt->wake_index);
84 for (i = 0; i < BT_WAIT_QUEUES; i++) {
85 struct bt_wait_state *bs = &bt->bs[wake_index];
86
87 if (waitqueue_active(&bs->wait))
88 wake_up(&bs->wait);
89
90 wake_index = bt_index_inc(wake_index);
91 }
92
93 if (include_reserve) {
94 bt = &tags->breserved_tags;
95 if (waitqueue_active(&bt->bs[0].wait))
96 wake_up(&bt->bs[0].wait);
97 }
98}
99
100/*
101 * If a previously busy queue goes inactive, potential waiters could now
102 * be allowed to queue. Wake them up and check.
103 */
104void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
105{
106 struct blk_mq_tags *tags = hctx->tags;
107
108 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
109 return;
110
111 atomic_dec(&tags->active_queues);
112
113 blk_mq_tag_wakeup_all(tags, false);
114}
115
116/*
117 * For shared tag users, we track the number of currently active users
118 * and attempt to provide a fair share of the tag depth for each of them.
119 */
120static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
121 struct blk_mq_bitmap_tags *bt)
122{
123 unsigned int depth, users;
124
125 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
126 return true;
127 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
128 return true;
129
130 /*
131 * Don't try dividing an ant
132 */
133 if (bt->depth == 1)
134 return true;
135
136 users = atomic_read(&hctx->tags->active_queues);
137 if (!users)
138 return true;
139
140 /*
141 * Allow at least some tags
142 */
143 depth = max((bt->depth + users - 1) / users, 4U);
144 return atomic_read(&hctx->nr_active) < depth;
145}
146
147static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag,
148 bool nowrap)
149{
150 int tag, org_last_tag = last_tag;
151
152 while (1) {
153 tag = find_next_zero_bit(&bm->word, bm->depth, last_tag);
154 if (unlikely(tag >= bm->depth)) {
155 /*
156 * We started with an offset, and we didn't reset the
157 * offset to 0 in a failure case, so start from 0 to
158 * exhaust the map.
159 */
160 if (org_last_tag && last_tag && !nowrap) {
161 last_tag = org_last_tag = 0;
162 continue;
163 }
164 return -1;
165 }
166
167 if (!test_and_set_bit(tag, &bm->word))
168 break;
169
170 last_tag = tag + 1;
171 if (last_tag >= bm->depth - 1)
172 last_tag = 0;
173 }
174
175 return tag;
176}
177
178#define BT_ALLOC_RR(tags) (tags->alloc_policy == BLK_TAG_ALLOC_RR)
179
180/*
181 * Straight forward bitmap tag implementation, where each bit is a tag
182 * (cleared == free, and set == busy). The small twist is using per-cpu
183 * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
184 * contexts. This enables us to drastically limit the space searched,
185 * without dirtying an extra shared cacheline like we would if we stored
186 * the cache value inside the shared blk_mq_bitmap_tags structure. On top
187 * of that, each word of tags is in a separate cacheline. This means that
188 * multiple users will tend to stick to different cachelines, at least
189 * until the map is exhausted.
190 */
191static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt,
192 unsigned int *tag_cache, struct blk_mq_tags *tags)
193{
194 unsigned int last_tag, org_last_tag;
195 int index, i, tag;
196
197 if (!hctx_may_queue(hctx, bt))
198 return -1;
199
200 last_tag = org_last_tag = *tag_cache;
201 index = TAG_TO_INDEX(bt, last_tag);
202
203 for (i = 0; i < bt->map_nr; i++) {
204 tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag),
205 BT_ALLOC_RR(tags));
206 if (tag != -1) {
207 tag += (index << bt->bits_per_word);
208 goto done;
209 }
210
211 /*
212 * Jump to next index, and reset the last tag to be the
213 * first tag of that index
214 */
215 index++;
216 last_tag = (index << bt->bits_per_word);
217
218 if (index >= bt->map_nr) {
219 index = 0;
220 last_tag = 0;
221 }
222 }
223
224 *tag_cache = 0;
225 return -1;
226
227 /*
228 * Only update the cache from the allocation path, if we ended
229 * up using the specific cached tag.
230 */
231done:
232 if (tag == org_last_tag || unlikely(BT_ALLOC_RR(tags))) {
233 last_tag = tag + 1;
234 if (last_tag >= bt->depth - 1)
235 last_tag = 0;
236
237 *tag_cache = last_tag;
238 }
239
240 return tag;
241}
242
243static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
244 struct blk_mq_hw_ctx *hctx)
245{
246 struct bt_wait_state *bs;
247 int wait_index;
248
249 if (!hctx)
250 return &bt->bs[0];
251
252 wait_index = atomic_read(&hctx->wait_index);
253 bs = &bt->bs[wait_index];
254 bt_index_atomic_inc(&hctx->wait_index);
255 return bs;
256}
257
258static int bt_get(struct blk_mq_alloc_data *data,
259 struct blk_mq_bitmap_tags *bt,
260 struct blk_mq_hw_ctx *hctx,
261 unsigned int *last_tag, struct blk_mq_tags *tags)
262{
263 struct bt_wait_state *bs;
264 DEFINE_WAIT(wait);
265 int tag;
266
267 tag = __bt_get(hctx, bt, last_tag, tags);
268 if (tag != -1)
269 return tag;
270
271 if (data->flags & BLK_MQ_REQ_NOWAIT)
272 return -1;
273
274 bs = bt_wait_ptr(bt, hctx);
275 do {
276 prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
277
278 tag = __bt_get(hctx, bt, last_tag, tags);
279 if (tag != -1)
280 break;
281
282 /*
283 * We're out of tags on this hardware queue, kick any
284 * pending IO submits before going to sleep waiting for
285 * some to complete. Note that hctx can be NULL here for
286 * reserved tag allocation.
287 */
288 if (hctx)
289 blk_mq_run_hw_queue(hctx, false);
290
291 /*
292 * Retry tag allocation after running the hardware queue,
293 * as running the queue may also have found completions.
294 */
295 tag = __bt_get(hctx, bt, last_tag, tags);
296 if (tag != -1)
297 break;
298
299 blk_mq_put_ctx(data->ctx);
300
301 io_schedule();
302
303 data->ctx = blk_mq_get_ctx(data->q);
304 data->hctx = data->q->mq_ops->map_queue(data->q,
305 data->ctx->cpu);
306 if (data->flags & BLK_MQ_REQ_RESERVED) {
307 bt = &data->hctx->tags->breserved_tags;
308 } else {
309 last_tag = &data->ctx->last_tag;
310 hctx = data->hctx;
311 bt = &hctx->tags->bitmap_tags;
312 }
313 finish_wait(&bs->wait, &wait);
314 bs = bt_wait_ptr(bt, hctx);
315 } while (1);
316
317 finish_wait(&bs->wait, &wait);
318 return tag;
319}
320
321static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
322{
323 int tag;
324
325 tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
326 &data->ctx->last_tag, data->hctx->tags);
327 if (tag >= 0)
328 return tag + data->hctx->tags->nr_reserved_tags;
329
330 return BLK_MQ_TAG_FAIL;
331}
332
333static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
334{
335 int tag, zero = 0;
336
337 if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
338 WARN_ON_ONCE(1);
339 return BLK_MQ_TAG_FAIL;
340 }
341
342 tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero,
343 data->hctx->tags);
344 if (tag < 0)
345 return BLK_MQ_TAG_FAIL;
346
347 return tag;
348}
349
350unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
351{
352 if (data->flags & BLK_MQ_REQ_RESERVED)
353 return __blk_mq_get_reserved_tag(data);
354 return __blk_mq_get_tag(data);
355}
356
357static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
358{
359 int i, wake_index;
360
361 wake_index = atomic_read(&bt->wake_index);
362 for (i = 0; i < BT_WAIT_QUEUES; i++) {
363 struct bt_wait_state *bs = &bt->bs[wake_index];
364
365 if (waitqueue_active(&bs->wait)) {
366 int o = atomic_read(&bt->wake_index);
367 if (wake_index != o)
368 atomic_cmpxchg(&bt->wake_index, o, wake_index);
369
370 return bs;
371 }
372
373 wake_index = bt_index_inc(wake_index);
374 }
375
376 return NULL;
377}
378
379static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
380{
381 const int index = TAG_TO_INDEX(bt, tag);
382 struct bt_wait_state *bs;
383 int wait_cnt;
384
385 clear_bit(TAG_TO_BIT(bt, tag), &bt->map[index].word);
386
387 /* Ensure that the wait list checks occur after clear_bit(). */
388 smp_mb();
389
390 bs = bt_wake_ptr(bt);
391 if (!bs)
392 return;
393
394 wait_cnt = atomic_dec_return(&bs->wait_cnt);
395 if (unlikely(wait_cnt < 0))
396 wait_cnt = atomic_inc_return(&bs->wait_cnt);
397 if (wait_cnt == 0) {
398 atomic_add(bt->wake_cnt, &bs->wait_cnt);
399 bt_index_atomic_inc(&bt->wake_index);
400 wake_up(&bs->wait);
401 }
402}
403
404void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
405 unsigned int *last_tag)
406{
407 struct blk_mq_tags *tags = hctx->tags;
408
409 if (tag >= tags->nr_reserved_tags) {
410 const int real_tag = tag - tags->nr_reserved_tags;
411
412 BUG_ON(real_tag >= tags->nr_tags);
413 bt_clear_tag(&tags->bitmap_tags, real_tag);
414 if (likely(tags->alloc_policy == BLK_TAG_ALLOC_FIFO))
415 *last_tag = real_tag;
416 } else {
417 BUG_ON(tag >= tags->nr_reserved_tags);
418 bt_clear_tag(&tags->breserved_tags, tag);
419 }
420}
421
422static void bt_for_each(struct blk_mq_hw_ctx *hctx,
423 struct blk_mq_bitmap_tags *bt, unsigned int off,
424 busy_iter_fn *fn, void *data, bool reserved)
425{
426 struct request *rq;
427 int bit, i;
428
429 for (i = 0; i < bt->map_nr; i++) {
430 struct blk_align_bitmap *bm = &bt->map[i];
431
432 for (bit = find_first_bit(&bm->word, bm->depth);
433 bit < bm->depth;
434 bit = find_next_bit(&bm->word, bm->depth, bit + 1)) {
435 rq = hctx->tags->rqs[off + bit];
436 if (rq->q == hctx->queue)
437 fn(hctx, rq, data, reserved);
438 }
439
440 off += (1 << bt->bits_per_word);
441 }
442}
443
444static void bt_tags_for_each(struct blk_mq_tags *tags,
445 struct blk_mq_bitmap_tags *bt, unsigned int off,
446 busy_tag_iter_fn *fn, void *data, bool reserved)
447{
448 struct request *rq;
449 int bit, i;
450
451 if (!tags->rqs)
452 return;
453 for (i = 0; i < bt->map_nr; i++) {
454 struct blk_align_bitmap *bm = &bt->map[i];
455
456 for (bit = find_first_bit(&bm->word, bm->depth);
457 bit < bm->depth;
458 bit = find_next_bit(&bm->word, bm->depth, bit + 1)) {
459 rq = tags->rqs[off + bit];
460 fn(rq, data, reserved);
461 }
462
463 off += (1 << bt->bits_per_word);
464 }
465}
466
467void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
468 void *priv)
469{
470 if (tags->nr_reserved_tags)
471 bt_tags_for_each(tags, &tags->breserved_tags, 0, fn, priv, true);
472 bt_tags_for_each(tags, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv,
473 false);
474}
475EXPORT_SYMBOL(blk_mq_all_tag_busy_iter);
476
477void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
478 void *priv)
479{
480 struct blk_mq_hw_ctx *hctx;
481 int i;
482
483
484 queue_for_each_hw_ctx(q, hctx, i) {
485 struct blk_mq_tags *tags = hctx->tags;
486
487 /*
488 * If not software queues are currently mapped to this
489 * hardware queue, there's nothing to check
490 */
491 if (!blk_mq_hw_queue_mapped(hctx))
492 continue;
493
494 if (tags->nr_reserved_tags)
495 bt_for_each(hctx, &tags->breserved_tags, 0, fn, priv, true);
496 bt_for_each(hctx, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv,
497 false);
498 }
499
500}
501
502static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
503{
504 unsigned int i, used;
505
506 for (i = 0, used = 0; i < bt->map_nr; i++) {
507 struct blk_align_bitmap *bm = &bt->map[i];
508
509 used += bitmap_weight(&bm->word, bm->depth);
510 }
511
512 return bt->depth - used;
513}
514
515static void bt_update_count(struct blk_mq_bitmap_tags *bt,
516 unsigned int depth)
517{
518 unsigned int tags_per_word = 1U << bt->bits_per_word;
519 unsigned int map_depth = depth;
520
521 if (depth) {
522 int i;
523
524 for (i = 0; i < bt->map_nr; i++) {
525 bt->map[i].depth = min(map_depth, tags_per_word);
526 map_depth -= bt->map[i].depth;
527 }
528 }
529
530 bt->wake_cnt = BT_WAIT_BATCH;
531 if (bt->wake_cnt > depth / BT_WAIT_QUEUES)
532 bt->wake_cnt = max(1U, depth / BT_WAIT_QUEUES);
533
534 bt->depth = depth;
535}
536
537static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
538 int node, bool reserved)
539{
540 int i;
541
542 bt->bits_per_word = ilog2(BITS_PER_LONG);
543
544 /*
545 * Depth can be zero for reserved tags, that's not a failure
546 * condition.
547 */
548 if (depth) {
549 unsigned int nr, tags_per_word;
550
551 tags_per_word = (1 << bt->bits_per_word);
552
553 /*
554 * If the tag space is small, shrink the number of tags
555 * per word so we spread over a few cachelines, at least.
556 * If less than 4 tags, just forget about it, it's not
557 * going to work optimally anyway.
558 */
559 if (depth >= 4) {
560 while (tags_per_word * 4 > depth) {
561 bt->bits_per_word--;
562 tags_per_word = (1 << bt->bits_per_word);
563 }
564 }
565
566 nr = ALIGN(depth, tags_per_word) / tags_per_word;
567 bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
568 GFP_KERNEL, node);
569 if (!bt->map)
570 return -ENOMEM;
571
572 bt->map_nr = nr;
573 }
574
575 bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
576 if (!bt->bs) {
577 kfree(bt->map);
578 bt->map = NULL;
579 return -ENOMEM;
580 }
581
582 bt_update_count(bt, depth);
583
584 for (i = 0; i < BT_WAIT_QUEUES; i++) {
585 init_waitqueue_head(&bt->bs[i].wait);
586 atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt);
587 }
588
589 return 0;
590}
591
592static void bt_free(struct blk_mq_bitmap_tags *bt)
593{
594 kfree(bt->map);
595 kfree(bt->bs);
596}
597
598static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
599 int node, int alloc_policy)
600{
601 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
602
603 tags->alloc_policy = alloc_policy;
604
605 if (bt_alloc(&tags->bitmap_tags, depth, node, false))
606 goto enomem;
607 if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
608 goto enomem;
609
610 return tags;
611enomem:
612 bt_free(&tags->bitmap_tags);
613 kfree(tags);
614 return NULL;
615}
616
617struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
618 unsigned int reserved_tags,
619 int node, int alloc_policy)
620{
621 struct blk_mq_tags *tags;
622
623 if (total_tags > BLK_MQ_TAG_MAX) {
624 pr_err("blk-mq: tag depth too large\n");
625 return NULL;
626 }
627
628 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
629 if (!tags)
630 return NULL;
631
632 if (!zalloc_cpumask_var(&tags->cpumask, GFP_KERNEL)) {
633 kfree(tags);
634 return NULL;
635 }
636
637 tags->nr_tags = total_tags;
638 tags->nr_reserved_tags = reserved_tags;
639
640 return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
641}
642
643void blk_mq_free_tags(struct blk_mq_tags *tags)
644{
645 bt_free(&tags->bitmap_tags);
646 bt_free(&tags->breserved_tags);
647 free_cpumask_var(tags->cpumask);
648 kfree(tags);
649}
650
651void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
652{
653 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
654
655 *tag = prandom_u32() % depth;
656}
657
658int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
659{
660 tdepth -= tags->nr_reserved_tags;
661 if (tdepth > tags->nr_tags)
662 return -EINVAL;
663
664 /*
665 * Don't need (or can't) update reserved tags here, they remain
666 * static and should never need resizing.
667 */
668 bt_update_count(&tags->bitmap_tags, tdepth);
669 blk_mq_tag_wakeup_all(tags, false);
670 return 0;
671}
672
673/**
674 * blk_mq_unique_tag() - return a tag that is unique queue-wide
675 * @rq: request for which to compute a unique tag
676 *
677 * The tag field in struct request is unique per hardware queue but not over
678 * all hardware queues. Hence this function that returns a tag with the
679 * hardware context index in the upper bits and the per hardware queue tag in
680 * the lower bits.
681 *
682 * Note: When called for a request that is queued on a non-multiqueue request
683 * queue, the hardware context index is set to zero.
684 */
685u32 blk_mq_unique_tag(struct request *rq)
686{
687 struct request_queue *q = rq->q;
688 struct blk_mq_hw_ctx *hctx;
689 int hwq = 0;
690
691 if (q->mq_ops) {
692 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
693 hwq = hctx->queue_num;
694 }
695
696 return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
697 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
698}
699EXPORT_SYMBOL(blk_mq_unique_tag);
700
701ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
702{
703 char *orig_page = page;
704 unsigned int free, res;
705
706 if (!tags)
707 return 0;
708
709 page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
710 "bits_per_word=%u\n",
711 tags->nr_tags, tags->nr_reserved_tags,
712 tags->bitmap_tags.bits_per_word);
713
714 free = bt_unused_tags(&tags->bitmap_tags);
715 res = bt_unused_tags(&tags->breserved_tags);
716
717 page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
718 page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
719
720 return page - orig_page;
721}