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// SPDX-License-Identifier: GPL-2.0
2/*
3 * Tag allocation using scalable bitmaps. Uses active queue tracking to support
4 * fairer distribution of tags between multiple submitters when a shared tag map
5 * is used.
6 *
7 * Copyright (C) 2013-2014 Jens Axboe
8 */
9#include <linux/kernel.h>
10#include <linux/module.h>
11
12#include <linux/blk-mq.h>
13#include <linux/delay.h>
14#include "blk.h"
15#include "blk-mq.h"
16#include "blk-mq-sched.h"
17#include "blk-mq-tag.h"
18
19/*
20 * If a previously inactive queue goes active, bump the active user count.
21 * We need to do this before try to allocate driver tag, then even if fail
22 * to get tag when first time, the other shared-tag users could reserve
23 * budget for it.
24 */
25bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
26{
27 if (blk_mq_is_sbitmap_shared(hctx->flags)) {
28 struct request_queue *q = hctx->queue;
29 struct blk_mq_tag_set *set = q->tag_set;
30
31 if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags) &&
32 !test_and_set_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
33 atomic_inc(&set->active_queues_shared_sbitmap);
34 } else {
35 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
36 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
37 atomic_inc(&hctx->tags->active_queues);
38 }
39
40 return true;
41}
42
43/*
44 * Wakeup all potentially sleeping on tags
45 */
46void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
47{
48 sbitmap_queue_wake_all(tags->bitmap_tags);
49 if (include_reserve)
50 sbitmap_queue_wake_all(tags->breserved_tags);
51}
52
53/*
54 * If a previously busy queue goes inactive, potential waiters could now
55 * be allowed to queue. Wake them up and check.
56 */
57void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
58{
59 struct blk_mq_tags *tags = hctx->tags;
60 struct request_queue *q = hctx->queue;
61 struct blk_mq_tag_set *set = q->tag_set;
62
63 if (blk_mq_is_sbitmap_shared(hctx->flags)) {
64 if (!test_and_clear_bit(QUEUE_FLAG_HCTX_ACTIVE,
65 &q->queue_flags))
66 return;
67 atomic_dec(&set->active_queues_shared_sbitmap);
68 } else {
69 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
70 return;
71 atomic_dec(&tags->active_queues);
72 }
73
74 blk_mq_tag_wakeup_all(tags, false);
75}
76
77static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
78 struct sbitmap_queue *bt)
79{
80 if (!data->q->elevator && !(data->flags & BLK_MQ_REQ_RESERVED) &&
81 !hctx_may_queue(data->hctx, bt))
82 return BLK_MQ_NO_TAG;
83
84 if (data->shallow_depth)
85 return __sbitmap_queue_get_shallow(bt, data->shallow_depth);
86 else
87 return __sbitmap_queue_get(bt);
88}
89
90unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
91{
92 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
93 struct sbitmap_queue *bt;
94 struct sbq_wait_state *ws;
95 DEFINE_SBQ_WAIT(wait);
96 unsigned int tag_offset;
97 int tag;
98
99 if (data->flags & BLK_MQ_REQ_RESERVED) {
100 if (unlikely(!tags->nr_reserved_tags)) {
101 WARN_ON_ONCE(1);
102 return BLK_MQ_NO_TAG;
103 }
104 bt = tags->breserved_tags;
105 tag_offset = 0;
106 } else {
107 bt = tags->bitmap_tags;
108 tag_offset = tags->nr_reserved_tags;
109 }
110
111 tag = __blk_mq_get_tag(data, bt);
112 if (tag != BLK_MQ_NO_TAG)
113 goto found_tag;
114
115 if (data->flags & BLK_MQ_REQ_NOWAIT)
116 return BLK_MQ_NO_TAG;
117
118 ws = bt_wait_ptr(bt, data->hctx);
119 do {
120 struct sbitmap_queue *bt_prev;
121
122 /*
123 * We're out of tags on this hardware queue, kick any
124 * pending IO submits before going to sleep waiting for
125 * some to complete.
126 */
127 blk_mq_run_hw_queue(data->hctx, false);
128
129 /*
130 * Retry tag allocation after running the hardware queue,
131 * as running the queue may also have found completions.
132 */
133 tag = __blk_mq_get_tag(data, bt);
134 if (tag != BLK_MQ_NO_TAG)
135 break;
136
137 sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
138
139 tag = __blk_mq_get_tag(data, bt);
140 if (tag != BLK_MQ_NO_TAG)
141 break;
142
143 bt_prev = bt;
144 io_schedule();
145
146 sbitmap_finish_wait(bt, ws, &wait);
147
148 data->ctx = blk_mq_get_ctx(data->q);
149 data->hctx = blk_mq_map_queue(data->q, data->cmd_flags,
150 data->ctx);
151 tags = blk_mq_tags_from_data(data);
152 if (data->flags & BLK_MQ_REQ_RESERVED)
153 bt = tags->breserved_tags;
154 else
155 bt = tags->bitmap_tags;
156
157 /*
158 * If destination hw queue is changed, fake wake up on
159 * previous queue for compensating the wake up miss, so
160 * other allocations on previous queue won't be starved.
161 */
162 if (bt != bt_prev)
163 sbitmap_queue_wake_up(bt_prev);
164
165 ws = bt_wait_ptr(bt, data->hctx);
166 } while (1);
167
168 sbitmap_finish_wait(bt, ws, &wait);
169
170found_tag:
171 /*
172 * Give up this allocation if the hctx is inactive. The caller will
173 * retry on an active hctx.
174 */
175 if (unlikely(test_bit(BLK_MQ_S_INACTIVE, &data->hctx->state))) {
176 blk_mq_put_tag(tags, data->ctx, tag + tag_offset);
177 return BLK_MQ_NO_TAG;
178 }
179 return tag + tag_offset;
180}
181
182void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
183 unsigned int tag)
184{
185 if (!blk_mq_tag_is_reserved(tags, tag)) {
186 const int real_tag = tag - tags->nr_reserved_tags;
187
188 BUG_ON(real_tag >= tags->nr_tags);
189 sbitmap_queue_clear(tags->bitmap_tags, real_tag, ctx->cpu);
190 } else {
191 BUG_ON(tag >= tags->nr_reserved_tags);
192 sbitmap_queue_clear(tags->breserved_tags, tag, ctx->cpu);
193 }
194}
195
196struct bt_iter_data {
197 struct blk_mq_hw_ctx *hctx;
198 busy_iter_fn *fn;
199 void *data;
200 bool reserved;
201};
202
203static struct request *blk_mq_find_and_get_req(struct blk_mq_tags *tags,
204 unsigned int bitnr)
205{
206 struct request *rq;
207 unsigned long flags;
208
209 spin_lock_irqsave(&tags->lock, flags);
210 rq = tags->rqs[bitnr];
211 if (!rq || rq->tag != bitnr || !refcount_inc_not_zero(&rq->ref))
212 rq = NULL;
213 spin_unlock_irqrestore(&tags->lock, flags);
214 return rq;
215}
216
217static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
218{
219 struct bt_iter_data *iter_data = data;
220 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
221 struct blk_mq_tags *tags = hctx->tags;
222 bool reserved = iter_data->reserved;
223 struct request *rq;
224 bool ret = true;
225
226 if (!reserved)
227 bitnr += tags->nr_reserved_tags;
228 /*
229 * We can hit rq == NULL here, because the tagging functions
230 * test and set the bit before assigning ->rqs[].
231 */
232 rq = blk_mq_find_and_get_req(tags, bitnr);
233 if (!rq)
234 return true;
235
236 if (rq->q == hctx->queue && rq->mq_hctx == hctx)
237 ret = iter_data->fn(hctx, rq, iter_data->data, reserved);
238 blk_mq_put_rq_ref(rq);
239 return ret;
240}
241
242/**
243 * bt_for_each - iterate over the requests associated with a hardware queue
244 * @hctx: Hardware queue to examine.
245 * @bt: sbitmap to examine. This is either the breserved_tags member
246 * or the bitmap_tags member of struct blk_mq_tags.
247 * @fn: Pointer to the function that will be called for each request
248 * associated with @hctx that has been assigned a driver tag.
249 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
250 * where rq is a pointer to a request. Return true to continue
251 * iterating tags, false to stop.
252 * @data: Will be passed as third argument to @fn.
253 * @reserved: Indicates whether @bt is the breserved_tags member or the
254 * bitmap_tags member of struct blk_mq_tags.
255 */
256static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
257 busy_iter_fn *fn, void *data, bool reserved)
258{
259 struct bt_iter_data iter_data = {
260 .hctx = hctx,
261 .fn = fn,
262 .data = data,
263 .reserved = reserved,
264 };
265
266 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
267}
268
269struct bt_tags_iter_data {
270 struct blk_mq_tags *tags;
271 busy_tag_iter_fn *fn;
272 void *data;
273 unsigned int flags;
274};
275
276#define BT_TAG_ITER_RESERVED (1 << 0)
277#define BT_TAG_ITER_STARTED (1 << 1)
278#define BT_TAG_ITER_STATIC_RQS (1 << 2)
279
280static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
281{
282 struct bt_tags_iter_data *iter_data = data;
283 struct blk_mq_tags *tags = iter_data->tags;
284 bool reserved = iter_data->flags & BT_TAG_ITER_RESERVED;
285 struct request *rq;
286 bool ret = true;
287 bool iter_static_rqs = !!(iter_data->flags & BT_TAG_ITER_STATIC_RQS);
288
289 if (!reserved)
290 bitnr += tags->nr_reserved_tags;
291
292 /*
293 * We can hit rq == NULL here, because the tagging functions
294 * test and set the bit before assigning ->rqs[].
295 */
296 if (iter_static_rqs)
297 rq = tags->static_rqs[bitnr];
298 else
299 rq = blk_mq_find_and_get_req(tags, bitnr);
300 if (!rq)
301 return true;
302
303 if (!(iter_data->flags & BT_TAG_ITER_STARTED) ||
304 blk_mq_request_started(rq))
305 ret = iter_data->fn(rq, iter_data->data, reserved);
306 if (!iter_static_rqs)
307 blk_mq_put_rq_ref(rq);
308 return ret;
309}
310
311/**
312 * bt_tags_for_each - iterate over the requests in a tag map
313 * @tags: Tag map to iterate over.
314 * @bt: sbitmap to examine. This is either the breserved_tags member
315 * or the bitmap_tags member of struct blk_mq_tags.
316 * @fn: Pointer to the function that will be called for each started
317 * request. @fn will be called as follows: @fn(rq, @data,
318 * @reserved) where rq is a pointer to a request. Return true
319 * to continue iterating tags, false to stop.
320 * @data: Will be passed as second argument to @fn.
321 * @flags: BT_TAG_ITER_*
322 */
323static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
324 busy_tag_iter_fn *fn, void *data, unsigned int flags)
325{
326 struct bt_tags_iter_data iter_data = {
327 .tags = tags,
328 .fn = fn,
329 .data = data,
330 .flags = flags,
331 };
332
333 if (tags->rqs)
334 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
335}
336
337static void __blk_mq_all_tag_iter(struct blk_mq_tags *tags,
338 busy_tag_iter_fn *fn, void *priv, unsigned int flags)
339{
340 WARN_ON_ONCE(flags & BT_TAG_ITER_RESERVED);
341
342 if (tags->nr_reserved_tags)
343 bt_tags_for_each(tags, tags->breserved_tags, fn, priv,
344 flags | BT_TAG_ITER_RESERVED);
345 bt_tags_for_each(tags, tags->bitmap_tags, fn, priv, flags);
346}
347
348/**
349 * blk_mq_all_tag_iter - iterate over all requests in a tag map
350 * @tags: Tag map to iterate over.
351 * @fn: Pointer to the function that will be called for each
352 * request. @fn will be called as follows: @fn(rq, @priv,
353 * reserved) where rq is a pointer to a request. 'reserved'
354 * indicates whether or not @rq is a reserved request. Return
355 * true to continue iterating tags, false to stop.
356 * @priv: Will be passed as second argument to @fn.
357 *
358 * Caller has to pass the tag map from which requests are allocated.
359 */
360void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
361 void *priv)
362{
363 __blk_mq_all_tag_iter(tags, fn, priv, BT_TAG_ITER_STATIC_RQS);
364}
365
366/**
367 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
368 * @tagset: Tag set to iterate over.
369 * @fn: Pointer to the function that will be called for each started
370 * request. @fn will be called as follows: @fn(rq, @priv,
371 * reserved) where rq is a pointer to a request. 'reserved'
372 * indicates whether or not @rq is a reserved request. Return
373 * true to continue iterating tags, false to stop.
374 * @priv: Will be passed as second argument to @fn.
375 *
376 * We grab one request reference before calling @fn and release it after
377 * @fn returns.
378 */
379void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
380 busy_tag_iter_fn *fn, void *priv)
381{
382 int i;
383
384 for (i = 0; i < tagset->nr_hw_queues; i++) {
385 if (tagset->tags && tagset->tags[i])
386 __blk_mq_all_tag_iter(tagset->tags[i], fn, priv,
387 BT_TAG_ITER_STARTED);
388 }
389}
390EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
391
392static bool blk_mq_tagset_count_completed_rqs(struct request *rq,
393 void *data, bool reserved)
394{
395 unsigned *count = data;
396
397 if (blk_mq_request_completed(rq))
398 (*count)++;
399 return true;
400}
401
402/**
403 * blk_mq_tagset_wait_completed_request - Wait until all scheduled request
404 * completions have finished.
405 * @tagset: Tag set to drain completed request
406 *
407 * Note: This function has to be run after all IO queues are shutdown
408 */
409void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset)
410{
411 while (true) {
412 unsigned count = 0;
413
414 blk_mq_tagset_busy_iter(tagset,
415 blk_mq_tagset_count_completed_rqs, &count);
416 if (!count)
417 break;
418 msleep(5);
419 }
420}
421EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request);
422
423/**
424 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
425 * @q: Request queue to examine.
426 * @fn: Pointer to the function that will be called for each request
427 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
428 * reserved) where rq is a pointer to a request and hctx points
429 * to the hardware queue associated with the request. 'reserved'
430 * indicates whether or not @rq is a reserved request.
431 * @priv: Will be passed as third argument to @fn.
432 *
433 * Note: if @q->tag_set is shared with other request queues then @fn will be
434 * called for all requests on all queues that share that tag set and not only
435 * for requests associated with @q.
436 */
437void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
438 void *priv)
439{
440 struct blk_mq_hw_ctx *hctx;
441 int i;
442
443 /*
444 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and queue_hw_ctx
445 * while the queue is frozen. So we can use q_usage_counter to avoid
446 * racing with it.
447 */
448 if (!percpu_ref_tryget(&q->q_usage_counter))
449 return;
450
451 queue_for_each_hw_ctx(q, hctx, i) {
452 struct blk_mq_tags *tags = hctx->tags;
453
454 /*
455 * If no software queues are currently mapped to this
456 * hardware queue, there's nothing to check
457 */
458 if (!blk_mq_hw_queue_mapped(hctx))
459 continue;
460
461 if (tags->nr_reserved_tags)
462 bt_for_each(hctx, tags->breserved_tags, fn, priv, true);
463 bt_for_each(hctx, tags->bitmap_tags, fn, priv, false);
464 }
465 blk_queue_exit(q);
466}
467
468static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
469 bool round_robin, int node)
470{
471 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
472 node);
473}
474
475int blk_mq_init_bitmaps(struct sbitmap_queue *bitmap_tags,
476 struct sbitmap_queue *breserved_tags,
477 unsigned int queue_depth, unsigned int reserved,
478 int node, int alloc_policy)
479{
480 unsigned int depth = queue_depth - reserved;
481 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
482
483 if (bt_alloc(bitmap_tags, depth, round_robin, node))
484 return -ENOMEM;
485 if (bt_alloc(breserved_tags, reserved, round_robin, node))
486 goto free_bitmap_tags;
487
488 return 0;
489
490free_bitmap_tags:
491 sbitmap_queue_free(bitmap_tags);
492 return -ENOMEM;
493}
494
495static int blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
496 int node, int alloc_policy)
497{
498 int ret;
499
500 ret = blk_mq_init_bitmaps(&tags->__bitmap_tags,
501 &tags->__breserved_tags,
502 tags->nr_tags, tags->nr_reserved_tags,
503 node, alloc_policy);
504 if (ret)
505 return ret;
506
507 tags->bitmap_tags = &tags->__bitmap_tags;
508 tags->breserved_tags = &tags->__breserved_tags;
509
510 return 0;
511}
512
513int blk_mq_init_shared_sbitmap(struct blk_mq_tag_set *set)
514{
515 int alloc_policy = BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags);
516 int i, ret;
517
518 ret = blk_mq_init_bitmaps(&set->__bitmap_tags, &set->__breserved_tags,
519 set->queue_depth, set->reserved_tags,
520 set->numa_node, alloc_policy);
521 if (ret)
522 return ret;
523
524 for (i = 0; i < set->nr_hw_queues; i++) {
525 struct blk_mq_tags *tags = set->tags[i];
526
527 tags->bitmap_tags = &set->__bitmap_tags;
528 tags->breserved_tags = &set->__breserved_tags;
529 }
530
531 return 0;
532}
533
534void blk_mq_exit_shared_sbitmap(struct blk_mq_tag_set *set)
535{
536 sbitmap_queue_free(&set->__bitmap_tags);
537 sbitmap_queue_free(&set->__breserved_tags);
538}
539
540struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
541 unsigned int reserved_tags,
542 int node, unsigned int flags)
543{
544 int alloc_policy = BLK_MQ_FLAG_TO_ALLOC_POLICY(flags);
545 struct blk_mq_tags *tags;
546
547 if (total_tags > BLK_MQ_TAG_MAX) {
548 pr_err("blk-mq: tag depth too large\n");
549 return NULL;
550 }
551
552 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
553 if (!tags)
554 return NULL;
555
556 tags->nr_tags = total_tags;
557 tags->nr_reserved_tags = reserved_tags;
558 spin_lock_init(&tags->lock);
559
560 if (blk_mq_is_sbitmap_shared(flags))
561 return tags;
562
563 if (blk_mq_init_bitmap_tags(tags, node, alloc_policy) < 0) {
564 kfree(tags);
565 return NULL;
566 }
567 return tags;
568}
569
570void blk_mq_free_tags(struct blk_mq_tags *tags, unsigned int flags)
571{
572 if (!blk_mq_is_sbitmap_shared(flags)) {
573 sbitmap_queue_free(tags->bitmap_tags);
574 sbitmap_queue_free(tags->breserved_tags);
575 }
576 kfree(tags);
577}
578
579int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
580 struct blk_mq_tags **tagsptr, unsigned int tdepth,
581 bool can_grow)
582{
583 struct blk_mq_tags *tags = *tagsptr;
584
585 if (tdepth <= tags->nr_reserved_tags)
586 return -EINVAL;
587
588 /*
589 * If we are allowed to grow beyond the original size, allocate
590 * a new set of tags before freeing the old one.
591 */
592 if (tdepth > tags->nr_tags) {
593 struct blk_mq_tag_set *set = hctx->queue->tag_set;
594 struct blk_mq_tags *new;
595 bool ret;
596
597 if (!can_grow)
598 return -EINVAL;
599
600 /*
601 * We need some sort of upper limit, set it high enough that
602 * no valid use cases should require more.
603 */
604 if (tdepth > MAX_SCHED_RQ)
605 return -EINVAL;
606
607 new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth,
608 tags->nr_reserved_tags, set->flags);
609 if (!new)
610 return -ENOMEM;
611 ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth);
612 if (ret) {
613 blk_mq_free_rq_map(new, set->flags);
614 return -ENOMEM;
615 }
616
617 blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
618 blk_mq_free_rq_map(*tagsptr, set->flags);
619 *tagsptr = new;
620 } else {
621 /*
622 * Don't need (or can't) update reserved tags here, they
623 * remain static and should never need resizing.
624 */
625 sbitmap_queue_resize(tags->bitmap_tags,
626 tdepth - tags->nr_reserved_tags);
627 }
628
629 return 0;
630}
631
632void blk_mq_tag_resize_shared_sbitmap(struct blk_mq_tag_set *set, unsigned int size)
633{
634 sbitmap_queue_resize(&set->__bitmap_tags, size - set->reserved_tags);
635}
636
637/**
638 * blk_mq_unique_tag() - return a tag that is unique queue-wide
639 * @rq: request for which to compute a unique tag
640 *
641 * The tag field in struct request is unique per hardware queue but not over
642 * all hardware queues. Hence this function that returns a tag with the
643 * hardware context index in the upper bits and the per hardware queue tag in
644 * the lower bits.
645 *
646 * Note: When called for a request that is queued on a non-multiqueue request
647 * queue, the hardware context index is set to zero.
648 */
649u32 blk_mq_unique_tag(struct request *rq)
650{
651 return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) |
652 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
653}
654EXPORT_SYMBOL(blk_mq_unique_tag);