Loading...
1#include <linux/kernel.h>
2#include <linux/module.h>
3#include <linux/percpu_ida.h>
4
5#include <linux/blk-mq.h>
6#include "blk.h"
7#include "blk-mq.h"
8#include "blk-mq-tag.h"
9
10/*
11 * Per tagged queue (tag address space) map
12 */
13struct blk_mq_tags {
14 unsigned int nr_tags;
15 unsigned int nr_reserved_tags;
16 unsigned int nr_batch_move;
17 unsigned int nr_max_cache;
18
19 struct percpu_ida free_tags;
20 struct percpu_ida reserved_tags;
21};
22
23void blk_mq_wait_for_tags(struct blk_mq_tags *tags)
24{
25 int tag = blk_mq_get_tag(tags, __GFP_WAIT, false);
26 blk_mq_put_tag(tags, tag);
27}
28
29bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
30{
31 return !tags ||
32 percpu_ida_free_tags(&tags->free_tags, nr_cpu_ids) != 0;
33}
34
35static unsigned int __blk_mq_get_tag(struct blk_mq_tags *tags, gfp_t gfp)
36{
37 int tag;
38
39 tag = percpu_ida_alloc(&tags->free_tags, (gfp & __GFP_WAIT) ?
40 TASK_UNINTERRUPTIBLE : TASK_RUNNING);
41 if (tag < 0)
42 return BLK_MQ_TAG_FAIL;
43 return tag + tags->nr_reserved_tags;
44}
45
46static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_tags *tags,
47 gfp_t gfp)
48{
49 int tag;
50
51 if (unlikely(!tags->nr_reserved_tags)) {
52 WARN_ON_ONCE(1);
53 return BLK_MQ_TAG_FAIL;
54 }
55
56 tag = percpu_ida_alloc(&tags->reserved_tags, (gfp & __GFP_WAIT) ?
57 TASK_UNINTERRUPTIBLE : TASK_RUNNING);
58 if (tag < 0)
59 return BLK_MQ_TAG_FAIL;
60 return tag;
61}
62
63unsigned int blk_mq_get_tag(struct blk_mq_tags *tags, gfp_t gfp, bool reserved)
64{
65 if (!reserved)
66 return __blk_mq_get_tag(tags, gfp);
67
68 return __blk_mq_get_reserved_tag(tags, gfp);
69}
70
71static void __blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
72{
73 BUG_ON(tag >= tags->nr_tags);
74
75 percpu_ida_free(&tags->free_tags, tag - tags->nr_reserved_tags);
76}
77
78static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags,
79 unsigned int tag)
80{
81 BUG_ON(tag >= tags->nr_reserved_tags);
82
83 percpu_ida_free(&tags->reserved_tags, tag);
84}
85
86void blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
87{
88 if (tag >= tags->nr_reserved_tags)
89 __blk_mq_put_tag(tags, tag);
90 else
91 __blk_mq_put_reserved_tag(tags, tag);
92}
93
94static int __blk_mq_tag_iter(unsigned id, void *data)
95{
96 unsigned long *tag_map = data;
97 __set_bit(id, tag_map);
98 return 0;
99}
100
101void blk_mq_tag_busy_iter(struct blk_mq_tags *tags,
102 void (*fn)(void *, unsigned long *), void *data)
103{
104 unsigned long *tag_map;
105 size_t map_size;
106
107 map_size = ALIGN(tags->nr_tags, BITS_PER_LONG) / BITS_PER_LONG;
108 tag_map = kzalloc(map_size * sizeof(unsigned long), GFP_ATOMIC);
109 if (!tag_map)
110 return;
111
112 percpu_ida_for_each_free(&tags->free_tags, __blk_mq_tag_iter, tag_map);
113 if (tags->nr_reserved_tags)
114 percpu_ida_for_each_free(&tags->reserved_tags, __blk_mq_tag_iter,
115 tag_map);
116
117 fn(data, tag_map);
118 kfree(tag_map);
119}
120
121struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
122 unsigned int reserved_tags, int node)
123{
124 unsigned int nr_tags, nr_cache;
125 struct blk_mq_tags *tags;
126 int ret;
127
128 if (total_tags > BLK_MQ_TAG_MAX) {
129 pr_err("blk-mq: tag depth too large\n");
130 return NULL;
131 }
132
133 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
134 if (!tags)
135 return NULL;
136
137 nr_tags = total_tags - reserved_tags;
138 nr_cache = nr_tags / num_possible_cpus();
139
140 if (nr_cache < BLK_MQ_TAG_CACHE_MIN)
141 nr_cache = BLK_MQ_TAG_CACHE_MIN;
142 else if (nr_cache > BLK_MQ_TAG_CACHE_MAX)
143 nr_cache = BLK_MQ_TAG_CACHE_MAX;
144
145 tags->nr_tags = total_tags;
146 tags->nr_reserved_tags = reserved_tags;
147 tags->nr_max_cache = nr_cache;
148 tags->nr_batch_move = max(1u, nr_cache / 2);
149
150 ret = __percpu_ida_init(&tags->free_tags, tags->nr_tags -
151 tags->nr_reserved_tags,
152 tags->nr_max_cache,
153 tags->nr_batch_move);
154 if (ret)
155 goto err_free_tags;
156
157 if (reserved_tags) {
158 /*
159 * With max_cahe and batch set to 1, the allocator fallbacks to
160 * no cached. It's fine reserved tags allocation is slow.
161 */
162 ret = __percpu_ida_init(&tags->reserved_tags, reserved_tags,
163 1, 1);
164 if (ret)
165 goto err_reserved_tags;
166 }
167
168 return tags;
169
170err_reserved_tags:
171 percpu_ida_destroy(&tags->free_tags);
172err_free_tags:
173 kfree(tags);
174 return NULL;
175}
176
177void blk_mq_free_tags(struct blk_mq_tags *tags)
178{
179 percpu_ida_destroy(&tags->free_tags);
180 percpu_ida_destroy(&tags->reserved_tags);
181 kfree(tags);
182}
183
184ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
185{
186 char *orig_page = page;
187 unsigned int cpu;
188
189 if (!tags)
190 return 0;
191
192 page += sprintf(page, "nr_tags=%u, reserved_tags=%u, batch_move=%u,"
193 " max_cache=%u\n", tags->nr_tags, tags->nr_reserved_tags,
194 tags->nr_batch_move, tags->nr_max_cache);
195
196 page += sprintf(page, "nr_free=%u, nr_reserved=%u\n",
197 percpu_ida_free_tags(&tags->free_tags, nr_cpu_ids),
198 percpu_ida_free_tags(&tags->reserved_tags, nr_cpu_ids));
199
200 for_each_possible_cpu(cpu) {
201 page += sprintf(page, " cpu%02u: nr_free=%u\n", cpu,
202 percpu_ida_free_tags(&tags->free_tags, cpu));
203 }
204
205 return page - orig_page;
206}
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/delay.h>
13#include "blk.h"
14#include "blk-mq.h"
15#include "blk-mq-sched.h"
16
17/*
18 * Recalculate wakeup batch when tag is shared by hctx.
19 */
20static void blk_mq_update_wake_batch(struct blk_mq_tags *tags,
21 unsigned int users)
22{
23 if (!users)
24 return;
25
26 sbitmap_queue_recalculate_wake_batch(&tags->bitmap_tags,
27 users);
28 sbitmap_queue_recalculate_wake_batch(&tags->breserved_tags,
29 users);
30}
31
32/*
33 * If a previously inactive queue goes active, bump the active user count.
34 * We need to do this before try to allocate driver tag, then even if fail
35 * to get tag when first time, the other shared-tag users could reserve
36 * budget for it.
37 */
38void __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
39{
40 unsigned int users;
41 struct blk_mq_tags *tags = hctx->tags;
42
43 /*
44 * calling test_bit() prior to test_and_set_bit() is intentional,
45 * it avoids dirtying the cacheline if the queue is already active.
46 */
47 if (blk_mq_is_shared_tags(hctx->flags)) {
48 struct request_queue *q = hctx->queue;
49
50 if (test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags) ||
51 test_and_set_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
52 return;
53 } else {
54 if (test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) ||
55 test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
56 return;
57 }
58
59 spin_lock_irq(&tags->lock);
60 users = tags->active_queues + 1;
61 WRITE_ONCE(tags->active_queues, users);
62 blk_mq_update_wake_batch(tags, users);
63 spin_unlock_irq(&tags->lock);
64}
65
66/*
67 * Wakeup all potentially sleeping on tags
68 */
69void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
70{
71 sbitmap_queue_wake_all(&tags->bitmap_tags);
72 if (include_reserve)
73 sbitmap_queue_wake_all(&tags->breserved_tags);
74}
75
76/*
77 * If a previously busy queue goes inactive, potential waiters could now
78 * be allowed to queue. Wake them up and check.
79 */
80void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
81{
82 struct blk_mq_tags *tags = hctx->tags;
83 unsigned int users;
84
85 if (blk_mq_is_shared_tags(hctx->flags)) {
86 struct request_queue *q = hctx->queue;
87
88 if (!test_and_clear_bit(QUEUE_FLAG_HCTX_ACTIVE,
89 &q->queue_flags))
90 return;
91 } else {
92 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
93 return;
94 }
95
96 spin_lock_irq(&tags->lock);
97 users = tags->active_queues - 1;
98 WRITE_ONCE(tags->active_queues, users);
99 blk_mq_update_wake_batch(tags, users);
100 spin_unlock_irq(&tags->lock);
101
102 blk_mq_tag_wakeup_all(tags, false);
103}
104
105static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
106 struct sbitmap_queue *bt)
107{
108 if (!data->q->elevator && !(data->flags & BLK_MQ_REQ_RESERVED) &&
109 !hctx_may_queue(data->hctx, bt))
110 return BLK_MQ_NO_TAG;
111
112 if (data->shallow_depth)
113 return sbitmap_queue_get_shallow(bt, data->shallow_depth);
114 else
115 return __sbitmap_queue_get(bt);
116}
117
118unsigned long blk_mq_get_tags(struct blk_mq_alloc_data *data, int nr_tags,
119 unsigned int *offset)
120{
121 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
122 struct sbitmap_queue *bt = &tags->bitmap_tags;
123 unsigned long ret;
124
125 if (data->shallow_depth ||data->flags & BLK_MQ_REQ_RESERVED ||
126 data->hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
127 return 0;
128 ret = __sbitmap_queue_get_batch(bt, nr_tags, offset);
129 *offset += tags->nr_reserved_tags;
130 return ret;
131}
132
133unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
134{
135 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
136 struct sbitmap_queue *bt;
137 struct sbq_wait_state *ws;
138 DEFINE_SBQ_WAIT(wait);
139 unsigned int tag_offset;
140 int tag;
141
142 if (data->flags & BLK_MQ_REQ_RESERVED) {
143 if (unlikely(!tags->nr_reserved_tags)) {
144 WARN_ON_ONCE(1);
145 return BLK_MQ_NO_TAG;
146 }
147 bt = &tags->breserved_tags;
148 tag_offset = 0;
149 } else {
150 bt = &tags->bitmap_tags;
151 tag_offset = tags->nr_reserved_tags;
152 }
153
154 tag = __blk_mq_get_tag(data, bt);
155 if (tag != BLK_MQ_NO_TAG)
156 goto found_tag;
157
158 if (data->flags & BLK_MQ_REQ_NOWAIT)
159 return BLK_MQ_NO_TAG;
160
161 ws = bt_wait_ptr(bt, data->hctx);
162 do {
163 struct sbitmap_queue *bt_prev;
164
165 /*
166 * We're out of tags on this hardware queue, kick any
167 * pending IO submits before going to sleep waiting for
168 * some to complete.
169 */
170 blk_mq_run_hw_queue(data->hctx, false);
171
172 /*
173 * Retry tag allocation after running the hardware queue,
174 * as running the queue may also have found completions.
175 */
176 tag = __blk_mq_get_tag(data, bt);
177 if (tag != BLK_MQ_NO_TAG)
178 break;
179
180 sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
181
182 tag = __blk_mq_get_tag(data, bt);
183 if (tag != BLK_MQ_NO_TAG)
184 break;
185
186 bt_prev = bt;
187 io_schedule();
188
189 sbitmap_finish_wait(bt, ws, &wait);
190
191 data->ctx = blk_mq_get_ctx(data->q);
192 data->hctx = blk_mq_map_queue(data->q, data->cmd_flags,
193 data->ctx);
194 tags = blk_mq_tags_from_data(data);
195 if (data->flags & BLK_MQ_REQ_RESERVED)
196 bt = &tags->breserved_tags;
197 else
198 bt = &tags->bitmap_tags;
199
200 /*
201 * If destination hw queue is changed, fake wake up on
202 * previous queue for compensating the wake up miss, so
203 * other allocations on previous queue won't be starved.
204 */
205 if (bt != bt_prev)
206 sbitmap_queue_wake_up(bt_prev, 1);
207
208 ws = bt_wait_ptr(bt, data->hctx);
209 } while (1);
210
211 sbitmap_finish_wait(bt, ws, &wait);
212
213found_tag:
214 /*
215 * Give up this allocation if the hctx is inactive. The caller will
216 * retry on an active hctx.
217 */
218 if (unlikely(test_bit(BLK_MQ_S_INACTIVE, &data->hctx->state))) {
219 blk_mq_put_tag(tags, data->ctx, tag + tag_offset);
220 return BLK_MQ_NO_TAG;
221 }
222 return tag + tag_offset;
223}
224
225void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
226 unsigned int tag)
227{
228 if (!blk_mq_tag_is_reserved(tags, tag)) {
229 const int real_tag = tag - tags->nr_reserved_tags;
230
231 BUG_ON(real_tag >= tags->nr_tags);
232 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
233 } else {
234 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
235 }
236}
237
238void blk_mq_put_tags(struct blk_mq_tags *tags, int *tag_array, int nr_tags)
239{
240 sbitmap_queue_clear_batch(&tags->bitmap_tags, tags->nr_reserved_tags,
241 tag_array, nr_tags);
242}
243
244struct bt_iter_data {
245 struct blk_mq_hw_ctx *hctx;
246 struct request_queue *q;
247 busy_tag_iter_fn *fn;
248 void *data;
249 bool reserved;
250};
251
252static struct request *blk_mq_find_and_get_req(struct blk_mq_tags *tags,
253 unsigned int bitnr)
254{
255 struct request *rq;
256 unsigned long flags;
257
258 spin_lock_irqsave(&tags->lock, flags);
259 rq = tags->rqs[bitnr];
260 if (!rq || rq->tag != bitnr || !req_ref_inc_not_zero(rq))
261 rq = NULL;
262 spin_unlock_irqrestore(&tags->lock, flags);
263 return rq;
264}
265
266static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
267{
268 struct bt_iter_data *iter_data = data;
269 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
270 struct request_queue *q = iter_data->q;
271 struct blk_mq_tag_set *set = q->tag_set;
272 struct blk_mq_tags *tags;
273 struct request *rq;
274 bool ret = true;
275
276 if (blk_mq_is_shared_tags(set->flags))
277 tags = set->shared_tags;
278 else
279 tags = hctx->tags;
280
281 if (!iter_data->reserved)
282 bitnr += tags->nr_reserved_tags;
283 /*
284 * We can hit rq == NULL here, because the tagging functions
285 * test and set the bit before assigning ->rqs[].
286 */
287 rq = blk_mq_find_and_get_req(tags, bitnr);
288 if (!rq)
289 return true;
290
291 if (rq->q == q && (!hctx || rq->mq_hctx == hctx))
292 ret = iter_data->fn(rq, iter_data->data);
293 blk_mq_put_rq_ref(rq);
294 return ret;
295}
296
297/**
298 * bt_for_each - iterate over the requests associated with a hardware queue
299 * @hctx: Hardware queue to examine.
300 * @q: Request queue to examine.
301 * @bt: sbitmap to examine. This is either the breserved_tags member
302 * or the bitmap_tags member of struct blk_mq_tags.
303 * @fn: Pointer to the function that will be called for each request
304 * associated with @hctx that has been assigned a driver tag.
305 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
306 * where rq is a pointer to a request. Return true to continue
307 * iterating tags, false to stop.
308 * @data: Will be passed as third argument to @fn.
309 * @reserved: Indicates whether @bt is the breserved_tags member or the
310 * bitmap_tags member of struct blk_mq_tags.
311 */
312static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct request_queue *q,
313 struct sbitmap_queue *bt, busy_tag_iter_fn *fn,
314 void *data, bool reserved)
315{
316 struct bt_iter_data iter_data = {
317 .hctx = hctx,
318 .fn = fn,
319 .data = data,
320 .reserved = reserved,
321 .q = q,
322 };
323
324 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
325}
326
327struct bt_tags_iter_data {
328 struct blk_mq_tags *tags;
329 busy_tag_iter_fn *fn;
330 void *data;
331 unsigned int flags;
332};
333
334#define BT_TAG_ITER_RESERVED (1 << 0)
335#define BT_TAG_ITER_STARTED (1 << 1)
336#define BT_TAG_ITER_STATIC_RQS (1 << 2)
337
338static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
339{
340 struct bt_tags_iter_data *iter_data = data;
341 struct blk_mq_tags *tags = iter_data->tags;
342 struct request *rq;
343 bool ret = true;
344 bool iter_static_rqs = !!(iter_data->flags & BT_TAG_ITER_STATIC_RQS);
345
346 if (!(iter_data->flags & BT_TAG_ITER_RESERVED))
347 bitnr += tags->nr_reserved_tags;
348
349 /*
350 * We can hit rq == NULL here, because the tagging functions
351 * test and set the bit before assigning ->rqs[].
352 */
353 if (iter_static_rqs)
354 rq = tags->static_rqs[bitnr];
355 else
356 rq = blk_mq_find_and_get_req(tags, bitnr);
357 if (!rq)
358 return true;
359
360 if (!(iter_data->flags & BT_TAG_ITER_STARTED) ||
361 blk_mq_request_started(rq))
362 ret = iter_data->fn(rq, iter_data->data);
363 if (!iter_static_rqs)
364 blk_mq_put_rq_ref(rq);
365 return ret;
366}
367
368/**
369 * bt_tags_for_each - iterate over the requests in a tag map
370 * @tags: Tag map to iterate over.
371 * @bt: sbitmap to examine. This is either the breserved_tags member
372 * or the bitmap_tags member of struct blk_mq_tags.
373 * @fn: Pointer to the function that will be called for each started
374 * request. @fn will be called as follows: @fn(rq, @data,
375 * @reserved) where rq is a pointer to a request. Return true
376 * to continue iterating tags, false to stop.
377 * @data: Will be passed as second argument to @fn.
378 * @flags: BT_TAG_ITER_*
379 */
380static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
381 busy_tag_iter_fn *fn, void *data, unsigned int flags)
382{
383 struct bt_tags_iter_data iter_data = {
384 .tags = tags,
385 .fn = fn,
386 .data = data,
387 .flags = flags,
388 };
389
390 if (tags->rqs)
391 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
392}
393
394static void __blk_mq_all_tag_iter(struct blk_mq_tags *tags,
395 busy_tag_iter_fn *fn, void *priv, unsigned int flags)
396{
397 WARN_ON_ONCE(flags & BT_TAG_ITER_RESERVED);
398
399 if (tags->nr_reserved_tags)
400 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv,
401 flags | BT_TAG_ITER_RESERVED);
402 bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, flags);
403}
404
405/**
406 * blk_mq_all_tag_iter - iterate over all requests in a tag map
407 * @tags: Tag map to iterate over.
408 * @fn: Pointer to the function that will be called for each
409 * request. @fn will be called as follows: @fn(rq, @priv,
410 * reserved) where rq is a pointer to a request. 'reserved'
411 * indicates whether or not @rq is a reserved request. Return
412 * true to continue iterating tags, false to stop.
413 * @priv: Will be passed as second argument to @fn.
414 *
415 * Caller has to pass the tag map from which requests are allocated.
416 */
417void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
418 void *priv)
419{
420 __blk_mq_all_tag_iter(tags, fn, priv, BT_TAG_ITER_STATIC_RQS);
421}
422
423/**
424 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
425 * @tagset: Tag set to iterate over.
426 * @fn: Pointer to the function that will be called for each started
427 * request. @fn will be called as follows: @fn(rq, @priv,
428 * reserved) where rq is a pointer to a request. 'reserved'
429 * indicates whether or not @rq is a reserved request. Return
430 * true to continue iterating tags, false to stop.
431 * @priv: Will be passed as second argument to @fn.
432 *
433 * We grab one request reference before calling @fn and release it after
434 * @fn returns.
435 */
436void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
437 busy_tag_iter_fn *fn, void *priv)
438{
439 unsigned int flags = tagset->flags;
440 int i, nr_tags;
441
442 nr_tags = blk_mq_is_shared_tags(flags) ? 1 : tagset->nr_hw_queues;
443
444 for (i = 0; i < nr_tags; i++) {
445 if (tagset->tags && tagset->tags[i])
446 __blk_mq_all_tag_iter(tagset->tags[i], fn, priv,
447 BT_TAG_ITER_STARTED);
448 }
449}
450EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
451
452static bool blk_mq_tagset_count_completed_rqs(struct request *rq, void *data)
453{
454 unsigned *count = data;
455
456 if (blk_mq_request_completed(rq))
457 (*count)++;
458 return true;
459}
460
461/**
462 * blk_mq_tagset_wait_completed_request - Wait until all scheduled request
463 * completions have finished.
464 * @tagset: Tag set to drain completed request
465 *
466 * Note: This function has to be run after all IO queues are shutdown
467 */
468void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset)
469{
470 while (true) {
471 unsigned count = 0;
472
473 blk_mq_tagset_busy_iter(tagset,
474 blk_mq_tagset_count_completed_rqs, &count);
475 if (!count)
476 break;
477 msleep(5);
478 }
479}
480EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request);
481
482/**
483 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
484 * @q: Request queue to examine.
485 * @fn: Pointer to the function that will be called for each request
486 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
487 * reserved) where rq is a pointer to a request and hctx points
488 * to the hardware queue associated with the request. 'reserved'
489 * indicates whether or not @rq is a reserved request.
490 * @priv: Will be passed as third argument to @fn.
491 *
492 * Note: if @q->tag_set is shared with other request queues then @fn will be
493 * called for all requests on all queues that share that tag set and not only
494 * for requests associated with @q.
495 */
496void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_tag_iter_fn *fn,
497 void *priv)
498{
499 /*
500 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and hctx_table
501 * while the queue is frozen. So we can use q_usage_counter to avoid
502 * racing with it.
503 */
504 if (!percpu_ref_tryget(&q->q_usage_counter))
505 return;
506
507 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
508 struct blk_mq_tags *tags = q->tag_set->shared_tags;
509 struct sbitmap_queue *bresv = &tags->breserved_tags;
510 struct sbitmap_queue *btags = &tags->bitmap_tags;
511
512 if (tags->nr_reserved_tags)
513 bt_for_each(NULL, q, bresv, fn, priv, true);
514 bt_for_each(NULL, q, btags, fn, priv, false);
515 } else {
516 struct blk_mq_hw_ctx *hctx;
517 unsigned long i;
518
519 queue_for_each_hw_ctx(q, hctx, i) {
520 struct blk_mq_tags *tags = hctx->tags;
521 struct sbitmap_queue *bresv = &tags->breserved_tags;
522 struct sbitmap_queue *btags = &tags->bitmap_tags;
523
524 /*
525 * If no software queues are currently mapped to this
526 * hardware queue, there's nothing to check
527 */
528 if (!blk_mq_hw_queue_mapped(hctx))
529 continue;
530
531 if (tags->nr_reserved_tags)
532 bt_for_each(hctx, q, bresv, fn, priv, true);
533 bt_for_each(hctx, q, btags, fn, priv, false);
534 }
535 }
536 blk_queue_exit(q);
537}
538
539static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
540 bool round_robin, int node)
541{
542 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
543 node);
544}
545
546int blk_mq_init_bitmaps(struct sbitmap_queue *bitmap_tags,
547 struct sbitmap_queue *breserved_tags,
548 unsigned int queue_depth, unsigned int reserved,
549 int node, int alloc_policy)
550{
551 unsigned int depth = queue_depth - reserved;
552 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
553
554 if (bt_alloc(bitmap_tags, depth, round_robin, node))
555 return -ENOMEM;
556 if (bt_alloc(breserved_tags, reserved, round_robin, node))
557 goto free_bitmap_tags;
558
559 return 0;
560
561free_bitmap_tags:
562 sbitmap_queue_free(bitmap_tags);
563 return -ENOMEM;
564}
565
566struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
567 unsigned int reserved_tags,
568 int node, int alloc_policy)
569{
570 struct blk_mq_tags *tags;
571
572 if (total_tags > BLK_MQ_TAG_MAX) {
573 pr_err("blk-mq: tag depth too large\n");
574 return NULL;
575 }
576
577 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
578 if (!tags)
579 return NULL;
580
581 tags->nr_tags = total_tags;
582 tags->nr_reserved_tags = reserved_tags;
583 spin_lock_init(&tags->lock);
584
585 if (blk_mq_init_bitmaps(&tags->bitmap_tags, &tags->breserved_tags,
586 total_tags, reserved_tags, node,
587 alloc_policy) < 0) {
588 kfree(tags);
589 return NULL;
590 }
591 return tags;
592}
593
594void blk_mq_free_tags(struct blk_mq_tags *tags)
595{
596 sbitmap_queue_free(&tags->bitmap_tags);
597 sbitmap_queue_free(&tags->breserved_tags);
598 kfree(tags);
599}
600
601int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
602 struct blk_mq_tags **tagsptr, unsigned int tdepth,
603 bool can_grow)
604{
605 struct blk_mq_tags *tags = *tagsptr;
606
607 if (tdepth <= tags->nr_reserved_tags)
608 return -EINVAL;
609
610 /*
611 * If we are allowed to grow beyond the original size, allocate
612 * a new set of tags before freeing the old one.
613 */
614 if (tdepth > tags->nr_tags) {
615 struct blk_mq_tag_set *set = hctx->queue->tag_set;
616 struct blk_mq_tags *new;
617
618 if (!can_grow)
619 return -EINVAL;
620
621 /*
622 * We need some sort of upper limit, set it high enough that
623 * no valid use cases should require more.
624 */
625 if (tdepth > MAX_SCHED_RQ)
626 return -EINVAL;
627
628 /*
629 * Only the sbitmap needs resizing since we allocated the max
630 * initially.
631 */
632 if (blk_mq_is_shared_tags(set->flags))
633 return 0;
634
635 new = blk_mq_alloc_map_and_rqs(set, hctx->queue_num, tdepth);
636 if (!new)
637 return -ENOMEM;
638
639 blk_mq_free_map_and_rqs(set, *tagsptr, hctx->queue_num);
640 *tagsptr = new;
641 } else {
642 /*
643 * Don't need (or can't) update reserved tags here, they
644 * remain static and should never need resizing.
645 */
646 sbitmap_queue_resize(&tags->bitmap_tags,
647 tdepth - tags->nr_reserved_tags);
648 }
649
650 return 0;
651}
652
653void blk_mq_tag_resize_shared_tags(struct blk_mq_tag_set *set, unsigned int size)
654{
655 struct blk_mq_tags *tags = set->shared_tags;
656
657 sbitmap_queue_resize(&tags->bitmap_tags, size - set->reserved_tags);
658}
659
660void blk_mq_tag_update_sched_shared_tags(struct request_queue *q)
661{
662 sbitmap_queue_resize(&q->sched_shared_tags->bitmap_tags,
663 q->nr_requests - q->tag_set->reserved_tags);
664}
665
666/**
667 * blk_mq_unique_tag() - return a tag that is unique queue-wide
668 * @rq: request for which to compute a unique tag
669 *
670 * The tag field in struct request is unique per hardware queue but not over
671 * all hardware queues. Hence this function that returns a tag with the
672 * hardware context index in the upper bits and the per hardware queue tag in
673 * the lower bits.
674 *
675 * Note: When called for a request that is queued on a non-multiqueue request
676 * queue, the hardware context index is set to zero.
677 */
678u32 blk_mq_unique_tag(struct request *rq)
679{
680 return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) |
681 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
682}
683EXPORT_SYMBOL(blk_mq_unique_tag);