1 // SPDX-License-Identifier: GPL-2.0
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
7 * Copyright (C) 2013-2014 Jens Axboe
9 #include <linux/kernel.h>
10 #include <linux/module.h>
12 #include <linux/blk-mq.h>
15 #include "blk-mq-tag.h"
17 bool blk_mq_has_free_tags(struct blk_mq_tags
*tags
)
22 return sbitmap_any_bit_clear(&tags
->bitmap_tags
.sb
);
26 * If a previously inactive queue goes active, bump the active user count.
27 * We need to do this before try to allocate driver tag, then even if fail
28 * to get tag when first time, the other shared-tag users could reserve
31 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx
*hctx
)
33 if (!test_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
) &&
34 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
35 atomic_inc(&hctx
->tags
->active_queues
);
41 * Wakeup all potentially sleeping on tags
43 void blk_mq_tag_wakeup_all(struct blk_mq_tags
*tags
, bool include_reserve
)
45 sbitmap_queue_wake_all(&tags
->bitmap_tags
);
47 sbitmap_queue_wake_all(&tags
->breserved_tags
);
51 * If a previously busy queue goes inactive, potential waiters could now
52 * be allowed to queue. Wake them up and check.
54 void __blk_mq_tag_idle(struct blk_mq_hw_ctx
*hctx
)
56 struct blk_mq_tags
*tags
= hctx
->tags
;
58 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
61 atomic_dec(&tags
->active_queues
);
63 blk_mq_tag_wakeup_all(tags
, false);
67 * For shared tag users, we track the number of currently active users
68 * and attempt to provide a fair share of the tag depth for each of them.
70 static inline bool hctx_may_queue(struct blk_mq_hw_ctx
*hctx
,
71 struct sbitmap_queue
*bt
)
73 unsigned int depth
, users
;
75 if (!hctx
|| !(hctx
->flags
& BLK_MQ_F_TAG_SHARED
))
77 if (!test_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
81 * Don't try dividing an ant
83 if (bt
->sb
.depth
== 1)
86 users
= atomic_read(&hctx
->tags
->active_queues
);
91 * Allow at least some tags
93 depth
= max((bt
->sb
.depth
+ users
- 1) / users
, 4U);
94 return atomic_read(&hctx
->nr_active
) < depth
;
97 static int __blk_mq_get_tag(struct blk_mq_alloc_data
*data
,
98 struct sbitmap_queue
*bt
)
100 if (!(data
->flags
& BLK_MQ_REQ_INTERNAL
) &&
101 !hctx_may_queue(data
->hctx
, bt
))
103 if (data
->shallow_depth
)
104 return __sbitmap_queue_get_shallow(bt
, data
->shallow_depth
);
106 return __sbitmap_queue_get(bt
);
109 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data
*data
)
111 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
112 struct sbitmap_queue
*bt
;
113 struct sbq_wait_state
*ws
;
114 DEFINE_SBQ_WAIT(wait
);
115 unsigned int tag_offset
;
119 if (data
->flags
& BLK_MQ_REQ_RESERVED
) {
120 if (unlikely(!tags
->nr_reserved_tags
)) {
122 return BLK_MQ_TAG_FAIL
;
124 bt
= &tags
->breserved_tags
;
127 bt
= &tags
->bitmap_tags
;
128 tag_offset
= tags
->nr_reserved_tags
;
131 tag
= __blk_mq_get_tag(data
, bt
);
135 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
136 return BLK_MQ_TAG_FAIL
;
138 ws
= bt_wait_ptr(bt
, data
->hctx
);
139 drop_ctx
= data
->ctx
== NULL
;
141 struct sbitmap_queue
*bt_prev
;
144 * We're out of tags on this hardware queue, kick any
145 * pending IO submits before going to sleep waiting for
148 blk_mq_run_hw_queue(data
->hctx
, false);
151 * Retry tag allocation after running the hardware queue,
152 * as running the queue may also have found completions.
154 tag
= __blk_mq_get_tag(data
, bt
);
158 sbitmap_prepare_to_wait(bt
, ws
, &wait
, TASK_UNINTERRUPTIBLE
);
160 tag
= __blk_mq_get_tag(data
, bt
);
165 blk_mq_put_ctx(data
->ctx
);
170 sbitmap_finish_wait(bt
, ws
, &wait
);
172 data
->ctx
= blk_mq_get_ctx(data
->q
);
173 data
->hctx
= blk_mq_map_queue(data
->q
, data
->cmd_flags
,
175 tags
= blk_mq_tags_from_data(data
);
176 if (data
->flags
& BLK_MQ_REQ_RESERVED
)
177 bt
= &tags
->breserved_tags
;
179 bt
= &tags
->bitmap_tags
;
182 * If destination hw queue is changed, fake wake up on
183 * previous queue for compensating the wake up miss, so
184 * other allocations on previous queue won't be starved.
187 sbitmap_queue_wake_up(bt_prev
);
189 ws
= bt_wait_ptr(bt
, data
->hctx
);
192 if (drop_ctx
&& data
->ctx
)
193 blk_mq_put_ctx(data
->ctx
);
195 sbitmap_finish_wait(bt
, ws
, &wait
);
198 return tag
+ tag_offset
;
201 void blk_mq_put_tag(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_tags
*tags
,
202 struct blk_mq_ctx
*ctx
, unsigned int tag
)
204 if (!blk_mq_tag_is_reserved(tags
, tag
)) {
205 const int real_tag
= tag
- tags
->nr_reserved_tags
;
207 BUG_ON(real_tag
>= tags
->nr_tags
);
208 sbitmap_queue_clear(&tags
->bitmap_tags
, real_tag
, ctx
->cpu
);
210 BUG_ON(tag
>= tags
->nr_reserved_tags
);
211 sbitmap_queue_clear(&tags
->breserved_tags
, tag
, ctx
->cpu
);
215 struct bt_iter_data
{
216 struct blk_mq_hw_ctx
*hctx
;
222 static bool bt_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
224 struct bt_iter_data
*iter_data
= data
;
225 struct blk_mq_hw_ctx
*hctx
= iter_data
->hctx
;
226 struct blk_mq_tags
*tags
= hctx
->tags
;
227 bool reserved
= iter_data
->reserved
;
231 bitnr
+= tags
->nr_reserved_tags
;
232 rq
= tags
->rqs
[bitnr
];
235 * We can hit rq == NULL here, because the tagging functions
236 * test and set the bit before assigning ->rqs[].
238 if (rq
&& rq
->q
== hctx
->queue
)
239 return iter_data
->fn(hctx
, rq
, iter_data
->data
, reserved
);
244 * bt_for_each - iterate over the requests associated with a hardware queue
245 * @hctx: Hardware queue to examine.
246 * @bt: sbitmap to examine. This is either the breserved_tags member
247 * or the bitmap_tags member of struct blk_mq_tags.
248 * @fn: Pointer to the function that will be called for each request
249 * associated with @hctx that has been assigned a driver tag.
250 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
251 * where rq is a pointer to a request. Return true to continue
252 * iterating tags, false to stop.
253 * @data: Will be passed as third argument to @fn.
254 * @reserved: Indicates whether @bt is the breserved_tags member or the
255 * bitmap_tags member of struct blk_mq_tags.
257 static void bt_for_each(struct blk_mq_hw_ctx
*hctx
, struct sbitmap_queue
*bt
,
258 busy_iter_fn
*fn
, void *data
, bool reserved
)
260 struct bt_iter_data iter_data
= {
264 .reserved
= reserved
,
267 sbitmap_for_each_set(&bt
->sb
, bt_iter
, &iter_data
);
270 struct bt_tags_iter_data
{
271 struct blk_mq_tags
*tags
;
272 busy_tag_iter_fn
*fn
;
277 static bool bt_tags_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
279 struct bt_tags_iter_data
*iter_data
= data
;
280 struct blk_mq_tags
*tags
= iter_data
->tags
;
281 bool reserved
= iter_data
->reserved
;
285 bitnr
+= tags
->nr_reserved_tags
;
288 * We can hit rq == NULL here, because the tagging functions
289 * test and set the bit before assining ->rqs[].
291 rq
= tags
->rqs
[bitnr
];
292 if (rq
&& blk_mq_request_started(rq
))
293 return iter_data
->fn(rq
, iter_data
->data
, reserved
);
299 * bt_tags_for_each - iterate over the requests in a tag map
300 * @tags: Tag map to iterate over.
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 started
304 * request. @fn will be called as follows: @fn(rq, @data,
305 * @reserved) where rq is a pointer to a request. Return true
306 * to continue iterating tags, false to stop.
307 * @data: Will be passed as second argument to @fn.
308 * @reserved: Indicates whether @bt is the breserved_tags member or the
309 * bitmap_tags member of struct blk_mq_tags.
311 static void bt_tags_for_each(struct blk_mq_tags
*tags
, struct sbitmap_queue
*bt
,
312 busy_tag_iter_fn
*fn
, void *data
, bool reserved
)
314 struct bt_tags_iter_data iter_data
= {
318 .reserved
= reserved
,
322 sbitmap_for_each_set(&bt
->sb
, bt_tags_iter
, &iter_data
);
326 * blk_mq_all_tag_busy_iter - iterate over all started requests in a tag map
327 * @tags: Tag map to iterate over.
328 * @fn: Pointer to the function that will be called for each started
329 * request. @fn will be called as follows: @fn(rq, @priv,
330 * reserved) where rq is a pointer to a request. 'reserved'
331 * indicates whether or not @rq is a reserved request. Return
332 * true to continue iterating tags, false to stop.
333 * @priv: Will be passed as second argument to @fn.
335 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags
*tags
,
336 busy_tag_iter_fn
*fn
, void *priv
)
338 if (tags
->nr_reserved_tags
)
339 bt_tags_for_each(tags
, &tags
->breserved_tags
, fn
, priv
, true);
340 bt_tags_for_each(tags
, &tags
->bitmap_tags
, fn
, priv
, false);
344 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
345 * @tagset: Tag set to iterate over.
346 * @fn: Pointer to the function that will be called for each started
347 * request. @fn will be called as follows: @fn(rq, @priv,
348 * reserved) where rq is a pointer to a request. 'reserved'
349 * indicates whether or not @rq is a reserved request. Return
350 * true to continue iterating tags, false to stop.
351 * @priv: Will be passed as second argument to @fn.
353 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set
*tagset
,
354 busy_tag_iter_fn
*fn
, void *priv
)
358 for (i
= 0; i
< tagset
->nr_hw_queues
; i
++) {
359 if (tagset
->tags
&& tagset
->tags
[i
])
360 blk_mq_all_tag_busy_iter(tagset
->tags
[i
], fn
, priv
);
363 EXPORT_SYMBOL(blk_mq_tagset_busy_iter
);
366 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
367 * @q: Request queue to examine.
368 * @fn: Pointer to the function that will be called for each request
369 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
370 * reserved) where rq is a pointer to a request and hctx points
371 * to the hardware queue associated with the request. 'reserved'
372 * indicates whether or not @rq is a reserved request.
373 * @priv: Will be passed as third argument to @fn.
375 * Note: if @q->tag_set is shared with other request queues then @fn will be
376 * called for all requests on all queues that share that tag set and not only
377 * for requests associated with @q.
379 void blk_mq_queue_tag_busy_iter(struct request_queue
*q
, busy_iter_fn
*fn
,
382 struct blk_mq_hw_ctx
*hctx
;
386 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and queue_hw_ctx
387 * while the queue is frozen. So we can use q_usage_counter to avoid
388 * racing with it. __blk_mq_update_nr_hw_queues() uses
389 * synchronize_rcu() to ensure this function left the critical section
392 if (!percpu_ref_tryget(&q
->q_usage_counter
))
395 queue_for_each_hw_ctx(q
, hctx
, i
) {
396 struct blk_mq_tags
*tags
= hctx
->tags
;
399 * If no software queues are currently mapped to this
400 * hardware queue, there's nothing to check
402 if (!blk_mq_hw_queue_mapped(hctx
))
405 if (tags
->nr_reserved_tags
)
406 bt_for_each(hctx
, &tags
->breserved_tags
, fn
, priv
, true);
407 bt_for_each(hctx
, &tags
->bitmap_tags
, fn
, priv
, false);
412 static int bt_alloc(struct sbitmap_queue
*bt
, unsigned int depth
,
413 bool round_robin
, int node
)
415 return sbitmap_queue_init_node(bt
, depth
, -1, round_robin
, GFP_KERNEL
,
419 static struct blk_mq_tags
*blk_mq_init_bitmap_tags(struct blk_mq_tags
*tags
,
420 int node
, int alloc_policy
)
422 unsigned int depth
= tags
->nr_tags
- tags
->nr_reserved_tags
;
423 bool round_robin
= alloc_policy
== BLK_TAG_ALLOC_RR
;
425 if (bt_alloc(&tags
->bitmap_tags
, depth
, round_robin
, node
))
427 if (bt_alloc(&tags
->breserved_tags
, tags
->nr_reserved_tags
, round_robin
,
429 goto free_bitmap_tags
;
433 sbitmap_queue_free(&tags
->bitmap_tags
);
439 struct blk_mq_tags
*blk_mq_init_tags(unsigned int total_tags
,
440 unsigned int reserved_tags
,
441 int node
, int alloc_policy
)
443 struct blk_mq_tags
*tags
;
445 if (total_tags
> BLK_MQ_TAG_MAX
) {
446 pr_err("blk-mq: tag depth too large\n");
450 tags
= kzalloc_node(sizeof(*tags
), GFP_KERNEL
, node
);
454 tags
->nr_tags
= total_tags
;
455 tags
->nr_reserved_tags
= reserved_tags
;
457 return blk_mq_init_bitmap_tags(tags
, node
, alloc_policy
);
460 void blk_mq_free_tags(struct blk_mq_tags
*tags
)
462 sbitmap_queue_free(&tags
->bitmap_tags
);
463 sbitmap_queue_free(&tags
->breserved_tags
);
467 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx
*hctx
,
468 struct blk_mq_tags
**tagsptr
, unsigned int tdepth
,
471 struct blk_mq_tags
*tags
= *tagsptr
;
473 if (tdepth
<= tags
->nr_reserved_tags
)
477 * If we are allowed to grow beyond the original size, allocate
478 * a new set of tags before freeing the old one.
480 if (tdepth
> tags
->nr_tags
) {
481 struct blk_mq_tag_set
*set
= hctx
->queue
->tag_set
;
482 struct blk_mq_tags
*new;
489 * We need some sort of upper limit, set it high enough that
490 * no valid use cases should require more.
492 if (tdepth
> 16 * BLKDEV_MAX_RQ
)
495 new = blk_mq_alloc_rq_map(set
, hctx
->queue_num
, tdepth
,
496 tags
->nr_reserved_tags
);
499 ret
= blk_mq_alloc_rqs(set
, new, hctx
->queue_num
, tdepth
);
501 blk_mq_free_rq_map(new);
505 blk_mq_free_rqs(set
, *tagsptr
, hctx
->queue_num
);
506 blk_mq_free_rq_map(*tagsptr
);
510 * Don't need (or can't) update reserved tags here, they
511 * remain static and should never need resizing.
513 sbitmap_queue_resize(&tags
->bitmap_tags
,
514 tdepth
- tags
->nr_reserved_tags
);
521 * blk_mq_unique_tag() - return a tag that is unique queue-wide
522 * @rq: request for which to compute a unique tag
524 * The tag field in struct request is unique per hardware queue but not over
525 * all hardware queues. Hence this function that returns a tag with the
526 * hardware context index in the upper bits and the per hardware queue tag in
529 * Note: When called for a request that is queued on a non-multiqueue request
530 * queue, the hardware context index is set to zero.
532 u32
blk_mq_unique_tag(struct request
*rq
)
534 return (rq
->mq_hctx
->queue_num
<< BLK_MQ_UNIQUE_TAG_BITS
) |
535 (rq
->tag
& BLK_MQ_UNIQUE_TAG_MASK
);
537 EXPORT_SYMBOL(blk_mq_unique_tag
);