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>
13 #include <linux/delay.h>
16 #include "blk-mq-sched.h"
17 #include "blk-mq-tag.h"
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
25 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx
*hctx
)
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
;
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
);
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
);
44 * Wakeup all potentially sleeping on tags
46 void blk_mq_tag_wakeup_all(struct blk_mq_tags
*tags
, bool include_reserve
)
48 sbitmap_queue_wake_all(tags
->bitmap_tags
);
50 sbitmap_queue_wake_all(tags
->breserved_tags
);
54 * If a previously busy queue goes inactive, potential waiters could now
55 * be allowed to queue. Wake them up and check.
57 void __blk_mq_tag_idle(struct blk_mq_hw_ctx
*hctx
)
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
;
63 if (blk_mq_is_sbitmap_shared(hctx
->flags
)) {
64 if (!test_and_clear_bit(QUEUE_FLAG_HCTX_ACTIVE
,
67 atomic_dec(&set
->active_queues_shared_sbitmap
);
69 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
71 atomic_dec(&tags
->active_queues
);
74 blk_mq_tag_wakeup_all(tags
, false);
77 static int __blk_mq_get_tag(struct blk_mq_alloc_data
*data
,
78 struct sbitmap_queue
*bt
)
80 if (!data
->q
->elevator
&& !(data
->flags
& BLK_MQ_REQ_RESERVED
) &&
81 !hctx_may_queue(data
->hctx
, bt
))
84 if (data
->shallow_depth
)
85 return __sbitmap_queue_get_shallow(bt
, data
->shallow_depth
);
87 return __sbitmap_queue_get(bt
);
90 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data
*data
)
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
;
99 if (data
->flags
& BLK_MQ_REQ_RESERVED
) {
100 if (unlikely(!tags
->nr_reserved_tags
)) {
102 return BLK_MQ_NO_TAG
;
104 bt
= tags
->breserved_tags
;
107 bt
= tags
->bitmap_tags
;
108 tag_offset
= tags
->nr_reserved_tags
;
111 tag
= __blk_mq_get_tag(data
, bt
);
112 if (tag
!= BLK_MQ_NO_TAG
)
115 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
116 return BLK_MQ_NO_TAG
;
118 ws
= bt_wait_ptr(bt
, data
->hctx
);
120 struct sbitmap_queue
*bt_prev
;
123 * We're out of tags on this hardware queue, kick any
124 * pending IO submits before going to sleep waiting for
127 blk_mq_run_hw_queue(data
->hctx
, false);
130 * Retry tag allocation after running the hardware queue,
131 * as running the queue may also have found completions.
133 tag
= __blk_mq_get_tag(data
, bt
);
134 if (tag
!= BLK_MQ_NO_TAG
)
137 sbitmap_prepare_to_wait(bt
, ws
, &wait
, TASK_UNINTERRUPTIBLE
);
139 tag
= __blk_mq_get_tag(data
, bt
);
140 if (tag
!= BLK_MQ_NO_TAG
)
146 sbitmap_finish_wait(bt
, ws
, &wait
);
148 data
->ctx
= blk_mq_get_ctx(data
->q
);
149 data
->hctx
= blk_mq_map_queue(data
->q
, data
->cmd_flags
,
151 tags
= blk_mq_tags_from_data(data
);
152 if (data
->flags
& BLK_MQ_REQ_RESERVED
)
153 bt
= tags
->breserved_tags
;
155 bt
= tags
->bitmap_tags
;
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.
163 sbitmap_queue_wake_up(bt_prev
);
165 ws
= bt_wait_ptr(bt
, data
->hctx
);
168 sbitmap_finish_wait(bt
, ws
, &wait
);
172 * Give up this allocation if the hctx is inactive. The caller will
173 * retry on an active hctx.
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
;
179 return tag
+ tag_offset
;
182 void blk_mq_put_tag(struct blk_mq_tags
*tags
, struct blk_mq_ctx
*ctx
,
185 if (!blk_mq_tag_is_reserved(tags
, tag
)) {
186 const int real_tag
= tag
- tags
->nr_reserved_tags
;
188 BUG_ON(real_tag
>= tags
->nr_tags
);
189 sbitmap_queue_clear(tags
->bitmap_tags
, real_tag
, ctx
->cpu
);
191 BUG_ON(tag
>= tags
->nr_reserved_tags
);
192 sbitmap_queue_clear(tags
->breserved_tags
, tag
, ctx
->cpu
);
196 struct bt_iter_data
{
197 struct blk_mq_hw_ctx
*hctx
;
203 static struct request
*blk_mq_find_and_get_req(struct blk_mq_tags
*tags
,
209 spin_lock_irqsave(&tags
->lock
, flags
);
210 rq
= tags
->rqs
[bitnr
];
211 if (!rq
|| rq
->tag
!= bitnr
|| !refcount_inc_not_zero(&rq
->ref
))
213 spin_unlock_irqrestore(&tags
->lock
, flags
);
217 static bool bt_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
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
;
227 bitnr
+= tags
->nr_reserved_tags
;
229 * We can hit rq == NULL here, because the tagging functions
230 * test and set the bit before assigning ->rqs[].
232 rq
= blk_mq_find_and_get_req(tags
, bitnr
);
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
);
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.
256 static void bt_for_each(struct blk_mq_hw_ctx
*hctx
, struct sbitmap_queue
*bt
,
257 busy_iter_fn
*fn
, void *data
, bool reserved
)
259 struct bt_iter_data iter_data
= {
263 .reserved
= reserved
,
266 sbitmap_for_each_set(&bt
->sb
, bt_iter
, &iter_data
);
269 struct bt_tags_iter_data
{
270 struct blk_mq_tags
*tags
;
271 busy_tag_iter_fn
*fn
;
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)
280 static bool bt_tags_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
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
;
287 bool iter_static_rqs
= !!(iter_data
->flags
& BT_TAG_ITER_STATIC_RQS
);
290 bitnr
+= tags
->nr_reserved_tags
;
293 * We can hit rq == NULL here, because the tagging functions
294 * test and set the bit before assigning ->rqs[].
297 rq
= tags
->static_rqs
[bitnr
];
299 rq
= blk_mq_find_and_get_req(tags
, bitnr
);
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
);
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_*
323 static 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
)
326 struct bt_tags_iter_data iter_data
= {
334 sbitmap_for_each_set(&bt
->sb
, bt_tags_iter
, &iter_data
);
337 static void __blk_mq_all_tag_iter(struct blk_mq_tags
*tags
,
338 busy_tag_iter_fn
*fn
, void *priv
, unsigned int flags
)
340 WARN_ON_ONCE(flags
& BT_TAG_ITER_RESERVED
);
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
);
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.
358 * Caller has to pass the tag map from which requests are allocated.
360 void blk_mq_all_tag_iter(struct blk_mq_tags
*tags
, busy_tag_iter_fn
*fn
,
363 __blk_mq_all_tag_iter(tags
, fn
, priv
, BT_TAG_ITER_STATIC_RQS
);
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.
376 * We grab one request reference before calling @fn and release it after
379 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set
*tagset
,
380 busy_tag_iter_fn
*fn
, void *priv
)
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
);
390 EXPORT_SYMBOL(blk_mq_tagset_busy_iter
);
392 static bool blk_mq_tagset_count_completed_rqs(struct request
*rq
,
393 void *data
, bool reserved
)
395 unsigned *count
= data
;
397 if (blk_mq_request_completed(rq
))
403 * blk_mq_tagset_wait_completed_request - Wait until all scheduled request
404 * completions have finished.
405 * @tagset: Tag set to drain completed request
407 * Note: This function has to be run after all IO queues are shutdown
409 void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set
*tagset
)
414 blk_mq_tagset_busy_iter(tagset
,
415 blk_mq_tagset_count_completed_rqs
, &count
);
421 EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request
);
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.
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.
437 void blk_mq_queue_tag_busy_iter(struct request_queue
*q
, busy_iter_fn
*fn
,
440 struct blk_mq_hw_ctx
*hctx
;
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
448 if (!percpu_ref_tryget(&q
->q_usage_counter
))
451 queue_for_each_hw_ctx(q
, hctx
, i
) {
452 struct blk_mq_tags
*tags
= hctx
->tags
;
455 * If no software queues are currently mapped to this
456 * hardware queue, there's nothing to check
458 if (!blk_mq_hw_queue_mapped(hctx
))
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);
468 static int bt_alloc(struct sbitmap_queue
*bt
, unsigned int depth
,
469 bool round_robin
, int node
)
471 return sbitmap_queue_init_node(bt
, depth
, -1, round_robin
, GFP_KERNEL
,
475 int 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
)
480 unsigned int depth
= queue_depth
- reserved
;
481 bool round_robin
= alloc_policy
== BLK_TAG_ALLOC_RR
;
483 if (bt_alloc(bitmap_tags
, depth
, round_robin
, node
))
485 if (bt_alloc(breserved_tags
, reserved
, round_robin
, node
))
486 goto free_bitmap_tags
;
491 sbitmap_queue_free(bitmap_tags
);
495 static int blk_mq_init_bitmap_tags(struct blk_mq_tags
*tags
,
496 int node
, int alloc_policy
)
500 ret
= blk_mq_init_bitmaps(&tags
->__bitmap_tags
,
501 &tags
->__breserved_tags
,
502 tags
->nr_tags
, tags
->nr_reserved_tags
,
507 tags
->bitmap_tags
= &tags
->__bitmap_tags
;
508 tags
->breserved_tags
= &tags
->__breserved_tags
;
513 int blk_mq_init_shared_sbitmap(struct blk_mq_tag_set
*set
)
515 int alloc_policy
= BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
);
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
);
524 for (i
= 0; i
< set
->nr_hw_queues
; i
++) {
525 struct blk_mq_tags
*tags
= set
->tags
[i
];
527 tags
->bitmap_tags
= &set
->__bitmap_tags
;
528 tags
->breserved_tags
= &set
->__breserved_tags
;
534 void blk_mq_exit_shared_sbitmap(struct blk_mq_tag_set
*set
)
536 sbitmap_queue_free(&set
->__bitmap_tags
);
537 sbitmap_queue_free(&set
->__breserved_tags
);
540 struct blk_mq_tags
*blk_mq_init_tags(unsigned int total_tags
,
541 unsigned int reserved_tags
,
542 int node
, unsigned int flags
)
544 int alloc_policy
= BLK_MQ_FLAG_TO_ALLOC_POLICY(flags
);
545 struct blk_mq_tags
*tags
;
547 if (total_tags
> BLK_MQ_TAG_MAX
) {
548 pr_err("blk-mq: tag depth too large\n");
552 tags
= kzalloc_node(sizeof(*tags
), GFP_KERNEL
, node
);
556 tags
->nr_tags
= total_tags
;
557 tags
->nr_reserved_tags
= reserved_tags
;
558 spin_lock_init(&tags
->lock
);
560 if (blk_mq_is_sbitmap_shared(flags
))
563 if (blk_mq_init_bitmap_tags(tags
, node
, alloc_policy
) < 0) {
570 void blk_mq_free_tags(struct blk_mq_tags
*tags
, unsigned int flags
)
572 if (!blk_mq_is_sbitmap_shared(flags
)) {
573 sbitmap_queue_free(tags
->bitmap_tags
);
574 sbitmap_queue_free(tags
->breserved_tags
);
579 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx
*hctx
,
580 struct blk_mq_tags
**tagsptr
, unsigned int tdepth
,
583 struct blk_mq_tags
*tags
= *tagsptr
;
585 if (tdepth
<= tags
->nr_reserved_tags
)
589 * If we are allowed to grow beyond the original size, allocate
590 * a new set of tags before freeing the old one.
592 if (tdepth
> tags
->nr_tags
) {
593 struct blk_mq_tag_set
*set
= hctx
->queue
->tag_set
;
594 struct blk_mq_tags
*new;
601 * We need some sort of upper limit, set it high enough that
602 * no valid use cases should require more.
604 if (tdepth
> MAX_SCHED_RQ
)
607 new = blk_mq_alloc_rq_map(set
, hctx
->queue_num
, tdepth
,
608 tags
->nr_reserved_tags
, set
->flags
);
611 ret
= blk_mq_alloc_rqs(set
, new, hctx
->queue_num
, tdepth
);
613 blk_mq_free_rq_map(new, set
->flags
);
617 blk_mq_free_rqs(set
, *tagsptr
, hctx
->queue_num
);
618 blk_mq_free_rq_map(*tagsptr
, set
->flags
);
622 * Don't need (or can't) update reserved tags here, they
623 * remain static and should never need resizing.
625 sbitmap_queue_resize(tags
->bitmap_tags
,
626 tdepth
- tags
->nr_reserved_tags
);
632 void blk_mq_tag_resize_shared_sbitmap(struct blk_mq_tag_set
*set
, unsigned int size
)
634 sbitmap_queue_resize(&set
->__bitmap_tags
, size
- set
->reserved_tags
);
638 * blk_mq_unique_tag() - return a tag that is unique queue-wide
639 * @rq: request for which to compute a unique tag
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
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.
649 u32
blk_mq_unique_tag(struct request
*rq
)
651 return (rq
->mq_hctx
->queue_num
<< BLK_MQ_UNIQUE_TAG_BITS
) |
652 (rq
->tag
& BLK_MQ_UNIQUE_TAG_MASK
);
654 EXPORT_SYMBOL(blk_mq_unique_tag
);