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 * Recalculate wakeup batch when tag is shared by hctx.
22 static void blk_mq_update_wake_batch(struct blk_mq_tags
*tags
,
28 sbitmap_queue_recalculate_wake_batch(&tags
->bitmap_tags
,
30 sbitmap_queue_recalculate_wake_batch(&tags
->breserved_tags
,
35 * If a previously inactive queue goes active, bump the active user count.
36 * We need to do this before try to allocate driver tag, then even if fail
37 * to get tag when first time, the other shared-tag users could reserve
40 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx
*hctx
)
44 if (blk_mq_is_shared_tags(hctx
->flags
)) {
45 struct request_queue
*q
= hctx
->queue
;
47 if (test_bit(QUEUE_FLAG_HCTX_ACTIVE
, &q
->queue_flags
) ||
48 test_and_set_bit(QUEUE_FLAG_HCTX_ACTIVE
, &q
->queue_flags
)) {
52 if (test_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
) ||
53 test_and_set_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
)) {
58 users
= atomic_inc_return(&hctx
->tags
->active_queues
);
60 blk_mq_update_wake_batch(hctx
->tags
, users
);
66 * Wakeup all potentially sleeping on tags
68 void blk_mq_tag_wakeup_all(struct blk_mq_tags
*tags
, bool include_reserve
)
70 sbitmap_queue_wake_all(&tags
->bitmap_tags
);
72 sbitmap_queue_wake_all(&tags
->breserved_tags
);
76 * If a previously busy queue goes inactive, potential waiters could now
77 * be allowed to queue. Wake them up and check.
79 void __blk_mq_tag_idle(struct blk_mq_hw_ctx
*hctx
)
81 struct blk_mq_tags
*tags
= hctx
->tags
;
84 if (blk_mq_is_shared_tags(hctx
->flags
)) {
85 struct request_queue
*q
= hctx
->queue
;
87 if (!test_and_clear_bit(QUEUE_FLAG_HCTX_ACTIVE
,
91 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
95 users
= atomic_dec_return(&tags
->active_queues
);
97 blk_mq_update_wake_batch(tags
, users
);
99 blk_mq_tag_wakeup_all(tags
, false);
102 static int __blk_mq_get_tag(struct blk_mq_alloc_data
*data
,
103 struct sbitmap_queue
*bt
)
105 if (!data
->q
->elevator
&& !(data
->flags
& BLK_MQ_REQ_RESERVED
) &&
106 !hctx_may_queue(data
->hctx
, bt
))
107 return BLK_MQ_NO_TAG
;
109 if (data
->shallow_depth
)
110 return sbitmap_queue_get_shallow(bt
, data
->shallow_depth
);
112 return __sbitmap_queue_get(bt
);
115 unsigned long blk_mq_get_tags(struct blk_mq_alloc_data
*data
, int nr_tags
,
116 unsigned int *offset
)
118 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
119 struct sbitmap_queue
*bt
= &tags
->bitmap_tags
;
122 if (data
->shallow_depth
||data
->flags
& BLK_MQ_REQ_RESERVED
||
123 data
->hctx
->flags
& BLK_MQ_F_TAG_QUEUE_SHARED
)
125 ret
= __sbitmap_queue_get_batch(bt
, nr_tags
, offset
);
126 *offset
+= tags
->nr_reserved_tags
;
130 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data
*data
)
132 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
133 struct sbitmap_queue
*bt
;
134 struct sbq_wait_state
*ws
;
135 DEFINE_SBQ_WAIT(wait
);
136 unsigned int tag_offset
;
139 if (data
->flags
& BLK_MQ_REQ_RESERVED
) {
140 if (unlikely(!tags
->nr_reserved_tags
)) {
142 return BLK_MQ_NO_TAG
;
144 bt
= &tags
->breserved_tags
;
147 bt
= &tags
->bitmap_tags
;
148 tag_offset
= tags
->nr_reserved_tags
;
151 tag
= __blk_mq_get_tag(data
, bt
);
152 if (tag
!= BLK_MQ_NO_TAG
)
155 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
156 return BLK_MQ_NO_TAG
;
158 ws
= bt_wait_ptr(bt
, data
->hctx
);
160 struct sbitmap_queue
*bt_prev
;
163 * We're out of tags on this hardware queue, kick any
164 * pending IO submits before going to sleep waiting for
167 blk_mq_run_hw_queue(data
->hctx
, false);
170 * Retry tag allocation after running the hardware queue,
171 * as running the queue may also have found completions.
173 tag
= __blk_mq_get_tag(data
, bt
);
174 if (tag
!= BLK_MQ_NO_TAG
)
177 sbitmap_prepare_to_wait(bt
, ws
, &wait
, TASK_UNINTERRUPTIBLE
);
179 tag
= __blk_mq_get_tag(data
, bt
);
180 if (tag
!= BLK_MQ_NO_TAG
)
186 sbitmap_finish_wait(bt
, ws
, &wait
);
188 data
->ctx
= blk_mq_get_ctx(data
->q
);
189 data
->hctx
= blk_mq_map_queue(data
->q
, data
->cmd_flags
,
191 tags
= blk_mq_tags_from_data(data
);
192 if (data
->flags
& BLK_MQ_REQ_RESERVED
)
193 bt
= &tags
->breserved_tags
;
195 bt
= &tags
->bitmap_tags
;
198 * If destination hw queue is changed, fake wake up on
199 * previous queue for compensating the wake up miss, so
200 * other allocations on previous queue won't be starved.
203 sbitmap_queue_wake_up(bt_prev
);
205 ws
= bt_wait_ptr(bt
, data
->hctx
);
208 sbitmap_finish_wait(bt
, ws
, &wait
);
212 * Give up this allocation if the hctx is inactive. The caller will
213 * retry on an active hctx.
215 if (unlikely(test_bit(BLK_MQ_S_INACTIVE
, &data
->hctx
->state
))) {
216 blk_mq_put_tag(tags
, data
->ctx
, tag
+ tag_offset
);
217 return BLK_MQ_NO_TAG
;
219 return tag
+ tag_offset
;
222 void blk_mq_put_tag(struct blk_mq_tags
*tags
, struct blk_mq_ctx
*ctx
,
225 if (!blk_mq_tag_is_reserved(tags
, tag
)) {
226 const int real_tag
= tag
- tags
->nr_reserved_tags
;
228 BUG_ON(real_tag
>= tags
->nr_tags
);
229 sbitmap_queue_clear(&tags
->bitmap_tags
, real_tag
, ctx
->cpu
);
231 sbitmap_queue_clear(&tags
->breserved_tags
, tag
, ctx
->cpu
);
235 void blk_mq_put_tags(struct blk_mq_tags
*tags
, int *tag_array
, int nr_tags
)
237 sbitmap_queue_clear_batch(&tags
->bitmap_tags
, tags
->nr_reserved_tags
,
241 struct bt_iter_data
{
242 struct blk_mq_hw_ctx
*hctx
;
243 struct request_queue
*q
;
244 busy_tag_iter_fn
*fn
;
249 static struct request
*blk_mq_find_and_get_req(struct blk_mq_tags
*tags
,
255 spin_lock_irqsave(&tags
->lock
, flags
);
256 rq
= tags
->rqs
[bitnr
];
257 if (!rq
|| rq
->tag
!= bitnr
|| !req_ref_inc_not_zero(rq
))
259 spin_unlock_irqrestore(&tags
->lock
, flags
);
263 static bool bt_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
265 struct bt_iter_data
*iter_data
= data
;
266 struct blk_mq_hw_ctx
*hctx
= iter_data
->hctx
;
267 struct request_queue
*q
= iter_data
->q
;
268 struct blk_mq_tag_set
*set
= q
->tag_set
;
269 bool reserved
= iter_data
->reserved
;
270 struct blk_mq_tags
*tags
;
274 if (blk_mq_is_shared_tags(set
->flags
))
275 tags
= set
->shared_tags
;
280 bitnr
+= tags
->nr_reserved_tags
;
282 * We can hit rq == NULL here, because the tagging functions
283 * test and set the bit before assigning ->rqs[].
285 rq
= blk_mq_find_and_get_req(tags
, bitnr
);
289 if (rq
->q
== q
&& (!hctx
|| rq
->mq_hctx
== hctx
))
290 ret
= iter_data
->fn(rq
, iter_data
->data
, reserved
);
291 blk_mq_put_rq_ref(rq
);
296 * bt_for_each - iterate over the requests associated with a hardware queue
297 * @hctx: Hardware queue to examine.
298 * @q: Request queue to examine.
299 * @bt: sbitmap to examine. This is either the breserved_tags member
300 * or the bitmap_tags member of struct blk_mq_tags.
301 * @fn: Pointer to the function that will be called for each request
302 * associated with @hctx that has been assigned a driver tag.
303 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
304 * where rq is a pointer to a request. Return true to continue
305 * iterating tags, false to stop.
306 * @data: Will be passed as third argument to @fn.
307 * @reserved: Indicates whether @bt is the breserved_tags member or the
308 * bitmap_tags member of struct blk_mq_tags.
310 static void bt_for_each(struct blk_mq_hw_ctx
*hctx
, struct request_queue
*q
,
311 struct sbitmap_queue
*bt
, busy_tag_iter_fn
*fn
,
312 void *data
, bool reserved
)
314 struct bt_iter_data iter_data
= {
318 .reserved
= reserved
,
322 sbitmap_for_each_set(&bt
->sb
, bt_iter
, &iter_data
);
325 struct bt_tags_iter_data
{
326 struct blk_mq_tags
*tags
;
327 busy_tag_iter_fn
*fn
;
332 #define BT_TAG_ITER_RESERVED (1 << 0)
333 #define BT_TAG_ITER_STARTED (1 << 1)
334 #define BT_TAG_ITER_STATIC_RQS (1 << 2)
336 static bool bt_tags_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
338 struct bt_tags_iter_data
*iter_data
= data
;
339 struct blk_mq_tags
*tags
= iter_data
->tags
;
340 bool reserved
= iter_data
->flags
& BT_TAG_ITER_RESERVED
;
343 bool iter_static_rqs
= !!(iter_data
->flags
& BT_TAG_ITER_STATIC_RQS
);
346 bitnr
+= tags
->nr_reserved_tags
;
349 * We can hit rq == NULL here, because the tagging functions
350 * test and set the bit before assigning ->rqs[].
353 rq
= tags
->static_rqs
[bitnr
];
355 rq
= blk_mq_find_and_get_req(tags
, bitnr
);
359 if (!(iter_data
->flags
& BT_TAG_ITER_STARTED
) ||
360 blk_mq_request_started(rq
))
361 ret
= iter_data
->fn(rq
, iter_data
->data
, reserved
);
362 if (!iter_static_rqs
)
363 blk_mq_put_rq_ref(rq
);
368 * bt_tags_for_each - iterate over the requests in a tag map
369 * @tags: Tag map to iterate over.
370 * @bt: sbitmap to examine. This is either the breserved_tags member
371 * or the bitmap_tags member of struct blk_mq_tags.
372 * @fn: Pointer to the function that will be called for each started
373 * request. @fn will be called as follows: @fn(rq, @data,
374 * @reserved) where rq is a pointer to a request. Return true
375 * to continue iterating tags, false to stop.
376 * @data: Will be passed as second argument to @fn.
377 * @flags: BT_TAG_ITER_*
379 static void bt_tags_for_each(struct blk_mq_tags
*tags
, struct sbitmap_queue
*bt
,
380 busy_tag_iter_fn
*fn
, void *data
, unsigned int flags
)
382 struct bt_tags_iter_data iter_data
= {
390 sbitmap_for_each_set(&bt
->sb
, bt_tags_iter
, &iter_data
);
393 static void __blk_mq_all_tag_iter(struct blk_mq_tags
*tags
,
394 busy_tag_iter_fn
*fn
, void *priv
, unsigned int flags
)
396 WARN_ON_ONCE(flags
& BT_TAG_ITER_RESERVED
);
398 if (tags
->nr_reserved_tags
)
399 bt_tags_for_each(tags
, &tags
->breserved_tags
, fn
, priv
,
400 flags
| BT_TAG_ITER_RESERVED
);
401 bt_tags_for_each(tags
, &tags
->bitmap_tags
, fn
, priv
, flags
);
405 * blk_mq_all_tag_iter - iterate over all requests in a tag map
406 * @tags: Tag map to iterate over.
407 * @fn: Pointer to the function that will be called for each
408 * request. @fn will be called as follows: @fn(rq, @priv,
409 * reserved) where rq is a pointer to a request. 'reserved'
410 * indicates whether or not @rq is a reserved request. Return
411 * true to continue iterating tags, false to stop.
412 * @priv: Will be passed as second argument to @fn.
414 * Caller has to pass the tag map from which requests are allocated.
416 void blk_mq_all_tag_iter(struct blk_mq_tags
*tags
, busy_tag_iter_fn
*fn
,
419 __blk_mq_all_tag_iter(tags
, fn
, priv
, BT_TAG_ITER_STATIC_RQS
);
423 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
424 * @tagset: Tag set to iterate over.
425 * @fn: Pointer to the function that will be called for each started
426 * request. @fn will be called as follows: @fn(rq, @priv,
427 * reserved) where rq is a pointer to a request. 'reserved'
428 * indicates whether or not @rq is a reserved request. Return
429 * true to continue iterating tags, false to stop.
430 * @priv: Will be passed as second argument to @fn.
432 * We grab one request reference before calling @fn and release it after
435 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set
*tagset
,
436 busy_tag_iter_fn
*fn
, void *priv
)
438 unsigned int flags
= tagset
->flags
;
441 nr_tags
= blk_mq_is_shared_tags(flags
) ? 1 : tagset
->nr_hw_queues
;
443 for (i
= 0; i
< nr_tags
; i
++) {
444 if (tagset
->tags
&& tagset
->tags
[i
])
445 __blk_mq_all_tag_iter(tagset
->tags
[i
], fn
, priv
,
446 BT_TAG_ITER_STARTED
);
449 EXPORT_SYMBOL(blk_mq_tagset_busy_iter
);
451 static bool blk_mq_tagset_count_completed_rqs(struct request
*rq
,
452 void *data
, bool reserved
)
454 unsigned *count
= data
;
456 if (blk_mq_request_completed(rq
))
462 * blk_mq_tagset_wait_completed_request - Wait until all scheduled request
463 * completions have finished.
464 * @tagset: Tag set to drain completed request
466 * Note: This function has to be run after all IO queues are shutdown
468 void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set
*tagset
)
473 blk_mq_tagset_busy_iter(tagset
,
474 blk_mq_tagset_count_completed_rqs
, &count
);
480 EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request
);
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.
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.
496 void blk_mq_queue_tag_busy_iter(struct request_queue
*q
, busy_tag_iter_fn
*fn
,
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
504 if (!percpu_ref_tryget(&q
->q_usage_counter
))
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
;
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);
516 struct blk_mq_hw_ctx
*hctx
;
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
;
525 * If no software queues are currently mapped to this
526 * hardware queue, there's nothing to check
528 if (!blk_mq_hw_queue_mapped(hctx
))
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);
539 static int bt_alloc(struct sbitmap_queue
*bt
, unsigned int depth
,
540 bool round_robin
, int node
)
542 return sbitmap_queue_init_node(bt
, depth
, -1, round_robin
, GFP_KERNEL
,
546 int 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
)
551 unsigned int depth
= queue_depth
- reserved
;
552 bool round_robin
= alloc_policy
== BLK_TAG_ALLOC_RR
;
554 if (bt_alloc(bitmap_tags
, depth
, round_robin
, node
))
556 if (bt_alloc(breserved_tags
, reserved
, round_robin
, node
))
557 goto free_bitmap_tags
;
562 sbitmap_queue_free(bitmap_tags
);
566 struct blk_mq_tags
*blk_mq_init_tags(unsigned int total_tags
,
567 unsigned int reserved_tags
,
568 int node
, int alloc_policy
)
570 struct blk_mq_tags
*tags
;
572 if (total_tags
> BLK_MQ_TAG_MAX
) {
573 pr_err("blk-mq: tag depth too large\n");
577 tags
= kzalloc_node(sizeof(*tags
), GFP_KERNEL
, node
);
581 tags
->nr_tags
= total_tags
;
582 tags
->nr_reserved_tags
= reserved_tags
;
583 spin_lock_init(&tags
->lock
);
585 if (blk_mq_init_bitmaps(&tags
->bitmap_tags
, &tags
->breserved_tags
,
586 total_tags
, reserved_tags
, node
,
594 void blk_mq_free_tags(struct blk_mq_tags
*tags
)
596 sbitmap_queue_free(&tags
->bitmap_tags
);
597 sbitmap_queue_free(&tags
->breserved_tags
);
601 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx
*hctx
,
602 struct blk_mq_tags
**tagsptr
, unsigned int tdepth
,
605 struct blk_mq_tags
*tags
= *tagsptr
;
607 if (tdepth
<= tags
->nr_reserved_tags
)
611 * If we are allowed to grow beyond the original size, allocate
612 * a new set of tags before freeing the old one.
614 if (tdepth
> tags
->nr_tags
) {
615 struct blk_mq_tag_set
*set
= hctx
->queue
->tag_set
;
616 struct blk_mq_tags
*new;
622 * We need some sort of upper limit, set it high enough that
623 * no valid use cases should require more.
625 if (tdepth
> MAX_SCHED_RQ
)
629 * Only the sbitmap needs resizing since we allocated the max
632 if (blk_mq_is_shared_tags(set
->flags
))
635 new = blk_mq_alloc_map_and_rqs(set
, hctx
->queue_num
, tdepth
);
639 blk_mq_free_map_and_rqs(set
, *tagsptr
, hctx
->queue_num
);
643 * Don't need (or can't) update reserved tags here, they
644 * remain static and should never need resizing.
646 sbitmap_queue_resize(&tags
->bitmap_tags
,
647 tdepth
- tags
->nr_reserved_tags
);
653 void blk_mq_tag_resize_shared_tags(struct blk_mq_tag_set
*set
, unsigned int size
)
655 struct blk_mq_tags
*tags
= set
->shared_tags
;
657 sbitmap_queue_resize(&tags
->bitmap_tags
, size
- set
->reserved_tags
);
660 void blk_mq_tag_update_sched_shared_tags(struct request_queue
*q
)
662 sbitmap_queue_resize(&q
->sched_shared_tags
->bitmap_tags
,
663 q
->nr_requests
- q
->tag_set
->reserved_tags
);
667 * blk_mq_unique_tag() - return a tag that is unique queue-wide
668 * @rq: request for which to compute a unique tag
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
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.
678 u32
blk_mq_unique_tag(struct request
*rq
)
680 return (rq
->mq_hctx
->queue_num
<< BLK_MQ_UNIQUE_TAG_BITS
) |
681 (rq
->tag
& BLK_MQ_UNIQUE_TAG_MASK
);
683 EXPORT_SYMBOL(blk_mq_unique_tag
);