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
;
118 if (data
->flags
& BLK_MQ_REQ_RESERVED
) {
119 if (unlikely(!tags
->nr_reserved_tags
)) {
121 return BLK_MQ_TAG_FAIL
;
123 bt
= &tags
->breserved_tags
;
126 bt
= &tags
->bitmap_tags
;
127 tag_offset
= tags
->nr_reserved_tags
;
130 tag
= __blk_mq_get_tag(data
, bt
);
134 if (data
->flags
& BLK_MQ_REQ_NOWAIT
)
135 return BLK_MQ_TAG_FAIL
;
137 ws
= bt_wait_ptr(bt
, data
->hctx
);
139 struct sbitmap_queue
*bt_prev
;
142 * We're out of tags on this hardware queue, kick any
143 * pending IO submits before going to sleep waiting for
146 blk_mq_run_hw_queue(data
->hctx
, false);
149 * Retry tag allocation after running the hardware queue,
150 * as running the queue may also have found completions.
152 tag
= __blk_mq_get_tag(data
, bt
);
156 sbitmap_prepare_to_wait(bt
, ws
, &wait
, TASK_UNINTERRUPTIBLE
);
158 tag
= __blk_mq_get_tag(data
, bt
);
165 sbitmap_finish_wait(bt
, ws
, &wait
);
167 data
->ctx
= blk_mq_get_ctx(data
->q
);
168 data
->hctx
= blk_mq_map_queue(data
->q
, data
->cmd_flags
,
170 tags
= blk_mq_tags_from_data(data
);
171 if (data
->flags
& BLK_MQ_REQ_RESERVED
)
172 bt
= &tags
->breserved_tags
;
174 bt
= &tags
->bitmap_tags
;
177 * If destination hw queue is changed, fake wake up on
178 * previous queue for compensating the wake up miss, so
179 * other allocations on previous queue won't be starved.
182 sbitmap_queue_wake_up(bt_prev
);
184 ws
= bt_wait_ptr(bt
, data
->hctx
);
187 sbitmap_finish_wait(bt
, ws
, &wait
);
190 return tag
+ tag_offset
;
193 void blk_mq_put_tag(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_tags
*tags
,
194 struct blk_mq_ctx
*ctx
, unsigned int tag
)
196 if (!blk_mq_tag_is_reserved(tags
, tag
)) {
197 const int real_tag
= tag
- tags
->nr_reserved_tags
;
199 BUG_ON(real_tag
>= tags
->nr_tags
);
200 sbitmap_queue_clear(&tags
->bitmap_tags
, real_tag
, ctx
->cpu
);
202 BUG_ON(tag
>= tags
->nr_reserved_tags
);
203 sbitmap_queue_clear(&tags
->breserved_tags
, tag
, ctx
->cpu
);
207 struct bt_iter_data
{
208 struct blk_mq_hw_ctx
*hctx
;
214 static bool bt_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
216 struct bt_iter_data
*iter_data
= data
;
217 struct blk_mq_hw_ctx
*hctx
= iter_data
->hctx
;
218 struct blk_mq_tags
*tags
= hctx
->tags
;
219 bool reserved
= iter_data
->reserved
;
223 bitnr
+= tags
->nr_reserved_tags
;
224 rq
= tags
->rqs
[bitnr
];
227 * We can hit rq == NULL here, because the tagging functions
228 * test and set the bit before assigning ->rqs[].
230 if (rq
&& rq
->q
== hctx
->queue
)
231 return iter_data
->fn(hctx
, rq
, iter_data
->data
, reserved
);
236 * bt_for_each - iterate over the requests associated with a hardware queue
237 * @hctx: Hardware queue to examine.
238 * @bt: sbitmap to examine. This is either the breserved_tags member
239 * or the bitmap_tags member of struct blk_mq_tags.
240 * @fn: Pointer to the function that will be called for each request
241 * associated with @hctx that has been assigned a driver tag.
242 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
243 * where rq is a pointer to a request. Return true to continue
244 * iterating tags, false to stop.
245 * @data: Will be passed as third argument to @fn.
246 * @reserved: Indicates whether @bt is the breserved_tags member or the
247 * bitmap_tags member of struct blk_mq_tags.
249 static void bt_for_each(struct blk_mq_hw_ctx
*hctx
, struct sbitmap_queue
*bt
,
250 busy_iter_fn
*fn
, void *data
, bool reserved
)
252 struct bt_iter_data iter_data
= {
256 .reserved
= reserved
,
259 sbitmap_for_each_set(&bt
->sb
, bt_iter
, &iter_data
);
262 struct bt_tags_iter_data
{
263 struct blk_mq_tags
*tags
;
264 busy_tag_iter_fn
*fn
;
269 static bool bt_tags_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
271 struct bt_tags_iter_data
*iter_data
= data
;
272 struct blk_mq_tags
*tags
= iter_data
->tags
;
273 bool reserved
= iter_data
->reserved
;
277 bitnr
+= tags
->nr_reserved_tags
;
280 * We can hit rq == NULL here, because the tagging functions
281 * test and set the bit before assining ->rqs[].
283 rq
= tags
->rqs
[bitnr
];
284 if (rq
&& blk_mq_request_started(rq
))
285 return iter_data
->fn(rq
, iter_data
->data
, reserved
);
291 * bt_tags_for_each - iterate over the requests in a tag map
292 * @tags: Tag map to iterate over.
293 * @bt: sbitmap to examine. This is either the breserved_tags member
294 * or the bitmap_tags member of struct blk_mq_tags.
295 * @fn: Pointer to the function that will be called for each started
296 * request. @fn will be called as follows: @fn(rq, @data,
297 * @reserved) where rq is a pointer to a request. Return true
298 * to continue iterating tags, false to stop.
299 * @data: Will be passed as second argument to @fn.
300 * @reserved: Indicates whether @bt is the breserved_tags member or the
301 * bitmap_tags member of struct blk_mq_tags.
303 static void bt_tags_for_each(struct blk_mq_tags
*tags
, struct sbitmap_queue
*bt
,
304 busy_tag_iter_fn
*fn
, void *data
, bool reserved
)
306 struct bt_tags_iter_data iter_data
= {
310 .reserved
= reserved
,
314 sbitmap_for_each_set(&bt
->sb
, bt_tags_iter
, &iter_data
);
318 * blk_mq_all_tag_busy_iter - iterate over all started requests in a tag map
319 * @tags: Tag map to iterate over.
320 * @fn: Pointer to the function that will be called for each started
321 * request. @fn will be called as follows: @fn(rq, @priv,
322 * reserved) where rq is a pointer to a request. 'reserved'
323 * indicates whether or not @rq is a reserved request. Return
324 * true to continue iterating tags, false to stop.
325 * @priv: Will be passed as second argument to @fn.
327 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags
*tags
,
328 busy_tag_iter_fn
*fn
, void *priv
)
330 if (tags
->nr_reserved_tags
)
331 bt_tags_for_each(tags
, &tags
->breserved_tags
, fn
, priv
, true);
332 bt_tags_for_each(tags
, &tags
->bitmap_tags
, fn
, priv
, false);
336 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
337 * @tagset: Tag set to iterate over.
338 * @fn: Pointer to the function that will be called for each started
339 * request. @fn will be called as follows: @fn(rq, @priv,
340 * reserved) where rq is a pointer to a request. 'reserved'
341 * indicates whether or not @rq is a reserved request. Return
342 * true to continue iterating tags, false to stop.
343 * @priv: Will be passed as second argument to @fn.
345 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set
*tagset
,
346 busy_tag_iter_fn
*fn
, void *priv
)
350 for (i
= 0; i
< tagset
->nr_hw_queues
; i
++) {
351 if (tagset
->tags
&& tagset
->tags
[i
])
352 blk_mq_all_tag_busy_iter(tagset
->tags
[i
], fn
, priv
);
355 EXPORT_SYMBOL(blk_mq_tagset_busy_iter
);
358 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
359 * @q: Request queue to examine.
360 * @fn: Pointer to the function that will be called for each request
361 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
362 * reserved) where rq is a pointer to a request and hctx points
363 * to the hardware queue associated with the request. 'reserved'
364 * indicates whether or not @rq is a reserved request.
365 * @priv: Will be passed as third argument to @fn.
367 * Note: if @q->tag_set is shared with other request queues then @fn will be
368 * called for all requests on all queues that share that tag set and not only
369 * for requests associated with @q.
371 void blk_mq_queue_tag_busy_iter(struct request_queue
*q
, busy_iter_fn
*fn
,
374 struct blk_mq_hw_ctx
*hctx
;
378 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and queue_hw_ctx
379 * while the queue is frozen. So we can use q_usage_counter to avoid
380 * racing with it. __blk_mq_update_nr_hw_queues() uses
381 * synchronize_rcu() to ensure this function left the critical section
384 if (!percpu_ref_tryget(&q
->q_usage_counter
))
387 queue_for_each_hw_ctx(q
, hctx
, i
) {
388 struct blk_mq_tags
*tags
= hctx
->tags
;
391 * If no software queues are currently mapped to this
392 * hardware queue, there's nothing to check
394 if (!blk_mq_hw_queue_mapped(hctx
))
397 if (tags
->nr_reserved_tags
)
398 bt_for_each(hctx
, &tags
->breserved_tags
, fn
, priv
, true);
399 bt_for_each(hctx
, &tags
->bitmap_tags
, fn
, priv
, false);
404 static int bt_alloc(struct sbitmap_queue
*bt
, unsigned int depth
,
405 bool round_robin
, int node
)
407 return sbitmap_queue_init_node(bt
, depth
, -1, round_robin
, GFP_KERNEL
,
411 static struct blk_mq_tags
*blk_mq_init_bitmap_tags(struct blk_mq_tags
*tags
,
412 int node
, int alloc_policy
)
414 unsigned int depth
= tags
->nr_tags
- tags
->nr_reserved_tags
;
415 bool round_robin
= alloc_policy
== BLK_TAG_ALLOC_RR
;
417 if (bt_alloc(&tags
->bitmap_tags
, depth
, round_robin
, node
))
419 if (bt_alloc(&tags
->breserved_tags
, tags
->nr_reserved_tags
, round_robin
,
421 goto free_bitmap_tags
;
425 sbitmap_queue_free(&tags
->bitmap_tags
);
431 struct blk_mq_tags
*blk_mq_init_tags(unsigned int total_tags
,
432 unsigned int reserved_tags
,
433 int node
, int alloc_policy
)
435 struct blk_mq_tags
*tags
;
437 if (total_tags
> BLK_MQ_TAG_MAX
) {
438 pr_err("blk-mq: tag depth too large\n");
442 tags
= kzalloc_node(sizeof(*tags
), GFP_KERNEL
, node
);
446 tags
->nr_tags
= total_tags
;
447 tags
->nr_reserved_tags
= reserved_tags
;
449 return blk_mq_init_bitmap_tags(tags
, node
, alloc_policy
);
452 void blk_mq_free_tags(struct blk_mq_tags
*tags
)
454 sbitmap_queue_free(&tags
->bitmap_tags
);
455 sbitmap_queue_free(&tags
->breserved_tags
);
459 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx
*hctx
,
460 struct blk_mq_tags
**tagsptr
, unsigned int tdepth
,
463 struct blk_mq_tags
*tags
= *tagsptr
;
465 if (tdepth
<= tags
->nr_reserved_tags
)
469 * If we are allowed to grow beyond the original size, allocate
470 * a new set of tags before freeing the old one.
472 if (tdepth
> tags
->nr_tags
) {
473 struct blk_mq_tag_set
*set
= hctx
->queue
->tag_set
;
474 struct blk_mq_tags
*new;
481 * We need some sort of upper limit, set it high enough that
482 * no valid use cases should require more.
484 if (tdepth
> 16 * BLKDEV_MAX_RQ
)
487 new = blk_mq_alloc_rq_map(set
, hctx
->queue_num
, tdepth
,
488 tags
->nr_reserved_tags
);
491 ret
= blk_mq_alloc_rqs(set
, new, hctx
->queue_num
, tdepth
);
493 blk_mq_free_rq_map(new);
497 blk_mq_free_rqs(set
, *tagsptr
, hctx
->queue_num
);
498 blk_mq_free_rq_map(*tagsptr
);
502 * Don't need (or can't) update reserved tags here, they
503 * remain static and should never need resizing.
505 sbitmap_queue_resize(&tags
->bitmap_tags
,
506 tdepth
- tags
->nr_reserved_tags
);
513 * blk_mq_unique_tag() - return a tag that is unique queue-wide
514 * @rq: request for which to compute a unique tag
516 * The tag field in struct request is unique per hardware queue but not over
517 * all hardware queues. Hence this function that returns a tag with the
518 * hardware context index in the upper bits and the per hardware queue tag in
521 * Note: When called for a request that is queued on a non-multiqueue request
522 * queue, the hardware context index is set to zero.
524 u32
blk_mq_unique_tag(struct request
*rq
)
526 return (rq
->mq_hctx
->queue_num
<< BLK_MQ_UNIQUE_TAG_BITS
) |
527 (rq
->tag
& BLK_MQ_UNIQUE_TAG_MASK
);
529 EXPORT_SYMBOL(blk_mq_unique_tag
);