2 * Tag allocation using scalable bitmaps. Uses active queue tracking to support
3 * fairer distribution of tags between multiple submitters when a shared tag map
6 * Copyright (C) 2013-2014 Jens Axboe
8 #include <linux/kernel.h>
9 #include <linux/module.h>
11 #include <linux/blk-mq.h>
14 #include "blk-mq-tag.h"
16 bool blk_mq_has_free_tags(struct blk_mq_tags
*tags
)
21 return sbitmap_any_bit_clear(&tags
->bitmap_tags
.sb
);
25 * If a previously inactive queue goes active, bump the active user count.
26 * We need to do this before try to allocate driver tag, then even if fail
27 * to get tag when first time, the other shared-tag users could reserve
30 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx
*hctx
)
32 if (!test_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
) &&
33 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
34 atomic_inc(&hctx
->tags
->active_queues
);
40 * Wakeup all potentially sleeping on tags
42 void blk_mq_tag_wakeup_all(struct blk_mq_tags
*tags
, bool include_reserve
)
44 sbitmap_queue_wake_all(&tags
->bitmap_tags
);
46 sbitmap_queue_wake_all(&tags
->breserved_tags
);
50 * If a previously busy queue goes inactive, potential waiters could now
51 * be allowed to queue. Wake them up and check.
53 void __blk_mq_tag_idle(struct blk_mq_hw_ctx
*hctx
)
55 struct blk_mq_tags
*tags
= hctx
->tags
;
57 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
60 atomic_dec(&tags
->active_queues
);
62 blk_mq_tag_wakeup_all(tags
, false);
66 * For shared tag users, we track the number of currently active users
67 * and attempt to provide a fair share of the tag depth for each of them.
69 static inline bool hctx_may_queue(struct blk_mq_hw_ctx
*hctx
,
70 struct sbitmap_queue
*bt
)
72 unsigned int depth
, users
;
74 if (!hctx
|| !(hctx
->flags
& BLK_MQ_F_TAG_SHARED
))
76 if (!test_bit(BLK_MQ_S_TAG_ACTIVE
, &hctx
->state
))
80 * Don't try dividing an ant
82 if (bt
->sb
.depth
== 1)
85 users
= atomic_read(&hctx
->tags
->active_queues
);
90 * Allow at least some tags
92 depth
= max((bt
->sb
.depth
+ users
- 1) / users
, 4U);
93 return atomic_read(&hctx
->nr_active
) < depth
;
96 static int __blk_mq_get_tag(struct blk_mq_alloc_data
*data
,
97 struct sbitmap_queue
*bt
)
99 if (!(data
->flags
& BLK_MQ_REQ_INTERNAL
) &&
100 !hctx_may_queue(data
->hctx
, bt
))
102 if (data
->shallow_depth
)
103 return __sbitmap_queue_get_shallow(bt
, data
->shallow_depth
);
105 return __sbitmap_queue_get(bt
);
108 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data
*data
)
110 struct blk_mq_tags
*tags
= blk_mq_tags_from_data(data
);
111 struct sbitmap_queue
*bt
;
112 struct sbq_wait_state
*ws
;
114 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
);
138 drop_ctx
= data
->ctx
== NULL
;
140 struct sbitmap_queue
*bt_prev
;
143 * We're out of tags on this hardware queue, kick any
144 * pending IO submits before going to sleep waiting for
147 blk_mq_run_hw_queue(data
->hctx
, false);
150 * Retry tag allocation after running the hardware queue,
151 * as running the queue may also have found completions.
153 tag
= __blk_mq_get_tag(data
, bt
);
157 prepare_to_wait_exclusive(&ws
->wait
, &wait
,
158 TASK_UNINTERRUPTIBLE
);
160 tag
= __blk_mq_get_tag(data
, bt
);
165 blk_mq_put_ctx(data
->ctx
);
170 data
->ctx
= blk_mq_get_ctx(data
->q
);
171 data
->hctx
= blk_mq_map_queue(data
->q
, data
->ctx
->cpu
);
172 tags
= blk_mq_tags_from_data(data
);
173 if (data
->flags
& BLK_MQ_REQ_RESERVED
)
174 bt
= &tags
->breserved_tags
;
176 bt
= &tags
->bitmap_tags
;
178 finish_wait(&ws
->wait
, &wait
);
181 * If destination hw queue is changed, fake wake up on
182 * previous queue for compensating the wake up miss, so
183 * other allocations on previous queue won't be starved.
186 sbitmap_queue_wake_up(bt_prev
);
188 ws
= bt_wait_ptr(bt
, data
->hctx
);
191 if (drop_ctx
&& data
->ctx
)
192 blk_mq_put_ctx(data
->ctx
);
194 finish_wait(&ws
->wait
, &wait
);
197 return tag
+ tag_offset
;
200 void blk_mq_put_tag(struct blk_mq_hw_ctx
*hctx
, struct blk_mq_tags
*tags
,
201 struct blk_mq_ctx
*ctx
, unsigned int tag
)
203 if (!blk_mq_tag_is_reserved(tags
, tag
)) {
204 const int real_tag
= tag
- tags
->nr_reserved_tags
;
206 BUG_ON(real_tag
>= tags
->nr_tags
);
207 sbitmap_queue_clear(&tags
->bitmap_tags
, real_tag
, ctx
->cpu
);
209 BUG_ON(tag
>= tags
->nr_reserved_tags
);
210 sbitmap_queue_clear(&tags
->breserved_tags
, tag
, ctx
->cpu
);
214 struct bt_iter_data
{
215 struct blk_mq_hw_ctx
*hctx
;
221 static bool bt_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
223 struct bt_iter_data
*iter_data
= data
;
224 struct blk_mq_hw_ctx
*hctx
= iter_data
->hctx
;
225 struct blk_mq_tags
*tags
= hctx
->tags
;
226 bool reserved
= iter_data
->reserved
;
230 bitnr
+= tags
->nr_reserved_tags
;
231 rq
= tags
->rqs
[bitnr
];
234 * We can hit rq == NULL here, because the tagging functions
235 * test and set the bit before assigning ->rqs[].
237 if (rq
&& rq
->q
== hctx
->queue
)
238 iter_data
->fn(hctx
, rq
, iter_data
->data
, reserved
);
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.
251 * @data: Will be passed as third argument to @fn.
252 * @reserved: Indicates whether @bt is the breserved_tags member or the
253 * bitmap_tags member of struct blk_mq_tags.
255 static void bt_for_each(struct blk_mq_hw_ctx
*hctx
, struct sbitmap_queue
*bt
,
256 busy_iter_fn
*fn
, void *data
, bool reserved
)
258 struct bt_iter_data iter_data
= {
262 .reserved
= reserved
,
265 sbitmap_for_each_set(&bt
->sb
, bt_iter
, &iter_data
);
268 struct bt_tags_iter_data
{
269 struct blk_mq_tags
*tags
;
270 busy_tag_iter_fn
*fn
;
275 static bool bt_tags_iter(struct sbitmap
*bitmap
, unsigned int bitnr
, void *data
)
277 struct bt_tags_iter_data
*iter_data
= data
;
278 struct blk_mq_tags
*tags
= iter_data
->tags
;
279 bool reserved
= iter_data
->reserved
;
283 bitnr
+= tags
->nr_reserved_tags
;
286 * We can hit rq == NULL here, because the tagging functions
287 * test and set the bit before assining ->rqs[].
289 rq
= tags
->rqs
[bitnr
];
290 if (rq
&& blk_mq_request_started(rq
))
291 iter_data
->fn(rq
, iter_data
->data
, reserved
);
297 * bt_tags_for_each - iterate over the requests in a tag map
298 * @tags: Tag map to iterate over.
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 started
302 * request. @fn will be called as follows: @fn(rq, @data,
303 * @reserved) where rq is a pointer to a request.
304 * @data: Will be passed as second argument to @fn.
305 * @reserved: Indicates whether @bt is the breserved_tags member or the
306 * bitmap_tags member of struct blk_mq_tags.
308 static void bt_tags_for_each(struct blk_mq_tags
*tags
, struct sbitmap_queue
*bt
,
309 busy_tag_iter_fn
*fn
, void *data
, bool reserved
)
311 struct bt_tags_iter_data iter_data
= {
315 .reserved
= reserved
,
319 sbitmap_for_each_set(&bt
->sb
, bt_tags_iter
, &iter_data
);
323 * blk_mq_all_tag_busy_iter - iterate over all started requests in a tag map
324 * @tags: Tag map to iterate over.
325 * @fn: Pointer to the function that will be called for each started
326 * request. @fn will be called as follows: @fn(rq, @priv,
327 * reserved) where rq is a pointer to a request. 'reserved'
328 * indicates whether or not @rq is a reserved request.
329 * @priv: Will be passed as second argument to @fn.
331 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags
*tags
,
332 busy_tag_iter_fn
*fn
, void *priv
)
334 if (tags
->nr_reserved_tags
)
335 bt_tags_for_each(tags
, &tags
->breserved_tags
, fn
, priv
, true);
336 bt_tags_for_each(tags
, &tags
->bitmap_tags
, fn
, priv
, false);
340 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
341 * @tagset: Tag set to iterate over.
342 * @fn: Pointer to the function that will be called for each started
343 * request. @fn will be called as follows: @fn(rq, @priv,
344 * reserved) where rq is a pointer to a request. 'reserved'
345 * indicates whether or not @rq is a reserved request.
346 * @priv: Will be passed as second argument to @fn.
348 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set
*tagset
,
349 busy_tag_iter_fn
*fn
, void *priv
)
353 for (i
= 0; i
< tagset
->nr_hw_queues
; i
++) {
354 if (tagset
->tags
&& tagset
->tags
[i
])
355 blk_mq_all_tag_busy_iter(tagset
->tags
[i
], fn
, priv
);
358 EXPORT_SYMBOL(blk_mq_tagset_busy_iter
);
361 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
362 * @q: Request queue to examine.
363 * @fn: Pointer to the function that will be called for each request
364 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
365 * reserved) where rq is a pointer to a request and hctx points
366 * to the hardware queue associated with the request. 'reserved'
367 * indicates whether or not @rq is a reserved request.
368 * @priv: Will be passed as third argument to @fn.
370 * Note: if @q->tag_set is shared with other request queues then @fn will be
371 * called for all requests on all queues that share that tag set and not only
372 * for requests associated with @q.
374 void blk_mq_queue_tag_busy_iter(struct request_queue
*q
, busy_iter_fn
*fn
,
377 struct blk_mq_hw_ctx
*hctx
;
381 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and queue_hw_ctx
382 * while the queue is frozen. So we can use q_usage_counter to avoid
383 * racing with it. __blk_mq_update_nr_hw_queues() uses
384 * synchronize_rcu() to ensure this function left the critical section
387 if (!percpu_ref_tryget(&q
->q_usage_counter
))
390 queue_for_each_hw_ctx(q
, hctx
, i
) {
391 struct blk_mq_tags
*tags
= hctx
->tags
;
394 * If no software queues are currently mapped to this
395 * hardware queue, there's nothing to check
397 if (!blk_mq_hw_queue_mapped(hctx
))
400 if (tags
->nr_reserved_tags
)
401 bt_for_each(hctx
, &tags
->breserved_tags
, fn
, priv
, true);
402 bt_for_each(hctx
, &tags
->bitmap_tags
, fn
, priv
, false);
407 static int bt_alloc(struct sbitmap_queue
*bt
, unsigned int depth
,
408 bool round_robin
, int node
)
410 return sbitmap_queue_init_node(bt
, depth
, -1, round_robin
, GFP_KERNEL
,
414 static struct blk_mq_tags
*blk_mq_init_bitmap_tags(struct blk_mq_tags
*tags
,
415 int node
, int alloc_policy
)
417 unsigned int depth
= tags
->nr_tags
- tags
->nr_reserved_tags
;
418 bool round_robin
= alloc_policy
== BLK_TAG_ALLOC_RR
;
420 if (bt_alloc(&tags
->bitmap_tags
, depth
, round_robin
, node
))
422 if (bt_alloc(&tags
->breserved_tags
, tags
->nr_reserved_tags
, round_robin
,
424 goto free_bitmap_tags
;
428 sbitmap_queue_free(&tags
->bitmap_tags
);
434 struct blk_mq_tags
*blk_mq_init_tags(unsigned int total_tags
,
435 unsigned int reserved_tags
,
436 int node
, int alloc_policy
)
438 struct blk_mq_tags
*tags
;
440 if (total_tags
> BLK_MQ_TAG_MAX
) {
441 pr_err("blk-mq: tag depth too large\n");
445 tags
= kzalloc_node(sizeof(*tags
), GFP_KERNEL
, node
);
449 tags
->nr_tags
= total_tags
;
450 tags
->nr_reserved_tags
= reserved_tags
;
452 return blk_mq_init_bitmap_tags(tags
, node
, alloc_policy
);
455 void blk_mq_free_tags(struct blk_mq_tags
*tags
)
457 sbitmap_queue_free(&tags
->bitmap_tags
);
458 sbitmap_queue_free(&tags
->breserved_tags
);
462 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx
*hctx
,
463 struct blk_mq_tags
**tagsptr
, unsigned int tdepth
,
466 struct blk_mq_tags
*tags
= *tagsptr
;
468 if (tdepth
<= tags
->nr_reserved_tags
)
472 * If we are allowed to grow beyond the original size, allocate
473 * a new set of tags before freeing the old one.
475 if (tdepth
> tags
->nr_tags
) {
476 struct blk_mq_tag_set
*set
= hctx
->queue
->tag_set
;
477 struct blk_mq_tags
*new;
484 * We need some sort of upper limit, set it high enough that
485 * no valid use cases should require more.
487 if (tdepth
> 16 * BLKDEV_MAX_RQ
)
490 new = blk_mq_alloc_rq_map(set
, hctx
->queue_num
, tdepth
,
491 tags
->nr_reserved_tags
);
494 ret
= blk_mq_alloc_rqs(set
, new, hctx
->queue_num
, tdepth
);
496 blk_mq_free_rq_map(new);
500 blk_mq_free_rqs(set
, *tagsptr
, hctx
->queue_num
);
501 blk_mq_free_rq_map(*tagsptr
);
505 * Don't need (or can't) update reserved tags here, they
506 * remain static and should never need resizing.
508 sbitmap_queue_resize(&tags
->bitmap_tags
,
509 tdepth
- tags
->nr_reserved_tags
);
516 * blk_mq_unique_tag() - return a tag that is unique queue-wide
517 * @rq: request for which to compute a unique tag
519 * The tag field in struct request is unique per hardware queue but not over
520 * all hardware queues. Hence this function that returns a tag with the
521 * hardware context index in the upper bits and the per hardware queue tag in
524 * Note: When called for a request that is queued on a non-multiqueue request
525 * queue, the hardware context index is set to zero.
527 u32
blk_mq_unique_tag(struct request
*rq
)
529 struct request_queue
*q
= rq
->q
;
530 struct blk_mq_hw_ctx
*hctx
;
534 hctx
= blk_mq_map_queue(q
, rq
->mq_ctx
->cpu
);
535 hwq
= hctx
->queue_num
;
538 return (hwq
<< BLK_MQ_UNIQUE_TAG_BITS
) |
539 (rq
->tag
& BLK_MQ_UNIQUE_TAG_MASK
);
541 EXPORT_SYMBOL(blk_mq_unique_tag
);