1 /* SPDX-License-Identifier: GPL-2.0 */
6 #include "blk-mq-tag.h"
12 struct blk_mq_ctx __percpu *queue_ctx;
16 * struct blk_mq_ctx - State for a software queue facing the submitting CPUs
21 struct list_head rq_lists[HCTX_MAX_TYPES];
22 } ____cacheline_aligned_in_smp;
25 unsigned short index_hw[HCTX_MAX_TYPES];
26 struct blk_mq_hw_ctx *hctxs[HCTX_MAX_TYPES];
28 /* incremented at dispatch time */
29 unsigned long rq_dispatched[2];
30 unsigned long rq_merged;
32 /* incremented at completion time */
33 unsigned long ____cacheline_aligned_in_smp rq_completed[2];
35 struct request_queue *queue;
36 struct blk_mq_ctxs *ctxs;
38 } ____cacheline_aligned_in_smp;
40 void blk_mq_exit_queue(struct request_queue *q);
41 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr);
42 void blk_mq_wake_waiters(struct request_queue *q);
43 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *,
45 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
46 bool kick_requeue_list);
47 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list);
48 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
49 struct blk_mq_ctx *start);
52 * Internal helpers for allocating/freeing the request map
54 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
55 unsigned int hctx_idx);
56 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags);
57 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
58 unsigned int hctx_idx,
60 unsigned int reserved_tags,
62 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
63 unsigned int hctx_idx, unsigned int depth);
66 * Internal helpers for request insertion into sw queues
68 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
70 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
72 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
73 struct list_head *list);
75 /* Used by blk_insert_cloned_request() to issue request directly */
76 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last);
77 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
78 struct list_head *list);
81 * CPU -> queue mappings
83 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
86 * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
88 * @type: the hctx type index
91 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
95 return q->queue_hw_ctx[q->tag_set->map[type].mq_map[cpu]];
99 * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
101 * @flags: request command flags
102 * @ctx: software queue cpu ctx
104 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
106 struct blk_mq_ctx *ctx)
108 enum hctx_type type = HCTX_TYPE_DEFAULT;
111 * The caller ensure that if REQ_HIPRI, poll must be enabled.
113 if (flags & REQ_HIPRI)
114 type = HCTX_TYPE_POLL;
115 else if ((flags & REQ_OP_MASK) == REQ_OP_READ)
116 type = HCTX_TYPE_READ;
118 return ctx->hctxs[type];
124 extern void blk_mq_sysfs_init(struct request_queue *q);
125 extern void blk_mq_sysfs_deinit(struct request_queue *q);
126 extern int __blk_mq_register_dev(struct device *dev, struct request_queue *q);
127 extern int blk_mq_sysfs_register(struct request_queue *q);
128 extern void blk_mq_sysfs_unregister(struct request_queue *q);
129 extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
131 void blk_mq_release(struct request_queue *q);
133 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
136 return per_cpu_ptr(q->queue_ctx, cpu);
140 * This assumes per-cpu software queueing queues. They could be per-node
141 * as well, for instance. For now this is hardcoded as-is. Note that we don't
142 * care about preemption, since we know the ctx's are persistent. This does
143 * mean that we can't rely on ctx always matching the currently running CPU.
145 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
147 return __blk_mq_get_ctx(q, raw_smp_processor_id());
150 struct blk_mq_alloc_data {
151 /* input parameter */
152 struct request_queue *q;
153 blk_mq_req_flags_t flags;
154 unsigned int shallow_depth;
155 unsigned int cmd_flags;
157 /* input & output parameter */
158 struct blk_mq_ctx *ctx;
159 struct blk_mq_hw_ctx *hctx;
162 static inline bool blk_mq_is_sbitmap_shared(unsigned int flags)
164 return flags & BLK_MQ_F_TAG_HCTX_SHARED;
167 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
169 if (data->q->elevator)
170 return data->hctx->sched_tags;
172 return data->hctx->tags;
175 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
177 return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
180 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
182 return hctx->nr_ctx && hctx->tags;
185 unsigned int blk_mq_in_flight(struct request_queue *q,
186 struct block_device *part);
187 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
188 unsigned int inflight[2]);
190 static inline void blk_mq_put_dispatch_budget(struct request_queue *q)
192 if (q->mq_ops->put_budget)
193 q->mq_ops->put_budget(q);
196 static inline bool blk_mq_get_dispatch_budget(struct request_queue *q)
198 if (q->mq_ops->get_budget)
199 return q->mq_ops->get_budget(q);
203 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
205 if (blk_mq_is_sbitmap_shared(hctx->flags))
206 atomic_inc(&hctx->queue->nr_active_requests_shared_sbitmap);
208 atomic_inc(&hctx->nr_active);
211 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
213 if (blk_mq_is_sbitmap_shared(hctx->flags))
214 atomic_dec(&hctx->queue->nr_active_requests_shared_sbitmap);
216 atomic_dec(&hctx->nr_active);
219 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
221 if (blk_mq_is_sbitmap_shared(hctx->flags))
222 return atomic_read(&hctx->queue->nr_active_requests_shared_sbitmap);
223 return atomic_read(&hctx->nr_active);
225 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
228 blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
229 rq->tag = BLK_MQ_NO_TAG;
231 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
232 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
233 __blk_mq_dec_active_requests(hctx);
237 static inline void blk_mq_put_driver_tag(struct request *rq)
239 if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
242 __blk_mq_put_driver_tag(rq->mq_hctx, rq);
245 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
249 for_each_possible_cpu(cpu)
250 qmap->mq_map[cpu] = 0;
254 * blk_mq_plug() - Get caller context plug
256 * @bio : the bio being submitted by the caller context
258 * Plugging, by design, may delay the insertion of BIOs into the elevator in
259 * order to increase BIO merging opportunities. This however can cause BIO
260 * insertion order to change from the order in which submit_bio() is being
261 * executed in the case of multiple contexts concurrently issuing BIOs to a
262 * device, even if these context are synchronized to tightly control BIO issuing
263 * order. While this is not a problem with regular block devices, this ordering
264 * change can cause write BIO failures with zoned block devices as these
265 * require sequential write patterns to zones. Prevent this from happening by
266 * ignoring the plug state of a BIO issuing context if the target request queue
267 * is for a zoned block device and the BIO to plug is a write operation.
269 * Return current->plug if the bio can be plugged and NULL otherwise
271 static inline struct blk_plug *blk_mq_plug(struct request_queue *q,
275 * For regular block devices or read operations, use the context plug
276 * which may be NULL if blk_start_plug() was not executed.
278 if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio)))
279 return current->plug;
281 /* Zoned block device write operation case: do not plug the BIO */
286 * For shared tag users, we track the number of currently active users
287 * and attempt to provide a fair share of the tag depth for each of them.
289 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
290 struct sbitmap_queue *bt)
292 unsigned int depth, users;
294 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
298 * Don't try dividing an ant
300 if (bt->sb.depth == 1)
303 if (blk_mq_is_sbitmap_shared(hctx->flags)) {
304 struct request_queue *q = hctx->queue;
305 struct blk_mq_tag_set *set = q->tag_set;
307 if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
309 users = atomic_read(&set->active_queues_shared_sbitmap);
311 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
313 users = atomic_read(&hctx->tags->active_queues);
320 * Allow at least some tags
322 depth = max((bt->sb.depth + users - 1) / users, 4U);
323 return __blk_mq_active_requests(hctx) < depth;