xref: /kernel/linux/linux-5.10/block/blk-mq.h (revision 8c2ecf20)
1/* SPDX-License-Identifier: GPL-2.0 */
2#ifndef INT_BLK_MQ_H
3#define INT_BLK_MQ_H
4
5#include "blk-stat.h"
6#include "blk-mq-tag.h"
7
8struct blk_mq_tag_set;
9
10struct blk_mq_ctxs {
11	struct kobject kobj;
12	struct blk_mq_ctx __percpu	*queue_ctx;
13};
14
15/**
16 * struct blk_mq_ctx - State for a software queue facing the submitting CPUs
17 */
18struct blk_mq_ctx {
19	struct {
20		spinlock_t		lock;
21		struct list_head	rq_lists[HCTX_MAX_TYPES];
22	} ____cacheline_aligned_in_smp;
23
24	unsigned int		cpu;
25	unsigned short		index_hw[HCTX_MAX_TYPES];
26	struct blk_mq_hw_ctx 	*hctxs[HCTX_MAX_TYPES];
27
28	/* incremented at dispatch time */
29	unsigned long		rq_dispatched[2];
30	unsigned long		rq_merged;
31
32	/* incremented at completion time */
33	unsigned long		____cacheline_aligned_in_smp rq_completed[2];
34
35	struct request_queue	*queue;
36	struct blk_mq_ctxs      *ctxs;
37	struct kobject		kobj;
38} ____cacheline_aligned_in_smp;
39
40void blk_mq_exit_queue(struct request_queue *q);
41int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr);
42void blk_mq_wake_waiters(struct request_queue *q);
43bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *,
44			     unsigned int);
45void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
46				bool kick_requeue_list);
47void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list);
48struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
49					struct blk_mq_ctx *start);
50void blk_mq_put_rq_ref(struct request *rq);
51
52/*
53 * Internal helpers for allocating/freeing the request map
54 */
55void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
56		     unsigned int hctx_idx);
57void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags);
58struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
59					unsigned int hctx_idx,
60					unsigned int nr_tags,
61					unsigned int reserved_tags,
62					unsigned int flags);
63int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
64		     unsigned int hctx_idx, unsigned int depth);
65
66/*
67 * Internal helpers for request insertion into sw queues
68 */
69void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
70				bool at_head);
71void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
72				  bool run_queue);
73void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
74				struct list_head *list);
75
76/* Used by blk_insert_cloned_request() to issue request directly */
77blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last);
78void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
79				    struct list_head *list);
80
81/*
82 * CPU -> queue mappings
83 */
84extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
85
86/*
87 * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
88 * @q: request queue
89 * @type: the hctx type index
90 * @cpu: CPU
91 */
92static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
93							  enum hctx_type type,
94							  unsigned int cpu)
95{
96	return q->queue_hw_ctx[q->tag_set->map[type].mq_map[cpu]];
97}
98
99/*
100 * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
101 * @q: request queue
102 * @flags: request command flags
103 * @cpu: cpu ctx
104 */
105static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
106						     unsigned int flags,
107						     struct blk_mq_ctx *ctx)
108{
109	enum hctx_type type = HCTX_TYPE_DEFAULT;
110
111	/*
112	 * The caller ensure that if REQ_HIPRI, poll must be enabled.
113	 */
114	if (flags & REQ_HIPRI)
115		type = HCTX_TYPE_POLL;
116	else if ((flags & REQ_OP_MASK) == REQ_OP_READ)
117		type = HCTX_TYPE_READ;
118
119	return ctx->hctxs[type];
120}
121
122/*
123 * sysfs helpers
124 */
125extern void blk_mq_sysfs_init(struct request_queue *q);
126extern void blk_mq_sysfs_deinit(struct request_queue *q);
127extern int __blk_mq_register_dev(struct device *dev, struct request_queue *q);
128extern int blk_mq_sysfs_register(struct request_queue *q);
129extern void blk_mq_sysfs_unregister(struct request_queue *q);
130extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
131
132void blk_mq_cancel_work_sync(struct request_queue *q);
133void blk_mq_release(struct request_queue *q);
134
135static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
136					   unsigned int cpu)
137{
138	return per_cpu_ptr(q->queue_ctx, cpu);
139}
140
141/*
142 * This assumes per-cpu software queueing queues. They could be per-node
143 * as well, for instance. For now this is hardcoded as-is. Note that we don't
144 * care about preemption, since we know the ctx's are persistent. This does
145 * mean that we can't rely on ctx always matching the currently running CPU.
146 */
147static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
148{
149	return __blk_mq_get_ctx(q, raw_smp_processor_id());
150}
151
152struct blk_mq_alloc_data {
153	/* input parameter */
154	struct request_queue *q;
155	blk_mq_req_flags_t flags;
156	unsigned int shallow_depth;
157	unsigned int cmd_flags;
158
159	/* input & output parameter */
160	struct blk_mq_ctx *ctx;
161	struct blk_mq_hw_ctx *hctx;
162};
163
164static inline bool blk_mq_is_sbitmap_shared(unsigned int flags)
165{
166	return flags & BLK_MQ_F_TAG_HCTX_SHARED;
167}
168
169static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
170{
171	if (data->q->elevator)
172		return data->hctx->sched_tags;
173
174	return data->hctx->tags;
175}
176
177static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
178{
179	return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
180}
181
182static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
183{
184	return hctx->nr_ctx && hctx->tags;
185}
186
187unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part);
188void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
189			 unsigned int inflight[2]);
190
191static inline void blk_mq_put_dispatch_budget(struct request_queue *q)
192{
193	if (q->mq_ops->put_budget)
194		q->mq_ops->put_budget(q);
195}
196
197static inline bool blk_mq_get_dispatch_budget(struct request_queue *q)
198{
199	if (q->mq_ops->get_budget)
200		return q->mq_ops->get_budget(q);
201	return true;
202}
203
204static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
205{
206	if (blk_mq_is_sbitmap_shared(hctx->flags))
207		atomic_inc(&hctx->queue->nr_active_requests_shared_sbitmap);
208	else
209		atomic_inc(&hctx->nr_active);
210}
211
212static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
213{
214	if (blk_mq_is_sbitmap_shared(hctx->flags))
215		atomic_dec(&hctx->queue->nr_active_requests_shared_sbitmap);
216	else
217		atomic_dec(&hctx->nr_active);
218}
219
220static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
221{
222	if (blk_mq_is_sbitmap_shared(hctx->flags))
223		return atomic_read(&hctx->queue->nr_active_requests_shared_sbitmap);
224	return atomic_read(&hctx->nr_active);
225}
226static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
227					   struct request *rq)
228{
229	blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
230	rq->tag = BLK_MQ_NO_TAG;
231
232	if (rq->rq_flags & RQF_MQ_INFLIGHT) {
233		rq->rq_flags &= ~RQF_MQ_INFLIGHT;
234		__blk_mq_dec_active_requests(hctx);
235	}
236}
237
238static inline void blk_mq_put_driver_tag(struct request *rq)
239{
240	if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
241		return;
242
243	__blk_mq_put_driver_tag(rq->mq_hctx, rq);
244}
245
246static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
247{
248	int cpu;
249
250	for_each_possible_cpu(cpu)
251		qmap->mq_map[cpu] = 0;
252}
253
254/*
255 * blk_mq_plug() - Get caller context plug
256 * @q: request queue
257 * @bio : the bio being submitted by the caller context
258 *
259 * Plugging, by design, may delay the insertion of BIOs into the elevator in
260 * order to increase BIO merging opportunities. This however can cause BIO
261 * insertion order to change from the order in which submit_bio() is being
262 * executed in the case of multiple contexts concurrently issuing BIOs to a
263 * device, even if these context are synchronized to tightly control BIO issuing
264 * order. While this is not a problem with regular block devices, this ordering
265 * change can cause write BIO failures with zoned block devices as these
266 * require sequential write patterns to zones. Prevent this from happening by
267 * ignoring the plug state of a BIO issuing context if the target request queue
268 * is for a zoned block device and the BIO to plug is a write operation.
269 *
270 * Return current->plug if the bio can be plugged and NULL otherwise
271 */
272static inline struct blk_plug *blk_mq_plug(struct request_queue *q,
273					   struct bio *bio)
274{
275	/*
276	 * For regular block devices or read operations, use the context plug
277	 * which may be NULL if blk_start_plug() was not executed.
278	 */
279	if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio)))
280		return current->plug;
281
282	/* Zoned block device write operation case: do not plug the BIO */
283	return NULL;
284}
285
286/* Free all requests on the list */
287static inline void blk_mq_free_requests(struct list_head *list)
288{
289	while (!list_empty(list)) {
290		struct request *rq = list_entry_rq(list->next);
291
292		list_del_init(&rq->queuelist);
293		blk_mq_free_request(rq);
294	}
295}
296
297/*
298 * For shared tag users, we track the number of currently active users
299 * and attempt to provide a fair share of the tag depth for each of them.
300 */
301static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
302				  struct sbitmap_queue *bt)
303{
304	unsigned int depth, users;
305
306	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
307		return true;
308
309	/*
310	 * Don't try dividing an ant
311	 */
312	if (bt->sb.depth == 1)
313		return true;
314
315	if (blk_mq_is_sbitmap_shared(hctx->flags)) {
316		struct request_queue *q = hctx->queue;
317		struct blk_mq_tag_set *set = q->tag_set;
318
319		if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
320			return true;
321		users = atomic_read(&set->active_queues_shared_sbitmap);
322	} else {
323		if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
324			return true;
325		users = atomic_read(&hctx->tags->active_queues);
326	}
327
328	if (!users)
329		return true;
330
331	/*
332	 * Allow at least some tags
333	 */
334	depth = max((bt->sb.depth + users - 1) / users, 4U);
335	return __blk_mq_active_requests(hctx) < depth;
336}
337
338
339#endif
340