blob: 37d1602c4f7aa08b464577c675910046a4db3dde [file] [log] [blame]
Jens Axboe320ae512013-10-24 09:20:05 +01001#ifndef BLK_MQ_H
2#define BLK_MQ_H
3
4#include <linux/blkdev.h>
5
6struct blk_mq_tags;
Ming Leif70ced02014-09-25 23:23:47 +08007struct blk_flush_queue;
Jens Axboe320ae512013-10-24 09:20:05 +01008
9struct blk_mq_cpu_notifier {
10 struct list_head list;
11 void *data;
Jens Axboee814e712014-05-21 13:59:08 -060012 int (*notify)(void *data, unsigned long action, unsigned int cpu);
Jens Axboe320ae512013-10-24 09:20:05 +010013};
14
Jens Axboe1429d7c2014-05-19 09:23:55 -060015struct blk_mq_ctxmap {
Jens Axboe569fd0c2015-04-17 08:28:50 -060016 unsigned int size;
Jens Axboe1429d7c2014-05-19 09:23:55 -060017 unsigned int bits_per_word;
18 struct blk_align_bitmap *map;
19};
20
Jens Axboe320ae512013-10-24 09:20:05 +010021struct blk_mq_hw_ctx {
22 struct {
23 spinlock_t lock;
24 struct list_head dispatch;
25 } ____cacheline_aligned_in_smp;
26
27 unsigned long state; /* BLK_MQ_S_* flags */
Christoph Hellwig70f4db62014-04-16 10:48:08 -060028 struct delayed_work run_work;
29 struct delayed_work delay_work;
Jens Axboee4043dc2014-04-09 10:18:23 -060030 cpumask_var_t cpumask;
Jens Axboe506e9312014-05-07 10:26:44 -060031 int next_cpu;
32 int next_cpu_batch;
Jens Axboe320ae512013-10-24 09:20:05 +010033
34 unsigned long flags; /* BLK_MQ_F_* flags */
35
36 struct request_queue *queue;
Ming Leif70ced02014-09-25 23:23:47 +080037 struct blk_flush_queue *fq;
Jens Axboe320ae512013-10-24 09:20:05 +010038
39 void *driver_data;
40
Jens Axboe1429d7c2014-05-19 09:23:55 -060041 struct blk_mq_ctxmap ctx_map;
42
Jens Axboe4bb659b2014-05-09 09:36:49 -060043 unsigned int nr_ctx;
44 struct blk_mq_ctx **ctxs;
45
Alexander Gordeev8537b122014-06-17 22:12:35 -070046 atomic_t wait_index;
Jens Axboe320ae512013-10-24 09:20:05 +010047
Jens Axboe320ae512013-10-24 09:20:05 +010048 struct blk_mq_tags *tags;
49
50 unsigned long queued;
51 unsigned long run;
52#define BLK_MQ_MAX_DISPATCH_ORDER 10
53 unsigned long dispatched[BLK_MQ_MAX_DISPATCH_ORDER];
54
Jens Axboe320ae512013-10-24 09:20:05 +010055 unsigned int numa_node;
Jens Axboe17ded322015-01-07 10:44:04 -070056 unsigned int queue_num;
Jens Axboe320ae512013-10-24 09:20:05 +010057
Jens Axboe0d2602c2014-05-13 15:10:52 -060058 atomic_t nr_active;
59
Jens Axboe320ae512013-10-24 09:20:05 +010060 struct blk_mq_cpu_notifier cpu_notifier;
61 struct kobject kobj;
62};
63
Christoph Hellwig24d2f902014-04-15 14:14:00 -060064struct blk_mq_tag_set {
Jens Axboe320ae512013-10-24 09:20:05 +010065 struct blk_mq_ops *ops;
66 unsigned int nr_hw_queues;
Jens Axboee3a2b3f2014-05-20 11:49:02 -060067 unsigned int queue_depth; /* max hw supported */
Jens Axboe320ae512013-10-24 09:20:05 +010068 unsigned int reserved_tags;
69 unsigned int cmd_size; /* per-request extra data */
70 int numa_node;
71 unsigned int timeout;
72 unsigned int flags; /* BLK_MQ_F_* */
Christoph Hellwig24d2f902014-04-15 14:14:00 -060073 void *driver_data;
74
75 struct blk_mq_tags **tags;
Jens Axboe0d2602c2014-05-13 15:10:52 -060076
77 struct mutex tag_list_lock;
78 struct list_head tag_list;
Jens Axboe320ae512013-10-24 09:20:05 +010079};
80
Jens Axboe74c45052014-10-29 11:14:52 -060081struct blk_mq_queue_data {
82 struct request *rq;
83 struct list_head *list;
84 bool last;
85};
86
87typedef int (queue_rq_fn)(struct blk_mq_hw_ctx *, const struct blk_mq_queue_data *);
Jens Axboe320ae512013-10-24 09:20:05 +010088typedef struct blk_mq_hw_ctx *(map_queue_fn)(struct request_queue *, const int);
Christoph Hellwig0152fb62014-09-13 16:40:13 -070089typedef enum blk_eh_timer_return (timeout_fn)(struct request *, bool);
Jens Axboe320ae512013-10-24 09:20:05 +010090typedef int (init_hctx_fn)(struct blk_mq_hw_ctx *, void *, unsigned int);
91typedef void (exit_hctx_fn)(struct blk_mq_hw_ctx *, unsigned int);
Christoph Hellwig24d2f902014-04-15 14:14:00 -060092typedef int (init_request_fn)(void *, struct request *, unsigned int,
93 unsigned int, unsigned int);
94typedef void (exit_request_fn)(void *, struct request *, unsigned int,
95 unsigned int);
Jens Axboe320ae512013-10-24 09:20:05 +010096
Christoph Hellwig81481eb2014-09-13 16:40:11 -070097typedef void (busy_iter_fn)(struct blk_mq_hw_ctx *, struct request *, void *,
98 bool);
Keith Buschf26cdc82015-06-01 09:29:53 -060099typedef void (busy_tag_iter_fn)(struct request *, void *, bool);
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700100
Jens Axboe320ae512013-10-24 09:20:05 +0100101struct blk_mq_ops {
102 /*
103 * Queue request
104 */
105 queue_rq_fn *queue_rq;
106
107 /*
108 * Map to specific hardware queue
109 */
110 map_queue_fn *map_queue;
111
112 /*
113 * Called on request timeout
114 */
Christoph Hellwig0152fb62014-09-13 16:40:13 -0700115 timeout_fn *timeout;
Jens Axboe320ae512013-10-24 09:20:05 +0100116
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800117 softirq_done_fn *complete;
118
Jens Axboe320ae512013-10-24 09:20:05 +0100119 /*
Jens Axboe320ae512013-10-24 09:20:05 +0100120 * Called when the block layer side of a hardware queue has been
121 * set up, allowing the driver to allocate/init matching structures.
122 * Ditto for exit/teardown.
123 */
124 init_hctx_fn *init_hctx;
125 exit_hctx_fn *exit_hctx;
Christoph Hellwige9b267d2014-04-15 13:59:10 -0600126
127 /*
128 * Called for every command allocated by the block layer to allow
129 * the driver to set up driver specific data.
Ming Leif70ced02014-09-25 23:23:47 +0800130 *
131 * Tag greater than or equal to queue_depth is for setting up
132 * flush request.
133 *
Christoph Hellwige9b267d2014-04-15 13:59:10 -0600134 * Ditto for exit/teardown.
135 */
136 init_request_fn *init_request;
137 exit_request_fn *exit_request;
Jens Axboe320ae512013-10-24 09:20:05 +0100138};
139
140enum {
141 BLK_MQ_RQ_QUEUE_OK = 0, /* queued fine */
142 BLK_MQ_RQ_QUEUE_BUSY = 1, /* requeue IO for later */
143 BLK_MQ_RQ_QUEUE_ERROR = 2, /* end IO with error */
144
145 BLK_MQ_F_SHOULD_MERGE = 1 << 0,
Jens Axboe8a58d1f2014-08-15 12:38:41 -0600146 BLK_MQ_F_TAG_SHARED = 1 << 1,
147 BLK_MQ_F_SG_MERGE = 1 << 2,
148 BLK_MQ_F_SYSFS_UP = 1 << 3,
Jens Axboee167dfb2014-10-29 11:18:26 -0600149 BLK_MQ_F_DEFER_ISSUE = 1 << 4,
Shaohua Li24391c02015-01-23 14:18:00 -0700150 BLK_MQ_F_ALLOC_POLICY_START_BIT = 8,
151 BLK_MQ_F_ALLOC_POLICY_BITS = 1,
Jens Axboe320ae512013-10-24 09:20:05 +0100152
Jens Axboe5d12f902014-03-19 15:25:02 -0600153 BLK_MQ_S_STOPPED = 0,
Jens Axboe0d2602c2014-05-13 15:10:52 -0600154 BLK_MQ_S_TAG_ACTIVE = 1,
Jens Axboe320ae512013-10-24 09:20:05 +0100155
Jens Axboea4391c62014-06-05 15:21:56 -0600156 BLK_MQ_MAX_DEPTH = 10240,
Jens Axboe506e9312014-05-07 10:26:44 -0600157
158 BLK_MQ_CPU_WORK_BATCH = 8,
Jens Axboe320ae512013-10-24 09:20:05 +0100159};
Shaohua Li24391c02015-01-23 14:18:00 -0700160#define BLK_MQ_FLAG_TO_ALLOC_POLICY(flags) \
161 ((flags >> BLK_MQ_F_ALLOC_POLICY_START_BIT) & \
162 ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1))
163#define BLK_ALLOC_POLICY_TO_MQ_FLAG(policy) \
164 ((policy & ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1)) \
165 << BLK_MQ_F_ALLOC_POLICY_START_BIT)
Jens Axboe320ae512013-10-24 09:20:05 +0100166
Christoph Hellwig24d2f902014-04-15 14:14:00 -0600167struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *);
Mike Snitzerb62c21b2015-03-12 23:56:02 -0400168struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
169 struct request_queue *q);
Tejun Heo17497ac2014-09-24 13:31:50 -0400170void blk_mq_finish_init(struct request_queue *q);
Jens Axboe320ae512013-10-24 09:20:05 +0100171int blk_mq_register_disk(struct gendisk *);
172void blk_mq_unregister_disk(struct gendisk *);
Jens Axboe320ae512013-10-24 09:20:05 +0100173
Christoph Hellwig24d2f902014-04-15 14:14:00 -0600174int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set);
175void blk_mq_free_tag_set(struct blk_mq_tag_set *set);
176
Jens Axboe320ae512013-10-24 09:20:05 +0100177void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule);
178
Christoph Hellwigfeb71da2014-02-20 15:32:37 -0800179void blk_mq_insert_request(struct request *, bool, bool, bool);
Jens Axboe320ae512013-10-24 09:20:05 +0100180void blk_mq_free_request(struct request *rq);
Jens Axboe7c7f2f22014-11-17 10:41:57 -0700181void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *, struct request *rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100182bool blk_mq_can_queue(struct blk_mq_hw_ctx *);
Christoph Hellwig4ce01dd2014-05-27 20:59:46 +0200183struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
184 gfp_t gfp, bool reserved);
Jens Axboe0e62f512014-06-04 10:23:49 -0600185struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag);
Keith Buschf26cdc82015-06-01 09:29:53 -0600186struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags);
Jens Axboe320ae512013-10-24 09:20:05 +0100187
Bart Van Assche205fb5f2014-10-30 14:45:11 +0100188enum {
189 BLK_MQ_UNIQUE_TAG_BITS = 16,
190 BLK_MQ_UNIQUE_TAG_MASK = (1 << BLK_MQ_UNIQUE_TAG_BITS) - 1,
191};
192
193u32 blk_mq_unique_tag(struct request *rq);
194
195static inline u16 blk_mq_unique_tag_to_hwq(u32 unique_tag)
196{
197 return unique_tag >> BLK_MQ_UNIQUE_TAG_BITS;
198}
199
200static inline u16 blk_mq_unique_tag_to_tag(u32 unique_tag)
201{
202 return unique_tag & BLK_MQ_UNIQUE_TAG_MASK;
203}
204
Jens Axboe320ae512013-10-24 09:20:05 +0100205struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *, const int ctx_index);
Jens Axboef14bbe72014-05-27 12:06:53 -0600206struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *, unsigned int, int);
Jens Axboe320ae512013-10-24 09:20:05 +0100207
Keith Busch973c0192015-01-07 18:55:43 -0700208int blk_mq_request_started(struct request *rq);
Christoph Hellwige2490072014-09-13 16:40:09 -0700209void blk_mq_start_request(struct request *rq);
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700210void blk_mq_end_request(struct request *rq, int error);
211void __blk_mq_end_request(struct request *rq, int error);
Jens Axboe320ae512013-10-24 09:20:05 +0100212
Christoph Hellwiged0791b2014-04-16 09:44:57 +0200213void blk_mq_requeue_request(struct request *rq);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600214void blk_mq_add_to_requeue_list(struct request *rq, bool at_head);
Keith Buschc68ed592015-01-07 18:55:44 -0700215void blk_mq_cancel_requeue_work(struct request_queue *q);
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600216void blk_mq_kick_requeue_list(struct request_queue *q);
Jens Axboe1885b242015-01-07 18:55:45 -0700217void blk_mq_abort_requeue_list(struct request_queue *q);
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800218void blk_mq_complete_request(struct request *rq);
219
Jens Axboe320ae512013-10-24 09:20:05 +0100220void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx);
221void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx);
Christoph Hellwig280d45f2013-10-25 14:45:58 +0100222void blk_mq_stop_hw_queues(struct request_queue *q);
Christoph Hellwig2f268552014-04-16 09:44:56 +0200223void blk_mq_start_hw_queues(struct request_queue *q);
Christoph Hellwig1b4a3252014-04-16 09:44:54 +0200224void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async);
Mike Snitzerb94ec292015-03-11 23:56:38 -0400225void blk_mq_run_hw_queues(struct request_queue *q, bool async);
Christoph Hellwig70f4db62014-04-16 10:48:08 -0600226void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs);
Christoph Hellwig81481eb2014-09-13 16:40:11 -0700227void blk_mq_tag_busy_iter(struct blk_mq_hw_ctx *hctx, busy_iter_fn *fn,
228 void *priv);
Keith Buschf26cdc82015-06-01 09:29:53 -0600229void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
230 void *priv);
Jens Axboec761d962015-01-02 15:05:12 -0700231void blk_mq_freeze_queue(struct request_queue *q);
Keith Buschb4c6a022014-12-19 17:54:14 -0700232void blk_mq_unfreeze_queue(struct request_queue *q);
233void blk_mq_freeze_queue_start(struct request_queue *q);
Jens Axboe320ae512013-10-24 09:20:05 +0100234
235/*
236 * Driver command data is immediately after the request. So subtract request
Jens Axboe2963e3f2015-04-09 15:54:05 -0600237 * size to get back to the original request, add request size to get the PDU.
Jens Axboe320ae512013-10-24 09:20:05 +0100238 */
239static inline struct request *blk_mq_rq_from_pdu(void *pdu)
240{
241 return pdu - sizeof(struct request);
242}
243static inline void *blk_mq_rq_to_pdu(struct request *rq)
244{
Jens Axboe2963e3f2015-04-09 15:54:05 -0600245 return rq + 1;
Jens Axboe320ae512013-10-24 09:20:05 +0100246}
247
Jens Axboe320ae512013-10-24 09:20:05 +0100248#define queue_for_each_hw_ctx(q, hctx, i) \
Jose Alonso0d0b7d42014-01-28 08:09:46 -0700249 for ((i) = 0; (i) < (q)->nr_hw_queues && \
250 ({ hctx = (q)->queue_hw_ctx[i]; 1; }); (i)++)
Jens Axboe320ae512013-10-24 09:20:05 +0100251
252#define queue_for_each_ctx(q, ctx, i) \
Jose Alonso0d0b7d42014-01-28 08:09:46 -0700253 for ((i) = 0; (i) < (q)->nr_queues && \
254 ({ ctx = per_cpu_ptr((q)->queue_ctx, (i)); 1; }); (i)++)
Jens Axboe320ae512013-10-24 09:20:05 +0100255
256#define hctx_for_each_ctx(hctx, ctx, i) \
Jose Alonso0d0b7d42014-01-28 08:09:46 -0700257 for ((i) = 0; (i) < (hctx)->nr_ctx && \
258 ({ ctx = (hctx)->ctxs[(i)]; 1; }); (i)++)
Jens Axboe320ae512013-10-24 09:20:05 +0100259
260#define blk_ctx_sum(q, sum) \
261({ \
262 struct blk_mq_ctx *__x; \
263 unsigned int __ret = 0, __i; \
264 \
265 queue_for_each_ctx((q), __x, __i) \
266 __ret += sum; \
267 __ret; \
268})
269
270#endif