blk-mq: improve support for shared tags maps
[linux-2.6-block.git] / include / linux / blk-mq.h
1 #ifndef BLK_MQ_H
2 #define BLK_MQ_H
3
4 #include <linux/blkdev.h>
5
6 struct blk_mq_tags;
7
8 struct blk_mq_cpu_notifier {
9         struct list_head list;
10         void *data;
11         void (*notify)(void *data, unsigned long action, unsigned int cpu);
12 };
13
14 struct blk_mq_hw_ctx {
15         struct {
16                 spinlock_t              lock;
17                 struct list_head        dispatch;
18         } ____cacheline_aligned_in_smp;
19
20         unsigned long           state;          /* BLK_MQ_S_* flags */
21         struct delayed_work     run_work;
22         struct delayed_work     delay_work;
23         cpumask_var_t           cpumask;
24         int                     next_cpu;
25         int                     next_cpu_batch;
26
27         unsigned long           flags;          /* BLK_MQ_F_* flags */
28
29         struct request_queue    *queue;
30         unsigned int            queue_num;
31
32         void                    *driver_data;
33
34         unsigned int            nr_ctx_map;
35         unsigned long           *ctx_map;
36         unsigned int            nr_ctx;
37         struct blk_mq_ctx       **ctxs;
38
39         unsigned int            wait_index;
40
41         struct blk_mq_tags      *tags;
42
43         unsigned long           queued;
44         unsigned long           run;
45 #define BLK_MQ_MAX_DISPATCH_ORDER       10
46         unsigned long           dispatched[BLK_MQ_MAX_DISPATCH_ORDER];
47
48         unsigned int            numa_node;
49         unsigned int            cmd_size;       /* per-request extra data */
50
51         atomic_t                nr_active;
52
53         struct blk_mq_cpu_notifier      cpu_notifier;
54         struct kobject          kobj;
55 };
56
57 struct blk_mq_tag_set {
58         struct blk_mq_ops       *ops;
59         unsigned int            nr_hw_queues;
60         unsigned int            queue_depth;
61         unsigned int            reserved_tags;
62         unsigned int            cmd_size;       /* per-request extra data */
63         int                     numa_node;
64         unsigned int            timeout;
65         unsigned int            flags;          /* BLK_MQ_F_* */
66         void                    *driver_data;
67
68         struct blk_mq_tags      **tags;
69
70         struct mutex            tag_list_lock;
71         struct list_head        tag_list;
72 };
73
74 typedef int (queue_rq_fn)(struct blk_mq_hw_ctx *, struct request *);
75 typedef struct blk_mq_hw_ctx *(map_queue_fn)(struct request_queue *, const int);
76 typedef struct blk_mq_hw_ctx *(alloc_hctx_fn)(struct blk_mq_tag_set *,
77                 unsigned int);
78 typedef void (free_hctx_fn)(struct blk_mq_hw_ctx *, unsigned int);
79 typedef int (init_hctx_fn)(struct blk_mq_hw_ctx *, void *, unsigned int);
80 typedef void (exit_hctx_fn)(struct blk_mq_hw_ctx *, unsigned int);
81 typedef int (init_request_fn)(void *, struct request *, unsigned int,
82                 unsigned int, unsigned int);
83 typedef void (exit_request_fn)(void *, struct request *, unsigned int,
84                 unsigned int);
85
86 struct blk_mq_ops {
87         /*
88          * Queue request
89          */
90         queue_rq_fn             *queue_rq;
91
92         /*
93          * Map to specific hardware queue
94          */
95         map_queue_fn            *map_queue;
96
97         /*
98          * Called on request timeout
99          */
100         rq_timed_out_fn         *timeout;
101
102         softirq_done_fn         *complete;
103
104         /*
105          * Override for hctx allocations (should probably go)
106          */
107         alloc_hctx_fn           *alloc_hctx;
108         free_hctx_fn            *free_hctx;
109
110         /*
111          * Called when the block layer side of a hardware queue has been
112          * set up, allowing the driver to allocate/init matching structures.
113          * Ditto for exit/teardown.
114          */
115         init_hctx_fn            *init_hctx;
116         exit_hctx_fn            *exit_hctx;
117
118         /*
119          * Called for every command allocated by the block layer to allow
120          * the driver to set up driver specific data.
121          * Ditto for exit/teardown.
122          */
123         init_request_fn         *init_request;
124         exit_request_fn         *exit_request;
125 };
126
127 enum {
128         BLK_MQ_RQ_QUEUE_OK      = 0,    /* queued fine */
129         BLK_MQ_RQ_QUEUE_BUSY    = 1,    /* requeue IO for later */
130         BLK_MQ_RQ_QUEUE_ERROR   = 2,    /* end IO with error */
131
132         BLK_MQ_F_SHOULD_MERGE   = 1 << 0,
133         BLK_MQ_F_SHOULD_SORT    = 1 << 1,
134         BLK_MQ_F_TAG_SHARED     = 1 << 2,
135
136         BLK_MQ_S_STOPPED        = 0,
137         BLK_MQ_S_TAG_ACTIVE     = 1,
138
139         BLK_MQ_MAX_DEPTH        = 2048,
140
141         BLK_MQ_CPU_WORK_BATCH   = 8,
142 };
143
144 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *);
145 int blk_mq_register_disk(struct gendisk *);
146 void blk_mq_unregister_disk(struct gendisk *);
147
148 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set);
149 void blk_mq_free_tag_set(struct blk_mq_tag_set *set);
150
151 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule);
152
153 void blk_mq_insert_request(struct request *, bool, bool, bool);
154 void blk_mq_run_queues(struct request_queue *q, bool async);
155 void blk_mq_free_request(struct request *rq);
156 bool blk_mq_can_queue(struct blk_mq_hw_ctx *);
157 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp);
158 struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw, gfp_t gfp);
159 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag);
160
161 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *, const int ctx_index);
162 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *, unsigned int);
163 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *, unsigned int);
164
165 void blk_mq_end_io(struct request *rq, int error);
166 void __blk_mq_end_io(struct request *rq, int error);
167
168 void blk_mq_requeue_request(struct request *rq);
169
170 void blk_mq_complete_request(struct request *rq);
171
172 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx);
173 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx);
174 void blk_mq_stop_hw_queues(struct request_queue *q);
175 void blk_mq_start_hw_queues(struct request_queue *q);
176 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async);
177 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs);
178
179 /*
180  * Driver command data is immediately after the request. So subtract request
181  * size to get back to the original request.
182  */
183 static inline struct request *blk_mq_rq_from_pdu(void *pdu)
184 {
185         return pdu - sizeof(struct request);
186 }
187 static inline void *blk_mq_rq_to_pdu(struct request *rq)
188 {
189         return (void *) rq + sizeof(*rq);
190 }
191
192 #define queue_for_each_hw_ctx(q, hctx, i)                               \
193         for ((i) = 0; (i) < (q)->nr_hw_queues &&                        \
194              ({ hctx = (q)->queue_hw_ctx[i]; 1; }); (i)++)
195
196 #define queue_for_each_ctx(q, ctx, i)                                   \
197         for ((i) = 0; (i) < (q)->nr_queues &&                           \
198              ({ ctx = per_cpu_ptr((q)->queue_ctx, (i)); 1; }); (i)++)
199
200 #define hctx_for_each_ctx(hctx, ctx, i)                                 \
201         for ((i) = 0; (i) < (hctx)->nr_ctx &&                           \
202              ({ ctx = (hctx)->ctxs[(i)]; 1; }); (i)++)
203
204 #define blk_ctx_sum(q, sum)                                             \
205 ({                                                                      \
206         struct blk_mq_ctx *__x;                                         \
207         unsigned int __ret = 0, __i;                                    \
208                                                                         \
209         queue_for_each_ctx((q), __x, __i)                               \
210                 __ret += sum;                                           \
211         __ret;                                                          \
212 })
213
214 #endif