1 #ifndef BLK_MQ_H 2 #define BLK_MQ_H 3 4 #include <linux/blkdev.h> 5 6 struct blk_mq_tags; 7 struct blk_flush_queue; 8 9 struct blk_mq_cpu_notifier { 10 struct list_head list; 11 void *data; 12 int (*notify)(void *data, unsigned long action, unsigned int cpu); 13 }; 14 15 struct blk_mq_ctxmap { 16 unsigned int map_size; 17 unsigned int bits_per_word; 18 struct blk_align_bitmap *map; 19 }; 20 21 struct 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 */ 28 struct delayed_work run_work; 29 struct delayed_work delay_work; 30 cpumask_var_t cpumask; 31 int next_cpu; 32 int next_cpu_batch; 33 34 unsigned long flags; /* BLK_MQ_F_* flags */ 35 36 struct request_queue *queue; 37 unsigned int queue_num; 38 struct blk_flush_queue *fq; 39 40 void *driver_data; 41 42 struct blk_mq_ctxmap ctx_map; 43 44 unsigned int nr_ctx; 45 struct blk_mq_ctx **ctxs; 46 47 atomic_t wait_index; 48 49 struct blk_mq_tags *tags; 50 51 unsigned long queued; 52 unsigned long run; 53 #define BLK_MQ_MAX_DISPATCH_ORDER 10 54 unsigned long dispatched[BLK_MQ_MAX_DISPATCH_ORDER]; 55 56 unsigned int numa_node; 57 unsigned int cmd_size; /* per-request extra data */ 58 59 atomic_t nr_active; 60 61 struct blk_mq_cpu_notifier cpu_notifier; 62 struct kobject kobj; 63 }; 64 65 struct blk_mq_tag_set { 66 struct blk_mq_ops *ops; 67 unsigned int nr_hw_queues; 68 unsigned int queue_depth; /* max hw supported */ 69 unsigned int reserved_tags; 70 unsigned int cmd_size; /* per-request extra data */ 71 int numa_node; 72 unsigned int timeout; 73 unsigned int flags; /* BLK_MQ_F_* */ 74 void *driver_data; 75 76 struct blk_mq_tags **tags; 77 78 struct mutex tag_list_lock; 79 struct list_head tag_list; 80 }; 81 82 struct blk_mq_queue_data { 83 struct request *rq; 84 struct list_head *list; 85 bool last; 86 }; 87 88 typedef int (queue_rq_fn)(struct blk_mq_hw_ctx *, const struct blk_mq_queue_data *); 89 typedef struct blk_mq_hw_ctx *(map_queue_fn)(struct request_queue *, const int); 90 typedef enum blk_eh_timer_return (timeout_fn)(struct request *, bool); 91 typedef int (init_hctx_fn)(struct blk_mq_hw_ctx *, void *, unsigned int); 92 typedef void (exit_hctx_fn)(struct blk_mq_hw_ctx *, unsigned int); 93 typedef int (init_request_fn)(void *, struct request *, unsigned int, 94 unsigned int, unsigned int); 95 typedef void (exit_request_fn)(void *, struct request *, unsigned int, 96 unsigned int); 97 98 typedef void (busy_iter_fn)(struct blk_mq_hw_ctx *, struct request *, void *, 99 bool); 100 101 struct 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 */ 115 timeout_fn *timeout; 116 117 softirq_done_fn *complete; 118 119 /* 120 * 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; 126 127 /* 128 * Called for every command allocated by the block layer to allow 129 * the driver to set up driver specific data. 130 * 131 * Tag greater than or equal to queue_depth is for setting up 132 * flush request. 133 * 134 * Ditto for exit/teardown. 135 */ 136 init_request_fn *init_request; 137 exit_request_fn *exit_request; 138 }; 139 140 enum { 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, 146 BLK_MQ_F_TAG_SHARED = 1 << 1, 147 BLK_MQ_F_SG_MERGE = 1 << 2, 148 BLK_MQ_F_SYSFS_UP = 1 << 3, 149 BLK_MQ_F_DEFER_ISSUE = 1 << 4, 150 151 BLK_MQ_S_STOPPED = 0, 152 BLK_MQ_S_TAG_ACTIVE = 1, 153 154 BLK_MQ_MAX_DEPTH = 10240, 155 156 BLK_MQ_CPU_WORK_BATCH = 8, 157 }; 158 159 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *); 160 void blk_mq_finish_init(struct request_queue *q); 161 int blk_mq_register_disk(struct gendisk *); 162 void blk_mq_unregister_disk(struct gendisk *); 163 164 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set); 165 void blk_mq_free_tag_set(struct blk_mq_tag_set *set); 166 167 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule); 168 169 void blk_mq_insert_request(struct request *, bool, bool, bool); 170 void blk_mq_run_queues(struct request_queue *q, bool async); 171 void blk_mq_free_request(struct request *rq); 172 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *, struct request *rq); 173 bool blk_mq_can_queue(struct blk_mq_hw_ctx *); 174 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, 175 gfp_t gfp, bool reserved); 176 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag); 177 178 enum { 179 BLK_MQ_UNIQUE_TAG_BITS = 16, 180 BLK_MQ_UNIQUE_TAG_MASK = (1 << BLK_MQ_UNIQUE_TAG_BITS) - 1, 181 }; 182 183 u32 blk_mq_unique_tag(struct request *rq); 184 185 static inline u16 blk_mq_unique_tag_to_hwq(u32 unique_tag) 186 { 187 return unique_tag >> BLK_MQ_UNIQUE_TAG_BITS; 188 } 189 190 static inline u16 blk_mq_unique_tag_to_tag(u32 unique_tag) 191 { 192 return unique_tag & BLK_MQ_UNIQUE_TAG_MASK; 193 } 194 195 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *, const int ctx_index); 196 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *, unsigned int, int); 197 198 void blk_mq_start_request(struct request *rq); 199 void blk_mq_end_request(struct request *rq, int error); 200 void __blk_mq_end_request(struct request *rq, int error); 201 202 void blk_mq_requeue_request(struct request *rq); 203 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head); 204 void blk_mq_kick_requeue_list(struct request_queue *q); 205 void blk_mq_complete_request(struct request *rq); 206 207 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx); 208 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx); 209 void blk_mq_stop_hw_queues(struct request_queue *q); 210 void blk_mq_start_hw_queues(struct request_queue *q); 211 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async); 212 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs); 213 void blk_mq_tag_busy_iter(struct blk_mq_hw_ctx *hctx, busy_iter_fn *fn, 214 void *priv); 215 216 /* 217 * Driver command data is immediately after the request. So subtract request 218 * size to get back to the original request. 219 */ 220 static inline struct request *blk_mq_rq_from_pdu(void *pdu) 221 { 222 return pdu - sizeof(struct request); 223 } 224 static inline void *blk_mq_rq_to_pdu(struct request *rq) 225 { 226 return (void *) rq + sizeof(*rq); 227 } 228 229 #define queue_for_each_hw_ctx(q, hctx, i) \ 230 for ((i) = 0; (i) < (q)->nr_hw_queues && \ 231 ({ hctx = (q)->queue_hw_ctx[i]; 1; }); (i)++) 232 233 #define queue_for_each_ctx(q, ctx, i) \ 234 for ((i) = 0; (i) < (q)->nr_queues && \ 235 ({ ctx = per_cpu_ptr((q)->queue_ctx, (i)); 1; }); (i)++) 236 237 #define hctx_for_each_ctx(hctx, ctx, i) \ 238 for ((i) = 0; (i) < (hctx)->nr_ctx && \ 239 ({ ctx = (hctx)->ctxs[(i)]; 1; }); (i)++) 240 241 #define blk_ctx_sum(q, sum) \ 242 ({ \ 243 struct blk_mq_ctx *__x; \ 244 unsigned int __ret = 0, __i; \ 245 \ 246 queue_for_each_ctx((q), __x, __i) \ 247 __ret += sum; \ 248 __ret; \ 249 }) 250 251 #endif 252