1#ifndef BLK_MQ_H 2#define BLK_MQ_H 3 4#include <linux/blkdev.h> 5 6struct blk_mq_tags; 7struct blk_flush_queue; 8 9struct 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 15struct blk_mq_ctxmap { 16 unsigned int size; 17 unsigned int bits_per_word; 18 struct blk_align_bitmap *map; 19}; 20 21struct 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 struct blk_flush_queue *fq; 38 39 void *driver_data; 40 41 struct blk_mq_ctxmap ctx_map; 42 43 unsigned int nr_ctx; 44 struct blk_mq_ctx **ctxs; 45 46 atomic_t wait_index; 47 48 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 55 unsigned int numa_node; 56 unsigned int queue_num; 57 58 atomic_t nr_active; 59 60 struct blk_mq_cpu_notifier cpu_notifier; 61 struct kobject kobj; 62}; 63 64struct blk_mq_tag_set { 65 struct blk_mq_ops *ops; 66 unsigned int nr_hw_queues; 67 unsigned int queue_depth; /* max hw supported */ 68 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_* */ 73 void *driver_data; 74 75 struct blk_mq_tags **tags; 76 77 struct mutex tag_list_lock; 78 struct list_head tag_list; 79}; 80 81struct 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 *); 88typedef struct blk_mq_hw_ctx *(map_queue_fn)(struct request_queue *, const int); 89typedef enum blk_eh_timer_return (timeout_fn)(struct request *, bool); 90typedef int (init_hctx_fn)(struct blk_mq_hw_ctx *, void *, unsigned int); 91typedef void (exit_hctx_fn)(struct blk_mq_hw_ctx *, unsigned int); 92typedef 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); 96 97typedef void (busy_iter_fn)(struct blk_mq_hw_ctx *, struct request *, void *, 98 bool); 99 100struct blk_mq_ops { 101 /* 102 * Queue request 103 */ 104 queue_rq_fn *queue_rq; 105 106 /* 107 * Map to specific hardware queue 108 */ 109 map_queue_fn *map_queue; 110 111 /* 112 * Called on request timeout 113 */ 114 timeout_fn *timeout; 115 116 softirq_done_fn *complete; 117 118 /* 119 * Called when the block layer side of a hardware queue has been 120 * set up, allowing the driver to allocate/init matching structures. 121 * Ditto for exit/teardown. 122 */ 123 init_hctx_fn *init_hctx; 124 exit_hctx_fn *exit_hctx; 125 126 /* 127 * Called for every command allocated by the block layer to allow 128 * the driver to set up driver specific data. 129 * 130 * Tag greater than or equal to queue_depth is for setting up 131 * flush request. 132 * 133 * Ditto for exit/teardown. 134 */ 135 init_request_fn *init_request; 136 exit_request_fn *exit_request; 137}; 138 139enum { 140 BLK_MQ_RQ_QUEUE_OK = 0, /* queued fine */ 141 BLK_MQ_RQ_QUEUE_BUSY = 1, /* requeue IO for later */ 142 BLK_MQ_RQ_QUEUE_ERROR = 2, /* end IO with error */ 143 144 BLK_MQ_F_SHOULD_MERGE = 1 << 0, 145 BLK_MQ_F_TAG_SHARED = 1 << 1, 146 BLK_MQ_F_SG_MERGE = 1 << 2, 147 BLK_MQ_F_SYSFS_UP = 1 << 3, 148 BLK_MQ_F_DEFER_ISSUE = 1 << 4, 149 BLK_MQ_F_ALLOC_POLICY_START_BIT = 8, 150 BLK_MQ_F_ALLOC_POLICY_BITS = 1, 151 152 BLK_MQ_S_STOPPED = 0, 153 BLK_MQ_S_TAG_ACTIVE = 1, 154 155 BLK_MQ_MAX_DEPTH = 10240, 156 157 BLK_MQ_CPU_WORK_BATCH = 8, 158}; 159#define BLK_MQ_FLAG_TO_ALLOC_POLICY(flags) \ 160 ((flags >> BLK_MQ_F_ALLOC_POLICY_START_BIT) & \ 161 ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1)) 162#define BLK_ALLOC_POLICY_TO_MQ_FLAG(policy) \ 163 ((policy & ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1)) \ 164 << BLK_MQ_F_ALLOC_POLICY_START_BIT) 165 166struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *); 167struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, 168 struct request_queue *q); 169void blk_mq_finish_init(struct request_queue *q); 170int blk_mq_register_disk(struct gendisk *); 171void blk_mq_unregister_disk(struct gendisk *); 172 173int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set); 174void blk_mq_free_tag_set(struct blk_mq_tag_set *set); 175 176void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule); 177 178void blk_mq_insert_request(struct request *, bool, bool, bool); 179void blk_mq_free_request(struct request *rq); 180void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *, struct request *rq); 181bool blk_mq_can_queue(struct blk_mq_hw_ctx *); 182struct request *blk_mq_alloc_request(struct request_queue *q, int rw, 183 gfp_t gfp, bool reserved); 184struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag); 185 186enum { 187 BLK_MQ_UNIQUE_TAG_BITS = 16, 188 BLK_MQ_UNIQUE_TAG_MASK = (1 << BLK_MQ_UNIQUE_TAG_BITS) - 1, 189}; 190 191u32 blk_mq_unique_tag(struct request *rq); 192 193static inline u16 blk_mq_unique_tag_to_hwq(u32 unique_tag) 194{ 195 return unique_tag >> BLK_MQ_UNIQUE_TAG_BITS; 196} 197 198static inline u16 blk_mq_unique_tag_to_tag(u32 unique_tag) 199{ 200 return unique_tag & BLK_MQ_UNIQUE_TAG_MASK; 201} 202 203struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *, const int ctx_index); 204struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *, unsigned int, int); 205 206int blk_mq_request_started(struct request *rq); 207void blk_mq_start_request(struct request *rq); 208void blk_mq_end_request(struct request *rq, int error); 209void __blk_mq_end_request(struct request *rq, int error); 210 211void blk_mq_requeue_request(struct request *rq); 212void blk_mq_add_to_requeue_list(struct request *rq, bool at_head); 213void blk_mq_cancel_requeue_work(struct request_queue *q); 214void blk_mq_kick_requeue_list(struct request_queue *q); 215void blk_mq_abort_requeue_list(struct request_queue *q); 216void blk_mq_complete_request(struct request *rq); 217 218void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx); 219void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx); 220void blk_mq_stop_hw_queues(struct request_queue *q); 221void blk_mq_start_hw_queues(struct request_queue *q); 222void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async); 223void blk_mq_run_hw_queues(struct request_queue *q, bool async); 224void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs); 225void blk_mq_tag_busy_iter(struct blk_mq_hw_ctx *hctx, busy_iter_fn *fn, 226 void *priv); 227void blk_mq_freeze_queue(struct request_queue *q); 228void blk_mq_unfreeze_queue(struct request_queue *q); 229void blk_mq_freeze_queue_start(struct request_queue *q); 230 231/* 232 * Driver command data is immediately after the request. So subtract request 233 * size to get back to the original request, add request size to get the PDU. 234 */ 235static inline struct request *blk_mq_rq_from_pdu(void *pdu) 236{ 237 return pdu - sizeof(struct request); 238} 239static inline void *blk_mq_rq_to_pdu(struct request *rq) 240{ 241 return rq + 1; 242} 243 244#define queue_for_each_hw_ctx(q, hctx, i) \ 245 for ((i) = 0; (i) < (q)->nr_hw_queues && \ 246 ({ hctx = (q)->queue_hw_ctx[i]; 1; }); (i)++) 247 248#define queue_for_each_ctx(q, ctx, i) \ 249 for ((i) = 0; (i) < (q)->nr_queues && \ 250 ({ ctx = per_cpu_ptr((q)->queue_ctx, (i)); 1; }); (i)++) 251 252#define hctx_for_each_ctx(hctx, ctx, i) \ 253 for ((i) = 0; (i) < (hctx)->nr_ctx && \ 254 ({ ctx = (hctx)->ctxs[(i)]; 1; }); (i)++) 255 256#define blk_ctx_sum(q, sum) \ 257({ \ 258 struct blk_mq_ctx *__x; \ 259 unsigned int __ret = 0, __i; \ 260 \ 261 queue_for_each_ctx((q), __x, __i) \ 262 __ret += sum; \ 263 __ret; \ 264}) 265 266#endif 267