include/linux/blk-mq.h

   1 #ifndef BLK_MQ_H
   2 #define BLK_MQ_H
   3
   4 #include <linux/blkdev.h>
   5 #include <linux/sbitmap.h>
   6 #include <linux/srcu.h>
   7
   8 struct blk_mq_tags;
   9 struct blk_flush_queue;
  10
  11 struct blk_mq_hw_ctx {
  12         struct {
  13                 spinlock_t              lock;
  14                 struct list_head        dispatch;
  15                 unsigned long           state;          /* BLK_MQ_S_* flags */
  16         } ____cacheline_aligned_in_smp;
  17
  18         struct delayed_work     run_work;
  19         cpumask_var_t           cpumask;
  20         int                     next_cpu;
  21         int                     next_cpu_batch;
  22
  23         unsigned long           flags;          /* BLK_MQ_F_* flags */
  24
  25         void                    *sched_data;
  26         struct request_queue    *queue;
  27         struct blk_flush_queue  *fq;
  28
  29         void                    *driver_data;
  30
  31         struct sbitmap          ctx_map;
  32
  33         struct blk_mq_ctx       **ctxs;
  34         unsigned int            nr_ctx;
  35
  36         wait_queue_entry_t              dispatch_wait;
  37         atomic_t                wait_index;
  38
  39         struct blk_mq_tags      *tags;
  40         struct blk_mq_tags      *sched_tags;
  41
  42         unsigned long           queued;
  43         unsigned long           run;
  44 #define BLK_MQ_MAX_DISPATCH_ORDER       7
  45         unsigned long           dispatched[BLK_MQ_MAX_DISPATCH_ORDER];
  46
  47         unsigned int            numa_node;
  48         unsigned int            queue_num;
  49
  50         atomic_t                nr_active;
  51
  52         struct hlist_node       cpuhp_dead;
  53         struct kobject          kobj;
  54
  55         unsigned long           poll_considered;
  56         unsigned long           poll_invoked;
  57         unsigned long           poll_success;
  58
  59 #ifdef CONFIG_BLK_DEBUG_FS
  60         struct dentry           *debugfs_dir;
  61         struct dentry           *sched_debugfs_dir;
  62 #endif
  63
  64         /* Must be the last member - see also blk_mq_hw_ctx_size(). */
  65         struct srcu_struct      queue_rq_srcu[0];
  66 };
  67
  68 struct blk_mq_tag_set {
  69         unsigned int            *mq_map;
  70         const struct blk_mq_ops *ops;
  71         unsigned int            nr_hw_queues;
  72         unsigned int            queue_depth;    /* max hw supported */
  73         unsigned int            reserved_tags;
  74         unsigned int            cmd_size;       /* per-request extra data */
  75         int                     numa_node;
  76         unsigned int            timeout;
  77         unsigned int            flags;          /* BLK_MQ_F_* */
  78         void                    *driver_data;
  79
  80         struct blk_mq_tags      **tags;
  81
  82         struct mutex            tag_list_lock;
  83         struct list_head        tag_list;
  84 };
  85
  86 struct blk_mq_queue_data {
  87         struct request *rq;
  88         bool last;
  89 };
  90
  91 typedef blk_status_t (queue_rq_fn)(struct blk_mq_hw_ctx *,
  92                 const struct blk_mq_queue_data *);
  93 typedef blk_status_t (get_budget_fn)(struct blk_mq_hw_ctx *);
  94 typedef void (put_budget_fn)(struct blk_mq_hw_ctx *);
  95 typedef enum blk_eh_timer_return (timeout_fn)(struct request *, bool);
  96 typedef int (init_hctx_fn)(struct blk_mq_hw_ctx *, void *, unsigned int);
  97 typedef void (exit_hctx_fn)(struct blk_mq_hw_ctx *, unsigned int);
  98 typedef int (init_request_fn)(struct blk_mq_tag_set *set, struct request *,
  99                 unsigned int, unsigned int);
 100 typedef void (exit_request_fn)(struct blk_mq_tag_set *set, struct request *,
 101                 unsigned int);
 102
 103 typedef void (busy_iter_fn)(struct blk_mq_hw_ctx *, struct request *, void *,
 104                 bool);
 105 typedef void (busy_tag_iter_fn)(struct request *, void *, bool);
 106 typedef int (poll_fn)(struct blk_mq_hw_ctx *, unsigned int);
 107 typedef int (map_queues_fn)(struct blk_mq_tag_set *set);
 108
 109
 110 struct blk_mq_ops {
 111         /*
 112          * Queue request
 113          */
 114         queue_rq_fn             *queue_rq;
 115
 116         /*
 117          * Reserve budget before queue request, once .queue_rq is
 118          * run, it is driver's responsibility to release the
 119          * reserved budget. Also we have to handle failure case
 120          * of .get_budget for avoiding I/O deadlock.
 121          */
 122         get_budget_fn           *get_budget;
 123         put_budget_fn           *put_budget;
 124
 125         /*
 126          * Called on request timeout
 127          */
 128         timeout_fn              *timeout;
 129
 130         /*
 131          * Called to poll for completion of a specific tag.
 132          */
 133         poll_fn                 *poll;
 134
 135         softirq_done_fn         *complete;
 136
 137         /*
 138          * Called when the block layer side of a hardware queue has been
 139          * set up, allowing the driver to allocate/init matching structures.
 140          * Ditto for exit/teardown.
 141          */
 142         init_hctx_fn            *init_hctx;
 143         exit_hctx_fn            *exit_hctx;
 144
 145         /*
 146          * Called for every command allocated by the block layer to allow
 147          * the driver to set up driver specific data.
 148          *
 149          * Tag greater than or equal to queue_depth is for setting up
 150          * flush request.
 151          *
 152          * Ditto for exit/teardown.
 153          */
 154         init_request_fn         *init_request;
 155         exit_request_fn         *exit_request;
 156         /* Called from inside blk_get_request() */
 157         void (*initialize_rq_fn)(struct request *rq);
 158
 159         map_queues_fn           *map_queues;
 160
 161 #ifdef CONFIG_BLK_DEBUG_FS
 162         /*
 163          * Used by the debugfs implementation to show driver-specific
 164          * information about a request.
 165          */
 166         void (*show_rq)(struct seq_file *m, struct request *rq);
 167 #endif
 168 };
 169
 170 enum {
 171         BLK_MQ_F_SHOULD_MERGE   = 1 << 0,
 172         BLK_MQ_F_TAG_SHARED     = 1 << 1,
 173         BLK_MQ_F_SG_MERGE       = 1 << 2,
 174         BLK_MQ_F_BLOCKING       = 1 << 5,
 175         BLK_MQ_F_NO_SCHED       = 1 << 6,
 176         BLK_MQ_F_ALLOC_POLICY_START_BIT = 8,
 177         BLK_MQ_F_ALLOC_POLICY_BITS = 1,
 178
 179         BLK_MQ_S_STOPPED        = 0,
 180         BLK_MQ_S_TAG_ACTIVE     = 1,
 181         BLK_MQ_S_SCHED_RESTART  = 2,
 182         BLK_MQ_S_TAG_WAITING    = 3,
 183         BLK_MQ_S_START_ON_RUN   = 4,
 184
 185         BLK_MQ_MAX_DEPTH        = 10240,
 186
 187         BLK_MQ_CPU_WORK_BATCH   = 8,
 188 };
 189 #define BLK_MQ_FLAG_TO_ALLOC_POLICY(flags) \
 190         ((flags >> BLK_MQ_F_ALLOC_POLICY_START_BIT) & \
 191                 ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1))
 192 #define BLK_ALLOC_POLICY_TO_MQ_FLAG(policy) \
 193         ((policy & ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1)) \
 194                 << BLK_MQ_F_ALLOC_POLICY_START_BIT)
 195
 196 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *);
 197 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
 198                                                   struct request_queue *q);
 199 int blk_mq_register_dev(struct device *, struct request_queue *);
 200 void blk_mq_unregister_dev(struct device *, struct request_queue *);
 201
 202 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set);
 203 void blk_mq_free_tag_set(struct blk_mq_tag_set *set);
 204
 205 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule);
 206
 207 void blk_mq_free_request(struct request *rq);
 208 bool blk_mq_can_queue(struct blk_mq_hw_ctx *);
 209
 210 enum {
 211         BLK_MQ_REQ_NOWAIT       = (1 << 0), /* return when out of requests */
 212         BLK_MQ_REQ_RESERVED     = (1 << 1), /* allocate from reserved pool */
 213         BLK_MQ_REQ_INTERNAL     = (1 << 2), /* allocate internal/sched tag */
 214 };
 215
 216 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 217                 unsigned int flags);
 218 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 219                 unsigned int op, unsigned int flags, unsigned int hctx_idx);
 220 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag);
 221
 222 enum {
 223         BLK_MQ_UNIQUE_TAG_BITS = 16,
 224         BLK_MQ_UNIQUE_TAG_MASK = (1 << BLK_MQ_UNIQUE_TAG_BITS) - 1,
 225 };
 226
 227 u32 blk_mq_unique_tag(struct request *rq);
 228
 229 static inline u16 blk_mq_unique_tag_to_hwq(u32 unique_tag)
 230 {
 231         return unique_tag >> BLK_MQ_UNIQUE_TAG_BITS;
 232 }
 233
 234 static inline u16 blk_mq_unique_tag_to_tag(u32 unique_tag)
 235 {
 236         return unique_tag & BLK_MQ_UNIQUE_TAG_MASK;
 237 }
 238
 239
 240 int blk_mq_request_started(struct request *rq);
 241 void blk_mq_start_request(struct request *rq);
 242 void blk_mq_end_request(struct request *rq, blk_status_t error);
 243 void __blk_mq_end_request(struct request *rq, blk_status_t error);
 244
 245 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list);
 246 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 247                                 bool kick_requeue_list);
 248 void blk_mq_kick_requeue_list(struct request_queue *q);
 249 void blk_mq_delay_kick_requeue_list(struct request_queue *q, unsigned long msecs);
 250 void blk_mq_complete_request(struct request *rq);
 251
 252 bool blk_mq_queue_stopped(struct request_queue *q);
 253 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx);
 254 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx);
 255 void blk_mq_stop_hw_queues(struct request_queue *q);
 256 void blk_mq_start_hw_queues(struct request_queue *q);
 257 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);
 258 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async);
 259 void blk_mq_quiesce_queue(struct request_queue *q);
 260 void blk_mq_unquiesce_queue(struct request_queue *q);
 261 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs);
 262 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);
 263 void blk_mq_run_hw_queues(struct request_queue *q, bool async);
 264 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs);
 265 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
 266                 busy_tag_iter_fn *fn, void *priv);
 267 void blk_mq_freeze_queue(struct request_queue *q);
 268 void blk_mq_unfreeze_queue(struct request_queue *q);
 269 void blk_freeze_queue_start(struct request_queue *q);
 270 void blk_mq_freeze_queue_wait(struct request_queue *q);
 271 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 272                                      unsigned long timeout);
 273 int blk_mq_reinit_tagset(struct blk_mq_tag_set *set,
 274                          int (reinit_request)(void *, struct request *));
 275
 276 int blk_mq_map_queues(struct blk_mq_tag_set *set);
 277 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues);
 278
 279 void blk_mq_quiesce_queue_nowait(struct request_queue *q);
 280
 281 /*
 282  * Driver command data is immediately after the request. So subtract request
 283  * size to get back to the original request, add request size to get the PDU.
 284  */
 285 static inline struct request *blk_mq_rq_from_pdu(void *pdu)
 286 {
 287         return pdu - sizeof(struct request);
 288 }
 289 static inline void *blk_mq_rq_to_pdu(struct request *rq)
 290 {
 291         return rq + 1;
 292 }
 293
 294 #define queue_for_each_hw_ctx(q, hctx, i)                               \
 295         for ((i) = 0; (i) < (q)->nr_hw_queues &&                        \
 296              ({ hctx = (q)->queue_hw_ctx[i]; 1; }); (i)++)
 297
 298 #define hctx_for_each_ctx(hctx, ctx, i)                                 \
 299         for ((i) = 0; (i) < (hctx)->nr_ctx &&                           \
 300              ({ ctx = (hctx)->ctxs[(i)]; 1; }); (i)++)
 301
 302 #endif