include/linux/blk-mq.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef BLK_MQ_H
   3 #define BLK_MQ_H
   4
   5 #include <linux/blkdev.h>
   6 #include <linux/sbitmap.h>
   7 #include <linux/srcu.h>
   8
   9 struct blk_mq_tags;
  10 struct blk_flush_queue;
  11
  12 struct blk_mq_hw_ctx {
  13         struct {
  14                 spinlock_t              lock;
  15                 struct list_head        dispatch;
  16                 unsigned long           state;          /* BLK_MQ_S_* flags */
  17         } ____cacheline_aligned_in_smp;
  18
  19         struct delayed_work     run_work;
  20         cpumask_var_t           cpumask;
  21         int                     next_cpu;
  22         int                     next_cpu_batch;
  23
  24         unsigned long           flags;          /* BLK_MQ_F_* flags */
  25
  26         void                    *sched_data;
  27         struct request_queue    *queue;
  28         struct blk_flush_queue  *fq;
  29
  30         void                    *driver_data;
  31
  32         struct sbitmap          ctx_map;
  33
  34         struct blk_mq_ctx       *dispatch_from;
  35         unsigned int            dispatch_busy;
  36
  37         unsigned int            nr_ctx;
  38         struct blk_mq_ctx       **ctxs;
  39
  40         wait_queue_entry_t      dispatch_wait;
  41         atomic_t                wait_index;
  42
  43         struct blk_mq_tags      *tags;
  44         struct blk_mq_tags      *sched_tags;
  45
  46         unsigned long           queued;
  47         unsigned long           run;
  48 #define BLK_MQ_MAX_DISPATCH_ORDER       7
  49         unsigned long           dispatched[BLK_MQ_MAX_DISPATCH_ORDER];
  50
  51         unsigned int            numa_node;
  52         unsigned int            queue_num;
  53
  54         atomic_t                nr_active;
  55
  56         struct hlist_node       cpuhp_dead;
  57         struct kobject          kobj;
  58
  59         unsigned long           poll_considered;
  60         unsigned long           poll_invoked;
  61         unsigned long           poll_success;
  62
  63 #ifdef CONFIG_BLK_DEBUG_FS
  64         struct dentry           *debugfs_dir;
  65         struct dentry           *sched_debugfs_dir;
  66 #endif
  67
  68         /* Must be the last member - see also blk_mq_hw_ctx_size(). */
  69         struct srcu_struct      queue_rq_srcu[0];
  70 };
  71
  72 struct blk_mq_tag_set {
  73         unsigned int            *mq_map;
  74         const struct blk_mq_ops *ops;
  75         unsigned int            nr_hw_queues;
  76         unsigned int            queue_depth;    /* max hw supported */
  77         unsigned int            reserved_tags;
  78         unsigned int            cmd_size;       /* per-request extra data */
  79         int                     numa_node;
  80         unsigned int            timeout;
  81         unsigned int            flags;          /* BLK_MQ_F_* */
  82         void                    *driver_data;
  83
  84         struct blk_mq_tags      **tags;
  85
  86         struct mutex            tag_list_lock;
  87         struct list_head        tag_list;
  88 };
  89
  90 struct blk_mq_queue_data {
  91         struct request *rq;
  92         bool last;
  93 };
  94
  95 typedef blk_status_t (queue_rq_fn)(struct blk_mq_hw_ctx *,
  96                 const struct blk_mq_queue_data *);
  97 typedef bool (get_budget_fn)(struct blk_mq_hw_ctx *);
  98 typedef void (put_budget_fn)(struct blk_mq_hw_ctx *);
  99 typedef enum blk_eh_timer_return (timeout_fn)(struct request *, bool);
 100 typedef int (init_hctx_fn)(struct blk_mq_hw_ctx *, void *, unsigned int);
 101 typedef void (exit_hctx_fn)(struct blk_mq_hw_ctx *, unsigned int);
 102 typedef int (init_request_fn)(struct blk_mq_tag_set *set, struct request *,
 103                 unsigned int, unsigned int);
 104 typedef void (exit_request_fn)(struct blk_mq_tag_set *set, struct request *,
 105                 unsigned int);
 106
 107 typedef void (busy_iter_fn)(struct blk_mq_hw_ctx *, struct request *, void *,
 108                 bool);
 109 typedef void (busy_tag_iter_fn)(struct request *, void *, bool);
 110 typedef int (poll_fn)(struct blk_mq_hw_ctx *, unsigned int);
 111 typedef int (map_queues_fn)(struct blk_mq_tag_set *set);
 112
 113
 114 struct blk_mq_ops {
 115         /*
 116          * Queue request
 117          */
 118         queue_rq_fn             *queue_rq;
 119
 120         /*
 121          * Reserve budget before queue request, once .queue_rq is
 122          * run, it is driver's responsibility to release the
 123          * reserved budget. Also we have to handle failure case
 124          * of .get_budget for avoiding I/O deadlock.
 125          */
 126         get_budget_fn           *get_budget;
 127         put_budget_fn           *put_budget;
 128
 129         /*
 130          * Called on request timeout
 131          */
 132         timeout_fn              *timeout;
 133
 134         /*
 135          * Called to poll for completion of a specific tag.
 136          */
 137         poll_fn                 *poll;
 138
 139         softirq_done_fn         *complete;
 140
 141         /*
 142          * Called when the block layer side of a hardware queue has been
 143          * set up, allowing the driver to allocate/init matching structures.
 144          * Ditto for exit/teardown.
 145          */
 146         init_hctx_fn            *init_hctx;
 147         exit_hctx_fn            *exit_hctx;
 148
 149         /*
 150          * Called for every command allocated by the block layer to allow
 151          * the driver to set up driver specific data.
 152          *
 153          * Tag greater than or equal to queue_depth is for setting up
 154          * flush request.
 155          *
 156          * Ditto for exit/teardown.
 157          */
 158         init_request_fn         *init_request;
 159         exit_request_fn         *exit_request;
 160         /* Called from inside blk_get_request() */
 161         void (*initialize_rq_fn)(struct request *rq);
 162
 163         map_queues_fn           *map_queues;
 164
 165 #ifdef CONFIG_BLK_DEBUG_FS
 166         /*
 167          * Used by the debugfs implementation to show driver-specific
 168          * information about a request.
 169          */
 170         void (*show_rq)(struct seq_file *m, struct request *rq);
 171 #endif
 172 };
 173
 174 enum {
 175         BLK_MQ_F_SHOULD_MERGE   = 1 << 0,
 176         BLK_MQ_F_TAG_SHARED     = 1 << 1,
 177         BLK_MQ_F_SG_MERGE       = 1 << 2,
 178         BLK_MQ_F_BLOCKING       = 1 << 5,
 179         BLK_MQ_F_NO_SCHED       = 1 << 6,
 180         BLK_MQ_F_ALLOC_POLICY_START_BIT = 8,
 181         BLK_MQ_F_ALLOC_POLICY_BITS = 1,
 182
 183         BLK_MQ_S_STOPPED        = 0,
 184         BLK_MQ_S_TAG_ACTIVE     = 1,
 185         BLK_MQ_S_SCHED_RESTART  = 2,
 186         BLK_MQ_S_START_ON_RUN   = 3,
 187
 188         BLK_MQ_MAX_DEPTH        = 10240,
 189
 190         BLK_MQ_CPU_WORK_BATCH   = 8,
 191 };
 192 #define BLK_MQ_FLAG_TO_ALLOC_POLICY(flags) \
 193         ((flags >> BLK_MQ_F_ALLOC_POLICY_START_BIT) & \
 194                 ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1))
 195 #define BLK_ALLOC_POLICY_TO_MQ_FLAG(policy) \
 196         ((policy & ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1)) \
 197                 << BLK_MQ_F_ALLOC_POLICY_START_BIT)
 198
 199 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *);
 200 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
 201                                                   struct request_queue *q);
 202 int blk_mq_register_dev(struct device *, struct request_queue *);
 203 void blk_mq_unregister_dev(struct device *, struct request_queue *);
 204
 205 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set);
 206 void blk_mq_free_tag_set(struct blk_mq_tag_set *set);
 207
 208 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule);
 209
 210 void blk_mq_free_request(struct request *rq);
 211 bool blk_mq_can_queue(struct blk_mq_hw_ctx *);
 212
 213 enum {
 214         /* return when out of requests */
 215         BLK_MQ_REQ_NOWAIT       = (__force blk_mq_req_flags_t)(1 << 0),
 216         /* allocate from reserved pool */
 217         BLK_MQ_REQ_RESERVED     = (__force blk_mq_req_flags_t)(1 << 1),
 218         /* allocate internal/sched tag */
 219         BLK_MQ_REQ_INTERNAL     = (__force blk_mq_req_flags_t)(1 << 2),
 220         /* set RQF_PREEMPT */
 221         BLK_MQ_REQ_PREEMPT      = (__force blk_mq_req_flags_t)(1 << 3),
 222 };
 223
 224 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 225                 blk_mq_req_flags_t flags);
 226 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 227                 unsigned int op, blk_mq_req_flags_t flags,
 228                 unsigned int hctx_idx);
 229 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag);
 230
 231 enum {
 232         BLK_MQ_UNIQUE_TAG_BITS = 16,
 233         BLK_MQ_UNIQUE_TAG_MASK = (1 << BLK_MQ_UNIQUE_TAG_BITS) - 1,
 234 };
 235
 236 u32 blk_mq_unique_tag(struct request *rq);
 237
 238 static inline u16 blk_mq_unique_tag_to_hwq(u32 unique_tag)
 239 {
 240         return unique_tag >> BLK_MQ_UNIQUE_TAG_BITS;
 241 }
 242
 243 static inline u16 blk_mq_unique_tag_to_tag(u32 unique_tag)
 244 {
 245         return unique_tag & BLK_MQ_UNIQUE_TAG_MASK;
 246 }
 247
 248
 249 int blk_mq_request_started(struct request *rq);
 250 void blk_mq_start_request(struct request *rq);
 251 void blk_mq_end_request(struct request *rq, blk_status_t error);
 252 void __blk_mq_end_request(struct request *rq, blk_status_t error);
 253
 254 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list);
 255 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 256                                 bool kick_requeue_list);
 257 void blk_mq_kick_requeue_list(struct request_queue *q);
 258 void blk_mq_delay_kick_requeue_list(struct request_queue *q, unsigned long msecs);
 259 void blk_mq_complete_request(struct request *rq);
 260
 261 bool blk_mq_queue_stopped(struct request_queue *q);
 262 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx);
 263 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx);
 264 void blk_mq_stop_hw_queues(struct request_queue *q);
 265 void blk_mq_start_hw_queues(struct request_queue *q);
 266 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);
 267 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async);
 268 void blk_mq_quiesce_queue(struct request_queue *q);
 269 void blk_mq_unquiesce_queue(struct request_queue *q);
 270 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs);
 271 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);
 272 void blk_mq_run_hw_queues(struct request_queue *q, bool async);
 273 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs);
 274 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
 275                 busy_tag_iter_fn *fn, void *priv);
 276 void blk_mq_freeze_queue(struct request_queue *q);
 277 void blk_mq_unfreeze_queue(struct request_queue *q);
 278 void blk_freeze_queue_start(struct request_queue *q);
 279 void blk_mq_freeze_queue_wait(struct request_queue *q);
 280 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 281                                      unsigned long timeout);
 282 int blk_mq_tagset_iter(struct blk_mq_tag_set *set, void *data,
 283                 int (reinit_request)(void *, struct request *));
 284
 285 int blk_mq_map_queues(struct blk_mq_tag_set *set);
 286 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues);
 287
 288 void blk_mq_quiesce_queue_nowait(struct request_queue *q);
 289
 290 /*
 291  * Driver command data is immediately after the request. So subtract request
 292  * size to get back to the original request, add request size to get the PDU.
 293  */
 294 static inline struct request *blk_mq_rq_from_pdu(void *pdu)
 295 {
 296         return pdu - sizeof(struct request);
 297 }
 298 static inline void *blk_mq_rq_to_pdu(struct request *rq)
 299 {
 300         return rq + 1;
 301 }
 302
 303 #define queue_for_each_hw_ctx(q, hctx, i)                               \
 304         for ((i) = 0; (i) < (q)->nr_hw_queues &&                        \
 305              ({ hctx = (q)->queue_hw_ctx[i]; 1; }); (i)++)
 306
 307 #define hctx_for_each_ctx(hctx, ctx, i)                                 \
 308         for ((i) = 0; (i) < (hctx)->nr_ctx &&                           \
 309              ({ ctx = (hctx)->ctxs[(i)]; 1; }); (i)++)
 310
 311 #endif