include/linux/iocontext.h

   1 #ifndef IOCONTEXT_H
   2 #define IOCONTEXT_H
   3
   4 #include <linux/radix-tree.h>
   5 #include <linux/rcupdate.h>
   6 #include <linux/workqueue.h>
   7
   8 enum {
   9         ICQ_EXITED              = 1 << 2,
  10 };
  11
  12 /*
  13  * An io_cq (icq) is association between an io_context (ioc) and a
  14  * request_queue (q).  This is used by elevators which need to track
  15  * information per ioc - q pair.
  16  *
  17  * Elevator can request use of icq by setting elevator_type->icq_size and
  18  * ->icq_align.  Both size and align must be larger than that of struct
  19  * io_cq and elevator can use the tail area for private information.  The
  20  * recommended way to do this is defining a struct which contains io_cq as
  21  * the first member followed by private members and using its size and
  22  * align.  For example,
  23  *
  24  *      struct snail_io_cq {
  25  *              struct io_cq    icq;
  26  *              int             poke_snail;
  27  *              int             feed_snail;
  28  *      };
  29  *
  30  *      struct elevator_type snail_elv_type {
  31  *              .ops =          { ... },
  32  *              .icq_size =     sizeof(struct snail_io_cq),
  33  *              .icq_align =    __alignof__(struct snail_io_cq),
  34  *              ...
  35  *      };
  36  *
  37  * If icq_size is set, block core will manage icq's.  All requests will
  38  * have its ->elv.icq field set before elevator_ops->elevator_set_req_fn()
  39  * is called and be holding a reference to the associated io_context.
  40  *
  41  * Whenever a new icq is created, elevator_ops->elevator_init_icq_fn() is
  42  * called and, on destruction, ->elevator_exit_icq_fn().  Both functions
  43  * are called with both the associated io_context and queue locks held.
  44  *
  45  * Elevator is allowed to lookup icq using ioc_lookup_icq() while holding
  46  * queue lock but the returned icq is valid only until the queue lock is
  47  * released.  Elevators can not and should not try to create or destroy
  48  * icq's.
  49  *
  50  * As icq's are linked from both ioc and q, the locking rules are a bit
  51  * complex.
  52  *
  53  * - ioc lock nests inside q lock.
  54  *
  55  * - ioc->icq_list and icq->ioc_node are protected by ioc lock.
  56  *   q->icq_list and icq->q_node by q lock.
  57  *
  58  * - ioc->icq_tree and ioc->icq_hint are protected by ioc lock, while icq
  59  *   itself is protected by q lock.  However, both the indexes and icq
  60  *   itself are also RCU managed and lookup can be performed holding only
  61  *   the q lock.
  62  *
  63  * - icq's are not reference counted.  They are destroyed when either the
  64  *   ioc or q goes away.  Each request with icq set holds an extra
  65  *   reference to ioc to ensure it stays until the request is completed.
  66  *
  67  * - Linking and unlinking icq's are performed while holding both ioc and q
  68  *   locks.  Due to the lock ordering, q exit is simple but ioc exit
  69  *   requires reverse-order double lock dance.
  70  */
  71 struct io_cq {
  72         struct request_queue    *q;
  73         struct io_context       *ioc;
  74
  75         /*
  76          * q_node and ioc_node link io_cq through icq_list of q and ioc
  77          * respectively.  Both fields are unused once ioc_exit_icq() is
  78          * called and shared with __rcu_icq_cache and __rcu_head which are
  79          * used for RCU free of io_cq.
  80          */
  81         union {
  82                 struct list_head        q_node;
  83                 struct kmem_cache       *__rcu_icq_cache;
  84         };
  85         union {
  86                 struct hlist_node       ioc_node;
  87                 struct rcu_head         __rcu_head;
  88         };
  89
  90         unsigned int            flags;
  91 };
  92
  93 /*
  94  * I/O subsystem state of the associated processes.  It is refcounted
  95  * and kmalloc'ed. These could be shared between processes.
  96  */
  97 struct io_context {
  98         atomic_long_t refcount;
  99         atomic_t active_ref;
 100         atomic_t nr_tasks;
 101
 102         /* all the fields below are protected by this lock */
 103         spinlock_t lock;
 104
 105         unsigned short ioprio;
 106
 107         /*
 108          * For request batching
 109          */
 110         int nr_batch_requests;     /* Number of requests left in the batch */
 111         unsigned long last_waited; /* Time last woken after wait for request */
 112
 113         struct radix_tree_root  icq_tree;
 114         struct io_cq __rcu      *icq_hint;
 115         struct hlist_head       icq_list;
 116
 117         struct work_struct release_work;
 118 };
 119
 120 /**
 121  * get_io_context_active - get active reference on ioc
 122  * @ioc: ioc of interest
 123  *
 124  * Only iocs with active reference can issue new IOs.  This function
 125  * acquires an active reference on @ioc.  The caller must already have an
 126  * active reference on @ioc.
 127  */
 128 static inline void get_io_context_active(struct io_context *ioc)
 129 {
 130         WARN_ON_ONCE(atomic_long_read(&ioc->refcount) <= 0);
 131         WARN_ON_ONCE(atomic_read(&ioc->active_ref) <= 0);
 132         atomic_long_inc(&ioc->refcount);
 133         atomic_inc(&ioc->active_ref);
 134 }
 135
 136 static inline void ioc_task_link(struct io_context *ioc)
 137 {
 138         get_io_context_active(ioc);
 139
 140         WARN_ON_ONCE(atomic_read(&ioc->nr_tasks) <= 0);
 141         atomic_inc(&ioc->nr_tasks);
 142 }
 143
 144 struct task_struct;
 145 #ifdef CONFIG_BLOCK
 146 void put_io_context(struct io_context *ioc);
 147 void put_io_context_active(struct io_context *ioc);
 148 void exit_io_context(struct task_struct *task);
 149 struct io_context *get_task_io_context(struct task_struct *task,
 150                                        gfp_t gfp_flags, int node);
 151 #else
 152 struct io_context;
 153 static inline void put_io_context(struct io_context *ioc) { }
 154 static inline void exit_io_context(struct task_struct *task) { }
 155 #endif
 156
 157 #endif