include/linux/iocontext.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef IOCONTEXT_H
   3 #define IOCONTEXT_H
   4
   5 #include <linux/radix-tree.h>
   6 #include <linux/rcupdate.h>
   7 #include <linux/workqueue.h>
   8
   9 enum {
  10         ICQ_EXITED              = 1 << 2,
  11 };
  12
  13 /*
  14  * An io_cq (icq) is association between an io_context (ioc) and a
  15  * request_queue (q).  This is used by elevators which need to track
  16  * information per ioc - q pair.
  17  *
  18  * Elevator can request use of icq by setting elevator_type->icq_size and
  19  * ->icq_align.  Both size and align must be larger than that of struct
  20  * io_cq and elevator can use the tail area for private information.  The
  21  * recommended way to do this is defining a struct which contains io_cq as
  22  * the first member followed by private members and using its size and
  23  * align.  For example,
  24  *
  25  *      struct snail_io_cq {
  26  *              struct io_cq    icq;
  27  *              int             poke_snail;
  28  *              int             feed_snail;
  29  *      };
  30  *
  31  *      struct elevator_type snail_elv_type {
  32  *              .ops =          { ... },
  33  *              .icq_size =     sizeof(struct snail_io_cq),
  34  *              .icq_align =    __alignof__(struct snail_io_cq),
  35  *              ...
  36  *      };
  37  *
  38  * If icq_size is set, block core will manage icq's.  All requests will
  39  * have its ->elv.icq field set before elevator_ops->elevator_set_req_fn()
  40  * is called and be holding a reference to the associated io_context.
  41  *
  42  * Whenever a new icq is created, elevator_ops->elevator_init_icq_fn() is
  43  * called and, on destruction, ->elevator_exit_icq_fn().  Both functions
  44  * are called with both the associated io_context and queue locks held.
  45  *
  46  * Elevator is allowed to lookup icq using ioc_lookup_icq() while holding
  47  * queue lock but the returned icq is valid only until the queue lock is
  48  * released.  Elevators can not and should not try to create or destroy
  49  * icq's.
  50  *
  51  * As icq's are linked from both ioc and q, the locking rules are a bit
  52  * complex.
  53  *
  54  * - ioc lock nests inside q lock.
  55  *
  56  * - ioc->icq_list and icq->ioc_node are protected by ioc lock.
  57  *   q->icq_list and icq->q_node by q lock.
  58  *
  59  * - ioc->icq_tree and ioc->icq_hint are protected by ioc lock, while icq
  60  *   itself is protected by q lock.  However, both the indexes and icq
  61  *   itself are also RCU managed and lookup can be performed holding only
  62  *   the q lock.
  63  *
  64  * - icq's are not reference counted.  They are destroyed when either the
  65  *   ioc or q goes away.  Each request with icq set holds an extra
  66  *   reference to ioc to ensure it stays until the request is completed.
  67  *
  68  * - Linking and unlinking icq's are performed while holding both ioc and q
  69  *   locks.  Due to the lock ordering, q exit is simple but ioc exit
  70  *   requires reverse-order double lock dance.
  71  */
  72 struct io_cq {
  73         struct request_queue    *q;
  74         struct io_context       *ioc;
  75
  76         /*
  77          * q_node and ioc_node link io_cq through icq_list of q and ioc
  78          * respectively.  Both fields are unused once ioc_exit_icq() is
  79          * called and shared with __rcu_icq_cache and __rcu_head which are
  80          * used for RCU free of io_cq.
  81          */
  82         union {
  83                 struct list_head        q_node;
  84                 struct kmem_cache       *__rcu_icq_cache;
  85         };
  86         union {
  87                 struct hlist_node       ioc_node;
  88                 struct rcu_head         __rcu_head;
  89         };
  90
  91         unsigned int            flags;
  92 };
  93
  94 /*
  95  * I/O subsystem state of the associated processes.  It is refcounted
  96  * and kmalloc'ed. These could be shared between processes.
  97  */
  98 struct io_context {
  99         atomic_long_t refcount;
 100         atomic_t active_ref;
 101         atomic_t nr_tasks;
 102
 103         /* all the fields below are protected by this lock */
 104         spinlock_t lock;
 105
 106         unsigned short ioprio;
 107
 108         /*
 109          * For request batching
 110          */
 111         int nr_batch_requests;     /* Number of requests left in the batch */
 112         unsigned long last_waited; /* Time last woken after wait for request */
 113
 114         struct radix_tree_root  icq_tree;
 115         struct io_cq __rcu      *icq_hint;
 116         struct hlist_head       icq_list;
 117
 118         struct work_struct release_work;
 119 };
 120
 121 /**
 122  * get_io_context_active - get active reference on ioc
 123  * @ioc: ioc of interest
 124  *
 125  * Only iocs with active reference can issue new IOs.  This function
 126  * acquires an active reference on @ioc.  The caller must already have an
 127  * active reference on @ioc.
 128  */
 129 static inline void get_io_context_active(struct io_context *ioc)
 130 {
 131         WARN_ON_ONCE(atomic_long_read(&ioc->refcount) <= 0);
 132         WARN_ON_ONCE(atomic_read(&ioc->active_ref) <= 0);
 133         atomic_long_inc(&ioc->refcount);
 134         atomic_inc(&ioc->active_ref);
 135 }
 136
 137 static inline void ioc_task_link(struct io_context *ioc)
 138 {
 139         get_io_context_active(ioc);
 140
 141         WARN_ON_ONCE(atomic_read(&ioc->nr_tasks) <= 0);
 142         atomic_inc(&ioc->nr_tasks);
 143 }
 144
 145 struct task_struct;
 146 #ifdef CONFIG_BLOCK
 147 void put_io_context(struct io_context *ioc);
 148 void put_io_context_active(struct io_context *ioc);
 149 void exit_io_context(struct task_struct *task);
 150 struct io_context *get_task_io_context(struct task_struct *task,
 151                                        gfp_t gfp_flags, int node);
 152 #else
 153 struct io_context;
 154 static inline void put_io_context(struct io_context *ioc) { }
 155 static inline void exit_io_context(struct task_struct *task) { }
 156 #endif
 157
 158 #endif