iocontext.h 5.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165
  1. #ifndef IOCONTEXT_H
  2. #define IOCONTEXT_H
  3. #include <linux/radix-tree.h>
  4. #include <linux/rcupdate.h>
  5. #include <linux/workqueue.h>
  6. enum {
  7. ICQ_IOPRIO_CHANGED = 1 << 0,
  8. ICQ_CGROUP_CHANGED = 1 << 1,
  9. ICQ_EXITED = 1 << 2,
  10. ICQ_CHANGED_MASK = ICQ_IOPRIO_CHANGED | ICQ_CGROUP_CHANGED,
  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. * q_node and ioc_node link io_cq through icq_list of q and ioc
  76. * respectively. Both fields are unused once ioc_exit_icq() is
  77. * called and shared with __rcu_icq_cache and __rcu_head which are
  78. * used for RCU free of io_cq.
  79. */
  80. union {
  81. struct list_head q_node;
  82. struct kmem_cache *__rcu_icq_cache;
  83. };
  84. union {
  85. struct hlist_node ioc_node;
  86. struct rcu_head __rcu_head;
  87. };
  88. unsigned int flags;
  89. };
  90. /*
  91. * I/O subsystem state of the associated processes. It is refcounted
  92. * and kmalloc'ed. These could be shared between processes.
  93. */
  94. struct io_context {
  95. atomic_long_t refcount;
  96. atomic_t active_ref;
  97. atomic_t nr_tasks;
  98. /* all the fields below are protected by this lock */
  99. spinlock_t lock;
  100. unsigned short ioprio;
  101. /*
  102. * For request batching
  103. */
  104. int nr_batch_requests; /* Number of requests left in the batch */
  105. unsigned long last_waited; /* Time last woken after wait for request */
  106. struct radix_tree_root icq_tree;
  107. struct io_cq __rcu *icq_hint;
  108. struct hlist_head icq_list;
  109. struct work_struct release_work;
  110. };
  111. /**
  112. * get_io_context_active - get active reference on ioc
  113. * @ioc: ioc of interest
  114. *
  115. * Only iocs with active reference can issue new IOs. This function
  116. * acquires an active reference on @ioc. The caller must already have an
  117. * active reference on @ioc.
  118. */
  119. static inline void get_io_context_active(struct io_context *ioc)
  120. {
  121. WARN_ON_ONCE(atomic_long_read(&ioc->refcount) <= 0);
  122. WARN_ON_ONCE(atomic_read(&ioc->active_ref) <= 0);
  123. atomic_long_inc(&ioc->refcount);
  124. atomic_inc(&ioc->active_ref);
  125. }
  126. static inline void ioc_task_link(struct io_context *ioc)
  127. {
  128. get_io_context_active(ioc);
  129. WARN_ON_ONCE(atomic_read(&ioc->nr_tasks) <= 0);
  130. atomic_inc(&ioc->nr_tasks);
  131. }
  132. struct task_struct;
  133. #ifdef CONFIG_BLOCK
  134. void put_io_context(struct io_context *ioc);
  135. void put_io_context_active(struct io_context *ioc);
  136. void exit_io_context(struct task_struct *task);
  137. struct io_context *get_task_io_context(struct task_struct *task,
  138. gfp_t gfp_flags, int node);
  139. void ioc_ioprio_changed(struct io_context *ioc, int ioprio);
  140. void ioc_cgroup_changed(struct io_context *ioc);
  141. unsigned int icq_get_changed(struct io_cq *icq);
  142. #else
  143. struct io_context;
  144. static inline void put_io_context(struct io_context *ioc) { }
  145. static inline void exit_io_context(struct task_struct *task) { }
  146. #endif
  147. #endif