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