xref: /linux-6.15/include/linux/iocontext.h (revision d705ae6b)
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 
12 	ICQ_CHANGED_MASK	= ICQ_IOPRIO_CHANGED | ICQ_CGROUP_CHANGED,
13 };
14 
15 /*
16  * An io_cq (icq) is association between an io_context (ioc) and a
17  * request_queue (q).  This is used by elevators which need to track
18  * information per ioc - q pair.
19  *
20  * Elevator can request use of icq by setting elevator_type->icq_size and
21  * ->icq_align.  Both size and align must be larger than that of struct
22  * io_cq and elevator can use the tail area for private information.  The
23  * recommended way to do this is defining a struct which contains io_cq as
24  * the first member followed by private members and using its size and
25  * align.  For example,
26  *
27  *	struct snail_io_cq {
28  *		struct io_cq	icq;
29  *		int		poke_snail;
30  *		int		feed_snail;
31  *	};
32  *
33  *	struct elevator_type snail_elv_type {
34  *		.ops =		{ ... },
35  *		.icq_size =	sizeof(struct snail_io_cq),
36  *		.icq_align =	__alignof__(struct snail_io_cq),
37  *		...
38  *	};
39  *
40  * If icq_size is set, block core will manage icq's.  All requests will
41  * have its ->elv.icq field set before elevator_ops->elevator_set_req_fn()
42  * is called and be holding a reference to the associated io_context.
43  *
44  * Whenever a new icq is created, elevator_ops->elevator_init_icq_fn() is
45  * called and, on destruction, ->elevator_exit_icq_fn().  Both functions
46  * are called with both the associated io_context and queue locks held.
47  *
48  * Elevator is allowed to lookup icq using ioc_lookup_icq() while holding
49  * queue lock but the returned icq is valid only until the queue lock is
50  * released.  Elevators can not and should not try to create or destroy
51  * icq's.
52  *
53  * As icq's are linked from both ioc and q, the locking rules are a bit
54  * complex.
55  *
56  * - ioc lock nests inside q lock.
57  *
58  * - ioc->icq_list and icq->ioc_node are protected by ioc lock.
59  *   q->icq_list and icq->q_node by q lock.
60  *
61  * - ioc->icq_tree and ioc->icq_hint are protected by ioc lock, while icq
62  *   itself is protected by q lock.  However, both the indexes and icq
63  *   itself are also RCU managed and lookup can be performed holding only
64  *   the q lock.
65  *
66  * - icq's are not reference counted.  They are destroyed when either the
67  *   ioc or q goes away.  Each request with icq set holds an extra
68  *   reference to ioc to ensure it stays until the request is completed.
69  *
70  * - Linking and unlinking icq's are performed while holding both ioc and q
71  *   locks.  Due to the lock ordering, q exit is simple but ioc exit
72  *   requires reverse-order double lock dance.
73  */
74 struct io_cq {
75 	struct request_queue	*q;
76 	struct io_context	*ioc;
77 
78 	/*
79 	 * q_node and ioc_node link io_cq through icq_list of q and ioc
80 	 * respectively.  Both fields are unused once ioc_exit_icq() is
81 	 * called and shared with __rcu_icq_cache and __rcu_head which are
82 	 * used for RCU free of io_cq.
83 	 */
84 	union {
85 		struct list_head	q_node;
86 		struct kmem_cache	*__rcu_icq_cache;
87 	};
88 	union {
89 		struct hlist_node	ioc_node;
90 		struct rcu_head		__rcu_head;
91 	};
92 
93 	unsigned int		flags;
94 };
95 
96 /*
97  * I/O subsystem state of the associated processes.  It is refcounted
98  * and kmalloc'ed. These could be shared between processes.
99  */
100 struct io_context {
101 	atomic_long_t refcount;
102 	atomic_t nr_tasks;
103 
104 	/* all the fields below are protected by this lock */
105 	spinlock_t lock;
106 
107 	unsigned short ioprio;
108 
109 	/*
110 	 * For request batching
111 	 */
112 	int nr_batch_requests;     /* Number of requests left in the batch */
113 	unsigned long last_waited; /* Time last woken after wait for request */
114 
115 	struct radix_tree_root	icq_tree;
116 	struct io_cq __rcu	*icq_hint;
117 	struct hlist_head	icq_list;
118 
119 	struct work_struct release_work;
120 };
121 
122 static inline struct io_context *ioc_task_link(struct io_context *ioc)
123 {
124 	/*
125 	 * if ref count is zero, don't allow sharing (ioc is going away, it's
126 	 * a race).
127 	 */
128 	if (ioc && atomic_long_inc_not_zero(&ioc->refcount)) {
129 		atomic_inc(&ioc->nr_tasks);
130 		return ioc;
131 	}
132 
133 	return NULL;
134 }
135 
136 struct task_struct;
137 #ifdef CONFIG_BLOCK
138 void put_io_context(struct io_context *ioc);
139 void exit_io_context(struct task_struct *task);
140 struct io_context *get_task_io_context(struct task_struct *task,
141 				       gfp_t gfp_flags, int node);
142 void ioc_ioprio_changed(struct io_context *ioc, int ioprio);
143 void ioc_cgroup_changed(struct io_context *ioc);
144 unsigned int icq_get_changed(struct io_cq *icq);
145 #else
146 struct io_context;
147 static inline void put_io_context(struct io_context *ioc) { }
148 static inline void exit_io_context(struct task_struct *task) { }
149 #endif
150 
151 #endif
152