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