1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2016 Trond Myklebust
4 * Copyright (c) 2019 Jeff Layton
5 *
6 * I/O and data path helper functionality.
7 *
8 * Heavily borrowed from equivalent code in fs/nfs/io.c
9 */
10
11#include <linux/ceph/ceph_debug.h>
12
13#include <linux/types.h>
14#include <linux/kernel.h>
15#include <linux/rwsem.h>
16#include <linux/fs.h>
17
18#include "super.h"
19#include "io.h"
20
21/* Call with exclusively locked inode->i_rwsem */
22static void ceph_block_o_direct(struct ceph_inode_info *ci, struct inode *inode)
23{
24	lockdep_assert_held_write(&inode->i_rwsem);
25
26	if (READ_ONCE(ci->i_ceph_flags) & CEPH_I_ODIRECT) {
27		spin_lock(&ci->i_ceph_lock);
28		ci->i_ceph_flags &= ~CEPH_I_ODIRECT;
29		spin_unlock(&ci->i_ceph_lock);
30		inode_dio_wait(inode);
31	}
32}
33
34/**
35 * ceph_start_io_read - declare the file is being used for buffered reads
36 * @inode: file inode
37 *
38 * Declare that a buffered read operation is about to start, and ensure
39 * that we block all direct I/O.
40 * On exit, the function ensures that the CEPH_I_ODIRECT flag is unset,
41 * and holds a shared lock on inode->i_rwsem to ensure that the flag
42 * cannot be changed.
43 * In practice, this means that buffered read operations are allowed to
44 * execute in parallel, thanks to the shared lock, whereas direct I/O
45 * operations need to wait to grab an exclusive lock in order to set
46 * CEPH_I_ODIRECT.
47 * Note that buffered writes and truncates both take a write lock on
48 * inode->i_rwsem, meaning that those are serialised w.r.t. the reads.
49 */
50void
51ceph_start_io_read(struct inode *inode)
52{
53	struct ceph_inode_info *ci = ceph_inode(inode);
54
55	/* Be an optimist! */
56	down_read(&inode->i_rwsem);
57	if (!(READ_ONCE(ci->i_ceph_flags) & CEPH_I_ODIRECT))
58		return;
59	up_read(&inode->i_rwsem);
60	/* Slow path.... */
61	down_write(&inode->i_rwsem);
62	ceph_block_o_direct(ci, inode);
63	downgrade_write(&inode->i_rwsem);
64}
65
66/**
67 * ceph_end_io_read - declare that the buffered read operation is done
68 * @inode: file inode
69 *
70 * Declare that a buffered read operation is done, and release the shared
71 * lock on inode->i_rwsem.
72 */
73void
74ceph_end_io_read(struct inode *inode)
75{
76	up_read(&inode->i_rwsem);
77}
78
79/**
80 * ceph_start_io_write - declare the file is being used for buffered writes
81 * @inode: file inode
82 *
83 * Declare that a buffered write operation is about to start, and ensure
84 * that we block all direct I/O.
85 */
86void
87ceph_start_io_write(struct inode *inode)
88{
89	down_write(&inode->i_rwsem);
90	ceph_block_o_direct(ceph_inode(inode), inode);
91}
92
93/**
94 * ceph_end_io_write - declare that the buffered write operation is done
95 * @inode: file inode
96 *
97 * Declare that a buffered write operation is done, and release the
98 * lock on inode->i_rwsem.
99 */
100void
101ceph_end_io_write(struct inode *inode)
102{
103	up_write(&inode->i_rwsem);
104}
105
106/* Call with exclusively locked inode->i_rwsem */
107static void ceph_block_buffered(struct ceph_inode_info *ci, struct inode *inode)
108{
109	lockdep_assert_held_write(&inode->i_rwsem);
110
111	if (!(READ_ONCE(ci->i_ceph_flags) & CEPH_I_ODIRECT)) {
112		spin_lock(&ci->i_ceph_lock);
113		ci->i_ceph_flags |= CEPH_I_ODIRECT;
114		spin_unlock(&ci->i_ceph_lock);
115		/* FIXME: unmap_mapping_range? */
116		filemap_write_and_wait(inode->i_mapping);
117	}
118}
119
120/**
121 * ceph_start_io_direct - declare the file is being used for direct i/o
122 * @inode: file inode
123 *
124 * Declare that a direct I/O operation is about to start, and ensure
125 * that we block all buffered I/O.
126 * On exit, the function ensures that the CEPH_I_ODIRECT flag is set,
127 * and holds a shared lock on inode->i_rwsem to ensure that the flag
128 * cannot be changed.
129 * In practice, this means that direct I/O operations are allowed to
130 * execute in parallel, thanks to the shared lock, whereas buffered I/O
131 * operations need to wait to grab an exclusive lock in order to clear
132 * CEPH_I_ODIRECT.
133 * Note that buffered writes and truncates both take a write lock on
134 * inode->i_rwsem, meaning that those are serialised w.r.t. O_DIRECT.
135 */
136void
137ceph_start_io_direct(struct inode *inode)
138{
139	struct ceph_inode_info *ci = ceph_inode(inode);
140
141	/* Be an optimist! */
142	down_read(&inode->i_rwsem);
143	if (READ_ONCE(ci->i_ceph_flags) & CEPH_I_ODIRECT)
144		return;
145	up_read(&inode->i_rwsem);
146	/* Slow path.... */
147	down_write(&inode->i_rwsem);
148	ceph_block_buffered(ci, inode);
149	downgrade_write(&inode->i_rwsem);
150}
151
152/**
153 * ceph_end_io_direct - declare that the direct i/o operation is done
154 * @inode: file inode
155 *
156 * Declare that a direct I/O operation is done, and release the shared
157 * lock on inode->i_rwsem.
158 */
159void
160ceph_end_io_direct(struct inode *inode)
161{
162	up_read(&inode->i_rwsem);
163}
164