projects/linux·blob master · refs

251 lines · 6.4 KiB

1// SPDX-License-Identifier: GPL-2.0-or-later
2/* Direct I/O support.
3 *
4 * Copyright (C) 2023 Red Hat, Inc. All Rights Reserved.
5 * Written by David Howells ([email protected])
6 */
7
8#include <linux/export.h>
9#include <linux/fs.h>
10#include <linux/mm.h>
11#include <linux/pagemap.h>
12#include <linux/slab.h>
13#include <linux/uio.h>
14#include <linux/sched/mm.h>
15#include <linux/task_io_accounting_ops.h>
16#include <linux/netfs.h>
17#include "internal.h"
18
19static void netfs_prepare_dio_read_iterator(struct netfs_io_subrequest *subreq)
20{
21 struct netfs_io_request *rreq = subreq->rreq;
22 size_t rsize;
23
24 rsize = umin(subreq->len, rreq->io_streams[0].sreq_max_len);
25 subreq->len = rsize;
26
27 if (unlikely(rreq->io_streams[0].sreq_max_segs)) {
28 size_t limit = netfs_limit_iter(&rreq->buffer.iter, 0, rsize,
29 rreq->io_streams[0].sreq_max_segs);
30
31 if (limit < rsize) {
32 subreq->len = limit;
33 trace_netfs_sreq(subreq, netfs_sreq_trace_limited);
34 }
35 }
36
37 trace_netfs_sreq(subreq, netfs_sreq_trace_prepare);
38
39 subreq->io_iter = rreq->buffer.iter;
40 iov_iter_truncate(&subreq->io_iter, subreq->len);
41 iov_iter_advance(&rreq->buffer.iter, subreq->len);
42}
43
44/*
45 * Perform a read to a buffer from the server, slicing up the region to be read
46 * according to the network rsize.
47 */
48static void netfs_dispatch_unbuffered_reads(struct netfs_io_request *rreq)
49{
50 ssize_t size = rreq->len;
51 uoff_t start = rreq->start;
52 int ret;
53
54 do {
55 struct netfs_io_subrequest *subreq;
56 ssize_t slice;
57
58 subreq = netfs_alloc_subrequest(rreq, NETFS_DOWNLOAD_FROM_SERVER);
59 if (!subreq) {
60 /* Stash the error in the request if there's not
61 * already an error set.
62 */
63 cmpxchg(&rreq->error, 0, -ENOMEM);
64 break;
65 }
66
67 subreq->start = start;
68 subreq->len = size;
69
70 netfs_queue_read(rreq, subreq);
71
72 netfs_stat(&netfs_n_rh_download);
73 if (rreq->netfs_ops->prepare_read) {
74 ret = rreq->netfs_ops->prepare_read(subreq);
75 if (ret < 0) {
76 netfs_cancel_read(subreq, ret);
77 break;
78 }
79 }
80
81 netfs_prepare_dio_read_iterator(subreq);
82 slice = subreq->len;
83 size -= slice;
84 start += slice;
85 rreq->submitted += slice;
86 if (size <= 0)
87 netfs_all_subreqs_queued(rreq);
88
89 rreq->netfs_ops->issue_read(subreq);
90
91 if (test_bit(NETFS_RREQ_PAUSE, &rreq->flags))
92 netfs_wait_for_paused_read(rreq);
93 if (test_bit(NETFS_RREQ_FAILED, &rreq->flags))
94 break;
95 cond_resched();
96 } while (size > 0);
97
98 if (unlikely(size > 0)) {
99 netfs_all_subreqs_queued(rreq);
100 netfs_wake_collector(rreq);
101 }
102}
103
104/*
105 * Perform a read to an application buffer, bypassing the pagecache and the
106 * local disk cache.
107 */
108static ssize_t netfs_unbuffered_read(struct netfs_io_request *rreq, bool sync)
109{
110 ssize_t ret;
111
112 _enter("R=%x %llx-%llx",
113 rreq->debug_id, rreq->start, rreq->start + rreq->len - 1);
114
115 if (rreq->len == 0) {
116 pr_err("Zero-sized read [R=%x]\n", rreq->debug_id);
117 netfs_put_request(rreq, netfs_rreq_trace_put_discard);
118 return -EIO;
119 }
120
121 // TODO: Use bounce buffer if requested
122
123 inode_dio_begin(rreq->inode);
124 netfs_dispatch_unbuffered_reads(rreq);
125
126 /* The collector will get run, even if we don't manage to submit any
127 * subreqs, so we shouldn't call inode_dio_end() here.
128 */
129
130 if (sync)
131 ret = netfs_wait_for_read(rreq);
132 else
133 ret = -EIOCBQUEUED;
134
135 _leave(" = %zd", ret);
136 return ret;
137}
138
139/**
140 * netfs_unbuffered_read_iter_locked - Perform an unbuffered or direct I/O read
141 * @iocb: The I/O control descriptor describing the read
142 * @iter: The output buffer (also specifies read length)
143 *
144 * Perform an unbuffered I/O or direct I/O from the file in @iocb to the
145 * output buffer. No use is made of the pagecache.
146 *
147 * The caller must hold any appropriate locks.
148 */
149ssize_t netfs_unbuffered_read_iter_locked(struct kiocb *iocb, struct iov_iter *iter)
150{
151 struct netfs_io_request *rreq;
152 ssize_t ret;
153 size_t orig_count = iov_iter_count(iter);
154 bool sync = is_sync_kiocb(iocb);
155
156 _enter("");
157
158 if (!orig_count)
159 return 0; /* Don't update atime */
160
161 ret = kiocb_write_and_wait(iocb, orig_count);
162 if (ret < 0)
163 return ret;
164 file_accessed(iocb->ki_filp);
165
166 rreq = netfs_alloc_request(iocb->ki_filp->f_mapping, iocb->ki_filp,
167 iocb->ki_pos, orig_count,
168 iocb->ki_flags & IOCB_DIRECT ?
169 NETFS_DIO_READ : NETFS_UNBUFFERED_READ);
170 if (IS_ERR(rreq))
171 return PTR_ERR(rreq);
172
173 netfs_stat(&netfs_n_rh_dio_read);
174 trace_netfs_read(rreq, rreq->start, rreq->len, netfs_read_trace_dio_read);
175
176 /* If this is an async op, we have to keep track of the destination
177 * buffer for ourselves as the caller's iterator will be trashed when
178 * we return.
179 *
180 * In such a case, extract an iterator to represent as much of the the
181 * output buffer as we can manage. Note that the extraction might not
182 * be able to allocate a sufficiently large bvec array and may shorten
183 * the request.
184 */
185 if (user_backed_iter(iter)) {
186 ret = netfs_extract_user_iter(iter, rreq->len, &rreq->buffer.iter, 0);
187 if (ret < 0)
188 goto error_put;
189 rreq->direct_bv = (struct bio_vec *)rreq->buffer.iter.bvec;
190 rreq->direct_bv_count = ret;
191 rreq->direct_bv_unpin = iov_iter_extract_will_pin(iter);
192 rreq->len = iov_iter_count(&rreq->buffer.iter);
193 } else {
194 rreq->buffer.iter = *iter;
195 rreq->len = orig_count;
196 rreq->direct_bv_unpin = false;
197 iov_iter_advance(iter, orig_count);
198 }
199
200 // TODO: Set up bounce buffer if needed
201
202 if (!sync) {
203 rreq->iocb = iocb;
204 __set_bit(NETFS_RREQ_OFFLOAD_COLLECTION, &rreq->flags);
205 }
206
207 ret = netfs_unbuffered_read(rreq, sync);
208 if (ret < 0)
209 goto out; /* May be -EIOCBQUEUED */
210 if (sync) {
211 // TODO: Copy from bounce buffer
212 iocb->ki_pos += rreq->transferred;
213 ret = rreq->transferred;
214 }
215
216out:
217 netfs_put_request(rreq, netfs_rreq_trace_put_return);
218 if (ret > 0)
219 orig_count -= ret;
220 return ret;
221
222error_put:
223 netfs_put_failed_request(rreq);
224 return ret;
225}
226EXPORT_SYMBOL(netfs_unbuffered_read_iter_locked);
227
228/**
229 * netfs_unbuffered_read_iter - Perform an unbuffered or direct I/O read
230 * @iocb: The I/O control descriptor describing the read
231 * @iter: The output buffer (also specifies read length)
232 *
233 * Perform an unbuffered I/O or direct I/O from the file in @iocb to the
234 * output buffer. No use is made of the pagecache.
235 */
236ssize_t netfs_unbuffered_read_iter(struct kiocb *iocb, struct iov_iter *iter)
237{
238 struct inode *inode = file_inode(iocb->ki_filp);
239 ssize_t ret;
240
241 if (!iter->count)
242 return 0; /* Don't update atime */
243
244 ret = netfs_start_io_direct(inode);
245 if (ret == 0) {
246 ret = netfs_unbuffered_read_iter_locked(iocb, iter);
247 netfs_end_io_direct(inode);
248 }
249 return ret;
250}
251EXPORT_SYMBOL(netfs_unbuffered_read_iter);