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 (dhowells@redhat.com)
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(iter: &rreq->buffer.iter, start_offset: 0, max_size: rsize,
29 max_segs: rreq->io_streams[0].sreq_max_segs);
30
31 if (limit < rsize) {
32 subreq->len = limit;
33 trace_netfs_sreq(sreq: subreq, what: netfs_sreq_trace_limited);
34 }
35 }
36
37 trace_netfs_sreq(sreq: subreq, what: netfs_sreq_trace_prepare);
38
39 subreq->io_iter = rreq->buffer.iter;
40 iov_iter_truncate(i: &subreq->io_iter, count: subreq->len);
41 iov_iter_advance(i: &rreq->buffer.iter, bytes: 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 int netfs_dispatch_unbuffered_reads(struct netfs_io_request *rreq)
49{
50 struct netfs_io_stream *stream = &rreq->io_streams[0];
51 unsigned long long start = rreq->start;
52 ssize_t size = rreq->len;
53 int ret = 0;
54
55 do {
56 struct netfs_io_subrequest *subreq;
57 ssize_t slice;
58
59 subreq = netfs_alloc_subrequest(rreq);
60 if (!subreq) {
61 ret = -ENOMEM;
62 break;
63 }
64
65 subreq->source = NETFS_DOWNLOAD_FROM_SERVER;
66 subreq->start = start;
67 subreq->len = size;
68
69 __set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags);
70
71 spin_lock(lock: &rreq->lock);
72 list_add_tail(new: &subreq->rreq_link, head: &stream->subrequests);
73 if (list_is_first(list: &subreq->rreq_link, head: &stream->subrequests)) {
74 stream->front = subreq;
75 if (!stream->active) {
76 stream->collected_to = stream->front->start;
77 /* Store list pointers before active flag */
78 smp_store_release(&stream->active, true);
79 }
80 }
81 trace_netfs_sreq(sreq: subreq, what: netfs_sreq_trace_added);
82 spin_unlock(lock: &rreq->lock);
83
84 netfs_stat(&netfs_n_rh_download);
85 if (rreq->netfs_ops->prepare_read) {
86 ret = rreq->netfs_ops->prepare_read(subreq);
87 if (ret < 0) {
88 netfs_put_subrequest(subreq, what: netfs_sreq_trace_put_cancel);
89 break;
90 }
91 }
92
93 netfs_prepare_dio_read_iterator(subreq);
94 slice = subreq->len;
95 size -= slice;
96 start += slice;
97 rreq->submitted += slice;
98 if (size <= 0) {
99 smp_wmb(); /* Write lists before ALL_QUEUED. */
100 set_bit(NETFS_RREQ_ALL_QUEUED, addr: &rreq->flags);
101 }
102
103 rreq->netfs_ops->issue_read(subreq);
104
105 if (test_bit(NETFS_RREQ_PAUSE, &rreq->flags))
106 netfs_wait_for_paused_read(rreq);
107 if (test_bit(NETFS_RREQ_FAILED, &rreq->flags))
108 break;
109 cond_resched();
110 } while (size > 0);
111
112 if (unlikely(size > 0)) {
113 smp_wmb(); /* Write lists before ALL_QUEUED. */
114 set_bit(NETFS_RREQ_ALL_QUEUED, addr: &rreq->flags);
115 netfs_wake_collector(rreq);
116 }
117
118 return ret;
119}
120
121/*
122 * Perform a read to an application buffer, bypassing the pagecache and the
123 * local disk cache.
124 */
125static ssize_t netfs_unbuffered_read(struct netfs_io_request *rreq, bool sync)
126{
127 ssize_t ret;
128
129 _enter("R=%x %llx-%llx",
130 rreq->debug_id, rreq->start, rreq->start + rreq->len - 1);
131
132 if (rreq->len == 0) {
133 pr_err("Zero-sized read [R=%x]\n", rreq->debug_id);
134 netfs_put_request(rreq, what: netfs_rreq_trace_put_discard);
135 return -EIO;
136 }
137
138 // TODO: Use bounce buffer if requested
139
140 inode_dio_begin(inode: rreq->inode);
141
142 ret = netfs_dispatch_unbuffered_reads(rreq);
143
144 if (!rreq->submitted) {
145 netfs_put_request(rreq, what: netfs_rreq_trace_put_no_submit);
146 inode_dio_end(inode: rreq->inode);
147 ret = 0;
148 goto out;
149 }
150
151 if (sync)
152 ret = netfs_wait_for_read(rreq);
153 else
154 ret = -EIOCBQUEUED;
155out:
156 _leave(" = %zd", ret);
157 return ret;
158}
159
160/**
161 * netfs_unbuffered_read_iter_locked - Perform an unbuffered or direct I/O read
162 * @iocb: The I/O control descriptor describing the read
163 * @iter: The output buffer (also specifies read length)
164 *
165 * Perform an unbuffered I/O or direct I/O from the file in @iocb to the
166 * output buffer. No use is made of the pagecache.
167 *
168 * The caller must hold any appropriate locks.
169 */
170ssize_t netfs_unbuffered_read_iter_locked(struct kiocb *iocb, struct iov_iter *iter)
171{
172 struct netfs_io_request *rreq;
173 ssize_t ret;
174 size_t orig_count = iov_iter_count(i: iter);
175 bool sync = is_sync_kiocb(kiocb: iocb);
176
177 _enter("");
178
179 if (!orig_count)
180 return 0; /* Don't update atime */
181
182 ret = kiocb_write_and_wait(iocb, count: orig_count);
183 if (ret < 0)
184 return ret;
185 file_accessed(file: iocb->ki_filp);
186
187 rreq = netfs_alloc_request(mapping: iocb->ki_filp->f_mapping, file: iocb->ki_filp,
188 start: iocb->ki_pos, len: orig_count,
189 origin: iocb->ki_flags & IOCB_DIRECT ?
190 NETFS_DIO_READ : NETFS_UNBUFFERED_READ);
191 if (IS_ERR(ptr: rreq))
192 return PTR_ERR(ptr: rreq);
193
194 netfs_stat(&netfs_n_rh_dio_read);
195 trace_netfs_read(rreq, start: rreq->start, len: rreq->len, what: netfs_read_trace_dio_read);
196
197 /* If this is an async op, we have to keep track of the destination
198 * buffer for ourselves as the caller's iterator will be trashed when
199 * we return.
200 *
201 * In such a case, extract an iterator to represent as much of the the
202 * output buffer as we can manage. Note that the extraction might not
203 * be able to allocate a sufficiently large bvec array and may shorten
204 * the request.
205 */
206 if (user_backed_iter(i: iter)) {
207 ret = netfs_extract_user_iter(orig: iter, orig_len: rreq->len, new: &rreq->buffer.iter, extraction_flags: 0);
208 if (ret < 0)
209 goto error_put;
210 rreq->direct_bv = (struct bio_vec *)rreq->buffer.iter.bvec;
211 rreq->direct_bv_count = ret;
212 rreq->direct_bv_unpin = iov_iter_extract_will_pin(iter);
213 rreq->len = iov_iter_count(i: &rreq->buffer.iter);
214 } else {
215 rreq->buffer.iter = *iter;
216 rreq->len = orig_count;
217 rreq->direct_bv_unpin = false;
218 iov_iter_advance(i: iter, bytes: orig_count);
219 }
220
221 // TODO: Set up bounce buffer if needed
222
223 if (!sync) {
224 rreq->iocb = iocb;
225 __set_bit(NETFS_RREQ_OFFLOAD_COLLECTION, &rreq->flags);
226 }
227
228 ret = netfs_unbuffered_read(rreq, sync);
229 if (ret < 0)
230 goto out; /* May be -EIOCBQUEUED */
231 if (sync) {
232 // TODO: Copy from bounce buffer
233 iocb->ki_pos += rreq->transferred;
234 ret = rreq->transferred;
235 }
236
237out:
238 netfs_put_request(rreq, what: netfs_rreq_trace_put_return);
239 if (ret > 0)
240 orig_count -= ret;
241 return ret;
242
243error_put:
244 netfs_put_failed_request(rreq);
245 return ret;
246}
247EXPORT_SYMBOL(netfs_unbuffered_read_iter_locked);
248
249/**
250 * netfs_unbuffered_read_iter - Perform an unbuffered or direct I/O read
251 * @iocb: The I/O control descriptor describing the read
252 * @iter: The output buffer (also specifies read length)
253 *
254 * Perform an unbuffered I/O or direct I/O from the file in @iocb to the
255 * output buffer. No use is made of the pagecache.
256 */
257ssize_t netfs_unbuffered_read_iter(struct kiocb *iocb, struct iov_iter *iter)
258{
259 struct inode *inode = file_inode(f: iocb->ki_filp);
260 ssize_t ret;
261
262 if (!iter->count)
263 return 0; /* Don't update atime */
264
265 ret = netfs_start_io_direct(inode);
266 if (ret == 0) {
267 ret = netfs_unbuffered_read_iter_locked(iocb, iter);
268 netfs_end_io_direct(inode);
269 }
270 return ret;
271}
272EXPORT_SYMBOL(netfs_unbuffered_read_iter);
273