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