1// SPDX-License-Identifier: GPL-2.0-or-later
  2/* Unbuffered and direct write 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/uio.h>
 10#include "internal.h"
 11
 12static void netfs_cleanup_dio_write(struct netfs_io_request *wreq)
 13{
 14	struct inode *inode = wreq->inode;
 15	unsigned long long end = wreq->start + wreq->transferred;
 16
 17	if (!wreq->error &&
 18	    i_size_read(inode) < end) {
 19		if (wreq->netfs_ops->update_i_size)
 20			wreq->netfs_ops->update_i_size(inode, end);
 21		else
 22			i_size_write(inode, end);
 23	}
 24}
 25
 26/*
 27 * Perform an unbuffered write where we may have to do an RMW operation on an
 28 * encrypted file.  This can also be used for direct I/O writes.
 29 */
 30ssize_t netfs_unbuffered_write_iter_locked(struct kiocb *iocb, struct iov_iter *iter,
 31						  struct netfs_group *netfs_group)
 32{
 33	struct netfs_io_request *wreq;
 34	unsigned long long start = iocb->ki_pos;
 35	unsigned long long end = start + iov_iter_count(iter);
 36	ssize_t ret, n;
 37	size_t len = iov_iter_count(iter);
 38	bool async = !is_sync_kiocb(iocb);
 39
 40	_enter("");
 41
 42	/* We're going to need a bounce buffer if what we transmit is going to
 43	 * be different in some way to the source buffer, e.g. because it gets
 44	 * encrypted/compressed or because it needs expanding to a block size.
 45	 */
 46	// TODO
 47
 48	_debug("uw %llx-%llx", start, end);
 49
 50	wreq = netfs_create_write_req(iocb->ki_filp->f_mapping, iocb->ki_filp, start,
 51				      iocb->ki_flags & IOCB_DIRECT ?
 52				      NETFS_DIO_WRITE : NETFS_UNBUFFERED_WRITE);
 53	if (IS_ERR(wreq))
 54		return PTR_ERR(wreq);
 55
 56	wreq->io_streams[0].avail = true;
 57	trace_netfs_write(wreq, (iocb->ki_flags & IOCB_DIRECT ?
 58				 netfs_write_trace_dio_write :
 59				 netfs_write_trace_unbuffered_write));
 60
 61	{
 62		/* If this is an async op and we're not using a bounce buffer,
 63		 * we have to save the source buffer as the iterator is only
 64		 * good until we return.  In such a case, extract an iterator
 65		 * to represent as much of the the output buffer as we can
 66		 * manage.  Note that the extraction might not be able to
 67		 * allocate a sufficiently large bvec array and may shorten the
 68		 * request.
 69		 */
 70		if (user_backed_iter(iter)) {
 71			n = netfs_extract_user_iter(iter, len, &wreq->iter, 0);
 72			if (n < 0) {
 73				ret = n;
 74				goto out;
 75			}
 76			wreq->direct_bv = (struct bio_vec *)wreq->iter.bvec;
 77			wreq->direct_bv_count = n;
 78			wreq->direct_bv_unpin = iov_iter_extract_will_pin(iter);
 79		} else {
 80			/* If this is a kernel-generated async DIO request,
 81			 * assume that any resources the iterator points to
 82			 * (eg. a bio_vec array) will persist till the end of
 83			 * the op.
 84			 */
 85			wreq->iter = *iter;
 86		}
 87
 88		wreq->io_iter = wreq->iter;
 89	}
 90
 91	__set_bit(NETFS_RREQ_USE_IO_ITER, &wreq->flags);
 92
 93	/* Copy the data into the bounce buffer and encrypt it. */
 94	// TODO
 95
 96	/* Dispatch the write. */
 97	__set_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags);
 98	if (async)
 99		wreq->iocb = iocb;
100	wreq->len = iov_iter_count(&wreq->io_iter);
101	wreq->cleanup = netfs_cleanup_dio_write;
102	ret = netfs_unbuffered_write(wreq, is_sync_kiocb(iocb), wreq->len);
103	if (ret < 0) {
104		_debug("begin = %zd", ret);
105		goto out;
106	}
107
108	if (!async) {
109		trace_netfs_rreq(wreq, netfs_rreq_trace_wait_ip);
110		wait_on_bit(&wreq->flags, NETFS_RREQ_IN_PROGRESS,
111			    TASK_UNINTERRUPTIBLE);
112		ret = wreq->error;
113		if (ret == 0) {
114			ret = wreq->transferred;
115			iocb->ki_pos += ret;
116		}
117	} else {
118		ret = -EIOCBQUEUED;
119	}
120
121out:
122	netfs_put_request(wreq, false, netfs_rreq_trace_put_return);
123	return ret;
124}
125EXPORT_SYMBOL(netfs_unbuffered_write_iter_locked);
126
127/**
128 * netfs_unbuffered_write_iter - Unbuffered write to a file
129 * @iocb: IO state structure
130 * @from: iov_iter with data to write
131 *
132 * Do an unbuffered write to a file, writing the data directly to the server
133 * and not lodging the data in the pagecache.
134 *
135 * Return:
136 * * Negative error code if no data has been written at all of
137 *   vfs_fsync_range() failed for a synchronous write
138 * * Number of bytes written, even for truncated writes
139 */
140ssize_t netfs_unbuffered_write_iter(struct kiocb *iocb, struct iov_iter *from)
141{
142	struct file *file = iocb->ki_filp;
143	struct address_space *mapping = file->f_mapping;
144	struct inode *inode = mapping->host;
145	struct netfs_inode *ictx = netfs_inode(inode);
146	ssize_t ret;
147	loff_t pos = iocb->ki_pos;
148	unsigned long long end = pos + iov_iter_count(from) - 1;
149
150	_enter("%llx,%zx,%llx", pos, iov_iter_count(from), i_size_read(inode));
151
152	if (!iov_iter_count(from))
153		return 0;
154
155	trace_netfs_write_iter(iocb, from);
156	netfs_stat(&netfs_n_wh_dio_write);
157
158	ret = netfs_start_io_direct(inode);
159	if (ret < 0)
160		return ret;
161	ret = generic_write_checks(iocb, from);
162	if (ret <= 0)
163		goto out;
164	ret = file_remove_privs(file);
165	if (ret < 0)
166		goto out;
167	ret = file_update_time(file);
168	if (ret < 0)
169		goto out;
170	if (iocb->ki_flags & IOCB_NOWAIT) {
171		/* We could block if there are any pages in the range. */
172		ret = -EAGAIN;
173		if (filemap_range_has_page(mapping, pos, end))
174			if (filemap_invalidate_inode(inode, true, pos, end))
175				goto out;
176	} else {
177		ret = filemap_write_and_wait_range(mapping, pos, end);
178		if (ret < 0)
179			goto out;
180	}
181
182	/*
183	 * After a write we want buffered reads to be sure to go to disk to get
184	 * the new data.  We invalidate clean cached page from the region we're
185	 * about to write.  We do this *before* the write so that we can return
186	 * without clobbering -EIOCBQUEUED from ->direct_IO().
187	 */
188	ret = filemap_invalidate_inode(inode, true, pos, end);
189	if (ret < 0)
190		goto out;
191	end = iocb->ki_pos + iov_iter_count(from);
192	if (end > ictx->zero_point)
193		ictx->zero_point = end;
194
195	fscache_invalidate(netfs_i_cookie(ictx), NULL, i_size_read(inode),
196			   FSCACHE_INVAL_DIO_WRITE);
197	ret = netfs_unbuffered_write_iter_locked(iocb, from, NULL);
198out:
199	netfs_end_io_direct(inode);
200	return ret;
201}
202EXPORT_SYMBOL(netfs_unbuffered_write_iter);