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 12 static 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->len; 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 */ 30 ssize_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 (async || 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 wreq->iter = *iter; 81 } 82 83 wreq->io_iter = wreq->iter; 84 } 85 86 __set_bit(NETFS_RREQ_USE_IO_ITER, &wreq->flags); 87 88 /* Copy the data into the bounce buffer and encrypt it. */ 89 // TODO 90 91 /* Dispatch the write. */ 92 __set_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags); 93 if (async) 94 wreq->iocb = iocb; 95 wreq->cleanup = netfs_cleanup_dio_write; 96 ret = netfs_unbuffered_write(wreq, is_sync_kiocb(iocb), iov_iter_count(&wreq->io_iter)); 97 if (ret < 0) { 98 _debug("begin = %zd", ret); 99 goto out; 100 } 101 102 if (!async) { 103 trace_netfs_rreq(wreq, netfs_rreq_trace_wait_ip); 104 wait_on_bit(&wreq->flags, NETFS_RREQ_IN_PROGRESS, 105 TASK_UNINTERRUPTIBLE); 106 smp_rmb(); /* Read error/transferred after RIP flag */ 107 ret = wreq->error; 108 if (ret == 0) { 109 ret = wreq->transferred; 110 iocb->ki_pos += ret; 111 } 112 } else { 113 ret = -EIOCBQUEUED; 114 } 115 116 out: 117 netfs_put_request(wreq, false, netfs_rreq_trace_put_return); 118 return ret; 119 } 120 EXPORT_SYMBOL(netfs_unbuffered_write_iter_locked); 121 122 /** 123 * netfs_unbuffered_write_iter - Unbuffered write to a file 124 * @iocb: IO state structure 125 * @from: iov_iter with data to write 126 * 127 * Do an unbuffered write to a file, writing the data directly to the server 128 * and not lodging the data in the pagecache. 129 * 130 * Return: 131 * * Negative error code if no data has been written at all of 132 * vfs_fsync_range() failed for a synchronous write 133 * * Number of bytes written, even for truncated writes 134 */ 135 ssize_t netfs_unbuffered_write_iter(struct kiocb *iocb, struct iov_iter *from) 136 { 137 struct file *file = iocb->ki_filp; 138 struct address_space *mapping = file->f_mapping; 139 struct inode *inode = mapping->host; 140 struct netfs_inode *ictx = netfs_inode(inode); 141 ssize_t ret; 142 loff_t pos = iocb->ki_pos; 143 unsigned long long end = pos + iov_iter_count(from) - 1; 144 145 _enter("%llx,%zx,%llx", pos, iov_iter_count(from), i_size_read(inode)); 146 147 if (!iov_iter_count(from)) 148 return 0; 149 150 trace_netfs_write_iter(iocb, from); 151 netfs_stat(&netfs_n_wh_dio_write); 152 153 ret = netfs_start_io_direct(inode); 154 if (ret < 0) 155 return ret; 156 ret = generic_write_checks(iocb, from); 157 if (ret <= 0) 158 goto out; 159 ret = file_remove_privs(file); 160 if (ret < 0) 161 goto out; 162 ret = file_update_time(file); 163 if (ret < 0) 164 goto out; 165 if (iocb->ki_flags & IOCB_NOWAIT) { 166 /* We could block if there are any pages in the range. */ 167 ret = -EAGAIN; 168 if (filemap_range_has_page(mapping, pos, end)) 169 if (filemap_invalidate_inode(inode, true, pos, end)) 170 goto out; 171 } else { 172 ret = filemap_write_and_wait_range(mapping, pos, end); 173 if (ret < 0) 174 goto out; 175 } 176 177 /* 178 * After a write we want buffered reads to be sure to go to disk to get 179 * the new data. We invalidate clean cached page from the region we're 180 * about to write. We do this *before* the write so that we can return 181 * without clobbering -EIOCBQUEUED from ->direct_IO(). 182 */ 183 ret = filemap_invalidate_inode(inode, true, pos, end); 184 if (ret < 0) 185 goto out; 186 end = iocb->ki_pos + iov_iter_count(from); 187 if (end > ictx->zero_point) 188 ictx->zero_point = end; 189 190 fscache_invalidate(netfs_i_cookie(ictx), NULL, i_size_read(inode), 191 FSCACHE_INVAL_DIO_WRITE); 192 ret = netfs_unbuffered_write_iter_locked(iocb, from, NULL); 193 out: 194 netfs_end_io_direct(inode); 195 return ret; 196 } 197 EXPORT_SYMBOL(netfs_unbuffered_write_iter); 198