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 /* 13 * Perform the cleanup rituals after an unbuffered write is complete. 14 */ 15 static void netfs_unbuffered_write_done(struct netfs_io_request *wreq) 16 { 17 struct netfs_inode *ictx = netfs_inode(wreq->inode); 18 19 _enter("R=%x", wreq->debug_id); 20 21 /* Okay, declare that all I/O is complete. */ 22 trace_netfs_rreq(wreq, netfs_rreq_trace_write_done); 23 24 if (!wreq->error) 25 netfs_update_i_size(ictx, &ictx->inode, wreq->start, wreq->transferred); 26 27 if (wreq->origin == NETFS_DIO_WRITE && 28 wreq->mapping->nrpages) { 29 /* mmap may have got underfoot and we may now have folios 30 * locally covering the region we just wrote. Attempt to 31 * discard the folios, but leave in place any modified locally. 32 * ->write_iter() is prevented from interfering by the DIO 33 * counter. 34 */ 35 pgoff_t first = wreq->start >> PAGE_SHIFT; 36 pgoff_t last = (wreq->start + wreq->transferred - 1) >> PAGE_SHIFT; 37 38 invalidate_inode_pages2_range(wreq->mapping, first, last); 39 } 40 41 if (wreq->origin == NETFS_DIO_WRITE) 42 inode_dio_end(wreq->inode); 43 44 _debug("finished"); 45 netfs_wake_rreq_flag(wreq, NETFS_RREQ_IN_PROGRESS, netfs_rreq_trace_wake_ip); 46 /* As we cleared NETFS_RREQ_IN_PROGRESS, we acquired its ref. */ 47 48 if (wreq->iocb) { 49 size_t written = umin(wreq->transferred, wreq->len); 50 51 wreq->iocb->ki_pos += written; 52 if (wreq->iocb->ki_complete) { 53 trace_netfs_rreq(wreq, netfs_rreq_trace_ki_complete); 54 wreq->iocb->ki_complete(wreq->iocb, wreq->error ?: written); 55 } 56 wreq->iocb = VFS_PTR_POISON; 57 } 58 59 netfs_clear_subrequests(wreq); 60 } 61 62 /* 63 * Collect the subrequest results of unbuffered write subrequests. 64 */ 65 static void netfs_unbuffered_write_collect(struct netfs_io_request *wreq, 66 struct netfs_io_stream *stream, 67 struct netfs_io_subrequest *subreq) 68 { 69 trace_netfs_collect_sreq(wreq, subreq); 70 71 spin_lock(&wreq->lock); 72 list_del_init(&subreq->rreq_link); 73 spin_unlock(&wreq->lock); 74 75 wreq->transferred += subreq->transferred; 76 iov_iter_advance(&wreq->buffer.iter, subreq->transferred); 77 78 stream->collected_to = subreq->start + subreq->transferred; 79 wreq->collected_to = stream->collected_to; 80 netfs_put_subrequest(subreq, netfs_sreq_trace_put_done); 81 82 trace_netfs_collect_stream(wreq, stream); 83 trace_netfs_collect_state(wreq, wreq->collected_to, 0); 84 } 85 86 /* 87 * Write data to the server without going through the pagecache and without 88 * writing it to the local cache. We dispatch the subrequests serially and 89 * wait for each to complete before dispatching the next, lest we leave a gap 90 * in the data written due to a failure such as ENOSPC. We could, however 91 * attempt to do preparation such as content encryption for the next subreq 92 * whilst the current is in progress. 93 */ 94 static int netfs_unbuffered_write(struct netfs_io_request *wreq) 95 { 96 struct netfs_io_subrequest *subreq = NULL; 97 struct netfs_io_stream *stream = &wreq->io_streams[0]; 98 int ret; 99 100 _enter("%llx", wreq->len); 101 102 if (wreq->origin == NETFS_DIO_WRITE) 103 inode_dio_begin(wreq->inode); 104 105 stream->collected_to = wreq->start; 106 107 for (;;) { 108 bool retry = false; 109 110 if (!subreq) { 111 netfs_prepare_write(wreq, stream, wreq->start + wreq->transferred); 112 subreq = stream->construct; 113 stream->construct = NULL; 114 stream->front = NULL; 115 } 116 117 /* Check if (re-)preparation failed. */ 118 if (unlikely(test_bit(NETFS_SREQ_FAILED, &subreq->flags))) { 119 netfs_write_subrequest_terminated(subreq, subreq->error); 120 wreq->error = subreq->error; 121 break; 122 } 123 124 iov_iter_truncate(&subreq->io_iter, wreq->len - wreq->transferred); 125 if (!iov_iter_count(&subreq->io_iter)) 126 break; 127 128 subreq->len = netfs_limit_iter(&subreq->io_iter, 0, 129 stream->sreq_max_len, 130 stream->sreq_max_segs); 131 iov_iter_truncate(&subreq->io_iter, subreq->len); 132 stream->submit_extendable_to = subreq->len; 133 134 trace_netfs_sreq(subreq, netfs_sreq_trace_submit); 135 stream->issue_write(subreq); 136 137 /* Async, need to wait. */ 138 netfs_wait_for_in_progress_stream(wreq, stream); 139 140 if (test_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags)) { 141 retry = true; 142 } else if (test_bit(NETFS_SREQ_FAILED, &subreq->flags)) { 143 ret = subreq->error; 144 wreq->error = ret; 145 netfs_see_subrequest(subreq, netfs_sreq_trace_see_failed); 146 subreq = NULL; 147 break; 148 } 149 ret = 0; 150 151 if (!retry) { 152 netfs_unbuffered_write_collect(wreq, stream, subreq); 153 subreq = NULL; 154 if (wreq->transferred >= wreq->len) 155 break; 156 if (!wreq->iocb && signal_pending(current)) { 157 ret = wreq->transferred ? -EINTR : -ERESTARTSYS; 158 trace_netfs_rreq(wreq, netfs_rreq_trace_intr); 159 break; 160 } 161 continue; 162 } 163 164 /* We need to retry the last subrequest, so first reset the 165 * iterator, taking into account what, if anything, we managed 166 * to transfer. 167 */ 168 subreq->error = -EAGAIN; 169 trace_netfs_sreq(subreq, netfs_sreq_trace_retry); 170 if (subreq->transferred > 0) 171 iov_iter_advance(&wreq->buffer.iter, subreq->transferred); 172 173 if (stream->source == NETFS_UPLOAD_TO_SERVER && 174 wreq->netfs_ops->retry_request) 175 wreq->netfs_ops->retry_request(wreq, stream); 176 177 __clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags); 178 __clear_bit(NETFS_SREQ_BOUNDARY, &subreq->flags); 179 __clear_bit(NETFS_SREQ_FAILED, &subreq->flags); 180 subreq->io_iter = wreq->buffer.iter; 181 subreq->start = wreq->start + wreq->transferred; 182 subreq->len = wreq->len - wreq->transferred; 183 subreq->transferred = 0; 184 subreq->retry_count += 1; 185 stream->sreq_max_len = UINT_MAX; 186 stream->sreq_max_segs = INT_MAX; 187 188 netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit); 189 stream->prepare_write(subreq); 190 191 __set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags); 192 netfs_stat(&netfs_n_wh_retry_write_subreq); 193 } 194 195 netfs_unbuffered_write_done(wreq); 196 _leave(" = %d", ret); 197 return ret; 198 } 199 200 static void netfs_unbuffered_write_async(struct work_struct *work) 201 { 202 struct netfs_io_request *wreq = container_of(work, struct netfs_io_request, work); 203 204 netfs_unbuffered_write(wreq); 205 netfs_put_request(wreq, netfs_rreq_trace_put_complete); 206 } 207 208 /* 209 * Perform an unbuffered write where we may have to do an RMW operation on an 210 * encrypted file. This can also be used for direct I/O writes. 211 */ 212 ssize_t netfs_unbuffered_write_iter_locked(struct kiocb *iocb, struct iov_iter *iter, 213 struct netfs_group *netfs_group) 214 { 215 struct netfs_io_request *wreq; 216 unsigned long long start = iocb->ki_pos; 217 unsigned long long end = start + iov_iter_count(iter); 218 ssize_t ret, n; 219 size_t len = iov_iter_count(iter); 220 bool async = !is_sync_kiocb(iocb); 221 222 _enter(""); 223 224 /* We're going to need a bounce buffer if what we transmit is going to 225 * be different in some way to the source buffer, e.g. because it gets 226 * encrypted/compressed or because it needs expanding to a block size. 227 */ 228 // TODO 229 230 _debug("uw %llx-%llx", start, end); 231 232 wreq = netfs_create_write_req(iocb->ki_filp->f_mapping, iocb->ki_filp, start, 233 iocb->ki_flags & IOCB_DIRECT ? 234 NETFS_DIO_WRITE : NETFS_UNBUFFERED_WRITE); 235 if (IS_ERR(wreq)) 236 return PTR_ERR(wreq); 237 238 wreq->io_streams[0].avail = true; 239 trace_netfs_write(wreq, (iocb->ki_flags & IOCB_DIRECT ? 240 netfs_write_trace_dio_write : 241 netfs_write_trace_unbuffered_write)); 242 243 { 244 /* If this is an async op and we're not using a bounce buffer, 245 * we have to save the source buffer as the iterator is only 246 * good until we return. In such a case, extract an iterator 247 * to represent as much of the the output buffer as we can 248 * manage. Note that the extraction might not be able to 249 * allocate a sufficiently large bvec array and may shorten the 250 * request. 251 */ 252 if (user_backed_iter(iter)) { 253 n = netfs_extract_user_iter(iter, len, &wreq->buffer.iter, 0); 254 if (n < 0) { 255 ret = n; 256 goto error_put; 257 } 258 wreq->direct_bv = (struct bio_vec *)wreq->buffer.iter.bvec; 259 wreq->direct_bv_count = n; 260 wreq->direct_bv_unpin = iov_iter_extract_will_pin(iter); 261 } else { 262 /* If this is a kernel-generated async DIO request, 263 * assume that any resources the iterator points to 264 * (eg. a bio_vec array) will persist till the end of 265 * the op. 266 */ 267 wreq->buffer.iter = *iter; 268 } 269 270 wreq->len = iov_iter_count(&wreq->buffer.iter); 271 } 272 273 __set_bit(NETFS_RREQ_USE_IO_ITER, &wreq->flags); 274 275 /* Copy the data into the bounce buffer and encrypt it. */ 276 // TODO 277 278 /* Dispatch the write. */ 279 __set_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags); 280 281 if (async) { 282 INIT_WORK(&wreq->work, netfs_unbuffered_write_async); 283 wreq->iocb = iocb; 284 queue_work(system_dfl_wq, &wreq->work); 285 ret = -EIOCBQUEUED; 286 } else { 287 ret = netfs_unbuffered_write(wreq); 288 if (ret < 0) { 289 _debug("begin = %zd", ret); 290 } else { 291 iocb->ki_pos += wreq->transferred; 292 ret = wreq->transferred ?: wreq->error; 293 } 294 295 netfs_put_request(wreq, netfs_rreq_trace_put_complete); 296 } 297 298 netfs_put_request(wreq, netfs_rreq_trace_put_return); 299 return ret; 300 301 error_put: 302 netfs_put_failed_request(wreq); 303 return ret; 304 } 305 EXPORT_SYMBOL(netfs_unbuffered_write_iter_locked); 306 307 /** 308 * netfs_unbuffered_write_iter - Unbuffered write to a file 309 * @iocb: IO state structure 310 * @from: iov_iter with data to write 311 * 312 * Do an unbuffered write to a file, writing the data directly to the server 313 * and not lodging the data in the pagecache. 314 * 315 * Return: 316 * * Negative error code if no data has been written at all of 317 * vfs_fsync_range() failed for a synchronous write 318 * * Number of bytes written, even for truncated writes 319 */ 320 ssize_t netfs_unbuffered_write_iter(struct kiocb *iocb, struct iov_iter *from) 321 { 322 struct file *file = iocb->ki_filp; 323 struct address_space *mapping = file->f_mapping; 324 struct inode *inode = mapping->host; 325 struct netfs_inode *ictx = netfs_inode(inode); 326 ssize_t ret; 327 loff_t pos = iocb->ki_pos; 328 unsigned long long end = pos + iov_iter_count(from) - 1; 329 330 _enter("%llx,%zx,%llx", pos, iov_iter_count(from), i_size_read(inode)); 331 332 if (!iov_iter_count(from)) 333 return 0; 334 335 trace_netfs_write_iter(iocb, from); 336 netfs_stat(&netfs_n_wh_dio_write); 337 338 ret = netfs_start_io_direct(inode); 339 if (ret < 0) 340 return ret; 341 ret = generic_write_checks(iocb, from); 342 if (ret <= 0) 343 goto out; 344 ret = file_remove_privs(file); 345 if (ret < 0) 346 goto out; 347 ret = file_update_time(file); 348 if (ret < 0) 349 goto out; 350 if (iocb->ki_flags & IOCB_NOWAIT) { 351 /* We could block if there are any pages in the range. */ 352 ret = -EAGAIN; 353 if (filemap_range_has_page(mapping, pos, end)) 354 if (filemap_invalidate_inode(inode, true, pos, end)) 355 goto out; 356 } else { 357 ret = filemap_write_and_wait_range(mapping, pos, end); 358 if (ret < 0) 359 goto out; 360 } 361 362 /* 363 * After a write we want buffered reads to be sure to go to disk to get 364 * the new data. We invalidate clean cached page from the region we're 365 * about to write. We do this *before* the write so that we can return 366 * without clobbering -EIOCBQUEUED from ->direct_IO(). 367 */ 368 ret = filemap_invalidate_inode(inode, true, pos, end); 369 if (ret < 0) 370 goto out; 371 end = iocb->ki_pos + iov_iter_count(from); 372 if (end > ictx->zero_point) 373 ictx->zero_point = end; 374 375 fscache_invalidate(netfs_i_cookie(ictx), NULL, i_size_read(inode), 376 FSCACHE_INVAL_DIO_WRITE); 377 ret = netfs_unbuffered_write_iter_locked(iocb, from, NULL); 378 out: 379 netfs_end_io_direct(inode); 380 return ret; 381 } 382 EXPORT_SYMBOL(netfs_unbuffered_write_iter); 383