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 } 115 116 /* Check if (re-)preparation failed. */ 117 if (unlikely(test_bit(NETFS_SREQ_FAILED, &subreq->flags))) { 118 netfs_write_subrequest_terminated(subreq, subreq->error); 119 wreq->error = subreq->error; 120 break; 121 } 122 123 iov_iter_truncate(&subreq->io_iter, wreq->len - wreq->transferred); 124 if (!iov_iter_count(&subreq->io_iter)) 125 break; 126 127 subreq->len = netfs_limit_iter(&subreq->io_iter, 0, 128 stream->sreq_max_len, 129 stream->sreq_max_segs); 130 iov_iter_truncate(&subreq->io_iter, subreq->len); 131 stream->submit_extendable_to = subreq->len; 132 133 trace_netfs_sreq(subreq, netfs_sreq_trace_submit); 134 stream->issue_write(subreq); 135 136 /* Async, need to wait. */ 137 netfs_wait_for_in_progress_stream(wreq, stream); 138 139 if (test_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags)) { 140 retry = true; 141 } else if (test_bit(NETFS_SREQ_FAILED, &subreq->flags)) { 142 ret = subreq->error; 143 wreq->error = ret; 144 netfs_see_subrequest(subreq, netfs_sreq_trace_see_failed); 145 subreq = NULL; 146 break; 147 } 148 ret = 0; 149 150 if (!retry) { 151 netfs_unbuffered_write_collect(wreq, stream, subreq); 152 subreq = NULL; 153 if (wreq->transferred >= wreq->len) 154 break; 155 if (!wreq->iocb && signal_pending(current)) { 156 ret = wreq->transferred ? -EINTR : -ERESTARTSYS; 157 trace_netfs_rreq(wreq, netfs_rreq_trace_intr); 158 break; 159 } 160 continue; 161 } 162 163 /* We need to retry the last subrequest, so first reset the 164 * iterator, taking into account what, if anything, we managed 165 * to transfer. 166 */ 167 subreq->error = -EAGAIN; 168 trace_netfs_sreq(subreq, netfs_sreq_trace_retry); 169 if (subreq->transferred > 0) { 170 iov_iter_advance(&wreq->buffer.iter, subreq->transferred); 171 wreq->transferred += subreq->transferred; 172 } 173 174 if (stream->source == NETFS_UPLOAD_TO_SERVER && 175 wreq->netfs_ops->retry_request) 176 wreq->netfs_ops->retry_request(wreq, stream); 177 178 __clear_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags); 179 __clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags); 180 __clear_bit(NETFS_SREQ_BOUNDARY, &subreq->flags); 181 __clear_bit(NETFS_SREQ_FAILED, &subreq->flags); 182 subreq->io_iter = wreq->buffer.iter; 183 subreq->start = wreq->start + wreq->transferred; 184 subreq->len = wreq->len - wreq->transferred; 185 subreq->transferred = 0; 186 subreq->retry_count += 1; 187 stream->sreq_max_len = UINT_MAX; 188 stream->sreq_max_segs = INT_MAX; 189 190 netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit); 191 192 if (stream->prepare_write) 193 stream->prepare_write(subreq); 194 __set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags); 195 netfs_stat(&netfs_n_wh_retry_write_subreq); 196 } 197 198 netfs_unbuffered_write_done(wreq); 199 _leave(" = %d", ret); 200 return ret; 201 } 202 203 static void netfs_unbuffered_write_async(struct work_struct *work) 204 { 205 struct netfs_io_request *wreq = container_of(work, struct netfs_io_request, work); 206 207 netfs_unbuffered_write(wreq); 208 netfs_put_request(wreq, netfs_rreq_trace_put_complete); 209 } 210 211 /* 212 * Perform an unbuffered write where we may have to do an RMW operation on an 213 * encrypted file. This can also be used for direct I/O writes. 214 */ 215 ssize_t netfs_unbuffered_write_iter_locked(struct kiocb *iocb, struct iov_iter *iter, 216 struct netfs_group *netfs_group) 217 { 218 struct netfs_io_request *wreq; 219 unsigned long long start = iocb->ki_pos; 220 unsigned long long end = start + iov_iter_count(iter); 221 ssize_t ret, n; 222 size_t len = iov_iter_count(iter); 223 bool async = !is_sync_kiocb(iocb); 224 225 _enter(""); 226 227 /* We're going to need a bounce buffer if what we transmit is going to 228 * be different in some way to the source buffer, e.g. because it gets 229 * encrypted/compressed or because it needs expanding to a block size. 230 */ 231 // TODO 232 233 _debug("uw %llx-%llx", start, end); 234 235 wreq = netfs_create_write_req(iocb->ki_filp->f_mapping, iocb->ki_filp, start, 236 iocb->ki_flags & IOCB_DIRECT ? 237 NETFS_DIO_WRITE : NETFS_UNBUFFERED_WRITE); 238 if (IS_ERR(wreq)) 239 return PTR_ERR(wreq); 240 241 wreq->io_streams[0].avail = true; 242 trace_netfs_write(wreq, (iocb->ki_flags & IOCB_DIRECT ? 243 netfs_write_trace_dio_write : 244 netfs_write_trace_unbuffered_write)); 245 246 { 247 /* If this is an async op and we're not using a bounce buffer, 248 * we have to save the source buffer as the iterator is only 249 * good until we return. In such a case, extract an iterator 250 * to represent as much of the the output buffer as we can 251 * manage. Note that the extraction might not be able to 252 * allocate a sufficiently large bvec array and may shorten the 253 * request. 254 */ 255 if (user_backed_iter(iter)) { 256 n = netfs_extract_user_iter(iter, len, &wreq->buffer.iter, 0); 257 if (n < 0) { 258 ret = n; 259 goto error_put; 260 } 261 wreq->direct_bv = (struct bio_vec *)wreq->buffer.iter.bvec; 262 wreq->direct_bv_count = n; 263 wreq->direct_bv_unpin = iov_iter_extract_will_pin(iter); 264 } else { 265 /* If this is a kernel-generated async DIO request, 266 * assume that any resources the iterator points to 267 * (eg. a bio_vec array) will persist till the end of 268 * the op. 269 */ 270 wreq->buffer.iter = *iter; 271 } 272 273 wreq->len = iov_iter_count(&wreq->buffer.iter); 274 } 275 276 __set_bit(NETFS_RREQ_USE_IO_ITER, &wreq->flags); 277 278 /* Copy the data into the bounce buffer and encrypt it. */ 279 // TODO 280 281 /* Dispatch the write. */ 282 __set_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags); 283 284 if (async) { 285 INIT_WORK(&wreq->work, netfs_unbuffered_write_async); 286 wreq->iocb = iocb; 287 queue_work(system_dfl_wq, &wreq->work); 288 ret = -EIOCBQUEUED; 289 } else { 290 ret = netfs_unbuffered_write(wreq); 291 if (ret < 0) { 292 _debug("begin = %zd", ret); 293 } else { 294 iocb->ki_pos += wreq->transferred; 295 ret = wreq->transferred ?: wreq->error; 296 } 297 298 netfs_put_request(wreq, netfs_rreq_trace_put_complete); 299 } 300 301 netfs_put_request(wreq, netfs_rreq_trace_put_return); 302 return ret; 303 304 error_put: 305 netfs_put_failed_request(wreq); 306 return ret; 307 } 308 EXPORT_SYMBOL(netfs_unbuffered_write_iter_locked); 309 310 /** 311 * netfs_unbuffered_write_iter - Unbuffered write to a file 312 * @iocb: IO state structure 313 * @from: iov_iter with data to write 314 * 315 * Do an unbuffered write to a file, writing the data directly to the server 316 * and not lodging the data in the pagecache. 317 * 318 * Return: 319 * * Negative error code if no data has been written at all of 320 * vfs_fsync_range() failed for a synchronous write 321 * * Number of bytes written, even for truncated writes 322 */ 323 ssize_t netfs_unbuffered_write_iter(struct kiocb *iocb, struct iov_iter *from) 324 { 325 struct file *file = iocb->ki_filp; 326 struct address_space *mapping = file->f_mapping; 327 struct inode *inode = mapping->host; 328 struct netfs_inode *ictx = netfs_inode(inode); 329 ssize_t ret; 330 loff_t pos = iocb->ki_pos; 331 unsigned long long end = pos + iov_iter_count(from) - 1; 332 333 _enter("%llx,%zx,%llx", pos, iov_iter_count(from), i_size_read(inode)); 334 335 if (!iov_iter_count(from)) 336 return 0; 337 338 trace_netfs_write_iter(iocb, from); 339 netfs_stat(&netfs_n_wh_dio_write); 340 341 ret = netfs_start_io_direct(inode); 342 if (ret < 0) 343 return ret; 344 ret = generic_write_checks(iocb, from); 345 if (ret <= 0) 346 goto out; 347 ret = file_remove_privs(file); 348 if (ret < 0) 349 goto out; 350 ret = file_update_time(file); 351 if (ret < 0) 352 goto out; 353 if (iocb->ki_flags & IOCB_NOWAIT) { 354 /* We could block if there are any pages in the range. */ 355 ret = -EAGAIN; 356 if (filemap_range_has_page(mapping, pos, end)) 357 if (filemap_invalidate_inode(inode, true, pos, end)) 358 goto out; 359 } else { 360 ret = filemap_write_and_wait_range(mapping, pos, end); 361 if (ret < 0) 362 goto out; 363 } 364 365 /* 366 * After a write we want buffered reads to be sure to go to disk to get 367 * the new data. We invalidate clean cached page from the region we're 368 * about to write. We do this *before* the write so that we can return 369 * without clobbering -EIOCBQUEUED from ->direct_IO(). 370 */ 371 ret = filemap_invalidate_inode(inode, true, pos, end); 372 if (ret < 0) 373 goto out; 374 end = iocb->ki_pos + iov_iter_count(from); 375 spin_lock(&inode->i_lock); 376 if (end > ictx->_zero_point) 377 netfs_write_zero_point(inode, end); 378 spin_unlock(&inode->i_lock); 379 380 fscache_invalidate(netfs_i_cookie(ictx), NULL, i_size_read(inode), 381 FSCACHE_INVAL_DIO_WRITE); 382 ret = netfs_unbuffered_write_iter_locked(iocb, from, NULL); 383 out: 384 netfs_end_io_direct(inode); 385 return ret; 386 } 387 EXPORT_SYMBOL(netfs_unbuffered_write_iter); 388