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->transferred) 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, written ?: wreq->error); 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 = 0; 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 if (!subreq) { 114 wreq->error = -ENOMEM; 115 ret = -ENOMEM; 116 break; 117 } 118 stream->construct = NULL; 119 } 120 121 /* Check if (re-)preparation failed. */ 122 if (unlikely(test_bit(NETFS_SREQ_FAILED, &subreq->flags))) { 123 netfs_write_subrequest_terminated(subreq, subreq->error); 124 wreq->error = subreq->error; 125 break; 126 } 127 128 iov_iter_truncate(&subreq->io_iter, wreq->len - wreq->transferred); 129 if (!iov_iter_count(&subreq->io_iter)) { 130 pr_warn("netfs: Unexpected zero-length iterator R=%08x\n", 131 wreq->debug_id); 132 __set_bit(NETFS_SREQ_FAILED, &subreq->flags); 133 netfs_write_subrequest_terminated(subreq, -EIO); 134 wreq->error = -EIO; 135 break; 136 } 137 138 subreq->len = netfs_limit_iter(&subreq->io_iter, 0, 139 stream->sreq_max_len, 140 stream->sreq_max_segs); 141 iov_iter_truncate(&subreq->io_iter, subreq->len); 142 stream->submit_extendable_to = subreq->len; 143 144 trace_netfs_sreq(subreq, netfs_sreq_trace_submit); 145 stream->issue_write(subreq); 146 147 /* Async, need to wait. */ 148 netfs_wait_for_in_progress_stream(wreq, stream); 149 150 if (test_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags)) { 151 retry = true; 152 } else if (test_bit(NETFS_SREQ_FAILED, &subreq->flags)) { 153 wreq->error = subreq->error; 154 netfs_see_subrequest(subreq, netfs_sreq_trace_see_failed); 155 subreq = NULL; 156 break; 157 } 158 159 if (!retry) { 160 netfs_unbuffered_write_collect(wreq, stream, subreq); 161 subreq = NULL; 162 if (wreq->transferred >= wreq->len) 163 break; 164 if (!wreq->iocb && signal_pending(current)) { 165 ret = wreq->transferred ? -EINTR : -ERESTARTSYS; 166 trace_netfs_rreq(wreq, netfs_rreq_trace_intr); 167 break; 168 } 169 continue; 170 } 171 172 /* We need to retry the last subrequest, so first reset the 173 * iterator, taking into account what, if anything, we managed 174 * to transfer. 175 */ 176 subreq->error = -EAGAIN; 177 trace_netfs_sreq(subreq, netfs_sreq_trace_retry); 178 if (subreq->transferred > 0) { 179 iov_iter_advance(&wreq->buffer.iter, subreq->transferred); 180 wreq->transferred += subreq->transferred; 181 } 182 183 if (stream->source == NETFS_UPLOAD_TO_SERVER && 184 wreq->netfs_ops->retry_request) 185 wreq->netfs_ops->retry_request(wreq, stream); 186 187 __clear_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags); 188 __clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags); 189 __clear_bit(NETFS_SREQ_BOUNDARY, &subreq->flags); 190 __clear_bit(NETFS_SREQ_FAILED, &subreq->flags); 191 subreq->io_iter = wreq->buffer.iter; 192 subreq->start = wreq->start + wreq->transferred; 193 subreq->len = wreq->len - wreq->transferred; 194 subreq->transferred = 0; 195 subreq->retry_count += 1; 196 stream->sreq_max_len = UINT_MAX; 197 stream->sreq_max_segs = INT_MAX; 198 199 netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit); 200 201 if (stream->prepare_write) 202 stream->prepare_write(subreq); 203 __set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags); 204 netfs_stat(&netfs_n_wh_retry_write_subreq); 205 } 206 207 netfs_unbuffered_write_done(wreq); 208 _leave(" = %d", ret); 209 return ret; 210 } 211 212 static void netfs_unbuffered_write_async(struct work_struct *work) 213 { 214 struct netfs_io_request *wreq = container_of(work, struct netfs_io_request, work); 215 216 netfs_unbuffered_write(wreq); 217 netfs_put_request(wreq, netfs_rreq_trace_put_complete); 218 } 219 220 /* 221 * Perform an unbuffered write where we may have to do an RMW operation on an 222 * encrypted file. This can also be used for direct I/O writes. 223 */ 224 ssize_t netfs_unbuffered_write_iter_locked(struct kiocb *iocb, struct iov_iter *iter, 225 struct netfs_group *netfs_group) 226 { 227 struct netfs_io_request *wreq; 228 unsigned long long start = iocb->ki_pos; 229 unsigned long long end = start + iov_iter_count(iter); 230 ssize_t ret, n; 231 size_t len = iov_iter_count(iter); 232 bool async = !is_sync_kiocb(iocb); 233 234 _enter(""); 235 236 /* We're going to need a bounce buffer if what we transmit is going to 237 * be different in some way to the source buffer, e.g. because it gets 238 * encrypted/compressed or because it needs expanding to a block size. 239 */ 240 // TODO 241 242 _debug("uw %llx-%llx", start, end); 243 244 wreq = netfs_create_write_req(iocb->ki_filp->f_mapping, iocb->ki_filp, start, 245 iocb->ki_flags & IOCB_DIRECT ? 246 NETFS_DIO_WRITE : NETFS_UNBUFFERED_WRITE); 247 if (IS_ERR(wreq)) 248 return PTR_ERR(wreq); 249 250 wreq->io_streams[0].avail = true; 251 trace_netfs_write(wreq, (iocb->ki_flags & IOCB_DIRECT ? 252 netfs_write_trace_dio_write : 253 netfs_write_trace_unbuffered_write)); 254 255 { 256 /* If this is an async op and we're not using a bounce buffer, 257 * we have to save the source buffer as the iterator is only 258 * good until we return. In such a case, extract an iterator 259 * to represent as much of the the output buffer as we can 260 * manage. Note that the extraction might not be able to 261 * allocate a sufficiently large bvec array and may shorten the 262 * request. 263 */ 264 if (user_backed_iter(iter)) { 265 n = netfs_extract_user_iter(iter, len, &wreq->buffer.iter, 0); 266 if (n < 0) { 267 ret = n; 268 goto error_put; 269 } 270 wreq->direct_bv = (struct bio_vec *)wreq->buffer.iter.bvec; 271 wreq->direct_bv_count = n; 272 wreq->direct_bv_unpin = iov_iter_extract_will_pin(iter); 273 } else { 274 /* If this is a kernel-generated async DIO request, 275 * assume that any resources the iterator points to 276 * (eg. a bio_vec array) will persist till the end of 277 * the op. 278 */ 279 wreq->buffer.iter = *iter; 280 } 281 282 wreq->len = iov_iter_count(&wreq->buffer.iter); 283 } 284 285 __set_bit(NETFS_RREQ_USE_IO_ITER, &wreq->flags); 286 287 /* Copy the data into the bounce buffer and encrypt it. */ 288 // TODO 289 290 /* Dispatch the write. */ 291 __set_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags); 292 293 if (async) { 294 INIT_WORK(&wreq->work, netfs_unbuffered_write_async); 295 wreq->iocb = iocb; 296 queue_work(system_dfl_wq, &wreq->work); 297 ret = -EIOCBQUEUED; 298 } else { 299 ret = netfs_unbuffered_write(wreq); 300 if (wreq->transferred) { 301 iocb->ki_pos += wreq->transferred; 302 ret = wreq->transferred; 303 } else if (wreq->error) { 304 ret = wreq->error; 305 } 306 307 netfs_put_request(wreq, netfs_rreq_trace_put_complete); 308 } 309 310 netfs_put_request(wreq, netfs_rreq_trace_put_return); 311 return ret; 312 313 error_put: 314 netfs_put_failed_request(wreq); 315 return ret; 316 } 317 EXPORT_SYMBOL(netfs_unbuffered_write_iter_locked); 318 319 /** 320 * netfs_unbuffered_write_iter - Unbuffered write to a file 321 * @iocb: IO state structure 322 * @from: iov_iter with data to write 323 * 324 * Do an unbuffered write to a file, writing the data directly to the server 325 * and not lodging the data in the pagecache. 326 * 327 * Return: 328 * * Negative error code if no data has been written at all of 329 * vfs_fsync_range() failed for a synchronous write 330 * * Number of bytes written, even for truncated writes 331 */ 332 ssize_t netfs_unbuffered_write_iter(struct kiocb *iocb, struct iov_iter *from) 333 { 334 struct file *file = iocb->ki_filp; 335 struct address_space *mapping = file->f_mapping; 336 struct inode *inode = mapping->host; 337 struct netfs_inode *ictx = netfs_inode(inode); 338 ssize_t ret; 339 loff_t pos = iocb->ki_pos; 340 unsigned long long end = pos + iov_iter_count(from) - 1; 341 342 _enter("%llx,%zx,%llx", pos, iov_iter_count(from), i_size_read(inode)); 343 344 if (!iov_iter_count(from)) 345 return 0; 346 347 trace_netfs_write_iter(iocb, from); 348 netfs_stat(&netfs_n_wh_dio_write); 349 350 ret = netfs_start_io_direct(inode); 351 if (ret < 0) 352 return ret; 353 ret = generic_write_checks(iocb, from); 354 if (ret <= 0) 355 goto out; 356 ret = file_remove_privs(file); 357 if (ret < 0) 358 goto out; 359 ret = file_update_time(file); 360 if (ret < 0) 361 goto out; 362 if (iocb->ki_flags & IOCB_NOWAIT) { 363 /* We could block if there are any pages in the range. */ 364 ret = -EAGAIN; 365 if (filemap_range_has_page(mapping, pos, end)) 366 if (filemap_invalidate_inode(inode, true, pos, end)) 367 goto out; 368 } else { 369 ret = filemap_write_and_wait_range(mapping, pos, end); 370 if (ret < 0) 371 goto out; 372 } 373 374 /* 375 * After a write we want buffered reads to be sure to go to disk to get 376 * the new data. We invalidate clean cached page from the region we're 377 * about to write. We do this *before* the write so that we can return 378 * without clobbering -EIOCBQUEUED from ->direct_IO(). 379 */ 380 ret = filemap_invalidate_inode(inode, true, pos, end); 381 if (ret < 0) 382 goto out; 383 end = iocb->ki_pos + iov_iter_count(from); 384 spin_lock(&inode->i_lock); 385 if (end > ictx->_zero_point) 386 netfs_write_zero_point(inode, end); 387 spin_unlock(&inode->i_lock); 388 389 fscache_invalidate(netfs_i_cookie(ictx), NULL, i_size_read(inode), 390 FSCACHE_INVAL_DIO_WRITE); 391 ret = netfs_unbuffered_write_iter_locked(iocb, from, NULL); 392 out: 393 netfs_end_io_direct(inode); 394 return ret; 395 } 396 EXPORT_SYMBOL(netfs_unbuffered_write_iter); 397