direct_write.c

  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 */
 15static 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 */
 65static 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 */
 94static 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
172		if (stream->source == NETFS_UPLOAD_TO_SERVER &&
173		    wreq->netfs_ops->retry_request)
174			wreq->netfs_ops->retry_request(wreq, stream);
175
176		__clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
177		__clear_bit(NETFS_SREQ_BOUNDARY, &subreq->flags);
178		__clear_bit(NETFS_SREQ_FAILED, &subreq->flags);
179		subreq->io_iter		= wreq->buffer.iter;
180		subreq->start		= wreq->start + wreq->transferred;
181		subreq->len		= wreq->len   - wreq->transferred;
182		subreq->transferred	= 0;
183		subreq->retry_count	+= 1;
184		stream->sreq_max_len	= UINT_MAX;
185		stream->sreq_max_segs	= INT_MAX;
186
187		netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit);
188
189		if (stream->prepare_write) {
190			stream->prepare_write(subreq);
191			__set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags);
192			netfs_stat(&netfs_n_wh_retry_write_subreq);
193		} else {
194			struct iov_iter source;
195
196			netfs_reset_iter(subreq);
197			source = subreq->io_iter;
198			netfs_reissue_write(stream, subreq, &source);
199		}
200	}
201
202	netfs_unbuffered_write_done(wreq);
203	_leave(" = %d", ret);
204	return ret;
205}
206
207static void netfs_unbuffered_write_async(struct work_struct *work)
208{
209	struct netfs_io_request *wreq = container_of(work, struct netfs_io_request, work);
210
211	netfs_unbuffered_write(wreq);
212	netfs_put_request(wreq, netfs_rreq_trace_put_complete);
213}
214
215/*
216 * Perform an unbuffered write where we may have to do an RMW operation on an
217 * encrypted file.  This can also be used for direct I/O writes.
218 */
219ssize_t netfs_unbuffered_write_iter_locked(struct kiocb *iocb, struct iov_iter *iter,
220						  struct netfs_group *netfs_group)
221{
222	struct netfs_io_request *wreq;
223	unsigned long long start = iocb->ki_pos;
224	unsigned long long end = start + iov_iter_count(iter);
225	ssize_t ret, n;
226	size_t len = iov_iter_count(iter);
227	bool async = !is_sync_kiocb(iocb);
228
229	_enter("");
230
231	/* We're going to need a bounce buffer if what we transmit is going to
232	 * be different in some way to the source buffer, e.g. because it gets
233	 * encrypted/compressed or because it needs expanding to a block size.
234	 */
235	// TODO
236
237	_debug("uw %llx-%llx", start, end);
238
239	wreq = netfs_create_write_req(iocb->ki_filp->f_mapping, iocb->ki_filp, start,
240				      iocb->ki_flags & IOCB_DIRECT ?
241				      NETFS_DIO_WRITE : NETFS_UNBUFFERED_WRITE);
242	if (IS_ERR(wreq))
243		return PTR_ERR(wreq);
244
245	wreq->io_streams[0].avail = true;
246	trace_netfs_write(wreq, (iocb->ki_flags & IOCB_DIRECT ?
247				 netfs_write_trace_dio_write :
248				 netfs_write_trace_unbuffered_write));
249
250	{
251		/* If this is an async op and we're not using a bounce buffer,
252		 * we have to save the source buffer as the iterator is only
253		 * good until we return.  In such a case, extract an iterator
254		 * to represent as much of the the output buffer as we can
255		 * manage.  Note that the extraction might not be able to
256		 * allocate a sufficiently large bvec array and may shorten the
257		 * request.
258		 */
259		if (user_backed_iter(iter)) {
260			n = netfs_extract_user_iter(iter, len, &wreq->buffer.iter, 0);
261			if (n < 0) {
262				ret = n;
263				goto error_put;
264			}
265			wreq->direct_bv = (struct bio_vec *)wreq->buffer.iter.bvec;
266			wreq->direct_bv_count = n;
267			wreq->direct_bv_unpin = iov_iter_extract_will_pin(iter);
268		} else {
269			/* If this is a kernel-generated async DIO request,
270			 * assume that any resources the iterator points to
271			 * (eg. a bio_vec array) will persist till the end of
272			 * the op.
273			 */
274			wreq->buffer.iter = *iter;
275		}
276
277		wreq->len = iov_iter_count(&wreq->buffer.iter);
278	}
279
280	__set_bit(NETFS_RREQ_USE_IO_ITER, &wreq->flags);
281
282	/* Copy the data into the bounce buffer and encrypt it. */
283	// TODO
284
285	/* Dispatch the write. */
286	__set_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags);
287
288	if (async) {
289		INIT_WORK(&wreq->work, netfs_unbuffered_write_async);
290		wreq->iocb = iocb;
291		queue_work(system_dfl_wq, &wreq->work);
292		ret = -EIOCBQUEUED;
293	} else {
294		ret = netfs_unbuffered_write(wreq);
295		if (ret < 0) {
296			_debug("begin = %zd", ret);
297		} else {
298			iocb->ki_pos += wreq->transferred;
299			ret = wreq->transferred ?: wreq->error;
300		}
301
302		netfs_put_request(wreq, netfs_rreq_trace_put_complete);
303	}
304
305	netfs_put_request(wreq, netfs_rreq_trace_put_return);
306	return ret;
307
308error_put:
309	netfs_put_failed_request(wreq);
310	return ret;
311}
312EXPORT_SYMBOL(netfs_unbuffered_write_iter_locked);
313
314/**
315 * netfs_unbuffered_write_iter - Unbuffered write to a file
316 * @iocb: IO state structure
317 * @from: iov_iter with data to write
318 *
319 * Do an unbuffered write to a file, writing the data directly to the server
320 * and not lodging the data in the pagecache.
321 *
322 * Return:
323 * * Negative error code if no data has been written at all of
324 *   vfs_fsync_range() failed for a synchronous write
325 * * Number of bytes written, even for truncated writes
326 */
327ssize_t netfs_unbuffered_write_iter(struct kiocb *iocb, struct iov_iter *from)
328{
329	struct file *file = iocb->ki_filp;
330	struct address_space *mapping = file->f_mapping;
331	struct inode *inode = mapping->host;
332	struct netfs_inode *ictx = netfs_inode(inode);
333	ssize_t ret;
334	loff_t pos = iocb->ki_pos;
335	unsigned long long end = pos + iov_iter_count(from) - 1;
336
337	_enter("%llx,%zx,%llx", pos, iov_iter_count(from), i_size_read(inode));
338
339	if (!iov_iter_count(from))
340		return 0;
341
342	trace_netfs_write_iter(iocb, from);
343	netfs_stat(&netfs_n_wh_dio_write);
344
345	ret = netfs_start_io_direct(inode);
346	if (ret < 0)
347		return ret;
348	ret = generic_write_checks(iocb, from);
349	if (ret <= 0)
350		goto out;
351	ret = file_remove_privs(file);
352	if (ret < 0)
353		goto out;
354	ret = file_update_time(file);
355	if (ret < 0)
356		goto out;
357	if (iocb->ki_flags & IOCB_NOWAIT) {
358		/* We could block if there are any pages in the range. */
359		ret = -EAGAIN;
360		if (filemap_range_has_page(mapping, pos, end))
361			if (filemap_invalidate_inode(inode, true, pos, end))
362				goto out;
363	} else {
364		ret = filemap_write_and_wait_range(mapping, pos, end);
365		if (ret < 0)
366			goto out;
367	}
368
369	/*
370	 * After a write we want buffered reads to be sure to go to disk to get
371	 * the new data.  We invalidate clean cached page from the region we're
372	 * about to write.  We do this *before* the write so that we can return
373	 * without clobbering -EIOCBQUEUED from ->direct_IO().
374	 */
375	ret = filemap_invalidate_inode(inode, true, pos, end);
376	if (ret < 0)
377		goto out;
378	end = iocb->ki_pos + iov_iter_count(from);
379	if (end > ictx->zero_point)
380		ictx->zero_point = end;
381
382	fscache_invalidate(netfs_i_cookie(ictx), NULL, i_size_read(inode),
383			   FSCACHE_INVAL_DIO_WRITE);
384	ret = netfs_unbuffered_write_iter_locked(iocb, from, NULL);
385out:
386	netfs_end_io_direct(inode);
387	return ret;
388}
389EXPORT_SYMBOL(netfs_unbuffered_write_iter);