1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Network filesystem high-level (buffered) writeback.
4 * Copyright (C) 2024 Red Hat, Inc. All Rights Reserved.
8 * To support network filesystems with local caching, we manage a situation
9 * that can be envisioned like the following:
11 * +---+---+-----+-----+---+----------+
12 * Folios: | | | | | | |
13 * +---+---+-----+-----+---+----------+
15 * +------+------+ +----+----+
16 * Upload: | | |.....| | |
17 * (Stream 0) +------+------+ +----+----+
19 * +------+------+------+------+------+
21 * (Stream 1) +------+------+------+------+------+
23 * Where we have a sequence of folios of varying sizes that we need to overlay
24 * with multiple parallel streams of I/O requests, where the I/O requests in a
25 * stream may also be of various sizes (in cifs, for example, the sizes are
26 * negotiated with the server; in something like ceph, they may represent the
27 * sizes of storage objects).
29 * The sequence in each stream may contain gaps and noncontiguous subrequests
30 * may be glued together into single vectored write RPCs.
33 #include <linux/export.h>
36 #include <linux/pagemap.h>
40 * Kill all dirty folios in the event of an unrecoverable error, starting with
41 * a locked folio we've already obtained from writeback_iter().
43 static void netfs_kill_dirty_pages(struct address_space *mapping,
44 struct writeback_control *wbc,
50 enum netfs_folio_trace why = netfs_folio_trace_kill;
51 struct netfs_group *group = NULL;
52 struct netfs_folio *finfo = NULL;
55 priv = folio_detach_private(folio);
57 finfo = __netfs_folio_info(priv);
59 /* Kill folio from streaming write. */
60 group = finfo->netfs_group;
61 why = netfs_folio_trace_kill_s;
64 if (group == NETFS_FOLIO_COPY_TO_CACHE) {
65 /* Kill copy-to-cache folio */
66 why = netfs_folio_trace_kill_cc;
69 /* Kill folio with group */
70 why = netfs_folio_trace_kill_g;
75 trace_netfs_folio(folio, why);
77 folio_start_writeback(folio);
79 folio_end_writeback(folio);
81 netfs_put_group(group);
84 } while ((folio = writeback_iter(mapping, wbc, folio, &error)));
88 * Create a write request and set it up appropriately for the origin type.
90 struct netfs_io_request *netfs_create_write_req(struct address_space *mapping,
93 enum netfs_io_origin origin)
95 struct netfs_io_request *wreq;
96 struct netfs_inode *ictx;
97 bool is_buffered = (origin == NETFS_WRITEBACK ||
98 origin == NETFS_WRITETHROUGH ||
99 origin == NETFS_PGPRIV2_COPY_TO_CACHE);
101 wreq = netfs_alloc_request(mapping, file, start, 0, origin);
105 _enter("R=%x", wreq->debug_id);
107 ictx = netfs_inode(wreq->inode);
108 if (is_buffered && netfs_is_cache_enabled(ictx))
109 fscache_begin_write_operation(&wreq->cache_resources, netfs_i_cookie(ictx));
111 wreq->cleaned_to = wreq->start;
113 wreq->io_streams[0].stream_nr = 0;
114 wreq->io_streams[0].source = NETFS_UPLOAD_TO_SERVER;
115 wreq->io_streams[0].prepare_write = ictx->ops->prepare_write;
116 wreq->io_streams[0].issue_write = ictx->ops->issue_write;
117 wreq->io_streams[0].collected_to = start;
118 wreq->io_streams[0].transferred = LONG_MAX;
120 wreq->io_streams[1].stream_nr = 1;
121 wreq->io_streams[1].source = NETFS_WRITE_TO_CACHE;
122 wreq->io_streams[1].collected_to = start;
123 wreq->io_streams[1].transferred = LONG_MAX;
124 if (fscache_resources_valid(&wreq->cache_resources)) {
125 wreq->io_streams[1].avail = true;
126 wreq->io_streams[1].active = true;
127 wreq->io_streams[1].prepare_write = wreq->cache_resources.ops->prepare_write_subreq;
128 wreq->io_streams[1].issue_write = wreq->cache_resources.ops->issue_write;
135 * netfs_prepare_write_failed - Note write preparation failed
136 * @subreq: The subrequest to mark
138 * Mark a subrequest to note that preparation for write failed.
140 void netfs_prepare_write_failed(struct netfs_io_subrequest *subreq)
142 __set_bit(NETFS_SREQ_FAILED, &subreq->flags);
143 trace_netfs_sreq(subreq, netfs_sreq_trace_prep_failed);
145 EXPORT_SYMBOL(netfs_prepare_write_failed);
148 * Prepare a write subrequest. We need to allocate a new subrequest
149 * if we don't have one.
151 static void netfs_prepare_write(struct netfs_io_request *wreq,
152 struct netfs_io_stream *stream,
155 struct netfs_io_subrequest *subreq;
156 struct iov_iter *wreq_iter = &wreq->io_iter;
158 /* Make sure we don't point the iterator at a used-up folio_queue
159 * struct being used as a placeholder to prevent the queue from
160 * collapsing. In such a case, extend the queue.
162 if (iov_iter_is_folioq(wreq_iter) &&
163 wreq_iter->folioq_slot >= folioq_nr_slots(wreq_iter->folioq)) {
164 netfs_buffer_make_space(wreq);
167 subreq = netfs_alloc_subrequest(wreq);
168 subreq->source = stream->source;
169 subreq->start = start;
170 subreq->stream_nr = stream->stream_nr;
171 subreq->io_iter = *wreq_iter;
173 _enter("R=%x[%x]", wreq->debug_id, subreq->debug_index);
175 trace_netfs_sreq(subreq, netfs_sreq_trace_prepare);
177 stream->sreq_max_len = UINT_MAX;
178 stream->sreq_max_segs = INT_MAX;
179 switch (stream->source) {
180 case NETFS_UPLOAD_TO_SERVER:
181 netfs_stat(&netfs_n_wh_upload);
182 stream->sreq_max_len = wreq->wsize;
184 case NETFS_WRITE_TO_CACHE:
185 netfs_stat(&netfs_n_wh_write);
192 if (stream->prepare_write)
193 stream->prepare_write(subreq);
195 __set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags);
197 /* We add to the end of the list whilst the collector may be walking
198 * the list. The collector only goes nextwards and uses the lock to
199 * remove entries off of the front.
201 spin_lock_bh(&wreq->lock);
202 list_add_tail(&subreq->rreq_link, &stream->subrequests);
203 if (list_is_first(&subreq->rreq_link, &stream->subrequests)) {
204 stream->front = subreq;
205 if (!stream->active) {
206 stream->collected_to = stream->front->start;
207 /* Write list pointers before active flag */
208 smp_store_release(&stream->active, true);
212 spin_unlock_bh(&wreq->lock);
214 stream->construct = subreq;
218 * Set the I/O iterator for the filesystem/cache to use and dispatch the I/O
219 * operation. The operation may be asynchronous and should call
220 * netfs_write_subrequest_terminated() when complete.
222 static void netfs_do_issue_write(struct netfs_io_stream *stream,
223 struct netfs_io_subrequest *subreq)
225 struct netfs_io_request *wreq = subreq->rreq;
227 _enter("R=%x[%x],%zx", wreq->debug_id, subreq->debug_index, subreq->len);
229 if (test_bit(NETFS_SREQ_FAILED, &subreq->flags))
230 return netfs_write_subrequest_terminated(subreq, subreq->error, false);
232 trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
233 stream->issue_write(subreq);
236 void netfs_reissue_write(struct netfs_io_stream *stream,
237 struct netfs_io_subrequest *subreq,
238 struct iov_iter *source)
240 size_t size = subreq->len - subreq->transferred;
242 // TODO: Use encrypted buffer
243 subreq->io_iter = *source;
244 iov_iter_advance(source, size);
245 iov_iter_truncate(&subreq->io_iter, size);
247 subreq->retry_count++;
248 __clear_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags);
249 __set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags);
250 netfs_do_issue_write(stream, subreq);
253 void netfs_issue_write(struct netfs_io_request *wreq,
254 struct netfs_io_stream *stream)
256 struct netfs_io_subrequest *subreq = stream->construct;
260 stream->construct = NULL;
261 subreq->io_iter.count = subreq->len;
262 netfs_do_issue_write(stream, subreq);
266 * Add data to the write subrequest, dispatching each as we fill it up or if it
267 * is discontiguous with the previous. We only fill one part at a time so that
268 * we can avoid overrunning the credits obtained (cifs) and try to parallelise
269 * content-crypto preparation with network writes.
271 int netfs_advance_write(struct netfs_io_request *wreq,
272 struct netfs_io_stream *stream,
273 loff_t start, size_t len, bool to_eof)
275 struct netfs_io_subrequest *subreq = stream->construct;
278 if (!stream->avail) {
283 _enter("R=%x[%x]", wreq->debug_id, subreq ? subreq->debug_index : 0);
285 if (subreq && start != subreq->start + subreq->len) {
286 netfs_issue_write(wreq, stream);
290 if (!stream->construct)
291 netfs_prepare_write(wreq, stream, start);
292 subreq = stream->construct;
294 part = umin(stream->sreq_max_len - subreq->len, len);
295 _debug("part %zx/%zx %zx/%zx", subreq->len, stream->sreq_max_len, part, len);
298 stream->submit_extendable_to -= part;
300 if (subreq->len >= stream->sreq_max_len ||
301 subreq->nr_segs >= stream->sreq_max_segs ||
303 netfs_issue_write(wreq, stream);
311 * Write some of a pending folio data back to the server.
313 static int netfs_write_folio(struct netfs_io_request *wreq,
314 struct writeback_control *wbc,
317 struct netfs_io_stream *upload = &wreq->io_streams[0];
318 struct netfs_io_stream *cache = &wreq->io_streams[1];
319 struct netfs_io_stream *stream;
320 struct netfs_group *fgroup; /* TODO: Use this with ceph */
321 struct netfs_folio *finfo;
323 size_t fsize = folio_size(folio), flen = fsize, foff = 0;
324 loff_t fpos = folio_pos(folio), i_size;
325 bool to_eof = false, streamw = false;
330 /* netfs_perform_write() may shift i_size around the page or from out
331 * of the page to beyond it, but cannot move i_size into or through the
332 * page since we have it locked.
334 i_size = i_size_read(wreq->inode);
336 if (fpos >= i_size) {
337 /* mmap beyond eof. */
338 _debug("beyond eof");
339 folio_start_writeback(folio);
341 wreq->nr_group_rel += netfs_folio_written_back(folio);
342 netfs_put_group_many(wreq->group, wreq->nr_group_rel);
343 wreq->nr_group_rel = 0;
347 if (fpos + fsize > wreq->i_size)
348 wreq->i_size = i_size;
350 fgroup = netfs_folio_group(folio);
351 finfo = netfs_folio_info(folio);
353 foff = finfo->dirty_offset;
354 flen = foff + finfo->dirty_len;
358 if (wreq->origin == NETFS_WRITETHROUGH) {
360 if (flen > i_size - fpos)
361 flen = i_size - fpos;
362 } else if (flen > i_size - fpos) {
363 flen = i_size - fpos;
365 folio_zero_segment(folio, flen, fsize);
367 } else if (flen == i_size - fpos) {
372 _debug("folio %zx %zx %zx", foff, flen, fsize);
374 /* Deal with discontinuities in the stream of dirty pages. These can
375 * arise from a number of sources:
377 * (1) Intervening non-dirty pages from random-access writes, multiple
378 * flushers writing back different parts simultaneously and manual
381 * (2) Partially-written pages from write-streaming.
383 * (3) Pages that belong to a different write-back group (eg. Ceph
386 * (4) Actually-clean pages that were marked for write to the cache
387 * when they were read. Note that these appear as a special
390 if (fgroup == NETFS_FOLIO_COPY_TO_CACHE) {
391 netfs_issue_write(wreq, upload);
392 } else if (fgroup != wreq->group) {
393 /* We can't write this page to the server yet. */
394 kdebug("wrong group");
395 folio_redirty_for_writepage(wbc, folio);
397 netfs_issue_write(wreq, upload);
398 netfs_issue_write(wreq, cache);
403 netfs_issue_write(wreq, upload);
405 netfs_issue_write(wreq, cache);
407 /* Flip the page to the writeback state and unlock. If we're called
408 * from write-through, then the page has already been put into the wb
411 if (wreq->origin == NETFS_WRITEBACK)
412 folio_start_writeback(folio);
415 if (fgroup == NETFS_FOLIO_COPY_TO_CACHE) {
417 trace_netfs_folio(folio, netfs_folio_trace_cancel_copy);
418 netfs_issue_write(wreq, upload);
419 netfs_folio_written_back(folio);
422 trace_netfs_folio(folio, netfs_folio_trace_store_copy);
423 } else if (!upload->avail && !cache->avail) {
424 trace_netfs_folio(folio, netfs_folio_trace_cancel_store);
425 netfs_folio_written_back(folio);
427 } else if (!upload->construct) {
428 trace_netfs_folio(folio, netfs_folio_trace_store);
430 trace_netfs_folio(folio, netfs_folio_trace_store_plus);
433 /* Attach the folio to the rolling buffer. */
434 netfs_buffer_append_folio(wreq, folio, false);
436 /* Move the submission point forward to allow for write-streaming data
437 * not starting at the front of the page. We don't do write-streaming
438 * with the cache as the cache requires DIO alignment.
440 * Also skip uploading for data that's been read and just needs copying
443 for (int s = 0; s < NR_IO_STREAMS; s++) {
444 stream = &wreq->io_streams[s];
445 stream->submit_off = foff;
446 stream->submit_len = flen;
447 if ((stream->source == NETFS_WRITE_TO_CACHE && streamw) ||
448 (stream->source == NETFS_UPLOAD_TO_SERVER &&
449 fgroup == NETFS_FOLIO_COPY_TO_CACHE)) {
450 stream->submit_off = UINT_MAX;
451 stream->submit_len = 0;
455 /* Attach the folio to one or more subrequests. For a big folio, we
456 * could end up with thousands of subrequests if the wsize is small -
457 * but we might need to wait during the creation of subrequests for
458 * network resources (eg. SMB credits).
462 size_t lowest_off = ULONG_MAX;
465 /* Always add to the lowest-submitted stream first. */
466 for (int s = 0; s < NR_IO_STREAMS; s++) {
467 stream = &wreq->io_streams[s];
468 if (stream->submit_len > 0 &&
469 stream->submit_off < lowest_off) {
470 lowest_off = stream->submit_off;
477 stream = &wreq->io_streams[choose_s];
479 /* Advance the iterator(s). */
480 if (stream->submit_off > iter_off) {
481 iov_iter_advance(&wreq->io_iter, stream->submit_off - iter_off);
482 iter_off = stream->submit_off;
485 atomic64_set(&wreq->issued_to, fpos + stream->submit_off);
486 stream->submit_extendable_to = fsize - stream->submit_off;
487 part = netfs_advance_write(wreq, stream, fpos + stream->submit_off,
488 stream->submit_len, to_eof);
489 stream->submit_off += part;
490 if (part > stream->submit_len)
491 stream->submit_len = 0;
493 stream->submit_len -= part;
498 if (fsize > iter_off)
499 iov_iter_advance(&wreq->io_iter, fsize - iter_off);
500 atomic64_set(&wreq->issued_to, fpos + fsize);
503 kdebug("R=%x: No submit", wreq->debug_id);
505 if (foff + flen < fsize)
506 for (int s = 0; s < NR_IO_STREAMS; s++)
507 netfs_issue_write(wreq, &wreq->io_streams[s]);
514 * End the issuing of writes, letting the collector know we're done.
516 static void netfs_end_issue_write(struct netfs_io_request *wreq)
518 bool needs_poke = true;
520 smp_wmb(); /* Write subreq lists before ALL_QUEUED. */
521 set_bit(NETFS_RREQ_ALL_QUEUED, &wreq->flags);
523 for (int s = 0; s < NR_IO_STREAMS; s++) {
524 struct netfs_io_stream *stream = &wreq->io_streams[s];
528 if (!list_empty(&stream->subrequests))
530 netfs_issue_write(wreq, stream);
534 netfs_wake_write_collector(wreq, false);
538 * Write some of the pending data back to the server
540 int netfs_writepages(struct address_space *mapping,
541 struct writeback_control *wbc)
543 struct netfs_inode *ictx = netfs_inode(mapping->host);
544 struct netfs_io_request *wreq = NULL;
548 if (!mutex_trylock(&ictx->wb_lock)) {
549 if (wbc->sync_mode == WB_SYNC_NONE) {
550 netfs_stat(&netfs_n_wb_lock_skip);
553 netfs_stat(&netfs_n_wb_lock_wait);
554 mutex_lock(&ictx->wb_lock);
557 /* Need the first folio to be able to set up the op. */
558 folio = writeback_iter(mapping, wbc, NULL, &error);
562 wreq = netfs_create_write_req(mapping, NULL, folio_pos(folio), NETFS_WRITEBACK);
564 error = PTR_ERR(wreq);
568 trace_netfs_write(wreq, netfs_write_trace_writeback);
569 netfs_stat(&netfs_n_wh_writepages);
572 _debug("wbiter %lx %llx", folio->index, atomic64_read(&wreq->issued_to));
574 /* It appears we don't have to handle cyclic writeback wrapping. */
575 WARN_ON_ONCE(wreq && folio_pos(folio) < atomic64_read(&wreq->issued_to));
577 if (netfs_folio_group(folio) != NETFS_FOLIO_COPY_TO_CACHE &&
578 unlikely(!test_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags))) {
579 set_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags);
580 wreq->netfs_ops->begin_writeback(wreq);
583 error = netfs_write_folio(wreq, wbc, folio);
586 } while ((folio = writeback_iter(mapping, wbc, folio, &error)));
588 netfs_end_issue_write(wreq);
590 mutex_unlock(&ictx->wb_lock);
592 netfs_put_request(wreq, false, netfs_rreq_trace_put_return);
593 _leave(" = %d", error);
597 netfs_kill_dirty_pages(mapping, wbc, folio);
599 mutex_unlock(&ictx->wb_lock);
600 _leave(" = %d", error);
603 EXPORT_SYMBOL(netfs_writepages);
606 * Begin a write operation for writing through the pagecache.
608 struct netfs_io_request *netfs_begin_writethrough(struct kiocb *iocb, size_t len)
610 struct netfs_io_request *wreq = NULL;
611 struct netfs_inode *ictx = netfs_inode(file_inode(iocb->ki_filp));
613 mutex_lock(&ictx->wb_lock);
615 wreq = netfs_create_write_req(iocb->ki_filp->f_mapping, iocb->ki_filp,
616 iocb->ki_pos, NETFS_WRITETHROUGH);
618 mutex_unlock(&ictx->wb_lock);
622 wreq->io_streams[0].avail = true;
623 trace_netfs_write(wreq, netfs_write_trace_writethrough);
628 * Advance the state of the write operation used when writing through the
629 * pagecache. Data has been copied into the pagecache that we need to append
630 * to the request. If we've added more than wsize then we need to create a new
633 int netfs_advance_writethrough(struct netfs_io_request *wreq, struct writeback_control *wbc,
634 struct folio *folio, size_t copied, bool to_page_end,
635 struct folio **writethrough_cache)
637 _enter("R=%x ic=%zu ws=%u cp=%zu tp=%u",
638 wreq->debug_id, wreq->iter.count, wreq->wsize, copied, to_page_end);
640 if (!*writethrough_cache) {
641 if (folio_test_dirty(folio))
643 folio_clear_dirty_for_io(folio);
645 /* We can make multiple writes to the folio... */
646 folio_start_writeback(folio);
648 trace_netfs_folio(folio, netfs_folio_trace_wthru);
650 trace_netfs_folio(folio, netfs_folio_trace_wthru_plus);
651 *writethrough_cache = folio;
658 *writethrough_cache = NULL;
659 return netfs_write_folio(wreq, wbc, folio);
663 * End a write operation used when writing through the pagecache.
665 int netfs_end_writethrough(struct netfs_io_request *wreq, struct writeback_control *wbc,
666 struct folio *writethrough_cache)
668 struct netfs_inode *ictx = netfs_inode(wreq->inode);
671 _enter("R=%x", wreq->debug_id);
673 if (writethrough_cache)
674 netfs_write_folio(wreq, wbc, writethrough_cache);
676 netfs_end_issue_write(wreq);
678 mutex_unlock(&ictx->wb_lock);
683 wait_on_bit(&wreq->flags, NETFS_RREQ_IN_PROGRESS, TASK_UNINTERRUPTIBLE);
686 netfs_put_request(wreq, false, netfs_rreq_trace_put_return);
691 * Write data to the server without going through the pagecache and without
692 * writing it to the local cache.
694 int netfs_unbuffered_write(struct netfs_io_request *wreq, bool may_wait, size_t len)
696 struct netfs_io_stream *upload = &wreq->io_streams[0];
698 loff_t start = wreq->start;
703 if (wreq->origin == NETFS_DIO_WRITE)
704 inode_dio_begin(wreq->inode);
707 // TODO: Prepare content encryption
709 _debug("unbuffered %zx", len);
710 part = netfs_advance_write(wreq, upload, start, len, false);
713 iov_iter_advance(&wreq->io_iter, part);
714 if (test_bit(NETFS_RREQ_PAUSE, &wreq->flags)) {
715 trace_netfs_rreq(wreq, netfs_rreq_trace_wait_pause);
716 wait_on_bit(&wreq->flags, NETFS_RREQ_PAUSE, TASK_UNINTERRUPTIBLE);
718 if (test_bit(NETFS_RREQ_FAILED, &wreq->flags))
722 netfs_end_issue_write(wreq);
723 _leave(" = %d", error);