1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Network filesystem high-level (buffered) writeback.
3 *
4 * Copyright (C) 2024 Red Hat, Inc. All Rights Reserved.
5 * Written by David Howells (dhowells@redhat.com)
6 *
7 *
8 * To support network filesystems with local caching, we manage a situation
9 * that can be envisioned like the following:
10 *
11 * +---+---+-----+-----+---+----------+
12 * Folios: | | | | | | |
13 * +---+---+-----+-----+---+----------+
14 *
15 * +------+------+ +----+----+
16 * Upload: | | |.....| | |
17 * (Stream 0) +------+------+ +----+----+
18 *
19 * +------+------+------+------+------+
20 * Cache: | | | | | |
21 * (Stream 1) +------+------+------+------+------+
22 *
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).
28 *
29 * The sequence in each stream may contain gaps and noncontiguous subrequests
30 * may be glued together into single vectored write RPCs.
31 */
32
33 #include <linux/export.h>
34 #include <linux/fs.h>
35 #include <linux/mm.h>
36 #include <linux/pagemap.h>
37 #include "internal.h"
38
39 /*
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().
42 */
netfs_kill_dirty_pages(struct address_space * mapping,struct writeback_control * wbc,struct folio * folio)43 static void netfs_kill_dirty_pages(struct address_space *mapping,
44 struct writeback_control *wbc,
45 struct folio *folio)
46 {
47 int error = 0;
48
49 do {
50 enum netfs_folio_trace why = netfs_folio_trace_kill;
51 struct netfs_group *group = NULL;
52 struct netfs_folio *finfo = NULL;
53 void *priv;
54
55 priv = folio_detach_private(folio);
56 if (priv) {
57 finfo = __netfs_folio_info(priv);
58 if (finfo) {
59 /* Kill folio from streaming write. */
60 group = finfo->netfs_group;
61 why = netfs_folio_trace_kill_s;
62 } else {
63 group = priv;
64 if (group == NETFS_FOLIO_COPY_TO_CACHE) {
65 /* Kill copy-to-cache folio */
66 why = netfs_folio_trace_kill_cc;
67 group = NULL;
68 } else {
69 /* Kill folio with group */
70 why = netfs_folio_trace_kill_g;
71 }
72 }
73 }
74
75 trace_netfs_folio(folio, why);
76
77 folio_start_writeback(folio);
78 folio_unlock(folio);
79 folio_end_writeback(folio);
80
81 netfs_put_group(group);
82 kfree(finfo);
83
84 } while ((folio = writeback_iter(mapping, wbc, folio, &error)));
85 }
86
87 /*
88 * Create a write request and set it up appropriately for the origin type.
89 */
netfs_create_write_req(struct address_space * mapping,struct file * file,loff_t start,enum netfs_io_origin origin)90 struct netfs_io_request *netfs_create_write_req(struct address_space *mapping,
91 struct file *file,
92 loff_t start,
93 enum netfs_io_origin origin)
94 {
95 struct netfs_io_request *wreq;
96 struct netfs_inode *ictx;
97 bool is_cacheable = (origin == NETFS_WRITEBACK ||
98 origin == NETFS_WRITEBACK_SINGLE ||
99 origin == NETFS_WRITETHROUGH ||
100 origin == NETFS_PGPRIV2_COPY_TO_CACHE);
101
102 wreq = netfs_alloc_request(mapping, file, start, 0, origin);
103 if (IS_ERR(wreq))
104 return wreq;
105
106 _enter("R=%x", wreq->debug_id);
107
108 ictx = netfs_inode(wreq->inode);
109 if (is_cacheable && netfs_is_cache_enabled(ictx))
110 fscache_begin_write_operation(&wreq->cache_resources, netfs_i_cookie(ictx));
111 if (rolling_buffer_init(&wreq->buffer, wreq->debug_id, ITER_SOURCE) < 0)
112 goto nomem;
113
114 wreq->cleaned_to = wreq->start;
115
116 wreq->io_streams[0].stream_nr = 0;
117 wreq->io_streams[0].source = NETFS_UPLOAD_TO_SERVER;
118 wreq->io_streams[0].prepare_write = ictx->ops->prepare_write;
119 wreq->io_streams[0].issue_write = ictx->ops->issue_write;
120 wreq->io_streams[0].collected_to = start;
121 wreq->io_streams[0].transferred = 0;
122
123 wreq->io_streams[1].stream_nr = 1;
124 wreq->io_streams[1].source = NETFS_WRITE_TO_CACHE;
125 wreq->io_streams[1].collected_to = start;
126 wreq->io_streams[1].transferred = 0;
127 if (fscache_resources_valid(&wreq->cache_resources)) {
128 wreq->io_streams[1].avail = true;
129 wreq->io_streams[1].active = true;
130 wreq->io_streams[1].prepare_write = wreq->cache_resources.ops->prepare_write_subreq;
131 wreq->io_streams[1].issue_write = wreq->cache_resources.ops->issue_write;
132 }
133
134 return wreq;
135 nomem:
136 netfs_put_failed_request(wreq);
137 return ERR_PTR(-ENOMEM);
138 }
139
140 /**
141 * netfs_prepare_write_failed - Note write preparation failed
142 * @subreq: The subrequest to mark
143 *
144 * Mark a subrequest to note that preparation for write failed.
145 */
netfs_prepare_write_failed(struct netfs_io_subrequest * subreq)146 void netfs_prepare_write_failed(struct netfs_io_subrequest *subreq)
147 {
148 __set_bit(NETFS_SREQ_FAILED, &subreq->flags);
149 trace_netfs_sreq(subreq, netfs_sreq_trace_prep_failed);
150 }
151 EXPORT_SYMBOL(netfs_prepare_write_failed);
152
153 /*
154 * Prepare a write subrequest. We need to allocate a new subrequest
155 * if we don't have one.
156 */
netfs_prepare_write(struct netfs_io_request * wreq,struct netfs_io_stream * stream,loff_t start)157 void netfs_prepare_write(struct netfs_io_request *wreq,
158 struct netfs_io_stream *stream,
159 loff_t start)
160 {
161 struct netfs_io_subrequest *subreq;
162 struct iov_iter *wreq_iter = &wreq->buffer.iter;
163
164 /* Make sure we don't point the iterator at a used-up folio_queue
165 * struct being used as a placeholder to prevent the queue from
166 * collapsing. In such a case, extend the queue.
167 */
168 if (iov_iter_is_folioq(wreq_iter) &&
169 wreq_iter->folioq_slot >= folioq_nr_slots(wreq_iter->folioq))
170 rolling_buffer_make_space(&wreq->buffer);
171
172 subreq = netfs_alloc_subrequest(wreq);
173 subreq->source = stream->source;
174 subreq->start = start;
175 subreq->stream_nr = stream->stream_nr;
176 subreq->io_iter = *wreq_iter;
177
178 _enter("R=%x[%x]", wreq->debug_id, subreq->debug_index);
179
180 trace_netfs_sreq(subreq, netfs_sreq_trace_prepare);
181
182 stream->sreq_max_len = UINT_MAX;
183 stream->sreq_max_segs = INT_MAX;
184 switch (stream->source) {
185 case NETFS_UPLOAD_TO_SERVER:
186 netfs_stat(&netfs_n_wh_upload);
187 stream->sreq_max_len = wreq->wsize;
188 break;
189 case NETFS_WRITE_TO_CACHE:
190 netfs_stat(&netfs_n_wh_write);
191 break;
192 default:
193 WARN_ON_ONCE(1);
194 break;
195 }
196
197 if (stream->prepare_write)
198 stream->prepare_write(subreq);
199
200 __set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags);
201
202 /* We add to the end of the list whilst the collector may be walking
203 * the list. The collector only goes nextwards and uses the lock to
204 * remove entries off of the front.
205 */
206 spin_lock(&wreq->lock);
207 list_add_tail(&subreq->rreq_link, &stream->subrequests);
208 if (list_is_first(&subreq->rreq_link, &stream->subrequests)) {
209 stream->front = subreq;
210 if (!stream->active) {
211 stream->collected_to = stream->front->start;
212 /* Write list pointers before active flag */
213 smp_store_release(&stream->active, true);
214 }
215 }
216
217 spin_unlock(&wreq->lock);
218
219 stream->construct = subreq;
220 }
221
222 /*
223 * Set the I/O iterator for the filesystem/cache to use and dispatch the I/O
224 * operation. The operation may be asynchronous and should call
225 * netfs_write_subrequest_terminated() when complete.
226 */
netfs_do_issue_write(struct netfs_io_stream * stream,struct netfs_io_subrequest * subreq)227 static void netfs_do_issue_write(struct netfs_io_stream *stream,
228 struct netfs_io_subrequest *subreq)
229 {
230 struct netfs_io_request *wreq = subreq->rreq;
231
232 _enter("R=%x[%x],%zx", wreq->debug_id, subreq->debug_index, subreq->len);
233
234 if (test_bit(NETFS_SREQ_FAILED, &subreq->flags))
235 return netfs_write_subrequest_terminated(subreq, subreq->error);
236
237 trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
238 stream->issue_write(subreq);
239 }
240
netfs_reissue_write(struct netfs_io_stream * stream,struct netfs_io_subrequest * subreq,struct iov_iter * source)241 void netfs_reissue_write(struct netfs_io_stream *stream,
242 struct netfs_io_subrequest *subreq,
243 struct iov_iter *source)
244 {
245 size_t size = subreq->len - subreq->transferred;
246
247 // TODO: Use encrypted buffer
248 subreq->io_iter = *source;
249 iov_iter_advance(source, size);
250 iov_iter_truncate(&subreq->io_iter, size);
251
252 subreq->retry_count++;
253 subreq->error = 0;
254 __clear_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags);
255 __set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags);
256 netfs_stat(&netfs_n_wh_retry_write_subreq);
257 netfs_do_issue_write(stream, subreq);
258 }
259
netfs_issue_write(struct netfs_io_request * wreq,struct netfs_io_stream * stream)260 void netfs_issue_write(struct netfs_io_request *wreq,
261 struct netfs_io_stream *stream)
262 {
263 struct netfs_io_subrequest *subreq = stream->construct;
264
265 if (!subreq)
266 return;
267 stream->construct = NULL;
268 subreq->io_iter.count = subreq->len;
269 netfs_do_issue_write(stream, subreq);
270 }
271
272 /*
273 * Add data to the write subrequest, dispatching each as we fill it up or if it
274 * is discontiguous with the previous. We only fill one part at a time so that
275 * we can avoid overrunning the credits obtained (cifs) and try to parallelise
276 * content-crypto preparation with network writes.
277 */
netfs_advance_write(struct netfs_io_request * wreq,struct netfs_io_stream * stream,loff_t start,size_t len,bool to_eof)278 size_t netfs_advance_write(struct netfs_io_request *wreq,
279 struct netfs_io_stream *stream,
280 loff_t start, size_t len, bool to_eof)
281 {
282 struct netfs_io_subrequest *subreq = stream->construct;
283 size_t part;
284
285 if (!stream->avail) {
286 _leave("no write");
287 return len;
288 }
289
290 _enter("R=%x[%x]", wreq->debug_id, subreq ? subreq->debug_index : 0);
291
292 if (subreq && start != subreq->start + subreq->len) {
293 netfs_issue_write(wreq, stream);
294 subreq = NULL;
295 }
296
297 if (!stream->construct)
298 netfs_prepare_write(wreq, stream, start);
299 subreq = stream->construct;
300
301 part = umin(stream->sreq_max_len - subreq->len, len);
302 _debug("part %zx/%zx %zx/%zx", subreq->len, stream->sreq_max_len, part, len);
303 subreq->len += part;
304 subreq->nr_segs++;
305 stream->submit_extendable_to -= part;
306
307 if (subreq->len >= stream->sreq_max_len ||
308 subreq->nr_segs >= stream->sreq_max_segs ||
309 to_eof) {
310 netfs_issue_write(wreq, stream);
311 subreq = NULL;
312 }
313
314 return part;
315 }
316
317 /*
318 * Write some of a pending folio data back to the server.
319 */
netfs_write_folio(struct netfs_io_request * wreq,struct writeback_control * wbc,struct folio * folio)320 static int netfs_write_folio(struct netfs_io_request *wreq,
321 struct writeback_control *wbc,
322 struct folio *folio)
323 {
324 struct netfs_io_stream *upload = &wreq->io_streams[0];
325 struct netfs_io_stream *cache = &wreq->io_streams[1];
326 struct netfs_io_stream *stream;
327 struct netfs_group *fgroup; /* TODO: Use this with ceph */
328 struct netfs_folio *finfo;
329 size_t iter_off = 0;
330 size_t fsize = folio_size(folio), flen = fsize, foff = 0;
331 loff_t fpos = folio_pos(folio), i_size;
332 bool to_eof = false, streamw = false;
333 bool debug = false;
334
335 _enter("");
336
337 if (rolling_buffer_make_space(&wreq->buffer) < 0)
338 return -ENOMEM;
339
340 /* netfs_perform_write() may shift i_size around the page or from out
341 * of the page to beyond it, but cannot move i_size into or through the
342 * page since we have it locked.
343 */
344 i_size = i_size_read(wreq->inode);
345
346 if (fpos >= i_size) {
347 /* mmap beyond eof. */
348 _debug("beyond eof");
349 folio_start_writeback(folio);
350 folio_unlock(folio);
351 wreq->nr_group_rel += netfs_folio_written_back(folio);
352 netfs_put_group_many(wreq->group, wreq->nr_group_rel);
353 wreq->nr_group_rel = 0;
354 return 0;
355 }
356
357 if (fpos + fsize > wreq->i_size)
358 wreq->i_size = i_size;
359
360 fgroup = netfs_folio_group(folio);
361 finfo = netfs_folio_info(folio);
362 if (finfo) {
363 foff = finfo->dirty_offset;
364 flen = foff + finfo->dirty_len;
365 streamw = true;
366 }
367
368 if (wreq->origin == NETFS_WRITETHROUGH) {
369 to_eof = false;
370 if (flen > i_size - fpos)
371 flen = i_size - fpos;
372 } else if (flen > i_size - fpos) {
373 flen = i_size - fpos;
374 if (!streamw)
375 folio_zero_segment(folio, flen, fsize);
376 to_eof = true;
377 } else if (flen == i_size - fpos) {
378 to_eof = true;
379 }
380 flen -= foff;
381
382 _debug("folio %zx %zx %zx", foff, flen, fsize);
383
384 /* Deal with discontinuities in the stream of dirty pages. These can
385 * arise from a number of sources:
386 *
387 * (1) Intervening non-dirty pages from random-access writes, multiple
388 * flushers writing back different parts simultaneously and manual
389 * syncing.
390 *
391 * (2) Partially-written pages from write-streaming.
392 *
393 * (3) Pages that belong to a different write-back group (eg. Ceph
394 * snapshots).
395 *
396 * (4) Actually-clean pages that were marked for write to the cache
397 * when they were read. Note that these appear as a special
398 * write-back group.
399 */
400 if (fgroup == NETFS_FOLIO_COPY_TO_CACHE) {
401 netfs_issue_write(wreq, upload);
402 } else if (fgroup != wreq->group) {
403 /* We can't write this page to the server yet. */
404 kdebug("wrong group");
405 folio_redirty_for_writepage(wbc, folio);
406 folio_unlock(folio);
407 netfs_issue_write(wreq, upload);
408 netfs_issue_write(wreq, cache);
409 return 0;
410 }
411
412 if (foff > 0)
413 netfs_issue_write(wreq, upload);
414 if (streamw)
415 netfs_issue_write(wreq, cache);
416
417 /* Flip the page to the writeback state and unlock. If we're called
418 * from write-through, then the page has already been put into the wb
419 * state.
420 */
421 if (wreq->origin == NETFS_WRITEBACK)
422 folio_start_writeback(folio);
423 folio_unlock(folio);
424
425 if (fgroup == NETFS_FOLIO_COPY_TO_CACHE) {
426 if (!cache->avail) {
427 trace_netfs_folio(folio, netfs_folio_trace_cancel_copy);
428 netfs_issue_write(wreq, upload);
429 netfs_folio_written_back(folio);
430 return 0;
431 }
432 trace_netfs_folio(folio, netfs_folio_trace_store_copy);
433 } else if (!upload->avail && !cache->avail) {
434 trace_netfs_folio(folio, netfs_folio_trace_cancel_store);
435 netfs_folio_written_back(folio);
436 return 0;
437 } else if (!upload->construct) {
438 trace_netfs_folio(folio, netfs_folio_trace_store);
439 } else {
440 trace_netfs_folio(folio, netfs_folio_trace_store_plus);
441 }
442
443 /* Attach the folio to the rolling buffer. */
444 rolling_buffer_append(&wreq->buffer, folio, 0);
445
446 /* Move the submission point forward to allow for write-streaming data
447 * not starting at the front of the page. We don't do write-streaming
448 * with the cache as the cache requires DIO alignment.
449 *
450 * Also skip uploading for data that's been read and just needs copying
451 * to the cache.
452 */
453 for (int s = 0; s < NR_IO_STREAMS; s++) {
454 stream = &wreq->io_streams[s];
455 stream->submit_off = foff;
456 stream->submit_len = flen;
457 if (!stream->avail ||
458 (stream->source == NETFS_WRITE_TO_CACHE && streamw) ||
459 (stream->source == NETFS_UPLOAD_TO_SERVER &&
460 fgroup == NETFS_FOLIO_COPY_TO_CACHE)) {
461 stream->submit_off = UINT_MAX;
462 stream->submit_len = 0;
463 }
464 }
465
466 /* Attach the folio to one or more subrequests. For a big folio, we
467 * could end up with thousands of subrequests if the wsize is small -
468 * but we might need to wait during the creation of subrequests for
469 * network resources (eg. SMB credits).
470 */
471 for (;;) {
472 ssize_t part;
473 size_t lowest_off = ULONG_MAX;
474 int choose_s = -1;
475
476 /* Always add to the lowest-submitted stream first. */
477 for (int s = 0; s < NR_IO_STREAMS; s++) {
478 stream = &wreq->io_streams[s];
479 if (stream->submit_len > 0 &&
480 stream->submit_off < lowest_off) {
481 lowest_off = stream->submit_off;
482 choose_s = s;
483 }
484 }
485
486 if (choose_s < 0)
487 break;
488 stream = &wreq->io_streams[choose_s];
489
490 /* Advance the iterator(s). */
491 if (stream->submit_off > iter_off) {
492 rolling_buffer_advance(&wreq->buffer, stream->submit_off - iter_off);
493 iter_off = stream->submit_off;
494 }
495
496 atomic64_set(&wreq->issued_to, fpos + stream->submit_off);
497 stream->submit_extendable_to = fsize - stream->submit_off;
498 part = netfs_advance_write(wreq, stream, fpos + stream->submit_off,
499 stream->submit_len, to_eof);
500 stream->submit_off += part;
501 if (part > stream->submit_len)
502 stream->submit_len = 0;
503 else
504 stream->submit_len -= part;
505 if (part > 0)
506 debug = true;
507 }
508
509 if (fsize > iter_off)
510 rolling_buffer_advance(&wreq->buffer, fsize - iter_off);
511 atomic64_set(&wreq->issued_to, fpos + fsize);
512
513 if (!debug)
514 kdebug("R=%x: No submit", wreq->debug_id);
515
516 if (foff + flen < fsize)
517 for (int s = 0; s < NR_IO_STREAMS; s++)
518 netfs_issue_write(wreq, &wreq->io_streams[s]);
519
520 _leave(" = 0");
521 return 0;
522 }
523
524 /*
525 * End the issuing of writes, letting the collector know we're done.
526 */
netfs_end_issue_write(struct netfs_io_request * wreq)527 static void netfs_end_issue_write(struct netfs_io_request *wreq)
528 {
529 bool needs_poke = true;
530
531 smp_wmb(); /* Write subreq lists before ALL_QUEUED. */
532 set_bit(NETFS_RREQ_ALL_QUEUED, &wreq->flags);
533
534 for (int s = 0; s < NR_IO_STREAMS; s++) {
535 struct netfs_io_stream *stream = &wreq->io_streams[s];
536
537 if (!stream->active)
538 continue;
539 if (!list_empty(&stream->subrequests))
540 needs_poke = false;
541 netfs_issue_write(wreq, stream);
542 }
543
544 if (needs_poke)
545 netfs_wake_collector(wreq);
546 }
547
548 /*
549 * Write some of the pending data back to the server
550 */
netfs_writepages(struct address_space * mapping,struct writeback_control * wbc)551 int netfs_writepages(struct address_space *mapping,
552 struct writeback_control *wbc)
553 {
554 struct netfs_inode *ictx = netfs_inode(mapping->host);
555 struct netfs_io_request *wreq = NULL;
556 struct folio *folio;
557 int error = 0;
558
559 if (!mutex_trylock(&ictx->wb_lock)) {
560 if (wbc->sync_mode == WB_SYNC_NONE) {
561 netfs_stat(&netfs_n_wb_lock_skip);
562 return 0;
563 }
564 netfs_stat(&netfs_n_wb_lock_wait);
565 mutex_lock(&ictx->wb_lock);
566 }
567
568 /* Need the first folio to be able to set up the op. */
569 folio = writeback_iter(mapping, wbc, NULL, &error);
570 if (!folio)
571 goto out;
572
573 wreq = netfs_create_write_req(mapping, NULL, folio_pos(folio), NETFS_WRITEBACK);
574 if (IS_ERR(wreq)) {
575 error = PTR_ERR(wreq);
576 goto couldnt_start;
577 }
578
579 __set_bit(NETFS_RREQ_OFFLOAD_COLLECTION, &wreq->flags);
580 trace_netfs_write(wreq, netfs_write_trace_writeback);
581 netfs_stat(&netfs_n_wh_writepages);
582
583 do {
584 _debug("wbiter %lx %llx", folio->index, atomic64_read(&wreq->issued_to));
585
586 /* It appears we don't have to handle cyclic writeback wrapping. */
587 WARN_ON_ONCE(wreq && folio_pos(folio) < atomic64_read(&wreq->issued_to));
588
589 if (netfs_folio_group(folio) != NETFS_FOLIO_COPY_TO_CACHE &&
590 unlikely(!test_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags))) {
591 set_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags);
592 wreq->netfs_ops->begin_writeback(wreq);
593 }
594
595 error = netfs_write_folio(wreq, wbc, folio);
596 if (error < 0)
597 break;
598 } while ((folio = writeback_iter(mapping, wbc, folio, &error)));
599
600 netfs_end_issue_write(wreq);
601
602 mutex_unlock(&ictx->wb_lock);
603 netfs_wake_collector(wreq);
604
605 netfs_put_request(wreq, netfs_rreq_trace_put_return);
606 _leave(" = %d", error);
607 return error;
608
609 couldnt_start:
610 netfs_kill_dirty_pages(mapping, wbc, folio);
611 out:
612 mutex_unlock(&ictx->wb_lock);
613 _leave(" = %d", error);
614 return error;
615 }
616 EXPORT_SYMBOL(netfs_writepages);
617
618 /*
619 * Begin a write operation for writing through the pagecache.
620 */
netfs_begin_writethrough(struct kiocb * iocb,size_t len)621 struct netfs_io_request *netfs_begin_writethrough(struct kiocb *iocb, size_t len)
622 {
623 struct netfs_io_request *wreq = NULL;
624 struct netfs_inode *ictx = netfs_inode(file_inode(iocb->ki_filp));
625
626 mutex_lock(&ictx->wb_lock);
627
628 wreq = netfs_create_write_req(iocb->ki_filp->f_mapping, iocb->ki_filp,
629 iocb->ki_pos, NETFS_WRITETHROUGH);
630 if (IS_ERR(wreq)) {
631 mutex_unlock(&ictx->wb_lock);
632 return wreq;
633 }
634
635 wreq->io_streams[0].avail = true;
636 trace_netfs_write(wreq, netfs_write_trace_writethrough);
637 return wreq;
638 }
639
640 /*
641 * Advance the state of the write operation used when writing through the
642 * pagecache. Data has been copied into the pagecache that we need to append
643 * to the request. If we've added more than wsize then we need to create a new
644 * subrequest.
645 */
netfs_advance_writethrough(struct netfs_io_request * wreq,struct writeback_control * wbc,struct folio * folio,size_t copied,bool to_page_end,struct folio ** writethrough_cache)646 int netfs_advance_writethrough(struct netfs_io_request *wreq, struct writeback_control *wbc,
647 struct folio *folio, size_t copied, bool to_page_end,
648 struct folio **writethrough_cache)
649 {
650 _enter("R=%x ic=%zu ws=%u cp=%zu tp=%u",
651 wreq->debug_id, wreq->buffer.iter.count, wreq->wsize, copied, to_page_end);
652
653 if (!*writethrough_cache) {
654 if (folio_test_dirty(folio))
655 /* Sigh. mmap. */
656 folio_clear_dirty_for_io(folio);
657
658 /* We can make multiple writes to the folio... */
659 folio_start_writeback(folio);
660 if (wreq->len == 0)
661 trace_netfs_folio(folio, netfs_folio_trace_wthru);
662 else
663 trace_netfs_folio(folio, netfs_folio_trace_wthru_plus);
664 *writethrough_cache = folio;
665 }
666
667 wreq->len += copied;
668 if (!to_page_end)
669 return 0;
670
671 *writethrough_cache = NULL;
672 return netfs_write_folio(wreq, wbc, folio);
673 }
674
675 /*
676 * End a write operation used when writing through the pagecache.
677 */
netfs_end_writethrough(struct netfs_io_request * wreq,struct writeback_control * wbc,struct folio * writethrough_cache)678 ssize_t netfs_end_writethrough(struct netfs_io_request *wreq, struct writeback_control *wbc,
679 struct folio *writethrough_cache)
680 {
681 struct netfs_inode *ictx = netfs_inode(wreq->inode);
682 ssize_t ret;
683
684 _enter("R=%x", wreq->debug_id);
685
686 if (writethrough_cache)
687 netfs_write_folio(wreq, wbc, writethrough_cache);
688
689 netfs_end_issue_write(wreq);
690
691 mutex_unlock(&ictx->wb_lock);
692
693 if (wreq->iocb)
694 ret = -EIOCBQUEUED;
695 else
696 ret = netfs_wait_for_write(wreq);
697 netfs_put_request(wreq, netfs_rreq_trace_put_return);
698 return ret;
699 }
700
701 /*
702 * Write some of a pending folio data back to the server and/or the cache.
703 */
netfs_write_folio_single(struct netfs_io_request * wreq,struct folio * folio)704 static int netfs_write_folio_single(struct netfs_io_request *wreq,
705 struct folio *folio)
706 {
707 struct netfs_io_stream *upload = &wreq->io_streams[0];
708 struct netfs_io_stream *cache = &wreq->io_streams[1];
709 struct netfs_io_stream *stream;
710 size_t iter_off = 0;
711 size_t fsize = folio_size(folio), flen;
712 loff_t fpos = folio_pos(folio);
713 bool to_eof = false;
714 bool no_debug = false;
715
716 _enter("");
717
718 flen = folio_size(folio);
719 if (flen > wreq->i_size - fpos) {
720 flen = wreq->i_size - fpos;
721 folio_zero_segment(folio, flen, fsize);
722 to_eof = true;
723 } else if (flen == wreq->i_size - fpos) {
724 to_eof = true;
725 }
726
727 _debug("folio %zx/%zx", flen, fsize);
728
729 if (!upload->avail && !cache->avail) {
730 trace_netfs_folio(folio, netfs_folio_trace_cancel_store);
731 return 0;
732 }
733
734 if (!upload->construct)
735 trace_netfs_folio(folio, netfs_folio_trace_store);
736 else
737 trace_netfs_folio(folio, netfs_folio_trace_store_plus);
738
739 /* Attach the folio to the rolling buffer. */
740 folio_get(folio);
741 rolling_buffer_append(&wreq->buffer, folio, NETFS_ROLLBUF_PUT_MARK);
742
743 /* Move the submission point forward to allow for write-streaming data
744 * not starting at the front of the page. We don't do write-streaming
745 * with the cache as the cache requires DIO alignment.
746 *
747 * Also skip uploading for data that's been read and just needs copying
748 * to the cache.
749 */
750 for (int s = 0; s < NR_IO_STREAMS; s++) {
751 stream = &wreq->io_streams[s];
752 stream->submit_off = 0;
753 stream->submit_len = flen;
754 if (!stream->avail) {
755 stream->submit_off = UINT_MAX;
756 stream->submit_len = 0;
757 }
758 }
759
760 /* Attach the folio to one or more subrequests. For a big folio, we
761 * could end up with thousands of subrequests if the wsize is small -
762 * but we might need to wait during the creation of subrequests for
763 * network resources (eg. SMB credits).
764 */
765 for (;;) {
766 ssize_t part;
767 size_t lowest_off = ULONG_MAX;
768 int choose_s = -1;
769
770 /* Always add to the lowest-submitted stream first. */
771 for (int s = 0; s < NR_IO_STREAMS; s++) {
772 stream = &wreq->io_streams[s];
773 if (stream->submit_len > 0 &&
774 stream->submit_off < lowest_off) {
775 lowest_off = stream->submit_off;
776 choose_s = s;
777 }
778 }
779
780 if (choose_s < 0)
781 break;
782 stream = &wreq->io_streams[choose_s];
783
784 /* Advance the iterator(s). */
785 if (stream->submit_off > iter_off) {
786 rolling_buffer_advance(&wreq->buffer, stream->submit_off - iter_off);
787 iter_off = stream->submit_off;
788 }
789
790 atomic64_set(&wreq->issued_to, fpos + stream->submit_off);
791 stream->submit_extendable_to = fsize - stream->submit_off;
792 part = netfs_advance_write(wreq, stream, fpos + stream->submit_off,
793 stream->submit_len, to_eof);
794 stream->submit_off += part;
795 if (part > stream->submit_len)
796 stream->submit_len = 0;
797 else
798 stream->submit_len -= part;
799 if (part > 0)
800 no_debug = true;
801 }
802
803 wreq->buffer.iter.iov_offset = 0;
804 if (fsize > iter_off)
805 rolling_buffer_advance(&wreq->buffer, fsize - iter_off);
806 atomic64_set(&wreq->issued_to, fpos + fsize);
807
808 if (!no_debug)
809 kdebug("R=%x: No submit", wreq->debug_id);
810 _leave(" = 0");
811 return 0;
812 }
813
814 /**
815 * netfs_writeback_single - Write back a monolithic payload
816 * @mapping: The mapping to write from
817 * @wbc: Hints from the VM
818 * @iter: Data to write, must be ITER_FOLIOQ.
819 *
820 * Write a monolithic, non-pagecache object back to the server and/or
821 * the cache.
822 */
netfs_writeback_single(struct address_space * mapping,struct writeback_control * wbc,struct iov_iter * iter)823 int netfs_writeback_single(struct address_space *mapping,
824 struct writeback_control *wbc,
825 struct iov_iter *iter)
826 {
827 struct netfs_io_request *wreq;
828 struct netfs_inode *ictx = netfs_inode(mapping->host);
829 struct folio_queue *fq;
830 size_t size = iov_iter_count(iter);
831 int ret;
832
833 if (WARN_ON_ONCE(!iov_iter_is_folioq(iter)))
834 return -EIO;
835
836 if (!mutex_trylock(&ictx->wb_lock)) {
837 if (wbc->sync_mode == WB_SYNC_NONE) {
838 netfs_stat(&netfs_n_wb_lock_skip);
839 return 0;
840 }
841 netfs_stat(&netfs_n_wb_lock_wait);
842 mutex_lock(&ictx->wb_lock);
843 }
844
845 wreq = netfs_create_write_req(mapping, NULL, 0, NETFS_WRITEBACK_SINGLE);
846 if (IS_ERR(wreq)) {
847 ret = PTR_ERR(wreq);
848 goto couldnt_start;
849 }
850
851 __set_bit(NETFS_RREQ_OFFLOAD_COLLECTION, &wreq->flags);
852 trace_netfs_write(wreq, netfs_write_trace_writeback_single);
853 netfs_stat(&netfs_n_wh_writepages);
854
855 if (__test_and_set_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags))
856 wreq->netfs_ops->begin_writeback(wreq);
857
858 for (fq = (struct folio_queue *)iter->folioq; fq; fq = fq->next) {
859 for (int slot = 0; slot < folioq_count(fq); slot++) {
860 struct folio *folio = folioq_folio(fq, slot);
861 size_t part = umin(folioq_folio_size(fq, slot), size);
862
863 _debug("wbiter %lx %llx", folio->index, atomic64_read(&wreq->issued_to));
864
865 ret = netfs_write_folio_single(wreq, folio);
866 if (ret < 0)
867 goto stop;
868 size -= part;
869 if (size <= 0)
870 goto stop;
871 }
872 }
873
874 stop:
875 for (int s = 0; s < NR_IO_STREAMS; s++)
876 netfs_issue_write(wreq, &wreq->io_streams[s]);
877 smp_wmb(); /* Write lists before ALL_QUEUED. */
878 set_bit(NETFS_RREQ_ALL_QUEUED, &wreq->flags);
879
880 mutex_unlock(&ictx->wb_lock);
881 netfs_wake_collector(wreq);
882
883 netfs_put_request(wreq, netfs_rreq_trace_put_return);
884 _leave(" = %d", ret);
885 return ret;
886
887 couldnt_start:
888 mutex_unlock(&ictx->wb_lock);
889 _leave(" = %d", ret);
890 return ret;
891 }
892 EXPORT_SYMBOL(netfs_writeback_single);
893