1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Network filesystem read subrequest result collection, assessment and
3 * retrying.
4 *
5 * Copyright (C) 2024 Red Hat, Inc. All Rights Reserved.
6 * Written by David Howells (dhowells@redhat.com)
7 */
8
9 #include <linux/export.h>
10 #include <linux/fs.h>
11 #include <linux/mm.h>
12 #include <linux/pagemap.h>
13 #include <linux/slab.h>
14 #include <linux/task_io_accounting_ops.h>
15 #include "internal.h"
16
17 /* Notes made in the collector */
18 #define HIT_PENDING 0x01 /* A front op was still pending */
19 #define MADE_PROGRESS 0x04 /* Made progress cleaning up a stream or the folio set */
20 #define BUFFERED 0x08 /* The pagecache needs cleaning up */
21 #define NEED_RETRY 0x10 /* A front op requests retrying */
22 #define COPY_TO_CACHE 0x40 /* Need to copy subrequest to cache */
23 #define ABANDON_SREQ 0x80 /* Need to abandon untransferred part of subrequest */
24
25 /*
26 * Clear the unread part of an I/O request.
27 */
netfs_clear_unread(struct netfs_io_subrequest * subreq)28 static void netfs_clear_unread(struct netfs_io_subrequest *subreq)
29 {
30 netfs_reset_iter(subreq);
31 WARN_ON_ONCE(subreq->len - subreq->transferred != iov_iter_count(&subreq->io_iter));
32 iov_iter_zero(iov_iter_count(&subreq->io_iter), &subreq->io_iter);
33 if (subreq->start + subreq->transferred >= subreq->rreq->i_size)
34 __set_bit(NETFS_SREQ_HIT_EOF, &subreq->flags);
35 }
36
37 /*
38 * Flush, mark and unlock a folio that's now completely read. If we want to
39 * cache the folio, we set the group to NETFS_FOLIO_COPY_TO_CACHE, mark it
40 * dirty and let writeback handle it.
41 */
netfs_unlock_read_folio(struct netfs_io_request * rreq,struct folio_queue * folioq,int slot)42 static void netfs_unlock_read_folio(struct netfs_io_request *rreq,
43 struct folio_queue *folioq,
44 int slot)
45 {
46 struct netfs_folio *finfo;
47 struct folio *folio = folioq_folio(folioq, slot);
48
49 if (unlikely(folio_pos(folio) < rreq->abandon_to)) {
50 trace_netfs_folio(folio, netfs_folio_trace_abandon);
51 goto just_unlock;
52 }
53
54 flush_dcache_folio(folio);
55 folio_mark_uptodate(folio);
56
57 if (!test_bit(NETFS_RREQ_USE_PGPRIV2, &rreq->flags)) {
58 finfo = netfs_folio_info(folio);
59 if (finfo) {
60 trace_netfs_folio(folio, netfs_folio_trace_filled_gaps);
61 if (finfo->netfs_group)
62 folio_change_private(folio, finfo->netfs_group);
63 else
64 folio_detach_private(folio);
65 kfree(finfo);
66 }
67
68 if (test_bit(NETFS_RREQ_FOLIO_COPY_TO_CACHE, &rreq->flags)) {
69 if (!WARN_ON_ONCE(folio_get_private(folio) != NULL)) {
70 trace_netfs_folio(folio, netfs_folio_trace_copy_to_cache);
71 folio_attach_private(folio, NETFS_FOLIO_COPY_TO_CACHE);
72 folio_mark_dirty(folio);
73 }
74 } else {
75 trace_netfs_folio(folio, netfs_folio_trace_read_done);
76 }
77
78 folioq_clear(folioq, slot);
79 } else {
80 // TODO: Use of PG_private_2 is deprecated.
81 if (test_bit(NETFS_RREQ_FOLIO_COPY_TO_CACHE, &rreq->flags))
82 netfs_pgpriv2_copy_to_cache(rreq, folio);
83 }
84
85 just_unlock:
86 if (folio->index == rreq->no_unlock_folio &&
87 test_bit(NETFS_RREQ_NO_UNLOCK_FOLIO, &rreq->flags)) {
88 _debug("no unlock");
89 } else {
90 trace_netfs_folio(folio, netfs_folio_trace_read_unlock);
91 folio_unlock(folio);
92 }
93
94 folioq_clear(folioq, slot);
95 }
96
97 /*
98 * Unlock any folios we've finished with.
99 */
netfs_read_unlock_folios(struct netfs_io_request * rreq,unsigned int * notes)100 static void netfs_read_unlock_folios(struct netfs_io_request *rreq,
101 unsigned int *notes)
102 {
103 struct folio_queue *folioq = rreq->buffer.tail;
104 unsigned long long collected_to = rreq->collected_to;
105 unsigned int slot = rreq->buffer.first_tail_slot;
106
107 if (rreq->cleaned_to >= rreq->collected_to)
108 return;
109
110 // TODO: Begin decryption
111
112 if (slot >= folioq_nr_slots(folioq)) {
113 folioq = rolling_buffer_delete_spent(&rreq->buffer);
114 if (!folioq) {
115 rreq->front_folio_order = 0;
116 return;
117 }
118 slot = 0;
119 }
120
121 for (;;) {
122 struct folio *folio;
123 unsigned long long fpos, fend;
124 unsigned int order;
125 size_t fsize;
126
127 if (*notes & COPY_TO_CACHE)
128 set_bit(NETFS_RREQ_FOLIO_COPY_TO_CACHE, &rreq->flags);
129
130 folio = folioq_folio(folioq, slot);
131 if (WARN_ONCE(!folio_test_locked(folio),
132 "R=%08x: folio %lx is not locked\n",
133 rreq->debug_id, folio->index))
134 trace_netfs_folio(folio, netfs_folio_trace_not_locked);
135
136 order = folioq_folio_order(folioq, slot);
137 rreq->front_folio_order = order;
138 fsize = PAGE_SIZE << order;
139 fpos = folio_pos(folio);
140 fend = umin(fpos + fsize, rreq->i_size);
141
142 trace_netfs_collect_folio(rreq, folio, fend, collected_to);
143
144 /* Unlock any folio we've transferred all of. */
145 if (collected_to < fend)
146 break;
147
148 netfs_unlock_read_folio(rreq, folioq, slot);
149 WRITE_ONCE(rreq->cleaned_to, fpos + fsize);
150 *notes |= MADE_PROGRESS;
151
152 clear_bit(NETFS_RREQ_FOLIO_COPY_TO_CACHE, &rreq->flags);
153
154 /* Clean up the head folioq. If we clear an entire folioq, then
155 * we can get rid of it provided it's not also the tail folioq
156 * being filled by the issuer.
157 */
158 folioq_clear(folioq, slot);
159 slot++;
160 if (slot >= folioq_nr_slots(folioq)) {
161 folioq = rolling_buffer_delete_spent(&rreq->buffer);
162 if (!folioq)
163 goto done;
164 slot = 0;
165 trace_netfs_folioq(folioq, netfs_trace_folioq_read_progress);
166 }
167
168 if (fpos + fsize >= collected_to)
169 break;
170 }
171
172 rreq->buffer.tail = folioq;
173 done:
174 rreq->buffer.first_tail_slot = slot;
175 }
176
177 /*
178 * Collect and assess the results of various read subrequests. We may need to
179 * retry some of the results.
180 *
181 * Note that we have a sequence of subrequests, which may be drawing on
182 * different sources and may or may not be the same size or starting position
183 * and may not even correspond in boundary alignment.
184 */
netfs_collect_read_results(struct netfs_io_request * rreq)185 static void netfs_collect_read_results(struct netfs_io_request *rreq)
186 {
187 struct netfs_io_subrequest *front, *remove;
188 struct netfs_io_stream *stream = &rreq->io_streams[0];
189 unsigned int notes;
190
191 _enter("%llx-%llx", rreq->start, rreq->start + rreq->len);
192 trace_netfs_rreq(rreq, netfs_rreq_trace_collect);
193 trace_netfs_collect(rreq);
194
195 reassess:
196 if (rreq->origin == NETFS_READAHEAD ||
197 rreq->origin == NETFS_READPAGE ||
198 rreq->origin == NETFS_READ_FOR_WRITE)
199 notes = BUFFERED;
200 else
201 notes = 0;
202
203 /* Remove completed subrequests from the front of the stream and
204 * advance the completion point. We stop when we hit something that's
205 * in progress. The issuer thread may be adding stuff to the tail
206 * whilst we're doing this.
207 */
208 front = READ_ONCE(stream->front);
209 while (front) {
210 size_t transferred;
211
212 trace_netfs_collect_sreq(rreq, front);
213 _debug("sreq [%x] %llx %zx/%zx",
214 front->debug_index, front->start, front->transferred, front->len);
215
216 if (stream->collected_to < front->start) {
217 trace_netfs_collect_gap(rreq, stream, front->start, 'F');
218 stream->collected_to = front->start;
219 }
220
221 if (netfs_check_subreq_in_progress(front))
222 notes |= HIT_PENDING;
223 smp_rmb(); /* Read counters after IN_PROGRESS flag. */
224 transferred = READ_ONCE(front->transferred);
225
226 /* If we can now collect the next folio, do so. We don't want
227 * to defer this as we have to decide whether we need to copy
228 * to the cache or not, and that may differ between adjacent
229 * subreqs.
230 */
231 if (notes & BUFFERED) {
232 size_t fsize = PAGE_SIZE << rreq->front_folio_order;
233
234 /* Clear the tail of a short read. */
235 if (!(notes & HIT_PENDING) &&
236 front->error == 0 &&
237 transferred < front->len &&
238 (test_bit(NETFS_SREQ_HIT_EOF, &front->flags) ||
239 test_bit(NETFS_SREQ_CLEAR_TAIL, &front->flags))) {
240 netfs_clear_unread(front);
241 transferred = front->transferred = front->len;
242 trace_netfs_sreq(front, netfs_sreq_trace_clear);
243 }
244
245 stream->collected_to = front->start + transferred;
246 rreq->collected_to = stream->collected_to;
247
248 if (test_bit(NETFS_SREQ_COPY_TO_CACHE, &front->flags))
249 notes |= COPY_TO_CACHE;
250
251 if (test_bit(NETFS_SREQ_FAILED, &front->flags)) {
252 rreq->abandon_to = front->start + front->len;
253 front->transferred = front->len;
254 transferred = front->len;
255 trace_netfs_rreq(rreq, netfs_rreq_trace_set_abandon);
256 }
257 if (front->start + transferred >= rreq->cleaned_to + fsize ||
258 test_bit(NETFS_SREQ_HIT_EOF, &front->flags))
259 netfs_read_unlock_folios(rreq, ¬es);
260 } else {
261 stream->collected_to = front->start + transferred;
262 rreq->collected_to = stream->collected_to;
263 }
264
265 /* Stall if the front is still undergoing I/O. */
266 if (notes & HIT_PENDING)
267 break;
268
269 if (test_bit(NETFS_SREQ_FAILED, &front->flags)) {
270 if (!stream->failed) {
271 stream->error = front->error;
272 rreq->error = front->error;
273 set_bit(NETFS_RREQ_FAILED, &rreq->flags);
274 stream->failed = true;
275 }
276 notes |= MADE_PROGRESS | ABANDON_SREQ;
277 } else if (test_bit(NETFS_SREQ_NEED_RETRY, &front->flags)) {
278 stream->need_retry = true;
279 notes |= NEED_RETRY | MADE_PROGRESS;
280 break;
281 } else if (test_bit(NETFS_RREQ_SHORT_TRANSFER, &rreq->flags)) {
282 notes |= MADE_PROGRESS;
283 } else {
284 if (!stream->failed)
285 stream->transferred += transferred;
286 if (front->transferred < front->len)
287 set_bit(NETFS_RREQ_SHORT_TRANSFER, &rreq->flags);
288 notes |= MADE_PROGRESS;
289 }
290
291 /* Remove if completely consumed. */
292 stream->source = front->source;
293 spin_lock(&rreq->lock);
294
295 remove = front;
296 trace_netfs_sreq(front,
297 notes & ABANDON_SREQ ?
298 netfs_sreq_trace_abandoned : netfs_sreq_trace_consumed);
299 list_del_init(&front->rreq_link);
300 front = list_first_entry_or_null(&stream->subrequests,
301 struct netfs_io_subrequest, rreq_link);
302 stream->front = front;
303 spin_unlock(&rreq->lock);
304 netfs_put_subrequest(remove,
305 notes & ABANDON_SREQ ?
306 netfs_sreq_trace_put_abandon :
307 netfs_sreq_trace_put_done);
308 }
309
310 trace_netfs_collect_stream(rreq, stream);
311 trace_netfs_collect_state(rreq, rreq->collected_to, notes);
312
313 if (!(notes & BUFFERED))
314 rreq->cleaned_to = rreq->collected_to;
315
316 if (notes & NEED_RETRY)
317 goto need_retry;
318 if (notes & MADE_PROGRESS) {
319 netfs_wake_rreq_flag(rreq, NETFS_RREQ_PAUSE, netfs_rreq_trace_unpause);
320 //cond_resched();
321 goto reassess;
322 }
323
324 out:
325 _leave(" = %x", notes);
326 return;
327
328 need_retry:
329 /* Okay... We're going to have to retry parts of the stream. Note
330 * that any partially completed op will have had any wholly transferred
331 * folios removed from it.
332 */
333 _debug("retry");
334 netfs_retry_reads(rreq);
335 goto out;
336 }
337
338 /*
339 * Do page flushing and suchlike after DIO.
340 */
netfs_rreq_assess_dio(struct netfs_io_request * rreq)341 static void netfs_rreq_assess_dio(struct netfs_io_request *rreq)
342 {
343 unsigned int i;
344
345 if (rreq->origin == NETFS_UNBUFFERED_READ ||
346 rreq->origin == NETFS_DIO_READ) {
347 for (i = 0; i < rreq->direct_bv_count; i++) {
348 flush_dcache_page(rreq->direct_bv[i].bv_page);
349 // TODO: cifs marks pages in the destination buffer
350 // dirty under some circumstances after a read. Do we
351 // need to do that too?
352 set_page_dirty(rreq->direct_bv[i].bv_page);
353 }
354 }
355
356 if (rreq->iocb) {
357 rreq->iocb->ki_pos += rreq->transferred;
358 if (rreq->iocb->ki_complete) {
359 trace_netfs_rreq(rreq, netfs_rreq_trace_ki_complete);
360 rreq->iocb->ki_complete(
361 rreq->iocb, rreq->error ? rreq->error : rreq->transferred);
362 }
363 }
364 if (rreq->netfs_ops->done)
365 rreq->netfs_ops->done(rreq);
366 if (rreq->origin == NETFS_UNBUFFERED_READ ||
367 rreq->origin == NETFS_DIO_READ)
368 inode_dio_end(rreq->inode);
369 }
370
371 /*
372 * Do processing after reading a monolithic single object.
373 */
netfs_rreq_assess_single(struct netfs_io_request * rreq)374 static void netfs_rreq_assess_single(struct netfs_io_request *rreq)
375 {
376 struct netfs_io_stream *stream = &rreq->io_streams[0];
377
378 if (!rreq->error && stream->source == NETFS_DOWNLOAD_FROM_SERVER &&
379 fscache_resources_valid(&rreq->cache_resources)) {
380 trace_netfs_rreq(rreq, netfs_rreq_trace_dirty);
381 netfs_single_mark_inode_dirty(rreq->inode);
382 }
383
384 if (rreq->iocb) {
385 rreq->iocb->ki_pos += rreq->transferred;
386 if (rreq->iocb->ki_complete) {
387 trace_netfs_rreq(rreq, netfs_rreq_trace_ki_complete);
388 rreq->iocb->ki_complete(
389 rreq->iocb, rreq->error ? rreq->error : rreq->transferred);
390 }
391 }
392 if (rreq->netfs_ops->done)
393 rreq->netfs_ops->done(rreq);
394 }
395
396 /*
397 * Perform the collection of subrequests and folios.
398 *
399 * Note that we're in normal kernel thread context at this point, possibly
400 * running on a workqueue.
401 */
netfs_read_collection(struct netfs_io_request * rreq)402 bool netfs_read_collection(struct netfs_io_request *rreq)
403 {
404 struct netfs_io_stream *stream = &rreq->io_streams[0];
405
406 netfs_collect_read_results(rreq);
407
408 /* We're done when the app thread has finished posting subreqs and the
409 * queue is empty.
410 */
411 if (!test_bit(NETFS_RREQ_ALL_QUEUED, &rreq->flags))
412 return false;
413 smp_rmb(); /* Read ALL_QUEUED before subreq lists. */
414
415 if (!list_empty(&stream->subrequests))
416 return false;
417
418 /* Okay, declare that all I/O is complete. */
419 rreq->transferred = stream->transferred;
420 trace_netfs_rreq(rreq, netfs_rreq_trace_complete);
421
422 //netfs_rreq_is_still_valid(rreq);
423
424 switch (rreq->origin) {
425 case NETFS_UNBUFFERED_READ:
426 case NETFS_DIO_READ:
427 case NETFS_READ_GAPS:
428 netfs_rreq_assess_dio(rreq);
429 break;
430 case NETFS_READ_SINGLE:
431 netfs_rreq_assess_single(rreq);
432 break;
433 default:
434 break;
435 }
436 task_io_account_read(rreq->transferred);
437
438 netfs_wake_rreq_flag(rreq, NETFS_RREQ_IN_PROGRESS, netfs_rreq_trace_wake_ip);
439 /* As we cleared NETFS_RREQ_IN_PROGRESS, we acquired its ref. */
440
441 trace_netfs_rreq(rreq, netfs_rreq_trace_done);
442 netfs_clear_subrequests(rreq);
443 netfs_unlock_abandoned_read_pages(rreq);
444 if (unlikely(rreq->copy_to_cache))
445 netfs_pgpriv2_end_copy_to_cache(rreq);
446 return true;
447 }
448
netfs_read_collection_worker(struct work_struct * work)449 void netfs_read_collection_worker(struct work_struct *work)
450 {
451 struct netfs_io_request *rreq = container_of(work, struct netfs_io_request, work);
452
453 netfs_see_request(rreq, netfs_rreq_trace_see_work);
454 if (netfs_check_rreq_in_progress(rreq)) {
455 if (netfs_read_collection(rreq))
456 /* Drop the ref from the IN_PROGRESS flag. */
457 netfs_put_request(rreq, netfs_rreq_trace_put_work_ip);
458 else
459 netfs_see_request(rreq, netfs_rreq_trace_see_work_complete);
460 }
461 }
462
463 /**
464 * netfs_read_subreq_progress - Note progress of a read operation.
465 * @subreq: The read request that has terminated.
466 *
467 * This tells the read side of netfs lib that a contributory I/O operation has
468 * made some progress and that it may be possible to unlock some folios.
469 *
470 * Before calling, the filesystem should update subreq->transferred to track
471 * the amount of data copied into the output buffer.
472 */
netfs_read_subreq_progress(struct netfs_io_subrequest * subreq)473 void netfs_read_subreq_progress(struct netfs_io_subrequest *subreq)
474 {
475 struct netfs_io_request *rreq = subreq->rreq;
476 struct netfs_io_stream *stream = &rreq->io_streams[0];
477 size_t fsize = PAGE_SIZE << rreq->front_folio_order;
478
479 trace_netfs_sreq(subreq, netfs_sreq_trace_progress);
480
481 /* If we are at the head of the queue, wake up the collector,
482 * getting a ref to it if we were the ones to do so.
483 */
484 if (subreq->start + subreq->transferred > rreq->cleaned_to + fsize &&
485 (rreq->origin == NETFS_READAHEAD ||
486 rreq->origin == NETFS_READPAGE ||
487 rreq->origin == NETFS_READ_FOR_WRITE) &&
488 list_is_first(&subreq->rreq_link, &stream->subrequests)
489 ) {
490 __set_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags);
491 netfs_wake_collector(rreq);
492 }
493 }
494 EXPORT_SYMBOL(netfs_read_subreq_progress);
495
496 /**
497 * netfs_read_subreq_terminated - Note the termination of an I/O operation.
498 * @subreq: The I/O request that has terminated.
499 *
500 * This tells the read helper that a contributory I/O operation has terminated,
501 * one way or another, and that it should integrate the results.
502 *
503 * The caller indicates the outcome of the operation through @subreq->error,
504 * supplying 0 to indicate a successful or retryable transfer (if
505 * NETFS_SREQ_NEED_RETRY is set) or a negative error code. The helper will
506 * look after reissuing I/O operations as appropriate and writing downloaded
507 * data to the cache.
508 *
509 * Before calling, the filesystem should update subreq->transferred to track
510 * the amount of data copied into the output buffer.
511 */
netfs_read_subreq_terminated(struct netfs_io_subrequest * subreq)512 void netfs_read_subreq_terminated(struct netfs_io_subrequest *subreq)
513 {
514 struct netfs_io_request *rreq = subreq->rreq;
515
516 switch (subreq->source) {
517 case NETFS_READ_FROM_CACHE:
518 netfs_stat(&netfs_n_rh_read_done);
519 break;
520 case NETFS_DOWNLOAD_FROM_SERVER:
521 netfs_stat(&netfs_n_rh_download_done);
522 break;
523 default:
524 break;
525 }
526
527 /* Deal with retry requests, short reads and errors. If we retry
528 * but don't make progress, we abandon the attempt.
529 */
530 if (!subreq->error && subreq->transferred < subreq->len) {
531 if (test_bit(NETFS_SREQ_HIT_EOF, &subreq->flags)) {
532 trace_netfs_sreq(subreq, netfs_sreq_trace_hit_eof);
533 } else if (test_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags)) {
534 trace_netfs_sreq(subreq, netfs_sreq_trace_need_clear);
535 } else if (test_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags)) {
536 trace_netfs_sreq(subreq, netfs_sreq_trace_need_retry);
537 } else if (test_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags)) {
538 __set_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
539 trace_netfs_sreq(subreq, netfs_sreq_trace_partial_read);
540 } else {
541 __set_bit(NETFS_SREQ_FAILED, &subreq->flags);
542 subreq->error = -ENODATA;
543 trace_netfs_sreq(subreq, netfs_sreq_trace_short);
544 }
545 }
546
547 if (unlikely(subreq->error < 0)) {
548 trace_netfs_failure(rreq, subreq, subreq->error, netfs_fail_read);
549 if (subreq->source == NETFS_READ_FROM_CACHE) {
550 netfs_stat(&netfs_n_rh_read_failed);
551 __set_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
552 } else {
553 netfs_stat(&netfs_n_rh_download_failed);
554 __set_bit(NETFS_SREQ_FAILED, &subreq->flags);
555 }
556 trace_netfs_rreq(rreq, netfs_rreq_trace_set_pause);
557 set_bit(NETFS_RREQ_PAUSE, &rreq->flags);
558 }
559
560 trace_netfs_sreq(subreq, netfs_sreq_trace_terminated);
561 netfs_subreq_clear_in_progress(subreq);
562 netfs_put_subrequest(subreq, netfs_sreq_trace_put_terminated);
563 }
564 EXPORT_SYMBOL(netfs_read_subreq_terminated);
565
566 /*
567 * Handle termination of a read from the cache.
568 */
netfs_cache_read_terminated(void * priv,ssize_t transferred_or_error)569 void netfs_cache_read_terminated(void *priv, ssize_t transferred_or_error)
570 {
571 struct netfs_io_subrequest *subreq = priv;
572
573 if (transferred_or_error > 0) {
574 subreq->error = 0;
575 if (transferred_or_error > 0) {
576 subreq->transferred += transferred_or_error;
577 __set_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags);
578 }
579 } else {
580 subreq->error = transferred_or_error;
581 }
582 netfs_read_subreq_terminated(subreq);
583 }
584