xref: /linux/fs/netfs/write_collect.c (revision 5dfa01ef37a8b944773aef8dee747cd76dec4234)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Network filesystem write subrequest result collection, assessment
3  * and 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 "internal.h"
15 
16 /* Notes made in the collector */
17 #define HIT_PENDING		0x01	/* A front op was still pending */
18 #define NEED_REASSESS		0x02	/* Need to loop round and reassess */
19 #define MADE_PROGRESS		0x04	/* Made progress cleaning up a stream or the folio set */
20 #define NEED_UNLOCK		0x08	/* The pagecache needs unlocking */
21 #define NEED_RETRY		0x10	/* A front op requests retrying */
22 #define SAW_FAILURE		0x20	/* One stream or hit a permanent failure */
23 
netfs_dump_request(const struct netfs_io_request * rreq)24 static void netfs_dump_request(const struct netfs_io_request *rreq)
25 {
26 	pr_err("Request R=%08x r=%d fl=%lx or=%x e=%ld\n",
27 	       rreq->debug_id, refcount_read(&rreq->ref), rreq->flags,
28 	       rreq->origin, rreq->error);
29 	pr_err("  st=%llx tsl=%zx/%llx/%llx\n",
30 	       rreq->start, rreq->transferred, rreq->submitted, rreq->len);
31 	pr_err("  cci=%llx/%llx/%llx\n",
32 	       rreq->cleaned_to, rreq->collected_to, atomic64_read(&rreq->issued_to));
33 	pr_err("  iw=%pSR\n", rreq->netfs_ops->issue_write);
34 	for (int i = 0; i < NR_IO_STREAMS; i++) {
35 		const struct netfs_io_subrequest *sreq;
36 		const struct netfs_io_stream *s = &rreq->io_streams[i];
37 
38 		pr_err("  str[%x] s=%x e=%d acnf=%u,%u,%u,%u\n",
39 		       s->stream_nr, s->source, s->error,
40 		       s->avail, s->active, s->need_retry, s->failed);
41 		pr_err("  str[%x] ct=%llx t=%zx\n",
42 		       s->stream_nr, s->collected_to, s->transferred);
43 		list_for_each_entry(sreq, &s->subrequests, rreq_link) {
44 			pr_err("  sreq[%x:%x] sc=%u s=%llx t=%zx/%zx r=%d f=%lx\n",
45 			       sreq->stream_nr, sreq->debug_index, sreq->source,
46 			       sreq->start, sreq->transferred, sreq->len,
47 			       refcount_read(&sreq->ref), sreq->flags);
48 		}
49 	}
50 }
51 
52 /*
53  * Successful completion of write of a folio to the server and/or cache.  Note
54  * that we are not allowed to lock the folio here on pain of deadlocking with
55  * truncate.
56  */
netfs_folio_written_back(struct folio * folio)57 int netfs_folio_written_back(struct folio *folio)
58 {
59 	enum netfs_folio_trace why = netfs_folio_trace_clear;
60 	struct inode *inode = folio_inode(folio);
61 	struct netfs_inode *ictx = netfs_inode(inode);
62 	struct netfs_folio *finfo;
63 	struct netfs_group *group = NULL;
64 	int gcount = 0;
65 
66 	if ((finfo = netfs_folio_info(folio))) {
67 		/* Streaming writes cannot be redirtied whilst under writeback,
68 		 * so discard the streaming record.
69 		 */
70 		unsigned long long fend;
71 
72 		fend = folio_pos(folio) + finfo->dirty_offset + finfo->dirty_len;
73 		spin_lock(&ictx->inode.i_lock);
74 		if (fend > ictx->_zero_point)
75 			netfs_write_zero_point(inode, fend);
76 		spin_unlock(&ictx->inode.i_lock);
77 
78 		folio_detach_private(folio);
79 		group = finfo->netfs_group;
80 		gcount++;
81 		kfree(finfo);
82 		why = netfs_folio_trace_clear_s;
83 		goto end_wb;
84 	}
85 
86 	if ((group = netfs_folio_group(folio))) {
87 		if (group == NETFS_FOLIO_COPY_TO_CACHE) {
88 			why = netfs_folio_trace_clear_cc;
89 			folio_detach_private(folio);
90 			goto end_wb;
91 		}
92 
93 		/* Need to detach the group pointer if the page didn't get
94 		 * redirtied.  If it has been redirtied, then it must be within
95 		 * the same group.
96 		 */
97 		why = netfs_folio_trace_redirtied;
98 		if (!folio_test_dirty(folio)) {
99 			folio_detach_private(folio);
100 			gcount++;
101 			why = netfs_folio_trace_clear_g;
102 		}
103 	}
104 
105 end_wb:
106 	trace_netfs_folio(folio, why);
107 	folio_end_writeback(folio);
108 	return gcount;
109 }
110 
111 /*
112  * Unlock any folios we've finished with.
113  */
netfs_writeback_unlock_folios(struct netfs_io_request * wreq,unsigned int * notes)114 static void netfs_writeback_unlock_folios(struct netfs_io_request *wreq,
115 					  unsigned int *notes)
116 {
117 	struct folio_queue *folioq = wreq->buffer.tail;
118 	unsigned long long collected_to = wreq->collected_to;
119 	unsigned int slot = wreq->buffer.first_tail_slot;
120 
121 	if (WARN_ON_ONCE(!folioq)) {
122 		pr_err("[!] Writeback unlock found empty rolling buffer!\n");
123 		netfs_dump_request(wreq);
124 		return;
125 	}
126 
127 	if (wreq->origin == NETFS_PGPRIV2_COPY_TO_CACHE) {
128 		if (netfs_pgpriv2_unlock_copied_folios(wreq))
129 			*notes |= MADE_PROGRESS;
130 		return;
131 	}
132 
133 	if (slot >= folioq_nr_slots(folioq)) {
134 		folioq = rolling_buffer_delete_spent(&wreq->buffer);
135 		if (!folioq)
136 			return;
137 		slot = 0;
138 	}
139 
140 	for (;;) {
141 		struct folio *folio;
142 		struct netfs_folio *finfo;
143 		unsigned long long fpos, fend;
144 		size_t fsize, flen;
145 
146 		folio = folioq_folio(folioq, slot);
147 		if (WARN_ONCE(!folio_test_writeback(folio),
148 			      "R=%08x: folio %lx is not under writeback\n",
149 			      wreq->debug_id, folio->index))
150 			trace_netfs_folio(folio, netfs_folio_trace_not_under_wback);
151 
152 		fpos = folio_pos(folio);
153 		fsize = folio_size(folio);
154 		finfo = netfs_folio_info(folio);
155 		flen = finfo ? finfo->dirty_offset + finfo->dirty_len : fsize;
156 
157 		fend = min_t(unsigned long long, fpos + flen, wreq->i_size);
158 
159 		trace_netfs_collect_folio(wreq, folio, fend, collected_to);
160 
161 		/* Unlock any folio we've transferred all of. */
162 		if (collected_to < fend)
163 			break;
164 
165 		wreq->nr_group_rel += netfs_folio_written_back(folio);
166 		wreq->cleaned_to = fpos + fsize;
167 		*notes |= MADE_PROGRESS;
168 
169 		/* Clean up the head folioq.  If we clear an entire folioq, then
170 		 * we can get rid of it provided it's not also the tail folioq
171 		 * being filled by the issuer.
172 		 */
173 		folioq_clear(folioq, slot);
174 		slot++;
175 		if (slot >= folioq_nr_slots(folioq)) {
176 			folioq = rolling_buffer_delete_spent(&wreq->buffer);
177 			if (!folioq)
178 				goto done;
179 			slot = 0;
180 		}
181 
182 		if (fpos + fsize >= collected_to)
183 			break;
184 	}
185 
186 	wreq->buffer.tail = folioq;
187 done:
188 	wreq->buffer.first_tail_slot = slot;
189 }
190 
191 /*
192  * Collect and assess the results of various write subrequests.  We may need to
193  * retry some of the results - or even do an RMW cycle for content crypto.
194  *
195  * Note that we have a number of parallel, overlapping lists of subrequests,
196  * one to the server and one to the local cache for example, which may not be
197  * the same size or starting position and may not even correspond in boundary
198  * alignment.
199  */
netfs_collect_write_results(struct netfs_io_request * wreq)200 static void netfs_collect_write_results(struct netfs_io_request *wreq)
201 {
202 	struct netfs_io_subrequest *front, *remove;
203 	struct netfs_io_stream *stream;
204 	unsigned long long collected_to, issued_to;
205 	unsigned int notes;
206 	int s;
207 
208 	_enter("%llx-%llx", wreq->start, wreq->start + wreq->len);
209 	trace_netfs_collect(wreq);
210 	trace_netfs_rreq(wreq, netfs_rreq_trace_collect);
211 
212 reassess_streams:
213 	issued_to = atomic64_read(&wreq->issued_to);
214 	smp_rmb();
215 	collected_to = ULLONG_MAX;
216 	if (wreq->origin == NETFS_WRITEBACK ||
217 	    wreq->origin == NETFS_WRITETHROUGH ||
218 	    wreq->origin == NETFS_PGPRIV2_COPY_TO_CACHE)
219 		notes = NEED_UNLOCK;
220 	else
221 		notes = 0;
222 
223 	/* Remove completed subrequests from the front of the streams and
224 	 * advance the completion point on each stream.  We stop when we hit
225 	 * something that's in progress.  The issuer thread may be adding stuff
226 	 * to the tail whilst we're doing this.
227 	 */
228 	for (s = 0; s < NR_IO_STREAMS; s++) {
229 		stream = &wreq->io_streams[s];
230 		/* Read active flag before list pointers */
231 		if (!smp_load_acquire(&stream->active))
232 			continue;
233 
234 		front = list_first_entry_or_null_acquire(&stream->subrequests,
235 							 struct netfs_io_subrequest, rreq_link);
236 		/* Read first subreq pointer before IN_PROGRESS flag. */
237 
238 		while (front) {
239 			trace_netfs_collect_sreq(wreq, front);
240 			//_debug("sreq [%x] %llx %zx/%zx",
241 			//       front->debug_index, front->start, front->transferred, front->len);
242 
243 			if (stream->collected_to < front->start) {
244 				trace_netfs_collect_gap(wreq, stream, issued_to, 'F');
245 				stream->collected_to = front->start;
246 			}
247 
248 			/* Stall if the front is still undergoing I/O. */
249 			if (netfs_check_subreq_in_progress(front)) {
250 				notes |= HIT_PENDING;
251 				break;
252 			}
253 			smp_rmb(); /* Read counters after I-P flag. */
254 
255 			if (stream->failed) {
256 				stream->collected_to = front->start + front->len;
257 				notes |= MADE_PROGRESS | SAW_FAILURE;
258 				goto cancel;
259 			}
260 			if (front->start + front->transferred > stream->collected_to) {
261 				stream->collected_to = front->start + front->transferred;
262 				stream->transferred = stream->collected_to - wreq->start;
263 				stream->transferred_valid = true;
264 				notes |= MADE_PROGRESS;
265 			}
266 			if (test_bit(NETFS_SREQ_FAILED, &front->flags)) {
267 				stream->failed = true;
268 				stream->error = front->error;
269 				if (stream->source == NETFS_UPLOAD_TO_SERVER)
270 					mapping_set_error(wreq->mapping, front->error);
271 				notes |= NEED_REASSESS | SAW_FAILURE;
272 				break;
273 			}
274 			if (front->transferred < front->len) {
275 				stream->need_retry = true;
276 				notes |= NEED_RETRY | MADE_PROGRESS;
277 				break;
278 			}
279 
280 		cancel:
281 			/* Remove if completely consumed. */
282 			spin_lock(&wreq->lock);
283 
284 			remove = front;
285 			list_del_init(&front->rreq_link);
286 			front = list_first_entry_or_null(&stream->subrequests,
287 							 struct netfs_io_subrequest, rreq_link);
288 			spin_unlock(&wreq->lock);
289 			netfs_put_subrequest(remove,
290 					     notes & SAW_FAILURE ?
291 					     netfs_sreq_trace_put_cancel :
292 					     netfs_sreq_trace_put_done);
293 		}
294 
295 		/* If we have an empty stream, we need to jump it forward
296 		 * otherwise the collection point will never advance.
297 		 */
298 		if (!front && issued_to > stream->collected_to) {
299 			trace_netfs_collect_gap(wreq, stream, issued_to, 'E');
300 			stream->collected_to = issued_to;
301 		}
302 
303 		if (stream->collected_to < collected_to)
304 			collected_to = stream->collected_to;
305 	}
306 
307 	if (collected_to != ULLONG_MAX && collected_to > wreq->collected_to)
308 		wreq->collected_to = collected_to;
309 
310 	for (s = 0; s < NR_IO_STREAMS; s++) {
311 		stream = &wreq->io_streams[s];
312 		if (stream->active)
313 			trace_netfs_collect_stream(wreq, stream);
314 	}
315 
316 	trace_netfs_collect_state(wreq, wreq->collected_to, notes);
317 
318 	/* Unlock any folios that we have now finished with. */
319 	if (notes & NEED_UNLOCK) {
320 		if (wreq->cleaned_to < wreq->collected_to)
321 			netfs_writeback_unlock_folios(wreq, &notes);
322 	} else {
323 		wreq->cleaned_to = wreq->collected_to;
324 	}
325 
326 	// TODO: Discard encryption buffers
327 
328 	if (notes & NEED_RETRY)
329 		goto need_retry;
330 
331 	if (notes & MADE_PROGRESS) {
332 		netfs_wake_rreq_flag(wreq, NETFS_RREQ_PAUSE, netfs_rreq_trace_unpause);
333 		//cond_resched();
334 		goto reassess_streams;
335 	}
336 
337 	if (notes & NEED_REASSESS) {
338 		//cond_resched();
339 		goto reassess_streams;
340 	}
341 
342 out:
343 	netfs_put_group_many(wreq->group, wreq->nr_group_rel);
344 	wreq->nr_group_rel = 0;
345 	_leave(" = %x", notes);
346 	return;
347 
348 need_retry:
349 	/* Okay...  We're going to have to retry one or both streams.  Note
350 	 * that any partially completed op will have had any wholly transferred
351 	 * folios removed from it.
352 	 */
353 	_debug("retry");
354 	netfs_retry_writes(wreq);
355 	goto out;
356 }
357 
358 /*
359  * Perform the collection of subrequests, folios and encryption buffers.
360  */
netfs_write_collection(struct netfs_io_request * wreq)361 bool netfs_write_collection(struct netfs_io_request *wreq)
362 {
363 	struct netfs_inode *ictx = netfs_inode(wreq->inode);
364 	size_t transferred;
365 	bool transferred_valid = false;
366 	int s;
367 
368 	_enter("R=%x", wreq->debug_id);
369 
370 	netfs_collect_write_results(wreq);
371 
372 	/* We're done when the app thread has finished posting subreqs and all
373 	 * the queues in all the streams are empty.
374 	 */
375 	if (!test_bit(NETFS_RREQ_ALL_QUEUED, &wreq->flags))
376 		return false;
377 	smp_rmb(); /* Read ALL_QUEUED before lists. */
378 
379 	transferred = LONG_MAX;
380 	for (s = 0; s < NR_IO_STREAMS; s++) {
381 		struct netfs_io_stream *stream = &wreq->io_streams[s];
382 		if (!stream->active)
383 			continue;
384 		if (!list_empty(&stream->subrequests))
385 			return false;
386 		if (stream->transferred_valid &&
387 		    stream->transferred < transferred) {
388 			transferred = stream->transferred;
389 			transferred_valid = true;
390 		}
391 	}
392 
393 	/* Okay, declare that all I/O is complete. */
394 	if (transferred_valid)
395 		wreq->transferred = transferred;
396 	trace_netfs_rreq(wreq, netfs_rreq_trace_write_done);
397 
398 	if (wreq->io_streams[1].active &&
399 	    wreq->io_streams[1].failed &&
400 	    ictx->ops->invalidate_cache) {
401 		/* Cache write failure doesn't prevent writeback completion
402 		 * unless we're in disconnected mode.
403 		 */
404 		ictx->ops->invalidate_cache(wreq);
405 	}
406 
407 	_debug("finished");
408 	netfs_wake_rreq_flag(wreq, NETFS_RREQ_IN_PROGRESS, netfs_rreq_trace_wake_ip);
409 	/* As we cleared NETFS_RREQ_IN_PROGRESS, we acquired its ref. */
410 
411 	if (wreq->iocb) {
412 		size_t written = min(wreq->transferred, wreq->len);
413 		wreq->iocb->ki_pos += written;
414 		if (wreq->iocb->ki_complete) {
415 			trace_netfs_rreq(wreq, netfs_rreq_trace_ki_complete);
416 			wreq->iocb->ki_complete(
417 				wreq->iocb, wreq->error ? wreq->error : written);
418 		}
419 		wreq->iocb = VFS_PTR_POISON;
420 	}
421 
422 	netfs_clear_subrequests(wreq);
423 	return true;
424 }
425 
netfs_write_collection_worker(struct work_struct * work)426 void netfs_write_collection_worker(struct work_struct *work)
427 {
428 	struct netfs_io_request *rreq = container_of(work, struct netfs_io_request, work);
429 
430 	netfs_see_request(rreq, netfs_rreq_trace_see_work);
431 	if (netfs_check_rreq_in_progress(rreq)) {
432 		if (netfs_write_collection(rreq))
433 			/* Drop the ref from the IN_PROGRESS flag. */
434 			netfs_put_request(rreq, netfs_rreq_trace_put_work_ip);
435 		else
436 			netfs_see_request(rreq, netfs_rreq_trace_see_work_complete);
437 	}
438 }
439 
440 /**
441  * netfs_write_subrequest_terminated - Note the termination of a write operation.
442  * @_op: The I/O request that has terminated.
443  * @transferred_or_error: The amount of data transferred or an error code.
444  *
445  * This tells the library that a contributory write I/O operation has
446  * terminated, one way or another, and that it should collect the results.
447  *
448  * The caller indicates in @transferred_or_error the outcome of the operation,
449  * supplying a positive value to indicate the number of bytes transferred or a
450  * negative error code.  The library will look after reissuing I/O operations
451  * as appropriate and writing downloaded data to the cache.
452  *
453  * When this is called, ownership of the subrequest is transferred back to the
454  * library, along with a ref.
455  *
456  * Note that %_op is a void* so that the function can be passed to
457  * kiocb::term_func without the need for a casting wrapper.
458  */
netfs_write_subrequest_terminated(void * _op,ssize_t transferred_or_error)459 void netfs_write_subrequest_terminated(void *_op, ssize_t transferred_or_error)
460 {
461 	struct netfs_io_subrequest *subreq = _op;
462 	struct netfs_io_request *wreq = subreq->rreq;
463 
464 	_enter("%x[%x] %zd", wreq->debug_id, subreq->debug_index, transferred_or_error);
465 
466 	switch (subreq->source) {
467 	case NETFS_UPLOAD_TO_SERVER:
468 		netfs_stat(&netfs_n_wh_upload_done);
469 		break;
470 	case NETFS_WRITE_TO_CACHE:
471 		netfs_stat(&netfs_n_wh_write_done);
472 		break;
473 	default:
474 		BUG();
475 	}
476 
477 	if (IS_ERR_VALUE(transferred_or_error)) {
478 		subreq->error = transferred_or_error;
479 		/* if need retry is set, error should not matter */
480 		if (!test_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags)) {
481 			set_bit(NETFS_SREQ_FAILED, &subreq->flags);
482 			trace_netfs_failure(wreq, subreq, transferred_or_error, netfs_fail_write);
483 		}
484 
485 		switch (subreq->source) {
486 		case NETFS_WRITE_TO_CACHE:
487 			netfs_stat(&netfs_n_wh_write_failed);
488 			break;
489 		case NETFS_UPLOAD_TO_SERVER:
490 			netfs_stat(&netfs_n_wh_upload_failed);
491 			break;
492 		default:
493 			break;
494 		}
495 		trace_netfs_rreq(wreq, netfs_rreq_trace_set_pause);
496 		set_bit(NETFS_RREQ_PAUSE, &wreq->flags);
497 	} else {
498 		if (WARN(transferred_or_error > subreq->len - subreq->transferred,
499 			 "Subreq excess write: R=%x[%x] %zd > %zu - %zu",
500 			 wreq->debug_id, subreq->debug_index,
501 			 transferred_or_error, subreq->len, subreq->transferred))
502 			transferred_or_error = subreq->len - subreq->transferred;
503 
504 		subreq->error = 0;
505 		subreq->transferred += transferred_or_error;
506 
507 		if (subreq->transferred < subreq->len)
508 			set_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
509 	}
510 
511 	trace_netfs_sreq(subreq, netfs_sreq_trace_terminated);
512 	netfs_subreq_clear_in_progress(subreq);
513 	netfs_put_subrequest(subreq, netfs_sreq_trace_put_terminated);
514 }
515 EXPORT_SYMBOL(netfs_write_subrequest_terminated);
516