1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Functions to sequence PREFLUSH and FUA writes.
4 *
5 * Copyright (C) 2011		Max Planck Institute for Gravitational Physics
6 * Copyright (C) 2011		Tejun Heo <tj@kernel.org>
7 *
8 * REQ_{PREFLUSH|FUA} requests are decomposed to sequences consisted of three
9 * optional steps - PREFLUSH, DATA and POSTFLUSH - according to the request
10 * properties and hardware capability.
11 *
12 * If a request doesn't have data, only REQ_PREFLUSH makes sense, which
13 * indicates a simple flush request.  If there is data, REQ_PREFLUSH indicates
14 * that the device cache should be flushed before the data is executed, and
15 * REQ_FUA means that the data must be on non-volatile media on request
16 * completion.
17 *
18 * If the device doesn't have writeback cache, PREFLUSH and FUA don't make any
19 * difference.  The requests are either completed immediately if there's no data
20 * or executed as normal requests otherwise.
21 *
22 * If the device has writeback cache and supports FUA, REQ_PREFLUSH is
23 * translated to PREFLUSH but REQ_FUA is passed down directly with DATA.
24 *
25 * If the device has writeback cache and doesn't support FUA, REQ_PREFLUSH
26 * is translated to PREFLUSH and REQ_FUA to POSTFLUSH.
27 *
28 * The actual execution of flush is double buffered.  Whenever a request
29 * needs to execute PRE or POSTFLUSH, it queues at
30 * fq->flush_queue[fq->flush_pending_idx].  Once certain criteria are met, a
31 * REQ_OP_FLUSH is issued and the pending_idx is toggled.  When the flush
32 * completes, all the requests which were pending are proceeded to the next
33 * step.  This allows arbitrary merging of different types of PREFLUSH/FUA
34 * requests.
35 *
36 * Currently, the following conditions are used to determine when to issue
37 * flush.
38 *
39 * C1. At any given time, only one flush shall be in progress.  This makes
40 *     double buffering sufficient.
41 *
42 * C2. Flush is deferred if any request is executing DATA of its sequence.
43 *     This avoids issuing separate POSTFLUSHes for requests which shared
44 *     PREFLUSH.
45 *
46 * C3. The second condition is ignored if there is a request which has
47 *     waited longer than FLUSH_PENDING_TIMEOUT.  This is to avoid
48 *     starvation in the unlikely case where there are continuous stream of
49 *     FUA (without PREFLUSH) requests.
50 *
51 * For devices which support FUA, it isn't clear whether C2 (and thus C3)
52 * is beneficial.
53 *
54 * Note that a sequenced PREFLUSH/FUA request with DATA is completed twice.
55 * Once while executing DATA and again after the whole sequence is
56 * complete.  The first completion updates the contained bio but doesn't
57 * finish it so that the bio submitter is notified only after the whole
58 * sequence is complete.  This is implemented by testing RQF_FLUSH_SEQ in
59 * req_bio_endio().
60 *
61 * The above peculiarity requires that each PREFLUSH/FUA request has only one
62 * bio attached to it, which is guaranteed as they aren't allowed to be
63 * merged in the usual way.
64 */
65
66#include <linux/kernel.h>
67#include <linux/module.h>
68#include <linux/bio.h>
69#include <linux/blkdev.h>
70#include <linux/gfp.h>
71#include <linux/part_stat.h>
72
73#include "blk.h"
74#include "blk-mq.h"
75#include "blk-mq-sched.h"
76
77/* PREFLUSH/FUA sequences */
78enum {
79	REQ_FSEQ_PREFLUSH	= (1 << 0), /* pre-flushing in progress */
80	REQ_FSEQ_DATA		= (1 << 1), /* data write in progress */
81	REQ_FSEQ_POSTFLUSH	= (1 << 2), /* post-flushing in progress */
82	REQ_FSEQ_DONE		= (1 << 3),
83
84	REQ_FSEQ_ACTIONS	= REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
85				  REQ_FSEQ_POSTFLUSH,
86
87	/*
88	 * If flush has been pending longer than the following timeout,
89	 * it's issued even if flush_data requests are still in flight.
90	 */
91	FLUSH_PENDING_TIMEOUT	= 5 * HZ,
92};
93
94static void blk_kick_flush(struct request_queue *q,
95			   struct blk_flush_queue *fq, blk_opf_t flags);
96
97static inline struct blk_flush_queue *
98blk_get_flush_queue(struct request_queue *q, struct blk_mq_ctx *ctx)
99{
100	return blk_mq_map_queue(q, REQ_OP_FLUSH, ctx)->fq;
101}
102
103static unsigned int blk_flush_policy(unsigned long fflags, struct request *rq)
104{
105	unsigned int policy = 0;
106
107	if (blk_rq_sectors(rq))
108		policy |= REQ_FSEQ_DATA;
109
110	if (fflags & (1UL << QUEUE_FLAG_WC)) {
111		if (rq->cmd_flags & REQ_PREFLUSH)
112			policy |= REQ_FSEQ_PREFLUSH;
113		if (!(fflags & (1UL << QUEUE_FLAG_FUA)) &&
114		    (rq->cmd_flags & REQ_FUA))
115			policy |= REQ_FSEQ_POSTFLUSH;
116	}
117	return policy;
118}
119
120static unsigned int blk_flush_cur_seq(struct request *rq)
121{
122	return 1 << ffz(rq->flush.seq);
123}
124
125static void blk_flush_restore_request(struct request *rq)
126{
127	/*
128	 * After flush data completion, @rq->bio is %NULL but we need to
129	 * complete the bio again.  @rq->biotail is guaranteed to equal the
130	 * original @rq->bio.  Restore it.
131	 */
132	rq->bio = rq->biotail;
133
134	/* make @rq a normal request */
135	rq->rq_flags &= ~RQF_FLUSH_SEQ;
136	rq->end_io = rq->flush.saved_end_io;
137}
138
139static void blk_account_io_flush(struct request *rq)
140{
141	struct block_device *part = rq->q->disk->part0;
142
143	part_stat_lock();
144	part_stat_inc(part, ios[STAT_FLUSH]);
145	part_stat_add(part, nsecs[STAT_FLUSH],
146		      blk_time_get_ns() - rq->start_time_ns);
147	part_stat_unlock();
148}
149
150/**
151 * blk_flush_complete_seq - complete flush sequence
152 * @rq: PREFLUSH/FUA request being sequenced
153 * @fq: flush queue
154 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
155 * @error: whether an error occurred
156 *
157 * @rq just completed @seq part of its flush sequence, record the
158 * completion and trigger the next step.
159 *
160 * CONTEXT:
161 * spin_lock_irq(fq->mq_flush_lock)
162 */
163static void blk_flush_complete_seq(struct request *rq,
164				   struct blk_flush_queue *fq,
165				   unsigned int seq, blk_status_t error)
166{
167	struct request_queue *q = rq->q;
168	struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
169	blk_opf_t cmd_flags;
170
171	BUG_ON(rq->flush.seq & seq);
172	rq->flush.seq |= seq;
173	cmd_flags = rq->cmd_flags;
174
175	if (likely(!error))
176		seq = blk_flush_cur_seq(rq);
177	else
178		seq = REQ_FSEQ_DONE;
179
180	switch (seq) {
181	case REQ_FSEQ_PREFLUSH:
182	case REQ_FSEQ_POSTFLUSH:
183		/* queue for flush */
184		if (list_empty(pending))
185			fq->flush_pending_since = jiffies;
186		list_move_tail(&rq->queuelist, pending);
187		break;
188
189	case REQ_FSEQ_DATA:
190		fq->flush_data_in_flight++;
191		spin_lock(&q->requeue_lock);
192		list_move(&rq->queuelist, &q->requeue_list);
193		spin_unlock(&q->requeue_lock);
194		blk_mq_kick_requeue_list(q);
195		break;
196
197	case REQ_FSEQ_DONE:
198		/*
199		 * @rq was previously adjusted by blk_insert_flush() for
200		 * flush sequencing and may already have gone through the
201		 * flush data request completion path.  Restore @rq for
202		 * normal completion and end it.
203		 */
204		list_del_init(&rq->queuelist);
205		blk_flush_restore_request(rq);
206		blk_mq_end_request(rq, error);
207		break;
208
209	default:
210		BUG();
211	}
212
213	blk_kick_flush(q, fq, cmd_flags);
214}
215
216static enum rq_end_io_ret flush_end_io(struct request *flush_rq,
217				       blk_status_t error)
218{
219	struct request_queue *q = flush_rq->q;
220	struct list_head *running;
221	struct request *rq, *n;
222	unsigned long flags = 0;
223	struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
224
225	/* release the tag's ownership to the req cloned from */
226	spin_lock_irqsave(&fq->mq_flush_lock, flags);
227
228	if (!req_ref_put_and_test(flush_rq)) {
229		fq->rq_status = error;
230		spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
231		return RQ_END_IO_NONE;
232	}
233
234	blk_account_io_flush(flush_rq);
235	/*
236	 * Flush request has to be marked as IDLE when it is really ended
237	 * because its .end_io() is called from timeout code path too for
238	 * avoiding use-after-free.
239	 */
240	WRITE_ONCE(flush_rq->state, MQ_RQ_IDLE);
241	if (fq->rq_status != BLK_STS_OK) {
242		error = fq->rq_status;
243		fq->rq_status = BLK_STS_OK;
244	}
245
246	if (!q->elevator) {
247		flush_rq->tag = BLK_MQ_NO_TAG;
248	} else {
249		blk_mq_put_driver_tag(flush_rq);
250		flush_rq->internal_tag = BLK_MQ_NO_TAG;
251	}
252
253	running = &fq->flush_queue[fq->flush_running_idx];
254	BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
255
256	/* account completion of the flush request */
257	fq->flush_running_idx ^= 1;
258
259	/* and push the waiting requests to the next stage */
260	list_for_each_entry_safe(rq, n, running, queuelist) {
261		unsigned int seq = blk_flush_cur_seq(rq);
262
263		BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
264		blk_flush_complete_seq(rq, fq, seq, error);
265	}
266
267	spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
268	return RQ_END_IO_NONE;
269}
270
271bool is_flush_rq(struct request *rq)
272{
273	return rq->end_io == flush_end_io;
274}
275
276/**
277 * blk_kick_flush - consider issuing flush request
278 * @q: request_queue being kicked
279 * @fq: flush queue
280 * @flags: cmd_flags of the original request
281 *
282 * Flush related states of @q have changed, consider issuing flush request.
283 * Please read the comment at the top of this file for more info.
284 *
285 * CONTEXT:
286 * spin_lock_irq(fq->mq_flush_lock)
287 *
288 */
289static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
290			   blk_opf_t flags)
291{
292	struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
293	struct request *first_rq =
294		list_first_entry(pending, struct request, queuelist);
295	struct request *flush_rq = fq->flush_rq;
296
297	/* C1 described at the top of this file */
298	if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
299		return;
300
301	/* C2 and C3 */
302	if (fq->flush_data_in_flight &&
303	    time_before(jiffies,
304			fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
305		return;
306
307	/*
308	 * Issue flush and toggle pending_idx.  This makes pending_idx
309	 * different from running_idx, which means flush is in flight.
310	 */
311	fq->flush_pending_idx ^= 1;
312
313	blk_rq_init(q, flush_rq);
314
315	/*
316	 * In case of none scheduler, borrow tag from the first request
317	 * since they can't be in flight at the same time. And acquire
318	 * the tag's ownership for flush req.
319	 *
320	 * In case of IO scheduler, flush rq need to borrow scheduler tag
321	 * just for cheating put/get driver tag.
322	 */
323	flush_rq->mq_ctx = first_rq->mq_ctx;
324	flush_rq->mq_hctx = first_rq->mq_hctx;
325
326	if (!q->elevator)
327		flush_rq->tag = first_rq->tag;
328	else
329		flush_rq->internal_tag = first_rq->internal_tag;
330
331	flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
332	flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK);
333	flush_rq->rq_flags |= RQF_FLUSH_SEQ;
334	flush_rq->end_io = flush_end_io;
335	/*
336	 * Order WRITE ->end_io and WRITE rq->ref, and its pair is the one
337	 * implied in refcount_inc_not_zero() called from
338	 * blk_mq_find_and_get_req(), which orders WRITE/READ flush_rq->ref
339	 * and READ flush_rq->end_io
340	 */
341	smp_wmb();
342	req_ref_set(flush_rq, 1);
343
344	spin_lock(&q->requeue_lock);
345	list_add_tail(&flush_rq->queuelist, &q->flush_list);
346	spin_unlock(&q->requeue_lock);
347
348	blk_mq_kick_requeue_list(q);
349}
350
351static enum rq_end_io_ret mq_flush_data_end_io(struct request *rq,
352					       blk_status_t error)
353{
354	struct request_queue *q = rq->q;
355	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
356	struct blk_mq_ctx *ctx = rq->mq_ctx;
357	unsigned long flags;
358	struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
359
360	if (q->elevator) {
361		WARN_ON(rq->tag < 0);
362		blk_mq_put_driver_tag(rq);
363	}
364
365	/*
366	 * After populating an empty queue, kick it to avoid stall.  Read
367	 * the comment in flush_end_io().
368	 */
369	spin_lock_irqsave(&fq->mq_flush_lock, flags);
370	fq->flush_data_in_flight--;
371	/*
372	 * May have been corrupted by rq->rq_next reuse, we need to
373	 * re-initialize rq->queuelist before reusing it here.
374	 */
375	INIT_LIST_HEAD(&rq->queuelist);
376	blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
377	spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
378
379	blk_mq_sched_restart(hctx);
380	return RQ_END_IO_NONE;
381}
382
383static void blk_rq_init_flush(struct request *rq)
384{
385	rq->flush.seq = 0;
386	rq->rq_flags |= RQF_FLUSH_SEQ;
387	rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
388	rq->end_io = mq_flush_data_end_io;
389}
390
391/*
392 * Insert a PREFLUSH/FUA request into the flush state machine.
393 * Returns true if the request has been consumed by the flush state machine,
394 * or false if the caller should continue to process it.
395 */
396bool blk_insert_flush(struct request *rq)
397{
398	struct request_queue *q = rq->q;
399	unsigned long fflags = q->queue_flags;	/* may change, cache */
400	unsigned int policy = blk_flush_policy(fflags, rq);
401	struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
402
403	/* FLUSH/FUA request must never be merged */
404	WARN_ON_ONCE(rq->bio != rq->biotail);
405
406	/*
407	 * @policy now records what operations need to be done.  Adjust
408	 * REQ_PREFLUSH and FUA for the driver.
409	 */
410	rq->cmd_flags &= ~REQ_PREFLUSH;
411	if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
412		rq->cmd_flags &= ~REQ_FUA;
413
414	/*
415	 * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
416	 * of those flags, we have to set REQ_SYNC to avoid skewing
417	 * the request accounting.
418	 */
419	rq->cmd_flags |= REQ_SYNC;
420
421	switch (policy) {
422	case 0:
423		/*
424		 * An empty flush handed down from a stacking driver may
425		 * translate into nothing if the underlying device does not
426		 * advertise a write-back cache.  In this case, simply
427		 * complete the request.
428		 */
429		blk_mq_end_request(rq, 0);
430		return true;
431	case REQ_FSEQ_DATA:
432		/*
433		 * If there's data, but no flush is necessary, the request can
434		 * be processed directly without going through flush machinery.
435		 * Queue for normal execution.
436		 */
437		return false;
438	case REQ_FSEQ_DATA | REQ_FSEQ_POSTFLUSH:
439		/*
440		 * Initialize the flush fields and completion handler to trigger
441		 * the post flush, and then just pass the command on.
442		 */
443		blk_rq_init_flush(rq);
444		rq->flush.seq |= REQ_FSEQ_PREFLUSH;
445		spin_lock_irq(&fq->mq_flush_lock);
446		fq->flush_data_in_flight++;
447		spin_unlock_irq(&fq->mq_flush_lock);
448		return false;
449	default:
450		/*
451		 * Mark the request as part of a flush sequence and submit it
452		 * for further processing to the flush state machine.
453		 */
454		blk_rq_init_flush(rq);
455		spin_lock_irq(&fq->mq_flush_lock);
456		blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
457		spin_unlock_irq(&fq->mq_flush_lock);
458		return true;
459	}
460}
461
462/**
463 * blkdev_issue_flush - queue a flush
464 * @bdev:	blockdev to issue flush for
465 *
466 * Description:
467 *    Issue a flush for the block device in question.
468 */
469int blkdev_issue_flush(struct block_device *bdev)
470{
471	struct bio bio;
472
473	bio_init(&bio, bdev, NULL, 0, REQ_OP_WRITE | REQ_PREFLUSH);
474	return submit_bio_wait(&bio);
475}
476EXPORT_SYMBOL(blkdev_issue_flush);
477
478struct blk_flush_queue *blk_alloc_flush_queue(int node, int cmd_size,
479					      gfp_t flags)
480{
481	struct blk_flush_queue *fq;
482	int rq_sz = sizeof(struct request);
483
484	fq = kzalloc_node(sizeof(*fq), flags, node);
485	if (!fq)
486		goto fail;
487
488	spin_lock_init(&fq->mq_flush_lock);
489
490	rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
491	fq->flush_rq = kzalloc_node(rq_sz, flags, node);
492	if (!fq->flush_rq)
493		goto fail_rq;
494
495	INIT_LIST_HEAD(&fq->flush_queue[0]);
496	INIT_LIST_HEAD(&fq->flush_queue[1]);
497
498	return fq;
499
500 fail_rq:
501	kfree(fq);
502 fail:
503	return NULL;
504}
505
506void blk_free_flush_queue(struct blk_flush_queue *fq)
507{
508	/* bio based request queue hasn't flush queue */
509	if (!fq)
510		return;
511
512	kfree(fq->flush_rq);
513	kfree(fq);
514}
515
516/*
517 * Allow driver to set its own lock class to fq->mq_flush_lock for
518 * avoiding lockdep complaint.
519 *
520 * flush_end_io() may be called recursively from some driver, such as
521 * nvme-loop, so lockdep may complain 'possible recursive locking' because
522 * all 'struct blk_flush_queue' instance share same mq_flush_lock lock class
523 * key. We need to assign different lock class for these driver's
524 * fq->mq_flush_lock for avoiding the lockdep warning.
525 *
526 * Use dynamically allocated lock class key for each 'blk_flush_queue'
527 * instance is over-kill, and more worse it introduces horrible boot delay
528 * issue because synchronize_rcu() is implied in lockdep_unregister_key which
529 * is called for each hctx release. SCSI probing may synchronously create and
530 * destroy lots of MQ request_queues for non-existent devices, and some robot
531 * test kernel always enable lockdep option. It is observed that more than half
532 * an hour is taken during SCSI MQ probe with per-fq lock class.
533 */
534void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx,
535		struct lock_class_key *key)
536{
537	lockdep_set_class(&hctx->fq->mq_flush_lock, key);
538}
539EXPORT_SYMBOL_GPL(blk_mq_hctx_set_fq_lock_class);
540