1/*
2 * Functions to sequence FLUSH and FUA writes.
3 *
4 * Copyright (C) 2011		Max Planck Institute for Gravitational Physics
5 * Copyright (C) 2011		Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * REQ_{FLUSH|FUA} requests are decomposed to sequences consisted of three
10 * optional steps - PREFLUSH, DATA and POSTFLUSH - according to the request
11 * properties and hardware capability.
12 *
13 * If a request doesn't have data, only REQ_FLUSH makes sense, which
14 * indicates a simple flush request.  If there is data, REQ_FLUSH indicates
15 * that the device cache should be flushed before the data is executed, and
16 * REQ_FUA means that the data must be on non-volatile media on request
17 * completion.
18 *
19 * If the device doesn't have writeback cache, FLUSH and FUA don't make any
20 * difference.  The requests are either completed immediately if there's no
21 * data or executed as normal requests otherwise.
22 *
23 * If the device has writeback cache and supports FUA, REQ_FLUSH is
24 * translated to PREFLUSH but REQ_FUA is passed down directly with DATA.
25 *
26 * If the device has writeback cache and doesn't support FUA, REQ_FLUSH is
27 * translated to PREFLUSH and REQ_FUA to POSTFLUSH.
28 *
29 * The actual execution of flush is double buffered.  Whenever a request
30 * needs to execute PRE or POSTFLUSH, it queues at
31 * fq->flush_queue[fq->flush_pending_idx].  Once certain criteria are met, a
32 * flush is issued and the pending_idx is toggled.  When the flush
33 * completes, all the requests which were pending are proceeded to the next
34 * step.  This allows arbitrary merging of different types of FLUSH/FUA
35 * requests.
36 *
37 * Currently, the following conditions are used to determine when to issue
38 * flush.
39 *
40 * C1. At any given time, only one flush shall be in progress.  This makes
41 *     double buffering sufficient.
42 *
43 * C2. Flush is deferred if any request is executing DATA of its sequence.
44 *     This avoids issuing separate POSTFLUSHes for requests which shared
45 *     PREFLUSH.
46 *
47 * C3. The second condition is ignored if there is a request which has
48 *     waited longer than FLUSH_PENDING_TIMEOUT.  This is to avoid
49 *     starvation in the unlikely case where there are continuous stream of
50 *     FUA (without FLUSH) requests.
51 *
52 * For devices which support FUA, it isn't clear whether C2 (and thus C3)
53 * is beneficial.
54 *
55 * Note that a sequenced FLUSH/FUA request with DATA is completed twice.
56 * Once while executing DATA and again after the whole sequence is
57 * complete.  The first completion updates the contained bio but doesn't
58 * finish it so that the bio submitter is notified only after the whole
59 * sequence is complete.  This is implemented by testing REQ_FLUSH_SEQ in
60 * req_bio_endio().
61 *
62 * The above peculiarity requires that each FLUSH/FUA request has only one
63 * bio attached to it, which is guaranteed as they aren't allowed to be
64 * merged in the usual way.
65 */
66
67#include <linux/kernel.h>
68#include <linux/module.h>
69#include <linux/bio.h>
70#include <linux/blkdev.h>
71#include <linux/gfp.h>
72#include <linux/blk-mq.h>
73
74#include "blk.h"
75#include "blk-mq.h"
76#include "blk-mq-tag.h"
77
78/* FLUSH/FUA sequences */
79enum {
80	REQ_FSEQ_PREFLUSH	= (1 << 0), /* pre-flushing in progress */
81	REQ_FSEQ_DATA		= (1 << 1), /* data write in progress */
82	REQ_FSEQ_POSTFLUSH	= (1 << 2), /* post-flushing in progress */
83	REQ_FSEQ_DONE		= (1 << 3),
84
85	REQ_FSEQ_ACTIONS	= REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
86				  REQ_FSEQ_POSTFLUSH,
87
88	/*
89	 * If flush has been pending longer than the following timeout,
90	 * it's issued even if flush_data requests are still in flight.
91	 */
92	FLUSH_PENDING_TIMEOUT	= 5 * HZ,
93};
94
95static bool blk_kick_flush(struct request_queue *q,
96			   struct blk_flush_queue *fq);
97
98static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
99{
100	unsigned int policy = 0;
101
102	if (blk_rq_sectors(rq))
103		policy |= REQ_FSEQ_DATA;
104
105	if (fflags & REQ_FLUSH) {
106		if (rq->cmd_flags & REQ_FLUSH)
107			policy |= REQ_FSEQ_PREFLUSH;
108		if (!(fflags & REQ_FUA) && (rq->cmd_flags & REQ_FUA))
109			policy |= REQ_FSEQ_POSTFLUSH;
110	}
111	return policy;
112}
113
114static unsigned int blk_flush_cur_seq(struct request *rq)
115{
116	return 1 << ffz(rq->flush.seq);
117}
118
119static void blk_flush_restore_request(struct request *rq)
120{
121	/*
122	 * After flush data completion, @rq->bio is %NULL but we need to
123	 * complete the bio again.  @rq->biotail is guaranteed to equal the
124	 * original @rq->bio.  Restore it.
125	 */
126	rq->bio = rq->biotail;
127
128	/* make @rq a normal request */
129	rq->cmd_flags &= ~REQ_FLUSH_SEQ;
130	rq->end_io = rq->flush.saved_end_io;
131}
132
133static bool blk_flush_queue_rq(struct request *rq, bool add_front)
134{
135	if (rq->q->mq_ops) {
136		struct request_queue *q = rq->q;
137
138		blk_mq_add_to_requeue_list(rq, add_front);
139		blk_mq_kick_requeue_list(q);
140		return false;
141	} else {
142		if (add_front)
143			list_add(&rq->queuelist, &rq->q->queue_head);
144		else
145			list_add_tail(&rq->queuelist, &rq->q->queue_head);
146		return true;
147	}
148}
149
150/**
151 * blk_flush_complete_seq - complete flush sequence
152 * @rq: FLUSH/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(q->queue_lock or fq->mq_flush_lock)
162 *
163 * RETURNS:
164 * %true if requests were added to the dispatch queue, %false otherwise.
165 */
166static bool blk_flush_complete_seq(struct request *rq,
167				   struct blk_flush_queue *fq,
168				   unsigned int seq, int error)
169{
170	struct request_queue *q = rq->q;
171	struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
172	bool queued = false, kicked;
173
174	BUG_ON(rq->flush.seq & seq);
175	rq->flush.seq |= seq;
176
177	if (likely(!error))
178		seq = blk_flush_cur_seq(rq);
179	else
180		seq = REQ_FSEQ_DONE;
181
182	switch (seq) {
183	case REQ_FSEQ_PREFLUSH:
184	case REQ_FSEQ_POSTFLUSH:
185		/* queue for flush */
186		if (list_empty(pending))
187			fq->flush_pending_since = jiffies;
188		list_move_tail(&rq->flush.list, pending);
189		break;
190
191	case REQ_FSEQ_DATA:
192		list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
193		queued = blk_flush_queue_rq(rq, true);
194		break;
195
196	case REQ_FSEQ_DONE:
197		/*
198		 * @rq was previously adjusted by blk_flush_issue() for
199		 * flush sequencing and may already have gone through the
200		 * flush data request completion path.  Restore @rq for
201		 * normal completion and end it.
202		 */
203		BUG_ON(!list_empty(&rq->queuelist));
204		list_del_init(&rq->flush.list);
205		blk_flush_restore_request(rq);
206		if (q->mq_ops)
207			blk_mq_end_request(rq, error);
208		else
209			__blk_end_request_all(rq, error);
210		break;
211
212	default:
213		BUG();
214	}
215
216	kicked = blk_kick_flush(q, fq);
217	return kicked | queued;
218}
219
220static void flush_end_io(struct request *flush_rq, int error)
221{
222	struct request_queue *q = flush_rq->q;
223	struct list_head *running;
224	bool queued = false;
225	struct request *rq, *n;
226	unsigned long flags = 0;
227	struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
228
229	if (q->mq_ops) {
230		struct blk_mq_hw_ctx *hctx;
231
232		/* release the tag's ownership to the req cloned from */
233		spin_lock_irqsave(&fq->mq_flush_lock, flags);
234		hctx = q->mq_ops->map_queue(q, flush_rq->mq_ctx->cpu);
235		blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq);
236		flush_rq->tag = -1;
237	}
238
239	running = &fq->flush_queue[fq->flush_running_idx];
240	BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
241
242	/* account completion of the flush request */
243	fq->flush_running_idx ^= 1;
244
245	if (!q->mq_ops)
246		elv_completed_request(q, flush_rq);
247
248	/* and push the waiting requests to the next stage */
249	list_for_each_entry_safe(rq, n, running, flush.list) {
250		unsigned int seq = blk_flush_cur_seq(rq);
251
252		BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
253		queued |= blk_flush_complete_seq(rq, fq, seq, error);
254	}
255
256	/*
257	 * Kick the queue to avoid stall for two cases:
258	 * 1. Moving a request silently to empty queue_head may stall the
259	 * queue.
260	 * 2. When flush request is running in non-queueable queue, the
261	 * queue is hold. Restart the queue after flush request is finished
262	 * to avoid stall.
263	 * This function is called from request completion path and calling
264	 * directly into request_fn may confuse the driver.  Always use
265	 * kblockd.
266	 */
267	if (queued || fq->flush_queue_delayed) {
268		WARN_ON(q->mq_ops);
269		blk_run_queue_async(q);
270	}
271	fq->flush_queue_delayed = 0;
272	if (q->mq_ops)
273		spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
274}
275
276/**
277 * blk_kick_flush - consider issuing flush request
278 * @q: request_queue being kicked
279 * @fq: flush queue
280 *
281 * Flush related states of @q have changed, consider issuing flush request.
282 * Please read the comment at the top of this file for more info.
283 *
284 * CONTEXT:
285 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
286 *
287 * RETURNS:
288 * %true if flush was issued, %false otherwise.
289 */
290static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq)
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, flush.list);
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 false;
300
301	/* C2 and C3 */
302	if (!list_empty(&fq->flush_data_in_flight) &&
303	    time_before(jiffies,
304			fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
305		return false;
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	 * Borrow tag from the first request since they can't
317	 * be in flight at the same time. And acquire the tag's
318	 * ownership for flush req.
319	 */
320	if (q->mq_ops) {
321		struct blk_mq_hw_ctx *hctx;
322
323		flush_rq->mq_ctx = first_rq->mq_ctx;
324		flush_rq->tag = first_rq->tag;
325		fq->orig_rq = first_rq;
326
327		hctx = q->mq_ops->map_queue(q, first_rq->mq_ctx->cpu);
328		blk_mq_tag_set_rq(hctx, first_rq->tag, flush_rq);
329	}
330
331	flush_rq->cmd_type = REQ_TYPE_FS;
332	flush_rq->cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
333	flush_rq->rq_disk = first_rq->rq_disk;
334	flush_rq->end_io = flush_end_io;
335
336	return blk_flush_queue_rq(flush_rq, false);
337}
338
339static void flush_data_end_io(struct request *rq, int error)
340{
341	struct request_queue *q = rq->q;
342	struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL);
343
344	/*
345	 * After populating an empty queue, kick it to avoid stall.  Read
346	 * the comment in flush_end_io().
347	 */
348	if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
349		blk_run_queue_async(q);
350}
351
352static void mq_flush_data_end_io(struct request *rq, int error)
353{
354	struct request_queue *q = rq->q;
355	struct blk_mq_hw_ctx *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	hctx = q->mq_ops->map_queue(q, ctx->cpu);
361
362	/*
363	 * After populating an empty queue, kick it to avoid stall.  Read
364	 * the comment in flush_end_io().
365	 */
366	spin_lock_irqsave(&fq->mq_flush_lock, flags);
367	if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
368		blk_mq_run_hw_queue(hctx, true);
369	spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
370}
371
372/**
373 * blk_insert_flush - insert a new FLUSH/FUA request
374 * @rq: request to insert
375 *
376 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
377 * or __blk_mq_run_hw_queue() to dispatch request.
378 * @rq is being submitted.  Analyze what needs to be done and put it on the
379 * right queue.
380 *
381 * CONTEXT:
382 * spin_lock_irq(q->queue_lock) in !mq case
383 */
384void blk_insert_flush(struct request *rq)
385{
386	struct request_queue *q = rq->q;
387	unsigned int fflags = q->flush_flags;	/* may change, cache */
388	unsigned int policy = blk_flush_policy(fflags, rq);
389	struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
390
391	/*
392	 * @policy now records what operations need to be done.  Adjust
393	 * REQ_FLUSH and FUA for the driver.
394	 */
395	rq->cmd_flags &= ~REQ_FLUSH;
396	if (!(fflags & REQ_FUA))
397		rq->cmd_flags &= ~REQ_FUA;
398
399	/*
400	 * An empty flush handed down from a stacking driver may
401	 * translate into nothing if the underlying device does not
402	 * advertise a write-back cache.  In this case, simply
403	 * complete the request.
404	 */
405	if (!policy) {
406		if (q->mq_ops)
407			blk_mq_end_request(rq, 0);
408		else
409			__blk_end_bidi_request(rq, 0, 0, 0);
410		return;
411	}
412
413	BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
414
415	/*
416	 * If there's data but flush is not necessary, the request can be
417	 * processed directly without going through flush machinery.  Queue
418	 * for normal execution.
419	 */
420	if ((policy & REQ_FSEQ_DATA) &&
421	    !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
422		if (q->mq_ops) {
423			blk_mq_insert_request(rq, false, false, true);
424		} else
425			list_add_tail(&rq->queuelist, &q->queue_head);
426		return;
427	}
428
429	/*
430	 * @rq should go through flush machinery.  Mark it part of flush
431	 * sequence and submit for further processing.
432	 */
433	memset(&rq->flush, 0, sizeof(rq->flush));
434	INIT_LIST_HEAD(&rq->flush.list);
435	rq->cmd_flags |= REQ_FLUSH_SEQ;
436	rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
437	if (q->mq_ops) {
438		rq->end_io = mq_flush_data_end_io;
439
440		spin_lock_irq(&fq->mq_flush_lock);
441		blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
442		spin_unlock_irq(&fq->mq_flush_lock);
443		return;
444	}
445	rq->end_io = flush_data_end_io;
446
447	blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
448}
449
450/**
451 * blkdev_issue_flush - queue a flush
452 * @bdev:	blockdev to issue flush for
453 * @gfp_mask:	memory allocation flags (for bio_alloc)
454 * @error_sector:	error sector
455 *
456 * Description:
457 *    Issue a flush for the block device in question. Caller can supply
458 *    room for storing the error offset in case of a flush error, if they
459 *    wish to. If WAIT flag is not passed then caller may check only what
460 *    request was pushed in some internal queue for later handling.
461 */
462int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
463		sector_t *error_sector)
464{
465	struct request_queue *q;
466	struct bio *bio;
467	int ret = 0;
468
469	if (bdev->bd_disk == NULL)
470		return -ENXIO;
471
472	q = bdev_get_queue(bdev);
473	if (!q)
474		return -ENXIO;
475
476	/*
477	 * some block devices may not have their queue correctly set up here
478	 * (e.g. loop device without a backing file) and so issuing a flush
479	 * here will panic. Ensure there is a request function before issuing
480	 * the flush.
481	 */
482	if (!q->make_request_fn)
483		return -ENXIO;
484
485	bio = bio_alloc(gfp_mask, 0);
486	bio->bi_bdev = bdev;
487
488	ret = submit_bio_wait(WRITE_FLUSH, bio);
489
490	/*
491	 * The driver must store the error location in ->bi_sector, if
492	 * it supports it. For non-stacked drivers, this should be
493	 * copied from blk_rq_pos(rq).
494	 */
495	if (error_sector)
496		*error_sector = bio->bi_iter.bi_sector;
497
498	bio_put(bio);
499	return ret;
500}
501EXPORT_SYMBOL(blkdev_issue_flush);
502
503struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
504		int node, int cmd_size)
505{
506	struct blk_flush_queue *fq;
507	int rq_sz = sizeof(struct request);
508
509	fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node);
510	if (!fq)
511		goto fail;
512
513	if (q->mq_ops) {
514		spin_lock_init(&fq->mq_flush_lock);
515		rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
516	}
517
518	fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node);
519	if (!fq->flush_rq)
520		goto fail_rq;
521
522	INIT_LIST_HEAD(&fq->flush_queue[0]);
523	INIT_LIST_HEAD(&fq->flush_queue[1]);
524	INIT_LIST_HEAD(&fq->flush_data_in_flight);
525
526	return fq;
527
528 fail_rq:
529	kfree(fq);
530 fail:
531	return NULL;
532}
533
534void blk_free_flush_queue(struct blk_flush_queue *fq)
535{
536	/* bio based request queue hasn't flush queue */
537	if (!fq)
538		return;
539
540	kfree(fq->flush_rq);
541	kfree(fq);
542}
543