2 * Functions to sequence FLUSH and FUA writes.
4 * Copyright (C) 2011 Max Planck Institute for Gravitational Physics
5 * Copyright (C) 2011 Tejun Heo <tj@kernel.org>
7 * This file is released under the GPLv2.
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.
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
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.
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.
26 * If the device has writeback cache and doesn't support FUA, REQ_FLUSH is
27 * translated to PREFLUSH and REQ_FUA to POSTFLUSH.
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
37 * Currently, the following conditions are used to determine when to issue
40 * C1. At any given time, only one flush shall be in progress. This makes
41 * double buffering sufficient.
43 * C2. Flush is deferred if any request is executing DATA of its sequence.
44 * This avoids issuing separate POSTFLUSHes for requests which shared
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.
52 * For devices which support FUA, it isn't clear whether C2 (and thus C3)
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
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.
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>
77 /* FLUSH/FUA sequences */
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),
84 REQ_FSEQ_ACTIONS = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
88 * If flush has been pending longer than the following timeout,
89 * it's issued even if flush_data requests are still in flight.
91 FLUSH_PENDING_TIMEOUT = 5 * HZ,
94 static bool blk_kick_flush(struct request_queue *q,
95 struct blk_flush_queue *fq);
97 static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
99 unsigned int policy = 0;
101 if (blk_rq_sectors(rq))
102 policy |= REQ_FSEQ_DATA;
104 if (fflags & REQ_FLUSH) {
105 if (rq->cmd_flags & REQ_FLUSH)
106 policy |= REQ_FSEQ_PREFLUSH;
107 if (!(fflags & REQ_FUA) && (rq->cmd_flags & REQ_FUA))
108 policy |= REQ_FSEQ_POSTFLUSH;
113 static unsigned int blk_flush_cur_seq(struct request *rq)
115 return 1 << ffz(rq->flush.seq);
118 static void blk_flush_restore_request(struct request *rq)
121 * After flush data completion, @rq->bio is %NULL but we need to
122 * complete the bio again. @rq->biotail is guaranteed to equal the
123 * original @rq->bio. Restore it.
125 rq->bio = rq->biotail;
127 /* make @rq a normal request */
128 rq->cmd_flags &= ~REQ_FLUSH_SEQ;
129 rq->end_io = rq->flush.saved_end_io;
132 static bool blk_flush_queue_rq(struct request *rq, bool add_front)
135 struct request_queue *q = rq->q;
137 blk_mq_add_to_requeue_list(rq, add_front);
138 blk_mq_kick_requeue_list(q);
142 list_add(&rq->queuelist, &rq->q->queue_head);
144 list_add_tail(&rq->queuelist, &rq->q->queue_head);
150 * blk_flush_complete_seq - complete flush sequence
151 * @rq: FLUSH/FUA request being sequenced
153 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
154 * @error: whether an error occurred
156 * @rq just completed @seq part of its flush sequence, record the
157 * completion and trigger the next step.
160 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
163 * %true if requests were added to the dispatch queue, %false otherwise.
165 static bool blk_flush_complete_seq(struct request *rq,
166 struct blk_flush_queue *fq,
167 unsigned int seq, int error)
169 struct request_queue *q = rq->q;
170 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
171 bool queued = false, kicked;
173 BUG_ON(rq->flush.seq & seq);
174 rq->flush.seq |= seq;
177 seq = blk_flush_cur_seq(rq);
182 case REQ_FSEQ_PREFLUSH:
183 case REQ_FSEQ_POSTFLUSH:
184 /* queue for flush */
185 if (list_empty(pending))
186 fq->flush_pending_since = jiffies;
187 list_move_tail(&rq->flush.list, pending);
191 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
192 queued = blk_flush_queue_rq(rq, true);
197 * @rq was previously adjusted by blk_flush_issue() for
198 * flush sequencing and may already have gone through the
199 * flush data request completion path. Restore @rq for
200 * normal completion and end it.
202 BUG_ON(!list_empty(&rq->queuelist));
203 list_del_init(&rq->flush.list);
204 blk_flush_restore_request(rq);
206 blk_mq_end_request(rq, error);
208 __blk_end_request_all(rq, error);
215 kicked = blk_kick_flush(q, fq);
216 return kicked | queued;
219 static void flush_end_io(struct request *flush_rq, int error)
221 struct request_queue *q = flush_rq->q;
222 struct list_head *running;
224 struct request *rq, *n;
225 unsigned long flags = 0;
226 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
229 spin_lock_irqsave(&fq->mq_flush_lock, flags);
233 running = &fq->flush_queue[fq->flush_running_idx];
234 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
236 /* account completion of the flush request */
237 fq->flush_running_idx ^= 1;
240 elv_completed_request(q, flush_rq);
242 /* and push the waiting requests to the next stage */
243 list_for_each_entry_safe(rq, n, running, flush.list) {
244 unsigned int seq = blk_flush_cur_seq(rq);
246 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
247 queued |= blk_flush_complete_seq(rq, fq, seq, error);
251 * Kick the queue to avoid stall for two cases:
252 * 1. Moving a request silently to empty queue_head may stall the
254 * 2. When flush request is running in non-queueable queue, the
255 * queue is hold. Restart the queue after flush request is finished
257 * This function is called from request completion path and calling
258 * directly into request_fn may confuse the driver. Always use
261 if (queued || fq->flush_queue_delayed) {
263 blk_run_queue_async(q);
265 fq->flush_queue_delayed = 0;
267 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
271 * blk_kick_flush - consider issuing flush request
272 * @q: request_queue being kicked
275 * Flush related states of @q have changed, consider issuing flush request.
276 * Please read the comment at the top of this file for more info.
279 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
282 * %true if flush was issued, %false otherwise.
284 static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq)
286 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
287 struct request *first_rq =
288 list_first_entry(pending, struct request, flush.list);
289 struct request *flush_rq = fq->flush_rq;
291 /* C1 described at the top of this file */
292 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
296 if (!list_empty(&fq->flush_data_in_flight) &&
298 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
302 * Issue flush and toggle pending_idx. This makes pending_idx
303 * different from running_idx, which means flush is in flight.
305 fq->flush_pending_idx ^= 1;
307 blk_rq_init(q, flush_rq);
310 * Borrow tag from the first request since they can't
311 * be in flight at the same time.
314 flush_rq->mq_ctx = first_rq->mq_ctx;
315 flush_rq->tag = first_rq->tag;
318 flush_rq->cmd_type = REQ_TYPE_FS;
319 flush_rq->cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
320 flush_rq->rq_disk = first_rq->rq_disk;
321 flush_rq->end_io = flush_end_io;
323 return blk_flush_queue_rq(flush_rq, false);
326 static void flush_data_end_io(struct request *rq, int error)
328 struct request_queue *q = rq->q;
329 struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL);
332 * After populating an empty queue, kick it to avoid stall. Read
333 * the comment in flush_end_io().
335 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
336 blk_run_queue_async(q);
339 static void mq_flush_data_end_io(struct request *rq, int error)
341 struct request_queue *q = rq->q;
342 struct blk_mq_hw_ctx *hctx;
343 struct blk_mq_ctx *ctx = rq->mq_ctx;
345 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
347 hctx = q->mq_ops->map_queue(q, ctx->cpu);
350 * After populating an empty queue, kick it to avoid stall. Read
351 * the comment in flush_end_io().
353 spin_lock_irqsave(&fq->mq_flush_lock, flags);
354 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
355 blk_mq_run_hw_queue(hctx, true);
356 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
360 * blk_insert_flush - insert a new FLUSH/FUA request
361 * @rq: request to insert
363 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
364 * or __blk_mq_run_hw_queue() to dispatch request.
365 * @rq is being submitted. Analyze what needs to be done and put it on the
369 * spin_lock_irq(q->queue_lock) in !mq case
371 void blk_insert_flush(struct request *rq)
373 struct request_queue *q = rq->q;
374 unsigned int fflags = q->flush_flags; /* may change, cache */
375 unsigned int policy = blk_flush_policy(fflags, rq);
376 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
379 * @policy now records what operations need to be done. Adjust
380 * REQ_FLUSH and FUA for the driver.
382 rq->cmd_flags &= ~REQ_FLUSH;
383 if (!(fflags & REQ_FUA))
384 rq->cmd_flags &= ~REQ_FUA;
387 * An empty flush handed down from a stacking driver may
388 * translate into nothing if the underlying device does not
389 * advertise a write-back cache. In this case, simply
390 * complete the request.
394 blk_mq_end_request(rq, 0);
396 __blk_end_bidi_request(rq, 0, 0, 0);
400 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
403 * If there's data but flush is not necessary, the request can be
404 * processed directly without going through flush machinery. Queue
405 * for normal execution.
407 if ((policy & REQ_FSEQ_DATA) &&
408 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
410 blk_mq_insert_request(rq, false, false, true);
412 list_add_tail(&rq->queuelist, &q->queue_head);
417 * @rq should go through flush machinery. Mark it part of flush
418 * sequence and submit for further processing.
420 memset(&rq->flush, 0, sizeof(rq->flush));
421 INIT_LIST_HEAD(&rq->flush.list);
422 rq->cmd_flags |= REQ_FLUSH_SEQ;
423 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
425 rq->end_io = mq_flush_data_end_io;
427 spin_lock_irq(&fq->mq_flush_lock);
428 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
429 spin_unlock_irq(&fq->mq_flush_lock);
432 rq->end_io = flush_data_end_io;
434 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
438 * blkdev_issue_flush - queue a flush
439 * @bdev: blockdev to issue flush for
440 * @gfp_mask: memory allocation flags (for bio_alloc)
441 * @error_sector: error sector
444 * Issue a flush for the block device in question. Caller can supply
445 * room for storing the error offset in case of a flush error, if they
446 * wish to. If WAIT flag is not passed then caller may check only what
447 * request was pushed in some internal queue for later handling.
449 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
450 sector_t *error_sector)
452 struct request_queue *q;
456 if (bdev->bd_disk == NULL)
459 q = bdev_get_queue(bdev);
464 * some block devices may not have their queue correctly set up here
465 * (e.g. loop device without a backing file) and so issuing a flush
466 * here will panic. Ensure there is a request function before issuing
469 if (!q->make_request_fn)
472 bio = bio_alloc(gfp_mask, 0);
475 ret = submit_bio_wait(WRITE_FLUSH, bio);
478 * The driver must store the error location in ->bi_sector, if
479 * it supports it. For non-stacked drivers, this should be
480 * copied from blk_rq_pos(rq).
483 *error_sector = bio->bi_iter.bi_sector;
488 EXPORT_SYMBOL(blkdev_issue_flush);
490 struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
491 int node, int cmd_size)
493 struct blk_flush_queue *fq;
494 int rq_sz = sizeof(struct request);
496 fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node);
501 spin_lock_init(&fq->mq_flush_lock);
502 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
505 fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node);
509 INIT_LIST_HEAD(&fq->flush_queue[0]);
510 INIT_LIST_HEAD(&fq->flush_queue[1]);
511 INIT_LIST_HEAD(&fq->flush_data_in_flight);
521 void blk_free_flush_queue(struct blk_flush_queue *fq)
523 /* bio based request queue hasn't flush queue */