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 * q->flush_queue[q->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);
96 static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
98 unsigned int policy = 0;
100 if (blk_rq_sectors(rq))
101 policy |= REQ_FSEQ_DATA;
103 if (fflags & REQ_FLUSH) {
104 if (rq->cmd_flags & REQ_FLUSH)
105 policy |= REQ_FSEQ_PREFLUSH;
106 if (!(fflags & REQ_FUA) && (rq->cmd_flags & REQ_FUA))
107 policy |= REQ_FSEQ_POSTFLUSH;
112 static unsigned int blk_flush_cur_seq(struct request *rq)
114 return 1 << ffz(rq->flush.seq);
117 static void blk_flush_restore_request(struct request *rq)
120 * After flush data completion, @rq->bio is %NULL but we need to
121 * complete the bio again. @rq->biotail is guaranteed to equal the
122 * original @rq->bio. Restore it.
124 rq->bio = rq->biotail;
126 /* make @rq a normal request */
127 rq->cmd_flags &= ~REQ_FLUSH_SEQ;
128 rq->end_io = rq->flush.saved_end_io;
130 blk_clear_rq_complete(rq);
133 static void mq_flush_data_run(struct work_struct *work)
137 rq = container_of(work, struct request, mq_flush_data);
139 memset(&rq->csd, 0, sizeof(rq->csd));
140 blk_mq_run_request(rq, true, false);
143 static void blk_mq_flush_data_insert(struct request *rq)
145 INIT_WORK(&rq->mq_flush_data, mq_flush_data_run);
146 kblockd_schedule_work(rq->q, &rq->mq_flush_data);
150 * blk_flush_complete_seq - complete flush sequence
151 * @rq: FLUSH/FUA request being sequenced
152 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
153 * @error: whether an error occurred
155 * @rq just completed @seq part of its flush sequence, record the
156 * completion and trigger the next step.
159 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
162 * %true if requests were added to the dispatch queue, %false otherwise.
164 static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
167 struct request_queue *q = rq->q;
168 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
169 bool queued = false, kicked;
171 BUG_ON(rq->flush.seq & seq);
172 rq->flush.seq |= seq;
175 seq = blk_flush_cur_seq(rq);
180 case REQ_FSEQ_PREFLUSH:
181 case REQ_FSEQ_POSTFLUSH:
182 /* queue for flush */
183 if (list_empty(pending))
184 q->flush_pending_since = jiffies;
185 list_move_tail(&rq->flush.list, pending);
189 list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
191 blk_mq_flush_data_insert(rq);
193 list_add(&rq->queuelist, &q->queue_head);
200 * @rq was previously adjusted by blk_flush_issue() for
201 * flush sequencing and may already have gone through the
202 * flush data request completion path. Restore @rq for
203 * normal completion and end it.
205 BUG_ON(!list_empty(&rq->queuelist));
206 list_del_init(&rq->flush.list);
207 blk_flush_restore_request(rq);
209 blk_mq_end_io(rq, error);
211 __blk_end_request_all(rq, error);
218 kicked = blk_kick_flush(q);
219 /* blk_mq_run_flush will run queue */
222 return kicked | queued;
225 static void flush_end_io(struct request *flush_rq, int error)
227 struct request_queue *q = flush_rq->q;
228 struct list_head *running;
230 struct request *rq, *n;
231 unsigned long flags = 0;
234 blk_mq_free_request(flush_rq);
235 spin_lock_irqsave(&q->mq_flush_lock, flags);
237 running = &q->flush_queue[q->flush_running_idx];
238 BUG_ON(q->flush_pending_idx == q->flush_running_idx);
240 /* account completion of the flush request */
241 q->flush_running_idx ^= 1;
244 elv_completed_request(q, flush_rq);
246 /* and push the waiting requests to the next stage */
247 list_for_each_entry_safe(rq, n, running, flush.list) {
248 unsigned int seq = blk_flush_cur_seq(rq);
250 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
251 queued |= blk_flush_complete_seq(rq, seq, error);
255 * Kick the queue to avoid stall for two cases:
256 * 1. Moving a request silently to empty queue_head may stall the
258 * 2. When flush request is running in non-queueable queue, the
259 * queue is hold. Restart the queue after flush request is finished
261 * This function is called from request completion path and calling
262 * directly into request_fn may confuse the driver. Always use
265 if (queued || q->flush_queue_delayed) {
267 blk_run_queue_async(q);
270 * This can be optimized to only run queues with requests
271 * queued if necessary.
273 blk_mq_run_queues(q, true);
275 q->flush_queue_delayed = 0;
277 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
280 static void mq_flush_work(struct work_struct *work)
282 struct request_queue *q;
285 q = container_of(work, struct request_queue, mq_flush_work);
287 /* We don't need set REQ_FLUSH_SEQ, it's for consistency */
288 rq = blk_mq_alloc_request(q, WRITE_FLUSH|REQ_FLUSH_SEQ,
289 __GFP_WAIT|GFP_ATOMIC, true);
290 rq->cmd_type = REQ_TYPE_FS;
291 rq->end_io = flush_end_io;
293 blk_mq_run_request(rq, true, false);
297 * We can't directly use q->flush_rq, because it doesn't have tag and is not in
298 * hctx->rqs[]. so we must allocate a new request, since we can't sleep here,
299 * so offload the work to workqueue.
301 * Note: we assume a flush request finished in any hardware queue will flush
302 * the whole disk cache.
304 static void mq_run_flush(struct request_queue *q)
306 kblockd_schedule_work(q, &q->mq_flush_work);
310 * blk_kick_flush - consider issuing flush request
311 * @q: request_queue being kicked
313 * Flush related states of @q have changed, consider issuing flush request.
314 * Please read the comment at the top of this file for more info.
317 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
320 * %true if flush was issued, %false otherwise.
322 static bool blk_kick_flush(struct request_queue *q)
324 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
325 struct request *first_rq =
326 list_first_entry(pending, struct request, flush.list);
328 /* C1 described at the top of this file */
329 if (q->flush_pending_idx != q->flush_running_idx || list_empty(pending))
333 if (!list_empty(&q->flush_data_in_flight) &&
335 q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
339 * Issue flush and toggle pending_idx. This makes pending_idx
340 * different from running_idx, which means flush is in flight.
342 q->flush_pending_idx ^= 1;
348 blk_rq_init(q, &q->flush_rq);
349 q->flush_rq.cmd_type = REQ_TYPE_FS;
350 q->flush_rq.cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
351 q->flush_rq.rq_disk = first_rq->rq_disk;
352 q->flush_rq.end_io = flush_end_io;
354 list_add_tail(&q->flush_rq.queuelist, &q->queue_head);
358 static void flush_data_end_io(struct request *rq, int error)
360 struct request_queue *q = rq->q;
363 * After populating an empty queue, kick it to avoid stall. Read
364 * the comment in flush_end_io().
366 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
367 blk_run_queue_async(q);
370 static void mq_flush_data_end_io(struct request *rq, int error)
372 struct request_queue *q = rq->q;
373 struct blk_mq_hw_ctx *hctx;
374 struct blk_mq_ctx *ctx;
378 hctx = q->mq_ops->map_queue(q, ctx->cpu);
381 * After populating an empty queue, kick it to avoid stall. Read
382 * the comment in flush_end_io().
384 spin_lock_irqsave(&q->mq_flush_lock, flags);
385 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
386 blk_mq_run_hw_queue(hctx, true);
387 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
391 * blk_insert_flush - insert a new FLUSH/FUA request
392 * @rq: request to insert
394 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
395 * or __blk_mq_run_hw_queue() to dispatch request.
396 * @rq is being submitted. Analyze what needs to be done and put it on the
400 * spin_lock_irq(q->queue_lock) in !mq case
402 void blk_insert_flush(struct request *rq)
404 struct request_queue *q = rq->q;
405 unsigned int fflags = q->flush_flags; /* may change, cache */
406 unsigned int policy = blk_flush_policy(fflags, rq);
409 * @policy now records what operations need to be done. Adjust
410 * REQ_FLUSH and FUA for the driver.
412 rq->cmd_flags &= ~REQ_FLUSH;
413 if (!(fflags & REQ_FUA))
414 rq->cmd_flags &= ~REQ_FUA;
417 * An empty flush handed down from a stacking driver may
418 * translate into nothing if the underlying device does not
419 * advertise a write-back cache. In this case, simply
420 * complete the request.
424 blk_mq_end_io(rq, 0);
426 __blk_end_bidi_request(rq, 0, 0, 0);
430 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
433 * If there's data but flush is not necessary, the request can be
434 * processed directly without going through flush machinery. Queue
435 * for normal execution.
437 if ((policy & REQ_FSEQ_DATA) &&
438 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
440 blk_mq_run_request(rq, false, true);
442 list_add_tail(&rq->queuelist, &q->queue_head);
447 * @rq should go through flush machinery. Mark it part of flush
448 * sequence and submit for further processing.
450 memset(&rq->flush, 0, sizeof(rq->flush));
451 INIT_LIST_HEAD(&rq->flush.list);
452 rq->cmd_flags |= REQ_FLUSH_SEQ;
453 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
455 rq->end_io = mq_flush_data_end_io;
457 spin_lock_irq(&q->mq_flush_lock);
458 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
459 spin_unlock_irq(&q->mq_flush_lock);
462 rq->end_io = flush_data_end_io;
464 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
468 * blk_abort_flushes - @q is being aborted, abort flush requests
469 * @q: request_queue being aborted
471 * To be called from elv_abort_queue(). @q is being aborted. Prepare all
472 * FLUSH/FUA requests for abortion.
475 * spin_lock_irq(q->queue_lock)
477 void blk_abort_flushes(struct request_queue *q)
479 struct request *rq, *n;
483 * Requests in flight for data are already owned by the dispatch
484 * queue or the device driver. Just restore for normal completion.
486 list_for_each_entry_safe(rq, n, &q->flush_data_in_flight, flush.list) {
487 list_del_init(&rq->flush.list);
488 blk_flush_restore_request(rq);
492 * We need to give away requests on flush queues. Restore for
493 * normal completion and put them on the dispatch queue.
495 for (i = 0; i < ARRAY_SIZE(q->flush_queue); i++) {
496 list_for_each_entry_safe(rq, n, &q->flush_queue[i],
498 list_del_init(&rq->flush.list);
499 blk_flush_restore_request(rq);
500 list_add_tail(&rq->queuelist, &q->queue_head);
505 static void bio_end_flush(struct bio *bio, int err)
508 clear_bit(BIO_UPTODATE, &bio->bi_flags);
510 complete(bio->bi_private);
515 * blkdev_issue_flush - queue a flush
516 * @bdev: blockdev to issue flush for
517 * @gfp_mask: memory allocation flags (for bio_alloc)
518 * @error_sector: error sector
521 * Issue a flush for the block device in question. Caller can supply
522 * room for storing the error offset in case of a flush error, if they
523 * wish to. If WAIT flag is not passed then caller may check only what
524 * request was pushed in some internal queue for later handling.
526 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
527 sector_t *error_sector)
529 DECLARE_COMPLETION_ONSTACK(wait);
530 struct request_queue *q;
534 if (bdev->bd_disk == NULL)
537 q = bdev_get_queue(bdev);
542 * some block devices may not have their queue correctly set up here
543 * (e.g. loop device without a backing file) and so issuing a flush
544 * here will panic. Ensure there is a request function before issuing
547 if (!q->make_request_fn)
550 bio = bio_alloc(gfp_mask, 0);
551 bio->bi_end_io = bio_end_flush;
553 bio->bi_private = &wait;
556 submit_bio(WRITE_FLUSH, bio);
557 wait_for_completion_io(&wait);
560 * The driver must store the error location in ->bi_sector, if
561 * it supports it. For non-stacked drivers, this should be
562 * copied from blk_rq_pos(rq).
565 *error_sector = bio->bi_sector;
567 if (!bio_flagged(bio, BIO_UPTODATE))
573 EXPORT_SYMBOL(blkdev_issue_flush);
575 void blk_mq_init_flush(struct request_queue *q)
577 spin_lock_init(&q->mq_flush_lock);
578 INIT_WORK(&q->mq_flush_work, mq_flush_work);