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>
75 /* FLUSH/FUA sequences */
77 REQ_FSEQ_PREFLUSH = (1 << 0), /* pre-flushing in progress */
78 REQ_FSEQ_DATA = (1 << 1), /* data write in progress */
79 REQ_FSEQ_POSTFLUSH = (1 << 2), /* post-flushing in progress */
80 REQ_FSEQ_DONE = (1 << 3),
82 REQ_FSEQ_ACTIONS = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
86 * If flush has been pending longer than the following timeout,
87 * it's issued even if flush_data requests are still in flight.
89 FLUSH_PENDING_TIMEOUT = 5 * HZ,
92 static bool blk_kick_flush(struct request_queue *q);
94 static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
96 unsigned int policy = 0;
98 if (blk_rq_sectors(rq))
99 policy |= REQ_FSEQ_DATA;
101 if (fflags & REQ_FLUSH) {
102 if (rq->cmd_flags & REQ_FLUSH)
103 policy |= REQ_FSEQ_PREFLUSH;
104 if (!(fflags & REQ_FUA) && (rq->cmd_flags & REQ_FUA))
105 policy |= REQ_FSEQ_POSTFLUSH;
110 static unsigned int blk_flush_cur_seq(struct request *rq)
112 return 1 << ffz(rq->flush.seq);
115 static void blk_flush_restore_request(struct request *rq)
118 * After flush data completion, @rq->bio is %NULL but we need to
119 * complete the bio again. @rq->biotail is guaranteed to equal the
120 * original @rq->bio. Restore it.
122 rq->bio = rq->biotail;
124 /* make @rq a normal request */
125 rq->cmd_flags &= ~REQ_FLUSH_SEQ;
126 rq->end_io = rq->flush.saved_end_io;
130 * blk_flush_complete_seq - complete flush sequence
131 * @rq: FLUSH/FUA request being sequenced
132 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
133 * @error: whether an error occurred
135 * @rq just completed @seq part of its flush sequence, record the
136 * completion and trigger the next step.
139 * spin_lock_irq(q->queue_lock)
142 * %true if requests were added to the dispatch queue, %false otherwise.
144 static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
147 struct request_queue *q = rq->q;
148 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
151 BUG_ON(rq->flush.seq & seq);
152 rq->flush.seq |= seq;
155 seq = blk_flush_cur_seq(rq);
160 case REQ_FSEQ_PREFLUSH:
161 case REQ_FSEQ_POSTFLUSH:
162 /* queue for flush */
163 if (list_empty(pending))
164 q->flush_pending_since = jiffies;
165 list_move_tail(&rq->flush.list, pending);
169 list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
170 list_add(&rq->queuelist, &q->queue_head);
176 * @rq was previously adjusted by blk_flush_issue() for
177 * flush sequencing and may already have gone through the
178 * flush data request completion path. Restore @rq for
179 * normal completion and end it.
181 BUG_ON(!list_empty(&rq->queuelist));
182 list_del_init(&rq->flush.list);
183 blk_flush_restore_request(rq);
184 __blk_end_request_all(rq, error);
191 return blk_kick_flush(q) | queued;
194 static void flush_end_io(struct request *flush_rq, int error)
196 struct request_queue *q = flush_rq->q;
197 struct list_head *running = &q->flush_queue[q->flush_running_idx];
199 struct request *rq, *n;
201 BUG_ON(q->flush_pending_idx == q->flush_running_idx);
203 /* account completion of the flush request */
204 q->flush_running_idx ^= 1;
205 elv_completed_request(q, flush_rq);
207 /* and push the waiting requests to the next stage */
208 list_for_each_entry_safe(rq, n, running, flush.list) {
209 unsigned int seq = blk_flush_cur_seq(rq);
211 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
212 queued |= blk_flush_complete_seq(rq, seq, error);
216 * Kick the queue to avoid stall for two cases:
217 * 1. Moving a request silently to empty queue_head may stall the
219 * 2. When flush request is running in non-queueable queue, the
220 * queue is hold. Restart the queue after flush request is finished
222 * This function is called from request completion path and calling
223 * directly into request_fn may confuse the driver. Always use
226 if (queued || q->flush_queue_delayed)
227 blk_run_queue_async(q);
228 q->flush_queue_delayed = 0;
232 * blk_kick_flush - consider issuing flush request
233 * @q: request_queue being kicked
235 * Flush related states of @q have changed, consider issuing flush request.
236 * Please read the comment at the top of this file for more info.
239 * spin_lock_irq(q->queue_lock)
242 * %true if flush was issued, %false otherwise.
244 static bool blk_kick_flush(struct request_queue *q)
246 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
247 struct request *first_rq =
248 list_first_entry(pending, struct request, flush.list);
250 /* C1 described at the top of this file */
251 if (q->flush_pending_idx != q->flush_running_idx || list_empty(pending))
255 if (!list_empty(&q->flush_data_in_flight) &&
257 q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
261 * Issue flush and toggle pending_idx. This makes pending_idx
262 * different from running_idx, which means flush is in flight.
264 blk_rq_init(q, &q->flush_rq);
265 q->flush_rq.cmd_type = REQ_TYPE_FS;
266 q->flush_rq.cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
267 q->flush_rq.rq_disk = first_rq->rq_disk;
268 q->flush_rq.end_io = flush_end_io;
270 q->flush_pending_idx ^= 1;
271 list_add_tail(&q->flush_rq.queuelist, &q->queue_head);
275 static void flush_data_end_io(struct request *rq, int error)
277 struct request_queue *q = rq->q;
280 * After populating an empty queue, kick it to avoid stall. Read
281 * the comment in flush_end_io().
283 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
284 blk_run_queue_async(q);
288 * blk_insert_flush - insert a new FLUSH/FUA request
289 * @rq: request to insert
291 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
292 * @rq is being submitted. Analyze what needs to be done and put it on the
296 * spin_lock_irq(q->queue_lock)
298 void blk_insert_flush(struct request *rq)
300 struct request_queue *q = rq->q;
301 unsigned int fflags = q->flush_flags; /* may change, cache */
302 unsigned int policy = blk_flush_policy(fflags, rq);
305 * @policy now records what operations need to be done. Adjust
306 * REQ_FLUSH and FUA for the driver.
308 rq->cmd_flags &= ~REQ_FLUSH;
309 if (!(fflags & REQ_FUA))
310 rq->cmd_flags &= ~REQ_FUA;
313 * An empty flush handed down from a stacking driver may
314 * translate into nothing if the underlying device does not
315 * advertise a write-back cache. In this case, simply
316 * complete the request.
319 __blk_end_bidi_request(rq, 0, 0, 0);
323 BUG_ON(!rq->bio || rq->bio != rq->biotail);
326 * If there's data but flush is not necessary, the request can be
327 * processed directly without going through flush machinery. Queue
328 * for normal execution.
330 if ((policy & REQ_FSEQ_DATA) &&
331 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
332 list_add_tail(&rq->queuelist, &q->queue_head);
333 blk_run_queue_async(q);
338 * @rq should go through flush machinery. Mark it part of flush
339 * sequence and submit for further processing.
341 memset(&rq->flush, 0, sizeof(rq->flush));
342 INIT_LIST_HEAD(&rq->flush.list);
343 rq->cmd_flags |= REQ_FLUSH_SEQ;
344 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
345 rq->end_io = flush_data_end_io;
347 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
351 * blk_abort_flushes - @q is being aborted, abort flush requests
352 * @q: request_queue being aborted
354 * To be called from elv_abort_queue(). @q is being aborted. Prepare all
355 * FLUSH/FUA requests for abortion.
358 * spin_lock_irq(q->queue_lock)
360 void blk_abort_flushes(struct request_queue *q)
362 struct request *rq, *n;
366 * Requests in flight for data are already owned by the dispatch
367 * queue or the device driver. Just restore for normal completion.
369 list_for_each_entry_safe(rq, n, &q->flush_data_in_flight, flush.list) {
370 list_del_init(&rq->flush.list);
371 blk_flush_restore_request(rq);
375 * We need to give away requests on flush queues. Restore for
376 * normal completion and put them on the dispatch queue.
378 for (i = 0; i < ARRAY_SIZE(q->flush_queue); i++) {
379 list_for_each_entry_safe(rq, n, &q->flush_queue[i],
381 list_del_init(&rq->flush.list);
382 blk_flush_restore_request(rq);
383 list_add_tail(&rq->queuelist, &q->queue_head);
388 static void bio_end_flush(struct bio *bio, int err)
391 clear_bit(BIO_UPTODATE, &bio->bi_flags);
393 complete(bio->bi_private);
398 * blkdev_issue_flush - queue a flush
399 * @bdev: blockdev to issue flush for
400 * @gfp_mask: memory allocation flags (for bio_alloc)
401 * @error_sector: error sector
404 * Issue a flush for the block device in question. Caller can supply
405 * room for storing the error offset in case of a flush error, if they
406 * wish to. If WAIT flag is not passed then caller may check only what
407 * request was pushed in some internal queue for later handling.
409 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
410 sector_t *error_sector)
412 DECLARE_COMPLETION_ONSTACK(wait);
413 struct request_queue *q;
417 if (bdev->bd_disk == NULL)
420 q = bdev_get_queue(bdev);
425 * some block devices may not have their queue correctly set up here
426 * (e.g. loop device without a backing file) and so issuing a flush
427 * here will panic. Ensure there is a request function before issuing
430 if (!q->make_request_fn)
433 bio = bio_alloc(gfp_mask, 0);
434 bio->bi_end_io = bio_end_flush;
436 bio->bi_private = &wait;
439 submit_bio(WRITE_FLUSH, bio);
440 wait_for_completion(&wait);
443 * The driver must store the error location in ->bi_sector, if
444 * it supports it. For non-stacked drivers, this should be
445 * copied from blk_rq_pos(rq).
448 *error_sector = bio->bi_sector;
450 if (!bio_flagged(bio, BIO_UPTODATE))
456 EXPORT_SYMBOL(blkdev_issue_flush);