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 bool blk_flush_queue_rq(struct request
*rq
, bool add_front
)
136 struct request_queue
*q
= rq
->q
;
138 blk_mq_add_to_requeue_list(rq
, add_front
);
139 blk_mq_kick_requeue_list(q
);
143 list_add(&rq
->queuelist
, &rq
->q
->queue_head
);
145 list_add_tail(&rq
->queuelist
, &rq
->q
->queue_head
);
151 * blk_flush_complete_seq - complete flush sequence
152 * @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 q->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
, unsigned int seq
,
168 struct request_queue
*q
= rq
->q
;
169 struct list_head
*pending
= &q
->flush_queue
[q
->flush_pending_idx
];
170 bool queued
= false, kicked
;
172 BUG_ON(rq
->flush
.seq
& seq
);
173 rq
->flush
.seq
|= seq
;
176 seq
= blk_flush_cur_seq(rq
);
181 case REQ_FSEQ_PREFLUSH
:
182 case REQ_FSEQ_POSTFLUSH
:
183 /* queue for flush */
184 if (list_empty(pending
))
185 q
->flush_pending_since
= jiffies
;
186 list_move_tail(&rq
->flush
.list
, pending
);
190 list_move_tail(&rq
->flush
.list
, &q
->flush_data_in_flight
);
191 queued
= blk_flush_queue_rq(rq
, true);
196 * @rq was previously adjusted by blk_flush_issue() for
197 * flush sequencing and may already have gone through the
198 * flush data request completion path. Restore @rq for
199 * normal completion and end it.
201 BUG_ON(!list_empty(&rq
->queuelist
));
202 list_del_init(&rq
->flush
.list
);
203 blk_flush_restore_request(rq
);
205 blk_mq_end_io(rq
, error
);
207 __blk_end_request_all(rq
, error
);
214 kicked
= blk_kick_flush(q
);
215 return kicked
| queued
;
218 static void flush_end_io(struct request
*flush_rq
, int error
)
220 struct request_queue
*q
= flush_rq
->q
;
221 struct list_head
*running
;
223 struct request
*rq
, *n
;
224 unsigned long flags
= 0;
227 spin_lock_irqsave(&q
->mq_flush_lock
, flags
);
229 running
= &q
->flush_queue
[q
->flush_running_idx
];
230 BUG_ON(q
->flush_pending_idx
== q
->flush_running_idx
);
232 /* account completion of the flush request */
233 q
->flush_running_idx
^= 1;
236 elv_completed_request(q
, flush_rq
);
238 /* and push the waiting requests to the next stage */
239 list_for_each_entry_safe(rq
, n
, running
, flush
.list
) {
240 unsigned int seq
= blk_flush_cur_seq(rq
);
242 BUG_ON(seq
!= REQ_FSEQ_PREFLUSH
&& seq
!= REQ_FSEQ_POSTFLUSH
);
243 queued
|= blk_flush_complete_seq(rq
, seq
, error
);
247 * Kick the queue to avoid stall for two cases:
248 * 1. Moving a request silently to empty queue_head may stall the
250 * 2. When flush request is running in non-queueable queue, the
251 * queue is hold. Restart the queue after flush request is finished
253 * This function is called from request completion path and calling
254 * directly into request_fn may confuse the driver. Always use
257 if (queued
|| q
->flush_queue_delayed
) {
259 blk_run_queue_async(q
);
261 q
->flush_queue_delayed
= 0;
263 spin_unlock_irqrestore(&q
->mq_flush_lock
, flags
);
267 * blk_kick_flush - consider issuing flush request
268 * @q: request_queue being kicked
270 * Flush related states of @q have changed, consider issuing flush request.
271 * Please read the comment at the top of this file for more info.
274 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
277 * %true if flush was issued, %false otherwise.
279 static bool blk_kick_flush(struct request_queue
*q
)
281 struct list_head
*pending
= &q
->flush_queue
[q
->flush_pending_idx
];
282 struct request
*first_rq
=
283 list_first_entry(pending
, struct request
, flush
.list
);
285 /* C1 described at the top of this file */
286 if (q
->flush_pending_idx
!= q
->flush_running_idx
|| list_empty(pending
))
290 if (!list_empty(&q
->flush_data_in_flight
) &&
292 q
->flush_pending_since
+ FLUSH_PENDING_TIMEOUT
))
296 * Issue flush and toggle pending_idx. This makes pending_idx
297 * different from running_idx, which means flush is in flight.
299 q
->flush_pending_idx
^= 1;
301 blk_rq_init(q
, q
->flush_rq
);
303 blk_mq_clone_flush_request(q
->flush_rq
, first_rq
);
305 q
->flush_rq
->cmd_type
= REQ_TYPE_FS
;
306 q
->flush_rq
->cmd_flags
= WRITE_FLUSH
| REQ_FLUSH_SEQ
;
307 q
->flush_rq
->rq_disk
= first_rq
->rq_disk
;
308 q
->flush_rq
->end_io
= flush_end_io
;
310 return blk_flush_queue_rq(q
->flush_rq
, false);
313 static void flush_data_end_io(struct request
*rq
, int error
)
315 struct request_queue
*q
= rq
->q
;
318 * After populating an empty queue, kick it to avoid stall. Read
319 * the comment in flush_end_io().
321 if (blk_flush_complete_seq(rq
, REQ_FSEQ_DATA
, error
))
322 blk_run_queue_async(q
);
325 static void mq_flush_data_end_io(struct request
*rq
, int error
)
327 struct request_queue
*q
= rq
->q
;
328 struct blk_mq_hw_ctx
*hctx
;
329 struct blk_mq_ctx
*ctx
;
333 hctx
= q
->mq_ops
->map_queue(q
, ctx
->cpu
);
336 * After populating an empty queue, kick it to avoid stall. Read
337 * the comment in flush_end_io().
339 spin_lock_irqsave(&q
->mq_flush_lock
, flags
);
340 if (blk_flush_complete_seq(rq
, REQ_FSEQ_DATA
, error
))
341 blk_mq_run_hw_queue(hctx
, true);
342 spin_unlock_irqrestore(&q
->mq_flush_lock
, flags
);
346 * blk_insert_flush - insert a new FLUSH/FUA request
347 * @rq: request to insert
349 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
350 * or __blk_mq_run_hw_queue() to dispatch request.
351 * @rq is being submitted. Analyze what needs to be done and put it on the
355 * spin_lock_irq(q->queue_lock) in !mq case
357 void blk_insert_flush(struct request
*rq
)
359 struct request_queue
*q
= rq
->q
;
360 unsigned int fflags
= q
->flush_flags
; /* may change, cache */
361 unsigned int policy
= blk_flush_policy(fflags
, rq
);
364 * @policy now records what operations need to be done. Adjust
365 * REQ_FLUSH and FUA for the driver.
367 rq
->cmd_flags
&= ~REQ_FLUSH
;
368 if (!(fflags
& REQ_FUA
))
369 rq
->cmd_flags
&= ~REQ_FUA
;
372 * An empty flush handed down from a stacking driver may
373 * translate into nothing if the underlying device does not
374 * advertise a write-back cache. In this case, simply
375 * complete the request.
379 blk_mq_end_io(rq
, 0);
381 __blk_end_bidi_request(rq
, 0, 0, 0);
385 BUG_ON(rq
->bio
!= rq
->biotail
); /*assumes zero or single bio rq */
388 * If there's data but flush is not necessary, the request can be
389 * processed directly without going through flush machinery. Queue
390 * for normal execution.
392 if ((policy
& REQ_FSEQ_DATA
) &&
393 !(policy
& (REQ_FSEQ_PREFLUSH
| REQ_FSEQ_POSTFLUSH
))) {
395 blk_mq_insert_request(rq
, false, false, true);
397 list_add_tail(&rq
->queuelist
, &q
->queue_head
);
402 * @rq should go through flush machinery. Mark it part of flush
403 * sequence and submit for further processing.
405 memset(&rq
->flush
, 0, sizeof(rq
->flush
));
406 INIT_LIST_HEAD(&rq
->flush
.list
);
407 rq
->cmd_flags
|= REQ_FLUSH_SEQ
;
408 rq
->flush
.saved_end_io
= rq
->end_io
; /* Usually NULL */
410 rq
->end_io
= mq_flush_data_end_io
;
412 spin_lock_irq(&q
->mq_flush_lock
);
413 blk_flush_complete_seq(rq
, REQ_FSEQ_ACTIONS
& ~policy
, 0);
414 spin_unlock_irq(&q
->mq_flush_lock
);
417 rq
->end_io
= flush_data_end_io
;
419 blk_flush_complete_seq(rq
, REQ_FSEQ_ACTIONS
& ~policy
, 0);
423 * blk_abort_flushes - @q is being aborted, abort flush requests
424 * @q: request_queue being aborted
426 * To be called from elv_abort_queue(). @q is being aborted. Prepare all
427 * FLUSH/FUA requests for abortion.
430 * spin_lock_irq(q->queue_lock)
432 void blk_abort_flushes(struct request_queue
*q
)
434 struct request
*rq
, *n
;
438 * Requests in flight for data are already owned by the dispatch
439 * queue or the device driver. Just restore for normal completion.
441 list_for_each_entry_safe(rq
, n
, &q
->flush_data_in_flight
, flush
.list
) {
442 list_del_init(&rq
->flush
.list
);
443 blk_flush_restore_request(rq
);
447 * We need to give away requests on flush queues. Restore for
448 * normal completion and put them on the dispatch queue.
450 for (i
= 0; i
< ARRAY_SIZE(q
->flush_queue
); i
++) {
451 list_for_each_entry_safe(rq
, n
, &q
->flush_queue
[i
],
453 list_del_init(&rq
->flush
.list
);
454 blk_flush_restore_request(rq
);
455 list_add_tail(&rq
->queuelist
, &q
->queue_head
);
461 * blkdev_issue_flush - queue a flush
462 * @bdev: blockdev to issue flush for
463 * @gfp_mask: memory allocation flags (for bio_alloc)
464 * @error_sector: error sector
467 * Issue a flush for the block device in question. Caller can supply
468 * room for storing the error offset in case of a flush error, if they
469 * wish to. If WAIT flag is not passed then caller may check only what
470 * request was pushed in some internal queue for later handling.
472 int blkdev_issue_flush(struct block_device
*bdev
, gfp_t gfp_mask
,
473 sector_t
*error_sector
)
475 struct request_queue
*q
;
479 if (bdev
->bd_disk
== NULL
)
482 q
= bdev_get_queue(bdev
);
487 * some block devices may not have their queue correctly set up here
488 * (e.g. loop device without a backing file) and so issuing a flush
489 * here will panic. Ensure there is a request function before issuing
492 if (!q
->make_request_fn
)
495 bio
= bio_alloc(gfp_mask
, 0);
498 ret
= submit_bio_wait(WRITE_FLUSH
, bio
);
501 * The driver must store the error location in ->bi_sector, if
502 * it supports it. For non-stacked drivers, this should be
503 * copied from blk_rq_pos(rq).
506 *error_sector
= bio
->bi_iter
.bi_sector
;
511 EXPORT_SYMBOL(blkdev_issue_flush
);
513 void blk_mq_init_flush(struct request_queue
*q
)
515 spin_lock_init(&q
->mq_flush_lock
);
This page took 0.042928 seconds and 5 git commands to generate.