]>
git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/blob - block/blk-flush.c
1 // SPDX-License-Identifier: GPL-2.0
3 * Functions to sequence PREFLUSH and FUA writes.
5 * Copyright (C) 2011 Max Planck Institute for Gravitational Physics
6 * Copyright (C) 2011 Tejun Heo <tj@kernel.org>
8 * REQ_{PREFLUSH|FUA} requests are decomposed to sequences consisted of three
9 * optional steps - PREFLUSH, DATA and POSTFLUSH - according to the request
10 * properties and hardware capability.
12 * If a request doesn't have data, only REQ_PREFLUSH makes sense, which
13 * indicates a simple flush request. If there is data, REQ_PREFLUSH indicates
14 * that the device cache should be flushed before the data is executed, and
15 * REQ_FUA means that the data must be on non-volatile media on request
18 * If the device doesn't have writeback cache, PREFLUSH and FUA don't make any
19 * difference. The requests are either completed immediately if there's no data
20 * or executed as normal requests otherwise.
22 * If the device has writeback cache and supports FUA, REQ_PREFLUSH is
23 * translated to PREFLUSH but REQ_FUA is passed down directly with DATA.
25 * If the device has writeback cache and doesn't support FUA, REQ_PREFLUSH
26 * is translated to PREFLUSH and REQ_FUA to POSTFLUSH.
28 * The actual execution of flush is double buffered. Whenever a request
29 * needs to execute PRE or POSTFLUSH, it queues at
30 * fq->flush_queue[fq->flush_pending_idx]. Once certain criteria are met, a
31 * REQ_OP_FLUSH is issued and the pending_idx is toggled. When the flush
32 * completes, all the requests which were pending are proceeded to the next
33 * step. This allows arbitrary merging of different types of PREFLUSH/FUA
36 * Currently, the following conditions are used to determine when to issue
39 * C1. At any given time, only one flush shall be in progress. This makes
40 * double buffering sufficient.
42 * C2. Flush is deferred if any request is executing DATA of its sequence.
43 * This avoids issuing separate POSTFLUSHes for requests which shared
46 * C3. The second condition is ignored if there is a request which has
47 * waited longer than FLUSH_PENDING_TIMEOUT. This is to avoid
48 * starvation in the unlikely case where there are continuous stream of
49 * FUA (without PREFLUSH) requests.
51 * For devices which support FUA, it isn't clear whether C2 (and thus C3)
54 * Note that a sequenced PREFLUSH/FUA request with DATA is completed twice.
55 * Once while executing DATA and again after the whole sequence is
56 * complete. The first completion updates the contained bio but doesn't
57 * finish it so that the bio submitter is notified only after the whole
58 * sequence is complete. This is implemented by testing RQF_FLUSH_SEQ in
61 * The above peculiarity requires that each PREFLUSH/FUA request has only one
62 * bio attached to it, which is guaranteed as they aren't allowed to be
63 * merged in the usual way.
66 #include <linux/kernel.h>
67 #include <linux/module.h>
68 #include <linux/bio.h>
69 #include <linux/blkdev.h>
70 #include <linux/gfp.h>
71 #include <linux/blk-mq.h>
72 #include <linux/lockdep.h>
76 #include "blk-mq-tag.h"
77 #include "blk-mq-sched.h"
79 /* PREFLUSH/FUA sequences */
81 REQ_FSEQ_PREFLUSH
= (1 << 0), /* pre-flushing in progress */
82 REQ_FSEQ_DATA
= (1 << 1), /* data write in progress */
83 REQ_FSEQ_POSTFLUSH
= (1 << 2), /* post-flushing in progress */
84 REQ_FSEQ_DONE
= (1 << 3),
86 REQ_FSEQ_ACTIONS
= REQ_FSEQ_PREFLUSH
| REQ_FSEQ_DATA
|
90 * If flush has been pending longer than the following timeout,
91 * it's issued even if flush_data requests are still in flight.
93 FLUSH_PENDING_TIMEOUT
= 5 * HZ
,
96 static void blk_kick_flush(struct request_queue
*q
,
97 struct blk_flush_queue
*fq
, unsigned int flags
);
99 static unsigned int blk_flush_policy(unsigned long fflags
, struct request
*rq
)
101 unsigned int policy
= 0;
103 if (blk_rq_sectors(rq
))
104 policy
|= REQ_FSEQ_DATA
;
106 if (fflags
& (1UL << QUEUE_FLAG_WC
)) {
107 if (rq
->cmd_flags
& REQ_PREFLUSH
)
108 policy
|= REQ_FSEQ_PREFLUSH
;
109 if (!(fflags
& (1UL << QUEUE_FLAG_FUA
)) &&
110 (rq
->cmd_flags
& REQ_FUA
))
111 policy
|= REQ_FSEQ_POSTFLUSH
;
116 static unsigned int blk_flush_cur_seq(struct request
*rq
)
118 return 1 << ffz(rq
->flush
.seq
);
121 static void blk_flush_restore_request(struct request
*rq
)
124 * After flush data completion, @rq->bio is %NULL but we need to
125 * complete the bio again. @rq->biotail is guaranteed to equal the
126 * original @rq->bio. Restore it.
128 rq
->bio
= rq
->biotail
;
130 /* make @rq a normal request */
131 rq
->rq_flags
&= ~RQF_FLUSH_SEQ
;
132 rq
->end_io
= rq
->flush
.saved_end_io
;
135 static void blk_flush_queue_rq(struct request
*rq
, bool add_front
)
137 blk_mq_add_to_requeue_list(rq
, add_front
, true);
140 static void blk_account_io_flush(struct request
*rq
)
142 struct hd_struct
*part
= &rq
->rq_disk
->part0
;
145 part_stat_inc(part
, ios
[STAT_FLUSH
]);
146 part_stat_add(part
, nsecs
[STAT_FLUSH
],
147 ktime_get_ns() - rq
->start_time_ns
);
152 * blk_flush_complete_seq - complete flush sequence
153 * @rq: PREFLUSH/FUA request being sequenced
155 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
156 * @error: whether an error occurred
158 * @rq just completed @seq part of its flush sequence, record the
159 * completion and trigger the next step.
162 * spin_lock_irq(fq->mq_flush_lock)
164 static void blk_flush_complete_seq(struct request
*rq
,
165 struct blk_flush_queue
*fq
,
166 unsigned int seq
, blk_status_t error
)
168 struct request_queue
*q
= rq
->q
;
169 struct list_head
*pending
= &fq
->flush_queue
[fq
->flush_pending_idx
];
170 unsigned int cmd_flags
;
172 BUG_ON(rq
->flush
.seq
& seq
);
173 rq
->flush
.seq
|= seq
;
174 cmd_flags
= rq
->cmd_flags
;
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 blk_flush_queue_rq(rq
, true);
197 * @rq was previously adjusted by blk_insert_flush() 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
);
205 blk_mq_end_request(rq
, error
);
212 blk_kick_flush(q
, fq
, cmd_flags
);
215 static void flush_end_io(struct request
*flush_rq
, blk_status_t error
)
217 struct request_queue
*q
= flush_rq
->q
;
218 struct list_head
*running
;
219 struct request
*rq
, *n
;
220 unsigned long flags
= 0;
221 struct blk_flush_queue
*fq
= blk_get_flush_queue(q
, flush_rq
->mq_ctx
);
223 blk_account_io_flush(flush_rq
);
225 /* release the tag's ownership to the req cloned from */
226 spin_lock_irqsave(&fq
->mq_flush_lock
, flags
);
228 if (!refcount_dec_and_test(&flush_rq
->ref
)) {
229 fq
->rq_status
= error
;
230 spin_unlock_irqrestore(&fq
->mq_flush_lock
, flags
);
234 if (fq
->rq_status
!= BLK_STS_OK
)
235 error
= fq
->rq_status
;
238 flush_rq
->tag
= BLK_MQ_NO_TAG
;
240 blk_mq_put_driver_tag(flush_rq
);
241 flush_rq
->internal_tag
= BLK_MQ_NO_TAG
;
244 running
= &fq
->flush_queue
[fq
->flush_running_idx
];
245 BUG_ON(fq
->flush_pending_idx
== fq
->flush_running_idx
);
247 /* account completion of the flush request */
248 fq
->flush_running_idx
^= 1;
250 /* and push the waiting requests to the next stage */
251 list_for_each_entry_safe(rq
, n
, running
, flush
.list
) {
252 unsigned int seq
= blk_flush_cur_seq(rq
);
254 BUG_ON(seq
!= REQ_FSEQ_PREFLUSH
&& seq
!= REQ_FSEQ_POSTFLUSH
);
255 blk_flush_complete_seq(rq
, fq
, seq
, error
);
258 spin_unlock_irqrestore(&fq
->mq_flush_lock
, flags
);
262 * blk_kick_flush - consider issuing flush request
263 * @q: request_queue being kicked
265 * @flags: cmd_flags of the original request
267 * Flush related states of @q have changed, consider issuing flush request.
268 * Please read the comment at the top of this file for more info.
271 * spin_lock_irq(fq->mq_flush_lock)
274 static void blk_kick_flush(struct request_queue
*q
, struct blk_flush_queue
*fq
,
277 struct list_head
*pending
= &fq
->flush_queue
[fq
->flush_pending_idx
];
278 struct request
*first_rq
=
279 list_first_entry(pending
, struct request
, flush
.list
);
280 struct request
*flush_rq
= fq
->flush_rq
;
282 /* C1 described at the top of this file */
283 if (fq
->flush_pending_idx
!= fq
->flush_running_idx
|| list_empty(pending
))
287 if (!list_empty(&fq
->flush_data_in_flight
) &&
289 fq
->flush_pending_since
+ FLUSH_PENDING_TIMEOUT
))
293 * Issue flush and toggle pending_idx. This makes pending_idx
294 * different from running_idx, which means flush is in flight.
296 fq
->flush_pending_idx
^= 1;
298 blk_rq_init(q
, flush_rq
);
301 * In case of none scheduler, borrow tag from the first request
302 * since they can't be in flight at the same time. And acquire
303 * the tag's ownership for flush req.
305 * In case of IO scheduler, flush rq need to borrow scheduler tag
306 * just for cheating put/get driver tag.
308 flush_rq
->mq_ctx
= first_rq
->mq_ctx
;
309 flush_rq
->mq_hctx
= first_rq
->mq_hctx
;
312 flush_rq
->tag
= first_rq
->tag
;
314 flush_rq
->internal_tag
= first_rq
->internal_tag
;
316 flush_rq
->cmd_flags
= REQ_OP_FLUSH
| REQ_PREFLUSH
;
317 flush_rq
->cmd_flags
|= (flags
& REQ_DRV
) | (flags
& REQ_FAILFAST_MASK
);
318 flush_rq
->rq_flags
|= RQF_FLUSH_SEQ
;
319 flush_rq
->rq_disk
= first_rq
->rq_disk
;
320 flush_rq
->end_io
= flush_end_io
;
322 blk_flush_queue_rq(flush_rq
, false);
325 static void mq_flush_data_end_io(struct request
*rq
, blk_status_t error
)
327 struct request_queue
*q
= rq
->q
;
328 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
329 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
331 struct blk_flush_queue
*fq
= blk_get_flush_queue(q
, ctx
);
334 WARN_ON(rq
->tag
< 0);
335 blk_mq_put_driver_tag(rq
);
339 * After populating an empty queue, kick it to avoid stall. Read
340 * the comment in flush_end_io().
342 spin_lock_irqsave(&fq
->mq_flush_lock
, flags
);
343 blk_flush_complete_seq(rq
, fq
, REQ_FSEQ_DATA
, error
);
344 spin_unlock_irqrestore(&fq
->mq_flush_lock
, flags
);
346 blk_mq_sched_restart(hctx
);
350 * blk_insert_flush - insert a new PREFLUSH/FUA request
351 * @rq: request to insert
353 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
354 * or __blk_mq_run_hw_queue() to dispatch request.
355 * @rq is being submitted. Analyze what needs to be done and put it on the
358 void blk_insert_flush(struct request
*rq
)
360 struct request_queue
*q
= rq
->q
;
361 unsigned long fflags
= q
->queue_flags
; /* may change, cache */
362 unsigned int policy
= blk_flush_policy(fflags
, rq
);
363 struct blk_flush_queue
*fq
= blk_get_flush_queue(q
, rq
->mq_ctx
);
366 * @policy now records what operations need to be done. Adjust
367 * REQ_PREFLUSH and FUA for the driver.
369 rq
->cmd_flags
&= ~REQ_PREFLUSH
;
370 if (!(fflags
& (1UL << QUEUE_FLAG_FUA
)))
371 rq
->cmd_flags
&= ~REQ_FUA
;
374 * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
375 * of those flags, we have to set REQ_SYNC to avoid skewing
376 * the request accounting.
378 rq
->cmd_flags
|= REQ_SYNC
;
381 * An empty flush handed down from a stacking driver may
382 * translate into nothing if the underlying device does not
383 * advertise a write-back cache. In this case, simply
384 * complete the request.
387 blk_mq_end_request(rq
, 0);
391 BUG_ON(rq
->bio
!= rq
->biotail
); /*assumes zero or single bio rq */
394 * If there's data but flush is not necessary, the request can be
395 * processed directly without going through flush machinery. Queue
396 * for normal execution.
398 if ((policy
& REQ_FSEQ_DATA
) &&
399 !(policy
& (REQ_FSEQ_PREFLUSH
| REQ_FSEQ_POSTFLUSH
))) {
400 blk_mq_request_bypass_insert(rq
, false, false);
405 * @rq should go through flush machinery. Mark it part of flush
406 * sequence and submit for further processing.
408 memset(&rq
->flush
, 0, sizeof(rq
->flush
));
409 INIT_LIST_HEAD(&rq
->flush
.list
);
410 rq
->rq_flags
|= RQF_FLUSH_SEQ
;
411 rq
->flush
.saved_end_io
= rq
->end_io
; /* Usually NULL */
413 rq
->end_io
= mq_flush_data_end_io
;
415 spin_lock_irq(&fq
->mq_flush_lock
);
416 blk_flush_complete_seq(rq
, fq
, REQ_FSEQ_ACTIONS
& ~policy
, 0);
417 spin_unlock_irq(&fq
->mq_flush_lock
);
421 * blkdev_issue_flush - queue a flush
422 * @bdev: blockdev to issue flush for
423 * @gfp_mask: memory allocation flags (for bio_alloc)
426 * Issue a flush for the block device in question.
428 int blkdev_issue_flush(struct block_device
*bdev
, gfp_t gfp_mask
)
433 bio
= bio_alloc(gfp_mask
, 0);
434 bio_set_dev(bio
, bdev
);
435 bio
->bi_opf
= REQ_OP_WRITE
| REQ_PREFLUSH
;
437 ret
= submit_bio_wait(bio
);
441 EXPORT_SYMBOL(blkdev_issue_flush
);
443 struct blk_flush_queue
*blk_alloc_flush_queue(int node
, int cmd_size
,
446 struct blk_flush_queue
*fq
;
447 int rq_sz
= sizeof(struct request
);
449 fq
= kzalloc_node(sizeof(*fq
), flags
, node
);
453 spin_lock_init(&fq
->mq_flush_lock
);
455 rq_sz
= round_up(rq_sz
+ cmd_size
, cache_line_size());
456 fq
->flush_rq
= kzalloc_node(rq_sz
, flags
, node
);
460 INIT_LIST_HEAD(&fq
->flush_queue
[0]);
461 INIT_LIST_HEAD(&fq
->flush_queue
[1]);
462 INIT_LIST_HEAD(&fq
->flush_data_in_flight
);
464 lockdep_register_key(&fq
->key
);
465 lockdep_set_class(&fq
->mq_flush_lock
, &fq
->key
);
475 void blk_free_flush_queue(struct blk_flush_queue
*fq
)
477 /* bio based request queue hasn't flush queue */
481 lockdep_unregister_key(&fq
->key
);