blob: 6e1543c10493399fe23a2d60194e90e0d2ec828c [file] [log] [blame]
Christoph Hellwig8c165672019-04-30 14:42:39 -04001// SPDX-License-Identifier: GPL-2.0
Jens Axboe86db1e22008-01-29 14:53:40 +01002/*
Omar Sandoval3140c3c2017-08-24 11:09:25 -07003 * Functions to sequence PREFLUSH and FUA writes.
Tejun Heoae1b1532011-01-25 12:43:54 +01004 *
5 * Copyright (C) 2011 Max Planck Institute for Gravitational Physics
6 * Copyright (C) 2011 Tejun Heo <tj@kernel.org>
7 *
Omar Sandoval3140c3c2017-08-24 11:09:25 -07008 * REQ_{PREFLUSH|FUA} requests are decomposed to sequences consisted of three
Tejun Heoae1b1532011-01-25 12:43:54 +01009 * optional steps - PREFLUSH, DATA and POSTFLUSH - according to the request
10 * properties and hardware capability.
11 *
Mike Christie28a8f0d2016-06-05 14:32:25 -050012 * 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
Tejun Heoae1b1532011-01-25 12:43:54 +010014 * 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
16 * completion.
17 *
Omar Sandoval3140c3c2017-08-24 11:09:25 -070018 * 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.
Tejun Heoae1b1532011-01-25 12:43:54 +010021 *
Mike Christie28a8f0d2016-06-05 14:32:25 -050022 * If the device has writeback cache and supports FUA, REQ_PREFLUSH is
Tejun Heoae1b1532011-01-25 12:43:54 +010023 * translated to PREFLUSH but REQ_FUA is passed down directly with DATA.
24 *
Mike Christie28a8f0d2016-06-05 14:32:25 -050025 * If the device has writeback cache and doesn't support FUA, REQ_PREFLUSH
26 * is translated to PREFLUSH and REQ_FUA to POSTFLUSH.
Tejun Heoae1b1532011-01-25 12:43:54 +010027 *
28 * The actual execution of flush is double buffered. Whenever a request
29 * needs to execute PRE or POSTFLUSH, it queues at
Ming Lei7c94e1c2014-09-25 23:23:43 +080030 * fq->flush_queue[fq->flush_pending_idx]. Once certain criteria are met, a
Mike Christie3a5e02c2016-06-05 14:32:23 -050031 * REQ_OP_FLUSH is issued and the pending_idx is toggled. When the flush
Tejun Heoae1b1532011-01-25 12:43:54 +010032 * completes, all the requests which were pending are proceeded to the next
Omar Sandoval3140c3c2017-08-24 11:09:25 -070033 * step. This allows arbitrary merging of different types of PREFLUSH/FUA
Tejun Heoae1b1532011-01-25 12:43:54 +010034 * requests.
35 *
36 * Currently, the following conditions are used to determine when to issue
37 * flush.
38 *
39 * C1. At any given time, only one flush shall be in progress. This makes
40 * double buffering sufficient.
41 *
42 * C2. Flush is deferred if any request is executing DATA of its sequence.
43 * This avoids issuing separate POSTFLUSHes for requests which shared
44 * PREFLUSH.
45 *
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
Omar Sandoval3140c3c2017-08-24 11:09:25 -070049 * FUA (without PREFLUSH) requests.
Tejun Heoae1b1532011-01-25 12:43:54 +010050 *
51 * For devices which support FUA, it isn't clear whether C2 (and thus C3)
52 * is beneficial.
53 *
Omar Sandoval3140c3c2017-08-24 11:09:25 -070054 * Note that a sequenced PREFLUSH/FUA request with DATA is completed twice.
Tejun Heoae1b1532011-01-25 12:43:54 +010055 * 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
Christoph Hellwige8064022016-10-20 15:12:13 +020058 * sequence is complete. This is implemented by testing RQF_FLUSH_SEQ in
Tejun Heoae1b1532011-01-25 12:43:54 +010059 * req_bio_endio().
60 *
Omar Sandoval3140c3c2017-08-24 11:09:25 -070061 * The above peculiarity requires that each PREFLUSH/FUA request has only one
Tejun Heoae1b1532011-01-25 12:43:54 +010062 * bio attached to it, which is guaranteed as they aren't allowed to be
63 * merged in the usual way.
Jens Axboe86db1e22008-01-29 14:53:40 +010064 */
Tejun Heoae1b1532011-01-25 12:43:54 +010065
Jens Axboe86db1e22008-01-29 14:53:40 +010066#include <linux/kernel.h>
67#include <linux/module.h>
68#include <linux/bio.h>
69#include <linux/blkdev.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090070#include <linux/gfp.h>
Jens Axboe320ae512013-10-24 09:20:05 +010071#include <linux/blk-mq.h>
Bart Van Asscheb3c6a592019-12-17 16:24:35 -080072#include <linux/lockdep.h>
Jens Axboe86db1e22008-01-29 14:53:40 +010073
74#include "blk.h"
Jens Axboe320ae512013-10-24 09:20:05 +010075#include "blk-mq.h"
Ming Lei0048b482015-08-09 03:41:51 -040076#include "blk-mq-tag.h"
Jens Axboebd166ef2017-01-17 06:03:22 -070077#include "blk-mq-sched.h"
Jens Axboe86db1e22008-01-29 14:53:40 +010078
Omar Sandoval3140c3c2017-08-24 11:09:25 -070079/* PREFLUSH/FUA sequences */
Tejun Heo4fed9472010-09-03 11:56:17 +020080enum {
Tejun Heoae1b1532011-01-25 12:43:54 +010081 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),
Tejun Heo4fed9472010-09-03 11:56:17 +020085
Tejun Heoae1b1532011-01-25 12:43:54 +010086 REQ_FSEQ_ACTIONS = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
87 REQ_FSEQ_POSTFLUSH,
Tejun Heo47f70d52010-09-03 11:56:17 +020088
89 /*
Tejun Heoae1b1532011-01-25 12:43:54 +010090 * If flush has been pending longer than the following timeout,
91 * it's issued even if flush_data requests are still in flight.
Tejun Heo47f70d52010-09-03 11:56:17 +020092 */
Tejun Heoae1b1532011-01-25 12:43:54 +010093 FLUSH_PENDING_TIMEOUT = 5 * HZ,
94};
95
Jens Axboe404b8f52018-10-31 12:43:24 -060096static void blk_kick_flush(struct request_queue *q,
Hannes Reinecke84fca1b2018-06-06 16:21:40 +020097 struct blk_flush_queue *fq, unsigned int flags);
Tejun Heoae1b1532011-01-25 12:43:54 +010098
Jens Axboec888a8f2016-04-13 13:33:19 -060099static unsigned int blk_flush_policy(unsigned long fflags, struct request *rq)
Tejun Heoae1b1532011-01-25 12:43:54 +0100100{
101 unsigned int policy = 0;
102
Jeff Moyerfa1bf422011-08-09 20:32:09 +0200103 if (blk_rq_sectors(rq))
104 policy |= REQ_FSEQ_DATA;
105
Jens Axboec888a8f2016-04-13 13:33:19 -0600106 if (fflags & (1UL << QUEUE_FLAG_WC)) {
Mike Christie28a8f0d2016-06-05 14:32:25 -0500107 if (rq->cmd_flags & REQ_PREFLUSH)
Tejun Heoae1b1532011-01-25 12:43:54 +0100108 policy |= REQ_FSEQ_PREFLUSH;
Jens Axboec888a8f2016-04-13 13:33:19 -0600109 if (!(fflags & (1UL << QUEUE_FLAG_FUA)) &&
110 (rq->cmd_flags & REQ_FUA))
Tejun Heoae1b1532011-01-25 12:43:54 +0100111 policy |= REQ_FSEQ_POSTFLUSH;
112 }
113 return policy;
Tejun Heo47f70d52010-09-03 11:56:17 +0200114}
115
Tejun Heoae1b1532011-01-25 12:43:54 +0100116static unsigned int blk_flush_cur_seq(struct request *rq)
Jens Axboe86db1e22008-01-29 14:53:40 +0100117{
Tejun Heoae1b1532011-01-25 12:43:54 +0100118 return 1 << ffz(rq->flush.seq);
Jens Axboe86db1e22008-01-29 14:53:40 +0100119}
120
Tejun Heoae1b1532011-01-25 12:43:54 +0100121static void blk_flush_restore_request(struct request *rq)
Jens Axboe86db1e22008-01-29 14:53:40 +0100122{
Tejun Heoae1b1532011-01-25 12:43:54 +0100123 /*
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.
127 */
128 rq->bio = rq->biotail;
129
130 /* make @rq a normal request */
Christoph Hellwige8064022016-10-20 15:12:13 +0200131 rq->rq_flags &= ~RQF_FLUSH_SEQ;
Jeff Moyer4853aba2011-08-15 21:37:25 +0200132 rq->end_io = rq->flush.saved_end_io;
Jens Axboe320ae512013-10-24 09:20:05 +0100133}
134
Jens Axboe404b8f52018-10-31 12:43:24 -0600135static void blk_flush_queue_rq(struct request *rq, bool add_front)
Jens Axboe320ae512013-10-24 09:20:05 +0100136{
Jens Axboe7e992f82018-10-13 15:44:12 -0600137 blk_mq_add_to_requeue_list(rq, add_front, true);
Jens Axboe86db1e22008-01-29 14:53:40 +0100138}
139
Konstantin Khlebnikovb6866312019-11-21 13:40:26 +0300140static void blk_account_io_flush(struct request *rq)
141{
142 struct hd_struct *part = &rq->rq_disk->part0;
143
144 part_stat_lock();
145 part_stat_inc(part, ios[STAT_FLUSH]);
146 part_stat_add(part, nsecs[STAT_FLUSH],
147 ktime_get_ns() - rq->start_time_ns);
148 part_stat_unlock();
149}
150
Tejun Heoae1b1532011-01-25 12:43:54 +0100151/**
152 * blk_flush_complete_seq - complete flush sequence
Omar Sandoval3140c3c2017-08-24 11:09:25 -0700153 * @rq: PREFLUSH/FUA request being sequenced
Ming Lei0bae3522014-09-25 23:23:45 +0800154 * @fq: flush queue
Tejun Heoae1b1532011-01-25 12:43:54 +0100155 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
156 * @error: whether an error occurred
157 *
158 * @rq just completed @seq part of its flush sequence, record the
159 * completion and trigger the next step.
160 *
161 * CONTEXT:
Christoph Hellwig9809b4e2018-11-14 17:02:10 +0100162 * spin_lock_irq(fq->mq_flush_lock)
Tejun Heoae1b1532011-01-25 12:43:54 +0100163 */
Jens Axboe404b8f52018-10-31 12:43:24 -0600164static void blk_flush_complete_seq(struct request *rq,
Ming Lei0bae3522014-09-25 23:23:45 +0800165 struct blk_flush_queue *fq,
Christoph Hellwig2a842ac2017-06-03 09:38:04 +0200166 unsigned int seq, blk_status_t error)
Jens Axboe86db1e22008-01-29 14:53:40 +0100167{
Tejun Heoae1b1532011-01-25 12:43:54 +0100168 struct request_queue *q = rq->q;
Ming Lei7c94e1c2014-09-25 23:23:43 +0800169 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
Jens Axboe190b02e2018-06-09 06:37:14 -0600170 unsigned int cmd_flags;
Jens Axboe86db1e22008-01-29 14:53:40 +0100171
Tejun Heoae1b1532011-01-25 12:43:54 +0100172 BUG_ON(rq->flush.seq & seq);
173 rq->flush.seq |= seq;
Jens Axboe190b02e2018-06-09 06:37:14 -0600174 cmd_flags = rq->cmd_flags;
Jens Axboe86db1e22008-01-29 14:53:40 +0100175
Tejun Heoae1b1532011-01-25 12:43:54 +0100176 if (likely(!error))
177 seq = blk_flush_cur_seq(rq);
178 else
179 seq = REQ_FSEQ_DONE;
Tejun Heo28e7d182010-09-03 11:56:16 +0200180
Tejun Heoae1b1532011-01-25 12:43:54 +0100181 switch (seq) {
182 case REQ_FSEQ_PREFLUSH:
183 case REQ_FSEQ_POSTFLUSH:
184 /* queue for flush */
185 if (list_empty(pending))
Ming Lei7c94e1c2014-09-25 23:23:43 +0800186 fq->flush_pending_since = jiffies;
Tejun Heoae1b1532011-01-25 12:43:54 +0100187 list_move_tail(&rq->flush.list, pending);
Tejun Heo28e7d182010-09-03 11:56:16 +0200188 break;
Tejun Heoae1b1532011-01-25 12:43:54 +0100189
190 case REQ_FSEQ_DATA:
Ming Lei7c94e1c2014-09-25 23:23:43 +0800191 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
Jens Axboe404b8f52018-10-31 12:43:24 -0600192 blk_flush_queue_rq(rq, true);
Tejun Heoae1b1532011-01-25 12:43:54 +0100193 break;
194
195 case REQ_FSEQ_DONE:
Tejun Heo09d60c72010-09-03 11:56:17 +0200196 /*
Konstantin Khlebnikovb6866312019-11-21 13:40:26 +0300197 * @rq was previously adjusted by blk_insert_flush() for
Tejun Heoae1b1532011-01-25 12:43:54 +0100198 * flush sequencing and may already have gone through the
199 * flush data request completion path. Restore @rq for
200 * normal completion and end it.
Tejun Heo09d60c72010-09-03 11:56:17 +0200201 */
Tejun Heoae1b1532011-01-25 12:43:54 +0100202 BUG_ON(!list_empty(&rq->queuelist));
203 list_del_init(&rq->flush.list);
204 blk_flush_restore_request(rq);
Jens Axboe7e992f82018-10-13 15:44:12 -0600205 blk_mq_end_request(rq, error);
Tejun Heo28e7d182010-09-03 11:56:16 +0200206 break;
Tejun Heoae1b1532011-01-25 12:43:54 +0100207
Tejun Heo28e7d182010-09-03 11:56:16 +0200208 default:
209 BUG();
210 }
Christoph Hellwigcde4c402010-09-03 11:56:17 +0200211
Jens Axboe404b8f52018-10-31 12:43:24 -0600212 blk_kick_flush(q, fq, cmd_flags);
Tejun Heo28e7d182010-09-03 11:56:16 +0200213}
214
Christoph Hellwig2a842ac2017-06-03 09:38:04 +0200215static void flush_end_io(struct request *flush_rq, blk_status_t error)
Jens Axboe86db1e22008-01-29 14:53:40 +0100216{
Tejun Heoae1b1532011-01-25 12:43:54 +0100217 struct request_queue *q = flush_rq->q;
Jens Axboe320ae512013-10-24 09:20:05 +0100218 struct list_head *running;
Tejun Heoae1b1532011-01-25 12:43:54 +0100219 struct request *rq, *n;
Jens Axboe320ae512013-10-24 09:20:05 +0100220 unsigned long flags = 0;
Ming Leie97c2932014-09-25 23:23:46 +0800221 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900222
Konstantin Khlebnikovb6866312019-11-21 13:40:26 +0300223 blk_account_io_flush(flush_rq);
224
Jens Axboe7e992f82018-10-13 15:44:12 -0600225 /* release the tag's ownership to the req cloned from */
226 spin_lock_irqsave(&fq->mq_flush_lock, flags);
Yufen Yu8d699662019-09-27 16:19:55 +0800227
228 if (!refcount_dec_and_test(&flush_rq->ref)) {
229 fq->rq_status = error;
230 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
231 return;
232 }
233
234 if (fq->rq_status != BLK_STS_OK)
235 error = fq->rq_status;
236
Jens Axboe4e2f62e52020-07-01 22:58:32 -0600237 if (!q->elevator) {
Ming Lei568f2702020-07-06 22:41:11 +0800238 flush_rq->tag = BLK_MQ_NO_TAG;
Jens Axboe4e2f62e52020-07-01 22:58:32 -0600239 } else {
240 blk_mq_put_driver_tag(flush_rq);
Ming Lei568f2702020-07-06 22:41:11 +0800241 flush_rq->internal_tag = BLK_MQ_NO_TAG;
Jens Axboe4e2f62e52020-07-01 22:58:32 -0600242 }
Christoph Hellwig18741982014-02-10 09:29:00 -0700243
Ming Lei7c94e1c2014-09-25 23:23:43 +0800244 running = &fq->flush_queue[fq->flush_running_idx];
245 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
Tejun Heoae1b1532011-01-25 12:43:54 +0100246
247 /* account completion of the flush request */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800248 fq->flush_running_idx ^= 1;
Jens Axboe320ae512013-10-24 09:20:05 +0100249
Tejun Heoae1b1532011-01-25 12:43:54 +0100250 /* 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);
253
254 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
Jens Axboe404b8f52018-10-31 12:43:24 -0600255 blk_flush_complete_seq(rq, fq, seq, error);
Tejun Heo28e7d182010-09-03 11:56:16 +0200256 }
257
Jens Axboe7e992f82018-10-13 15:44:12 -0600258 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
Jens Axboe320ae512013-10-24 09:20:05 +0100259}
260
Tejun Heoae1b1532011-01-25 12:43:54 +0100261/**
262 * blk_kick_flush - consider issuing flush request
263 * @q: request_queue being kicked
Ming Lei0bae3522014-09-25 23:23:45 +0800264 * @fq: flush queue
Hannes Reinecke84fca1b2018-06-06 16:21:40 +0200265 * @flags: cmd_flags of the original request
Tejun Heoae1b1532011-01-25 12:43:54 +0100266 *
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.
269 *
270 * CONTEXT:
Christoph Hellwig9809b4e2018-11-14 17:02:10 +0100271 * spin_lock_irq(fq->mq_flush_lock)
Tejun Heoae1b1532011-01-25 12:43:54 +0100272 *
Tejun Heoae1b1532011-01-25 12:43:54 +0100273 */
Jens Axboe404b8f52018-10-31 12:43:24 -0600274static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
Hannes Reinecke84fca1b2018-06-06 16:21:40 +0200275 unsigned int flags)
Tejun Heoae1b1532011-01-25 12:43:54 +0100276{
Ming Lei7c94e1c2014-09-25 23:23:43 +0800277 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
Tejun Heoae1b1532011-01-25 12:43:54 +0100278 struct request *first_rq =
279 list_first_entry(pending, struct request, flush.list);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800280 struct request *flush_rq = fq->flush_rq;
Tejun Heoae1b1532011-01-25 12:43:54 +0100281
282 /* C1 described at the top of this file */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800283 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
Jens Axboe404b8f52018-10-31 12:43:24 -0600284 return;
Tejun Heoae1b1532011-01-25 12:43:54 +0100285
Yufen Yub5718d62020-07-16 02:52:01 -0400286 /* C2 and C3 */
287 if (!list_empty(&fq->flush_data_in_flight) &&
Tejun Heoae1b1532011-01-25 12:43:54 +0100288 time_before(jiffies,
Ming Lei7c94e1c2014-09-25 23:23:43 +0800289 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
Jens Axboe404b8f52018-10-31 12:43:24 -0600290 return;
Tejun Heo28e7d182010-09-03 11:56:16 +0200291
292 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100293 * Issue flush and toggle pending_idx. This makes pending_idx
294 * different from running_idx, which means flush is in flight.
Tejun Heo28e7d182010-09-03 11:56:16 +0200295 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800296 fq->flush_pending_idx ^= 1;
Christoph Hellwig18741982014-02-10 09:29:00 -0700297
Ming Lei7ddab5d2014-09-25 23:23:42 +0800298 blk_rq_init(q, flush_rq);
Ming Leif70ced02014-09-25 23:23:47 +0800299
300 /*
Ming Lei923218f2017-11-02 23:24:38 +0800301 * 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.
304 *
305 * In case of IO scheduler, flush rq need to borrow scheduler tag
306 * just for cheating put/get driver tag.
Ming Leif70ced02014-09-25 23:23:47 +0800307 */
Jens Axboe7e992f82018-10-13 15:44:12 -0600308 flush_rq->mq_ctx = first_rq->mq_ctx;
Jens Axboeea4f9952018-10-29 15:06:13 -0600309 flush_rq->mq_hctx = first_rq->mq_hctx;
Ming Lei0048b482015-08-09 03:41:51 -0400310
Ming Lei568f2702020-07-06 22:41:11 +0800311 if (!q->elevator)
Jens Axboe7e992f82018-10-13 15:44:12 -0600312 flush_rq->tag = first_rq->tag;
Ming Lei568f2702020-07-06 22:41:11 +0800313 else
Jens Axboe7e992f82018-10-13 15:44:12 -0600314 flush_rq->internal_tag = first_rq->internal_tag;
Jens Axboe320ae512013-10-24 09:20:05 +0100315
Christoph Hellwig70fd7612016-11-01 07:40:10 -0600316 flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
Hannes Reinecke84fca1b2018-06-06 16:21:40 +0200317 flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK);
Christoph Hellwige8064022016-10-20 15:12:13 +0200318 flush_rq->rq_flags |= RQF_FLUSH_SEQ;
Ming Lei7ddab5d2014-09-25 23:23:42 +0800319 flush_rq->rq_disk = first_rq->rq_disk;
320 flush_rq->end_io = flush_end_io;
Jens Axboe86db1e22008-01-29 14:53:40 +0100321
Jens Axboe404b8f52018-10-31 12:43:24 -0600322 blk_flush_queue_rq(flush_rq, false);
Tejun Heoae1b1532011-01-25 12:43:54 +0100323}
324
Christoph Hellwig2a842ac2017-06-03 09:38:04 +0200325static void mq_flush_data_end_io(struct request *rq, blk_status_t error)
Jens Axboe320ae512013-10-24 09:20:05 +0100326{
327 struct request_queue *q = rq->q;
Jens Axboeea4f9952018-10-29 15:06:13 -0600328 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
Ming Leie97c2932014-09-25 23:23:46 +0800329 struct blk_mq_ctx *ctx = rq->mq_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +0100330 unsigned long flags;
Ming Leie97c2932014-09-25 23:23:46 +0800331 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +0100332
Jens Axboe4e2f62e52020-07-01 22:58:32 -0600333 if (q->elevator) {
334 WARN_ON(rq->tag < 0);
335 blk_mq_put_driver_tag(rq);
336 }
337
Jens Axboe320ae512013-10-24 09:20:05 +0100338 /*
339 * After populating an empty queue, kick it to avoid stall. Read
340 * the comment in flush_end_io().
341 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800342 spin_lock_irqsave(&fq->mq_flush_lock, flags);
Jens Axboebd166ef2017-01-17 06:03:22 -0700343 blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800344 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
Jens Axboebd166ef2017-01-17 06:03:22 -0700345
Jianchao Wang85bd6e62019-01-30 17:01:56 +0800346 blk_mq_sched_restart(hctx);
Jens Axboe320ae512013-10-24 09:20:05 +0100347}
348
Tejun Heoae1b1532011-01-25 12:43:54 +0100349/**
Omar Sandoval3140c3c2017-08-24 11:09:25 -0700350 * blk_insert_flush - insert a new PREFLUSH/FUA request
Tejun Heoae1b1532011-01-25 12:43:54 +0100351 * @rq: request to insert
352 *
Jens Axboeb710a482011-03-30 09:52:30 +0200353 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
Jens Axboe320ae512013-10-24 09:20:05 +0100354 * or __blk_mq_run_hw_queue() to dispatch request.
Tejun Heoae1b1532011-01-25 12:43:54 +0100355 * @rq is being submitted. Analyze what needs to be done and put it on the
356 * right queue.
Tejun Heoae1b1532011-01-25 12:43:54 +0100357 */
358void blk_insert_flush(struct request *rq)
359{
360 struct request_queue *q = rq->q;
Jens Axboec888a8f2016-04-13 13:33:19 -0600361 unsigned long fflags = q->queue_flags; /* may change, cache */
Tejun Heoae1b1532011-01-25 12:43:54 +0100362 unsigned int policy = blk_flush_policy(fflags, rq);
Ming Leie97c2932014-09-25 23:23:46 +0800363 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
Tejun Heoae1b1532011-01-25 12:43:54 +0100364
Tejun Heoae1b1532011-01-25 12:43:54 +0100365 /*
366 * @policy now records what operations need to be done. Adjust
Mike Christie28a8f0d2016-06-05 14:32:25 -0500367 * REQ_PREFLUSH and FUA for the driver.
Tejun Heoae1b1532011-01-25 12:43:54 +0100368 */
Mike Christie28a8f0d2016-06-05 14:32:25 -0500369 rq->cmd_flags &= ~REQ_PREFLUSH;
Jens Axboec888a8f2016-04-13 13:33:19 -0600370 if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
Tejun Heo4fed9472010-09-03 11:56:17 +0200371 rq->cmd_flags &= ~REQ_FUA;
Jens Axboe86db1e22008-01-29 14:53:40 +0100372
Tejun Heoae1b1532011-01-25 12:43:54 +0100373 /*
Jens Axboeae5b2ec2016-11-08 19:39:28 -0700374 * 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.
377 */
378 rq->cmd_flags |= REQ_SYNC;
379
380 /*
Jeff Moyer4853aba2011-08-15 21:37:25 +0200381 * 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.
385 */
386 if (!policy) {
Jens Axboe7e992f82018-10-13 15:44:12 -0600387 blk_mq_end_request(rq, 0);
Jeff Moyer4853aba2011-08-15 21:37:25 +0200388 return;
389 }
390
Jeff Moyer834f9f62011-10-17 12:57:22 +0200391 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
Jeff Moyer4853aba2011-08-15 21:37:25 +0200392
393 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100394 * 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.
397 */
398 if ((policy & REQ_FSEQ_DATA) &&
399 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
Ming Lei01e99ae2020-02-25 09:04:32 +0800400 blk_mq_request_bypass_insert(rq, false, false);
Tejun Heoae1b1532011-01-25 12:43:54 +0100401 return;
402 }
403
404 /*
405 * @rq should go through flush machinery. Mark it part of flush
406 * sequence and submit for further processing.
407 */
408 memset(&rq->flush, 0, sizeof(rq->flush));
409 INIT_LIST_HEAD(&rq->flush.list);
Christoph Hellwige8064022016-10-20 15:12:13 +0200410 rq->rq_flags |= RQF_FLUSH_SEQ;
Jeff Moyer4853aba2011-08-15 21:37:25 +0200411 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
Jens Axboe320ae512013-10-24 09:20:05 +0100412
Jens Axboe7e992f82018-10-13 15:44:12 -0600413 rq->end_io = mq_flush_data_end_io;
Tejun Heoae1b1532011-01-25 12:43:54 +0100414
Jens Axboe7e992f82018-10-13 15:44:12 -0600415 spin_lock_irq(&fq->mq_flush_lock);
Ming Lei0bae3522014-09-25 23:23:45 +0800416 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
Jens Axboe7e992f82018-10-13 15:44:12 -0600417 spin_unlock_irq(&fq->mq_flush_lock);
Tejun Heoae1b1532011-01-25 12:43:54 +0100418}
419
420/**
Jens Axboe86db1e22008-01-29 14:53:40 +0100421 * blkdev_issue_flush - queue a flush
422 * @bdev: blockdev to issue flush for
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400423 * @gfp_mask: memory allocation flags (for bio_alloc)
Jens Axboe86db1e22008-01-29 14:53:40 +0100424 *
425 * Description:
Christoph Hellwig93985542020-05-13 14:36:00 +0200426 * Issue a flush for the block device in question.
Jens Axboe86db1e22008-01-29 14:53:40 +0100427 */
Christoph Hellwig93985542020-05-13 14:36:00 +0200428int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask)
Jens Axboe86db1e22008-01-29 14:53:40 +0100429{
Jens Axboe86db1e22008-01-29 14:53:40 +0100430 struct bio *bio;
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400431 int ret = 0;
Jens Axboe86db1e22008-01-29 14:53:40 +0100432
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400433 bio = bio_alloc(gfp_mask, 0);
Christoph Hellwig74d46992017-08-23 19:10:32 +0200434 bio_set_dev(bio, bdev);
Christoph Hellwig70fd7612016-11-01 07:40:10 -0600435 bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
Dmitry Monakhovf17e2322010-04-28 17:55:07 +0400436
Mike Christie4e49ea42016-06-05 14:31:41 -0500437 ret = submit_bio_wait(bio);
Jens Axboe86db1e22008-01-29 14:53:40 +0100438 bio_put(bio);
439 return ret;
440}
Jens Axboe86db1e22008-01-29 14:53:40 +0100441EXPORT_SYMBOL(blkdev_issue_flush);
Jens Axboe320ae512013-10-24 09:20:05 +0100442
Guoqing Jiang754a1572020-03-09 22:41:37 +0100443struct blk_flush_queue *blk_alloc_flush_queue(int node, int cmd_size,
444 gfp_t flags)
Jens Axboe320ae512013-10-24 09:20:05 +0100445{
Ming Lei7c94e1c2014-09-25 23:23:43 +0800446 struct blk_flush_queue *fq;
447 int rq_sz = sizeof(struct request);
Ming Lei1bcb1ea2014-09-25 23:23:39 +0800448
Jianchao Wang5b202852018-10-12 18:07:26 +0800449 fq = kzalloc_node(sizeof(*fq), flags, node);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800450 if (!fq)
451 goto fail;
Ming Lei1bcb1ea2014-09-25 23:23:39 +0800452
Jens Axboe7e992f82018-10-13 15:44:12 -0600453 spin_lock_init(&fq->mq_flush_lock);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800454
Christoph Hellwig6d247d72017-01-27 09:51:45 -0700455 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
Jianchao Wang5b202852018-10-12 18:07:26 +0800456 fq->flush_rq = kzalloc_node(rq_sz, flags, node);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800457 if (!fq->flush_rq)
458 goto fail_rq;
459
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);
463
Bart Van Asscheb3c6a592019-12-17 16:24:35 -0800464 lockdep_register_key(&fq->key);
465 lockdep_set_class(&fq->mq_flush_lock, &fq->key);
466
Ming Lei7c94e1c2014-09-25 23:23:43 +0800467 return fq;
468
469 fail_rq:
470 kfree(fq);
471 fail:
472 return NULL;
473}
474
Ming Leiba483382014-09-25 23:23:44 +0800475void blk_free_flush_queue(struct blk_flush_queue *fq)
Ming Lei7c94e1c2014-09-25 23:23:43 +0800476{
477 /* bio based request queue hasn't flush queue */
478 if (!fq)
479 return;
480
Bart Van Asscheb3c6a592019-12-17 16:24:35 -0800481 lockdep_unregister_key(&fq->key);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800482 kfree(fq->flush_rq);
483 kfree(fq);
Jens Axboe320ae512013-10-24 09:20:05 +0100484}