blob: 6a14b68b91358bdd28d90fb5ab23100a77af31dd [file] [log] [blame]
Jens Axboe86db1e22008-01-29 14:53:40 +01001/*
Tejun Heo4fed9472010-09-03 11:56:17 +02002 * Functions to sequence FLUSH and FUA writes.
Tejun Heoae1b1532011-01-25 12:43:54 +01003 *
4 * Copyright (C) 2011 Max Planck Institute for Gravitational Physics
5 * Copyright (C) 2011 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
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.
12 *
Mike Christie28a8f0d2016-06-05 14:32:25 -050013 * If a request doesn't have data, only REQ_PREFLUSH makes sense, which
14 * indicates a simple flush request. If there is data, REQ_PREFLUSH indicates
Tejun Heoae1b1532011-01-25 12:43:54 +010015 * 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
17 * completion.
18 *
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.
22 *
Mike Christie28a8f0d2016-06-05 14:32:25 -050023 * If the device has writeback cache and supports FUA, REQ_PREFLUSH is
Tejun Heoae1b1532011-01-25 12:43:54 +010024 * translated to PREFLUSH but REQ_FUA is passed down directly with DATA.
25 *
Mike Christie28a8f0d2016-06-05 14:32:25 -050026 * If the device has writeback cache and doesn't support FUA, REQ_PREFLUSH
27 * is translated to PREFLUSH and REQ_FUA to POSTFLUSH.
Tejun Heoae1b1532011-01-25 12:43:54 +010028 *
29 * The actual execution of flush is double buffered. Whenever a request
30 * needs to execute PRE or POSTFLUSH, it queues at
Ming Lei7c94e1c2014-09-25 23:23:43 +080031 * fq->flush_queue[fq->flush_pending_idx]. Once certain criteria are met, a
Mike Christie3a5e02c2016-06-05 14:32:23 -050032 * REQ_OP_FLUSH is issued and the pending_idx is toggled. When the flush
Tejun Heoae1b1532011-01-25 12:43:54 +010033 * completes, all the requests which were pending are proceeded to the next
34 * step. This allows arbitrary merging of different types of FLUSH/FUA
35 * requests.
36 *
37 * Currently, the following conditions are used to determine when to issue
38 * flush.
39 *
40 * C1. At any given time, only one flush shall be in progress. This makes
41 * double buffering sufficient.
42 *
43 * C2. Flush is deferred if any request is executing DATA of its sequence.
44 * This avoids issuing separate POSTFLUSHes for requests which shared
45 * PREFLUSH.
46 *
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.
51 *
52 * For devices which support FUA, it isn't clear whether C2 (and thus C3)
53 * is beneficial.
54 *
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
60 * req_bio_endio().
61 *
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.
Jens Axboe86db1e22008-01-29 14:53:40 +010065 */
Tejun Heoae1b1532011-01-25 12:43:54 +010066
Jens Axboe86db1e22008-01-29 14:53:40 +010067#include <linux/kernel.h>
68#include <linux/module.h>
69#include <linux/bio.h>
70#include <linux/blkdev.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090071#include <linux/gfp.h>
Jens Axboe320ae512013-10-24 09:20:05 +010072#include <linux/blk-mq.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 Axboe86db1e22008-01-29 14:53:40 +010077
Tejun Heo4fed9472010-09-03 11:56:17 +020078/* FLUSH/FUA sequences */
79enum {
Tejun Heoae1b1532011-01-25 12:43:54 +010080 REQ_FSEQ_PREFLUSH = (1 << 0), /* pre-flushing in progress */
81 REQ_FSEQ_DATA = (1 << 1), /* data write in progress */
82 REQ_FSEQ_POSTFLUSH = (1 << 2), /* post-flushing in progress */
83 REQ_FSEQ_DONE = (1 << 3),
Tejun Heo4fed9472010-09-03 11:56:17 +020084
Tejun Heoae1b1532011-01-25 12:43:54 +010085 REQ_FSEQ_ACTIONS = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
86 REQ_FSEQ_POSTFLUSH,
Tejun Heo47f70d52010-09-03 11:56:17 +020087
88 /*
Tejun Heoae1b1532011-01-25 12:43:54 +010089 * If flush has been pending longer than the following timeout,
90 * it's issued even if flush_data requests are still in flight.
Tejun Heo47f70d52010-09-03 11:56:17 +020091 */
Tejun Heoae1b1532011-01-25 12:43:54 +010092 FLUSH_PENDING_TIMEOUT = 5 * HZ,
93};
94
Ming Lei0bae3522014-09-25 23:23:45 +080095static bool blk_kick_flush(struct request_queue *q,
96 struct blk_flush_queue *fq);
Tejun Heoae1b1532011-01-25 12:43:54 +010097
Jens Axboec888a8f2016-04-13 13:33:19 -060098static unsigned int blk_flush_policy(unsigned long fflags, struct request *rq)
Tejun Heoae1b1532011-01-25 12:43:54 +010099{
100 unsigned int policy = 0;
101
Jeff Moyerfa1bf422011-08-09 20:32:09 +0200102 if (blk_rq_sectors(rq))
103 policy |= REQ_FSEQ_DATA;
104
Jens Axboec888a8f2016-04-13 13:33:19 -0600105 if (fflags & (1UL << QUEUE_FLAG_WC)) {
Mike Christie28a8f0d2016-06-05 14:32:25 -0500106 if (rq->cmd_flags & REQ_PREFLUSH)
Tejun Heoae1b1532011-01-25 12:43:54 +0100107 policy |= REQ_FSEQ_PREFLUSH;
Jens Axboec888a8f2016-04-13 13:33:19 -0600108 if (!(fflags & (1UL << QUEUE_FLAG_FUA)) &&
109 (rq->cmd_flags & REQ_FUA))
Tejun Heoae1b1532011-01-25 12:43:54 +0100110 policy |= REQ_FSEQ_POSTFLUSH;
111 }
112 return policy;
Tejun Heo47f70d52010-09-03 11:56:17 +0200113}
114
Tejun Heoae1b1532011-01-25 12:43:54 +0100115static unsigned int blk_flush_cur_seq(struct request *rq)
Jens Axboe86db1e22008-01-29 14:53:40 +0100116{
Tejun Heoae1b1532011-01-25 12:43:54 +0100117 return 1 << ffz(rq->flush.seq);
Jens Axboe86db1e22008-01-29 14:53:40 +0100118}
119
Tejun Heoae1b1532011-01-25 12:43:54 +0100120static void blk_flush_restore_request(struct request *rq)
Jens Axboe86db1e22008-01-29 14:53:40 +0100121{
Tejun Heoae1b1532011-01-25 12:43:54 +0100122 /*
123 * After flush data completion, @rq->bio is %NULL but we need to
124 * complete the bio again. @rq->biotail is guaranteed to equal the
125 * original @rq->bio. Restore it.
126 */
127 rq->bio = rq->biotail;
128
129 /* make @rq a normal request */
130 rq->cmd_flags &= ~REQ_FLUSH_SEQ;
Jeff Moyer4853aba2011-08-15 21:37:25 +0200131 rq->end_io = rq->flush.saved_end_io;
Jens Axboe320ae512013-10-24 09:20:05 +0100132}
133
Mike Snitzer10beafc2014-03-08 20:19:20 -0700134static bool blk_flush_queue_rq(struct request *rq, bool add_front)
Jens Axboe320ae512013-10-24 09:20:05 +0100135{
Christoph Hellwig18741982014-02-10 09:29:00 -0700136 if (rq->q->mq_ops) {
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600137 struct request_queue *q = rq->q;
138
139 blk_mq_add_to_requeue_list(rq, add_front);
140 blk_mq_kick_requeue_list(q);
Christoph Hellwig18741982014-02-10 09:29:00 -0700141 return false;
142 } else {
Mike Snitzer10beafc2014-03-08 20:19:20 -0700143 if (add_front)
144 list_add(&rq->queuelist, &rq->q->queue_head);
145 else
146 list_add_tail(&rq->queuelist, &rq->q->queue_head);
Christoph Hellwig18741982014-02-10 09:29:00 -0700147 return true;
148 }
Jens Axboe86db1e22008-01-29 14:53:40 +0100149}
150
Tejun Heoae1b1532011-01-25 12:43:54 +0100151/**
152 * blk_flush_complete_seq - complete flush sequence
153 * @rq: FLUSH/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:
Ming Lei7c94e1c2014-09-25 23:23:43 +0800162 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
Tejun Heoae1b1532011-01-25 12:43:54 +0100163 *
164 * RETURNS:
165 * %true if requests were added to the dispatch queue, %false otherwise.
166 */
Ming Lei0bae3522014-09-25 23:23:45 +0800167static bool blk_flush_complete_seq(struct request *rq,
168 struct blk_flush_queue *fq,
169 unsigned int seq, int error)
Jens Axboe86db1e22008-01-29 14:53:40 +0100170{
Tejun Heoae1b1532011-01-25 12:43:54 +0100171 struct request_queue *q = rq->q;
Ming Lei7c94e1c2014-09-25 23:23:43 +0800172 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
Jens Axboe320ae512013-10-24 09:20:05 +0100173 bool queued = false, kicked;
Jens Axboe86db1e22008-01-29 14:53:40 +0100174
Tejun Heoae1b1532011-01-25 12:43:54 +0100175 BUG_ON(rq->flush.seq & seq);
176 rq->flush.seq |= seq;
Jens Axboe86db1e22008-01-29 14:53:40 +0100177
Tejun Heoae1b1532011-01-25 12:43:54 +0100178 if (likely(!error))
179 seq = blk_flush_cur_seq(rq);
180 else
181 seq = REQ_FSEQ_DONE;
Tejun Heo28e7d182010-09-03 11:56:16 +0200182
Tejun Heoae1b1532011-01-25 12:43:54 +0100183 switch (seq) {
184 case REQ_FSEQ_PREFLUSH:
185 case REQ_FSEQ_POSTFLUSH:
186 /* queue for flush */
187 if (list_empty(pending))
Ming Lei7c94e1c2014-09-25 23:23:43 +0800188 fq->flush_pending_since = jiffies;
Tejun Heoae1b1532011-01-25 12:43:54 +0100189 list_move_tail(&rq->flush.list, pending);
Tejun Heo28e7d182010-09-03 11:56:16 +0200190 break;
Tejun Heoae1b1532011-01-25 12:43:54 +0100191
192 case REQ_FSEQ_DATA:
Ming Lei7c94e1c2014-09-25 23:23:43 +0800193 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
Mike Snitzer10beafc2014-03-08 20:19:20 -0700194 queued = blk_flush_queue_rq(rq, true);
Tejun Heoae1b1532011-01-25 12:43:54 +0100195 break;
196
197 case REQ_FSEQ_DONE:
Tejun Heo09d60c72010-09-03 11:56:17 +0200198 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100199 * @rq was previously adjusted by blk_flush_issue() for
200 * flush sequencing and may already have gone through the
201 * flush data request completion path. Restore @rq for
202 * normal completion and end it.
Tejun Heo09d60c72010-09-03 11:56:17 +0200203 */
Tejun Heoae1b1532011-01-25 12:43:54 +0100204 BUG_ON(!list_empty(&rq->queuelist));
205 list_del_init(&rq->flush.list);
206 blk_flush_restore_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100207 if (q->mq_ops)
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700208 blk_mq_end_request(rq, error);
Jens Axboe320ae512013-10-24 09:20:05 +0100209 else
210 __blk_end_request_all(rq, error);
Tejun Heo28e7d182010-09-03 11:56:16 +0200211 break;
Tejun Heoae1b1532011-01-25 12:43:54 +0100212
Tejun Heo28e7d182010-09-03 11:56:16 +0200213 default:
214 BUG();
215 }
Christoph Hellwigcde4c402010-09-03 11:56:17 +0200216
Ming Lei0bae3522014-09-25 23:23:45 +0800217 kicked = blk_kick_flush(q, fq);
Jens Axboe320ae512013-10-24 09:20:05 +0100218 return kicked | queued;
Tejun Heo28e7d182010-09-03 11:56:16 +0200219}
220
Tejun Heoae1b1532011-01-25 12:43:54 +0100221static void flush_end_io(struct request *flush_rq, int error)
Jens Axboe86db1e22008-01-29 14:53:40 +0100222{
Tejun Heoae1b1532011-01-25 12:43:54 +0100223 struct request_queue *q = flush_rq->q;
Jens Axboe320ae512013-10-24 09:20:05 +0100224 struct list_head *running;
Tejun Heoae1b1532011-01-25 12:43:54 +0100225 bool queued = false;
226 struct request *rq, *n;
Jens Axboe320ae512013-10-24 09:20:05 +0100227 unsigned long flags = 0;
Ming Leie97c2932014-09-25 23:23:46 +0800228 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900229
Shaohua Li22302372014-05-30 08:06:42 -0600230 if (q->mq_ops) {
Ming Lei0048b482015-08-09 03:41:51 -0400231 struct blk_mq_hw_ctx *hctx;
232
233 /* release the tag's ownership to the req cloned from */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800234 spin_lock_irqsave(&fq->mq_flush_lock, flags);
Christoph Hellwig7d7e0f92016-09-14 16:18:54 +0200235 hctx = blk_mq_map_queue(q, flush_rq->mq_ctx->cpu);
Ming Lei0048b482015-08-09 03:41:51 -0400236 blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq);
Ming Lei7ddab5d2014-09-25 23:23:42 +0800237 flush_rq->tag = -1;
Shaohua Li22302372014-05-30 08:06:42 -0600238 }
Christoph Hellwig18741982014-02-10 09:29:00 -0700239
Ming Lei7c94e1c2014-09-25 23:23:43 +0800240 running = &fq->flush_queue[fq->flush_running_idx];
241 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
Tejun Heoae1b1532011-01-25 12:43:54 +0100242
243 /* account completion of the flush request */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800244 fq->flush_running_idx ^= 1;
Jens Axboe320ae512013-10-24 09:20:05 +0100245
246 if (!q->mq_ops)
247 elv_completed_request(q, flush_rq);
Tejun Heoae1b1532011-01-25 12:43:54 +0100248
249 /* and push the waiting requests to the next stage */
250 list_for_each_entry_safe(rq, n, running, flush.list) {
251 unsigned int seq = blk_flush_cur_seq(rq);
252
253 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
Ming Lei0bae3522014-09-25 23:23:45 +0800254 queued |= blk_flush_complete_seq(rq, fq, seq, error);
Tejun Heo28e7d182010-09-03 11:56:16 +0200255 }
256
Jens Axboe86db1e22008-01-29 14:53:40 +0100257 /*
shaohua.li@intel.com3ac0cc42011-05-06 11:34:41 -0600258 * Kick the queue to avoid stall for two cases:
259 * 1. Moving a request silently to empty queue_head may stall the
260 * queue.
261 * 2. When flush request is running in non-queueable queue, the
262 * queue is hold. Restart the queue after flush request is finished
263 * to avoid stall.
264 * This function is called from request completion path and calling
265 * directly into request_fn may confuse the driver. Always use
266 * kblockd.
Jens Axboe86db1e22008-01-29 14:53:40 +0100267 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800268 if (queued || fq->flush_queue_delayed) {
Christoph Hellwig18741982014-02-10 09:29:00 -0700269 WARN_ON(q->mq_ops);
270 blk_run_queue_async(q);
Jens Axboe320ae512013-10-24 09:20:05 +0100271 }
Ming Lei7c94e1c2014-09-25 23:23:43 +0800272 fq->flush_queue_delayed = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100273 if (q->mq_ops)
Ming Lei7c94e1c2014-09-25 23:23:43 +0800274 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
Jens Axboe320ae512013-10-24 09:20:05 +0100275}
276
Tejun Heoae1b1532011-01-25 12:43:54 +0100277/**
278 * blk_kick_flush - consider issuing flush request
279 * @q: request_queue being kicked
Ming Lei0bae3522014-09-25 23:23:45 +0800280 * @fq: flush queue
Tejun Heoae1b1532011-01-25 12:43:54 +0100281 *
282 * Flush related states of @q have changed, consider issuing flush request.
283 * Please read the comment at the top of this file for more info.
284 *
285 * CONTEXT:
Ming Lei7c94e1c2014-09-25 23:23:43 +0800286 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
Tejun Heoae1b1532011-01-25 12:43:54 +0100287 *
288 * RETURNS:
289 * %true if flush was issued, %false otherwise.
290 */
Ming Lei0bae3522014-09-25 23:23:45 +0800291static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq)
Tejun Heoae1b1532011-01-25 12:43:54 +0100292{
Ming Lei7c94e1c2014-09-25 23:23:43 +0800293 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
Tejun Heoae1b1532011-01-25 12:43:54 +0100294 struct request *first_rq =
295 list_first_entry(pending, struct request, flush.list);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800296 struct request *flush_rq = fq->flush_rq;
Tejun Heoae1b1532011-01-25 12:43:54 +0100297
298 /* C1 described at the top of this file */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800299 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
Tejun Heoae1b1532011-01-25 12:43:54 +0100300 return false;
301
302 /* C2 and C3 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800303 if (!list_empty(&fq->flush_data_in_flight) &&
Tejun Heoae1b1532011-01-25 12:43:54 +0100304 time_before(jiffies,
Ming Lei7c94e1c2014-09-25 23:23:43 +0800305 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
Tejun Heoae1b1532011-01-25 12:43:54 +0100306 return false;
Tejun Heo28e7d182010-09-03 11:56:16 +0200307
308 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100309 * Issue flush and toggle pending_idx. This makes pending_idx
310 * different from running_idx, which means flush is in flight.
Tejun Heo28e7d182010-09-03 11:56:16 +0200311 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800312 fq->flush_pending_idx ^= 1;
Christoph Hellwig18741982014-02-10 09:29:00 -0700313
Ming Lei7ddab5d2014-09-25 23:23:42 +0800314 blk_rq_init(q, flush_rq);
Ming Leif70ced02014-09-25 23:23:47 +0800315
316 /*
317 * Borrow tag from the first request since they can't
Ming Lei0048b482015-08-09 03:41:51 -0400318 * be in flight at the same time. And acquire the tag's
319 * ownership for flush req.
Ming Leif70ced02014-09-25 23:23:47 +0800320 */
321 if (q->mq_ops) {
Ming Lei0048b482015-08-09 03:41:51 -0400322 struct blk_mq_hw_ctx *hctx;
323
Ming Leif70ced02014-09-25 23:23:47 +0800324 flush_rq->mq_ctx = first_rq->mq_ctx;
325 flush_rq->tag = first_rq->tag;
Ming Lei0048b482015-08-09 03:41:51 -0400326 fq->orig_rq = first_rq;
327
Christoph Hellwig7d7e0f92016-09-14 16:18:54 +0200328 hctx = blk_mq_map_queue(q, first_rq->mq_ctx->cpu);
Ming Lei0048b482015-08-09 03:41:51 -0400329 blk_mq_tag_set_rq(hctx, first_rq->tag, flush_rq);
Ming Leif70ced02014-09-25 23:23:47 +0800330 }
Jens Axboe320ae512013-10-24 09:20:05 +0100331
Ming Lei7ddab5d2014-09-25 23:23:42 +0800332 flush_rq->cmd_type = REQ_TYPE_FS;
Mike Christie3a5e02c2016-06-05 14:32:23 -0500333 req_set_op_attrs(flush_rq, REQ_OP_FLUSH, WRITE_FLUSH | REQ_FLUSH_SEQ);
Ming Lei7ddab5d2014-09-25 23:23:42 +0800334 flush_rq->rq_disk = first_rq->rq_disk;
335 flush_rq->end_io = flush_end_io;
Jens Axboe86db1e22008-01-29 14:53:40 +0100336
Ming Lei7ddab5d2014-09-25 23:23:42 +0800337 return blk_flush_queue_rq(flush_rq, false);
Tejun Heoae1b1532011-01-25 12:43:54 +0100338}
339
340static void flush_data_end_io(struct request *rq, int error)
341{
342 struct request_queue *q = rq->q;
Ming Leie97c2932014-09-25 23:23:46 +0800343 struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL);
Tejun Heoae1b1532011-01-25 12:43:54 +0100344
Jens Axboe86db1e22008-01-29 14:53:40 +0100345 /*
Tejun Heoe83a46b2011-03-04 19:09:02 +0100346 * After populating an empty queue, kick it to avoid stall. Read
347 * the comment in flush_end_io().
Jens Axboe86db1e22008-01-29 14:53:40 +0100348 */
Ming Lei0bae3522014-09-25 23:23:45 +0800349 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +0200350 blk_run_queue_async(q);
Tejun Heoae1b1532011-01-25 12:43:54 +0100351}
352
Jens Axboe320ae512013-10-24 09:20:05 +0100353static void mq_flush_data_end_io(struct request *rq, int error)
354{
355 struct request_queue *q = rq->q;
356 struct blk_mq_hw_ctx *hctx;
Ming Leie97c2932014-09-25 23:23:46 +0800357 struct blk_mq_ctx *ctx = rq->mq_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +0100358 unsigned long flags;
Ming Leie97c2932014-09-25 23:23:46 +0800359 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +0100360
Christoph Hellwig7d7e0f92016-09-14 16:18:54 +0200361 hctx = blk_mq_map_queue(q, ctx->cpu);
Jens Axboe320ae512013-10-24 09:20:05 +0100362
363 /*
364 * After populating an empty queue, kick it to avoid stall. Read
365 * the comment in flush_end_io().
366 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800367 spin_lock_irqsave(&fq->mq_flush_lock, flags);
Ming Lei0bae3522014-09-25 23:23:45 +0800368 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
Jens Axboe320ae512013-10-24 09:20:05 +0100369 blk_mq_run_hw_queue(hctx, true);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800370 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
Jens Axboe320ae512013-10-24 09:20:05 +0100371}
372
Tejun Heoae1b1532011-01-25 12:43:54 +0100373/**
374 * blk_insert_flush - insert a new FLUSH/FUA request
375 * @rq: request to insert
376 *
Jens Axboeb710a482011-03-30 09:52:30 +0200377 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
Jens Axboe320ae512013-10-24 09:20:05 +0100378 * or __blk_mq_run_hw_queue() to dispatch request.
Tejun Heoae1b1532011-01-25 12:43:54 +0100379 * @rq is being submitted. Analyze what needs to be done and put it on the
380 * right queue.
381 *
382 * CONTEXT:
Jens Axboe320ae512013-10-24 09:20:05 +0100383 * spin_lock_irq(q->queue_lock) in !mq case
Tejun Heoae1b1532011-01-25 12:43:54 +0100384 */
385void blk_insert_flush(struct request *rq)
386{
387 struct request_queue *q = rq->q;
Jens Axboec888a8f2016-04-13 13:33:19 -0600388 unsigned long fflags = q->queue_flags; /* may change, cache */
Tejun Heoae1b1532011-01-25 12:43:54 +0100389 unsigned int policy = blk_flush_policy(fflags, rq);
Ming Leie97c2932014-09-25 23:23:46 +0800390 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
Tejun Heoae1b1532011-01-25 12:43:54 +0100391
Tejun Heoae1b1532011-01-25 12:43:54 +0100392 /*
393 * @policy now records what operations need to be done. Adjust
Mike Christie28a8f0d2016-06-05 14:32:25 -0500394 * REQ_PREFLUSH and FUA for the driver.
Tejun Heoae1b1532011-01-25 12:43:54 +0100395 */
Mike Christie28a8f0d2016-06-05 14:32:25 -0500396 rq->cmd_flags &= ~REQ_PREFLUSH;
Jens Axboec888a8f2016-04-13 13:33:19 -0600397 if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
Tejun Heo4fed9472010-09-03 11:56:17 +0200398 rq->cmd_flags &= ~REQ_FUA;
Jens Axboe86db1e22008-01-29 14:53:40 +0100399
Tejun Heoae1b1532011-01-25 12:43:54 +0100400 /*
Jeff Moyer4853aba2011-08-15 21:37:25 +0200401 * An empty flush handed down from a stacking driver may
402 * translate into nothing if the underlying device does not
403 * advertise a write-back cache. In this case, simply
404 * complete the request.
405 */
406 if (!policy) {
Jens Axboe320ae512013-10-24 09:20:05 +0100407 if (q->mq_ops)
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700408 blk_mq_end_request(rq, 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100409 else
410 __blk_end_bidi_request(rq, 0, 0, 0);
Jeff Moyer4853aba2011-08-15 21:37:25 +0200411 return;
412 }
413
Jeff Moyer834f9f62011-10-17 12:57:22 +0200414 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
Jeff Moyer4853aba2011-08-15 21:37:25 +0200415
416 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100417 * If there's data but flush is not necessary, the request can be
418 * processed directly without going through flush machinery. Queue
419 * for normal execution.
420 */
421 if ((policy & REQ_FSEQ_DATA) &&
422 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
Jens Axboe320ae512013-10-24 09:20:05 +0100423 if (q->mq_ops) {
Christoph Hellwigfeb71da2014-02-20 15:32:37 -0800424 blk_mq_insert_request(rq, false, false, true);
Jens Axboe320ae512013-10-24 09:20:05 +0100425 } else
Jens Axboedcd83762015-11-25 10:12:54 -0700426 list_add_tail(&rq->queuelist, &q->queue_head);
Tejun Heoae1b1532011-01-25 12:43:54 +0100427 return;
428 }
429
430 /*
431 * @rq should go through flush machinery. Mark it part of flush
432 * sequence and submit for further processing.
433 */
434 memset(&rq->flush, 0, sizeof(rq->flush));
435 INIT_LIST_HEAD(&rq->flush.list);
436 rq->cmd_flags |= REQ_FLUSH_SEQ;
Jeff Moyer4853aba2011-08-15 21:37:25 +0200437 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
Jens Axboe320ae512013-10-24 09:20:05 +0100438 if (q->mq_ops) {
439 rq->end_io = mq_flush_data_end_io;
440
Ming Lei7c94e1c2014-09-25 23:23:43 +0800441 spin_lock_irq(&fq->mq_flush_lock);
Ming Lei0bae3522014-09-25 23:23:45 +0800442 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800443 spin_unlock_irq(&fq->mq_flush_lock);
Jens Axboe320ae512013-10-24 09:20:05 +0100444 return;
445 }
Tejun Heoae1b1532011-01-25 12:43:54 +0100446 rq->end_io = flush_data_end_io;
447
Ming Lei0bae3522014-09-25 23:23:45 +0800448 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
Tejun Heoae1b1532011-01-25 12:43:54 +0100449}
450
451/**
Jens Axboe86db1e22008-01-29 14:53:40 +0100452 * blkdev_issue_flush - queue a flush
453 * @bdev: blockdev to issue flush for
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400454 * @gfp_mask: memory allocation flags (for bio_alloc)
Jens Axboe86db1e22008-01-29 14:53:40 +0100455 * @error_sector: error sector
456 *
457 * Description:
458 * Issue a flush for the block device in question. Caller can supply
459 * room for storing the error offset in case of a flush error, if they
Dmitry Monakhovf17e2322010-04-28 17:55:07 +0400460 * wish to. If WAIT flag is not passed then caller may check only what
461 * request was pushed in some internal queue for later handling.
Jens Axboe86db1e22008-01-29 14:53:40 +0100462 */
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400463int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
Christoph Hellwigdd3932e2010-09-16 20:51:46 +0200464 sector_t *error_sector)
Jens Axboe86db1e22008-01-29 14:53:40 +0100465{
Jens Axboe86db1e22008-01-29 14:53:40 +0100466 struct request_queue *q;
467 struct bio *bio;
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400468 int ret = 0;
Jens Axboe86db1e22008-01-29 14:53:40 +0100469
470 if (bdev->bd_disk == NULL)
471 return -ENXIO;
472
473 q = bdev_get_queue(bdev);
474 if (!q)
475 return -ENXIO;
476
Dave Chinnerf10d9f62010-07-13 17:50:50 +1000477 /*
478 * some block devices may not have their queue correctly set up here
479 * (e.g. loop device without a backing file) and so issuing a flush
480 * here will panic. Ensure there is a request function before issuing
Tejun Heod391a2d2010-09-03 11:56:17 +0200481 * the flush.
Dave Chinnerf10d9f62010-07-13 17:50:50 +1000482 */
483 if (!q->make_request_fn)
484 return -ENXIO;
485
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400486 bio = bio_alloc(gfp_mask, 0);
Jens Axboe86db1e22008-01-29 14:53:40 +0100487 bio->bi_bdev = bdev;
Mike Christie95fe6c12016-06-05 14:31:48 -0500488 bio_set_op_attrs(bio, REQ_OP_WRITE, WRITE_FLUSH);
Dmitry Monakhovf17e2322010-04-28 17:55:07 +0400489
Mike Christie4e49ea42016-06-05 14:31:41 -0500490 ret = submit_bio_wait(bio);
Christoph Hellwigdd3932e2010-09-16 20:51:46 +0200491
492 /*
493 * The driver must store the error location in ->bi_sector, if
494 * it supports it. For non-stacked drivers, this should be
495 * copied from blk_rq_pos(rq).
496 */
497 if (error_sector)
Kent Overstreet4f024f32013-10-11 15:44:27 -0700498 *error_sector = bio->bi_iter.bi_sector;
Jens Axboe86db1e22008-01-29 14:53:40 +0100499
Jens Axboe86db1e22008-01-29 14:53:40 +0100500 bio_put(bio);
501 return ret;
502}
Jens Axboe86db1e22008-01-29 14:53:40 +0100503EXPORT_SYMBOL(blkdev_issue_flush);
Jens Axboe320ae512013-10-24 09:20:05 +0100504
Ming Leif70ced02014-09-25 23:23:47 +0800505struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
506 int node, int cmd_size)
Jens Axboe320ae512013-10-24 09:20:05 +0100507{
Ming Lei7c94e1c2014-09-25 23:23:43 +0800508 struct blk_flush_queue *fq;
509 int rq_sz = sizeof(struct request);
Ming Lei1bcb1ea2014-09-25 23:23:39 +0800510
Ming Leif70ced02014-09-25 23:23:47 +0800511 fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800512 if (!fq)
513 goto fail;
Ming Lei1bcb1ea2014-09-25 23:23:39 +0800514
Ming Lei7c94e1c2014-09-25 23:23:43 +0800515 if (q->mq_ops) {
516 spin_lock_init(&fq->mq_flush_lock);
Ming Leif70ced02014-09-25 23:23:47 +0800517 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
Ming Lei7c94e1c2014-09-25 23:23:43 +0800518 }
519
Ming Leif70ced02014-09-25 23:23:47 +0800520 fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800521 if (!fq->flush_rq)
522 goto fail_rq;
523
524 INIT_LIST_HEAD(&fq->flush_queue[0]);
525 INIT_LIST_HEAD(&fq->flush_queue[1]);
526 INIT_LIST_HEAD(&fq->flush_data_in_flight);
527
528 return fq;
529
530 fail_rq:
531 kfree(fq);
532 fail:
533 return NULL;
534}
535
Ming Leiba483382014-09-25 23:23:44 +0800536void blk_free_flush_queue(struct blk_flush_queue *fq)
Ming Lei7c94e1c2014-09-25 23:23:43 +0800537{
538 /* bio based request queue hasn't flush queue */
539 if (!fq)
540 return;
541
542 kfree(fq->flush_rq);
543 kfree(fq);
Jens Axboe320ae512013-10-24 09:20:05 +0100544}