blob: f598f794c3c679e8c7904bb88284c96434123d66 [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 *
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
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 *
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.
25 *
26 * If the device has writeback cache and doesn't support FUA, REQ_FLUSH is
27 * translated to PREFLUSH and REQ_FUA to POSTFLUSH.
28 *
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
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"
Jens Axboe86db1e22008-01-29 14:53:40 +010076
Tejun Heo4fed9472010-09-03 11:56:17 +020077/* FLUSH/FUA sequences */
78enum {
Tejun Heoae1b1532011-01-25 12:43:54 +010079 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),
Tejun Heo4fed9472010-09-03 11:56:17 +020083
Tejun Heoae1b1532011-01-25 12:43:54 +010084 REQ_FSEQ_ACTIONS = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
85 REQ_FSEQ_POSTFLUSH,
Tejun Heo47f70d52010-09-03 11:56:17 +020086
87 /*
Tejun Heoae1b1532011-01-25 12:43:54 +010088 * If flush has been pending longer than the following timeout,
89 * it's issued even if flush_data requests are still in flight.
Tejun Heo47f70d52010-09-03 11:56:17 +020090 */
Tejun Heoae1b1532011-01-25 12:43:54 +010091 FLUSH_PENDING_TIMEOUT = 5 * HZ,
92};
93
94static bool blk_kick_flush(struct request_queue *q);
95
96static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
97{
98 unsigned int policy = 0;
99
Jeff Moyerfa1bf422011-08-09 20:32:09 +0200100 if (blk_rq_sectors(rq))
101 policy |= REQ_FSEQ_DATA;
102
Tejun Heoae1b1532011-01-25 12:43:54 +0100103 if (fflags & REQ_FLUSH) {
104 if (rq->cmd_flags & REQ_FLUSH)
105 policy |= REQ_FSEQ_PREFLUSH;
Tejun Heoae1b1532011-01-25 12:43:54 +0100106 if (!(fflags & REQ_FUA) && (rq->cmd_flags & REQ_FUA))
107 policy |= REQ_FSEQ_POSTFLUSH;
108 }
109 return policy;
Tejun Heo47f70d52010-09-03 11:56:17 +0200110}
111
Tejun Heoae1b1532011-01-25 12:43:54 +0100112static unsigned int blk_flush_cur_seq(struct request *rq)
Jens Axboe86db1e22008-01-29 14:53:40 +0100113{
Tejun Heoae1b1532011-01-25 12:43:54 +0100114 return 1 << ffz(rq->flush.seq);
Jens Axboe86db1e22008-01-29 14:53:40 +0100115}
116
Tejun Heoae1b1532011-01-25 12:43:54 +0100117static void blk_flush_restore_request(struct request *rq)
Jens Axboe86db1e22008-01-29 14:53:40 +0100118{
Tejun Heoae1b1532011-01-25 12:43:54 +0100119 /*
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.
123 */
124 rq->bio = rq->biotail;
125
126 /* make @rq a normal request */
127 rq->cmd_flags &= ~REQ_FLUSH_SEQ;
Jeff Moyer4853aba2011-08-15 21:37:25 +0200128 rq->end_io = rq->flush.saved_end_io;
Jens Axboe320ae512013-10-24 09:20:05 +0100129
130 blk_clear_rq_complete(rq);
131}
132
Christoph Hellwig18741982014-02-10 09:29:00 -0700133static void mq_flush_run(struct work_struct *work)
Jens Axboe320ae512013-10-24 09:20:05 +0100134{
135 struct request *rq;
136
Christoph Hellwig18741982014-02-10 09:29:00 -0700137 rq = container_of(work, struct request, mq_flush_work);
Jens Axboe320ae512013-10-24 09:20:05 +0100138
139 memset(&rq->csd, 0, sizeof(rq->csd));
Christoph Hellwigfeb71da2014-02-20 15:32:37 -0800140 blk_mq_insert_request(rq, false, true, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100141}
142
Christoph Hellwig18741982014-02-10 09:29:00 -0700143static bool blk_flush_queue_rq(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100144{
Christoph Hellwig18741982014-02-10 09:29:00 -0700145 if (rq->q->mq_ops) {
146 INIT_WORK(&rq->mq_flush_work, mq_flush_run);
147 kblockd_schedule_work(rq->q, &rq->mq_flush_work);
148 return false;
149 } else {
150 list_add_tail(&rq->queuelist, &rq->q->queue_head);
151 return true;
152 }
Jens Axboe86db1e22008-01-29 14:53:40 +0100153}
154
Tejun Heoae1b1532011-01-25 12:43:54 +0100155/**
156 * blk_flush_complete_seq - complete flush sequence
157 * @rq: FLUSH/FUA request being sequenced
158 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
159 * @error: whether an error occurred
160 *
161 * @rq just completed @seq part of its flush sequence, record the
162 * completion and trigger the next step.
163 *
164 * CONTEXT:
Jens Axboe320ae512013-10-24 09:20:05 +0100165 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
Tejun Heoae1b1532011-01-25 12:43:54 +0100166 *
167 * RETURNS:
168 * %true if requests were added to the dispatch queue, %false otherwise.
169 */
170static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
171 int error)
Jens Axboe86db1e22008-01-29 14:53:40 +0100172{
Tejun Heoae1b1532011-01-25 12:43:54 +0100173 struct request_queue *q = rq->q;
174 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
Jens Axboe320ae512013-10-24 09:20:05 +0100175 bool queued = false, kicked;
Jens Axboe86db1e22008-01-29 14:53:40 +0100176
Tejun Heoae1b1532011-01-25 12:43:54 +0100177 BUG_ON(rq->flush.seq & seq);
178 rq->flush.seq |= seq;
Jens Axboe86db1e22008-01-29 14:53:40 +0100179
Tejun Heoae1b1532011-01-25 12:43:54 +0100180 if (likely(!error))
181 seq = blk_flush_cur_seq(rq);
182 else
183 seq = REQ_FSEQ_DONE;
Tejun Heo28e7d182010-09-03 11:56:16 +0200184
Tejun Heoae1b1532011-01-25 12:43:54 +0100185 switch (seq) {
186 case REQ_FSEQ_PREFLUSH:
187 case REQ_FSEQ_POSTFLUSH:
188 /* queue for flush */
189 if (list_empty(pending))
190 q->flush_pending_since = jiffies;
191 list_move_tail(&rq->flush.list, pending);
Tejun Heo28e7d182010-09-03 11:56:16 +0200192 break;
Tejun Heoae1b1532011-01-25 12:43:54 +0100193
194 case REQ_FSEQ_DATA:
195 list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
Christoph Hellwig18741982014-02-10 09:29:00 -0700196 queued = blk_flush_queue_rq(rq);
Tejun Heoae1b1532011-01-25 12:43:54 +0100197 break;
198
199 case REQ_FSEQ_DONE:
Tejun Heo09d60c72010-09-03 11:56:17 +0200200 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100201 * @rq was previously adjusted by blk_flush_issue() for
202 * flush sequencing and may already have gone through the
203 * flush data request completion path. Restore @rq for
204 * normal completion and end it.
Tejun Heo09d60c72010-09-03 11:56:17 +0200205 */
Tejun Heoae1b1532011-01-25 12:43:54 +0100206 BUG_ON(!list_empty(&rq->queuelist));
207 list_del_init(&rq->flush.list);
208 blk_flush_restore_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100209 if (q->mq_ops)
210 blk_mq_end_io(rq, error);
211 else
212 __blk_end_request_all(rq, error);
Tejun Heo28e7d182010-09-03 11:56:16 +0200213 break;
Tejun Heoae1b1532011-01-25 12:43:54 +0100214
Tejun Heo28e7d182010-09-03 11:56:16 +0200215 default:
216 BUG();
217 }
Christoph Hellwigcde4c402010-09-03 11:56:17 +0200218
Jens Axboe320ae512013-10-24 09:20:05 +0100219 kicked = blk_kick_flush(q);
Jens Axboe320ae512013-10-24 09:20:05 +0100220 return kicked | queued;
Tejun Heo28e7d182010-09-03 11:56:16 +0200221}
222
Tejun Heoae1b1532011-01-25 12:43:54 +0100223static void flush_end_io(struct request *flush_rq, int error)
Jens Axboe86db1e22008-01-29 14:53:40 +0100224{
Tejun Heoae1b1532011-01-25 12:43:54 +0100225 struct request_queue *q = flush_rq->q;
Jens Axboe320ae512013-10-24 09:20:05 +0100226 struct list_head *running;
Tejun Heoae1b1532011-01-25 12:43:54 +0100227 bool queued = false;
228 struct request *rq, *n;
Jens Axboe320ae512013-10-24 09:20:05 +0100229 unsigned long flags = 0;
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900230
Christoph Hellwig18741982014-02-10 09:29:00 -0700231 if (q->mq_ops)
Jens Axboe320ae512013-10-24 09:20:05 +0100232 spin_lock_irqsave(&q->mq_flush_lock, flags);
Christoph Hellwig18741982014-02-10 09:29:00 -0700233
Jens Axboe320ae512013-10-24 09:20:05 +0100234 running = &q->flush_queue[q->flush_running_idx];
Tejun Heoae1b1532011-01-25 12:43:54 +0100235 BUG_ON(q->flush_pending_idx == q->flush_running_idx);
236
237 /* account completion of the flush request */
238 q->flush_running_idx ^= 1;
Jens Axboe320ae512013-10-24 09:20:05 +0100239
240 if (!q->mq_ops)
241 elv_completed_request(q, flush_rq);
Tejun Heoae1b1532011-01-25 12:43:54 +0100242
243 /* and push the waiting requests to the next stage */
244 list_for_each_entry_safe(rq, n, running, flush.list) {
245 unsigned int seq = blk_flush_cur_seq(rq);
246
247 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
248 queued |= blk_flush_complete_seq(rq, seq, error);
Tejun Heo28e7d182010-09-03 11:56:16 +0200249 }
250
Jens Axboe86db1e22008-01-29 14:53:40 +0100251 /*
shaohua.li@intel.com3ac0cc42011-05-06 11:34:41 -0600252 * Kick the queue to avoid stall for two cases:
253 * 1. Moving a request silently to empty queue_head may stall the
254 * queue.
255 * 2. When flush request is running in non-queueable queue, the
256 * queue is hold. Restart the queue after flush request is finished
257 * to avoid stall.
258 * This function is called from request completion path and calling
259 * directly into request_fn may confuse the driver. Always use
260 * kblockd.
Jens Axboe86db1e22008-01-29 14:53:40 +0100261 */
Jens Axboe320ae512013-10-24 09:20:05 +0100262 if (queued || q->flush_queue_delayed) {
Christoph Hellwig18741982014-02-10 09:29:00 -0700263 WARN_ON(q->mq_ops);
264 blk_run_queue_async(q);
Jens Axboe320ae512013-10-24 09:20:05 +0100265 }
shaohua.li@intel.com3ac0cc42011-05-06 11:34:41 -0600266 q->flush_queue_delayed = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100267 if (q->mq_ops)
268 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
269}
270
Tejun Heoae1b1532011-01-25 12:43:54 +0100271/**
272 * blk_kick_flush - consider issuing flush request
273 * @q: request_queue being kicked
274 *
275 * Flush related states of @q have changed, consider issuing flush request.
276 * Please read the comment at the top of this file for more info.
277 *
278 * CONTEXT:
Jens Axboe320ae512013-10-24 09:20:05 +0100279 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
Tejun Heoae1b1532011-01-25 12:43:54 +0100280 *
281 * RETURNS:
282 * %true if flush was issued, %false otherwise.
283 */
284static bool blk_kick_flush(struct request_queue *q)
285{
286 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
287 struct request *first_rq =
288 list_first_entry(pending, struct request, flush.list);
289
290 /* C1 described at the top of this file */
291 if (q->flush_pending_idx != q->flush_running_idx || list_empty(pending))
292 return false;
293
294 /* C2 and C3 */
295 if (!list_empty(&q->flush_data_in_flight) &&
296 time_before(jiffies,
297 q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
298 return false;
Tejun Heo28e7d182010-09-03 11:56:16 +0200299
300 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100301 * Issue flush and toggle pending_idx. This makes pending_idx
302 * different from running_idx, which means flush is in flight.
Tejun Heo28e7d182010-09-03 11:56:16 +0200303 */
Jens Axboe320ae512013-10-24 09:20:05 +0100304 q->flush_pending_idx ^= 1;
Christoph Hellwig18741982014-02-10 09:29:00 -0700305
Jens Axboe320ae512013-10-24 09:20:05 +0100306 if (q->mq_ops) {
Christoph Hellwig18741982014-02-10 09:29:00 -0700307 struct blk_mq_ctx *ctx = first_rq->mq_ctx;
308 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
309
310 blk_mq_rq_init(hctx, q->flush_rq);
311 q->flush_rq->mq_ctx = ctx;
312
313 /*
314 * Reuse the tag value from the fist waiting request,
315 * with blk-mq the tag is generated during request
316 * allocation and drivers can rely on it being inside
317 * the range they asked for.
318 */
319 q->flush_rq->tag = first_rq->tag;
320 } else {
321 blk_rq_init(q, q->flush_rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100322 }
323
Christoph Hellwig18741982014-02-10 09:29:00 -0700324 q->flush_rq->cmd_type = REQ_TYPE_FS;
325 q->flush_rq->cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
326 q->flush_rq->rq_disk = first_rq->rq_disk;
327 q->flush_rq->end_io = flush_end_io;
Jens Axboe86db1e22008-01-29 14:53:40 +0100328
Christoph Hellwig18741982014-02-10 09:29:00 -0700329 return blk_flush_queue_rq(q->flush_rq);
Tejun Heoae1b1532011-01-25 12:43:54 +0100330}
331
332static void flush_data_end_io(struct request *rq, int error)
333{
334 struct request_queue *q = rq->q;
Tejun Heoae1b1532011-01-25 12:43:54 +0100335
Jens Axboe86db1e22008-01-29 14:53:40 +0100336 /*
Tejun Heoe83a46b2011-03-04 19:09:02 +0100337 * After populating an empty queue, kick it to avoid stall. Read
338 * the comment in flush_end_io().
Jens Axboe86db1e22008-01-29 14:53:40 +0100339 */
Jens Axboe73c10102011-03-08 13:19:51 +0100340 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +0200341 blk_run_queue_async(q);
Tejun Heoae1b1532011-01-25 12:43:54 +0100342}
343
Jens Axboe320ae512013-10-24 09:20:05 +0100344static void mq_flush_data_end_io(struct request *rq, int error)
345{
346 struct request_queue *q = rq->q;
347 struct blk_mq_hw_ctx *hctx;
348 struct blk_mq_ctx *ctx;
349 unsigned long flags;
350
351 ctx = rq->mq_ctx;
352 hctx = q->mq_ops->map_queue(q, ctx->cpu);
353
354 /*
355 * After populating an empty queue, kick it to avoid stall. Read
356 * the comment in flush_end_io().
357 */
358 spin_lock_irqsave(&q->mq_flush_lock, flags);
359 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
360 blk_mq_run_hw_queue(hctx, true);
361 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
362}
363
Tejun Heoae1b1532011-01-25 12:43:54 +0100364/**
365 * blk_insert_flush - insert a new FLUSH/FUA request
366 * @rq: request to insert
367 *
Jens Axboeb710a482011-03-30 09:52:30 +0200368 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
Jens Axboe320ae512013-10-24 09:20:05 +0100369 * or __blk_mq_run_hw_queue() to dispatch request.
Tejun Heoae1b1532011-01-25 12:43:54 +0100370 * @rq is being submitted. Analyze what needs to be done and put it on the
371 * right queue.
372 *
373 * CONTEXT:
Jens Axboe320ae512013-10-24 09:20:05 +0100374 * spin_lock_irq(q->queue_lock) in !mq case
Tejun Heoae1b1532011-01-25 12:43:54 +0100375 */
376void blk_insert_flush(struct request *rq)
377{
378 struct request_queue *q = rq->q;
379 unsigned int fflags = q->flush_flags; /* may change, cache */
380 unsigned int policy = blk_flush_policy(fflags, rq);
381
Tejun Heoae1b1532011-01-25 12:43:54 +0100382 /*
383 * @policy now records what operations need to be done. Adjust
384 * REQ_FLUSH and FUA for the driver.
385 */
Christoph Hellwig18741982014-02-10 09:29:00 -0700386 rq->cmd_flags &= ~REQ_FLUSH;
Tejun Heoae1b1532011-01-25 12:43:54 +0100387 if (!(fflags & REQ_FUA))
Tejun Heo4fed9472010-09-03 11:56:17 +0200388 rq->cmd_flags &= ~REQ_FUA;
Jens Axboe86db1e22008-01-29 14:53:40 +0100389
Tejun Heoae1b1532011-01-25 12:43:54 +0100390 /*
Jeff Moyer4853aba2011-08-15 21:37:25 +0200391 * An empty flush handed down from a stacking driver may
392 * translate into nothing if the underlying device does not
393 * advertise a write-back cache. In this case, simply
394 * complete the request.
395 */
396 if (!policy) {
Jens Axboe320ae512013-10-24 09:20:05 +0100397 if (q->mq_ops)
398 blk_mq_end_io(rq, 0);
399 else
400 __blk_end_bidi_request(rq, 0, 0, 0);
Jeff Moyer4853aba2011-08-15 21:37:25 +0200401 return;
402 }
403
Jeff Moyer834f9f62011-10-17 12:57:22 +0200404 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
Jeff Moyer4853aba2011-08-15 21:37:25 +0200405
406 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100407 * If there's data but flush is not necessary, the request can be
408 * processed directly without going through flush machinery. Queue
409 * for normal execution.
410 */
411 if ((policy & REQ_FSEQ_DATA) &&
412 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
Jens Axboe320ae512013-10-24 09:20:05 +0100413 if (q->mq_ops) {
Christoph Hellwigfeb71da2014-02-20 15:32:37 -0800414 blk_mq_insert_request(rq, false, false, true);
Jens Axboe320ae512013-10-24 09:20:05 +0100415 } else
416 list_add_tail(&rq->queuelist, &q->queue_head);
Tejun Heoae1b1532011-01-25 12:43:54 +0100417 return;
418 }
419
420 /*
421 * @rq should go through flush machinery. Mark it part of flush
422 * sequence and submit for further processing.
423 */
424 memset(&rq->flush, 0, sizeof(rq->flush));
425 INIT_LIST_HEAD(&rq->flush.list);
426 rq->cmd_flags |= REQ_FLUSH_SEQ;
Jeff Moyer4853aba2011-08-15 21:37:25 +0200427 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
Jens Axboe320ae512013-10-24 09:20:05 +0100428 if (q->mq_ops) {
429 rq->end_io = mq_flush_data_end_io;
430
431 spin_lock_irq(&q->mq_flush_lock);
432 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
433 spin_unlock_irq(&q->mq_flush_lock);
434 return;
435 }
Tejun Heoae1b1532011-01-25 12:43:54 +0100436 rq->end_io = flush_data_end_io;
437
438 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
439}
440
441/**
442 * blk_abort_flushes - @q is being aborted, abort flush requests
443 * @q: request_queue being aborted
444 *
445 * To be called from elv_abort_queue(). @q is being aborted. Prepare all
446 * FLUSH/FUA requests for abortion.
447 *
448 * CONTEXT:
449 * spin_lock_irq(q->queue_lock)
450 */
451void blk_abort_flushes(struct request_queue *q)
452{
453 struct request *rq, *n;
454 int i;
455
456 /*
457 * Requests in flight for data are already owned by the dispatch
458 * queue or the device driver. Just restore for normal completion.
459 */
460 list_for_each_entry_safe(rq, n, &q->flush_data_in_flight, flush.list) {
461 list_del_init(&rq->flush.list);
462 blk_flush_restore_request(rq);
463 }
464
465 /*
466 * We need to give away requests on flush queues. Restore for
467 * normal completion and put them on the dispatch queue.
468 */
469 for (i = 0; i < ARRAY_SIZE(q->flush_queue); i++) {
470 list_for_each_entry_safe(rq, n, &q->flush_queue[i],
471 flush.list) {
472 list_del_init(&rq->flush.list);
473 blk_flush_restore_request(rq);
474 list_add_tail(&rq->queuelist, &q->queue_head);
475 }
476 }
Jens Axboe86db1e22008-01-29 14:53:40 +0100477}
478
Jens Axboe86db1e22008-01-29 14:53:40 +0100479/**
480 * blkdev_issue_flush - queue a flush
481 * @bdev: blockdev to issue flush for
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400482 * @gfp_mask: memory allocation flags (for bio_alloc)
Jens Axboe86db1e22008-01-29 14:53:40 +0100483 * @error_sector: error sector
484 *
485 * Description:
486 * Issue a flush for the block device in question. Caller can supply
487 * room for storing the error offset in case of a flush error, if they
Dmitry Monakhovf17e2322010-04-28 17:55:07 +0400488 * wish to. If WAIT flag is not passed then caller may check only what
489 * request was pushed in some internal queue for later handling.
Jens Axboe86db1e22008-01-29 14:53:40 +0100490 */
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400491int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
Christoph Hellwigdd3932e2010-09-16 20:51:46 +0200492 sector_t *error_sector)
Jens Axboe86db1e22008-01-29 14:53:40 +0100493{
Jens Axboe86db1e22008-01-29 14:53:40 +0100494 struct request_queue *q;
495 struct bio *bio;
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400496 int ret = 0;
Jens Axboe86db1e22008-01-29 14:53:40 +0100497
498 if (bdev->bd_disk == NULL)
499 return -ENXIO;
500
501 q = bdev_get_queue(bdev);
502 if (!q)
503 return -ENXIO;
504
Dave Chinnerf10d9f62010-07-13 17:50:50 +1000505 /*
506 * some block devices may not have their queue correctly set up here
507 * (e.g. loop device without a backing file) and so issuing a flush
508 * here will panic. Ensure there is a request function before issuing
Tejun Heod391a2d2010-09-03 11:56:17 +0200509 * the flush.
Dave Chinnerf10d9f62010-07-13 17:50:50 +1000510 */
511 if (!q->make_request_fn)
512 return -ENXIO;
513
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400514 bio = bio_alloc(gfp_mask, 0);
Jens Axboe86db1e22008-01-29 14:53:40 +0100515 bio->bi_bdev = bdev;
Dmitry Monakhovf17e2322010-04-28 17:55:07 +0400516
Kent Overstreet33879d42013-11-23 22:33:32 -0800517 ret = submit_bio_wait(WRITE_FLUSH, bio);
Christoph Hellwigdd3932e2010-09-16 20:51:46 +0200518
519 /*
520 * The driver must store the error location in ->bi_sector, if
521 * it supports it. For non-stacked drivers, this should be
522 * copied from blk_rq_pos(rq).
523 */
524 if (error_sector)
Kent Overstreet4f024f32013-10-11 15:44:27 -0700525 *error_sector = bio->bi_iter.bi_sector;
Jens Axboe86db1e22008-01-29 14:53:40 +0100526
Jens Axboe86db1e22008-01-29 14:53:40 +0100527 bio_put(bio);
528 return ret;
529}
Jens Axboe86db1e22008-01-29 14:53:40 +0100530EXPORT_SYMBOL(blkdev_issue_flush);
Jens Axboe320ae512013-10-24 09:20:05 +0100531
532void blk_mq_init_flush(struct request_queue *q)
533{
534 spin_lock_init(&q->mq_flush_lock);
Jens Axboe320ae512013-10-24 09:20:05 +0100535}