blob: 331e627301eaacbced43537469cb9f9914b6747c [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
133static void mq_flush_data_run(struct work_struct *work)
134{
135 struct request *rq;
136
137 rq = container_of(work, struct request, mq_flush_data);
138
139 memset(&rq->csd, 0, sizeof(rq->csd));
140 blk_mq_run_request(rq, true, false);
141}
142
143static void blk_mq_flush_data_insert(struct request *rq)
144{
145 INIT_WORK(&rq->mq_flush_data, mq_flush_data_run);
146 kblockd_schedule_work(rq->q, &rq->mq_flush_data);
Jens Axboe86db1e22008-01-29 14:53:40 +0100147}
148
Tejun Heoae1b1532011-01-25 12:43:54 +0100149/**
150 * blk_flush_complete_seq - complete flush sequence
151 * @rq: FLUSH/FUA request being sequenced
152 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
153 * @error: whether an error occurred
154 *
155 * @rq just completed @seq part of its flush sequence, record the
156 * completion and trigger the next step.
157 *
158 * CONTEXT:
Jens Axboe320ae512013-10-24 09:20:05 +0100159 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
Tejun Heoae1b1532011-01-25 12:43:54 +0100160 *
161 * RETURNS:
162 * %true if requests were added to the dispatch queue, %false otherwise.
163 */
164static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
165 int error)
Jens Axboe86db1e22008-01-29 14:53:40 +0100166{
Tejun Heoae1b1532011-01-25 12:43:54 +0100167 struct request_queue *q = rq->q;
168 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
Jens Axboe320ae512013-10-24 09:20:05 +0100169 bool queued = false, kicked;
Jens Axboe86db1e22008-01-29 14:53:40 +0100170
Tejun Heoae1b1532011-01-25 12:43:54 +0100171 BUG_ON(rq->flush.seq & seq);
172 rq->flush.seq |= seq;
Jens Axboe86db1e22008-01-29 14:53:40 +0100173
Tejun Heoae1b1532011-01-25 12:43:54 +0100174 if (likely(!error))
175 seq = blk_flush_cur_seq(rq);
176 else
177 seq = REQ_FSEQ_DONE;
Tejun Heo28e7d182010-09-03 11:56:16 +0200178
Tejun Heoae1b1532011-01-25 12:43:54 +0100179 switch (seq) {
180 case REQ_FSEQ_PREFLUSH:
181 case REQ_FSEQ_POSTFLUSH:
182 /* queue for flush */
183 if (list_empty(pending))
184 q->flush_pending_since = jiffies;
185 list_move_tail(&rq->flush.list, pending);
Tejun Heo28e7d182010-09-03 11:56:16 +0200186 break;
Tejun Heoae1b1532011-01-25 12:43:54 +0100187
188 case REQ_FSEQ_DATA:
189 list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
Jens Axboe320ae512013-10-24 09:20:05 +0100190 if (q->mq_ops)
191 blk_mq_flush_data_insert(rq);
192 else {
193 list_add(&rq->queuelist, &q->queue_head);
194 queued = true;
195 }
Tejun Heoae1b1532011-01-25 12:43:54 +0100196 break;
197
198 case REQ_FSEQ_DONE:
Tejun Heo09d60c72010-09-03 11:56:17 +0200199 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100200 * @rq was previously adjusted by blk_flush_issue() for
201 * flush sequencing and may already have gone through the
202 * flush data request completion path. Restore @rq for
203 * normal completion and end it.
Tejun Heo09d60c72010-09-03 11:56:17 +0200204 */
Tejun Heoae1b1532011-01-25 12:43:54 +0100205 BUG_ON(!list_empty(&rq->queuelist));
206 list_del_init(&rq->flush.list);
207 blk_flush_restore_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100208 if (q->mq_ops)
209 blk_mq_end_io(rq, error);
210 else
211 __blk_end_request_all(rq, error);
Tejun Heo28e7d182010-09-03 11:56:16 +0200212 break;
Tejun Heoae1b1532011-01-25 12:43:54 +0100213
Tejun Heo28e7d182010-09-03 11:56:16 +0200214 default:
215 BUG();
216 }
Christoph Hellwigcde4c402010-09-03 11:56:17 +0200217
Jens Axboe320ae512013-10-24 09:20:05 +0100218 kicked = blk_kick_flush(q);
219 /* blk_mq_run_flush will run queue */
220 if (q->mq_ops)
221 return queued;
222 return kicked | queued;
Tejun Heo28e7d182010-09-03 11:56:16 +0200223}
224
Tejun Heoae1b1532011-01-25 12:43:54 +0100225static void flush_end_io(struct request *flush_rq, int error)
Jens Axboe86db1e22008-01-29 14:53:40 +0100226{
Tejun Heoae1b1532011-01-25 12:43:54 +0100227 struct request_queue *q = flush_rq->q;
Jens Axboe320ae512013-10-24 09:20:05 +0100228 struct list_head *running;
Tejun Heoae1b1532011-01-25 12:43:54 +0100229 bool queued = false;
230 struct request *rq, *n;
Jens Axboe320ae512013-10-24 09:20:05 +0100231 unsigned long flags = 0;
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900232
Jens Axboe320ae512013-10-24 09:20:05 +0100233 if (q->mq_ops) {
234 blk_mq_free_request(flush_rq);
235 spin_lock_irqsave(&q->mq_flush_lock, flags);
236 }
237 running = &q->flush_queue[q->flush_running_idx];
Tejun Heoae1b1532011-01-25 12:43:54 +0100238 BUG_ON(q->flush_pending_idx == q->flush_running_idx);
239
240 /* account completion of the flush request */
241 q->flush_running_idx ^= 1;
Jens Axboe320ae512013-10-24 09:20:05 +0100242
243 if (!q->mq_ops)
244 elv_completed_request(q, flush_rq);
Tejun Heoae1b1532011-01-25 12:43:54 +0100245
246 /* and push the waiting requests to the next stage */
247 list_for_each_entry_safe(rq, n, running, flush.list) {
248 unsigned int seq = blk_flush_cur_seq(rq);
249
250 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
251 queued |= blk_flush_complete_seq(rq, seq, error);
Tejun Heo28e7d182010-09-03 11:56:16 +0200252 }
253
Jens Axboe86db1e22008-01-29 14:53:40 +0100254 /*
shaohua.li@intel.com3ac0cc42011-05-06 11:34:41 -0600255 * Kick the queue to avoid stall for two cases:
256 * 1. Moving a request silently to empty queue_head may stall the
257 * queue.
258 * 2. When flush request is running in non-queueable queue, the
259 * queue is hold. Restart the queue after flush request is finished
260 * to avoid stall.
261 * This function is called from request completion path and calling
262 * directly into request_fn may confuse the driver. Always use
263 * kblockd.
Jens Axboe86db1e22008-01-29 14:53:40 +0100264 */
Jens Axboe320ae512013-10-24 09:20:05 +0100265 if (queued || q->flush_queue_delayed) {
266 if (!q->mq_ops)
267 blk_run_queue_async(q);
268 else
269 /*
270 * This can be optimized to only run queues with requests
271 * queued if necessary.
272 */
273 blk_mq_run_queues(q, true);
274 }
shaohua.li@intel.com3ac0cc42011-05-06 11:34:41 -0600275 q->flush_queue_delayed = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100276 if (q->mq_ops)
277 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
278}
279
280static void mq_flush_work(struct work_struct *work)
281{
282 struct request_queue *q;
283 struct request *rq;
284
285 q = container_of(work, struct request_queue, mq_flush_work);
286
287 /* We don't need set REQ_FLUSH_SEQ, it's for consistency */
288 rq = blk_mq_alloc_request(q, WRITE_FLUSH|REQ_FLUSH_SEQ,
Christoph Hellwig3228f482013-10-28 13:33:58 -0600289 __GFP_WAIT|GFP_ATOMIC, true);
Jens Axboe320ae512013-10-24 09:20:05 +0100290 rq->cmd_type = REQ_TYPE_FS;
291 rq->end_io = flush_end_io;
292
293 blk_mq_run_request(rq, true, false);
294}
295
296/*
297 * We can't directly use q->flush_rq, because it doesn't have tag and is not in
298 * hctx->rqs[]. so we must allocate a new request, since we can't sleep here,
299 * so offload the work to workqueue.
300 *
301 * Note: we assume a flush request finished in any hardware queue will flush
302 * the whole disk cache.
303 */
304static void mq_run_flush(struct request_queue *q)
305{
306 kblockd_schedule_work(q, &q->mq_flush_work);
Tejun Heoae1b1532011-01-25 12:43:54 +0100307}
308
309/**
310 * blk_kick_flush - consider issuing flush request
311 * @q: request_queue being kicked
312 *
313 * Flush related states of @q have changed, consider issuing flush request.
314 * Please read the comment at the top of this file for more info.
315 *
316 * CONTEXT:
Jens Axboe320ae512013-10-24 09:20:05 +0100317 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
Tejun Heoae1b1532011-01-25 12:43:54 +0100318 *
319 * RETURNS:
320 * %true if flush was issued, %false otherwise.
321 */
322static bool blk_kick_flush(struct request_queue *q)
323{
324 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
325 struct request *first_rq =
326 list_first_entry(pending, struct request, flush.list);
327
328 /* C1 described at the top of this file */
329 if (q->flush_pending_idx != q->flush_running_idx || list_empty(pending))
330 return false;
331
332 /* C2 and C3 */
333 if (!list_empty(&q->flush_data_in_flight) &&
334 time_before(jiffies,
335 q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
336 return false;
Tejun Heo28e7d182010-09-03 11:56:16 +0200337
338 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100339 * Issue flush and toggle pending_idx. This makes pending_idx
340 * different from running_idx, which means flush is in flight.
Tejun Heo28e7d182010-09-03 11:56:16 +0200341 */
Jens Axboe320ae512013-10-24 09:20:05 +0100342 q->flush_pending_idx ^= 1;
343 if (q->mq_ops) {
344 mq_run_flush(q);
345 return true;
346 }
347
Tejun Heoae1b1532011-01-25 12:43:54 +0100348 blk_rq_init(q, &q->flush_rq);
349 q->flush_rq.cmd_type = REQ_TYPE_FS;
350 q->flush_rq.cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
351 q->flush_rq.rq_disk = first_rq->rq_disk;
352 q->flush_rq.end_io = flush_end_io;
Jens Axboe86db1e22008-01-29 14:53:40 +0100353
Jens Axboe53d63e62011-03-30 13:27:09 +0200354 list_add_tail(&q->flush_rq.queuelist, &q->queue_head);
Tejun Heoae1b1532011-01-25 12:43:54 +0100355 return true;
356}
357
358static void flush_data_end_io(struct request *rq, int error)
359{
360 struct request_queue *q = rq->q;
Tejun Heoae1b1532011-01-25 12:43:54 +0100361
Jens Axboe86db1e22008-01-29 14:53:40 +0100362 /*
Tejun Heoe83a46b2011-03-04 19:09:02 +0100363 * After populating an empty queue, kick it to avoid stall. Read
364 * the comment in flush_end_io().
Jens Axboe86db1e22008-01-29 14:53:40 +0100365 */
Jens Axboe73c10102011-03-08 13:19:51 +0100366 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +0200367 blk_run_queue_async(q);
Tejun Heoae1b1532011-01-25 12:43:54 +0100368}
369
Jens Axboe320ae512013-10-24 09:20:05 +0100370static void mq_flush_data_end_io(struct request *rq, int error)
371{
372 struct request_queue *q = rq->q;
373 struct blk_mq_hw_ctx *hctx;
374 struct blk_mq_ctx *ctx;
375 unsigned long flags;
376
377 ctx = rq->mq_ctx;
378 hctx = q->mq_ops->map_queue(q, ctx->cpu);
379
380 /*
381 * After populating an empty queue, kick it to avoid stall. Read
382 * the comment in flush_end_io().
383 */
384 spin_lock_irqsave(&q->mq_flush_lock, flags);
385 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
386 blk_mq_run_hw_queue(hctx, true);
387 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
388}
389
Tejun Heoae1b1532011-01-25 12:43:54 +0100390/**
391 * blk_insert_flush - insert a new FLUSH/FUA request
392 * @rq: request to insert
393 *
Jens Axboeb710a482011-03-30 09:52:30 +0200394 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
Jens Axboe320ae512013-10-24 09:20:05 +0100395 * or __blk_mq_run_hw_queue() to dispatch request.
Tejun Heoae1b1532011-01-25 12:43:54 +0100396 * @rq is being submitted. Analyze what needs to be done and put it on the
397 * right queue.
398 *
399 * CONTEXT:
Jens Axboe320ae512013-10-24 09:20:05 +0100400 * spin_lock_irq(q->queue_lock) in !mq case
Tejun Heoae1b1532011-01-25 12:43:54 +0100401 */
402void blk_insert_flush(struct request *rq)
403{
404 struct request_queue *q = rq->q;
405 unsigned int fflags = q->flush_flags; /* may change, cache */
406 unsigned int policy = blk_flush_policy(fflags, rq);
407
Tejun Heoae1b1532011-01-25 12:43:54 +0100408 /*
409 * @policy now records what operations need to be done. Adjust
410 * REQ_FLUSH and FUA for the driver.
411 */
Tejun Heo4fed9472010-09-03 11:56:17 +0200412 rq->cmd_flags &= ~REQ_FLUSH;
Tejun Heoae1b1532011-01-25 12:43:54 +0100413 if (!(fflags & REQ_FUA))
Tejun Heo4fed9472010-09-03 11:56:17 +0200414 rq->cmd_flags &= ~REQ_FUA;
Jens Axboe86db1e22008-01-29 14:53:40 +0100415
Tejun Heoae1b1532011-01-25 12:43:54 +0100416 /*
Jeff Moyer4853aba2011-08-15 21:37:25 +0200417 * An empty flush handed down from a stacking driver may
418 * translate into nothing if the underlying device does not
419 * advertise a write-back cache. In this case, simply
420 * complete the request.
421 */
422 if (!policy) {
Jens Axboe320ae512013-10-24 09:20:05 +0100423 if (q->mq_ops)
424 blk_mq_end_io(rq, 0);
425 else
426 __blk_end_bidi_request(rq, 0, 0, 0);
Jeff Moyer4853aba2011-08-15 21:37:25 +0200427 return;
428 }
429
Jeff Moyer834f9f62011-10-17 12:57:22 +0200430 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
Jeff Moyer4853aba2011-08-15 21:37:25 +0200431
432 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100433 * If there's data but flush is not necessary, the request can be
434 * processed directly without going through flush machinery. Queue
435 * for normal execution.
436 */
437 if ((policy & REQ_FSEQ_DATA) &&
438 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
Jens Axboe320ae512013-10-24 09:20:05 +0100439 if (q->mq_ops) {
440 blk_mq_run_request(rq, false, true);
441 } else
442 list_add_tail(&rq->queuelist, &q->queue_head);
Tejun Heoae1b1532011-01-25 12:43:54 +0100443 return;
444 }
445
446 /*
447 * @rq should go through flush machinery. Mark it part of flush
448 * sequence and submit for further processing.
449 */
450 memset(&rq->flush, 0, sizeof(rq->flush));
451 INIT_LIST_HEAD(&rq->flush.list);
452 rq->cmd_flags |= REQ_FLUSH_SEQ;
Jeff Moyer4853aba2011-08-15 21:37:25 +0200453 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
Jens Axboe320ae512013-10-24 09:20:05 +0100454 if (q->mq_ops) {
455 rq->end_io = mq_flush_data_end_io;
456
457 spin_lock_irq(&q->mq_flush_lock);
458 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
459 spin_unlock_irq(&q->mq_flush_lock);
460 return;
461 }
Tejun Heoae1b1532011-01-25 12:43:54 +0100462 rq->end_io = flush_data_end_io;
463
464 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
465}
466
467/**
468 * blk_abort_flushes - @q is being aborted, abort flush requests
469 * @q: request_queue being aborted
470 *
471 * To be called from elv_abort_queue(). @q is being aborted. Prepare all
472 * FLUSH/FUA requests for abortion.
473 *
474 * CONTEXT:
475 * spin_lock_irq(q->queue_lock)
476 */
477void blk_abort_flushes(struct request_queue *q)
478{
479 struct request *rq, *n;
480 int i;
481
482 /*
483 * Requests in flight for data are already owned by the dispatch
484 * queue or the device driver. Just restore for normal completion.
485 */
486 list_for_each_entry_safe(rq, n, &q->flush_data_in_flight, flush.list) {
487 list_del_init(&rq->flush.list);
488 blk_flush_restore_request(rq);
489 }
490
491 /*
492 * We need to give away requests on flush queues. Restore for
493 * normal completion and put them on the dispatch queue.
494 */
495 for (i = 0; i < ARRAY_SIZE(q->flush_queue); i++) {
496 list_for_each_entry_safe(rq, n, &q->flush_queue[i],
497 flush.list) {
498 list_del_init(&rq->flush.list);
499 blk_flush_restore_request(rq);
500 list_add_tail(&rq->queuelist, &q->queue_head);
501 }
502 }
Jens Axboe86db1e22008-01-29 14:53:40 +0100503}
504
Tejun Heod391a2d2010-09-03 11:56:17 +0200505static void bio_end_flush(struct bio *bio, int err)
Jens Axboe86db1e22008-01-29 14:53:40 +0100506{
Tejun Heod391a2d2010-09-03 11:56:17 +0200507 if (err)
Jens Axboe86db1e22008-01-29 14:53:40 +0100508 clear_bit(BIO_UPTODATE, &bio->bi_flags);
Dmitry Monakhovf17e2322010-04-28 17:55:07 +0400509 if (bio->bi_private)
510 complete(bio->bi_private);
511 bio_put(bio);
Jens Axboe86db1e22008-01-29 14:53:40 +0100512}
513
514/**
515 * blkdev_issue_flush - queue a flush
516 * @bdev: blockdev to issue flush for
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400517 * @gfp_mask: memory allocation flags (for bio_alloc)
Jens Axboe86db1e22008-01-29 14:53:40 +0100518 * @error_sector: error sector
519 *
520 * Description:
521 * Issue a flush for the block device in question. Caller can supply
522 * room for storing the error offset in case of a flush error, if they
Dmitry Monakhovf17e2322010-04-28 17:55:07 +0400523 * wish to. If WAIT flag is not passed then caller may check only what
524 * request was pushed in some internal queue for later handling.
Jens Axboe86db1e22008-01-29 14:53:40 +0100525 */
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400526int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
Christoph Hellwigdd3932e2010-09-16 20:51:46 +0200527 sector_t *error_sector)
Jens Axboe86db1e22008-01-29 14:53:40 +0100528{
529 DECLARE_COMPLETION_ONSTACK(wait);
530 struct request_queue *q;
531 struct bio *bio;
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400532 int ret = 0;
Jens Axboe86db1e22008-01-29 14:53:40 +0100533
534 if (bdev->bd_disk == NULL)
535 return -ENXIO;
536
537 q = bdev_get_queue(bdev);
538 if (!q)
539 return -ENXIO;
540
Dave Chinnerf10d9f62010-07-13 17:50:50 +1000541 /*
542 * some block devices may not have their queue correctly set up here
543 * (e.g. loop device without a backing file) and so issuing a flush
544 * here will panic. Ensure there is a request function before issuing
Tejun Heod391a2d2010-09-03 11:56:17 +0200545 * the flush.
Dave Chinnerf10d9f62010-07-13 17:50:50 +1000546 */
547 if (!q->make_request_fn)
548 return -ENXIO;
549
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400550 bio = bio_alloc(gfp_mask, 0);
Tejun Heod391a2d2010-09-03 11:56:17 +0200551 bio->bi_end_io = bio_end_flush;
Jens Axboe86db1e22008-01-29 14:53:40 +0100552 bio->bi_bdev = bdev;
Christoph Hellwigdd3932e2010-09-16 20:51:46 +0200553 bio->bi_private = &wait;
Dmitry Monakhovf17e2322010-04-28 17:55:07 +0400554
555 bio_get(bio);
Tejun Heod391a2d2010-09-03 11:56:17 +0200556 submit_bio(WRITE_FLUSH, bio);
Vladimir Davydov55770222013-02-14 18:19:59 +0400557 wait_for_completion_io(&wait);
Christoph Hellwigdd3932e2010-09-16 20:51:46 +0200558
559 /*
560 * The driver must store the error location in ->bi_sector, if
561 * it supports it. For non-stacked drivers, this should be
562 * copied from blk_rq_pos(rq).
563 */
564 if (error_sector)
Alice Ferrazzif2fc7d02013-03-22 11:11:04 -0600565 *error_sector = bio->bi_sector;
Jens Axboe86db1e22008-01-29 14:53:40 +0100566
Tejun Heod391a2d2010-09-03 11:56:17 +0200567 if (!bio_flagged(bio, BIO_UPTODATE))
Jens Axboe86db1e22008-01-29 14:53:40 +0100568 ret = -EIO;
569
570 bio_put(bio);
571 return ret;
572}
Jens Axboe86db1e22008-01-29 14:53:40 +0100573EXPORT_SYMBOL(blkdev_issue_flush);
Jens Axboe320ae512013-10-24 09:20:05 +0100574
575void blk_mq_init_flush(struct request_queue *q)
576{
577 spin_lock_init(&q->mq_flush_lock);
578 INIT_WORK(&q->mq_flush_work, mq_flush_work);
579}