blob: 9c423e53324a29c1ba1213c6009822b15d75dac7 [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
Ming Lei7c94e1c2014-09-25 23:23:43 +080031 * fq->flush_queue[fq->flush_pending_idx]. Once certain criteria are met, a
Tejun Heoae1b1532011-01-25 12:43:54 +010032 * 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"
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
98static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
99{
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
Tejun Heoae1b1532011-01-25 12:43:54 +0100105 if (fflags & REQ_FLUSH) {
106 if (rq->cmd_flags & REQ_FLUSH)
107 policy |= REQ_FSEQ_PREFLUSH;
Tejun Heoae1b1532011-01-25 12:43:54 +0100108 if (!(fflags & REQ_FUA) && (rq->cmd_flags & REQ_FUA))
109 policy |= REQ_FSEQ_POSTFLUSH;
110 }
111 return policy;
Tejun Heo47f70d52010-09-03 11:56:17 +0200112}
113
Tejun Heoae1b1532011-01-25 12:43:54 +0100114static unsigned int blk_flush_cur_seq(struct request *rq)
Jens Axboe86db1e22008-01-29 14:53:40 +0100115{
Tejun Heoae1b1532011-01-25 12:43:54 +0100116 return 1 << ffz(rq->flush.seq);
Jens Axboe86db1e22008-01-29 14:53:40 +0100117}
118
Tejun Heoae1b1532011-01-25 12:43:54 +0100119static void blk_flush_restore_request(struct request *rq)
Jens Axboe86db1e22008-01-29 14:53:40 +0100120{
Tejun Heoae1b1532011-01-25 12:43:54 +0100121 /*
122 * After flush data completion, @rq->bio is %NULL but we need to
123 * complete the bio again. @rq->biotail is guaranteed to equal the
124 * original @rq->bio. Restore it.
125 */
126 rq->bio = rq->biotail;
127
128 /* make @rq a normal request */
129 rq->cmd_flags &= ~REQ_FLUSH_SEQ;
Jeff Moyer4853aba2011-08-15 21:37:25 +0200130 rq->end_io = rq->flush.saved_end_io;
Jens Axboe320ae512013-10-24 09:20:05 +0100131}
132
Mike Snitzer10beafc2014-03-08 20:19:20 -0700133static bool blk_flush_queue_rq(struct request *rq, bool add_front)
Jens Axboe320ae512013-10-24 09:20:05 +0100134{
Christoph Hellwig18741982014-02-10 09:29:00 -0700135 if (rq->q->mq_ops) {
Christoph Hellwig6fca6a62014-05-28 08:08:02 -0600136 struct request_queue *q = rq->q;
137
138 blk_mq_add_to_requeue_list(rq, add_front);
139 blk_mq_kick_requeue_list(q);
Christoph Hellwig18741982014-02-10 09:29:00 -0700140 return false;
141 } else {
Mike Snitzer10beafc2014-03-08 20:19:20 -0700142 if (add_front)
143 list_add(&rq->queuelist, &rq->q->queue_head);
144 else
145 list_add_tail(&rq->queuelist, &rq->q->queue_head);
Christoph Hellwig18741982014-02-10 09:29:00 -0700146 return true;
147 }
Jens Axboe86db1e22008-01-29 14:53:40 +0100148}
149
Tejun Heoae1b1532011-01-25 12:43:54 +0100150/**
151 * blk_flush_complete_seq - complete flush sequence
152 * @rq: FLUSH/FUA request being sequenced
Ming Lei0bae3522014-09-25 23:23:45 +0800153 * @fq: flush queue
Tejun Heoae1b1532011-01-25 12:43:54 +0100154 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
155 * @error: whether an error occurred
156 *
157 * @rq just completed @seq part of its flush sequence, record the
158 * completion and trigger the next step.
159 *
160 * CONTEXT:
Ming Lei7c94e1c2014-09-25 23:23:43 +0800161 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
Tejun Heoae1b1532011-01-25 12:43:54 +0100162 *
163 * RETURNS:
164 * %true if requests were added to the dispatch queue, %false otherwise.
165 */
Ming Lei0bae3522014-09-25 23:23:45 +0800166static bool blk_flush_complete_seq(struct request *rq,
167 struct blk_flush_queue *fq,
168 unsigned int seq, int error)
Jens Axboe86db1e22008-01-29 14:53:40 +0100169{
Tejun Heoae1b1532011-01-25 12:43:54 +0100170 struct request_queue *q = rq->q;
Ming Lei7c94e1c2014-09-25 23:23:43 +0800171 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
Jens Axboe320ae512013-10-24 09:20:05 +0100172 bool queued = false, kicked;
Jens Axboe86db1e22008-01-29 14:53:40 +0100173
Tejun Heoae1b1532011-01-25 12:43:54 +0100174 BUG_ON(rq->flush.seq & seq);
175 rq->flush.seq |= seq;
Jens Axboe86db1e22008-01-29 14:53:40 +0100176
Tejun Heoae1b1532011-01-25 12:43:54 +0100177 if (likely(!error))
178 seq = blk_flush_cur_seq(rq);
179 else
180 seq = REQ_FSEQ_DONE;
Tejun Heo28e7d182010-09-03 11:56:16 +0200181
Tejun Heoae1b1532011-01-25 12:43:54 +0100182 switch (seq) {
183 case REQ_FSEQ_PREFLUSH:
184 case REQ_FSEQ_POSTFLUSH:
185 /* queue for flush */
186 if (list_empty(pending))
Ming Lei7c94e1c2014-09-25 23:23:43 +0800187 fq->flush_pending_since = jiffies;
Tejun Heoae1b1532011-01-25 12:43:54 +0100188 list_move_tail(&rq->flush.list, pending);
Tejun Heo28e7d182010-09-03 11:56:16 +0200189 break;
Tejun Heoae1b1532011-01-25 12:43:54 +0100190
191 case REQ_FSEQ_DATA:
Ming Lei7c94e1c2014-09-25 23:23:43 +0800192 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
Mike Snitzer10beafc2014-03-08 20:19:20 -0700193 queued = blk_flush_queue_rq(rq, true);
Tejun Heoae1b1532011-01-25 12:43:54 +0100194 break;
195
196 case REQ_FSEQ_DONE:
Tejun Heo09d60c72010-09-03 11:56:17 +0200197 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100198 * @rq was previously adjusted by blk_flush_issue() for
199 * flush sequencing and may already have gone through the
200 * flush data request completion path. Restore @rq for
201 * normal completion and end it.
Tejun Heo09d60c72010-09-03 11:56:17 +0200202 */
Tejun Heoae1b1532011-01-25 12:43:54 +0100203 BUG_ON(!list_empty(&rq->queuelist));
204 list_del_init(&rq->flush.list);
205 blk_flush_restore_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100206 if (q->mq_ops)
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700207 blk_mq_end_request(rq, error);
Jens Axboe320ae512013-10-24 09:20:05 +0100208 else
209 __blk_end_request_all(rq, error);
Tejun Heo28e7d182010-09-03 11:56:16 +0200210 break;
Tejun Heoae1b1532011-01-25 12:43:54 +0100211
Tejun Heo28e7d182010-09-03 11:56:16 +0200212 default:
213 BUG();
214 }
Christoph Hellwigcde4c402010-09-03 11:56:17 +0200215
Ming Lei0bae3522014-09-25 23:23:45 +0800216 kicked = blk_kick_flush(q, fq);
Jens Axboe320ae512013-10-24 09:20:05 +0100217 return kicked | queued;
Tejun Heo28e7d182010-09-03 11:56:16 +0200218}
219
Tejun Heoae1b1532011-01-25 12:43:54 +0100220static void flush_end_io(struct request *flush_rq, int error)
Jens Axboe86db1e22008-01-29 14:53:40 +0100221{
Tejun Heoae1b1532011-01-25 12:43:54 +0100222 struct request_queue *q = flush_rq->q;
Jens Axboe320ae512013-10-24 09:20:05 +0100223 struct list_head *running;
Tejun Heoae1b1532011-01-25 12:43:54 +0100224 bool queued = false;
225 struct request *rq, *n;
Jens Axboe320ae512013-10-24 09:20:05 +0100226 unsigned long flags = 0;
Ming Leie97c2932014-09-25 23:23:46 +0800227 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900228
Shaohua Li22302372014-05-30 08:06:42 -0600229 if (q->mq_ops) {
Ming Lei0048b482015-08-09 03:41:51 -0400230 struct blk_mq_hw_ctx *hctx;
231
232 /* release the tag's ownership to the req cloned from */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800233 spin_lock_irqsave(&fq->mq_flush_lock, flags);
Ming Lei0048b482015-08-09 03:41:51 -0400234 hctx = q->mq_ops->map_queue(q, flush_rq->mq_ctx->cpu);
235 blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq);
Ming Lei7ddab5d2014-09-25 23:23:42 +0800236 flush_rq->tag = -1;
Shaohua Li22302372014-05-30 08:06:42 -0600237 }
Christoph Hellwig18741982014-02-10 09:29:00 -0700238
Ming Lei7c94e1c2014-09-25 23:23:43 +0800239 running = &fq->flush_queue[fq->flush_running_idx];
240 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
Tejun Heoae1b1532011-01-25 12:43:54 +0100241
242 /* account completion of the flush request */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800243 fq->flush_running_idx ^= 1;
Jens Axboe320ae512013-10-24 09:20:05 +0100244
245 if (!q->mq_ops)
246 elv_completed_request(q, flush_rq);
Tejun Heoae1b1532011-01-25 12:43:54 +0100247
248 /* and push the waiting requests to the next stage */
249 list_for_each_entry_safe(rq, n, running, flush.list) {
250 unsigned int seq = blk_flush_cur_seq(rq);
251
252 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
Ming Lei0bae3522014-09-25 23:23:45 +0800253 queued |= blk_flush_complete_seq(rq, fq, seq, error);
Tejun Heo28e7d182010-09-03 11:56:16 +0200254 }
255
Jens Axboe86db1e22008-01-29 14:53:40 +0100256 /*
shaohua.li@intel.com3ac0cc42011-05-06 11:34:41 -0600257 * Kick the queue to avoid stall for two cases:
258 * 1. Moving a request silently to empty queue_head may stall the
259 * queue.
260 * 2. When flush request is running in non-queueable queue, the
261 * queue is hold. Restart the queue after flush request is finished
262 * to avoid stall.
263 * This function is called from request completion path and calling
264 * directly into request_fn may confuse the driver. Always use
265 * kblockd.
Jens Axboe86db1e22008-01-29 14:53:40 +0100266 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800267 if (queued || fq->flush_queue_delayed) {
Christoph Hellwig18741982014-02-10 09:29:00 -0700268 WARN_ON(q->mq_ops);
269 blk_run_queue_async(q);
Jens Axboe320ae512013-10-24 09:20:05 +0100270 }
Ming Lei7c94e1c2014-09-25 23:23:43 +0800271 fq->flush_queue_delayed = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100272 if (q->mq_ops)
Ming Lei7c94e1c2014-09-25 23:23:43 +0800273 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
Jens Axboe320ae512013-10-24 09:20:05 +0100274}
275
Tejun Heoae1b1532011-01-25 12:43:54 +0100276/**
277 * blk_kick_flush - consider issuing flush request
278 * @q: request_queue being kicked
Ming Lei0bae3522014-09-25 23:23:45 +0800279 * @fq: flush queue
Tejun Heoae1b1532011-01-25 12:43:54 +0100280 *
281 * Flush related states of @q have changed, consider issuing flush request.
282 * Please read the comment at the top of this file for more info.
283 *
284 * CONTEXT:
Ming Lei7c94e1c2014-09-25 23:23:43 +0800285 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
Tejun Heoae1b1532011-01-25 12:43:54 +0100286 *
287 * RETURNS:
288 * %true if flush was issued, %false otherwise.
289 */
Ming Lei0bae3522014-09-25 23:23:45 +0800290static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq)
Tejun Heoae1b1532011-01-25 12:43:54 +0100291{
Ming Lei7c94e1c2014-09-25 23:23:43 +0800292 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
Tejun Heoae1b1532011-01-25 12:43:54 +0100293 struct request *first_rq =
294 list_first_entry(pending, struct request, flush.list);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800295 struct request *flush_rq = fq->flush_rq;
Tejun Heoae1b1532011-01-25 12:43:54 +0100296
297 /* C1 described at the top of this file */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800298 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
Tejun Heoae1b1532011-01-25 12:43:54 +0100299 return false;
300
301 /* C2 and C3 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800302 if (!list_empty(&fq->flush_data_in_flight) &&
Tejun Heoae1b1532011-01-25 12:43:54 +0100303 time_before(jiffies,
Ming Lei7c94e1c2014-09-25 23:23:43 +0800304 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
Tejun Heoae1b1532011-01-25 12:43:54 +0100305 return false;
Tejun Heo28e7d182010-09-03 11:56:16 +0200306
307 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100308 * Issue flush and toggle pending_idx. This makes pending_idx
309 * different from running_idx, which means flush is in flight.
Tejun Heo28e7d182010-09-03 11:56:16 +0200310 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800311 fq->flush_pending_idx ^= 1;
Christoph Hellwig18741982014-02-10 09:29:00 -0700312
Ming Lei7ddab5d2014-09-25 23:23:42 +0800313 blk_rq_init(q, flush_rq);
Ming Leif70ced02014-09-25 23:23:47 +0800314
315 /*
316 * Borrow tag from the first request since they can't
Ming Lei0048b482015-08-09 03:41:51 -0400317 * be in flight at the same time. And acquire the tag's
318 * ownership for flush req.
Ming Leif70ced02014-09-25 23:23:47 +0800319 */
320 if (q->mq_ops) {
Ming Lei0048b482015-08-09 03:41:51 -0400321 struct blk_mq_hw_ctx *hctx;
322
Ming Leif70ced02014-09-25 23:23:47 +0800323 flush_rq->mq_ctx = first_rq->mq_ctx;
324 flush_rq->tag = first_rq->tag;
Ming Lei0048b482015-08-09 03:41:51 -0400325 fq->orig_rq = first_rq;
326
327 hctx = q->mq_ops->map_queue(q, first_rq->mq_ctx->cpu);
328 blk_mq_tag_set_rq(hctx, first_rq->tag, flush_rq);
Ming Leif70ced02014-09-25 23:23:47 +0800329 }
Jens Axboe320ae512013-10-24 09:20:05 +0100330
Ming Lei7ddab5d2014-09-25 23:23:42 +0800331 flush_rq->cmd_type = REQ_TYPE_FS;
332 flush_rq->cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
333 flush_rq->rq_disk = first_rq->rq_disk;
334 flush_rq->end_io = flush_end_io;
Jens Axboe86db1e22008-01-29 14:53:40 +0100335
Ming Lei7ddab5d2014-09-25 23:23:42 +0800336 return blk_flush_queue_rq(flush_rq, false);
Tejun Heoae1b1532011-01-25 12:43:54 +0100337}
338
339static void flush_data_end_io(struct request *rq, int error)
340{
341 struct request_queue *q = rq->q;
Ming Leie97c2932014-09-25 23:23:46 +0800342 struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL);
Tejun Heoae1b1532011-01-25 12:43:54 +0100343
Jens Axboe86db1e22008-01-29 14:53:40 +0100344 /*
Tejun Heoe83a46b2011-03-04 19:09:02 +0100345 * After populating an empty queue, kick it to avoid stall. Read
346 * the comment in flush_end_io().
Jens Axboe86db1e22008-01-29 14:53:40 +0100347 */
Ming Lei0bae3522014-09-25 23:23:45 +0800348 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +0200349 blk_run_queue_async(q);
Tejun Heoae1b1532011-01-25 12:43:54 +0100350}
351
Jens Axboe320ae512013-10-24 09:20:05 +0100352static void mq_flush_data_end_io(struct request *rq, int error)
353{
354 struct request_queue *q = rq->q;
355 struct blk_mq_hw_ctx *hctx;
Ming Leie97c2932014-09-25 23:23:46 +0800356 struct blk_mq_ctx *ctx = rq->mq_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +0100357 unsigned long flags;
Ming Leie97c2932014-09-25 23:23:46 +0800358 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +0100359
Jens Axboe320ae512013-10-24 09:20:05 +0100360 hctx = q->mq_ops->map_queue(q, ctx->cpu);
361
362 /*
363 * After populating an empty queue, kick it to avoid stall. Read
364 * the comment in flush_end_io().
365 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800366 spin_lock_irqsave(&fq->mq_flush_lock, flags);
Ming Lei0bae3522014-09-25 23:23:45 +0800367 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
Jens Axboe320ae512013-10-24 09:20:05 +0100368 blk_mq_run_hw_queue(hctx, true);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800369 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
Jens Axboe320ae512013-10-24 09:20:05 +0100370}
371
Tejun Heoae1b1532011-01-25 12:43:54 +0100372/**
373 * blk_insert_flush - insert a new FLUSH/FUA request
374 * @rq: request to insert
375 *
Jens Axboeb710a482011-03-30 09:52:30 +0200376 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
Jens Axboe320ae512013-10-24 09:20:05 +0100377 * or __blk_mq_run_hw_queue() to dispatch request.
Tejun Heoae1b1532011-01-25 12:43:54 +0100378 * @rq is being submitted. Analyze what needs to be done and put it on the
379 * right queue.
380 *
381 * CONTEXT:
Jens Axboe320ae512013-10-24 09:20:05 +0100382 * spin_lock_irq(q->queue_lock) in !mq case
Tejun Heoae1b1532011-01-25 12:43:54 +0100383 */
384void blk_insert_flush(struct request *rq)
385{
386 struct request_queue *q = rq->q;
387 unsigned int fflags = q->flush_flags; /* may change, cache */
388 unsigned int policy = blk_flush_policy(fflags, rq);
Ming Leie97c2932014-09-25 23:23:46 +0800389 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
Tejun Heoae1b1532011-01-25 12:43:54 +0100390
Tejun Heoae1b1532011-01-25 12:43:54 +0100391 /*
392 * @policy now records what operations need to be done. Adjust
393 * REQ_FLUSH and FUA for the driver.
394 */
Christoph Hellwig18741982014-02-10 09:29:00 -0700395 rq->cmd_flags &= ~REQ_FLUSH;
Tejun Heoae1b1532011-01-25 12:43:54 +0100396 if (!(fflags & REQ_FUA))
Tejun Heo4fed9472010-09-03 11:56:17 +0200397 rq->cmd_flags &= ~REQ_FUA;
Jens Axboe86db1e22008-01-29 14:53:40 +0100398
Tejun Heoae1b1532011-01-25 12:43:54 +0100399 /*
Jeff Moyer4853aba2011-08-15 21:37:25 +0200400 * An empty flush handed down from a stacking driver may
401 * translate into nothing if the underlying device does not
402 * advertise a write-back cache. In this case, simply
403 * complete the request.
404 */
405 if (!policy) {
Jens Axboe320ae512013-10-24 09:20:05 +0100406 if (q->mq_ops)
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700407 blk_mq_end_request(rq, 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100408 else
409 __blk_end_bidi_request(rq, 0, 0, 0);
Jeff Moyer4853aba2011-08-15 21:37:25 +0200410 return;
411 }
412
Jeff Moyer834f9f62011-10-17 12:57:22 +0200413 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
Jeff Moyer4853aba2011-08-15 21:37:25 +0200414
415 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100416 * If there's data but flush is not necessary, the request can be
417 * processed directly without going through flush machinery. Queue
418 * for normal execution.
419 */
420 if ((policy & REQ_FSEQ_DATA) &&
421 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
Jens Axboe320ae512013-10-24 09:20:05 +0100422 if (q->mq_ops) {
Christoph Hellwigfeb71da2014-02-20 15:32:37 -0800423 blk_mq_insert_request(rq, false, false, true);
Jens Axboe320ae512013-10-24 09:20:05 +0100424 } else
425 list_add_tail(&rq->queuelist, &q->queue_head);
Tejun Heoae1b1532011-01-25 12:43:54 +0100426 return;
427 }
428
429 /*
430 * @rq should go through flush machinery. Mark it part of flush
431 * sequence and submit for further processing.
432 */
433 memset(&rq->flush, 0, sizeof(rq->flush));
434 INIT_LIST_HEAD(&rq->flush.list);
435 rq->cmd_flags |= REQ_FLUSH_SEQ;
Jeff Moyer4853aba2011-08-15 21:37:25 +0200436 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
Jens Axboe320ae512013-10-24 09:20:05 +0100437 if (q->mq_ops) {
438 rq->end_io = mq_flush_data_end_io;
439
Ming Lei7c94e1c2014-09-25 23:23:43 +0800440 spin_lock_irq(&fq->mq_flush_lock);
Ming Lei0bae3522014-09-25 23:23:45 +0800441 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800442 spin_unlock_irq(&fq->mq_flush_lock);
Jens Axboe320ae512013-10-24 09:20:05 +0100443 return;
444 }
Tejun Heoae1b1532011-01-25 12:43:54 +0100445 rq->end_io = flush_data_end_io;
446
Ming Lei0bae3522014-09-25 23:23:45 +0800447 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
Tejun Heoae1b1532011-01-25 12:43:54 +0100448}
449
450/**
Jens Axboe86db1e22008-01-29 14:53:40 +0100451 * blkdev_issue_flush - queue a flush
452 * @bdev: blockdev to issue flush for
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400453 * @gfp_mask: memory allocation flags (for bio_alloc)
Jens Axboe86db1e22008-01-29 14:53:40 +0100454 * @error_sector: error sector
455 *
456 * Description:
457 * Issue a flush for the block device in question. Caller can supply
458 * room for storing the error offset in case of a flush error, if they
Dmitry Monakhovf17e2322010-04-28 17:55:07 +0400459 * wish to. If WAIT flag is not passed then caller may check only what
460 * request was pushed in some internal queue for later handling.
Jens Axboe86db1e22008-01-29 14:53:40 +0100461 */
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400462int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
Christoph Hellwigdd3932e2010-09-16 20:51:46 +0200463 sector_t *error_sector)
Jens Axboe86db1e22008-01-29 14:53:40 +0100464{
Jens Axboe86db1e22008-01-29 14:53:40 +0100465 struct request_queue *q;
466 struct bio *bio;
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400467 int ret = 0;
Jens Axboe86db1e22008-01-29 14:53:40 +0100468
469 if (bdev->bd_disk == NULL)
470 return -ENXIO;
471
472 q = bdev_get_queue(bdev);
473 if (!q)
474 return -ENXIO;
475
Dave Chinnerf10d9f62010-07-13 17:50:50 +1000476 /*
477 * some block devices may not have their queue correctly set up here
478 * (e.g. loop device without a backing file) and so issuing a flush
479 * here will panic. Ensure there is a request function before issuing
Tejun Heod391a2d2010-09-03 11:56:17 +0200480 * the flush.
Dave Chinnerf10d9f62010-07-13 17:50:50 +1000481 */
482 if (!q->make_request_fn)
483 return -ENXIO;
484
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400485 bio = bio_alloc(gfp_mask, 0);
Jens Axboe86db1e22008-01-29 14:53:40 +0100486 bio->bi_bdev = bdev;
Dmitry Monakhovf17e2322010-04-28 17:55:07 +0400487
Kent Overstreet33879d42013-11-23 22:33:32 -0800488 ret = submit_bio_wait(WRITE_FLUSH, bio);
Christoph Hellwigdd3932e2010-09-16 20:51:46 +0200489
490 /*
491 * The driver must store the error location in ->bi_sector, if
492 * it supports it. For non-stacked drivers, this should be
493 * copied from blk_rq_pos(rq).
494 */
495 if (error_sector)
Kent Overstreet4f024f32013-10-11 15:44:27 -0700496 *error_sector = bio->bi_iter.bi_sector;
Jens Axboe86db1e22008-01-29 14:53:40 +0100497
Jens Axboe86db1e22008-01-29 14:53:40 +0100498 bio_put(bio);
499 return ret;
500}
Jens Axboe86db1e22008-01-29 14:53:40 +0100501EXPORT_SYMBOL(blkdev_issue_flush);
Jens Axboe320ae512013-10-24 09:20:05 +0100502
Ming Leif70ced02014-09-25 23:23:47 +0800503struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
504 int node, int cmd_size)
Jens Axboe320ae512013-10-24 09:20:05 +0100505{
Ming Lei7c94e1c2014-09-25 23:23:43 +0800506 struct blk_flush_queue *fq;
507 int rq_sz = sizeof(struct request);
Ming Lei1bcb1ea2014-09-25 23:23:39 +0800508
Ming Leif70ced02014-09-25 23:23:47 +0800509 fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800510 if (!fq)
511 goto fail;
Ming Lei1bcb1ea2014-09-25 23:23:39 +0800512
Ming Lei7c94e1c2014-09-25 23:23:43 +0800513 if (q->mq_ops) {
514 spin_lock_init(&fq->mq_flush_lock);
Ming Leif70ced02014-09-25 23:23:47 +0800515 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
Ming Lei7c94e1c2014-09-25 23:23:43 +0800516 }
517
Ming Leif70ced02014-09-25 23:23:47 +0800518 fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800519 if (!fq->flush_rq)
520 goto fail_rq;
521
522 INIT_LIST_HEAD(&fq->flush_queue[0]);
523 INIT_LIST_HEAD(&fq->flush_queue[1]);
524 INIT_LIST_HEAD(&fq->flush_data_in_flight);
525
526 return fq;
527
528 fail_rq:
529 kfree(fq);
530 fail:
531 return NULL;
532}
533
Ming Leiba483382014-09-25 23:23:44 +0800534void blk_free_flush_queue(struct blk_flush_queue *fq)
Ming Lei7c94e1c2014-09-25 23:23:43 +0800535{
536 /* bio based request queue hasn't flush queue */
537 if (!fq)
538 return;
539
540 kfree(fq->flush_rq);
541 kfree(fq);
Jens Axboe320ae512013-10-24 09:20:05 +0100542}