blob: 20b7c7a02f1cbdfe5a63c5918d3abd356bd4265b [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
Christoph Hellwige8064022016-10-20 15:12:13 +020059 * sequence is complete. This is implemented by testing RQF_FLUSH_SEQ in
Tejun Heoae1b1532011-01-25 12:43:54 +010060 * 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 */
Christoph Hellwige8064022016-10-20 15:12:13 +0200130 rq->rq_flags &= ~RQF_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) {
Bart Van Assche2b053ac2016-10-28 17:21:41 -0700137 blk_mq_add_to_requeue_list(rq, add_front, true);
Christoph Hellwig18741982014-02-10 09:29:00 -0700138 return false;
139 } else {
Mike Snitzer10beafc2014-03-08 20:19:20 -0700140 if (add_front)
141 list_add(&rq->queuelist, &rq->q->queue_head);
142 else
143 list_add_tail(&rq->queuelist, &rq->q->queue_head);
Christoph Hellwig18741982014-02-10 09:29:00 -0700144 return true;
145 }
Jens Axboe86db1e22008-01-29 14:53:40 +0100146}
147
Tejun Heoae1b1532011-01-25 12:43:54 +0100148/**
149 * blk_flush_complete_seq - complete flush sequence
150 * @rq: FLUSH/FUA request being sequenced
Ming Lei0bae3522014-09-25 23:23:45 +0800151 * @fq: flush queue
Tejun Heoae1b1532011-01-25 12:43:54 +0100152 * @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:
Ming Lei7c94e1c2014-09-25 23:23:43 +0800159 * spin_lock_irq(q->queue_lock or fq->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 */
Ming Lei0bae3522014-09-25 23:23:45 +0800164static bool blk_flush_complete_seq(struct request *rq,
165 struct blk_flush_queue *fq,
166 unsigned int seq, int error)
Jens Axboe86db1e22008-01-29 14:53:40 +0100167{
Tejun Heoae1b1532011-01-25 12:43:54 +0100168 struct request_queue *q = rq->q;
Ming Lei7c94e1c2014-09-25 23:23:43 +0800169 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
Jens Axboe320ae512013-10-24 09:20:05 +0100170 bool queued = false, kicked;
Jens Axboe86db1e22008-01-29 14:53:40 +0100171
Tejun Heoae1b1532011-01-25 12:43:54 +0100172 BUG_ON(rq->flush.seq & seq);
173 rq->flush.seq |= seq;
Jens Axboe86db1e22008-01-29 14:53:40 +0100174
Tejun Heoae1b1532011-01-25 12:43:54 +0100175 if (likely(!error))
176 seq = blk_flush_cur_seq(rq);
177 else
178 seq = REQ_FSEQ_DONE;
Tejun Heo28e7d182010-09-03 11:56:16 +0200179
Tejun Heoae1b1532011-01-25 12:43:54 +0100180 switch (seq) {
181 case REQ_FSEQ_PREFLUSH:
182 case REQ_FSEQ_POSTFLUSH:
183 /* queue for flush */
184 if (list_empty(pending))
Ming Lei7c94e1c2014-09-25 23:23:43 +0800185 fq->flush_pending_since = jiffies;
Tejun Heoae1b1532011-01-25 12:43:54 +0100186 list_move_tail(&rq->flush.list, pending);
Tejun Heo28e7d182010-09-03 11:56:16 +0200187 break;
Tejun Heoae1b1532011-01-25 12:43:54 +0100188
189 case REQ_FSEQ_DATA:
Ming Lei7c94e1c2014-09-25 23:23:43 +0800190 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
Mike Snitzer10beafc2014-03-08 20:19:20 -0700191 queued = blk_flush_queue_rq(rq, true);
Tejun Heoae1b1532011-01-25 12:43:54 +0100192 break;
193
194 case REQ_FSEQ_DONE:
Tejun Heo09d60c72010-09-03 11:56:17 +0200195 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100196 * @rq was previously adjusted by blk_flush_issue() for
197 * flush sequencing and may already have gone through the
198 * flush data request completion path. Restore @rq for
199 * normal completion and end it.
Tejun Heo09d60c72010-09-03 11:56:17 +0200200 */
Tejun Heoae1b1532011-01-25 12:43:54 +0100201 BUG_ON(!list_empty(&rq->queuelist));
202 list_del_init(&rq->flush.list);
203 blk_flush_restore_request(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100204 if (q->mq_ops)
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700205 blk_mq_end_request(rq, error);
Jens Axboe320ae512013-10-24 09:20:05 +0100206 else
207 __blk_end_request_all(rq, error);
Tejun Heo28e7d182010-09-03 11:56:16 +0200208 break;
Tejun Heoae1b1532011-01-25 12:43:54 +0100209
Tejun Heo28e7d182010-09-03 11:56:16 +0200210 default:
211 BUG();
212 }
Christoph Hellwigcde4c402010-09-03 11:56:17 +0200213
Ming Lei0bae3522014-09-25 23:23:45 +0800214 kicked = blk_kick_flush(q, fq);
Jens Axboe320ae512013-10-24 09:20:05 +0100215 return kicked | queued;
Tejun Heo28e7d182010-09-03 11:56:16 +0200216}
217
Tejun Heoae1b1532011-01-25 12:43:54 +0100218static void flush_end_io(struct request *flush_rq, int error)
Jens Axboe86db1e22008-01-29 14:53:40 +0100219{
Tejun Heoae1b1532011-01-25 12:43:54 +0100220 struct request_queue *q = flush_rq->q;
Jens Axboe320ae512013-10-24 09:20:05 +0100221 struct list_head *running;
Tejun Heoae1b1532011-01-25 12:43:54 +0100222 bool queued = false;
223 struct request *rq, *n;
Jens Axboe320ae512013-10-24 09:20:05 +0100224 unsigned long flags = 0;
Ming Leie97c2932014-09-25 23:23:46 +0800225 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
Tejun Heo8f11b3e2008-11-28 13:32:05 +0900226
Shaohua Li22302372014-05-30 08:06:42 -0600227 if (q->mq_ops) {
Ming Lei0048b482015-08-09 03:41:51 -0400228 struct blk_mq_hw_ctx *hctx;
229
230 /* release the tag's ownership to the req cloned from */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800231 spin_lock_irqsave(&fq->mq_flush_lock, flags);
Christoph Hellwig7d7e0f92016-09-14 16:18:54 +0200232 hctx = blk_mq_map_queue(q, flush_rq->mq_ctx->cpu);
Ming Lei0048b482015-08-09 03:41:51 -0400233 blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq);
Ming Lei7ddab5d2014-09-25 23:23:42 +0800234 flush_rq->tag = -1;
Shaohua Li22302372014-05-30 08:06:42 -0600235 }
Christoph Hellwig18741982014-02-10 09:29:00 -0700236
Ming Lei7c94e1c2014-09-25 23:23:43 +0800237 running = &fq->flush_queue[fq->flush_running_idx];
238 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
Tejun Heoae1b1532011-01-25 12:43:54 +0100239
240 /* account completion of the flush request */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800241 fq->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);
Ming Lei0bae3522014-09-25 23:23:45 +0800251 queued |= blk_flush_complete_seq(rq, fq, 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 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800265 if (queued || fq->flush_queue_delayed) {
Christoph Hellwig18741982014-02-10 09:29:00 -0700266 WARN_ON(q->mq_ops);
267 blk_run_queue_async(q);
Jens Axboe320ae512013-10-24 09:20:05 +0100268 }
Ming Lei7c94e1c2014-09-25 23:23:43 +0800269 fq->flush_queue_delayed = 0;
Jens Axboe320ae512013-10-24 09:20:05 +0100270 if (q->mq_ops)
Ming Lei7c94e1c2014-09-25 23:23:43 +0800271 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
Jens Axboe320ae512013-10-24 09:20:05 +0100272}
273
Tejun Heoae1b1532011-01-25 12:43:54 +0100274/**
275 * blk_kick_flush - consider issuing flush request
276 * @q: request_queue being kicked
Ming Lei0bae3522014-09-25 23:23:45 +0800277 * @fq: flush queue
Tejun Heoae1b1532011-01-25 12:43:54 +0100278 *
279 * Flush related states of @q have changed, consider issuing flush request.
280 * Please read the comment at the top of this file for more info.
281 *
282 * CONTEXT:
Ming Lei7c94e1c2014-09-25 23:23:43 +0800283 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
Tejun Heoae1b1532011-01-25 12:43:54 +0100284 *
285 * RETURNS:
286 * %true if flush was issued, %false otherwise.
287 */
Ming Lei0bae3522014-09-25 23:23:45 +0800288static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq)
Tejun Heoae1b1532011-01-25 12:43:54 +0100289{
Ming Lei7c94e1c2014-09-25 23:23:43 +0800290 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
Tejun Heoae1b1532011-01-25 12:43:54 +0100291 struct request *first_rq =
292 list_first_entry(pending, struct request, flush.list);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800293 struct request *flush_rq = fq->flush_rq;
Tejun Heoae1b1532011-01-25 12:43:54 +0100294
295 /* C1 described at the top of this file */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800296 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
Tejun Heoae1b1532011-01-25 12:43:54 +0100297 return false;
298
299 /* C2 and C3 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800300 if (!list_empty(&fq->flush_data_in_flight) &&
Tejun Heoae1b1532011-01-25 12:43:54 +0100301 time_before(jiffies,
Ming Lei7c94e1c2014-09-25 23:23:43 +0800302 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
Tejun Heoae1b1532011-01-25 12:43:54 +0100303 return false;
Tejun Heo28e7d182010-09-03 11:56:16 +0200304
305 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100306 * Issue flush and toggle pending_idx. This makes pending_idx
307 * different from running_idx, which means flush is in flight.
Tejun Heo28e7d182010-09-03 11:56:16 +0200308 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800309 fq->flush_pending_idx ^= 1;
Christoph Hellwig18741982014-02-10 09:29:00 -0700310
Ming Lei7ddab5d2014-09-25 23:23:42 +0800311 blk_rq_init(q, flush_rq);
Ming Leif70ced02014-09-25 23:23:47 +0800312
313 /*
314 * Borrow tag from the first request since they can't
Ming Lei0048b482015-08-09 03:41:51 -0400315 * be in flight at the same time. And acquire the tag's
316 * ownership for flush req.
Ming Leif70ced02014-09-25 23:23:47 +0800317 */
318 if (q->mq_ops) {
Ming Lei0048b482015-08-09 03:41:51 -0400319 struct blk_mq_hw_ctx *hctx;
320
Ming Leif70ced02014-09-25 23:23:47 +0800321 flush_rq->mq_ctx = first_rq->mq_ctx;
322 flush_rq->tag = first_rq->tag;
Ming Lei0048b482015-08-09 03:41:51 -0400323 fq->orig_rq = first_rq;
324
Christoph Hellwig7d7e0f92016-09-14 16:18:54 +0200325 hctx = blk_mq_map_queue(q, first_rq->mq_ctx->cpu);
Ming Lei0048b482015-08-09 03:41:51 -0400326 blk_mq_tag_set_rq(hctx, first_rq->tag, flush_rq);
Ming Leif70ced02014-09-25 23:23:47 +0800327 }
Jens Axboe320ae512013-10-24 09:20:05 +0100328
Ming Lei7ddab5d2014-09-25 23:23:42 +0800329 flush_rq->cmd_type = REQ_TYPE_FS;
Christoph Hellwig70fd7612016-11-01 07:40:10 -0600330 flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
Christoph Hellwige8064022016-10-20 15:12:13 +0200331 flush_rq->rq_flags |= RQF_FLUSH_SEQ;
Ming Lei7ddab5d2014-09-25 23:23:42 +0800332 flush_rq->rq_disk = first_rq->rq_disk;
333 flush_rq->end_io = flush_end_io;
Jens Axboe86db1e22008-01-29 14:53:40 +0100334
Ming Lei7ddab5d2014-09-25 23:23:42 +0800335 return blk_flush_queue_rq(flush_rq, false);
Tejun Heoae1b1532011-01-25 12:43:54 +0100336}
337
338static void flush_data_end_io(struct request *rq, int error)
339{
340 struct request_queue *q = rq->q;
Ming Leie97c2932014-09-25 23:23:46 +0800341 struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL);
Tejun Heoae1b1532011-01-25 12:43:54 +0100342
Jens Axboe86db1e22008-01-29 14:53:40 +0100343 /*
Ming Lei94d7dea2016-10-26 16:57:15 +0800344 * Updating q->in_flight[] here for making this tag usable
345 * early. Because in blk_queue_start_tag(),
346 * q->in_flight[BLK_RW_ASYNC] is used to limit async I/O and
347 * reserve tags for sync I/O.
348 *
349 * More importantly this way can avoid the following I/O
350 * deadlock:
351 *
352 * - suppose there are 40 fua requests comming to flush queue
353 * and queue depth is 31
354 * - 30 rqs are scheduled then blk_queue_start_tag() can't alloc
355 * tag for async I/O any more
356 * - all the 30 rqs are completed before FLUSH_PENDING_TIMEOUT
357 * and flush_data_end_io() is called
358 * - the other rqs still can't go ahead if not updating
359 * q->in_flight[BLK_RW_ASYNC] here, meantime these rqs
360 * are held in flush data queue and make no progress of
361 * handling post flush rq
362 * - only after the post flush rq is handled, all these rqs
363 * can be completed
364 */
365
366 elv_completed_request(q, rq);
367
368 /* for avoiding double accounting */
Linus Torvalds36869cb2016-12-13 10:19:16 -0800369 rq->rq_flags &= ~RQF_STARTED;
Ming Lei94d7dea2016-10-26 16:57:15 +0800370
371 /*
Tejun Heoe83a46b2011-03-04 19:09:02 +0100372 * After populating an empty queue, kick it to avoid stall. Read
373 * the comment in flush_end_io().
Jens Axboe86db1e22008-01-29 14:53:40 +0100374 */
Ming Lei0bae3522014-09-25 23:23:45 +0800375 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
Christoph Hellwig24ecfbe2011-04-18 11:41:33 +0200376 blk_run_queue_async(q);
Tejun Heoae1b1532011-01-25 12:43:54 +0100377}
378
Jens Axboe320ae512013-10-24 09:20:05 +0100379static void mq_flush_data_end_io(struct request *rq, int error)
380{
381 struct request_queue *q = rq->q;
382 struct blk_mq_hw_ctx *hctx;
Ming Leie97c2932014-09-25 23:23:46 +0800383 struct blk_mq_ctx *ctx = rq->mq_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +0100384 unsigned long flags;
Ming Leie97c2932014-09-25 23:23:46 +0800385 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
Jens Axboe320ae512013-10-24 09:20:05 +0100386
Christoph Hellwig7d7e0f92016-09-14 16:18:54 +0200387 hctx = blk_mq_map_queue(q, ctx->cpu);
Jens Axboe320ae512013-10-24 09:20:05 +0100388
389 /*
390 * After populating an empty queue, kick it to avoid stall. Read
391 * the comment in flush_end_io().
392 */
Ming Lei7c94e1c2014-09-25 23:23:43 +0800393 spin_lock_irqsave(&fq->mq_flush_lock, flags);
Ming Lei0bae3522014-09-25 23:23:45 +0800394 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
Jens Axboe320ae512013-10-24 09:20:05 +0100395 blk_mq_run_hw_queue(hctx, true);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800396 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
Jens Axboe320ae512013-10-24 09:20:05 +0100397}
398
Tejun Heoae1b1532011-01-25 12:43:54 +0100399/**
400 * blk_insert_flush - insert a new FLUSH/FUA request
401 * @rq: request to insert
402 *
Jens Axboeb710a482011-03-30 09:52:30 +0200403 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
Jens Axboe320ae512013-10-24 09:20:05 +0100404 * or __blk_mq_run_hw_queue() to dispatch request.
Tejun Heoae1b1532011-01-25 12:43:54 +0100405 * @rq is being submitted. Analyze what needs to be done and put it on the
406 * right queue.
407 *
408 * CONTEXT:
Jens Axboe320ae512013-10-24 09:20:05 +0100409 * spin_lock_irq(q->queue_lock) in !mq case
Tejun Heoae1b1532011-01-25 12:43:54 +0100410 */
411void blk_insert_flush(struct request *rq)
412{
413 struct request_queue *q = rq->q;
Jens Axboec888a8f2016-04-13 13:33:19 -0600414 unsigned long fflags = q->queue_flags; /* may change, cache */
Tejun Heoae1b1532011-01-25 12:43:54 +0100415 unsigned int policy = blk_flush_policy(fflags, rq);
Ming Leie97c2932014-09-25 23:23:46 +0800416 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
Tejun Heoae1b1532011-01-25 12:43:54 +0100417
Tejun Heoae1b1532011-01-25 12:43:54 +0100418 /*
419 * @policy now records what operations need to be done. Adjust
Mike Christie28a8f0d2016-06-05 14:32:25 -0500420 * REQ_PREFLUSH and FUA for the driver.
Tejun Heoae1b1532011-01-25 12:43:54 +0100421 */
Mike Christie28a8f0d2016-06-05 14:32:25 -0500422 rq->cmd_flags &= ~REQ_PREFLUSH;
Jens Axboec888a8f2016-04-13 13:33:19 -0600423 if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
Tejun Heo4fed9472010-09-03 11:56:17 +0200424 rq->cmd_flags &= ~REQ_FUA;
Jens Axboe86db1e22008-01-29 14:53:40 +0100425
Tejun Heoae1b1532011-01-25 12:43:54 +0100426 /*
Jens Axboeae5b2ec2016-11-08 19:39:28 -0700427 * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
428 * of those flags, we have to set REQ_SYNC to avoid skewing
429 * the request accounting.
430 */
431 rq->cmd_flags |= REQ_SYNC;
432
433 /*
Jeff Moyer4853aba2011-08-15 21:37:25 +0200434 * An empty flush handed down from a stacking driver may
435 * translate into nothing if the underlying device does not
436 * advertise a write-back cache. In this case, simply
437 * complete the request.
438 */
439 if (!policy) {
Jens Axboe320ae512013-10-24 09:20:05 +0100440 if (q->mq_ops)
Christoph Hellwigc8a446a2014-09-13 16:40:10 -0700441 blk_mq_end_request(rq, 0);
Jens Axboe320ae512013-10-24 09:20:05 +0100442 else
443 __blk_end_bidi_request(rq, 0, 0, 0);
Jeff Moyer4853aba2011-08-15 21:37:25 +0200444 return;
445 }
446
Jeff Moyer834f9f62011-10-17 12:57:22 +0200447 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
Jeff Moyer4853aba2011-08-15 21:37:25 +0200448
449 /*
Tejun Heoae1b1532011-01-25 12:43:54 +0100450 * If there's data but flush is not necessary, the request can be
451 * processed directly without going through flush machinery. Queue
452 * for normal execution.
453 */
454 if ((policy & REQ_FSEQ_DATA) &&
455 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
Jens Axboe320ae512013-10-24 09:20:05 +0100456 if (q->mq_ops) {
Jens Axboec8e52ba2016-12-07 15:53:18 -0700457 blk_mq_insert_request(rq, false, true, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100458 } else
Jens Axboedcd83762015-11-25 10:12:54 -0700459 list_add_tail(&rq->queuelist, &q->queue_head);
Tejun Heoae1b1532011-01-25 12:43:54 +0100460 return;
461 }
462
463 /*
464 * @rq should go through flush machinery. Mark it part of flush
465 * sequence and submit for further processing.
466 */
467 memset(&rq->flush, 0, sizeof(rq->flush));
468 INIT_LIST_HEAD(&rq->flush.list);
Christoph Hellwige8064022016-10-20 15:12:13 +0200469 rq->rq_flags |= RQF_FLUSH_SEQ;
Jeff Moyer4853aba2011-08-15 21:37:25 +0200470 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
Jens Axboe320ae512013-10-24 09:20:05 +0100471 if (q->mq_ops) {
472 rq->end_io = mq_flush_data_end_io;
473
Ming Lei7c94e1c2014-09-25 23:23:43 +0800474 spin_lock_irq(&fq->mq_flush_lock);
Ming Lei0bae3522014-09-25 23:23:45 +0800475 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800476 spin_unlock_irq(&fq->mq_flush_lock);
Jens Axboe320ae512013-10-24 09:20:05 +0100477 return;
478 }
Tejun Heoae1b1532011-01-25 12:43:54 +0100479 rq->end_io = flush_data_end_io;
480
Ming Lei0bae3522014-09-25 23:23:45 +0800481 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
Tejun Heoae1b1532011-01-25 12:43:54 +0100482}
483
484/**
Jens Axboe86db1e22008-01-29 14:53:40 +0100485 * blkdev_issue_flush - queue a flush
486 * @bdev: blockdev to issue flush for
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400487 * @gfp_mask: memory allocation flags (for bio_alloc)
Jens Axboe86db1e22008-01-29 14:53:40 +0100488 * @error_sector: error sector
489 *
490 * Description:
491 * Issue a flush for the block device in question. Caller can supply
492 * room for storing the error offset in case of a flush error, if they
Dmitry Monakhovf17e2322010-04-28 17:55:07 +0400493 * wish to. If WAIT flag is not passed then caller may check only what
494 * request was pushed in some internal queue for later handling.
Jens Axboe86db1e22008-01-29 14:53:40 +0100495 */
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400496int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
Christoph Hellwigdd3932e2010-09-16 20:51:46 +0200497 sector_t *error_sector)
Jens Axboe86db1e22008-01-29 14:53:40 +0100498{
Jens Axboe86db1e22008-01-29 14:53:40 +0100499 struct request_queue *q;
500 struct bio *bio;
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400501 int ret = 0;
Jens Axboe86db1e22008-01-29 14:53:40 +0100502
503 if (bdev->bd_disk == NULL)
504 return -ENXIO;
505
506 q = bdev_get_queue(bdev);
507 if (!q)
508 return -ENXIO;
509
Dave Chinnerf10d9f62010-07-13 17:50:50 +1000510 /*
511 * some block devices may not have their queue correctly set up here
512 * (e.g. loop device without a backing file) and so issuing a flush
513 * here will panic. Ensure there is a request function before issuing
Tejun Heod391a2d2010-09-03 11:56:17 +0200514 * the flush.
Dave Chinnerf10d9f62010-07-13 17:50:50 +1000515 */
516 if (!q->make_request_fn)
517 return -ENXIO;
518
Dmitry Monakhovfbd9b092010-04-28 17:55:06 +0400519 bio = bio_alloc(gfp_mask, 0);
Jens Axboe86db1e22008-01-29 14:53:40 +0100520 bio->bi_bdev = bdev;
Christoph Hellwig70fd7612016-11-01 07:40:10 -0600521 bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
Dmitry Monakhovf17e2322010-04-28 17:55:07 +0400522
Mike Christie4e49ea42016-06-05 14:31:41 -0500523 ret = submit_bio_wait(bio);
Christoph Hellwigdd3932e2010-09-16 20:51:46 +0200524
525 /*
526 * The driver must store the error location in ->bi_sector, if
527 * it supports it. For non-stacked drivers, this should be
528 * copied from blk_rq_pos(rq).
529 */
530 if (error_sector)
Kent Overstreet4f024f32013-10-11 15:44:27 -0700531 *error_sector = bio->bi_iter.bi_sector;
Jens Axboe86db1e22008-01-29 14:53:40 +0100532
Jens Axboe86db1e22008-01-29 14:53:40 +0100533 bio_put(bio);
534 return ret;
535}
Jens Axboe86db1e22008-01-29 14:53:40 +0100536EXPORT_SYMBOL(blkdev_issue_flush);
Jens Axboe320ae512013-10-24 09:20:05 +0100537
Ming Leif70ced02014-09-25 23:23:47 +0800538struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
539 int node, int cmd_size)
Jens Axboe320ae512013-10-24 09:20:05 +0100540{
Ming Lei7c94e1c2014-09-25 23:23:43 +0800541 struct blk_flush_queue *fq;
542 int rq_sz = sizeof(struct request);
Ming Lei1bcb1ea2014-09-25 23:23:39 +0800543
Ming Leif70ced02014-09-25 23:23:47 +0800544 fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800545 if (!fq)
546 goto fail;
Ming Lei1bcb1ea2014-09-25 23:23:39 +0800547
Ming Lei7c94e1c2014-09-25 23:23:43 +0800548 if (q->mq_ops) {
549 spin_lock_init(&fq->mq_flush_lock);
Ming Leif70ced02014-09-25 23:23:47 +0800550 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
Ming Lei7c94e1c2014-09-25 23:23:43 +0800551 }
552
Ming Leif70ced02014-09-25 23:23:47 +0800553 fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node);
Ming Lei7c94e1c2014-09-25 23:23:43 +0800554 if (!fq->flush_rq)
555 goto fail_rq;
556
557 INIT_LIST_HEAD(&fq->flush_queue[0]);
558 INIT_LIST_HEAD(&fq->flush_queue[1]);
559 INIT_LIST_HEAD(&fq->flush_data_in_flight);
560
561 return fq;
562
563 fail_rq:
564 kfree(fq);
565 fail:
566 return NULL;
567}
568
Ming Leiba483382014-09-25 23:23:44 +0800569void blk_free_flush_queue(struct blk_flush_queue *fq)
Ming Lei7c94e1c2014-09-25 23:23:43 +0800570{
571 /* bio based request queue hasn't flush queue */
572 if (!fq)
573 return;
574
575 kfree(fq->flush_rq);
576 kfree(fq);
Jens Axboe320ae512013-10-24 09:20:05 +0100577}