1 // SPDX-License-Identifier: GPL-2.0
3 * Functions to sequence PREFLUSH and FUA writes.
5 * Copyright (C) 2011 Max Planck Institute for Gravitational Physics
6 * Copyright (C) 2011 Tejun Heo <tj@kernel.org>
8 * REQ_{PREFLUSH|FUA} requests are decomposed to sequences consisted of three
9 * optional steps - PREFLUSH, DATA and POSTFLUSH - according to the request
10 * properties and hardware capability.
12 * If a request doesn't have data, only REQ_PREFLUSH makes sense, which
13 * indicates a simple flush request. If there is data, REQ_PREFLUSH indicates
14 * that the device cache should be flushed before the data is executed, and
15 * REQ_FUA means that the data must be on non-volatile media on request
18 * If the device doesn't have writeback cache, PREFLUSH and FUA don't make any
19 * difference. The requests are either completed immediately if there's no data
20 * or executed as normal requests otherwise.
22 * If the device has writeback cache and supports FUA, REQ_PREFLUSH is
23 * translated to PREFLUSH but REQ_FUA is passed down directly with DATA.
25 * If the device has writeback cache and doesn't support FUA, REQ_PREFLUSH
26 * is translated to PREFLUSH and REQ_FUA to POSTFLUSH.
28 * The actual execution of flush is double buffered. Whenever a request
29 * needs to execute PRE or POSTFLUSH, it queues at
30 * fq->flush_queue[fq->flush_pending_idx]. Once certain criteria are met, a
31 * REQ_OP_FLUSH is issued and the pending_idx is toggled. When the flush
32 * completes, all the requests which were pending are proceeded to the next
33 * step. This allows arbitrary merging of different types of PREFLUSH/FUA
36 * Currently, the following conditions are used to determine when to issue
39 * C1. At any given time, only one flush shall be in progress. This makes
40 * double buffering sufficient.
42 * C2. Flush is deferred if any request is executing DATA of its sequence.
43 * This avoids issuing separate POSTFLUSHes for requests which shared
46 * C3. The second condition is ignored if there is a request which has
47 * waited longer than FLUSH_PENDING_TIMEOUT. This is to avoid
48 * starvation in the unlikely case where there are continuous stream of
49 * FUA (without PREFLUSH) requests.
51 * For devices which support FUA, it isn't clear whether C2 (and thus C3)
54 * Note that a sequenced PREFLUSH/FUA request with DATA is completed twice.
55 * Once while executing DATA and again after the whole sequence is
56 * complete. The first completion updates the contained bio but doesn't
57 * finish it so that the bio submitter is notified only after the whole
58 * sequence is complete. This is implemented by testing RQF_FLUSH_SEQ in
61 * The above peculiarity requires that each PREFLUSH/FUA request has only one
62 * bio attached to it, which is guaranteed as they aren't allowed to be
63 * merged in the usual way.
66 #include <linux/kernel.h>
67 #include <linux/module.h>
68 #include <linux/bio.h>
69 #include <linux/blkdev.h>
70 #include <linux/gfp.h>
71 #include <linux/blk-mq.h>
75 #include "blk-mq-tag.h"
76 #include "blk-mq-sched.h"
78 /* PREFLUSH/FUA sequences */
80 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),
85 REQ_FSEQ_ACTIONS = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
89 * If flush has been pending longer than the following timeout,
90 * it's issued even if flush_data requests are still in flight.
92 FLUSH_PENDING_TIMEOUT = 5 * HZ,
95 static void blk_kick_flush(struct request_queue *q,
96 struct blk_flush_queue *fq, unsigned int flags);
98 static unsigned int blk_flush_policy(unsigned long fflags, struct request *rq)
100 unsigned int policy = 0;
102 if (blk_rq_sectors(rq))
103 policy |= REQ_FSEQ_DATA;
105 if (fflags & (1UL << QUEUE_FLAG_WC)) {
106 if (rq->cmd_flags & REQ_PREFLUSH)
107 policy |= REQ_FSEQ_PREFLUSH;
108 if (!(fflags & (1UL << QUEUE_FLAG_FUA)) &&
109 (rq->cmd_flags & REQ_FUA))
110 policy |= REQ_FSEQ_POSTFLUSH;
115 static unsigned int blk_flush_cur_seq(struct request *rq)
117 return 1 << ffz(rq->flush.seq);
120 static void blk_flush_restore_request(struct request *rq)
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.
127 rq->bio = rq->biotail;
129 /* make @rq a normal request */
130 rq->rq_flags &= ~RQF_FLUSH_SEQ;
131 rq->end_io = rq->flush.saved_end_io;
134 static void blk_flush_queue_rq(struct request *rq, bool add_front)
136 blk_mq_add_to_requeue_list(rq, add_front, true);
139 static void blk_account_io_flush(struct request *rq)
141 struct block_device *part = rq->rq_disk->part0;
144 part_stat_inc(part, ios[STAT_FLUSH]);
145 part_stat_add(part, nsecs[STAT_FLUSH],
146 ktime_get_ns() - rq->start_time_ns);
151 * blk_flush_complete_seq - complete flush sequence
152 * @rq: PREFLUSH/FUA request being sequenced
154 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
155 * @error: whether an error occurred
157 * @rq just completed @seq part of its flush sequence, record the
158 * completion and trigger the next step.
161 * spin_lock_irq(fq->mq_flush_lock)
163 static void blk_flush_complete_seq(struct request *rq,
164 struct blk_flush_queue *fq,
165 unsigned int seq, blk_status_t error)
167 struct request_queue *q = rq->q;
168 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
169 unsigned int cmd_flags;
171 BUG_ON(rq->flush.seq & seq);
172 rq->flush.seq |= seq;
173 cmd_flags = rq->cmd_flags;
176 seq = blk_flush_cur_seq(rq);
181 case REQ_FSEQ_PREFLUSH:
182 case REQ_FSEQ_POSTFLUSH:
183 /* queue for flush */
184 if (list_empty(pending))
185 fq->flush_pending_since = jiffies;
186 list_move_tail(&rq->flush.list, pending);
190 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
191 blk_flush_queue_rq(rq, true);
196 * @rq was previously adjusted by blk_insert_flush() 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.
201 BUG_ON(!list_empty(&rq->queuelist));
202 list_del_init(&rq->flush.list);
203 blk_flush_restore_request(rq);
204 blk_mq_end_request(rq, error);
211 blk_kick_flush(q, fq, cmd_flags);
214 static void flush_end_io(struct request *flush_rq, blk_status_t error)
216 struct request_queue *q = flush_rq->q;
217 struct list_head *running;
218 struct request *rq, *n;
219 unsigned long flags = 0;
220 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
222 /* release the tag's ownership to the req cloned from */
223 spin_lock_irqsave(&fq->mq_flush_lock, flags);
225 if (!refcount_dec_and_test(&flush_rq->ref)) {
226 fq->rq_status = error;
227 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
231 blk_account_io_flush(flush_rq);
233 * Flush request has to be marked as IDLE when it is really ended
234 * because its .end_io() is called from timeout code path too for
235 * avoiding use-after-free.
237 WRITE_ONCE(flush_rq->state, MQ_RQ_IDLE);
238 if (fq->rq_status != BLK_STS_OK) {
239 error = fq->rq_status;
240 fq->rq_status = BLK_STS_OK;
244 flush_rq->tag = BLK_MQ_NO_TAG;
246 blk_mq_put_driver_tag(flush_rq);
247 flush_rq->internal_tag = BLK_MQ_NO_TAG;
250 running = &fq->flush_queue[fq->flush_running_idx];
251 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
253 /* account completion of the flush request */
254 fq->flush_running_idx ^= 1;
256 /* and push the waiting requests to the next stage */
257 list_for_each_entry_safe(rq, n, running, flush.list) {
258 unsigned int seq = blk_flush_cur_seq(rq);
260 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
261 blk_flush_complete_seq(rq, fq, seq, error);
264 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
267 bool is_flush_rq(struct request *rq)
269 return rq->end_io == flush_end_io;
273 * blk_kick_flush - consider issuing flush request
274 * @q: request_queue being kicked
276 * @flags: cmd_flags of the original request
278 * Flush related states of @q have changed, consider issuing flush request.
279 * Please read the comment at the top of this file for more info.
282 * spin_lock_irq(fq->mq_flush_lock)
285 static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
288 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
289 struct request *first_rq =
290 list_first_entry(pending, struct request, flush.list);
291 struct request *flush_rq = fq->flush_rq;
293 /* C1 described at the top of this file */
294 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
298 if (!list_empty(&fq->flush_data_in_flight) &&
300 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
304 * Issue flush and toggle pending_idx. This makes pending_idx
305 * different from running_idx, which means flush is in flight.
307 fq->flush_pending_idx ^= 1;
309 blk_rq_init(q, flush_rq);
312 * In case of none scheduler, borrow tag from the first request
313 * since they can't be in flight at the same time. And acquire
314 * the tag's ownership for flush req.
316 * In case of IO scheduler, flush rq need to borrow scheduler tag
317 * just for cheating put/get driver tag.
319 flush_rq->mq_ctx = first_rq->mq_ctx;
320 flush_rq->mq_hctx = first_rq->mq_hctx;
323 flush_rq->tag = first_rq->tag;
326 * We borrow data request's driver tag, so have to mark
327 * this flush request as INFLIGHT for avoiding double
328 * account of this driver tag
330 flush_rq->rq_flags |= RQF_MQ_INFLIGHT;
332 flush_rq->internal_tag = first_rq->internal_tag;
334 flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
335 flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK);
336 flush_rq->rq_flags |= RQF_FLUSH_SEQ;
337 flush_rq->rq_disk = first_rq->rq_disk;
338 flush_rq->end_io = flush_end_io;
340 * Order WRITE ->end_io and WRITE rq->ref, and its pair is the one
341 * implied in refcount_inc_not_zero() called from
342 * blk_mq_find_and_get_req(), which orders WRITE/READ flush_rq->ref
343 * and READ flush_rq->end_io
346 refcount_set(&flush_rq->ref, 1);
348 blk_flush_queue_rq(flush_rq, false);
351 static void mq_flush_data_end_io(struct request *rq, blk_status_t error)
353 struct request_queue *q = rq->q;
354 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
355 struct blk_mq_ctx *ctx = rq->mq_ctx;
357 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
360 WARN_ON(rq->tag < 0);
361 blk_mq_put_driver_tag(rq);
365 * After populating an empty queue, kick it to avoid stall. Read
366 * the comment in flush_end_io().
368 spin_lock_irqsave(&fq->mq_flush_lock, flags);
369 blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
370 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
372 blk_mq_sched_restart(hctx);
376 * blk_insert_flush - insert a new PREFLUSH/FUA request
377 * @rq: request to insert
379 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
380 * or __blk_mq_run_hw_queue() to dispatch request.
381 * @rq is being submitted. Analyze what needs to be done and put it on the
384 void blk_insert_flush(struct request *rq)
386 struct request_queue *q = rq->q;
387 unsigned long fflags = q->queue_flags; /* may change, cache */
388 unsigned int policy = blk_flush_policy(fflags, rq);
389 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
392 * @policy now records what operations need to be done. Adjust
393 * REQ_PREFLUSH and FUA for the driver.
395 rq->cmd_flags &= ~REQ_PREFLUSH;
396 if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
397 rq->cmd_flags &= ~REQ_FUA;
400 * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
401 * of those flags, we have to set REQ_SYNC to avoid skewing
402 * the request accounting.
404 rq->cmd_flags |= REQ_SYNC;
407 * An empty flush handed down from a stacking driver may
408 * translate into nothing if the underlying device does not
409 * advertise a write-back cache. In this case, simply
410 * complete the request.
413 blk_mq_end_request(rq, 0);
417 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
420 * If there's data but flush is not necessary, the request can be
421 * processed directly without going through flush machinery. Queue
422 * for normal execution.
424 if ((policy & REQ_FSEQ_DATA) &&
425 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
426 blk_mq_request_bypass_insert(rq, false, false);
431 * @rq should go through flush machinery. Mark it part of flush
432 * sequence and submit for further processing.
434 memset(&rq->flush, 0, sizeof(rq->flush));
435 INIT_LIST_HEAD(&rq->flush.list);
436 rq->rq_flags |= RQF_FLUSH_SEQ;
437 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
439 rq->end_io = mq_flush_data_end_io;
441 spin_lock_irq(&fq->mq_flush_lock);
442 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
443 spin_unlock_irq(&fq->mq_flush_lock);
447 * blkdev_issue_flush - queue a flush
448 * @bdev: blockdev to issue flush for
451 * Issue a flush for the block device in question.
453 int blkdev_issue_flush(struct block_device *bdev)
457 bio_init(&bio, NULL, 0);
458 bio_set_dev(&bio, bdev);
459 bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
460 return submit_bio_wait(&bio);
462 EXPORT_SYMBOL(blkdev_issue_flush);
464 struct blk_flush_queue *blk_alloc_flush_queue(int node, int cmd_size,
467 struct blk_flush_queue *fq;
468 int rq_sz = sizeof(struct request);
470 fq = kzalloc_node(sizeof(*fq), flags, node);
474 spin_lock_init(&fq->mq_flush_lock);
476 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
477 fq->flush_rq = kzalloc_node(rq_sz, flags, node);
481 INIT_LIST_HEAD(&fq->flush_queue[0]);
482 INIT_LIST_HEAD(&fq->flush_queue[1]);
483 INIT_LIST_HEAD(&fq->flush_data_in_flight);
493 void blk_free_flush_queue(struct blk_flush_queue *fq)
495 /* bio based request queue hasn't flush queue */
504 * Allow driver to set its own lock class to fq->mq_flush_lock for
505 * avoiding lockdep complaint.
507 * flush_end_io() may be called recursively from some driver, such as
508 * nvme-loop, so lockdep may complain 'possible recursive locking' because
509 * all 'struct blk_flush_queue' instance share same mq_flush_lock lock class
510 * key. We need to assign different lock class for these driver's
511 * fq->mq_flush_lock for avoiding the lockdep warning.
513 * Use dynamically allocated lock class key for each 'blk_flush_queue'
514 * instance is over-kill, and more worse it introduces horrible boot delay
515 * issue because synchronize_rcu() is implied in lockdep_unregister_key which
516 * is called for each hctx release. SCSI probing may synchronously create and
517 * destroy lots of MQ request_queues for non-existent devices, and some robot
518 * test kernel always enable lockdep option. It is observed that more than half
519 * an hour is taken during SCSI MQ probe with per-fq lock class.
521 void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx,
522 struct lock_class_key *key)
524 lockdep_set_class(&hctx->fq->mq_flush_lock, key);
526 EXPORT_SYMBOL_GPL(blk_mq_hctx_set_fq_lock_class);