2 * Functions to sequence PREFLUSH and FUA writes.
4 * Copyright (C) 2011 Max Planck Institute for Gravitational Physics
5 * Copyright (C) 2011 Tejun Heo <tj@kernel.org>
7 * This file is released under the GPLv2.
9 * REQ_{PREFLUSH|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.
13 * 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
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
19 * If the device doesn't have writeback cache, PREFLUSH and FUA don't make any
20 * difference. The requests are either completed immediately if there's no data
21 * or executed as normal requests otherwise.
23 * If the device has writeback cache and supports FUA, REQ_PREFLUSH is
24 * translated to PREFLUSH but REQ_FUA is passed down directly with DATA.
26 * If the device has writeback cache and doesn't support FUA, REQ_PREFLUSH
27 * is translated to PREFLUSH and REQ_FUA to POSTFLUSH.
29 * The actual execution of flush is double buffered. Whenever a request
30 * needs to execute PRE or POSTFLUSH, it queues at
31 * fq->flush_queue[fq->flush_pending_idx]. Once certain criteria are met, a
32 * REQ_OP_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 PREFLUSH/FUA
37 * Currently, the following conditions are used to determine when to issue
40 * C1. At any given time, only one flush shall be in progress. This makes
41 * double buffering sufficient.
43 * C2. Flush is deferred if any request is executing DATA of its sequence.
44 * This avoids issuing separate POSTFLUSHes for requests which shared
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 PREFLUSH) requests.
52 * For devices which support FUA, it isn't clear whether C2 (and thus C3)
55 * Note that a sequenced PREFLUSH/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 RQF_FLUSH_SEQ in
62 * The above peculiarity requires that each PREFLUSH/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.
67 #include <linux/kernel.h>
68 #include <linux/module.h>
69 #include <linux/bio.h>
70 #include <linux/blkdev.h>
71 #include <linux/gfp.h>
72 #include <linux/blk-mq.h>
76 #include "blk-mq-tag.h"
77 #include "blk-mq-sched.h"
79 /* PREFLUSH/FUA sequences */
81 REQ_FSEQ_PREFLUSH = (1 << 0), /* pre-flushing in progress */
82 REQ_FSEQ_DATA = (1 << 1), /* data write in progress */
83 REQ_FSEQ_POSTFLUSH = (1 << 2), /* post-flushing in progress */
84 REQ_FSEQ_DONE = (1 << 3),
86 REQ_FSEQ_ACTIONS = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
90 * If flush has been pending longer than the following timeout,
91 * it's issued even if flush_data requests are still in flight.
93 FLUSH_PENDING_TIMEOUT = 5 * HZ,
96 static bool blk_kick_flush(struct request_queue *q,
97 struct blk_flush_queue *fq, unsigned int flags);
99 static unsigned int blk_flush_policy(unsigned long fflags, struct request *rq)
101 unsigned int policy = 0;
103 if (blk_rq_sectors(rq))
104 policy |= REQ_FSEQ_DATA;
106 if (fflags & (1UL << QUEUE_FLAG_WC)) {
107 if (rq->cmd_flags & REQ_PREFLUSH)
108 policy |= REQ_FSEQ_PREFLUSH;
109 if (!(fflags & (1UL << QUEUE_FLAG_FUA)) &&
110 (rq->cmd_flags & REQ_FUA))
111 policy |= REQ_FSEQ_POSTFLUSH;
116 static unsigned int blk_flush_cur_seq(struct request *rq)
118 return 1 << ffz(rq->flush.seq);
121 static void blk_flush_restore_request(struct request *rq)
124 * After flush data completion, @rq->bio is %NULL but we need to
125 * complete the bio again. @rq->biotail is guaranteed to equal the
126 * original @rq->bio. Restore it.
128 rq->bio = rq->biotail;
130 /* make @rq a normal request */
131 rq->rq_flags &= ~RQF_FLUSH_SEQ;
132 rq->end_io = rq->flush.saved_end_io;
135 static bool blk_flush_queue_rq(struct request *rq, bool add_front)
137 blk_mq_add_to_requeue_list(rq, add_front, true);
142 * blk_flush_complete_seq - complete flush sequence
143 * @rq: PREFLUSH/FUA request being sequenced
145 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
146 * @error: whether an error occurred
148 * @rq just completed @seq part of its flush sequence, record the
149 * completion and trigger the next step.
152 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
155 * %true if requests were added to the dispatch queue, %false otherwise.
157 static bool blk_flush_complete_seq(struct request *rq,
158 struct blk_flush_queue *fq,
159 unsigned int seq, blk_status_t error)
161 struct request_queue *q = rq->q;
162 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
163 bool queued = false, kicked;
164 unsigned int cmd_flags;
166 BUG_ON(rq->flush.seq & seq);
167 rq->flush.seq |= seq;
168 cmd_flags = rq->cmd_flags;
171 seq = blk_flush_cur_seq(rq);
176 case REQ_FSEQ_PREFLUSH:
177 case REQ_FSEQ_POSTFLUSH:
178 /* queue for flush */
179 if (list_empty(pending))
180 fq->flush_pending_since = jiffies;
181 list_move_tail(&rq->flush.list, pending);
185 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
186 queued = blk_flush_queue_rq(rq, true);
191 * @rq was previously adjusted by blk_flush_issue() for
192 * flush sequencing and may already have gone through the
193 * flush data request completion path. Restore @rq for
194 * normal completion and end it.
196 BUG_ON(!list_empty(&rq->queuelist));
197 list_del_init(&rq->flush.list);
198 blk_flush_restore_request(rq);
199 blk_mq_end_request(rq, error);
206 kicked = blk_kick_flush(q, fq, cmd_flags);
207 return kicked | queued;
210 static void flush_end_io(struct request *flush_rq, blk_status_t error)
212 struct request_queue *q = flush_rq->q;
213 struct list_head *running;
215 struct request *rq, *n;
216 unsigned long flags = 0;
217 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
218 struct blk_mq_hw_ctx *hctx;
220 /* release the tag's ownership to the req cloned from */
221 spin_lock_irqsave(&fq->mq_flush_lock, flags);
222 hctx = blk_mq_map_queue(q, flush_rq->mq_ctx->cpu);
224 blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq);
227 blk_mq_put_driver_tag_hctx(hctx, flush_rq);
228 flush_rq->internal_tag = -1;
231 running = &fq->flush_queue[fq->flush_running_idx];
232 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
234 /* account completion of the flush request */
235 fq->flush_running_idx ^= 1;
237 /* and push the waiting requests to the next stage */
238 list_for_each_entry_safe(rq, n, running, flush.list) {
239 unsigned int seq = blk_flush_cur_seq(rq);
241 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
242 queued |= blk_flush_complete_seq(rq, fq, seq, error);
245 fq->flush_queue_delayed = 0;
246 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
250 * blk_kick_flush - consider issuing flush request
251 * @q: request_queue being kicked
253 * @flags: cmd_flags of the original request
255 * Flush related states of @q have changed, consider issuing flush request.
256 * Please read the comment at the top of this file for more info.
259 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
262 * %true if flush was issued, %false otherwise.
264 static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
267 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
268 struct request *first_rq =
269 list_first_entry(pending, struct request, flush.list);
270 struct request *flush_rq = fq->flush_rq;
271 struct blk_mq_hw_ctx *hctx;
273 /* C1 described at the top of this file */
274 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
279 * For blk-mq + scheduling, we can risk having all driver tags
280 * assigned to empty flushes, and we deadlock if we are expecting
281 * other requests to make progress. Don't defer for that case.
283 if (!list_empty(&fq->flush_data_in_flight) &&
284 !(q->mq_ops && q->elevator) &&
286 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
290 * Issue flush and toggle pending_idx. This makes pending_idx
291 * different from running_idx, which means flush is in flight.
293 fq->flush_pending_idx ^= 1;
295 blk_rq_init(q, flush_rq);
298 * In case of none scheduler, borrow tag from the first request
299 * since they can't be in flight at the same time. And acquire
300 * the tag's ownership for flush req.
302 * In case of IO scheduler, flush rq need to borrow scheduler tag
303 * just for cheating put/get driver tag.
305 flush_rq->mq_ctx = first_rq->mq_ctx;
308 fq->orig_rq = first_rq;
309 flush_rq->tag = first_rq->tag;
310 hctx = blk_mq_map_queue(q, first_rq->mq_ctx->cpu);
311 blk_mq_tag_set_rq(hctx, first_rq->tag, flush_rq);
313 flush_rq->internal_tag = first_rq->internal_tag;
316 flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
317 flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK);
318 flush_rq->rq_flags |= RQF_FLUSH_SEQ;
319 flush_rq->rq_disk = first_rq->rq_disk;
320 flush_rq->end_io = flush_end_io;
322 return blk_flush_queue_rq(flush_rq, false);
325 static void mq_flush_data_end_io(struct request *rq, blk_status_t error)
327 struct request_queue *q = rq->q;
328 struct blk_mq_hw_ctx *hctx;
329 struct blk_mq_ctx *ctx = rq->mq_ctx;
331 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
333 hctx = blk_mq_map_queue(q, ctx->cpu);
336 WARN_ON(rq->tag < 0);
337 blk_mq_put_driver_tag_hctx(hctx, rq);
341 * After populating an empty queue, kick it to avoid stall. Read
342 * the comment in flush_end_io().
344 spin_lock_irqsave(&fq->mq_flush_lock, flags);
345 blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
346 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
348 blk_mq_run_hw_queue(hctx, true);
352 * blk_insert_flush - insert a new PREFLUSH/FUA request
353 * @rq: request to insert
355 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
356 * or __blk_mq_run_hw_queue() to dispatch request.
357 * @rq is being submitted. Analyze what needs to be done and put it on the
360 void blk_insert_flush(struct request *rq)
362 struct request_queue *q = rq->q;
363 unsigned long fflags = q->queue_flags; /* may change, cache */
364 unsigned int policy = blk_flush_policy(fflags, rq);
365 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
368 * @policy now records what operations need to be done. Adjust
369 * REQ_PREFLUSH and FUA for the driver.
371 rq->cmd_flags &= ~REQ_PREFLUSH;
372 if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
373 rq->cmd_flags &= ~REQ_FUA;
376 * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
377 * of those flags, we have to set REQ_SYNC to avoid skewing
378 * the request accounting.
380 rq->cmd_flags |= REQ_SYNC;
383 * An empty flush handed down from a stacking driver may
384 * translate into nothing if the underlying device does not
385 * advertise a write-back cache. In this case, simply
386 * complete the request.
389 blk_mq_end_request(rq, 0);
393 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
396 * If there's data but flush is not necessary, the request can be
397 * processed directly without going through flush machinery. Queue
398 * for normal execution.
400 if ((policy & REQ_FSEQ_DATA) &&
401 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
402 blk_mq_request_bypass_insert(rq, false);
407 * @rq should go through flush machinery. Mark it part of flush
408 * sequence and submit for further processing.
410 memset(&rq->flush, 0, sizeof(rq->flush));
411 INIT_LIST_HEAD(&rq->flush.list);
412 rq->rq_flags |= RQF_FLUSH_SEQ;
413 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
415 rq->end_io = mq_flush_data_end_io;
417 spin_lock_irq(&fq->mq_flush_lock);
418 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
419 spin_unlock_irq(&fq->mq_flush_lock);
423 * blkdev_issue_flush - queue a flush
424 * @bdev: blockdev to issue flush for
425 * @gfp_mask: memory allocation flags (for bio_alloc)
426 * @error_sector: error sector
429 * Issue a flush for the block device in question. Caller can supply
430 * room for storing the error offset in case of a flush error, if they
433 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
434 sector_t *error_sector)
436 struct request_queue *q;
440 if (bdev->bd_disk == NULL)
443 q = bdev_get_queue(bdev);
448 * some block devices may not have their queue correctly set up here
449 * (e.g. loop device without a backing file) and so issuing a flush
450 * here will panic. Ensure there is a request function before issuing
453 if (!q->make_request_fn)
456 bio = bio_alloc(gfp_mask, 0);
457 bio_set_dev(bio, bdev);
458 bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
460 ret = submit_bio_wait(bio);
463 * The driver must store the error location in ->bi_sector, if
464 * it supports it. For non-stacked drivers, this should be
465 * copied from blk_rq_pos(rq).
468 *error_sector = bio->bi_iter.bi_sector;
473 EXPORT_SYMBOL(blkdev_issue_flush);
475 struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
476 int node, int cmd_size, gfp_t flags)
478 struct blk_flush_queue *fq;
479 int rq_sz = sizeof(struct request);
481 fq = kzalloc_node(sizeof(*fq), flags, node);
485 spin_lock_init(&fq->mq_flush_lock);
487 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
488 fq->flush_rq = kzalloc_node(rq_sz, flags, node);
492 INIT_LIST_HEAD(&fq->flush_queue[0]);
493 INIT_LIST_HEAD(&fq->flush_queue[1]);
494 INIT_LIST_HEAD(&fq->flush_data_in_flight);
504 void blk_free_flush_queue(struct blk_flush_queue *fq)
506 /* bio based request queue hasn't flush queue */