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);
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)
138 blk_mq_add_to_requeue_list(rq, add_front, true);
142 list_add(&rq->queuelist, &rq->q->queue_head);
144 list_add_tail(&rq->queuelist, &rq->q->queue_head);
150 * blk_flush_complete_seq - complete flush sequence
151 * @rq: PREFLUSH/FUA request being sequenced
153 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
154 * @error: whether an error occurred
156 * @rq just completed @seq part of its flush sequence, record the
157 * completion and trigger the next step.
160 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
163 * %true if requests were added to the dispatch queue, %false otherwise.
165 static bool blk_flush_complete_seq(struct request *rq,
166 struct blk_flush_queue *fq,
167 unsigned int seq, blk_status_t error)
169 struct request_queue *q = rq->q;
170 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
171 bool queued = false, kicked;
173 BUG_ON(rq->flush.seq & seq);
174 rq->flush.seq |= seq;
177 seq = blk_flush_cur_seq(rq);
182 case REQ_FSEQ_PREFLUSH:
183 case REQ_FSEQ_POSTFLUSH:
184 /* queue for flush */
185 if (list_empty(pending))
186 fq->flush_pending_since = jiffies;
187 list_move_tail(&rq->flush.list, pending);
191 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
192 queued = blk_flush_queue_rq(rq, true);
197 * @rq was previously adjusted by blk_flush_issue() for
198 * flush sequencing and may already have gone through the
199 * flush data request completion path. Restore @rq for
200 * normal completion and end it.
202 BUG_ON(!list_empty(&rq->queuelist));
203 list_del_init(&rq->flush.list);
204 blk_flush_restore_request(rq);
206 blk_mq_end_request(rq, error);
208 __blk_end_request_all(rq, error);
215 kicked = blk_kick_flush(q, fq);
216 return kicked | queued;
219 static void flush_end_io(struct request *flush_rq, blk_status_t error)
221 struct request_queue *q = flush_rq->q;
222 struct list_head *running;
224 struct request *rq, *n;
225 unsigned long flags = 0;
226 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
229 struct blk_mq_hw_ctx *hctx;
231 /* release the tag's ownership to the req cloned from */
232 spin_lock_irqsave(&fq->mq_flush_lock, flags);
233 hctx = blk_mq_map_queue(q, flush_rq->mq_ctx->cpu);
234 blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq);
238 running = &fq->flush_queue[fq->flush_running_idx];
239 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
241 /* account completion of the flush request */
242 fq->flush_running_idx ^= 1;
245 elv_completed_request(q, flush_rq);
247 /* and push the waiting requests to the next stage */
248 list_for_each_entry_safe(rq, n, running, flush.list) {
249 unsigned int seq = blk_flush_cur_seq(rq);
251 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
252 queued |= blk_flush_complete_seq(rq, fq, seq, error);
256 * Kick the queue to avoid stall for two cases:
257 * 1. Moving a request silently to empty queue_head may stall the
259 * 2. When flush request is running in non-queueable queue, the
260 * queue is hold. Restart the queue after flush request is finished
262 * This function is called from request completion path and calling
263 * directly into request_fn may confuse the driver. Always use
266 if (queued || fq->flush_queue_delayed) {
268 blk_run_queue_async(q);
270 fq->flush_queue_delayed = 0;
272 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
276 * blk_kick_flush - consider issuing flush request
277 * @q: request_queue being kicked
280 * Flush related states of @q have changed, consider issuing flush request.
281 * Please read the comment at the top of this file for more info.
284 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
287 * %true if flush was issued, %false otherwise.
289 static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq)
291 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
292 struct request *first_rq =
293 list_first_entry(pending, struct request, flush.list);
294 struct request *flush_rq = fq->flush_rq;
296 /* C1 described at the top of this file */
297 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
302 * For blk-mq + scheduling, we can risk having all driver tags
303 * assigned to empty flushes, and we deadlock if we are expecting
304 * other requests to make progress. Don't defer for that case.
306 if (!list_empty(&fq->flush_data_in_flight) &&
307 !(q->mq_ops && q->elevator) &&
309 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
313 * Issue flush and toggle pending_idx. This makes pending_idx
314 * different from running_idx, which means flush is in flight.
316 fq->flush_pending_idx ^= 1;
318 blk_rq_init(q, flush_rq);
321 * Borrow tag from the first request since they can't
322 * be in flight at the same time. And acquire the tag's
323 * ownership for flush req.
326 struct blk_mq_hw_ctx *hctx;
328 flush_rq->mq_ctx = first_rq->mq_ctx;
329 flush_rq->tag = first_rq->tag;
330 fq->orig_rq = first_rq;
332 hctx = blk_mq_map_queue(q, first_rq->mq_ctx->cpu);
333 blk_mq_tag_set_rq(hctx, first_rq->tag, flush_rq);
336 flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
337 flush_rq->rq_flags |= RQF_FLUSH_SEQ;
338 flush_rq->rq_disk = first_rq->rq_disk;
339 flush_rq->end_io = flush_end_io;
341 return blk_flush_queue_rq(flush_rq, false);
344 static void flush_data_end_io(struct request *rq, blk_status_t error)
346 struct request_queue *q = rq->q;
347 struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL);
349 lockdep_assert_held(q->queue_lock);
352 * Updating q->in_flight[] here for making this tag usable
353 * early. Because in blk_queue_start_tag(),
354 * q->in_flight[BLK_RW_ASYNC] is used to limit async I/O and
355 * reserve tags for sync I/O.
357 * More importantly this way can avoid the following I/O
360 * - suppose there are 40 fua requests comming to flush queue
361 * and queue depth is 31
362 * - 30 rqs are scheduled then blk_queue_start_tag() can't alloc
363 * tag for async I/O any more
364 * - all the 30 rqs are completed before FLUSH_PENDING_TIMEOUT
365 * and flush_data_end_io() is called
366 * - the other rqs still can't go ahead if not updating
367 * q->in_flight[BLK_RW_ASYNC] here, meantime these rqs
368 * are held in flush data queue and make no progress of
369 * handling post flush rq
370 * - only after the post flush rq is handled, all these rqs
374 elv_completed_request(q, rq);
376 /* for avoiding double accounting */
377 rq->rq_flags &= ~RQF_STARTED;
380 * After populating an empty queue, kick it to avoid stall. Read
381 * the comment in flush_end_io().
383 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
384 blk_run_queue_async(q);
387 static void mq_flush_data_end_io(struct request *rq, blk_status_t error)
389 struct request_queue *q = rq->q;
390 struct blk_mq_hw_ctx *hctx;
391 struct blk_mq_ctx *ctx = rq->mq_ctx;
393 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
395 hctx = blk_mq_map_queue(q, ctx->cpu);
398 * After populating an empty queue, kick it to avoid stall. Read
399 * the comment in flush_end_io().
401 spin_lock_irqsave(&fq->mq_flush_lock, flags);
402 blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
403 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
405 blk_mq_run_hw_queue(hctx, true);
409 * blk_insert_flush - insert a new PREFLUSH/FUA request
410 * @rq: request to insert
412 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
413 * or __blk_mq_run_hw_queue() to dispatch request.
414 * @rq is being submitted. Analyze what needs to be done and put it on the
417 void blk_insert_flush(struct request *rq)
419 struct request_queue *q = rq->q;
420 unsigned long fflags = q->queue_flags; /* may change, cache */
421 unsigned int policy = blk_flush_policy(fflags, rq);
422 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
425 lockdep_assert_held(q->queue_lock);
428 * @policy now records what operations need to be done. Adjust
429 * REQ_PREFLUSH and FUA for the driver.
431 rq->cmd_flags &= ~REQ_PREFLUSH;
432 if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
433 rq->cmd_flags &= ~REQ_FUA;
436 * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
437 * of those flags, we have to set REQ_SYNC to avoid skewing
438 * the request accounting.
440 rq->cmd_flags |= REQ_SYNC;
443 * An empty flush handed down from a stacking driver may
444 * translate into nothing if the underlying device does not
445 * advertise a write-back cache. In this case, simply
446 * complete the request.
450 blk_mq_end_request(rq, 0);
452 __blk_end_request(rq, 0, 0);
456 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
459 * If there's data but flush is not necessary, the request can be
460 * processed directly without going through flush machinery. Queue
461 * for normal execution.
463 if ((policy & REQ_FSEQ_DATA) &&
464 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
466 blk_mq_sched_insert_request(rq, false, true, false, false);
468 list_add_tail(&rq->queuelist, &q->queue_head);
473 * @rq should go through flush machinery. Mark it part of flush
474 * sequence and submit for further processing.
476 memset(&rq->flush, 0, sizeof(rq->flush));
477 INIT_LIST_HEAD(&rq->flush.list);
478 rq->rq_flags |= RQF_FLUSH_SEQ;
479 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
481 rq->end_io = mq_flush_data_end_io;
483 spin_lock_irq(&fq->mq_flush_lock);
484 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
485 spin_unlock_irq(&fq->mq_flush_lock);
488 rq->end_io = flush_data_end_io;
490 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
494 * blkdev_issue_flush - queue a flush
495 * @bdev: blockdev to issue flush for
496 * @gfp_mask: memory allocation flags (for bio_alloc)
497 * @error_sector: error sector
500 * Issue a flush for the block device in question. Caller can supply
501 * room for storing the error offset in case of a flush error, if they
504 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
505 sector_t *error_sector)
507 struct request_queue *q;
511 if (bdev->bd_disk == NULL)
514 q = bdev_get_queue(bdev);
519 * some block devices may not have their queue correctly set up here
520 * (e.g. loop device without a backing file) and so issuing a flush
521 * here will panic. Ensure there is a request function before issuing
524 if (!q->make_request_fn)
527 bio = bio_alloc(gfp_mask, 0);
528 bio_set_dev(bio, bdev);
529 bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
531 ret = submit_bio_wait(bio);
534 * The driver must store the error location in ->bi_sector, if
535 * it supports it. For non-stacked drivers, this should be
536 * copied from blk_rq_pos(rq).
539 *error_sector = bio->bi_iter.bi_sector;
544 EXPORT_SYMBOL(blkdev_issue_flush);
546 struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
547 int node, int cmd_size)
549 struct blk_flush_queue *fq;
550 int rq_sz = sizeof(struct request);
552 fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node);
557 spin_lock_init(&fq->mq_flush_lock);
559 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
560 fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node);
564 INIT_LIST_HEAD(&fq->flush_queue[0]);
565 INIT_LIST_HEAD(&fq->flush_queue[1]);
566 INIT_LIST_HEAD(&fq->flush_data_in_flight);
576 void blk_free_flush_queue(struct blk_flush_queue *fq)
578 /* bio based request queue hasn't flush queue */