2 * Functions to sequence FLUSH 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_{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.
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
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.
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.
26 * If the device has writeback cache and doesn't support FUA, REQ_FLUSH is
27 * 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 * q->flush_queue[q->flush_pending_idx]. Once certain criteria are met, a
32 * 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
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 FLUSH) requests.
52 * For devices which support FUA, it isn't clear whether C2 (and thus C3)
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
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.
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>
77 /* FLUSH/FUA sequences */
79 REQ_FSEQ_PREFLUSH = (1 << 0), /* pre-flushing in progress */
80 REQ_FSEQ_DATA = (1 << 1), /* data write in progress */
81 REQ_FSEQ_POSTFLUSH = (1 << 2), /* post-flushing in progress */
82 REQ_FSEQ_DONE = (1 << 3),
84 REQ_FSEQ_ACTIONS = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
88 * If flush has been pending longer than the following timeout,
89 * it's issued even if flush_data requests are still in flight.
91 FLUSH_PENDING_TIMEOUT = 5 * HZ,
94 static bool blk_kick_flush(struct request_queue *q);
96 static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
98 unsigned int policy = 0;
100 if (blk_rq_sectors(rq))
101 policy |= REQ_FSEQ_DATA;
103 if (fflags & REQ_FLUSH) {
104 if (rq->cmd_flags & REQ_FLUSH)
105 policy |= REQ_FSEQ_PREFLUSH;
106 if (!(fflags & REQ_FUA) && (rq->cmd_flags & REQ_FUA))
107 policy |= REQ_FSEQ_POSTFLUSH;
112 static unsigned int blk_flush_cur_seq(struct request *rq)
114 return 1 << ffz(rq->flush.seq);
117 static void blk_flush_restore_request(struct request *rq)
120 * After flush data completion, @rq->bio is %NULL but we need to
121 * complete the bio again. @rq->biotail is guaranteed to equal the
122 * original @rq->bio. Restore it.
124 rq->bio = rq->biotail;
126 /* make @rq a normal request */
127 rq->cmd_flags &= ~REQ_FLUSH_SEQ;
128 rq->end_io = rq->flush.saved_end_io;
130 blk_clear_rq_complete(rq);
133 static void mq_flush_run(struct work_struct *work)
137 rq = container_of(work, struct request, mq_flush_work);
139 memset(&rq->csd, 0, sizeof(rq->csd));
140 blk_mq_insert_request(rq, false, true, false);
143 static bool blk_flush_queue_rq(struct request *rq)
146 INIT_WORK(&rq->mq_flush_work, mq_flush_run);
147 kblockd_schedule_work(rq->q, &rq->mq_flush_work);
150 list_add_tail(&rq->queuelist, &rq->q->queue_head);
156 * blk_flush_complete_seq - complete flush sequence
157 * @rq: FLUSH/FUA request being sequenced
158 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
159 * @error: whether an error occurred
161 * @rq just completed @seq part of its flush sequence, record the
162 * completion and trigger the next step.
165 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
168 * %true if requests were added to the dispatch queue, %false otherwise.
170 static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
173 struct request_queue *q = rq->q;
174 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
175 bool queued = false, kicked;
177 BUG_ON(rq->flush.seq & seq);
178 rq->flush.seq |= seq;
181 seq = blk_flush_cur_seq(rq);
186 case REQ_FSEQ_PREFLUSH:
187 case REQ_FSEQ_POSTFLUSH:
188 /* queue for flush */
189 if (list_empty(pending))
190 q->flush_pending_since = jiffies;
191 list_move_tail(&rq->flush.list, pending);
195 list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
196 queued = blk_flush_queue_rq(rq);
201 * @rq was previously adjusted by blk_flush_issue() for
202 * flush sequencing and may already have gone through the
203 * flush data request completion path. Restore @rq for
204 * normal completion and end it.
206 BUG_ON(!list_empty(&rq->queuelist));
207 list_del_init(&rq->flush.list);
208 blk_flush_restore_request(rq);
210 blk_mq_end_io(rq, error);
212 __blk_end_request_all(rq, error);
219 kicked = blk_kick_flush(q);
220 return kicked | queued;
223 static void flush_end_io(struct request *flush_rq, int error)
225 struct request_queue *q = flush_rq->q;
226 struct list_head *running;
228 struct request *rq, *n;
229 unsigned long flags = 0;
232 spin_lock_irqsave(&q->mq_flush_lock, flags);
234 running = &q->flush_queue[q->flush_running_idx];
235 BUG_ON(q->flush_pending_idx == q->flush_running_idx);
237 /* account completion of the flush request */
238 q->flush_running_idx ^= 1;
241 elv_completed_request(q, flush_rq);
243 /* and push the waiting requests to the next stage */
244 list_for_each_entry_safe(rq, n, running, flush.list) {
245 unsigned int seq = blk_flush_cur_seq(rq);
247 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
248 queued |= blk_flush_complete_seq(rq, seq, error);
252 * Kick the queue to avoid stall for two cases:
253 * 1. Moving a request silently to empty queue_head may stall the
255 * 2. When flush request is running in non-queueable queue, the
256 * queue is hold. Restart the queue after flush request is finished
258 * This function is called from request completion path and calling
259 * directly into request_fn may confuse the driver. Always use
262 if (queued || q->flush_queue_delayed) {
264 blk_run_queue_async(q);
266 q->flush_queue_delayed = 0;
268 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
272 * blk_kick_flush - consider issuing flush request
273 * @q: request_queue being kicked
275 * Flush related states of @q have changed, consider issuing flush request.
276 * Please read the comment at the top of this file for more info.
279 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
282 * %true if flush was issued, %false otherwise.
284 static bool blk_kick_flush(struct request_queue *q)
286 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
287 struct request *first_rq =
288 list_first_entry(pending, struct request, flush.list);
290 /* C1 described at the top of this file */
291 if (q->flush_pending_idx != q->flush_running_idx || list_empty(pending))
295 if (!list_empty(&q->flush_data_in_flight) &&
297 q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
301 * Issue flush and toggle pending_idx. This makes pending_idx
302 * different from running_idx, which means flush is in flight.
304 q->flush_pending_idx ^= 1;
307 struct blk_mq_ctx *ctx = first_rq->mq_ctx;
308 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
310 blk_mq_rq_init(hctx, q->flush_rq);
311 q->flush_rq->mq_ctx = ctx;
314 * Reuse the tag value from the fist waiting request,
315 * with blk-mq the tag is generated during request
316 * allocation and drivers can rely on it being inside
317 * the range they asked for.
319 q->flush_rq->tag = first_rq->tag;
321 blk_rq_init(q, q->flush_rq);
324 q->flush_rq->cmd_type = REQ_TYPE_FS;
325 q->flush_rq->cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
326 q->flush_rq->rq_disk = first_rq->rq_disk;
327 q->flush_rq->end_io = flush_end_io;
329 return blk_flush_queue_rq(q->flush_rq);
332 static void flush_data_end_io(struct request *rq, int error)
334 struct request_queue *q = rq->q;
337 * After populating an empty queue, kick it to avoid stall. Read
338 * the comment in flush_end_io().
340 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
341 blk_run_queue_async(q);
344 static void mq_flush_data_end_io(struct request *rq, int error)
346 struct request_queue *q = rq->q;
347 struct blk_mq_hw_ctx *hctx;
348 struct blk_mq_ctx *ctx;
352 hctx = q->mq_ops->map_queue(q, ctx->cpu);
355 * After populating an empty queue, kick it to avoid stall. Read
356 * the comment in flush_end_io().
358 spin_lock_irqsave(&q->mq_flush_lock, flags);
359 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
360 blk_mq_run_hw_queue(hctx, true);
361 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
365 * blk_insert_flush - insert a new FLUSH/FUA request
366 * @rq: request to insert
368 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
369 * or __blk_mq_run_hw_queue() to dispatch request.
370 * @rq is being submitted. Analyze what needs to be done and put it on the
374 * spin_lock_irq(q->queue_lock) in !mq case
376 void blk_insert_flush(struct request *rq)
378 struct request_queue *q = rq->q;
379 unsigned int fflags = q->flush_flags; /* may change, cache */
380 unsigned int policy = blk_flush_policy(fflags, rq);
383 * @policy now records what operations need to be done. Adjust
384 * REQ_FLUSH and FUA for the driver.
386 rq->cmd_flags &= ~REQ_FLUSH;
387 if (!(fflags & REQ_FUA))
388 rq->cmd_flags &= ~REQ_FUA;
391 * An empty flush handed down from a stacking driver may
392 * translate into nothing if the underlying device does not
393 * advertise a write-back cache. In this case, simply
394 * complete the request.
398 blk_mq_end_io(rq, 0);
400 __blk_end_bidi_request(rq, 0, 0, 0);
404 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
407 * If there's data but flush is not necessary, the request can be
408 * processed directly without going through flush machinery. Queue
409 * for normal execution.
411 if ((policy & REQ_FSEQ_DATA) &&
412 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
414 blk_mq_insert_request(rq, false, false, true);
416 list_add_tail(&rq->queuelist, &q->queue_head);
421 * @rq should go through flush machinery. Mark it part of flush
422 * sequence and submit for further processing.
424 memset(&rq->flush, 0, sizeof(rq->flush));
425 INIT_LIST_HEAD(&rq->flush.list);
426 rq->cmd_flags |= REQ_FLUSH_SEQ;
427 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
429 rq->end_io = mq_flush_data_end_io;
431 spin_lock_irq(&q->mq_flush_lock);
432 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
433 spin_unlock_irq(&q->mq_flush_lock);
436 rq->end_io = flush_data_end_io;
438 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
442 * blk_abort_flushes - @q is being aborted, abort flush requests
443 * @q: request_queue being aborted
445 * To be called from elv_abort_queue(). @q is being aborted. Prepare all
446 * FLUSH/FUA requests for abortion.
449 * spin_lock_irq(q->queue_lock)
451 void blk_abort_flushes(struct request_queue *q)
453 struct request *rq, *n;
457 * Requests in flight for data are already owned by the dispatch
458 * queue or the device driver. Just restore for normal completion.
460 list_for_each_entry_safe(rq, n, &q->flush_data_in_flight, flush.list) {
461 list_del_init(&rq->flush.list);
462 blk_flush_restore_request(rq);
466 * We need to give away requests on flush queues. Restore for
467 * normal completion and put them on the dispatch queue.
469 for (i = 0; i < ARRAY_SIZE(q->flush_queue); i++) {
470 list_for_each_entry_safe(rq, n, &q->flush_queue[i],
472 list_del_init(&rq->flush.list);
473 blk_flush_restore_request(rq);
474 list_add_tail(&rq->queuelist, &q->queue_head);
480 * blkdev_issue_flush - queue a flush
481 * @bdev: blockdev to issue flush for
482 * @gfp_mask: memory allocation flags (for bio_alloc)
483 * @error_sector: error sector
486 * Issue a flush for the block device in question. Caller can supply
487 * room for storing the error offset in case of a flush error, if they
488 * wish to. If WAIT flag is not passed then caller may check only what
489 * request was pushed in some internal queue for later handling.
491 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
492 sector_t *error_sector)
494 struct request_queue *q;
498 if (bdev->bd_disk == NULL)
501 q = bdev_get_queue(bdev);
506 * some block devices may not have their queue correctly set up here
507 * (e.g. loop device without a backing file) and so issuing a flush
508 * here will panic. Ensure there is a request function before issuing
511 if (!q->make_request_fn)
514 bio = bio_alloc(gfp_mask, 0);
517 ret = submit_bio_wait(WRITE_FLUSH, bio);
520 * The driver must store the error location in ->bi_sector, if
521 * it supports it. For non-stacked drivers, this should be
522 * copied from blk_rq_pos(rq).
525 *error_sector = bio->bi_iter.bi_sector;
530 EXPORT_SYMBOL(blkdev_issue_flush);
532 void blk_mq_init_flush(struct request_queue *q)
534 spin_lock_init(&q->mq_flush_lock);