Loading...
1/*
2 * Functions to sequence PREFLUSH and FUA writes.
3 *
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_{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.
12 *
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
17 * completion.
18 *
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.
22 *
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.
25 *
26 * If the device has writeback cache and doesn't support FUA, REQ_PREFLUSH
27 * is translated to PREFLUSH and REQ_FUA to POSTFLUSH.
28 *
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
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 PREFLUSH) 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 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
60 * req_bio_endio().
61 *
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.
65 */
66
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>
73
74#include "blk.h"
75#include "blk-mq.h"
76#include "blk-mq-tag.h"
77#include "blk-mq-sched.h"
78
79/* PREFLUSH/FUA sequences */
80enum {
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),
85
86 REQ_FSEQ_ACTIONS = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
87 REQ_FSEQ_POSTFLUSH,
88
89 /*
90 * If flush has been pending longer than the following timeout,
91 * it's issued even if flush_data requests are still in flight.
92 */
93 FLUSH_PENDING_TIMEOUT = 5 * HZ,
94};
95
96static bool blk_kick_flush(struct request_queue *q,
97 struct blk_flush_queue *fq);
98
99static unsigned int blk_flush_policy(unsigned long fflags, struct request *rq)
100{
101 unsigned int policy = 0;
102
103 if (blk_rq_sectors(rq))
104 policy |= REQ_FSEQ_DATA;
105
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;
112 }
113 return policy;
114}
115
116static unsigned int blk_flush_cur_seq(struct request *rq)
117{
118 return 1 << ffz(rq->flush.seq);
119}
120
121static void blk_flush_restore_request(struct request *rq)
122{
123 /*
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.
127 */
128 rq->bio = rq->biotail;
129
130 /* make @rq a normal request */
131 rq->rq_flags &= ~RQF_FLUSH_SEQ;
132 rq->end_io = rq->flush.saved_end_io;
133}
134
135static bool blk_flush_queue_rq(struct request *rq, bool add_front)
136{
137 if (rq->q->mq_ops) {
138 blk_mq_add_to_requeue_list(rq, add_front, true);
139 return false;
140 } else {
141 if (add_front)
142 list_add(&rq->queuelist, &rq->q->queue_head);
143 else
144 list_add_tail(&rq->queuelist, &rq->q->queue_head);
145 return true;
146 }
147}
148
149/**
150 * blk_flush_complete_seq - complete flush sequence
151 * @rq: PREFLUSH/FUA request being sequenced
152 * @fq: flush queue
153 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
154 * @error: whether an error occurred
155 *
156 * @rq just completed @seq part of its flush sequence, record the
157 * completion and trigger the next step.
158 *
159 * CONTEXT:
160 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
161 *
162 * RETURNS:
163 * %true if requests were added to the dispatch queue, %false otherwise.
164 */
165static bool blk_flush_complete_seq(struct request *rq,
166 struct blk_flush_queue *fq,
167 unsigned int seq, blk_status_t error)
168{
169 struct request_queue *q = rq->q;
170 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
171 bool queued = false, kicked;
172
173 BUG_ON(rq->flush.seq & seq);
174 rq->flush.seq |= seq;
175
176 if (likely(!error))
177 seq = blk_flush_cur_seq(rq);
178 else
179 seq = REQ_FSEQ_DONE;
180
181 switch (seq) {
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);
188 break;
189
190 case REQ_FSEQ_DATA:
191 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
192 queued = blk_flush_queue_rq(rq, true);
193 break;
194
195 case REQ_FSEQ_DONE:
196 /*
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.
201 */
202 BUG_ON(!list_empty(&rq->queuelist));
203 list_del_init(&rq->flush.list);
204 blk_flush_restore_request(rq);
205 if (q->mq_ops)
206 blk_mq_end_request(rq, error);
207 else
208 __blk_end_request_all(rq, error);
209 break;
210
211 default:
212 BUG();
213 }
214
215 kicked = blk_kick_flush(q, fq);
216 return kicked | queued;
217}
218
219static void flush_end_io(struct request *flush_rq, blk_status_t error)
220{
221 struct request_queue *q = flush_rq->q;
222 struct list_head *running;
223 bool queued = false;
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);
227
228 if (q->mq_ops) {
229 struct blk_mq_hw_ctx *hctx;
230
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 if (!q->elevator) {
235 blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq);
236 flush_rq->tag = -1;
237 } else {
238 blk_mq_put_driver_tag_hctx(hctx, flush_rq);
239 flush_rq->internal_tag = -1;
240 }
241 }
242
243 running = &fq->flush_queue[fq->flush_running_idx];
244 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
245
246 /* account completion of the flush request */
247 fq->flush_running_idx ^= 1;
248
249 if (!q->mq_ops)
250 elv_completed_request(q, flush_rq);
251
252 /* and push the waiting requests to the next stage */
253 list_for_each_entry_safe(rq, n, running, flush.list) {
254 unsigned int seq = blk_flush_cur_seq(rq);
255
256 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
257 queued |= blk_flush_complete_seq(rq, fq, seq, error);
258 }
259
260 /*
261 * Kick the queue to avoid stall for two cases:
262 * 1. Moving a request silently to empty queue_head may stall the
263 * queue.
264 * 2. When flush request is running in non-queueable queue, the
265 * queue is hold. Restart the queue after flush request is finished
266 * to avoid stall.
267 * This function is called from request completion path and calling
268 * directly into request_fn may confuse the driver. Always use
269 * kblockd.
270 */
271 if (queued || fq->flush_queue_delayed) {
272 WARN_ON(q->mq_ops);
273 blk_run_queue_async(q);
274 }
275 fq->flush_queue_delayed = 0;
276 if (q->mq_ops)
277 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
278}
279
280/**
281 * blk_kick_flush - consider issuing flush request
282 * @q: request_queue being kicked
283 * @fq: flush queue
284 *
285 * Flush related states of @q have changed, consider issuing flush request.
286 * Please read the comment at the top of this file for more info.
287 *
288 * CONTEXT:
289 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
290 *
291 * RETURNS:
292 * %true if flush was issued, %false otherwise.
293 */
294static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq)
295{
296 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
297 struct request *first_rq =
298 list_first_entry(pending, struct request, flush.list);
299 struct request *flush_rq = fq->flush_rq;
300
301 /* C1 described at the top of this file */
302 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
303 return false;
304
305 /* C2 and C3
306 *
307 * For blk-mq + scheduling, we can risk having all driver tags
308 * assigned to empty flushes, and we deadlock if we are expecting
309 * other requests to make progress. Don't defer for that case.
310 */
311 if (!list_empty(&fq->flush_data_in_flight) &&
312 !(q->mq_ops && q->elevator) &&
313 time_before(jiffies,
314 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
315 return false;
316
317 /*
318 * Issue flush and toggle pending_idx. This makes pending_idx
319 * different from running_idx, which means flush is in flight.
320 */
321 fq->flush_pending_idx ^= 1;
322
323 blk_rq_init(q, flush_rq);
324
325 /*
326 * In case of none scheduler, borrow tag from the first request
327 * since they can't be in flight at the same time. And acquire
328 * the tag's ownership for flush req.
329 *
330 * In case of IO scheduler, flush rq need to borrow scheduler tag
331 * just for cheating put/get driver tag.
332 */
333 if (q->mq_ops) {
334 struct blk_mq_hw_ctx *hctx;
335
336 flush_rq->mq_ctx = first_rq->mq_ctx;
337
338 if (!q->elevator) {
339 fq->orig_rq = first_rq;
340 flush_rq->tag = first_rq->tag;
341 hctx = blk_mq_map_queue(q, first_rq->mq_ctx->cpu);
342 blk_mq_tag_set_rq(hctx, first_rq->tag, flush_rq);
343 } else {
344 flush_rq->internal_tag = first_rq->internal_tag;
345 }
346 }
347
348 flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
349 flush_rq->rq_flags |= RQF_FLUSH_SEQ;
350 flush_rq->rq_disk = first_rq->rq_disk;
351 flush_rq->end_io = flush_end_io;
352
353 return blk_flush_queue_rq(flush_rq, false);
354}
355
356static void flush_data_end_io(struct request *rq, blk_status_t error)
357{
358 struct request_queue *q = rq->q;
359 struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL);
360
361 lockdep_assert_held(q->queue_lock);
362
363 /*
364 * Updating q->in_flight[] here for making this tag usable
365 * early. Because in blk_queue_start_tag(),
366 * q->in_flight[BLK_RW_ASYNC] is used to limit async I/O and
367 * reserve tags for sync I/O.
368 *
369 * More importantly this way can avoid the following I/O
370 * deadlock:
371 *
372 * - suppose there are 40 fua requests comming to flush queue
373 * and queue depth is 31
374 * - 30 rqs are scheduled then blk_queue_start_tag() can't alloc
375 * tag for async I/O any more
376 * - all the 30 rqs are completed before FLUSH_PENDING_TIMEOUT
377 * and flush_data_end_io() is called
378 * - the other rqs still can't go ahead if not updating
379 * q->in_flight[BLK_RW_ASYNC] here, meantime these rqs
380 * are held in flush data queue and make no progress of
381 * handling post flush rq
382 * - only after the post flush rq is handled, all these rqs
383 * can be completed
384 */
385
386 elv_completed_request(q, rq);
387
388 /* for avoiding double accounting */
389 rq->rq_flags &= ~RQF_STARTED;
390
391 /*
392 * After populating an empty queue, kick it to avoid stall. Read
393 * the comment in flush_end_io().
394 */
395 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
396 blk_run_queue_async(q);
397}
398
399static void mq_flush_data_end_io(struct request *rq, blk_status_t error)
400{
401 struct request_queue *q = rq->q;
402 struct blk_mq_hw_ctx *hctx;
403 struct blk_mq_ctx *ctx = rq->mq_ctx;
404 unsigned long flags;
405 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
406
407 hctx = blk_mq_map_queue(q, ctx->cpu);
408
409 if (q->elevator) {
410 WARN_ON(rq->tag < 0);
411 blk_mq_put_driver_tag_hctx(hctx, rq);
412 }
413
414 /*
415 * After populating an empty queue, kick it to avoid stall. Read
416 * the comment in flush_end_io().
417 */
418 spin_lock_irqsave(&fq->mq_flush_lock, flags);
419 blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
420 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
421
422 blk_mq_run_hw_queue(hctx, true);
423}
424
425/**
426 * blk_insert_flush - insert a new PREFLUSH/FUA request
427 * @rq: request to insert
428 *
429 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
430 * or __blk_mq_run_hw_queue() to dispatch request.
431 * @rq is being submitted. Analyze what needs to be done and put it on the
432 * right queue.
433 */
434void blk_insert_flush(struct request *rq)
435{
436 struct request_queue *q = rq->q;
437 unsigned long fflags = q->queue_flags; /* may change, cache */
438 unsigned int policy = blk_flush_policy(fflags, rq);
439 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
440
441 if (!q->mq_ops)
442 lockdep_assert_held(q->queue_lock);
443
444 /*
445 * @policy now records what operations need to be done. Adjust
446 * REQ_PREFLUSH and FUA for the driver.
447 */
448 rq->cmd_flags &= ~REQ_PREFLUSH;
449 if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
450 rq->cmd_flags &= ~REQ_FUA;
451
452 /*
453 * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
454 * of those flags, we have to set REQ_SYNC to avoid skewing
455 * the request accounting.
456 */
457 rq->cmd_flags |= REQ_SYNC;
458
459 /*
460 * An empty flush handed down from a stacking driver may
461 * translate into nothing if the underlying device does not
462 * advertise a write-back cache. In this case, simply
463 * complete the request.
464 */
465 if (!policy) {
466 if (q->mq_ops)
467 blk_mq_end_request(rq, 0);
468 else
469 __blk_end_request(rq, 0, 0);
470 return;
471 }
472
473 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
474
475 /*
476 * If there's data but flush is not necessary, the request can be
477 * processed directly without going through flush machinery. Queue
478 * for normal execution.
479 */
480 if ((policy & REQ_FSEQ_DATA) &&
481 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
482 if (q->mq_ops)
483 blk_mq_request_bypass_insert(rq, false);
484 else
485 list_add_tail(&rq->queuelist, &q->queue_head);
486 return;
487 }
488
489 /*
490 * @rq should go through flush machinery. Mark it part of flush
491 * sequence and submit for further processing.
492 */
493 memset(&rq->flush, 0, sizeof(rq->flush));
494 INIT_LIST_HEAD(&rq->flush.list);
495 rq->rq_flags |= RQF_FLUSH_SEQ;
496 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
497 if (q->mq_ops) {
498 rq->end_io = mq_flush_data_end_io;
499
500 spin_lock_irq(&fq->mq_flush_lock);
501 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
502 spin_unlock_irq(&fq->mq_flush_lock);
503 return;
504 }
505 rq->end_io = flush_data_end_io;
506
507 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
508}
509
510/**
511 * blkdev_issue_flush - queue a flush
512 * @bdev: blockdev to issue flush for
513 * @gfp_mask: memory allocation flags (for bio_alloc)
514 * @error_sector: error sector
515 *
516 * Description:
517 * Issue a flush for the block device in question. Caller can supply
518 * room for storing the error offset in case of a flush error, if they
519 * wish to.
520 */
521int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
522 sector_t *error_sector)
523{
524 struct request_queue *q;
525 struct bio *bio;
526 int ret = 0;
527
528 if (bdev->bd_disk == NULL)
529 return -ENXIO;
530
531 q = bdev_get_queue(bdev);
532 if (!q)
533 return -ENXIO;
534
535 /*
536 * some block devices may not have their queue correctly set up here
537 * (e.g. loop device without a backing file) and so issuing a flush
538 * here will panic. Ensure there is a request function before issuing
539 * the flush.
540 */
541 if (!q->make_request_fn)
542 return -ENXIO;
543
544 bio = bio_alloc(gfp_mask, 0);
545 bio_set_dev(bio, bdev);
546 bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
547
548 ret = submit_bio_wait(bio);
549
550 /*
551 * The driver must store the error location in ->bi_sector, if
552 * it supports it. For non-stacked drivers, this should be
553 * copied from blk_rq_pos(rq).
554 */
555 if (error_sector)
556 *error_sector = bio->bi_iter.bi_sector;
557
558 bio_put(bio);
559 return ret;
560}
561EXPORT_SYMBOL(blkdev_issue_flush);
562
563struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
564 int node, int cmd_size)
565{
566 struct blk_flush_queue *fq;
567 int rq_sz = sizeof(struct request);
568
569 fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node);
570 if (!fq)
571 goto fail;
572
573 if (q->mq_ops)
574 spin_lock_init(&fq->mq_flush_lock);
575
576 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
577 fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node);
578 if (!fq->flush_rq)
579 goto fail_rq;
580
581 INIT_LIST_HEAD(&fq->flush_queue[0]);
582 INIT_LIST_HEAD(&fq->flush_queue[1]);
583 INIT_LIST_HEAD(&fq->flush_data_in_flight);
584
585 return fq;
586
587 fail_rq:
588 kfree(fq);
589 fail:
590 return NULL;
591}
592
593void blk_free_flush_queue(struct blk_flush_queue *fq)
594{
595 /* bio based request queue hasn't flush queue */
596 if (!fq)
597 return;
598
599 kfree(fq->flush_rq);
600 kfree(fq);
601}
1/*
2 * Functions to sequence FLUSH and FUA writes.
3 *
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 *
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
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 *
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.
25 *
26 * If the device has writeback cache and doesn't support FUA, REQ_FLUSH is
27 * translated to PREFLUSH and REQ_FUA to POSTFLUSH.
28 *
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
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
59 * sequence is complete. This is implemented by testing REQ_FLUSH_SEQ in
60 * 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.
65 */
66
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>
73
74#include "blk.h"
75#include "blk-mq.h"
76
77/* FLUSH/FUA sequences */
78enum {
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),
83
84 REQ_FSEQ_ACTIONS = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
85 REQ_FSEQ_POSTFLUSH,
86
87 /*
88 * If flush has been pending longer than the following timeout,
89 * it's issued even if flush_data requests are still in flight.
90 */
91 FLUSH_PENDING_TIMEOUT = 5 * HZ,
92};
93
94static bool blk_kick_flush(struct request_queue *q);
95
96static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
97{
98 unsigned int policy = 0;
99
100 if (blk_rq_sectors(rq))
101 policy |= REQ_FSEQ_DATA;
102
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;
108 }
109 return policy;
110}
111
112static unsigned int blk_flush_cur_seq(struct request *rq)
113{
114 return 1 << ffz(rq->flush.seq);
115}
116
117static void blk_flush_restore_request(struct request *rq)
118{
119 /*
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.
123 */
124 rq->bio = rq->biotail;
125
126 /* make @rq a normal request */
127 rq->cmd_flags &= ~REQ_FLUSH_SEQ;
128 rq->end_io = rq->flush.saved_end_io;
129
130 blk_clear_rq_complete(rq);
131}
132
133static void mq_flush_run(struct work_struct *work)
134{
135 struct request *rq;
136
137 rq = container_of(work, struct request, mq_flush_work);
138
139 memset(&rq->csd, 0, sizeof(rq->csd));
140 blk_mq_insert_request(rq, false, true, false);
141}
142
143static bool blk_flush_queue_rq(struct request *rq, bool add_front)
144{
145 if (rq->q->mq_ops) {
146 INIT_WORK(&rq->mq_flush_work, mq_flush_run);
147 kblockd_schedule_work(rq->q, &rq->mq_flush_work);
148 return false;
149 } else {
150 if (add_front)
151 list_add(&rq->queuelist, &rq->q->queue_head);
152 else
153 list_add_tail(&rq->queuelist, &rq->q->queue_head);
154 return true;
155 }
156}
157
158/**
159 * blk_flush_complete_seq - complete flush sequence
160 * @rq: FLUSH/FUA request being sequenced
161 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
162 * @error: whether an error occurred
163 *
164 * @rq just completed @seq part of its flush sequence, record the
165 * completion and trigger the next step.
166 *
167 * CONTEXT:
168 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
169 *
170 * RETURNS:
171 * %true if requests were added to the dispatch queue, %false otherwise.
172 */
173static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
174 int error)
175{
176 struct request_queue *q = rq->q;
177 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
178 bool queued = false, kicked;
179
180 BUG_ON(rq->flush.seq & seq);
181 rq->flush.seq |= seq;
182
183 if (likely(!error))
184 seq = blk_flush_cur_seq(rq);
185 else
186 seq = REQ_FSEQ_DONE;
187
188 switch (seq) {
189 case REQ_FSEQ_PREFLUSH:
190 case REQ_FSEQ_POSTFLUSH:
191 /* queue for flush */
192 if (list_empty(pending))
193 q->flush_pending_since = jiffies;
194 list_move_tail(&rq->flush.list, pending);
195 break;
196
197 case REQ_FSEQ_DATA:
198 list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
199 queued = blk_flush_queue_rq(rq, true);
200 break;
201
202 case REQ_FSEQ_DONE:
203 /*
204 * @rq was previously adjusted by blk_flush_issue() for
205 * flush sequencing and may already have gone through the
206 * flush data request completion path. Restore @rq for
207 * normal completion and end it.
208 */
209 BUG_ON(!list_empty(&rq->queuelist));
210 list_del_init(&rq->flush.list);
211 blk_flush_restore_request(rq);
212 if (q->mq_ops)
213 blk_mq_end_io(rq, error);
214 else
215 __blk_end_request_all(rq, error);
216 break;
217
218 default:
219 BUG();
220 }
221
222 kicked = blk_kick_flush(q);
223 return kicked | queued;
224}
225
226static void flush_end_io(struct request *flush_rq, int error)
227{
228 struct request_queue *q = flush_rq->q;
229 struct list_head *running;
230 bool queued = false;
231 struct request *rq, *n;
232 unsigned long flags = 0;
233
234 if (q->mq_ops)
235 spin_lock_irqsave(&q->mq_flush_lock, flags);
236
237 running = &q->flush_queue[q->flush_running_idx];
238 BUG_ON(q->flush_pending_idx == q->flush_running_idx);
239
240 /* account completion of the flush request */
241 q->flush_running_idx ^= 1;
242
243 if (!q->mq_ops)
244 elv_completed_request(q, flush_rq);
245
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);
251 queued |= blk_flush_complete_seq(rq, seq, error);
252 }
253
254 /*
255 * 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.
264 */
265 if (queued || q->flush_queue_delayed) {
266 WARN_ON(q->mq_ops);
267 blk_run_queue_async(q);
268 }
269 q->flush_queue_delayed = 0;
270 if (q->mq_ops)
271 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
272}
273
274/**
275 * blk_kick_flush - consider issuing flush request
276 * @q: request_queue being kicked
277 *
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.
280 *
281 * CONTEXT:
282 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
283 *
284 * RETURNS:
285 * %true if flush was issued, %false otherwise.
286 */
287static bool blk_kick_flush(struct request_queue *q)
288{
289 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
290 struct request *first_rq =
291 list_first_entry(pending, struct request, flush.list);
292
293 /* C1 described at the top of this file */
294 if (q->flush_pending_idx != q->flush_running_idx || list_empty(pending))
295 return false;
296
297 /* C2 and C3 */
298 if (!list_empty(&q->flush_data_in_flight) &&
299 time_before(jiffies,
300 q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
301 return false;
302
303 /*
304 * Issue flush and toggle pending_idx. This makes pending_idx
305 * different from running_idx, which means flush is in flight.
306 */
307 q->flush_pending_idx ^= 1;
308
309 if (q->mq_ops) {
310 struct blk_mq_ctx *ctx = first_rq->mq_ctx;
311 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
312
313 blk_mq_rq_init(hctx, q->flush_rq);
314 q->flush_rq->mq_ctx = ctx;
315
316 /*
317 * Reuse the tag value from the fist waiting request,
318 * with blk-mq the tag is generated during request
319 * allocation and drivers can rely on it being inside
320 * the range they asked for.
321 */
322 q->flush_rq->tag = first_rq->tag;
323 } else {
324 blk_rq_init(q, q->flush_rq);
325 }
326
327 q->flush_rq->cmd_type = REQ_TYPE_FS;
328 q->flush_rq->cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
329 q->flush_rq->rq_disk = first_rq->rq_disk;
330 q->flush_rq->end_io = flush_end_io;
331
332 return blk_flush_queue_rq(q->flush_rq, false);
333}
334
335static void flush_data_end_io(struct request *rq, int error)
336{
337 struct request_queue *q = rq->q;
338
339 /*
340 * After populating an empty queue, kick it to avoid stall. Read
341 * the comment in flush_end_io().
342 */
343 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
344 blk_run_queue_async(q);
345}
346
347static void mq_flush_data_end_io(struct request *rq, int error)
348{
349 struct request_queue *q = rq->q;
350 struct blk_mq_hw_ctx *hctx;
351 struct blk_mq_ctx *ctx;
352 unsigned long flags;
353
354 ctx = rq->mq_ctx;
355 hctx = q->mq_ops->map_queue(q, ctx->cpu);
356
357 /*
358 * After populating an empty queue, kick it to avoid stall. Read
359 * the comment in flush_end_io().
360 */
361 spin_lock_irqsave(&q->mq_flush_lock, flags);
362 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
363 blk_mq_run_hw_queue(hctx, true);
364 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
365}
366
367/**
368 * blk_insert_flush - insert a new FLUSH/FUA request
369 * @rq: request to insert
370 *
371 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
372 * or __blk_mq_run_hw_queue() to dispatch request.
373 * @rq is being submitted. Analyze what needs to be done and put it on the
374 * right queue.
375 *
376 * CONTEXT:
377 * spin_lock_irq(q->queue_lock) in !mq case
378 */
379void blk_insert_flush(struct request *rq)
380{
381 struct request_queue *q = rq->q;
382 unsigned int fflags = q->flush_flags; /* may change, cache */
383 unsigned int policy = blk_flush_policy(fflags, rq);
384
385 /*
386 * @policy now records what operations need to be done. Adjust
387 * REQ_FLUSH and FUA for the driver.
388 */
389 rq->cmd_flags &= ~REQ_FLUSH;
390 if (!(fflags & REQ_FUA))
391 rq->cmd_flags &= ~REQ_FUA;
392
393 /*
394 * An empty flush handed down from a stacking driver may
395 * translate into nothing if the underlying device does not
396 * advertise a write-back cache. In this case, simply
397 * complete the request.
398 */
399 if (!policy) {
400 if (q->mq_ops)
401 blk_mq_end_io(rq, 0);
402 else
403 __blk_end_bidi_request(rq, 0, 0, 0);
404 return;
405 }
406
407 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
408
409 /*
410 * If there's data but flush is not necessary, the request can be
411 * processed directly without going through flush machinery. Queue
412 * for normal execution.
413 */
414 if ((policy & REQ_FSEQ_DATA) &&
415 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
416 if (q->mq_ops) {
417 blk_mq_insert_request(rq, false, false, true);
418 } else
419 list_add_tail(&rq->queuelist, &q->queue_head);
420 return;
421 }
422
423 /*
424 * @rq should go through flush machinery. Mark it part of flush
425 * sequence and submit for further processing.
426 */
427 memset(&rq->flush, 0, sizeof(rq->flush));
428 INIT_LIST_HEAD(&rq->flush.list);
429 rq->cmd_flags |= REQ_FLUSH_SEQ;
430 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
431 if (q->mq_ops) {
432 rq->end_io = mq_flush_data_end_io;
433
434 spin_lock_irq(&q->mq_flush_lock);
435 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
436 spin_unlock_irq(&q->mq_flush_lock);
437 return;
438 }
439 rq->end_io = flush_data_end_io;
440
441 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
442}
443
444/**
445 * blk_abort_flushes - @q is being aborted, abort flush requests
446 * @q: request_queue being aborted
447 *
448 * To be called from elv_abort_queue(). @q is being aborted. Prepare all
449 * FLUSH/FUA requests for abortion.
450 *
451 * CONTEXT:
452 * spin_lock_irq(q->queue_lock)
453 */
454void blk_abort_flushes(struct request_queue *q)
455{
456 struct request *rq, *n;
457 int i;
458
459 /*
460 * Requests in flight for data are already owned by the dispatch
461 * queue or the device driver. Just restore for normal completion.
462 */
463 list_for_each_entry_safe(rq, n, &q->flush_data_in_flight, flush.list) {
464 list_del_init(&rq->flush.list);
465 blk_flush_restore_request(rq);
466 }
467
468 /*
469 * We need to give away requests on flush queues. Restore for
470 * normal completion and put them on the dispatch queue.
471 */
472 for (i = 0; i < ARRAY_SIZE(q->flush_queue); i++) {
473 list_for_each_entry_safe(rq, n, &q->flush_queue[i],
474 flush.list) {
475 list_del_init(&rq->flush.list);
476 blk_flush_restore_request(rq);
477 list_add_tail(&rq->queuelist, &q->queue_head);
478 }
479 }
480}
481
482/**
483 * blkdev_issue_flush - queue a flush
484 * @bdev: blockdev to issue flush for
485 * @gfp_mask: memory allocation flags (for bio_alloc)
486 * @error_sector: error sector
487 *
488 * Description:
489 * Issue a flush for the block device in question. Caller can supply
490 * room for storing the error offset in case of a flush error, if they
491 * wish to. If WAIT flag is not passed then caller may check only what
492 * request was pushed in some internal queue for later handling.
493 */
494int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
495 sector_t *error_sector)
496{
497 struct request_queue *q;
498 struct bio *bio;
499 int ret = 0;
500
501 if (bdev->bd_disk == NULL)
502 return -ENXIO;
503
504 q = bdev_get_queue(bdev);
505 if (!q)
506 return -ENXIO;
507
508 /*
509 * some block devices may not have their queue correctly set up here
510 * (e.g. loop device without a backing file) and so issuing a flush
511 * here will panic. Ensure there is a request function before issuing
512 * the flush.
513 */
514 if (!q->make_request_fn)
515 return -ENXIO;
516
517 bio = bio_alloc(gfp_mask, 0);
518 bio->bi_bdev = bdev;
519
520 ret = submit_bio_wait(WRITE_FLUSH, bio);
521
522 /*
523 * The driver must store the error location in ->bi_sector, if
524 * it supports it. For non-stacked drivers, this should be
525 * copied from blk_rq_pos(rq).
526 */
527 if (error_sector)
528 *error_sector = bio->bi_iter.bi_sector;
529
530 bio_put(bio);
531 return ret;
532}
533EXPORT_SYMBOL(blkdev_issue_flush);
534
535void blk_mq_init_flush(struct request_queue *q)
536{
537 spin_lock_init(&q->mq_flush_lock);
538}