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 */
80 enum {
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
96 static bool blk_kick_flush(struct request_queue *q,
97 struct blk_flush_queue *fq, unsigned int flags);
98
blk_flush_policy(unsigned long fflags,struct request * rq)99 static 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
blk_flush_cur_seq(struct request * rq)116 static unsigned int blk_flush_cur_seq(struct request *rq)
117 {
118 return 1 << ffz(rq->flush.seq);
119 }
120
blk_flush_restore_request(struct request * rq)121 static 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
blk_flush_queue_rq(struct request * rq,bool add_front)135 static 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 */
blk_flush_complete_seq(struct request * rq,struct blk_flush_queue * fq,unsigned int seq,blk_status_t error)165 static 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 unsigned int cmd_flags;
173
174 BUG_ON(rq->flush.seq & seq);
175 rq->flush.seq |= seq;
176 cmd_flags = rq->cmd_flags;
177
178 if (likely(!error))
179 seq = blk_flush_cur_seq(rq);
180 else
181 seq = REQ_FSEQ_DONE;
182
183 switch (seq) {
184 case REQ_FSEQ_PREFLUSH:
185 case REQ_FSEQ_POSTFLUSH:
186 /* queue for flush */
187 if (list_empty(pending))
188 fq->flush_pending_since = jiffies;
189 list_move_tail(&rq->flush.list, pending);
190 break;
191
192 case REQ_FSEQ_DATA:
193 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
194 queued = blk_flush_queue_rq(rq, true);
195 break;
196
197 case REQ_FSEQ_DONE:
198 /*
199 * @rq was previously adjusted by blk_flush_issue() for
200 * flush sequencing and may already have gone through the
201 * flush data request completion path. Restore @rq for
202 * normal completion and end it.
203 */
204 BUG_ON(!list_empty(&rq->queuelist));
205 list_del_init(&rq->flush.list);
206 blk_flush_restore_request(rq);
207 if (q->mq_ops)
208 blk_mq_end_request(rq, error);
209 else
210 __blk_end_request_all(rq, error);
211 break;
212
213 default:
214 BUG();
215 }
216
217 kicked = blk_kick_flush(q, fq, cmd_flags);
218 return kicked | queued;
219 }
220
flush_end_io(struct request * flush_rq,blk_status_t error)221 static void flush_end_io(struct request *flush_rq, blk_status_t error)
222 {
223 struct request_queue *q = flush_rq->q;
224 struct list_head *running;
225 bool queued = false;
226 struct request *rq, *n;
227 unsigned long flags = 0;
228 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
229
230 if (q->mq_ops) {
231 struct blk_mq_hw_ctx *hctx;
232
233 /* release the tag's ownership to the req cloned from */
234 spin_lock_irqsave(&fq->mq_flush_lock, flags);
235 hctx = blk_mq_map_queue(q, flush_rq->mq_ctx->cpu);
236 if (!q->elevator) {
237 blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq);
238 flush_rq->tag = -1;
239 } else {
240 blk_mq_put_driver_tag_hctx(hctx, flush_rq);
241 flush_rq->internal_tag = -1;
242 }
243 }
244
245 running = &fq->flush_queue[fq->flush_running_idx];
246 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
247
248 /* account completion of the flush request */
249 fq->flush_running_idx ^= 1;
250
251 if (!q->mq_ops)
252 elv_completed_request(q, flush_rq);
253
254 /* and push the waiting requests to the next stage */
255 list_for_each_entry_safe(rq, n, running, flush.list) {
256 unsigned int seq = blk_flush_cur_seq(rq);
257
258 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
259 queued |= blk_flush_complete_seq(rq, fq, seq, error);
260 }
261
262 /*
263 * Kick the queue to avoid stall for two cases:
264 * 1. Moving a request silently to empty queue_head may stall the
265 * queue.
266 * 2. When flush request is running in non-queueable queue, the
267 * queue is hold. Restart the queue after flush request is finished
268 * to avoid stall.
269 * This function is called from request completion path and calling
270 * directly into request_fn may confuse the driver. Always use
271 * kblockd.
272 */
273 if (queued || fq->flush_queue_delayed) {
274 WARN_ON(q->mq_ops);
275 blk_run_queue_async(q);
276 }
277 fq->flush_queue_delayed = 0;
278 if (q->mq_ops)
279 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
280 }
281
282 /**
283 * blk_kick_flush - consider issuing flush request
284 * @q: request_queue being kicked
285 * @fq: flush queue
286 * @flags: cmd_flags of the original request
287 *
288 * Flush related states of @q have changed, consider issuing flush request.
289 * Please read the comment at the top of this file for more info.
290 *
291 * CONTEXT:
292 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
293 *
294 * RETURNS:
295 * %true if flush was issued, %false otherwise.
296 */
blk_kick_flush(struct request_queue * q,struct blk_flush_queue * fq,unsigned int flags)297 static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
298 unsigned int flags)
299 {
300 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
301 struct request *first_rq =
302 list_first_entry(pending, struct request, flush.list);
303 struct request *flush_rq = fq->flush_rq;
304
305 /* C1 described at the top of this file */
306 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
307 return false;
308
309 /* C2 and C3
310 *
311 * For blk-mq + scheduling, we can risk having all driver tags
312 * assigned to empty flushes, and we deadlock if we are expecting
313 * other requests to make progress. Don't defer for that case.
314 */
315 if (!list_empty(&fq->flush_data_in_flight) &&
316 !(q->mq_ops && q->elevator) &&
317 time_before(jiffies,
318 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
319 return false;
320
321 /*
322 * Issue flush and toggle pending_idx. This makes pending_idx
323 * different from running_idx, which means flush is in flight.
324 */
325 fq->flush_pending_idx ^= 1;
326
327 blk_rq_init(q, flush_rq);
328
329 /*
330 * In case of none scheduler, borrow tag from the first request
331 * since they can't be in flight at the same time. And acquire
332 * the tag's ownership for flush req.
333 *
334 * In case of IO scheduler, flush rq need to borrow scheduler tag
335 * just for cheating put/get driver tag.
336 */
337 if (q->mq_ops) {
338 struct blk_mq_hw_ctx *hctx;
339
340 flush_rq->mq_ctx = first_rq->mq_ctx;
341
342 if (!q->elevator) {
343 fq->orig_rq = first_rq;
344 flush_rq->tag = first_rq->tag;
345 hctx = blk_mq_map_queue(q, first_rq->mq_ctx->cpu);
346 blk_mq_tag_set_rq(hctx, first_rq->tag, flush_rq);
347 } else {
348 flush_rq->internal_tag = first_rq->internal_tag;
349 }
350 }
351
352 flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
353 flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK);
354 flush_rq->rq_flags |= RQF_FLUSH_SEQ;
355 flush_rq->rq_disk = first_rq->rq_disk;
356 flush_rq->end_io = flush_end_io;
357
358 return blk_flush_queue_rq(flush_rq, false);
359 }
360
flush_data_end_io(struct request * rq,blk_status_t error)361 static void flush_data_end_io(struct request *rq, blk_status_t error)
362 {
363 struct request_queue *q = rq->q;
364 struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL);
365
366 lockdep_assert_held(q->queue_lock);
367
368 /*
369 * Updating q->in_flight[] here for making this tag usable
370 * early. Because in blk_queue_start_tag(),
371 * q->in_flight[BLK_RW_ASYNC] is used to limit async I/O and
372 * reserve tags for sync I/O.
373 *
374 * More importantly this way can avoid the following I/O
375 * deadlock:
376 *
377 * - suppose there are 40 fua requests comming to flush queue
378 * and queue depth is 31
379 * - 30 rqs are scheduled then blk_queue_start_tag() can't alloc
380 * tag for async I/O any more
381 * - all the 30 rqs are completed before FLUSH_PENDING_TIMEOUT
382 * and flush_data_end_io() is called
383 * - the other rqs still can't go ahead if not updating
384 * q->in_flight[BLK_RW_ASYNC] here, meantime these rqs
385 * are held in flush data queue and make no progress of
386 * handling post flush rq
387 * - only after the post flush rq is handled, all these rqs
388 * can be completed
389 */
390
391 elv_completed_request(q, rq);
392
393 /* for avoiding double accounting */
394 rq->rq_flags &= ~RQF_STARTED;
395
396 /*
397 * After populating an empty queue, kick it to avoid stall. Read
398 * the comment in flush_end_io().
399 */
400 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
401 blk_run_queue_async(q);
402 }
403
mq_flush_data_end_io(struct request * rq,blk_status_t error)404 static void mq_flush_data_end_io(struct request *rq, blk_status_t error)
405 {
406 struct request_queue *q = rq->q;
407 struct blk_mq_hw_ctx *hctx;
408 struct blk_mq_ctx *ctx = rq->mq_ctx;
409 unsigned long flags;
410 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
411
412 hctx = blk_mq_map_queue(q, ctx->cpu);
413
414 if (q->elevator) {
415 WARN_ON(rq->tag < 0);
416 blk_mq_put_driver_tag_hctx(hctx, rq);
417 }
418
419 /*
420 * After populating an empty queue, kick it to avoid stall. Read
421 * the comment in flush_end_io().
422 */
423 spin_lock_irqsave(&fq->mq_flush_lock, flags);
424 blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
425 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
426
427 blk_mq_run_hw_queue(hctx, true);
428 }
429
430 /**
431 * blk_insert_flush - insert a new PREFLUSH/FUA request
432 * @rq: request to insert
433 *
434 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
435 * or __blk_mq_run_hw_queue() to dispatch request.
436 * @rq is being submitted. Analyze what needs to be done and put it on the
437 * right queue.
438 */
blk_insert_flush(struct request * rq)439 void blk_insert_flush(struct request *rq)
440 {
441 struct request_queue *q = rq->q;
442 unsigned long fflags = q->queue_flags; /* may change, cache */
443 unsigned int policy = blk_flush_policy(fflags, rq);
444 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
445
446 if (!q->mq_ops)
447 lockdep_assert_held(q->queue_lock);
448
449 /*
450 * @policy now records what operations need to be done. Adjust
451 * REQ_PREFLUSH and FUA for the driver.
452 */
453 rq->cmd_flags &= ~REQ_PREFLUSH;
454 if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
455 rq->cmd_flags &= ~REQ_FUA;
456
457 /*
458 * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
459 * of those flags, we have to set REQ_SYNC to avoid skewing
460 * the request accounting.
461 */
462 rq->cmd_flags |= REQ_SYNC;
463
464 /*
465 * An empty flush handed down from a stacking driver may
466 * translate into nothing if the underlying device does not
467 * advertise a write-back cache. In this case, simply
468 * complete the request.
469 */
470 if (!policy) {
471 if (q->mq_ops)
472 blk_mq_end_request(rq, 0);
473 else
474 __blk_end_request(rq, 0, 0);
475 return;
476 }
477
478 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
479
480 /*
481 * If there's data but flush is not necessary, the request can be
482 * processed directly without going through flush machinery. Queue
483 * for normal execution.
484 */
485 if ((policy & REQ_FSEQ_DATA) &&
486 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
487 if (q->mq_ops)
488 blk_mq_request_bypass_insert(rq, false);
489 else
490 list_add_tail(&rq->queuelist, &q->queue_head);
491 return;
492 }
493
494 /*
495 * @rq should go through flush machinery. Mark it part of flush
496 * sequence and submit for further processing.
497 */
498 memset(&rq->flush, 0, sizeof(rq->flush));
499 INIT_LIST_HEAD(&rq->flush.list);
500 rq->rq_flags |= RQF_FLUSH_SEQ;
501 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
502 if (q->mq_ops) {
503 rq->end_io = mq_flush_data_end_io;
504
505 spin_lock_irq(&fq->mq_flush_lock);
506 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
507 spin_unlock_irq(&fq->mq_flush_lock);
508 return;
509 }
510 rq->end_io = flush_data_end_io;
511
512 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
513 }
514
515 /**
516 * blkdev_issue_flush - queue a flush
517 * @bdev: blockdev to issue flush for
518 * @gfp_mask: memory allocation flags (for bio_alloc)
519 * @error_sector: error sector
520 *
521 * Description:
522 * Issue a flush for the block device in question. Caller can supply
523 * room for storing the error offset in case of a flush error, if they
524 * wish to.
525 */
blkdev_issue_flush(struct block_device * bdev,gfp_t gfp_mask,sector_t * error_sector)526 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
527 sector_t *error_sector)
528 {
529 struct request_queue *q;
530 struct bio *bio;
531 int ret = 0;
532
533 if (bdev->bd_disk == NULL)
534 return -ENXIO;
535
536 q = bdev_get_queue(bdev);
537 if (!q)
538 return -ENXIO;
539
540 /*
541 * some block devices may not have their queue correctly set up here
542 * (e.g. loop device without a backing file) and so issuing a flush
543 * here will panic. Ensure there is a request function before issuing
544 * the flush.
545 */
546 if (!q->make_request_fn)
547 return -ENXIO;
548
549 bio = bio_alloc(gfp_mask, 0);
550 bio_set_dev(bio, bdev);
551 bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
552
553 ret = submit_bio_wait(bio);
554
555 /*
556 * The driver must store the error location in ->bi_sector, if
557 * it supports it. For non-stacked drivers, this should be
558 * copied from blk_rq_pos(rq).
559 */
560 if (error_sector)
561 *error_sector = bio->bi_iter.bi_sector;
562
563 bio_put(bio);
564 return ret;
565 }
566 EXPORT_SYMBOL(blkdev_issue_flush);
567
blk_alloc_flush_queue(struct request_queue * q,int node,int cmd_size)568 struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
569 int node, int cmd_size)
570 {
571 struct blk_flush_queue *fq;
572 int rq_sz = sizeof(struct request);
573
574 fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node);
575 if (!fq)
576 goto fail;
577
578 if (q->mq_ops)
579 spin_lock_init(&fq->mq_flush_lock);
580
581 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
582 fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node);
583 if (!fq->flush_rq)
584 goto fail_rq;
585
586 INIT_LIST_HEAD(&fq->flush_queue[0]);
587 INIT_LIST_HEAD(&fq->flush_queue[1]);
588 INIT_LIST_HEAD(&fq->flush_data_in_flight);
589
590 return fq;
591
592 fail_rq:
593 kfree(fq);
594 fail:
595 return NULL;
596 }
597
blk_free_flush_queue(struct blk_flush_queue * fq)598 void blk_free_flush_queue(struct blk_flush_queue *fq)
599 {
600 /* bio based request queue hasn't flush queue */
601 if (!fq)
602 return;
603
604 kfree(fq->flush_rq);
605 kfree(fq);
606 }
607
