1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Functions to sequence PREFLUSH and FUA writes.
4 *
5 * Copyright (C) 2011 Max Planck Institute for Gravitational Physics
6 * Copyright (C) 2011 Tejun Heo <tj@kernel.org>
7 *
8 * REQ_{PREFLUSH|FUA} requests are decomposed to sequences consisted of three
9 * optional steps - PREFLUSH, DATA and POSTFLUSH - according to the request
10 * properties and hardware capability.
11 *
12 * If a request doesn't have data, only REQ_PREFLUSH makes sense, which
13 * indicates a simple flush request. If there is data, REQ_PREFLUSH indicates
14 * that the device cache should be flushed before the data is executed, and
15 * REQ_FUA means that the data must be on non-volatile media on request
16 * completion.
17 *
18 * If the device doesn't have writeback cache, PREFLUSH and FUA don't make any
19 * difference. The requests are either completed immediately if there's no data
20 * or executed as normal requests otherwise.
21 *
22 * If the device has writeback cache and supports FUA, REQ_PREFLUSH is
23 * translated to PREFLUSH but REQ_FUA is passed down directly with DATA.
24 *
25 * If the device has writeback cache and doesn't support FUA, REQ_PREFLUSH
26 * is translated to PREFLUSH and REQ_FUA to POSTFLUSH.
27 *
28 * The actual execution of flush is double buffered. Whenever a request
29 * needs to execute PRE or POSTFLUSH, it queues at
30 * fq->flush_queue[fq->flush_pending_idx]. Once certain criteria are met, a
31 * REQ_OP_FLUSH is issued and the pending_idx is toggled. When the flush
32 * completes, all the requests which were pending are proceeded to the next
33 * step. This allows arbitrary merging of different types of PREFLUSH/FUA
34 * requests.
35 *
36 * Currently, the following conditions are used to determine when to issue
37 * flush.
38 *
39 * C1. At any given time, only one flush shall be in progress. This makes
40 * double buffering sufficient.
41 *
42 * C2. Flush is deferred if any request is executing DATA of its sequence.
43 * This avoids issuing separate POSTFLUSHes for requests which shared
44 * PREFLUSH.
45 *
46 * C3. The second condition is ignored if there is a request which has
47 * waited longer than FLUSH_PENDING_TIMEOUT. This is to avoid
48 * starvation in the unlikely case where there are continuous stream of
49 * FUA (without PREFLUSH) requests.
50 *
51 * For devices which support FUA, it isn't clear whether C2 (and thus C3)
52 * is beneficial.
53 *
54 * Note that a sequenced PREFLUSH/FUA request with DATA is completed twice.
55 * Once while executing DATA and again after the whole sequence is
56 * complete. The first completion updates the contained bio but doesn't
57 * finish it so that the bio submitter is notified only after the whole
58 * sequence is complete. This is implemented by testing RQF_FLUSH_SEQ in
59 * req_bio_endio().
60 *
61 * The above peculiarity requires that each PREFLUSH/FUA request has only one
62 * bio attached to it, which is guaranteed as they aren't allowed to be
63 * merged in the usual way.
64 */
65
66 #include <linux/kernel.h>
67 #include <linux/module.h>
68 #include <linux/bio.h>
69 #include <linux/blkdev.h>
70 #include <linux/gfp.h>
71 #include <linux/part_stat.h>
72
73 #include "blk.h"
74 #include "blk-mq.h"
75 #include "blk-mq-sched.h"
76
77 /* PREFLUSH/FUA sequences */
78 enum {
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
94 static void blk_kick_flush(struct request_queue *q,
95 struct blk_flush_queue *fq, blk_opf_t flags);
96
97 static inline struct blk_flush_queue *
blk_get_flush_queue(struct request_queue * q,struct blk_mq_ctx * ctx)98 blk_get_flush_queue(struct request_queue *q, struct blk_mq_ctx *ctx)
99 {
100 return blk_mq_map_queue(q, REQ_OP_FLUSH, ctx)->fq;
101 }
102
blk_flush_cur_seq(struct request * rq)103 static unsigned int blk_flush_cur_seq(struct request *rq)
104 {
105 return 1 << ffz(rq->flush.seq);
106 }
107
blk_flush_restore_request(struct request * rq)108 static void blk_flush_restore_request(struct request *rq)
109 {
110 /*
111 * After flush data completion, @rq->bio is %NULL but we need to
112 * complete the bio again. @rq->biotail is guaranteed to equal the
113 * original @rq->bio. Restore it.
114 */
115 rq->bio = rq->biotail;
116 if (rq->bio)
117 rq->__sector = rq->bio->bi_iter.bi_sector;
118
119 /* make @rq a normal request */
120 rq->rq_flags &= ~RQF_FLUSH_SEQ;
121 rq->end_io = rq->flush.saved_end_io;
122 }
123
blk_account_io_flush(struct request * rq)124 static void blk_account_io_flush(struct request *rq)
125 {
126 struct block_device *part = rq->q->disk->part0;
127
128 part_stat_lock();
129 part_stat_inc(part, ios[STAT_FLUSH]);
130 part_stat_add(part, nsecs[STAT_FLUSH],
131 blk_time_get_ns() - rq->start_time_ns);
132 part_stat_unlock();
133 }
134
135 /**
136 * blk_flush_complete_seq - complete flush sequence
137 * @rq: PREFLUSH/FUA request being sequenced
138 * @fq: flush queue
139 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
140 * @error: whether an error occurred
141 *
142 * @rq just completed @seq part of its flush sequence, record the
143 * completion and trigger the next step.
144 *
145 * CONTEXT:
146 * spin_lock_irq(fq->mq_flush_lock)
147 */
blk_flush_complete_seq(struct request * rq,struct blk_flush_queue * fq,unsigned int seq,blk_status_t error)148 static void blk_flush_complete_seq(struct request *rq,
149 struct blk_flush_queue *fq,
150 unsigned int seq, blk_status_t error)
151 {
152 struct request_queue *q = rq->q;
153 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
154 blk_opf_t cmd_flags;
155
156 BUG_ON(rq->flush.seq & seq);
157 rq->flush.seq |= seq;
158 cmd_flags = rq->cmd_flags;
159
160 if (likely(!error))
161 seq = blk_flush_cur_seq(rq);
162 else
163 seq = REQ_FSEQ_DONE;
164
165 switch (seq) {
166 case REQ_FSEQ_PREFLUSH:
167 case REQ_FSEQ_POSTFLUSH:
168 /* queue for flush */
169 if (list_empty(pending))
170 fq->flush_pending_since = jiffies;
171 list_add_tail(&rq->queuelist, pending);
172 break;
173
174 case REQ_FSEQ_DATA:
175 fq->flush_data_in_flight++;
176 spin_lock(&q->requeue_lock);
177 list_move(&rq->queuelist, &q->requeue_list);
178 spin_unlock(&q->requeue_lock);
179 blk_mq_kick_requeue_list(q);
180 break;
181
182 case REQ_FSEQ_DONE:
183 /*
184 * @rq was previously adjusted by blk_insert_flush() for
185 * flush sequencing and may already have gone through the
186 * flush data request completion path. Restore @rq for
187 * normal completion and end it.
188 */
189 list_del_init(&rq->queuelist);
190 blk_flush_restore_request(rq);
191 blk_mq_end_request(rq, error);
192 break;
193
194 default:
195 BUG();
196 }
197
198 blk_kick_flush(q, fq, cmd_flags);
199 }
200
flush_end_io(struct request * flush_rq,blk_status_t error)201 static enum rq_end_io_ret flush_end_io(struct request *flush_rq,
202 blk_status_t error)
203 {
204 struct request_queue *q = flush_rq->q;
205 struct list_head *running;
206 struct request *rq, *n;
207 unsigned long flags = 0;
208 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
209
210 /* release the tag's ownership to the req cloned from */
211 spin_lock_irqsave(&fq->mq_flush_lock, flags);
212
213 if (!req_ref_put_and_test(flush_rq)) {
214 fq->rq_status = error;
215 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
216 return RQ_END_IO_NONE;
217 }
218
219 blk_account_io_flush(flush_rq);
220 /*
221 * Flush request has to be marked as IDLE when it is really ended
222 * because its .end_io() is called from timeout code path too for
223 * avoiding use-after-free.
224 */
225 WRITE_ONCE(flush_rq->state, MQ_RQ_IDLE);
226 if (fq->rq_status != BLK_STS_OK) {
227 error = fq->rq_status;
228 fq->rq_status = BLK_STS_OK;
229 }
230
231 if (!q->elevator) {
232 flush_rq->tag = BLK_MQ_NO_TAG;
233 } else {
234 blk_mq_put_driver_tag(flush_rq);
235 flush_rq->internal_tag = BLK_MQ_NO_TAG;
236 }
237
238 running = &fq->flush_queue[fq->flush_running_idx];
239 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
240
241 /* account completion of the flush request */
242 fq->flush_running_idx ^= 1;
243
244 /* and push the waiting requests to the next stage */
245 list_for_each_entry_safe(rq, n, running, queuelist) {
246 unsigned int seq = blk_flush_cur_seq(rq);
247
248 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
249 list_del_init(&rq->queuelist);
250 blk_flush_complete_seq(rq, fq, seq, error);
251 }
252
253 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
254 return RQ_END_IO_NONE;
255 }
256
is_flush_rq(struct request * rq)257 bool is_flush_rq(struct request *rq)
258 {
259 return rq->end_io == flush_end_io;
260 }
261
262 /**
263 * blk_kick_flush - consider issuing flush request
264 * @q: request_queue being kicked
265 * @fq: flush queue
266 * @flags: cmd_flags of the original request
267 *
268 * Flush related states of @q have changed, consider issuing flush request.
269 * Please read the comment at the top of this file for more info.
270 *
271 * CONTEXT:
272 * spin_lock_irq(fq->mq_flush_lock)
273 *
274 */
blk_kick_flush(struct request_queue * q,struct blk_flush_queue * fq,blk_opf_t flags)275 static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
276 blk_opf_t flags)
277 {
278 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
279 struct request *first_rq =
280 list_first_entry(pending, struct request, queuelist);
281 struct request *flush_rq = fq->flush_rq;
282
283 /* C1 described at the top of this file */
284 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
285 return;
286
287 /* C2 and C3 */
288 if (fq->flush_data_in_flight &&
289 time_before(jiffies,
290 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
291 return;
292
293 /*
294 * Issue flush and toggle pending_idx. This makes pending_idx
295 * different from running_idx, which means flush is in flight.
296 */
297 fq->flush_pending_idx ^= 1;
298
299 blk_rq_init(q, flush_rq);
300
301 /*
302 * In case of none scheduler, borrow tag from the first request
303 * since they can't be in flight at the same time. And acquire
304 * the tag's ownership for flush req.
305 *
306 * In case of IO scheduler, flush rq need to borrow scheduler tag
307 * just for cheating put/get driver tag.
308 */
309 flush_rq->mq_ctx = first_rq->mq_ctx;
310 flush_rq->mq_hctx = first_rq->mq_hctx;
311
312 if (!q->elevator)
313 flush_rq->tag = first_rq->tag;
314 else
315 flush_rq->internal_tag = first_rq->internal_tag;
316
317 flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
318 flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK);
319 flush_rq->rq_flags |= RQF_FLUSH_SEQ;
320 flush_rq->end_io = flush_end_io;
321 /*
322 * Order WRITE ->end_io and WRITE rq->ref, and its pair is the one
323 * implied in refcount_inc_not_zero() called from
324 * blk_mq_find_and_get_req(), which orders WRITE/READ flush_rq->ref
325 * and READ flush_rq->end_io
326 */
327 smp_wmb();
328 req_ref_set(flush_rq, 1);
329
330 spin_lock(&q->requeue_lock);
331 list_add_tail(&flush_rq->queuelist, &q->flush_list);
332 spin_unlock(&q->requeue_lock);
333
334 blk_mq_kick_requeue_list(q);
335 }
336
mq_flush_data_end_io(struct request * rq,blk_status_t error)337 static enum rq_end_io_ret mq_flush_data_end_io(struct request *rq,
338 blk_status_t error)
339 {
340 struct request_queue *q = rq->q;
341 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
342 struct blk_mq_ctx *ctx = rq->mq_ctx;
343 unsigned long flags;
344 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
345
346 if (q->elevator) {
347 WARN_ON(rq->tag < 0);
348 blk_mq_put_driver_tag(rq);
349 }
350
351 /*
352 * After populating an empty queue, kick it to avoid stall. Read
353 * the comment in flush_end_io().
354 */
355 spin_lock_irqsave(&fq->mq_flush_lock, flags);
356 fq->flush_data_in_flight--;
357 /*
358 * May have been corrupted by rq->rq_next reuse, we need to
359 * re-initialize rq->queuelist before reusing it here.
360 */
361 INIT_LIST_HEAD(&rq->queuelist);
362 blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
363 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
364
365 blk_mq_sched_restart(hctx);
366 return RQ_END_IO_NONE;
367 }
368
blk_rq_init_flush(struct request * rq)369 static void blk_rq_init_flush(struct request *rq)
370 {
371 rq->flush.seq = 0;
372 rq->rq_flags |= RQF_FLUSH_SEQ;
373 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
374 rq->end_io = mq_flush_data_end_io;
375 }
376
377 /*
378 * Insert a PREFLUSH/FUA request into the flush state machine.
379 * Returns true if the request has been consumed by the flush state machine,
380 * or false if the caller should continue to process it.
381 */
blk_insert_flush(struct request * rq)382 bool blk_insert_flush(struct request *rq)
383 {
384 struct request_queue *q = rq->q;
385 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
386 bool supports_fua = q->limits.features & BLK_FEAT_FUA;
387 unsigned int policy = 0;
388
389 /* FLUSH/FUA request must never be merged */
390 WARN_ON_ONCE(rq->bio != rq->biotail);
391
392 if (blk_rq_sectors(rq))
393 policy |= REQ_FSEQ_DATA;
394
395 /*
396 * Check which flushes we need to sequence for this operation.
397 */
398 if (blk_queue_write_cache(q)) {
399 if (rq->cmd_flags & REQ_PREFLUSH)
400 policy |= REQ_FSEQ_PREFLUSH;
401 if ((rq->cmd_flags & REQ_FUA) && !supports_fua)
402 policy |= REQ_FSEQ_POSTFLUSH;
403 }
404
405 /*
406 * @policy now records what operations need to be done. Adjust
407 * REQ_PREFLUSH and FUA for the driver.
408 */
409 rq->cmd_flags &= ~REQ_PREFLUSH;
410 if (!supports_fua)
411 rq->cmd_flags &= ~REQ_FUA;
412
413 /*
414 * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
415 * of those flags, we have to set REQ_SYNC to avoid skewing
416 * the request accounting.
417 */
418 rq->cmd_flags |= REQ_SYNC;
419
420 switch (policy) {
421 case 0:
422 /*
423 * An empty flush handed down from a stacking driver may
424 * translate into nothing if the underlying device does not
425 * advertise a write-back cache. In this case, simply
426 * complete the request.
427 */
428 blk_mq_end_request(rq, 0);
429 return true;
430 case REQ_FSEQ_DATA:
431 /*
432 * If there's data, but no flush is necessary, the request can
433 * be processed directly without going through flush machinery.
434 * Queue for normal execution.
435 */
436 return false;
437 case REQ_FSEQ_DATA | REQ_FSEQ_POSTFLUSH:
438 /*
439 * Initialize the flush fields and completion handler to trigger
440 * the post flush, and then just pass the command on.
441 */
442 blk_rq_init_flush(rq);
443 rq->flush.seq |= REQ_FSEQ_PREFLUSH;
444 spin_lock_irq(&fq->mq_flush_lock);
445 fq->flush_data_in_flight++;
446 spin_unlock_irq(&fq->mq_flush_lock);
447 return false;
448 default:
449 /*
450 * Mark the request as part of a flush sequence and submit it
451 * for further processing to the flush state machine.
452 */
453 blk_rq_init_flush(rq);
454 spin_lock_irq(&fq->mq_flush_lock);
455 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
456 spin_unlock_irq(&fq->mq_flush_lock);
457 return true;
458 }
459 }
460
461 /**
462 * blkdev_issue_flush - queue a flush
463 * @bdev: blockdev to issue flush for
464 *
465 * Description:
466 * Issue a flush for the block device in question.
467 */
blkdev_issue_flush(struct block_device * bdev)468 int blkdev_issue_flush(struct block_device *bdev)
469 {
470 struct bio bio;
471
472 bio_init(&bio, bdev, NULL, 0, REQ_OP_WRITE | REQ_PREFLUSH);
473 return submit_bio_wait(&bio);
474 }
475 EXPORT_SYMBOL(blkdev_issue_flush);
476
blk_alloc_flush_queue(int node,int cmd_size,gfp_t flags)477 struct blk_flush_queue *blk_alloc_flush_queue(int node, int cmd_size,
478 gfp_t flags)
479 {
480 struct blk_flush_queue *fq;
481 int rq_sz = sizeof(struct request);
482
483 fq = kzalloc_node(sizeof(*fq), flags, node);
484 if (!fq)
485 goto fail;
486
487 spin_lock_init(&fq->mq_flush_lock);
488
489 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
490 fq->flush_rq = kzalloc_node(rq_sz, flags, node);
491 if (!fq->flush_rq)
492 goto fail_rq;
493
494 INIT_LIST_HEAD(&fq->flush_queue[0]);
495 INIT_LIST_HEAD(&fq->flush_queue[1]);
496
497 return fq;
498
499 fail_rq:
500 kfree(fq);
501 fail:
502 return NULL;
503 }
504
blk_free_flush_queue(struct blk_flush_queue * fq)505 void blk_free_flush_queue(struct blk_flush_queue *fq)
506 {
507 /* bio based request queue hasn't flush queue */
508 if (!fq)
509 return;
510
511 kfree(fq->flush_rq);
512 kfree(fq);
513 }
514
515 /*
516 * Allow driver to set its own lock class to fq->mq_flush_lock for
517 * avoiding lockdep complaint.
518 *
519 * flush_end_io() may be called recursively from some driver, such as
520 * nvme-loop, so lockdep may complain 'possible recursive locking' because
521 * all 'struct blk_flush_queue' instance share same mq_flush_lock lock class
522 * key. We need to assign different lock class for these driver's
523 * fq->mq_flush_lock for avoiding the lockdep warning.
524 *
525 * Use dynamically allocated lock class key for each 'blk_flush_queue'
526 * instance is over-kill, and more worse it introduces horrible boot delay
527 * issue because synchronize_rcu() is implied in lockdep_unregister_key which
528 * is called for each hctx release. SCSI probing may synchronously create and
529 * destroy lots of MQ request_queues for non-existent devices, and some robot
530 * test kernel always enable lockdep option. It is observed that more than half
531 * an hour is taken during SCSI MQ probe with per-fq lock class.
532 */
blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx * hctx,struct lock_class_key * key)533 void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx,
534 struct lock_class_key *key)
535 {
536 lockdep_set_class(&hctx->fq->mq_flush_lock, key);
537 }
538 EXPORT_SYMBOL_GPL(blk_mq_hctx_set_fq_lock_class);
539