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/blk-mq.h>
72 #include <linux/lockdep.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 void blk_kick_flush(struct request_queue *q,
97 struct blk_flush_queue *fq, unsigned int flags);
98
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
116 static unsigned int blk_flush_cur_seq(struct request *rq)
117 {
118 return 1 << ffz(rq->flush.seq);
119 }
120
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
135 static void blk_flush_queue_rq(struct request *rq, bool add_front)
136 {
137 blk_mq_add_to_requeue_list(rq, add_front, true);
138 }
139
140 /**
141 * blk_flush_complete_seq - complete flush sequence
142 * @rq: PREFLUSH/FUA request being sequenced
143 * @fq: flush queue
144 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
145 * @error: whether an error occurred
146 *
147 * @rq just completed @seq part of its flush sequence, record the
148 * completion and trigger the next step.
149 *
150 * CONTEXT:
151 * spin_lock_irq(fq->mq_flush_lock)
152 *
153 * RETURNS:
154 * %true if requests were added to the dispatch queue, %false otherwise.
155 */
156 static void blk_flush_complete_seq(struct request *rq,
157 struct blk_flush_queue *fq,
158 unsigned int seq, blk_status_t error)
159 {
160 struct request_queue *q = rq->q;
161 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
162 unsigned int cmd_flags;
163
164 BUG_ON(rq->flush.seq & seq);
165 rq->flush.seq |= seq;
166 cmd_flags = rq->cmd_flags;
167
168 if (likely(!error))
169 seq = blk_flush_cur_seq(rq);
170 else
171 seq = REQ_FSEQ_DONE;
172
173 switch (seq) {
174 case REQ_FSEQ_PREFLUSH:
175 case REQ_FSEQ_POSTFLUSH:
176 /* queue for flush */
177 if (list_empty(pending))
178 fq->flush_pending_since = jiffies;
179 list_move_tail(&rq->flush.list, pending);
180 break;
181
182 case REQ_FSEQ_DATA:
183 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
184 blk_flush_queue_rq(rq, true);
185 break;
186
187 case REQ_FSEQ_DONE:
188 /*
189 * @rq was previously adjusted by blk_flush_issue() for
190 * flush sequencing and may already have gone through the
191 * flush data request completion path. Restore @rq for
192 * normal completion and end it.
193 */
194 BUG_ON(!list_empty(&rq->queuelist));
195 list_del_init(&rq->flush.list);
196 blk_flush_restore_request(rq);
197 blk_mq_end_request(rq, error);
198 break;
199
200 default:
201 BUG();
202 }
203
204 blk_kick_flush(q, fq, cmd_flags);
205 }
206
207 static void flush_end_io(struct request *flush_rq, blk_status_t error)
208 {
209 struct request_queue *q = flush_rq->q;
210 struct list_head *running;
211 struct request *rq, *n;
212 unsigned long flags = 0;
213 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
214 struct blk_mq_hw_ctx *hctx;
215
216 /* release the tag's ownership to the req cloned from */
217 spin_lock_irqsave(&fq->mq_flush_lock, flags);
218
219 if (!refcount_dec_and_test(&flush_rq->ref)) {
220 fq->rq_status = error;
221 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
222 return;
223 }
224
225 if (fq->rq_status != BLK_STS_OK)
226 error = fq->rq_status;
227
228 hctx = flush_rq->mq_hctx;
229 if (!q->elevator) {
230 blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq);
231 flush_rq->tag = -1;
232 } else {
233 blk_mq_put_driver_tag(flush_rq);
234 flush_rq->internal_tag = -1;
235 }
236
237 running = &fq->flush_queue[fq->flush_running_idx];
238 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
239
240 /* account completion of the flush request */
241 fq->flush_running_idx ^= 1;
242
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);
246
247 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
248 blk_flush_complete_seq(rq, fq, seq, error);
249 }
250
251 fq->flush_queue_delayed = 0;
252 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
253 }
254
255 /**
256 * blk_kick_flush - consider issuing flush request
257 * @q: request_queue being kicked
258 * @fq: flush queue
259 * @flags: cmd_flags of the original request
260 *
261 * Flush related states of @q have changed, consider issuing flush request.
262 * Please read the comment at the top of this file for more info.
263 *
264 * CONTEXT:
265 * spin_lock_irq(fq->mq_flush_lock)
266 *
267 */
268 static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
269 unsigned int flags)
270 {
271 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
272 struct request *first_rq =
273 list_first_entry(pending, struct request, flush.list);
274 struct request *flush_rq = fq->flush_rq;
275
276 /* C1 described at the top of this file */
277 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
278 return;
279
280 /* C2 and C3
281 *
282 * For blk-mq + scheduling, we can risk having all driver tags
283 * assigned to empty flushes, and we deadlock if we are expecting
284 * other requests to make progress. Don't defer for that case.
285 */
286 if (!list_empty(&fq->flush_data_in_flight) && q->elevator &&
287 time_before(jiffies,
288 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
289 return;
290
291 /*
292 * Issue flush and toggle pending_idx. This makes pending_idx
293 * different from running_idx, which means flush is in flight.
294 */
295 fq->flush_pending_idx ^= 1;
296
297 blk_rq_init(q, flush_rq);
298
299 /*
300 * In case of none scheduler, borrow tag from the first request
301 * since they can't be in flight at the same time. And acquire
302 * the tag's ownership for flush req.
303 *
304 * In case of IO scheduler, flush rq need to borrow scheduler tag
305 * just for cheating put/get driver tag.
306 */
307 flush_rq->mq_ctx = first_rq->mq_ctx;
308 flush_rq->mq_hctx = first_rq->mq_hctx;
309
310 if (!q->elevator) {
311 fq->orig_rq = first_rq;
312 flush_rq->tag = first_rq->tag;
313 blk_mq_tag_set_rq(flush_rq->mq_hctx, first_rq->tag, flush_rq);
314 } else {
315 flush_rq->internal_tag = first_rq->internal_tag;
316 }
317
318 flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
319 flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK);
320 flush_rq->rq_flags |= RQF_FLUSH_SEQ;
321 flush_rq->rq_disk = first_rq->rq_disk;
322 flush_rq->end_io = flush_end_io;
323
324 blk_flush_queue_rq(flush_rq, false);
325 }
326
327 static void mq_flush_data_end_io(struct request *rq, blk_status_t error)
328 {
329 struct request_queue *q = rq->q;
330 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
331 struct blk_mq_ctx *ctx = rq->mq_ctx;
332 unsigned long flags;
333 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
334
335 if (q->elevator) {
336 WARN_ON(rq->tag < 0);
337 blk_mq_put_driver_tag(rq);
338 }
339
340 /*
341 * After populating an empty queue, kick it to avoid stall. Read
342 * the comment in flush_end_io().
343 */
344 spin_lock_irqsave(&fq->mq_flush_lock, flags);
345 blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
346 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
347
348 blk_mq_sched_restart(hctx);
349 }
350
351 /**
352 * blk_insert_flush - insert a new PREFLUSH/FUA request
353 * @rq: request to insert
354 *
355 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
356 * or __blk_mq_run_hw_queue() to dispatch request.
357 * @rq is being submitted. Analyze what needs to be done and put it on the
358 * right queue.
359 */
360 void blk_insert_flush(struct request *rq)
361 {
362 struct request_queue *q = rq->q;
363 unsigned long fflags = q->queue_flags; /* may change, cache */
364 unsigned int policy = blk_flush_policy(fflags, rq);
365 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
366
367 /*
368 * @policy now records what operations need to be done. Adjust
369 * REQ_PREFLUSH and FUA for the driver.
370 */
371 rq->cmd_flags &= ~REQ_PREFLUSH;
372 if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
373 rq->cmd_flags &= ~REQ_FUA;
374
375 /*
376 * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
377 * of those flags, we have to set REQ_SYNC to avoid skewing
378 * the request accounting.
379 */
380 rq->cmd_flags |= REQ_SYNC;
381
382 /*
383 * An empty flush handed down from a stacking driver may
384 * translate into nothing if the underlying device does not
385 * advertise a write-back cache. In this case, simply
386 * complete the request.
387 */
388 if (!policy) {
389 blk_mq_end_request(rq, 0);
390 return;
391 }
392
393 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
394
395 /*
396 * If there's data but flush is not necessary, the request can be
397 * processed directly without going through flush machinery. Queue
398 * for normal execution.
399 */
400 if ((policy & REQ_FSEQ_DATA) &&
401 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
402 blk_mq_request_bypass_insert(rq, false, false);
403 return;
404 }
405
406 /*
407 * @rq should go through flush machinery. Mark it part of flush
408 * sequence and submit for further processing.
409 */
410 memset(&rq->flush, 0, sizeof(rq->flush));
411 INIT_LIST_HEAD(&rq->flush.list);
412 rq->rq_flags |= RQF_FLUSH_SEQ;
413 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
414
415 rq->end_io = mq_flush_data_end_io;
416
417 spin_lock_irq(&fq->mq_flush_lock);
418 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
419 spin_unlock_irq(&fq->mq_flush_lock);
420 }
421
422 /**
423 * blkdev_issue_flush - queue a flush
424 * @bdev: blockdev to issue flush for
425 * @gfp_mask: memory allocation flags (for bio_alloc)
426 * @error_sector: error sector
427 *
428 * Description:
429 * Issue a flush for the block device in question. Caller can supply
430 * room for storing the error offset in case of a flush error, if they
431 * wish to.
432 */
433 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
434 sector_t *error_sector)
435 {
436 struct request_queue *q;
437 struct bio *bio;
438 int ret = 0;
439
440 if (bdev->bd_disk == NULL)
441 return -ENXIO;
442
443 q = bdev_get_queue(bdev);
444 if (!q)
445 return -ENXIO;
446
447 /*
448 * some block devices may not have their queue correctly set up here
449 * (e.g. loop device without a backing file) and so issuing a flush
450 * here will panic. Ensure there is a request function before issuing
451 * the flush.
452 */
453 if (!q->make_request_fn)
454 return -ENXIO;
455
456 bio = bio_alloc(gfp_mask, 0);
457 bio_set_dev(bio, bdev);
458 bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
459
460 ret = submit_bio_wait(bio);
461
462 /*
463 * The driver must store the error location in ->bi_sector, if
464 * it supports it. For non-stacked drivers, this should be
465 * copied from blk_rq_pos(rq).
466 */
467 if (error_sector)
468 *error_sector = bio->bi_iter.bi_sector;
469
470 bio_put(bio);
471 return ret;
472 }
473 EXPORT_SYMBOL(blkdev_issue_flush);
474
475 struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
476 int node, int cmd_size, gfp_t flags)
477 {
478 struct blk_flush_queue *fq;
479 int rq_sz = sizeof(struct request);
480
481 fq = kzalloc_node(sizeof(*fq), flags, node);
482 if (!fq)
483 goto fail;
484
485 spin_lock_init(&fq->mq_flush_lock);
486
487 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
488 fq->flush_rq = kzalloc_node(rq_sz, flags, node);
489 if (!fq->flush_rq)
490 goto fail_rq;
491
492 INIT_LIST_HEAD(&fq->flush_queue[0]);
493 INIT_LIST_HEAD(&fq->flush_queue[1]);
494 INIT_LIST_HEAD(&fq->flush_data_in_flight);
495
496 lockdep_register_key(&fq->key);
497 lockdep_set_class(&fq->mq_flush_lock, &fq->key);
498
499 return fq;
500
501 fail_rq:
502 kfree(fq);
503 fail:
504 return NULL;
505 }
506
507 void blk_free_flush_queue(struct blk_flush_queue *fq)
508 {
509 /* bio based request queue hasn't flush queue */
510 if (!fq)
511 return;
512
513 lockdep_unregister_key(&fq->key);
514 kfree(fq->flush_rq);
515 kfree(fq);
516 }