Line data Source code
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/part_stat.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, blk_opf_t flags);
98 :
99 : static inline struct blk_flush_queue *
100 : blk_get_flush_queue(struct request_queue *q, struct blk_mq_ctx *ctx)
101 : {
102 0 : return blk_mq_map_queue(q, REQ_OP_FLUSH, ctx)->fq;
103 : }
104 :
105 : static unsigned int blk_flush_policy(unsigned long fflags, struct request *rq)
106 : {
107 0 : unsigned int policy = 0;
108 :
109 0 : if (blk_rq_sectors(rq))
110 0 : policy |= REQ_FSEQ_DATA;
111 :
112 0 : if (fflags & (1UL << QUEUE_FLAG_WC)) {
113 0 : if (rq->cmd_flags & REQ_PREFLUSH)
114 0 : policy |= REQ_FSEQ_PREFLUSH;
115 0 : if (!(fflags & (1UL << QUEUE_FLAG_FUA)) &&
116 0 : (rq->cmd_flags & REQ_FUA))
117 0 : policy |= REQ_FSEQ_POSTFLUSH;
118 : }
119 : return policy;
120 : }
121 :
122 : static unsigned int blk_flush_cur_seq(struct request *rq)
123 : {
124 0 : return 1 << ffz(rq->flush.seq);
125 : }
126 :
127 : static void blk_flush_restore_request(struct request *rq)
128 : {
129 : /*
130 : * After flush data completion, @rq->bio is %NULL but we need to
131 : * complete the bio again. @rq->biotail is guaranteed to equal the
132 : * original @rq->bio. Restore it.
133 : */
134 0 : rq->bio = rq->biotail;
135 :
136 : /* make @rq a normal request */
137 0 : rq->rq_flags &= ~RQF_FLUSH_SEQ;
138 0 : rq->end_io = rq->flush.saved_end_io;
139 : }
140 :
141 : static void blk_flush_queue_rq(struct request *rq, bool add_front)
142 : {
143 0 : blk_mq_add_to_requeue_list(rq, add_front, true);
144 : }
145 :
146 0 : static void blk_account_io_flush(struct request *rq)
147 : {
148 0 : struct block_device *part = rq->q->disk->part0;
149 :
150 0 : part_stat_lock();
151 0 : part_stat_inc(part, ios[STAT_FLUSH]);
152 0 : part_stat_add(part, nsecs[STAT_FLUSH],
153 : ktime_get_ns() - rq->start_time_ns);
154 0 : part_stat_unlock();
155 0 : }
156 :
157 : /**
158 : * blk_flush_complete_seq - complete flush sequence
159 : * @rq: PREFLUSH/FUA request being sequenced
160 : * @fq: flush queue
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(fq->mq_flush_lock)
169 : */
170 0 : static void blk_flush_complete_seq(struct request *rq,
171 : struct blk_flush_queue *fq,
172 : unsigned int seq, blk_status_t error)
173 : {
174 0 : struct request_queue *q = rq->q;
175 0 : struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
176 : blk_opf_t cmd_flags;
177 :
178 0 : BUG_ON(rq->flush.seq & seq);
179 0 : rq->flush.seq |= seq;
180 0 : cmd_flags = rq->cmd_flags;
181 :
182 0 : if (likely(!error))
183 0 : seq = blk_flush_cur_seq(rq);
184 : else
185 : seq = REQ_FSEQ_DONE;
186 :
187 0 : switch (seq) {
188 : case REQ_FSEQ_PREFLUSH:
189 : case REQ_FSEQ_POSTFLUSH:
190 : /* queue for flush */
191 0 : if (list_empty(pending))
192 0 : fq->flush_pending_since = jiffies;
193 0 : list_move_tail(&rq->flush.list, pending);
194 : break;
195 :
196 : case REQ_FSEQ_DATA:
197 0 : list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
198 : blk_flush_queue_rq(rq, true);
199 : break;
200 :
201 : case REQ_FSEQ_DONE:
202 : /*
203 : * @rq was previously adjusted by blk_insert_flush() for
204 : * flush sequencing and may already have gone through the
205 : * flush data request completion path. Restore @rq for
206 : * normal completion and end it.
207 : */
208 0 : list_del_init(&rq->flush.list);
209 0 : blk_flush_restore_request(rq);
210 0 : blk_mq_end_request(rq, error);
211 0 : break;
212 :
213 : default:
214 0 : BUG();
215 : }
216 :
217 0 : blk_kick_flush(q, fq, cmd_flags);
218 0 : }
219 :
220 0 : static enum rq_end_io_ret flush_end_io(struct request *flush_rq,
221 : blk_status_t error)
222 : {
223 0 : struct request_queue *q = flush_rq->q;
224 : struct list_head *running;
225 : struct request *rq, *n;
226 0 : unsigned long flags = 0;
227 0 : struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
228 :
229 : /* release the tag's ownership to the req cloned from */
230 0 : spin_lock_irqsave(&fq->mq_flush_lock, flags);
231 :
232 0 : if (!req_ref_put_and_test(flush_rq)) {
233 0 : fq->rq_status = error;
234 0 : spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
235 0 : return RQ_END_IO_NONE;
236 : }
237 :
238 0 : blk_account_io_flush(flush_rq);
239 : /*
240 : * Flush request has to be marked as IDLE when it is really ended
241 : * because its .end_io() is called from timeout code path too for
242 : * avoiding use-after-free.
243 : */
244 0 : WRITE_ONCE(flush_rq->state, MQ_RQ_IDLE);
245 0 : if (fq->rq_status != BLK_STS_OK) {
246 0 : error = fq->rq_status;
247 0 : fq->rq_status = BLK_STS_OK;
248 : }
249 :
250 0 : if (!q->elevator) {
251 0 : flush_rq->tag = BLK_MQ_NO_TAG;
252 : } else {
253 0 : blk_mq_put_driver_tag(flush_rq);
254 0 : flush_rq->internal_tag = BLK_MQ_NO_TAG;
255 : }
256 :
257 0 : running = &fq->flush_queue[fq->flush_running_idx];
258 0 : BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
259 :
260 : /* account completion of the flush request */
261 0 : fq->flush_running_idx ^= 1;
262 :
263 : /* and push the waiting requests to the next stage */
264 0 : list_for_each_entry_safe(rq, n, running, flush.list) {
265 0 : unsigned int seq = blk_flush_cur_seq(rq);
266 :
267 0 : BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
268 0 : blk_flush_complete_seq(rq, fq, seq, error);
269 : }
270 :
271 0 : spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
272 0 : return RQ_END_IO_NONE;
273 : }
274 :
275 0 : bool is_flush_rq(struct request *rq)
276 : {
277 0 : return rq->end_io == flush_end_io;
278 : }
279 :
280 : /**
281 : * blk_kick_flush - consider issuing flush request
282 : * @q: request_queue being kicked
283 : * @fq: flush queue
284 : * @flags: cmd_flags of the original request
285 : *
286 : * Flush related states of @q have changed, consider issuing flush request.
287 : * Please read the comment at the top of this file for more info.
288 : *
289 : * CONTEXT:
290 : * spin_lock_irq(fq->mq_flush_lock)
291 : *
292 : */
293 0 : static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
294 : blk_opf_t flags)
295 : {
296 0 : struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
297 0 : struct request *first_rq =
298 0 : list_first_entry(pending, struct request, flush.list);
299 0 : struct request *flush_rq = fq->flush_rq;
300 :
301 : /* C1 described at the top of this file */
302 0 : if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
303 : return;
304 :
305 : /* C2 and C3 */
306 0 : if (!list_empty(&fq->flush_data_in_flight) &&
307 0 : time_before(jiffies,
308 : fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
309 : return;
310 :
311 : /*
312 : * Issue flush and toggle pending_idx. This makes pending_idx
313 : * different from running_idx, which means flush is in flight.
314 : */
315 0 : fq->flush_pending_idx ^= 1;
316 :
317 0 : blk_rq_init(q, flush_rq);
318 :
319 : /*
320 : * In case of none scheduler, borrow tag from the first request
321 : * since they can't be in flight at the same time. And acquire
322 : * the tag's ownership for flush req.
323 : *
324 : * In case of IO scheduler, flush rq need to borrow scheduler tag
325 : * just for cheating put/get driver tag.
326 : */
327 0 : flush_rq->mq_ctx = first_rq->mq_ctx;
328 0 : flush_rq->mq_hctx = first_rq->mq_hctx;
329 :
330 0 : if (!q->elevator) {
331 0 : flush_rq->tag = first_rq->tag;
332 :
333 : /*
334 : * We borrow data request's driver tag, so have to mark
335 : * this flush request as INFLIGHT for avoiding double
336 : * account of this driver tag
337 : */
338 0 : flush_rq->rq_flags |= RQF_MQ_INFLIGHT;
339 : } else
340 0 : flush_rq->internal_tag = first_rq->internal_tag;
341 :
342 : flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
343 0 : flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK);
344 0 : flush_rq->rq_flags |= RQF_FLUSH_SEQ;
345 0 : flush_rq->end_io = flush_end_io;
346 : /*
347 : * Order WRITE ->end_io and WRITE rq->ref, and its pair is the one
348 : * implied in refcount_inc_not_zero() called from
349 : * blk_mq_find_and_get_req(), which orders WRITE/READ flush_rq->ref
350 : * and READ flush_rq->end_io
351 : */
352 0 : smp_wmb();
353 0 : req_ref_set(flush_rq, 1);
354 :
355 : blk_flush_queue_rq(flush_rq, false);
356 : }
357 :
358 0 : static enum rq_end_io_ret mq_flush_data_end_io(struct request *rq,
359 : blk_status_t error)
360 : {
361 0 : struct request_queue *q = rq->q;
362 0 : struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
363 0 : struct blk_mq_ctx *ctx = rq->mq_ctx;
364 : unsigned long flags;
365 0 : struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
366 :
367 0 : if (q->elevator) {
368 0 : WARN_ON(rq->tag < 0);
369 : blk_mq_put_driver_tag(rq);
370 : }
371 :
372 : /*
373 : * After populating an empty queue, kick it to avoid stall. Read
374 : * the comment in flush_end_io().
375 : */
376 0 : spin_lock_irqsave(&fq->mq_flush_lock, flags);
377 0 : blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
378 0 : spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
379 :
380 0 : blk_mq_sched_restart(hctx);
381 0 : return RQ_END_IO_NONE;
382 : }
383 :
384 : /**
385 : * blk_insert_flush - insert a new PREFLUSH/FUA request
386 : * @rq: request to insert
387 : *
388 : * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
389 : * or __blk_mq_run_hw_queue() to dispatch request.
390 : * @rq is being submitted. Analyze what needs to be done and put it on the
391 : * right queue.
392 : */
393 0 : void blk_insert_flush(struct request *rq)
394 : {
395 0 : struct request_queue *q = rq->q;
396 0 : unsigned long fflags = q->queue_flags; /* may change, cache */
397 0 : unsigned int policy = blk_flush_policy(fflags, rq);
398 0 : struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
399 :
400 : /*
401 : * @policy now records what operations need to be done. Adjust
402 : * REQ_PREFLUSH and FUA for the driver.
403 : */
404 0 : rq->cmd_flags &= ~REQ_PREFLUSH;
405 0 : if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
406 0 : rq->cmd_flags &= ~REQ_FUA;
407 :
408 : /*
409 : * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
410 : * of those flags, we have to set REQ_SYNC to avoid skewing
411 : * the request accounting.
412 : */
413 0 : rq->cmd_flags |= REQ_SYNC;
414 :
415 : /*
416 : * An empty flush handed down from a stacking driver may
417 : * translate into nothing if the underlying device does not
418 : * advertise a write-back cache. In this case, simply
419 : * complete the request.
420 : */
421 0 : if (!policy) {
422 0 : blk_mq_end_request(rq, 0);
423 0 : return;
424 : }
425 :
426 0 : BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
427 :
428 : /*
429 : * If there's data but flush is not necessary, the request can be
430 : * processed directly without going through flush machinery. Queue
431 : * for normal execution.
432 : */
433 0 : if ((policy & REQ_FSEQ_DATA) &&
434 : !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
435 0 : blk_mq_request_bypass_insert(rq, false, true);
436 0 : return;
437 : }
438 :
439 : /*
440 : * @rq should go through flush machinery. Mark it part of flush
441 : * sequence and submit for further processing.
442 : */
443 0 : memset(&rq->flush, 0, sizeof(rq->flush));
444 0 : INIT_LIST_HEAD(&rq->flush.list);
445 0 : rq->rq_flags |= RQF_FLUSH_SEQ;
446 0 : rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
447 :
448 0 : rq->end_io = mq_flush_data_end_io;
449 :
450 0 : spin_lock_irq(&fq->mq_flush_lock);
451 0 : blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
452 0 : spin_unlock_irq(&fq->mq_flush_lock);
453 : }
454 :
455 : /**
456 : * blkdev_issue_flush - queue a flush
457 : * @bdev: blockdev to issue flush for
458 : *
459 : * Description:
460 : * Issue a flush for the block device in question.
461 : */
462 0 : int blkdev_issue_flush(struct block_device *bdev)
463 : {
464 : struct bio bio;
465 :
466 0 : bio_init(&bio, bdev, NULL, 0, REQ_OP_WRITE | REQ_PREFLUSH);
467 0 : return submit_bio_wait(&bio);
468 : }
469 : EXPORT_SYMBOL(blkdev_issue_flush);
470 :
471 0 : struct blk_flush_queue *blk_alloc_flush_queue(int node, int cmd_size,
472 : gfp_t flags)
473 : {
474 : struct blk_flush_queue *fq;
475 0 : int rq_sz = sizeof(struct request);
476 :
477 0 : fq = kzalloc_node(sizeof(*fq), flags, node);
478 0 : if (!fq)
479 : goto fail;
480 :
481 0 : spin_lock_init(&fq->mq_flush_lock);
482 :
483 0 : rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
484 0 : fq->flush_rq = kzalloc_node(rq_sz, flags, node);
485 0 : if (!fq->flush_rq)
486 : goto fail_rq;
487 :
488 0 : INIT_LIST_HEAD(&fq->flush_queue[0]);
489 0 : INIT_LIST_HEAD(&fq->flush_queue[1]);
490 0 : INIT_LIST_HEAD(&fq->flush_data_in_flight);
491 :
492 0 : return fq;
493 :
494 : fail_rq:
495 0 : kfree(fq);
496 : fail:
497 : return NULL;
498 : }
499 :
500 0 : void blk_free_flush_queue(struct blk_flush_queue *fq)
501 : {
502 : /* bio based request queue hasn't flush queue */
503 0 : if (!fq)
504 : return;
505 :
506 0 : kfree(fq->flush_rq);
507 0 : kfree(fq);
508 : }
509 :
510 : /*
511 : * Allow driver to set its own lock class to fq->mq_flush_lock for
512 : * avoiding lockdep complaint.
513 : *
514 : * flush_end_io() may be called recursively from some driver, such as
515 : * nvme-loop, so lockdep may complain 'possible recursive locking' because
516 : * all 'struct blk_flush_queue' instance share same mq_flush_lock lock class
517 : * key. We need to assign different lock class for these driver's
518 : * fq->mq_flush_lock for avoiding the lockdep warning.
519 : *
520 : * Use dynamically allocated lock class key for each 'blk_flush_queue'
521 : * instance is over-kill, and more worse it introduces horrible boot delay
522 : * issue because synchronize_rcu() is implied in lockdep_unregister_key which
523 : * is called for each hctx release. SCSI probing may synchronously create and
524 : * destroy lots of MQ request_queues for non-existent devices, and some robot
525 : * test kernel always enable lockdep option. It is observed that more than half
526 : * an hour is taken during SCSI MQ probe with per-fq lock class.
527 : */
528 0 : void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx,
529 : struct lock_class_key *key)
530 : {
531 : lockdep_set_class(&hctx->fq->mq_flush_lock, key);
532 0 : }
533 : EXPORT_SYMBOL_GPL(blk_mq_hctx_set_fq_lock_class);
|
