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/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 *
98 : blk_get_flush_queue(struct request_queue *q, struct blk_mq_ctx *ctx)
99 : {
100 0 : return blk_mq_map_queue(q, REQ_OP_FLUSH, ctx)->fq;
101 : }
102 :
103 : static unsigned int blk_flush_policy(unsigned long fflags, struct request *rq)
104 : {
105 0 : unsigned int policy = 0;
106 :
107 0 : if (blk_rq_sectors(rq))
108 0 : policy |= REQ_FSEQ_DATA;
109 :
110 0 : if (fflags & (1UL << QUEUE_FLAG_WC)) {
111 0 : if (rq->cmd_flags & REQ_PREFLUSH)
112 0 : policy |= REQ_FSEQ_PREFLUSH;
113 0 : if (!(fflags & (1UL << QUEUE_FLAG_FUA)) &&
114 0 : (rq->cmd_flags & REQ_FUA))
115 0 : policy |= REQ_FSEQ_POSTFLUSH;
116 : }
117 : return policy;
118 : }
119 :
120 : static unsigned int blk_flush_cur_seq(struct request *rq)
121 : {
122 0 : return 1 << ffz(rq->flush.seq);
123 : }
124 :
125 : static void blk_flush_restore_request(struct request *rq)
126 : {
127 : /*
128 : * After flush data completion, @rq->bio is %NULL but we need to
129 : * complete the bio again. @rq->biotail is guaranteed to equal the
130 : * original @rq->bio. Restore it.
131 : */
132 0 : rq->bio = rq->biotail;
133 :
134 : /* make @rq a normal request */
135 0 : rq->rq_flags &= ~RQF_FLUSH_SEQ;
136 0 : rq->end_io = rq->flush.saved_end_io;
137 : }
138 :
139 0 : static void blk_account_io_flush(struct request *rq)
140 : {
141 0 : struct block_device *part = rq->q->disk->part0;
142 :
143 0 : part_stat_lock();
144 0 : part_stat_inc(part, ios[STAT_FLUSH]);
145 0 : part_stat_add(part, nsecs[STAT_FLUSH],
146 : ktime_get_ns() - rq->start_time_ns);
147 0 : part_stat_unlock();
148 0 : }
149 :
150 : /**
151 : * blk_flush_complete_seq - complete flush sequence
152 : * @rq: PREFLUSH/FUA request being sequenced
153 : * @fq: flush queue
154 : * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
155 : * @error: whether an error occurred
156 : *
157 : * @rq just completed @seq part of its flush sequence, record the
158 : * completion and trigger the next step.
159 : *
160 : * CONTEXT:
161 : * spin_lock_irq(fq->mq_flush_lock)
162 : */
163 0 : static void blk_flush_complete_seq(struct request *rq,
164 : struct blk_flush_queue *fq,
165 : unsigned int seq, blk_status_t error)
166 : {
167 0 : struct request_queue *q = rq->q;
168 0 : struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
169 : blk_opf_t cmd_flags;
170 :
171 0 : BUG_ON(rq->flush.seq & seq);
172 0 : rq->flush.seq |= seq;
173 0 : cmd_flags = rq->cmd_flags;
174 :
175 0 : if (likely(!error))
176 0 : seq = blk_flush_cur_seq(rq);
177 : else
178 : seq = REQ_FSEQ_DONE;
179 :
180 0 : switch (seq) {
181 : case REQ_FSEQ_PREFLUSH:
182 : case REQ_FSEQ_POSTFLUSH:
183 : /* queue for flush */
184 0 : if (list_empty(pending))
185 0 : fq->flush_pending_since = jiffies;
186 0 : list_move_tail(&rq->flush.list, pending);
187 : break;
188 :
189 : case REQ_FSEQ_DATA:
190 0 : list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
191 0 : blk_mq_add_to_requeue_list(rq, BLK_MQ_INSERT_AT_HEAD);
192 0 : blk_mq_kick_requeue_list(q);
193 0 : break;
194 :
195 : case REQ_FSEQ_DONE:
196 : /*
197 : * @rq was previously adjusted by blk_insert_flush() 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 0 : list_del_init(&rq->flush.list);
203 0 : blk_flush_restore_request(rq);
204 0 : blk_mq_end_request(rq, error);
205 0 : break;
206 :
207 : default:
208 0 : BUG();
209 : }
210 :
211 0 : blk_kick_flush(q, fq, cmd_flags);
212 0 : }
213 :
214 0 : static enum rq_end_io_ret flush_end_io(struct request *flush_rq,
215 : blk_status_t error)
216 : {
217 0 : struct request_queue *q = flush_rq->q;
218 : struct list_head *running;
219 : struct request *rq, *n;
220 0 : unsigned long flags = 0;
221 0 : struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
222 :
223 : /* release the tag's ownership to the req cloned from */
224 0 : spin_lock_irqsave(&fq->mq_flush_lock, flags);
225 :
226 0 : if (!req_ref_put_and_test(flush_rq)) {
227 0 : fq->rq_status = error;
228 0 : spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
229 0 : return RQ_END_IO_NONE;
230 : }
231 :
232 0 : blk_account_io_flush(flush_rq);
233 : /*
234 : * Flush request has to be marked as IDLE when it is really ended
235 : * because its .end_io() is called from timeout code path too for
236 : * avoiding use-after-free.
237 : */
238 0 : WRITE_ONCE(flush_rq->state, MQ_RQ_IDLE);
239 0 : if (fq->rq_status != BLK_STS_OK) {
240 0 : error = fq->rq_status;
241 0 : fq->rq_status = BLK_STS_OK;
242 : }
243 :
244 0 : if (!q->elevator) {
245 0 : flush_rq->tag = BLK_MQ_NO_TAG;
246 : } else {
247 0 : blk_mq_put_driver_tag(flush_rq);
248 0 : flush_rq->internal_tag = BLK_MQ_NO_TAG;
249 : }
250 :
251 0 : running = &fq->flush_queue[fq->flush_running_idx];
252 0 : BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
253 :
254 : /* account completion of the flush request */
255 0 : fq->flush_running_idx ^= 1;
256 :
257 : /* and push the waiting requests to the next stage */
258 0 : list_for_each_entry_safe(rq, n, running, flush.list) {
259 0 : unsigned int seq = blk_flush_cur_seq(rq);
260 :
261 0 : BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
262 0 : blk_flush_complete_seq(rq, fq, seq, error);
263 : }
264 :
265 0 : spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
266 0 : return RQ_END_IO_NONE;
267 : }
268 :
269 0 : bool is_flush_rq(struct request *rq)
270 : {
271 0 : return rq->end_io == flush_end_io;
272 : }
273 :
274 : /**
275 : * blk_kick_flush - consider issuing flush request
276 : * @q: request_queue being kicked
277 : * @fq: flush queue
278 : * @flags: cmd_flags of the original request
279 : *
280 : * Flush related states of @q have changed, consider issuing flush request.
281 : * Please read the comment at the top of this file for more info.
282 : *
283 : * CONTEXT:
284 : * spin_lock_irq(fq->mq_flush_lock)
285 : *
286 : */
287 0 : static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
288 : blk_opf_t flags)
289 : {
290 0 : struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
291 0 : struct request *first_rq =
292 0 : list_first_entry(pending, struct request, flush.list);
293 0 : struct request *flush_rq = fq->flush_rq;
294 :
295 : /* C1 described at the top of this file */
296 0 : if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
297 : return;
298 :
299 : /* C2 and C3 */
300 0 : if (!list_empty(&fq->flush_data_in_flight) &&
301 0 : time_before(jiffies,
302 : fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
303 : return;
304 :
305 : /*
306 : * Issue flush and toggle pending_idx. This makes pending_idx
307 : * different from running_idx, which means flush is in flight.
308 : */
309 0 : fq->flush_pending_idx ^= 1;
310 :
311 0 : blk_rq_init(q, flush_rq);
312 :
313 : /*
314 : * In case of none scheduler, borrow tag from the first request
315 : * since they can't be in flight at the same time. And acquire
316 : * the tag's ownership for flush req.
317 : *
318 : * In case of IO scheduler, flush rq need to borrow scheduler tag
319 : * just for cheating put/get driver tag.
320 : */
321 0 : flush_rq->mq_ctx = first_rq->mq_ctx;
322 0 : flush_rq->mq_hctx = first_rq->mq_hctx;
323 :
324 0 : if (!q->elevator) {
325 0 : flush_rq->tag = first_rq->tag;
326 :
327 : /*
328 : * We borrow data request's driver tag, so have to mark
329 : * this flush request as INFLIGHT for avoiding double
330 : * account of this driver tag
331 : */
332 0 : flush_rq->rq_flags |= RQF_MQ_INFLIGHT;
333 : } else
334 0 : flush_rq->internal_tag = first_rq->internal_tag;
335 :
336 : flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
337 0 : flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK);
338 0 : flush_rq->rq_flags |= RQF_FLUSH_SEQ;
339 0 : flush_rq->end_io = flush_end_io;
340 : /*
341 : * Order WRITE ->end_io and WRITE rq->ref, and its pair is the one
342 : * implied in refcount_inc_not_zero() called from
343 : * blk_mq_find_and_get_req(), which orders WRITE/READ flush_rq->ref
344 : * and READ flush_rq->end_io
345 : */
346 0 : smp_wmb();
347 0 : req_ref_set(flush_rq, 1);
348 :
349 0 : blk_mq_add_to_requeue_list(flush_rq, 0);
350 0 : blk_mq_kick_requeue_list(q);
351 : }
352 :
353 0 : static enum rq_end_io_ret mq_flush_data_end_io(struct request *rq,
354 : blk_status_t error)
355 : {
356 0 : struct request_queue *q = rq->q;
357 0 : struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
358 0 : struct blk_mq_ctx *ctx = rq->mq_ctx;
359 : unsigned long flags;
360 0 : struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
361 :
362 0 : if (q->elevator) {
363 0 : WARN_ON(rq->tag < 0);
364 : blk_mq_put_driver_tag(rq);
365 : }
366 :
367 : /*
368 : * After populating an empty queue, kick it to avoid stall. Read
369 : * the comment in flush_end_io().
370 : */
371 0 : spin_lock_irqsave(&fq->mq_flush_lock, flags);
372 0 : blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
373 0 : spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
374 :
375 0 : blk_mq_sched_restart(hctx);
376 0 : return RQ_END_IO_NONE;
377 : }
378 :
379 : /**
380 : * blk_insert_flush - insert a new PREFLUSH/FUA request
381 : * @rq: request to insert
382 : *
383 : * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
384 : * or __blk_mq_run_hw_queue() to dispatch request.
385 : * @rq is being submitted. Analyze what needs to be done and put it on the
386 : * right queue.
387 : */
388 0 : void blk_insert_flush(struct request *rq)
389 : {
390 0 : struct request_queue *q = rq->q;
391 0 : unsigned long fflags = q->queue_flags; /* may change, cache */
392 0 : unsigned int policy = blk_flush_policy(fflags, rq);
393 0 : struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
394 0 : struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
395 :
396 : /*
397 : * @policy now records what operations need to be done. Adjust
398 : * REQ_PREFLUSH and FUA for the driver.
399 : */
400 0 : rq->cmd_flags &= ~REQ_PREFLUSH;
401 0 : if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
402 0 : rq->cmd_flags &= ~REQ_FUA;
403 :
404 : /*
405 : * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
406 : * of those flags, we have to set REQ_SYNC to avoid skewing
407 : * the request accounting.
408 : */
409 0 : rq->cmd_flags |= REQ_SYNC;
410 :
411 : /*
412 : * An empty flush handed down from a stacking driver may
413 : * translate into nothing if the underlying device does not
414 : * advertise a write-back cache. In this case, simply
415 : * complete the request.
416 : */
417 0 : if (!policy) {
418 0 : blk_mq_end_request(rq, 0);
419 0 : return;
420 : }
421 :
422 0 : BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
423 :
424 : /*
425 : * If there's data but flush is not necessary, the request can be
426 : * processed directly without going through flush machinery. Queue
427 : * for normal execution.
428 : */
429 0 : if ((policy & REQ_FSEQ_DATA) &&
430 : !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
431 0 : blk_mq_request_bypass_insert(rq, 0);
432 0 : blk_mq_run_hw_queue(hctx, false);
433 0 : return;
434 : }
435 :
436 : /*
437 : * @rq should go through flush machinery. Mark it part of flush
438 : * sequence and submit for further processing.
439 : */
440 0 : memset(&rq->flush, 0, sizeof(rq->flush));
441 0 : INIT_LIST_HEAD(&rq->flush.list);
442 0 : rq->rq_flags |= RQF_FLUSH_SEQ;
443 0 : rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
444 :
445 0 : rq->end_io = mq_flush_data_end_io;
446 :
447 0 : spin_lock_irq(&fq->mq_flush_lock);
448 0 : blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
449 0 : spin_unlock_irq(&fq->mq_flush_lock);
450 : }
451 :
452 : /**
453 : * blkdev_issue_flush - queue a flush
454 : * @bdev: blockdev to issue flush for
455 : *
456 : * Description:
457 : * Issue a flush for the block device in question.
458 : */
459 0 : int blkdev_issue_flush(struct block_device *bdev)
460 : {
461 : struct bio bio;
462 :
463 0 : bio_init(&bio, bdev, NULL, 0, REQ_OP_WRITE | REQ_PREFLUSH);
464 0 : return submit_bio_wait(&bio);
465 : }
466 : EXPORT_SYMBOL(blkdev_issue_flush);
467 :
468 0 : struct blk_flush_queue *blk_alloc_flush_queue(int node, int cmd_size,
469 : gfp_t flags)
470 : {
471 : struct blk_flush_queue *fq;
472 0 : int rq_sz = sizeof(struct request);
473 :
474 0 : fq = kzalloc_node(sizeof(*fq), flags, node);
475 0 : if (!fq)
476 : goto fail;
477 :
478 0 : spin_lock_init(&fq->mq_flush_lock);
479 :
480 0 : rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
481 0 : fq->flush_rq = kzalloc_node(rq_sz, flags, node);
482 0 : if (!fq->flush_rq)
483 : goto fail_rq;
484 :
485 0 : INIT_LIST_HEAD(&fq->flush_queue[0]);
486 0 : INIT_LIST_HEAD(&fq->flush_queue[1]);
487 0 : INIT_LIST_HEAD(&fq->flush_data_in_flight);
488 :
489 0 : return fq;
490 :
491 : fail_rq:
492 0 : kfree(fq);
493 : fail:
494 : return NULL;
495 : }
496 :
497 0 : void blk_free_flush_queue(struct blk_flush_queue *fq)
498 : {
499 : /* bio based request queue hasn't flush queue */
500 0 : if (!fq)
501 : return;
502 :
503 0 : kfree(fq->flush_rq);
504 0 : kfree(fq);
505 : }
506 :
507 : /*
508 : * Allow driver to set its own lock class to fq->mq_flush_lock for
509 : * avoiding lockdep complaint.
510 : *
511 : * flush_end_io() may be called recursively from some driver, such as
512 : * nvme-loop, so lockdep may complain 'possible recursive locking' because
513 : * all 'struct blk_flush_queue' instance share same mq_flush_lock lock class
514 : * key. We need to assign different lock class for these driver's
515 : * fq->mq_flush_lock for avoiding the lockdep warning.
516 : *
517 : * Use dynamically allocated lock class key for each 'blk_flush_queue'
518 : * instance is over-kill, and more worse it introduces horrible boot delay
519 : * issue because synchronize_rcu() is implied in lockdep_unregister_key which
520 : * is called for each hctx release. SCSI probing may synchronously create and
521 : * destroy lots of MQ request_queues for non-existent devices, and some robot
522 : * test kernel always enable lockdep option. It is observed that more than half
523 : * an hour is taken during SCSI MQ probe with per-fq lock class.
524 : */
525 0 : void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx,
526 : struct lock_class_key *key)
527 : {
528 : lockdep_set_class(&hctx->fq->mq_flush_lock, key);
529 0 : }
530 : EXPORT_SYMBOL_GPL(blk_mq_hctx_set_fq_lock_class);
|
