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
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 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
blk_account_io_flush(struct request * rq)140 static void blk_account_io_flush(struct request *rq)
141 {
142 struct hd_struct *part = &rq->rq_disk->part0;
143
144 part_stat_lock();
145 part_stat_inc(part, ios[STAT_FLUSH]);
146 part_stat_add(part, nsecs[STAT_FLUSH],
147 ktime_get_ns() - rq->start_time_ns);
148 part_stat_unlock();
149 }
150
151 /**
152 * blk_flush_complete_seq - complete flush sequence
153 * @rq: PREFLUSH/FUA request being sequenced
154 * @fq: flush queue
155 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
156 * @error: whether an error occurred
157 *
158 * @rq just completed @seq part of its flush sequence, record the
159 * completion and trigger the next step.
160 *
161 * CONTEXT:
162 * spin_lock_irq(fq->mq_flush_lock)
163 */
blk_flush_complete_seq(struct request * rq,struct blk_flush_queue * fq,unsigned int seq,blk_status_t error)164 static void blk_flush_complete_seq(struct request *rq,
165 struct blk_flush_queue *fq,
166 unsigned int seq, blk_status_t error)
167 {
168 struct request_queue *q = rq->q;
169 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
170 unsigned int cmd_flags;
171
172 BUG_ON(rq->flush.seq & seq);
173 rq->flush.seq |= seq;
174 cmd_flags = rq->cmd_flags;
175
176 if (likely(!error))
177 seq = blk_flush_cur_seq(rq);
178 else
179 seq = REQ_FSEQ_DONE;
180
181 switch (seq) {
182 case REQ_FSEQ_PREFLUSH:
183 case REQ_FSEQ_POSTFLUSH:
184 /* queue for flush */
185 if (list_empty(pending))
186 fq->flush_pending_since = jiffies;
187 list_move_tail(&rq->flush.list, pending);
188 break;
189
190 case REQ_FSEQ_DATA:
191 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
192 blk_flush_queue_rq(rq, true);
193 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 BUG_ON(!list_empty(&rq->queuelist));
203 list_del_init(&rq->flush.list);
204 blk_flush_restore_request(rq);
205 blk_mq_end_request(rq, error);
206 break;
207
208 default:
209 BUG();
210 }
211
212 blk_kick_flush(q, fq, cmd_flags);
213 }
214
flush_end_io(struct request * flush_rq,blk_status_t error)215 static void flush_end_io(struct request *flush_rq, blk_status_t error)
216 {
217 struct request_queue *q = flush_rq->q;
218 struct list_head *running;
219 struct request *rq, *n;
220 unsigned long flags = 0;
221 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
222
223 blk_account_io_flush(flush_rq);
224
225 /* release the tag's ownership to the req cloned from */
226 spin_lock_irqsave(&fq->mq_flush_lock, flags);
227
228 if (!refcount_dec_and_test(&flush_rq->ref)) {
229 fq->rq_status = error;
230 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
231 return;
232 }
233
234 /*
235 * Flush request has to be marked as IDLE when it is really ended
236 * because its .end_io() is called from timeout code path too for
237 * avoiding use-after-free.
238 */
239 WRITE_ONCE(flush_rq->state, MQ_RQ_IDLE);
240 if (fq->rq_status != BLK_STS_OK)
241 error = fq->rq_status;
242
243 if (!q->elevator) {
244 flush_rq->tag = BLK_MQ_NO_TAG;
245 } else {
246 blk_mq_put_driver_tag(flush_rq);
247 flush_rq->internal_tag = BLK_MQ_NO_TAG;
248 }
249
250 running = &fq->flush_queue[fq->flush_running_idx];
251 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
252
253 /* account completion of the flush request */
254 fq->flush_running_idx ^= 1;
255
256 /* and push the waiting requests to the next stage */
257 list_for_each_entry_safe(rq, n, running, flush.list) {
258 unsigned int seq = blk_flush_cur_seq(rq);
259
260 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
261 blk_flush_complete_seq(rq, fq, seq, error);
262 }
263
264 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
265 }
266
267 /**
268 * blk_kick_flush - consider issuing flush request
269 * @q: request_queue being kicked
270 * @fq: flush queue
271 * @flags: cmd_flags of the original request
272 *
273 * Flush related states of @q have changed, consider issuing flush request.
274 * Please read the comment at the top of this file for more info.
275 *
276 * CONTEXT:
277 * spin_lock_irq(fq->mq_flush_lock)
278 *
279 */
blk_kick_flush(struct request_queue * q,struct blk_flush_queue * fq,unsigned int flags)280 static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
281 unsigned int flags)
282 {
283 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
284 struct request *first_rq =
285 list_first_entry(pending, struct request, flush.list);
286 struct request *flush_rq = fq->flush_rq;
287
288 /* C1 described at the top of this file */
289 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
290 return;
291
292 /* C2 and C3 */
293 if (!list_empty(&fq->flush_data_in_flight) &&
294 time_before(jiffies,
295 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
296 return;
297
298 /*
299 * Issue flush and toggle pending_idx. This makes pending_idx
300 * different from running_idx, which means flush is in flight.
301 */
302 fq->flush_pending_idx ^= 1;
303
304 blk_rq_init(q, flush_rq);
305
306 /*
307 * In case of none scheduler, borrow tag from the first request
308 * since they can't be in flight at the same time. And acquire
309 * the tag's ownership for flush req.
310 *
311 * In case of IO scheduler, flush rq need to borrow scheduler tag
312 * just for cheating put/get driver tag.
313 */
314 flush_rq->mq_ctx = first_rq->mq_ctx;
315 flush_rq->mq_hctx = first_rq->mq_hctx;
316
317 if (!q->elevator) {
318 flush_rq->tag = first_rq->tag;
319
320 /*
321 * We borrow data request's driver tag, so have to mark
322 * this flush request as INFLIGHT for avoiding double
323 * account of this driver tag
324 */
325 flush_rq->rq_flags |= RQF_MQ_INFLIGHT;
326 } else
327 flush_rq->internal_tag = first_rq->internal_tag;
328
329 flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
330 flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK);
331 flush_rq->rq_flags |= RQF_FLUSH_SEQ;
332 flush_rq->rq_disk = first_rq->rq_disk;
333 flush_rq->end_io = flush_end_io;
334
335 blk_flush_queue_rq(flush_rq, false);
336 }
337
mq_flush_data_end_io(struct request * rq,blk_status_t error)338 static void mq_flush_data_end_io(struct request *rq, 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 blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
357 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
358
359 blk_mq_sched_restart(hctx);
360 }
361
362 /**
363 * blk_insert_flush - insert a new PREFLUSH/FUA request
364 * @rq: request to insert
365 *
366 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
367 * or __blk_mq_run_hw_queue() to dispatch request.
368 * @rq is being submitted. Analyze what needs to be done and put it on the
369 * right queue.
370 */
blk_insert_flush(struct request * rq)371 void blk_insert_flush(struct request *rq)
372 {
373 struct request_queue *q = rq->q;
374 unsigned long fflags = q->queue_flags; /* may change, cache */
375 unsigned int policy = blk_flush_policy(fflags, rq);
376 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
377
378 /*
379 * @policy now records what operations need to be done. Adjust
380 * REQ_PREFLUSH and FUA for the driver.
381 */
382 rq->cmd_flags &= ~REQ_PREFLUSH;
383 if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
384 rq->cmd_flags &= ~REQ_FUA;
385
386 /*
387 * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
388 * of those flags, we have to set REQ_SYNC to avoid skewing
389 * the request accounting.
390 */
391 rq->cmd_flags |= REQ_SYNC;
392
393 /*
394 * An empty flush handed down from a stacking driver may
395 * translate into nothing if the underlying device does not
396 * advertise a write-back cache. In this case, simply
397 * complete the request.
398 */
399 if (!policy) {
400 blk_mq_end_request(rq, 0);
401 return;
402 }
403
404 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
405
406 /*
407 * If there's data but flush is not necessary, the request can be
408 * processed directly without going through flush machinery. Queue
409 * for normal execution.
410 */
411 if ((policy & REQ_FSEQ_DATA) &&
412 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
413 blk_mq_request_bypass_insert(rq, false, false);
414 return;
415 }
416
417 /*
418 * @rq should go through flush machinery. Mark it part of flush
419 * sequence and submit for further processing.
420 */
421 memset(&rq->flush, 0, sizeof(rq->flush));
422 INIT_LIST_HEAD(&rq->flush.list);
423 rq->rq_flags |= RQF_FLUSH_SEQ;
424 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
425
426 rq->end_io = mq_flush_data_end_io;
427
428 spin_lock_irq(&fq->mq_flush_lock);
429 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
430 spin_unlock_irq(&fq->mq_flush_lock);
431 }
432
433 /**
434 * blkdev_issue_flush - queue a flush
435 * @bdev: blockdev to issue flush for
436 * @gfp_mask: memory allocation flags (for bio_alloc)
437 *
438 * Description:
439 * Issue a flush for the block device in question.
440 */
blkdev_issue_flush(struct block_device * bdev,gfp_t gfp_mask)441 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask)
442 {
443 struct bio *bio;
444 int ret = 0;
445
446 bio = bio_alloc(gfp_mask, 0);
447 bio_set_dev(bio, bdev);
448 bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
449
450 ret = submit_bio_wait(bio);
451 bio_put(bio);
452 return ret;
453 }
454 EXPORT_SYMBOL(blkdev_issue_flush);
455
blk_alloc_flush_queue(int node,int cmd_size,gfp_t flags)456 struct blk_flush_queue *blk_alloc_flush_queue(int node, int cmd_size,
457 gfp_t flags)
458 {
459 struct blk_flush_queue *fq;
460 int rq_sz = sizeof(struct request);
461
462 fq = kzalloc_node(sizeof(*fq), flags, node);
463 if (!fq)
464 goto fail;
465
466 spin_lock_init(&fq->mq_flush_lock);
467
468 rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
469 fq->flush_rq = kzalloc_node(rq_sz, flags, node);
470 if (!fq->flush_rq)
471 goto fail_rq;
472
473 INIT_LIST_HEAD(&fq->flush_queue[0]);
474 INIT_LIST_HEAD(&fq->flush_queue[1]);
475 INIT_LIST_HEAD(&fq->flush_data_in_flight);
476
477 lockdep_register_key(&fq->key);
478 lockdep_set_class(&fq->mq_flush_lock, &fq->key);
479
480 return fq;
481
482 fail_rq:
483 kfree(fq);
484 fail:
485 return NULL;
486 }
487
blk_free_flush_queue(struct blk_flush_queue * fq)488 void blk_free_flush_queue(struct blk_flush_queue *fq)
489 {
490 /* bio based request queue hasn't flush queue */
491 if (!fq)
492 return;
493
494 lockdep_unregister_key(&fq->key);
495 kfree(fq->flush_rq);
496 kfree(fq);
497 }
498
