Branch data 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/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 : 28 : static unsigned int blk_flush_policy(unsigned long fflags, struct request *rq)
100 : : {
101 : 28 : unsigned int policy = 0;
102 : :
103 : 28 : if (blk_rq_sectors(rq))
104 : 28 : policy |= REQ_FSEQ_DATA;
105 : :
106 [ + - ]: 28 : if (fflags & (1UL << QUEUE_FLAG_WC)) {
107 [ + + ]: 28 : if (rq->cmd_flags & REQ_PREFLUSH)
108 : 22 : policy |= REQ_FSEQ_PREFLUSH;
109 [ + - + - ]: 28 : if (!(fflags & (1UL << QUEUE_FLAG_FUA)) &&
110 : : (rq->cmd_flags & REQ_FUA))
111 : 28 : policy |= REQ_FSEQ_POSTFLUSH;
112 : : }
113 : 28 : return policy;
114 : : }
115 : :
116 : 156 : static unsigned int blk_flush_cur_seq(struct request *rq)
117 : : {
118 : 156 : return 1 << ffz(rq->flush.seq);
119 : : }
120 : :
121 : 28 : 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 : 28 : rq->bio = rq->biotail;
129 : :
130 : : /* make @rq a normal request */
131 : 28 : rq->rq_flags &= ~RQF_FLUSH_SEQ;
132 : 28 : rq->end_io = rq->flush.saved_end_io;
133 : : }
134 : :
135 : 78 : static void blk_flush_queue_rq(struct request *rq, bool add_front)
136 : : {
137 : 78 : blk_mq_add_to_requeue_list(rq, add_front, true);
138 : 78 : }
139 : :
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 : : *
164 : : * RETURNS:
165 : : * %true if requests were added to the dispatch queue, %false otherwise.
166 : : */
167 : 106 : static void blk_flush_complete_seq(struct request *rq,
168 : : struct blk_flush_queue *fq,
169 : : unsigned int seq, blk_status_t error)
170 : : {
171 : 106 : struct request_queue *q = rq->q;
172 : 106 : struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
173 : 106 : unsigned int cmd_flags;
174 : :
175 [ - + ]: 106 : BUG_ON(rq->flush.seq & seq);
176 : 106 : rq->flush.seq |= seq;
177 : 106 : cmd_flags = rq->cmd_flags;
178 : :
179 [ + - ]: 106 : if (likely(!error))
180 [ + + + - ]: 106 : seq = blk_flush_cur_seq(rq);
181 : : else
182 : : seq = REQ_FSEQ_DONE;
183 : :
184 [ + + + - ]: 106 : switch (seq) {
185 : : case REQ_FSEQ_PREFLUSH:
186 : : case REQ_FSEQ_POSTFLUSH:
187 : : /* queue for flush */
188 [ + - ]: 50 : if (list_empty(pending))
189 : 50 : fq->flush_pending_since = jiffies;
190 : 50 : list_move_tail(&rq->flush.list, pending);
191 : : break;
192 : :
193 : 28 : case REQ_FSEQ_DATA:
194 : 28 : list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
195 : 28 : blk_flush_queue_rq(rq, true);
196 : : break;
197 : :
198 : 28 : case REQ_FSEQ_DONE:
199 : : /*
200 : : * @rq was previously adjusted by blk_insert_flush() for
201 : : * flush sequencing and may already have gone through the
202 : : * flush data request completion path. Restore @rq for
203 : : * normal completion and end it.
204 : : */
205 [ - + ]: 28 : BUG_ON(!list_empty(&rq->queuelist));
206 : 28 : list_del_init(&rq->flush.list);
207 : 28 : blk_flush_restore_request(rq);
208 : 28 : blk_mq_end_request(rq, error);
209 : 28 : break;
210 : :
211 : 0 : default:
212 : 0 : BUG();
213 : : }
214 : :
215 : 106 : blk_kick_flush(q, fq, cmd_flags);
216 : 106 : }
217 : :
218 : 50 : static void flush_end_io(struct request *flush_rq, blk_status_t error)
219 : : {
220 : 50 : struct request_queue *q = flush_rq->q;
221 : 50 : struct list_head *running;
222 : 50 : struct request *rq, *n;
223 : 50 : unsigned long flags = 0;
224 : 50 : struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
225 : 50 : struct blk_mq_hw_ctx *hctx;
226 : :
227 : 50 : blk_account_io_flush(flush_rq);
228 : :
229 : : /* release the tag's ownership to the req cloned from */
230 : 50 : spin_lock_irqsave(&fq->mq_flush_lock, flags);
231 : :
232 [ - + ]: 50 : if (!refcount_dec_and_test(&flush_rq->ref)) {
233 : 0 : fq->rq_status = error;
234 : 0 : spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
235 : 0 : return;
236 : : }
237 : :
238 [ - + ]: 50 : if (fq->rq_status != BLK_STS_OK)
239 : 0 : error = fq->rq_status;
240 : :
241 : 50 : hctx = flush_rq->mq_hctx;
242 [ - + ]: 50 : if (!q->elevator) {
243 : 0 : blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq);
244 : 0 : flush_rq->tag = -1;
245 : : } else {
246 [ + - ]: 50 : blk_mq_put_driver_tag(flush_rq);
247 : 50 : flush_rq->internal_tag = -1;
248 : : }
249 : :
250 : 50 : running = &fq->flush_queue[fq->flush_running_idx];
251 [ - + ]: 50 : BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
252 : :
253 : : /* account completion of the flush request */
254 : 50 : fq->flush_running_idx ^= 1;
255 : :
256 : : /* and push the waiting requests to the next stage */
257 [ + + ]: 100 : list_for_each_entry_safe(rq, n, running, flush.list) {
258 [ - + ]: 50 : unsigned int seq = blk_flush_cur_seq(rq);
259 : :
260 [ - + ]: 50 : BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
261 : 50 : blk_flush_complete_seq(rq, fq, seq, error);
262 : : }
263 : :
264 : 50 : fq->flush_queue_delayed = 0;
265 : 50 : spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
266 : : }
267 : :
268 : : /**
269 : : * blk_kick_flush - consider issuing flush request
270 : : * @q: request_queue being kicked
271 : : * @fq: flush queue
272 : : * @flags: cmd_flags of the original request
273 : : *
274 : : * Flush related states of @q have changed, consider issuing flush request.
275 : : * Please read the comment at the top of this file for more info.
276 : : *
277 : : * CONTEXT:
278 : : * spin_lock_irq(fq->mq_flush_lock)
279 : : *
280 : : */
281 : 106 : static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
282 : : unsigned int flags)
283 : : {
284 : 106 : struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
285 : 106 : struct request *first_rq =
286 : 106 : list_first_entry(pending, struct request, flush.list);
287 : 106 : struct request *flush_rq = fq->flush_rq;
288 : :
289 : : /* C1 described at the top of this file */
290 [ + - + + ]: 106 : if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
291 : : return;
292 : :
293 : : /* C2 and C3
294 : : *
295 : : * For blk-mq + scheduling, we can risk having all driver tags
296 : : * assigned to empty flushes, and we deadlock if we are expecting
297 : : * other requests to make progress. Don't defer for that case.
298 : : */
299 [ - + - - ]: 50 : if (!list_empty(&fq->flush_data_in_flight) && q->elevator &&
300 [ # # ]: 0 : time_before(jiffies,
301 : : fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
302 : : return;
303 : :
304 : : /*
305 : : * Issue flush and toggle pending_idx. This makes pending_idx
306 : : * different from running_idx, which means flush is in flight.
307 : : */
308 : 50 : fq->flush_pending_idx ^= 1;
309 : :
310 : 50 : blk_rq_init(q, flush_rq);
311 : :
312 : : /*
313 : : * In case of none scheduler, borrow tag from the first request
314 : : * since they can't be in flight at the same time. And acquire
315 : : * the tag's ownership for flush req.
316 : : *
317 : : * In case of IO scheduler, flush rq need to borrow scheduler tag
318 : : * just for cheating put/get driver tag.
319 : : */
320 : 50 : flush_rq->mq_ctx = first_rq->mq_ctx;
321 : 50 : flush_rq->mq_hctx = first_rq->mq_hctx;
322 : :
323 [ - + ]: 50 : if (!q->elevator) {
324 : 0 : fq->orig_rq = first_rq;
325 : 0 : flush_rq->tag = first_rq->tag;
326 : 0 : blk_mq_tag_set_rq(flush_rq->mq_hctx, first_rq->tag, flush_rq);
327 : : } else {
328 : 50 : flush_rq->internal_tag = first_rq->internal_tag;
329 : : }
330 : :
331 : 50 : flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
332 : 50 : flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK);
333 : 50 : flush_rq->rq_flags |= RQF_FLUSH_SEQ;
334 : 50 : flush_rq->rq_disk = first_rq->rq_disk;
335 : 50 : flush_rq->end_io = flush_end_io;
336 : :
337 : 50 : blk_flush_queue_rq(flush_rq, false);
338 : : }
339 : :
340 : 28 : static void mq_flush_data_end_io(struct request *rq, blk_status_t error)
341 : : {
342 : 28 : struct request_queue *q = rq->q;
343 : 28 : struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
344 : 28 : struct blk_mq_ctx *ctx = rq->mq_ctx;
345 : 28 : unsigned long flags;
346 [ + - ]: 28 : struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
347 : :
348 [ + - ]: 28 : if (q->elevator) {
349 [ - + ]: 28 : WARN_ON(rq->tag < 0);
350 [ + - ]: 28 : blk_mq_put_driver_tag(rq);
351 : : }
352 : :
353 : : /*
354 : : * After populating an empty queue, kick it to avoid stall. Read
355 : : * the comment in flush_end_io().
356 : : */
357 : 28 : spin_lock_irqsave(&fq->mq_flush_lock, flags);
358 : 28 : blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
359 : 28 : spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
360 : :
361 : 28 : blk_mq_sched_restart(hctx);
362 : 28 : }
363 : :
364 : : /**
365 : : * blk_insert_flush - insert a new PREFLUSH/FUA request
366 : : * @rq: request to insert
367 : : *
368 : : * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
369 : : * or __blk_mq_run_hw_queue() to dispatch request.
370 : : * @rq is being submitted. Analyze what needs to be done and put it on the
371 : : * right queue.
372 : : */
373 : 28 : void blk_insert_flush(struct request *rq)
374 : : {
375 : 28 : struct request_queue *q = rq->q;
376 : 28 : unsigned long fflags = q->queue_flags; /* may change, cache */
377 [ + - ]: 28 : unsigned int policy = blk_flush_policy(fflags, rq);
378 [ + - ]: 28 : struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
379 : :
380 : : /*
381 : : * @policy now records what operations need to be done. Adjust
382 : : * REQ_PREFLUSH and FUA for the driver.
383 : : */
384 : 28 : rq->cmd_flags &= ~REQ_PREFLUSH;
385 [ + - ]: 28 : if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
386 : 28 : rq->cmd_flags &= ~REQ_FUA;
387 : :
388 : : /*
389 : : * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
390 : : * of those flags, we have to set REQ_SYNC to avoid skewing
391 : : * the request accounting.
392 : : */
393 : 28 : rq->cmd_flags |= REQ_SYNC;
394 : :
395 : : /*
396 : : * An empty flush handed down from a stacking driver may
397 : : * translate into nothing if the underlying device does not
398 : : * advertise a write-back cache. In this case, simply
399 : : * complete the request.
400 : : */
401 [ - + ]: 28 : if (!policy) {
402 : 0 : blk_mq_end_request(rq, 0);
403 : 0 : return;
404 : : }
405 : :
406 [ - + ]: 28 : BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
407 : :
408 : : /*
409 : : * If there's data but flush is not necessary, the request can be
410 : : * processed directly without going through flush machinery. Queue
411 : : * for normal execution.
412 : : */
413 [ - + ]: 28 : if ((policy & REQ_FSEQ_DATA) &&
414 : : !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
415 : 0 : blk_mq_request_bypass_insert(rq, false, false);
416 : 0 : return;
417 : : }
418 : :
419 : : /*
420 : : * @rq should go through flush machinery. Mark it part of flush
421 : : * sequence and submit for further processing.
422 : : */
423 : 28 : memset(&rq->flush, 0, sizeof(rq->flush));
424 : 28 : INIT_LIST_HEAD(&rq->flush.list);
425 : 28 : rq->rq_flags |= RQF_FLUSH_SEQ;
426 : 28 : rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
427 : :
428 : 28 : rq->end_io = mq_flush_data_end_io;
429 : :
430 : 28 : spin_lock_irq(&fq->mq_flush_lock);
431 : 28 : blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
432 : 28 : spin_unlock_irq(&fq->mq_flush_lock);
433 : : }
434 : :
435 : : /**
436 : : * blkdev_issue_flush - queue a flush
437 : : * @bdev: blockdev to issue flush for
438 : : * @gfp_mask: memory allocation flags (for bio_alloc)
439 : : * @error_sector: error sector
440 : : *
441 : : * Description:
442 : : * Issue a flush for the block device in question. Caller can supply
443 : : * room for storing the error offset in case of a flush error, if they
444 : : * wish to.
445 : : */
446 : 0 : int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
447 : : sector_t *error_sector)
448 : : {
449 : 0 : struct request_queue *q;
450 : 0 : struct bio *bio;
451 : 0 : int ret = 0;
452 : :
453 [ # # ]: 0 : if (bdev->bd_disk == NULL)
454 : : return -ENXIO;
455 : :
456 [ # # ]: 0 : q = bdev_get_queue(bdev);
457 [ # # ]: 0 : if (!q)
458 : : return -ENXIO;
459 : :
460 : : /*
461 : : * some block devices may not have their queue correctly set up here
462 : : * (e.g. loop device without a backing file) and so issuing a flush
463 : : * here will panic. Ensure there is a request function before issuing
464 : : * the flush.
465 : : */
466 [ # # ]: 0 : if (!q->make_request_fn)
467 : : return -ENXIO;
468 : :
469 : 0 : bio = bio_alloc(gfp_mask, 0);
470 [ # # ]: 0 : bio_set_dev(bio, bdev);
471 : 0 : bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
472 : :
473 : 0 : ret = submit_bio_wait(bio);
474 : :
475 : : /*
476 : : * The driver must store the error location in ->bi_sector, if
477 : : * it supports it. For non-stacked drivers, this should be
478 : : * copied from blk_rq_pos(rq).
479 : : */
480 [ # # ]: 0 : if (error_sector)
481 : 0 : *error_sector = bio->bi_iter.bi_sector;
482 : :
483 : 0 : bio_put(bio);
484 : 0 : return ret;
485 : : }
486 : : EXPORT_SYMBOL(blkdev_issue_flush);
487 : :
488 : 33 : struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
489 : : int node, int cmd_size, gfp_t flags)
490 : : {
491 : 33 : struct blk_flush_queue *fq;
492 : 33 : int rq_sz = sizeof(struct request);
493 : :
494 : 33 : fq = kzalloc_node(sizeof(*fq), flags, node);
495 [ - + ]: 33 : if (!fq)
496 : 0 : goto fail;
497 : :
498 : 33 : spin_lock_init(&fq->mq_flush_lock);
499 : :
500 : 33 : rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
501 : 33 : fq->flush_rq = kzalloc_node(rq_sz, flags, node);
502 [ - + ]: 33 : if (!fq->flush_rq)
503 : 0 : goto fail_rq;
504 : :
505 : 33 : INIT_LIST_HEAD(&fq->flush_queue[0]);
506 : 33 : INIT_LIST_HEAD(&fq->flush_queue[1]);
507 : 33 : INIT_LIST_HEAD(&fq->flush_data_in_flight);
508 : :
509 : 33 : lockdep_register_key(&fq->key);
510 : 33 : lockdep_set_class(&fq->mq_flush_lock, &fq->key);
511 : :
512 : 33 : return fq;
513 : :
514 : : fail_rq:
515 : 0 : kfree(fq);
516 : : fail:
517 : : return NULL;
518 : : }
519 : :
520 : 0 : void blk_free_flush_queue(struct blk_flush_queue *fq)
521 : : {
522 : : /* bio based request queue hasn't flush queue */
523 [ # # ]: 0 : if (!fq)
524 : : return;
525 : :
526 : 0 : lockdep_unregister_key(&fq->key);
527 : 0 : kfree(fq->flush_rq);
528 : 0 : kfree(fq);
529 : : }
|
