Root/block/blk-flush.c

1/*
2 * Functions to sequence FLUSH and FUA writes.
3 *
4 * Copyright (C) 2011 Max Planck Institute for Gravitational Physics
5 * Copyright (C) 2011 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * REQ_{FLUSH|FUA} requests are decomposed to sequences consisted of three
10 * optional steps - PREFLUSH, DATA and POSTFLUSH - according to the request
11 * properties and hardware capability.
12 *
13 * If a request doesn't have data, only REQ_FLUSH makes sense, which
14 * indicates a simple flush request. If there is data, REQ_FLUSH indicates
15 * that the device cache should be flushed before the data is executed, and
16 * REQ_FUA means that the data must be on non-volatile media on request
17 * completion.
18 *
19 * If the device doesn't have writeback cache, FLUSH and FUA don't make any
20 * difference. The requests are either completed immediately if there's no
21 * data or executed as normal requests otherwise.
22 *
23 * If the device has writeback cache and supports FUA, REQ_FLUSH is
24 * translated to PREFLUSH but REQ_FUA is passed down directly with DATA.
25 *
26 * If the device has writeback cache and doesn't support FUA, REQ_FLUSH is
27 * translated to PREFLUSH and REQ_FUA to POSTFLUSH.
28 *
29 * The actual execution of flush is double buffered. Whenever a request
30 * needs to execute PRE or POSTFLUSH, it queues at
31 * q->flush_queue[q->flush_pending_idx]. Once certain criteria are met, a
32 * flush is issued and the pending_idx is toggled. When the flush
33 * completes, all the requests which were pending are proceeded to the next
34 * step. This allows arbitrary merging of different types of FLUSH/FUA
35 * requests.
36 *
37 * Currently, the following conditions are used to determine when to issue
38 * flush.
39 *
40 * C1. At any given time, only one flush shall be in progress. This makes
41 * double buffering sufficient.
42 *
43 * C2. Flush is deferred if any request is executing DATA of its sequence.
44 * This avoids issuing separate POSTFLUSHes for requests which shared
45 * PREFLUSH.
46 *
47 * C3. The second condition is ignored if there is a request which has
48 * waited longer than FLUSH_PENDING_TIMEOUT. This is to avoid
49 * starvation in the unlikely case where there are continuous stream of
50 * FUA (without FLUSH) requests.
51 *
52 * For devices which support FUA, it isn't clear whether C2 (and thus C3)
53 * is beneficial.
54 *
55 * Note that a sequenced FLUSH/FUA request with DATA is completed twice.
56 * Once while executing DATA and again after the whole sequence is
57 * complete. The first completion updates the contained bio but doesn't
58 * finish it so that the bio submitter is notified only after the whole
59 * sequence is complete. This is implemented by testing REQ_FLUSH_SEQ in
60 * req_bio_endio().
61 *
62 * The above peculiarity requires that each FLUSH/FUA request has only one
63 * bio attached to it, which is guaranteed as they aren't allowed to be
64 * merged in the usual way.
65 */
66
67#include <linux/kernel.h>
68#include <linux/module.h>
69#include <linux/bio.h>
70#include <linux/blkdev.h>
71#include <linux/gfp.h>
72#include <linux/blk-mq.h>
73
74#include "blk.h"
75#include "blk-mq.h"
76
77/* FLUSH/FUA sequences */
78enum {
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
94static bool blk_kick_flush(struct request_queue *q);
95
96static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
97{
98    unsigned int policy = 0;
99
100    if (blk_rq_sectors(rq))
101        policy |= REQ_FSEQ_DATA;
102
103    if (fflags & REQ_FLUSH) {
104        if (rq->cmd_flags & REQ_FLUSH)
105            policy |= REQ_FSEQ_PREFLUSH;
106        if (!(fflags & REQ_FUA) && (rq->cmd_flags & REQ_FUA))
107            policy |= REQ_FSEQ_POSTFLUSH;
108    }
109    return policy;
110}
111
112static unsigned int blk_flush_cur_seq(struct request *rq)
113{
114    return 1 << ffz(rq->flush.seq);
115}
116
117static void blk_flush_restore_request(struct request *rq)
118{
119    /*
120     * After flush data completion, @rq->bio is %NULL but we need to
121     * complete the bio again. @rq->biotail is guaranteed to equal the
122     * original @rq->bio. Restore it.
123     */
124    rq->bio = rq->biotail;
125
126    /* make @rq a normal request */
127    rq->cmd_flags &= ~REQ_FLUSH_SEQ;
128    rq->end_io = rq->flush.saved_end_io;
129
130    blk_clear_rq_complete(rq);
131}
132
133static bool blk_flush_queue_rq(struct request *rq, bool add_front)
134{
135    if (rq->q->mq_ops) {
136        struct request_queue *q = rq->q;
137
138        blk_mq_add_to_requeue_list(rq, add_front);
139        blk_mq_kick_requeue_list(q);
140        return false;
141    } else {
142        if (add_front)
143            list_add(&rq->queuelist, &rq->q->queue_head);
144        else
145            list_add_tail(&rq->queuelist, &rq->q->queue_head);
146        return true;
147    }
148}
149
150/**
151 * blk_flush_complete_seq - complete flush sequence
152 * @rq: FLUSH/FUA request being sequenced
153 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
154 * @error: whether an error occurred
155 *
156 * @rq just completed @seq part of its flush sequence, record the
157 * completion and trigger the next step.
158 *
159 * CONTEXT:
160 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
161 *
162 * RETURNS:
163 * %true if requests were added to the dispatch queue, %false otherwise.
164 */
165static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
166                   int error)
167{
168    struct request_queue *q = rq->q;
169    struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
170    bool queued = false, kicked;
171
172    BUG_ON(rq->flush.seq & seq);
173    rq->flush.seq |= seq;
174
175    if (likely(!error))
176        seq = blk_flush_cur_seq(rq);
177    else
178        seq = REQ_FSEQ_DONE;
179
180    switch (seq) {
181    case REQ_FSEQ_PREFLUSH:
182    case REQ_FSEQ_POSTFLUSH:
183        /* queue for flush */
184        if (list_empty(pending))
185            q->flush_pending_since = jiffies;
186        list_move_tail(&rq->flush.list, pending);
187        break;
188
189    case REQ_FSEQ_DATA:
190        list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
191        queued = blk_flush_queue_rq(rq, true);
192        break;
193
194    case REQ_FSEQ_DONE:
195        /*
196         * @rq was previously adjusted by blk_flush_issue() for
197         * flush sequencing and may already have gone through the
198         * flush data request completion path. Restore @rq for
199         * normal completion and end it.
200         */
201        BUG_ON(!list_empty(&rq->queuelist));
202        list_del_init(&rq->flush.list);
203        blk_flush_restore_request(rq);
204        if (q->mq_ops)
205            blk_mq_end_io(rq, error);
206        else
207            __blk_end_request_all(rq, error);
208        break;
209
210    default:
211        BUG();
212    }
213
214    kicked = blk_kick_flush(q);
215    return kicked | queued;
216}
217
218static void flush_end_io(struct request *flush_rq, int error)
219{
220    struct request_queue *q = flush_rq->q;
221    struct list_head *running;
222    bool queued = false;
223    struct request *rq, *n;
224    unsigned long flags = 0;
225
226    if (q->mq_ops) {
227        spin_lock_irqsave(&q->mq_flush_lock, flags);
228        q->flush_rq->tag = -1;
229    }
230
231    running = &q->flush_queue[q->flush_running_idx];
232    BUG_ON(q->flush_pending_idx == q->flush_running_idx);
233
234    /* account completion of the flush request */
235    q->flush_running_idx ^= 1;
236
237    if (!q->mq_ops)
238        elv_completed_request(q, flush_rq);
239
240    /* and push the waiting requests to the next stage */
241    list_for_each_entry_safe(rq, n, running, flush.list) {
242        unsigned int seq = blk_flush_cur_seq(rq);
243
244        BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
245        queued |= blk_flush_complete_seq(rq, seq, error);
246    }
247
248    /*
249     * Kick the queue to avoid stall for two cases:
250     * 1. Moving a request silently to empty queue_head may stall the
251     * queue.
252     * 2. When flush request is running in non-queueable queue, the
253     * queue is hold. Restart the queue after flush request is finished
254     * to avoid stall.
255     * This function is called from request completion path and calling
256     * directly into request_fn may confuse the driver. Always use
257     * kblockd.
258     */
259    if (queued || q->flush_queue_delayed) {
260        WARN_ON(q->mq_ops);
261        blk_run_queue_async(q);
262    }
263    q->flush_queue_delayed = 0;
264    if (q->mq_ops)
265        spin_unlock_irqrestore(&q->mq_flush_lock, flags);
266}
267
268/**
269 * blk_kick_flush - consider issuing flush request
270 * @q: request_queue being kicked
271 *
272 * Flush related states of @q have changed, consider issuing flush request.
273 * Please read the comment at the top of this file for more info.
274 *
275 * CONTEXT:
276 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
277 *
278 * RETURNS:
279 * %true if flush was issued, %false otherwise.
280 */
281static bool blk_kick_flush(struct request_queue *q)
282{
283    struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
284    struct request *first_rq =
285        list_first_entry(pending, struct request, flush.list);
286
287    /* C1 described at the top of this file */
288    if (q->flush_pending_idx != q->flush_running_idx || list_empty(pending))
289        return false;
290
291    /* C2 and C3 */
292    if (!list_empty(&q->flush_data_in_flight) &&
293        time_before(jiffies,
294            q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
295        return false;
296
297    /*
298     * Issue flush and toggle pending_idx. This makes pending_idx
299     * different from running_idx, which means flush is in flight.
300     */
301    q->flush_pending_idx ^= 1;
302
303    blk_rq_init(q, q->flush_rq);
304    if (q->mq_ops)
305        blk_mq_clone_flush_request(q->flush_rq, first_rq);
306
307    q->flush_rq->cmd_type = REQ_TYPE_FS;
308    q->flush_rq->cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
309    q->flush_rq->rq_disk = first_rq->rq_disk;
310    q->flush_rq->end_io = flush_end_io;
311
312    return blk_flush_queue_rq(q->flush_rq, false);
313}
314
315static void flush_data_end_io(struct request *rq, int error)
316{
317    struct request_queue *q = rq->q;
318
319    /*
320     * After populating an empty queue, kick it to avoid stall. Read
321     * the comment in flush_end_io().
322     */
323    if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
324        blk_run_queue_async(q);
325}
326
327static void mq_flush_data_end_io(struct request *rq, int error)
328{
329    struct request_queue *q = rq->q;
330    struct blk_mq_hw_ctx *hctx;
331    struct blk_mq_ctx *ctx;
332    unsigned long flags;
333
334    ctx = rq->mq_ctx;
335    hctx = q->mq_ops->map_queue(q, ctx->cpu);
336
337    /*
338     * After populating an empty queue, kick it to avoid stall. Read
339     * the comment in flush_end_io().
340     */
341    spin_lock_irqsave(&q->mq_flush_lock, flags);
342    if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
343        blk_mq_run_hw_queue(hctx, true);
344    spin_unlock_irqrestore(&q->mq_flush_lock, flags);
345}
346
347/**
348 * blk_insert_flush - insert a new FLUSH/FUA request
349 * @rq: request to insert
350 *
351 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
352 * or __blk_mq_run_hw_queue() to dispatch request.
353 * @rq is being submitted. Analyze what needs to be done and put it on the
354 * right queue.
355 *
356 * CONTEXT:
357 * spin_lock_irq(q->queue_lock) in !mq case
358 */
359void blk_insert_flush(struct request *rq)
360{
361    struct request_queue *q = rq->q;
362    unsigned int fflags = q->flush_flags; /* may change, cache */
363    unsigned int policy = blk_flush_policy(fflags, rq);
364
365    /*
366     * @policy now records what operations need to be done. Adjust
367     * REQ_FLUSH and FUA for the driver.
368     */
369    rq->cmd_flags &= ~REQ_FLUSH;
370    if (!(fflags & REQ_FUA))
371        rq->cmd_flags &= ~REQ_FUA;
372
373    /*
374     * An empty flush handed down from a stacking driver may
375     * translate into nothing if the underlying device does not
376     * advertise a write-back cache. In this case, simply
377     * complete the request.
378     */
379    if (!policy) {
380        if (q->mq_ops)
381            blk_mq_end_io(rq, 0);
382        else
383            __blk_end_bidi_request(rq, 0, 0, 0);
384        return;
385    }
386
387    BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
388
389    /*
390     * If there's data but flush is not necessary, the request can be
391     * processed directly without going through flush machinery. Queue
392     * for normal execution.
393     */
394    if ((policy & REQ_FSEQ_DATA) &&
395        !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
396        if (q->mq_ops) {
397            blk_mq_insert_request(rq, false, false, true);
398        } else
399            list_add_tail(&rq->queuelist, &q->queue_head);
400        return;
401    }
402
403    /*
404     * @rq should go through flush machinery. Mark it part of flush
405     * sequence and submit for further processing.
406     */
407    memset(&rq->flush, 0, sizeof(rq->flush));
408    INIT_LIST_HEAD(&rq->flush.list);
409    rq->cmd_flags |= REQ_FLUSH_SEQ;
410    rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
411    if (q->mq_ops) {
412        rq->end_io = mq_flush_data_end_io;
413
414        spin_lock_irq(&q->mq_flush_lock);
415        blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
416        spin_unlock_irq(&q->mq_flush_lock);
417        return;
418    }
419    rq->end_io = flush_data_end_io;
420
421    blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
422}
423
424/**
425 * blkdev_issue_flush - queue a flush
426 * @bdev: blockdev to issue flush for
427 * @gfp_mask: memory allocation flags (for bio_alloc)
428 * @error_sector: error sector
429 *
430 * Description:
431 * Issue a flush for the block device in question. Caller can supply
432 * room for storing the error offset in case of a flush error, if they
433 * wish to. If WAIT flag is not passed then caller may check only what
434 * request was pushed in some internal queue for later handling.
435 */
436int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
437        sector_t *error_sector)
438{
439    struct request_queue *q;
440    struct bio *bio;
441    int ret = 0;
442
443    if (bdev->bd_disk == NULL)
444        return -ENXIO;
445
446    q = bdev_get_queue(bdev);
447    if (!q)
448        return -ENXIO;
449
450    /*
451     * some block devices may not have their queue correctly set up here
452     * (e.g. loop device without a backing file) and so issuing a flush
453     * here will panic. Ensure there is a request function before issuing
454     * the flush.
455     */
456    if (!q->make_request_fn)
457        return -ENXIO;
458
459    bio = bio_alloc(gfp_mask, 0);
460    bio->bi_bdev = bdev;
461
462    ret = submit_bio_wait(WRITE_FLUSH, bio);
463
464    /*
465     * The driver must store the error location in ->bi_sector, if
466     * it supports it. For non-stacked drivers, this should be
467     * copied from blk_rq_pos(rq).
468     */
469    if (error_sector)
470        *error_sector = bio->bi_iter.bi_sector;
471
472    bio_put(bio);
473    return ret;
474}
475EXPORT_SYMBOL(blkdev_issue_flush);
476
477void blk_mq_init_flush(struct request_queue *q)
478{
479    spin_lock_init(&q->mq_flush_lock);
480}
481

Archive Download this file



interactive