linux/block/blk-flush.c
<<
>>
Prefs
   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
  73#include "blk.h"
  74
  75/* FLUSH/FUA sequences */
  76enum {
  77        REQ_FSEQ_PREFLUSH       = (1 << 0), /* pre-flushing in progress */
  78        REQ_FSEQ_DATA           = (1 << 1), /* data write in progress */
  79        REQ_FSEQ_POSTFLUSH      = (1 << 2), /* post-flushing in progress */
  80        REQ_FSEQ_DONE           = (1 << 3),
  81
  82        REQ_FSEQ_ACTIONS        = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA |
  83                                  REQ_FSEQ_POSTFLUSH,
  84
  85        /*
  86         * If flush has been pending longer than the following timeout,
  87         * it's issued even if flush_data requests are still in flight.
  88         */
  89        FLUSH_PENDING_TIMEOUT   = 5 * HZ,
  90};
  91
  92static bool blk_kick_flush(struct request_queue *q);
  93
  94static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
  95{
  96        unsigned int policy = 0;
  97
  98        if (fflags & REQ_FLUSH) {
  99                if (rq->cmd_flags & REQ_FLUSH)
 100                        policy |= REQ_FSEQ_PREFLUSH;
 101                if (blk_rq_sectors(rq))
 102                        policy |= REQ_FSEQ_DATA;
 103                if (!(fflags & REQ_FUA) && (rq->cmd_flags & REQ_FUA))
 104                        policy |= REQ_FSEQ_POSTFLUSH;
 105        }
 106        return policy;
 107}
 108
 109static unsigned int blk_flush_cur_seq(struct request *rq)
 110{
 111        return 1 << ffz(rq->flush.seq);
 112}
 113
 114static void blk_flush_restore_request(struct request *rq)
 115{
 116        /*
 117         * After flush data completion, @rq->bio is %NULL but we need to
 118         * complete the bio again.  @rq->biotail is guaranteed to equal the
 119         * original @rq->bio.  Restore it.
 120         */
 121        rq->bio = rq->biotail;
 122
 123        /* make @rq a normal request */
 124        rq->cmd_flags &= ~REQ_FLUSH_SEQ;
 125        rq->end_io = NULL;
 126}
 127
 128/**
 129 * blk_flush_complete_seq - complete flush sequence
 130 * @rq: FLUSH/FUA request being sequenced
 131 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
 132 * @error: whether an error occurred
 133 *
 134 * @rq just completed @seq part of its flush sequence, record the
 135 * completion and trigger the next step.
 136 *
 137 * CONTEXT:
 138 * spin_lock_irq(q->queue_lock)
 139 *
 140 * RETURNS:
 141 * %true if requests were added to the dispatch queue, %false otherwise.
 142 */
 143static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
 144                                   int error)
 145{
 146        struct request_queue *q = rq->q;
 147        struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
 148        bool queued = false;
 149
 150        BUG_ON(rq->flush.seq & seq);
 151        rq->flush.seq |= seq;
 152
 153        if (likely(!error))
 154                seq = blk_flush_cur_seq(rq);
 155        else
 156                seq = REQ_FSEQ_DONE;
 157
 158        switch (seq) {
 159        case REQ_FSEQ_PREFLUSH:
 160        case REQ_FSEQ_POSTFLUSH:
 161                /* queue for flush */
 162                if (list_empty(pending))
 163                        q->flush_pending_since = jiffies;
 164                list_move_tail(&rq->flush.list, pending);
 165                break;
 166
 167        case REQ_FSEQ_DATA:
 168                list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
 169                list_add(&rq->queuelist, &q->queue_head);
 170                queued = true;
 171                break;
 172
 173        case REQ_FSEQ_DONE:
 174                /*
 175                 * @rq was previously adjusted by blk_flush_issue() for
 176                 * flush sequencing and may already have gone through the
 177                 * flush data request completion path.  Restore @rq for
 178                 * normal completion and end it.
 179                 */
 180                BUG_ON(!list_empty(&rq->queuelist));
 181                list_del_init(&rq->flush.list);
 182                blk_flush_restore_request(rq);
 183                __blk_end_request_all(rq, error);
 184                break;
 185
 186        default:
 187                BUG();
 188        }
 189
 190        return blk_kick_flush(q) | queued;
 191}
 192
 193static void flush_end_io(struct request *flush_rq, int error)
 194{
 195        struct request_queue *q = flush_rq->q;
 196        struct list_head *running = &q->flush_queue[q->flush_running_idx];
 197        bool queued = false;
 198        struct request *rq, *n;
 199
 200        BUG_ON(q->flush_pending_idx == q->flush_running_idx);
 201
 202        /* account completion of the flush request */
 203        q->flush_running_idx ^= 1;
 204        elv_completed_request(q, flush_rq);
 205
 206        /* and push the waiting requests to the next stage */
 207        list_for_each_entry_safe(rq, n, running, flush.list) {
 208                unsigned int seq = blk_flush_cur_seq(rq);
 209
 210                BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
 211                queued |= blk_flush_complete_seq(rq, seq, error);
 212        }
 213
 214        /*
 215         * Kick the queue to avoid stall for two cases:
 216         * 1. Moving a request silently to empty queue_head may stall the
 217         * queue.
 218         * 2. When flush request is running in non-queueable queue, the
 219         * queue is hold. Restart the queue after flush request is finished
 220         * to avoid stall.
 221         * This function is called from request completion path and calling
 222         * directly into request_fn may confuse the driver.  Always use
 223         * kblockd.
 224         */
 225        if (queued || q->flush_queue_delayed)
 226                blk_run_queue_async(q);
 227        q->flush_queue_delayed = 0;
 228}
 229
 230/**
 231 * blk_kick_flush - consider issuing flush request
 232 * @q: request_queue being kicked
 233 *
 234 * Flush related states of @q have changed, consider issuing flush request.
 235 * Please read the comment at the top of this file for more info.
 236 *
 237 * CONTEXT:
 238 * spin_lock_irq(q->queue_lock)
 239 *
 240 * RETURNS:
 241 * %true if flush was issued, %false otherwise.
 242 */
 243static bool blk_kick_flush(struct request_queue *q)
 244{
 245        struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
 246        struct request *first_rq =
 247                list_first_entry(pending, struct request, flush.list);
 248
 249        /* C1 described at the top of this file */
 250        if (q->flush_pending_idx != q->flush_running_idx || list_empty(pending))
 251                return false;
 252
 253        /* C2 and C3 */
 254        if (!list_empty(&q->flush_data_in_flight) &&
 255            time_before(jiffies,
 256                        q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
 257                return false;
 258
 259        /*
 260         * Issue flush and toggle pending_idx.  This makes pending_idx
 261         * different from running_idx, which means flush is in flight.
 262         */
 263        blk_rq_init(q, &q->flush_rq);
 264        q->flush_rq.cmd_type = REQ_TYPE_FS;
 265        q->flush_rq.cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
 266        q->flush_rq.rq_disk = first_rq->rq_disk;
 267        q->flush_rq.end_io = flush_end_io;
 268
 269        q->flush_pending_idx ^= 1;
 270        list_add_tail(&q->flush_rq.queuelist, &q->queue_head);
 271        return true;
 272}
 273
 274static void flush_data_end_io(struct request *rq, int error)
 275{
 276        struct request_queue *q = rq->q;
 277
 278        /*
 279         * After populating an empty queue, kick it to avoid stall.  Read
 280         * the comment in flush_end_io().
 281         */
 282        if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
 283                blk_run_queue_async(q);
 284}
 285
 286/**
 287 * blk_insert_flush - insert a new FLUSH/FUA request
 288 * @rq: request to insert
 289 *
 290 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
 291 * @rq is being submitted.  Analyze what needs to be done and put it on the
 292 * right queue.
 293 *
 294 * CONTEXT:
 295 * spin_lock_irq(q->queue_lock)
 296 */
 297void blk_insert_flush(struct request *rq)
 298{
 299        struct request_queue *q = rq->q;
 300        unsigned int fflags = q->flush_flags;   /* may change, cache */
 301        unsigned int policy = blk_flush_policy(fflags, rq);
 302
 303        BUG_ON(rq->end_io);
 304        BUG_ON(!rq->bio || rq->bio != rq->biotail);
 305
 306        /*
 307         * @policy now records what operations need to be done.  Adjust
 308         * REQ_FLUSH and FUA for the driver.
 309         */
 310        rq->cmd_flags &= ~REQ_FLUSH;
 311        if (!(fflags & REQ_FUA))
 312                rq->cmd_flags &= ~REQ_FUA;
 313
 314        /*
 315         * If there's data but flush is not necessary, the request can be
 316         * processed directly without going through flush machinery.  Queue
 317         * for normal execution.
 318         */
 319        if ((policy & REQ_FSEQ_DATA) &&
 320            !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
 321                list_add_tail(&rq->queuelist, &q->queue_head);
 322                return;
 323        }
 324
 325        /*
 326         * @rq should go through flush machinery.  Mark it part of flush
 327         * sequence and submit for further processing.
 328         */
 329        memset(&rq->flush, 0, sizeof(rq->flush));
 330        INIT_LIST_HEAD(&rq->flush.list);
 331        rq->cmd_flags |= REQ_FLUSH_SEQ;
 332        rq->end_io = flush_data_end_io;
 333
 334        blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
 335}
 336
 337/**
 338 * blk_abort_flushes - @q is being aborted, abort flush requests
 339 * @q: request_queue being aborted
 340 *
 341 * To be called from elv_abort_queue().  @q is being aborted.  Prepare all
 342 * FLUSH/FUA requests for abortion.
 343 *
 344 * CONTEXT:
 345 * spin_lock_irq(q->queue_lock)
 346 */
 347void blk_abort_flushes(struct request_queue *q)
 348{
 349        struct request *rq, *n;
 350        int i;
 351
 352        /*
 353         * Requests in flight for data are already owned by the dispatch
 354         * queue or the device driver.  Just restore for normal completion.
 355         */
 356        list_for_each_entry_safe(rq, n, &q->flush_data_in_flight, flush.list) {
 357                list_del_init(&rq->flush.list);
 358                blk_flush_restore_request(rq);
 359        }
 360
 361        /*
 362         * We need to give away requests on flush queues.  Restore for
 363         * normal completion and put them on the dispatch queue.
 364         */
 365        for (i = 0; i < ARRAY_SIZE(q->flush_queue); i++) {
 366                list_for_each_entry_safe(rq, n, &q->flush_queue[i],
 367                                         flush.list) {
 368                        list_del_init(&rq->flush.list);
 369                        blk_flush_restore_request(rq);
 370                        list_add_tail(&rq->queuelist, &q->queue_head);
 371                }
 372        }
 373}
 374
 375static void bio_end_flush(struct bio *bio, int err)
 376{
 377        if (err)
 378                clear_bit(BIO_UPTODATE, &bio->bi_flags);
 379        if (bio->bi_private)
 380                complete(bio->bi_private);
 381        bio_put(bio);
 382}
 383
 384/**
 385 * blkdev_issue_flush - queue a flush
 386 * @bdev:       blockdev to issue flush for
 387 * @gfp_mask:   memory allocation flags (for bio_alloc)
 388 * @error_sector:       error sector
 389 *
 390 * Description:
 391 *    Issue a flush for the block device in question. Caller can supply
 392 *    room for storing the error offset in case of a flush error, if they
 393 *    wish to. If WAIT flag is not passed then caller may check only what
 394 *    request was pushed in some internal queue for later handling.
 395 */
 396int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
 397                sector_t *error_sector)
 398{
 399        DECLARE_COMPLETION_ONSTACK(wait);
 400        struct request_queue *q;
 401        struct bio *bio;
 402        int ret = 0;
 403
 404        if (bdev->bd_disk == NULL)
 405                return -ENXIO;
 406
 407        q = bdev_get_queue(bdev);
 408        if (!q)
 409                return -ENXIO;
 410
 411        /*
 412         * some block devices may not have their queue correctly set up here
 413         * (e.g. loop device without a backing file) and so issuing a flush
 414         * here will panic. Ensure there is a request function before issuing
 415         * the flush.
 416         */
 417        if (!q->make_request_fn)
 418                return -ENXIO;
 419
 420        bio = bio_alloc(gfp_mask, 0);
 421        bio->bi_end_io = bio_end_flush;
 422        bio->bi_bdev = bdev;
 423        bio->bi_private = &wait;
 424
 425        bio_get(bio);
 426        submit_bio(WRITE_FLUSH, bio);
 427        wait_for_completion(&wait);
 428
 429        /*
 430         * The driver must store the error location in ->bi_sector, if
 431         * it supports it. For non-stacked drivers, this should be
 432         * copied from blk_rq_pos(rq).
 433         */
 434        if (error_sector)
 435               *error_sector = bio->bi_sector;
 436
 437        if (!bio_flagged(bio, BIO_UPTODATE))
 438                ret = -EIO;
 439
 440        bio_put(bio);
 441        return ret;
 442}
 443EXPORT_SYMBOL(blkdev_issue_flush);
 444