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
  73 #include "blk.h"
  74
  75 /* FLUSH/FUA sequences */
  76 enum {
  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
  92 static bool blk_kick_flush(struct request_queue *q);
  93
  94 static unsigned int blk_flush_policy(unsigned int fflags, struct request *rq)
  95 {
  96         unsigned int policy = 0;
  97
  98         if (blk_rq_sectors(rq))
  99                 policy |= REQ_FSEQ_DATA;
 100
 101         if (fflags & REQ_FLUSH) {
 102                 if (rq->cmd_flags & REQ_FLUSH)
 103                         policy |= REQ_FSEQ_PREFLUSH;
 104                 if (!(fflags & REQ_FUA) && (rq->cmd_flags & REQ_FUA))
 105                         policy |= REQ_FSEQ_POSTFLUSH;
 106         }
 107         return policy;
 108 }
 109
 110 static unsigned int blk_flush_cur_seq(struct request *rq)
 111 {
 112         return 1 << ffz(rq->flush.seq);
 113 }
 114
 115 static void blk_flush_restore_request(struct request *rq)
 116 {
 117         /*
 118          * After flush data completion, @rq->bio is %NULL but we need to
 119          * complete the bio again.  @rq->biotail is guaranteed to equal the
 120          * original @rq->bio.  Restore it.
 121          */
 122         rq->bio = rq->biotail;
 123
 124         /* make @rq a normal request */
 125         rq->cmd_flags &= ~REQ_FLUSH_SEQ;
 126         rq->end_io = rq->flush.saved_end_io;
 127 }
 128
 129 /**
 130  * blk_flush_complete_seq - complete flush sequence
 131  * @rq: FLUSH/FUA request being sequenced
 132  * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
 133  * @error: whether an error occurred
 134  *
 135  * @rq just completed @seq part of its flush sequence, record the
 136  * completion and trigger the next step.
 137  *
 138  * CONTEXT:
 139  * spin_lock_irq(q->queue_lock)
 140  *
 141  * RETURNS:
 142  * %true if requests were added to the dispatch queue, %false otherwise.
 143  */
 144 static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
 145                                    int error)
 146 {
 147         struct request_queue *q = rq->q;
 148         struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
 149         bool queued = false;
 150
 151         BUG_ON(rq->flush.seq & seq);
 152         rq->flush.seq |= seq;
 153
 154         if (likely(!error))
 155                 seq = blk_flush_cur_seq(rq);
 156         else
 157                 seq = REQ_FSEQ_DONE;
 158
 159         switch (seq) {
 160         case REQ_FSEQ_PREFLUSH:
 161         case REQ_FSEQ_POSTFLUSH:
 162                 /* queue for flush */
 163                 if (list_empty(pending))
 164                         q->flush_pending_since = jiffies;
 165                 list_move_tail(&rq->flush.list, pending);
 166                 break;
 167
 168         case REQ_FSEQ_DATA:
 169                 list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
 170                 list_add(&rq->queuelist, &q->queue_head);
 171                 queued = true;
 172                 break;
 173
 174         case REQ_FSEQ_DONE:
 175                 /*
 176                  * @rq was previously adjusted by blk_flush_issue() for
 177                  * flush sequencing and may already have gone through the
 178                  * flush data request completion path.  Restore @rq for
 179                  * normal completion and end it.
 180                  */
 181                 BUG_ON(!list_empty(&rq->queuelist));
 182                 list_del_init(&rq->flush.list);
 183                 blk_flush_restore_request(rq);
 184                 __blk_end_request_all(rq, error);
 185                 break;
 186
 187         default:
 188                 BUG();
 189         }
 190
 191         return blk_kick_flush(q) | queued;
 192 }
 193
 194 static void flush_end_io(struct request *flush_rq, int error)
 195 {
 196         struct request_queue *q = flush_rq->q;
 197         struct list_head *running = &q->flush_queue[q->flush_running_idx];
 198         bool queued = false;
 199         struct request *rq, *n;
 200
 201         BUG_ON(q->flush_pending_idx == q->flush_running_idx);
 202
 203         /* account completion of the flush request */
 204         q->flush_running_idx ^= 1;
 205         elv_completed_request(q, flush_rq);
 206
 207         /* and push the waiting requests to the next stage */
 208         list_for_each_entry_safe(rq, n, running, flush.list) {
 209                 unsigned int seq = blk_flush_cur_seq(rq);
 210
 211                 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
 212                 queued |= blk_flush_complete_seq(rq, seq, error);
 213         }
 214
 215         /*
 216          * Kick the queue to avoid stall for two cases:
 217          * 1. Moving a request silently to empty queue_head may stall the
 218          * queue.
 219          * 2. When flush request is running in non-queueable queue, the
 220          * queue is hold. Restart the queue after flush request is finished
 221          * to avoid stall.
 222          * This function is called from request completion path and calling
 223          * directly into request_fn may confuse the driver.  Always use
 224          * kblockd.
 225          */
 226         if (queued || q->flush_queue_delayed)
 227                 blk_run_queue_async(q);
 228         q->flush_queue_delayed = 0;
 229 }
 230
 231 /**
 232  * blk_kick_flush - consider issuing flush request
 233  * @q: request_queue being kicked
 234  *
 235  * Flush related states of @q have changed, consider issuing flush request.
 236  * Please read the comment at the top of this file for more info.
 237  *
 238  * CONTEXT:
 239  * spin_lock_irq(q->queue_lock)
 240  *
 241  * RETURNS:
 242  * %true if flush was issued, %false otherwise.
 243  */
 244 static bool blk_kick_flush(struct request_queue *q)
 245 {
 246         struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
 247         struct request *first_rq =
 248                 list_first_entry(pending, struct request, flush.list);
 249
 250         /* C1 described at the top of this file */
 251         if (q->flush_pending_idx != q->flush_running_idx || list_empty(pending))
 252                 return false;
 253
 254         /* C2 and C3 */
 255         if (!list_empty(&q->flush_data_in_flight) &&
 256             time_before(jiffies,
 257                         q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
 258                 return false;
 259
 260         /*
 261          * Issue flush and toggle pending_idx.  This makes pending_idx
 262          * different from running_idx, which means flush is in flight.
 263          */
 264         blk_rq_init(q, &q->flush_rq);
 265         q->flush_rq.cmd_type = REQ_TYPE_FS;
 266         q->flush_rq.cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
 267         q->flush_rq.rq_disk = first_rq->rq_disk;
 268         q->flush_rq.end_io = flush_end_io;
 269
 270         q->flush_pending_idx ^= 1;
 271         list_add_tail(&q->flush_rq.queuelist, &q->queue_head);
 272         return true;
 273 }
 274
 275 static void flush_data_end_io(struct request *rq, int error)
 276 {
 277         struct request_queue *q = rq->q;
 278
 279         /*
 280          * After populating an empty queue, kick it to avoid stall.  Read
 281          * the comment in flush_end_io().
 282          */
 283         if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
 284                 blk_run_queue_async(q);
 285 }
 286
 287 /**
 288  * blk_insert_flush - insert a new FLUSH/FUA request
 289  * @rq: request to insert
 290  *
 291  * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
 292  * @rq is being submitted.  Analyze what needs to be done and put it on the
 293  * right queue.
 294  *
 295  * CONTEXT:
 296  * spin_lock_irq(q->queue_lock)
 297  */
 298 void blk_insert_flush(struct request *rq)
 299 {
 300         struct request_queue *q = rq->q;
 301         unsigned int fflags = q->flush_flags;   /* may change, cache */
 302         unsigned int policy = blk_flush_policy(fflags, rq);
 303
 304         /*
 305          * @policy now records what operations need to be done.  Adjust
 306          * REQ_FLUSH and FUA for the driver.
 307          */
 308         rq->cmd_flags &= ~REQ_FLUSH;
 309         if (!(fflags & REQ_FUA))
 310                 rq->cmd_flags &= ~REQ_FUA;
 311
 312         /*
 313          * An empty flush handed down from a stacking driver may
 314          * translate into nothing if the underlying device does not
 315          * advertise a write-back cache.  In this case, simply
 316          * complete the request.
 317          */
 318         if (!policy) {
 319                 __blk_end_bidi_request(rq, 0, 0, 0);
 320                 return;
 321         }
 322
 323         BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
 324
 325         /*
 326          * If there's data but flush is not necessary, the request can be
 327          * processed directly without going through flush machinery.  Queue
 328          * for normal execution.
 329          */
 330         if ((policy & REQ_FSEQ_DATA) &&
 331             !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
 332                 list_add_tail(&rq->queuelist, &q->queue_head);
 333                 return;
 334         }
 335
 336         /*
 337          * @rq should go through flush machinery.  Mark it part of flush
 338          * sequence and submit for further processing.
 339          */
 340         memset(&rq->flush, 0, sizeof(rq->flush));
 341         INIT_LIST_HEAD(&rq->flush.list);
 342         rq->cmd_flags |= REQ_FLUSH_SEQ;
 343         rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
 344         rq->end_io = flush_data_end_io;
 345
 346         blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
 347 }
 348
 349 /**
 350  * blk_abort_flushes - @q is being aborted, abort flush requests
 351  * @q: request_queue being aborted
 352  *
 353  * To be called from elv_abort_queue().  @q is being aborted.  Prepare all
 354  * FLUSH/FUA requests for abortion.
 355  *
 356  * CONTEXT:
 357  * spin_lock_irq(q->queue_lock)
 358  */
 359 void blk_abort_flushes(struct request_queue *q)
 360 {
 361         struct request *rq, *n;
 362         int i;
 363
 364         /*
 365          * Requests in flight for data are already owned by the dispatch
 366          * queue or the device driver.  Just restore for normal completion.
 367          */
 368         list_for_each_entry_safe(rq, n, &q->flush_data_in_flight, flush.list) {
 369                 list_del_init(&rq->flush.list);
 370                 blk_flush_restore_request(rq);
 371         }
 372
 373         /*
 374          * We need to give away requests on flush queues.  Restore for
 375          * normal completion and put them on the dispatch queue.
 376          */
 377         for (i = 0; i < ARRAY_SIZE(q->flush_queue); i++) {
 378                 list_for_each_entry_safe(rq, n, &q->flush_queue[i],
 379                                          flush.list) {
 380                         list_del_init(&rq->flush.list);
 381                         blk_flush_restore_request(rq);
 382                         list_add_tail(&rq->queuelist, &q->queue_head);
 383                 }
 384         }
 385 }
 386
 387 static void bio_end_flush(struct bio *bio, int err)
 388 {
 389         if (err)
 390                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 391         if (bio->bi_private)
 392                 complete(bio->bi_private);
 393         bio_put(bio);
 394 }
 395
 396 /**
 397  * blkdev_issue_flush - queue a flush
 398  * @bdev:       blockdev to issue flush for
 399  * @gfp_mask:   memory allocation flags (for bio_alloc)
 400  * @error_sector:       error sector
 401  *
 402  * Description:
 403  *    Issue a flush for the block device in question. Caller can supply
 404  *    room for storing the error offset in case of a flush error, if they
 405  *    wish to. If WAIT flag is not passed then caller may check only what
 406  *    request was pushed in some internal queue for later handling.
 407  */
 408 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
 409                 sector_t *error_sector)
 410 {
 411         DECLARE_COMPLETION_ONSTACK(wait);
 412         struct request_queue *q;
 413         struct bio *bio;
 414         int ret = 0;
 415
 416         if (bdev->bd_disk == NULL)
 417                 return -ENXIO;
 418
 419         q = bdev_get_queue(bdev);
 420         if (!q)
 421                 return -ENXIO;
 422
 423         /*
 424          * some block devices may not have their queue correctly set up here
 425          * (e.g. loop device without a backing file) and so issuing a flush
 426          * here will panic. Ensure there is a request function before issuing
 427          * the flush.
 428          */
 429         if (!q->make_request_fn)
 430                 return -ENXIO;
 431
 432         bio = bio_alloc(gfp_mask, 0);
 433         bio->bi_end_io = bio_end_flush;
 434         bio->bi_bdev = bdev;
 435         bio->bi_private = &wait;
 436
 437         bio_get(bio);
 438         submit_bio(WRITE_FLUSH, bio);
 439         wait_for_completion(&wait);
 440
 441         /*
 442          * The driver must store the error location in ->bi_sector, if
 443          * it supports it. For non-stacked drivers, this should be
 444          * copied from blk_rq_pos(rq).
 445          */
 446         if (error_sector)
 447                *error_sector = bio->bi_sector;
 448
 449         if (!bio_flagged(bio, BIO_UPTODATE))
 450                 ret = -EIO;
 451
 452         bio_put(bio);
 453         return ret;
 454 }
 455 EXPORT_SYMBOL(blkdev_issue_flush);