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 (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
 109 static unsigned int blk_flush_cur_seq(struct request *rq)
 110 {
 111         return 1 << ffz(rq->flush.seq);
 112 }
 113
 114 static 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  */
 143 static 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
 193 static 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          * Moving a request silently to empty queue_head may stall the
 216          * queue.  Kick the queue in those cases.  This function is called
 217          * from request completion path and calling directly into
 218          * request_fn may confuse the driver.  Always use kblockd.
 219          */
 220         if (queued)
 221                 __blk_run_queue(q, true);
 222 }
 223
 224 /**
 225  * blk_kick_flush - consider issuing flush request
 226  * @q: request_queue being kicked
 227  *
 228  * Flush related states of @q have changed, consider issuing flush request.
 229  * Please read the comment at the top of this file for more info.
 230  *
 231  * CONTEXT:
 232  * spin_lock_irq(q->queue_lock)
 233  *
 234  * RETURNS:
 235  * %true if flush was issued, %false otherwise.
 236  */
 237 static bool blk_kick_flush(struct request_queue *q)
 238 {
 239         struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
 240         struct request *first_rq =
 241                 list_first_entry(pending, struct request, flush.list);
 242
 243         /* C1 described at the top of this file */
 244         if (q->flush_pending_idx != q->flush_running_idx || list_empty(pending))
 245                 return false;
 246
 247         /* C2 and C3 */
 248         if (!list_empty(&q->flush_data_in_flight) &&
 249             time_before(jiffies,
 250                         q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
 251                 return false;
 252
 253         /*
 254          * Issue flush and toggle pending_idx.  This makes pending_idx
 255          * different from running_idx, which means flush is in flight.
 256          */
 257         blk_rq_init(q, &q->flush_rq);
 258         q->flush_rq.cmd_type = REQ_TYPE_FS;
 259         q->flush_rq.cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
 260         q->flush_rq.rq_disk = first_rq->rq_disk;
 261         q->flush_rq.end_io = flush_end_io;
 262
 263         q->flush_pending_idx ^= 1;
 264         elv_insert(q, &q->flush_rq, ELEVATOR_INSERT_REQUEUE);
 265         return true;
 266 }
 267
 268 static void flush_data_end_io(struct request *rq, int error)
 269 {
 270         struct request_queue *q = rq->q;
 271
 272         /*
 273          * After populating an empty queue, kick it to avoid stall.  Read
 274          * the comment in flush_end_io().
 275          */
 276         if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
 277                 __blk_run_queue(q, true);
 278 }
 279
 280 /**
 281  * blk_insert_flush - insert a new FLUSH/FUA request
 282  * @rq: request to insert
 283  *
 284  * To be called from elv_insert() for %ELEVATOR_INSERT_FLUSH insertions.
 285  * @rq is being submitted.  Analyze what needs to be done and put it on the
 286  * right queue.
 287  *
 288  * CONTEXT:
 289  * spin_lock_irq(q->queue_lock)
 290  */
 291 void blk_insert_flush(struct request *rq)
 292 {
 293         struct request_queue *q = rq->q;
 294         unsigned int fflags = q->flush_flags;   /* may change, cache */
 295         unsigned int policy = blk_flush_policy(fflags, rq);
 296
 297         BUG_ON(rq->end_io);
 298         BUG_ON(!rq->bio || rq->bio != rq->biotail);
 299
 300         /*
 301          * @policy now records what operations need to be done.  Adjust
 302          * REQ_FLUSH and FUA for the driver.
 303          */
 304         rq->cmd_flags &= ~REQ_FLUSH;
 305         if (!(fflags & REQ_FUA))
 306                 rq->cmd_flags &= ~REQ_FUA;
 307
 308         /*
 309          * If there's data but flush is not necessary, the request can be
 310          * processed directly without going through flush machinery.  Queue
 311          * for normal execution.
 312          */
 313         if ((policy & REQ_FSEQ_DATA) &&
 314             !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
 315                 list_add(&rq->queuelist, &q->queue_head);
 316                 return;
 317         }
 318
 319         /*
 320          * @rq should go through flush machinery.  Mark it part of flush
 321          * sequence and submit for further processing.
 322          */
 323         memset(&rq->flush, 0, sizeof(rq->flush));
 324         INIT_LIST_HEAD(&rq->flush.list);
 325         rq->cmd_flags |= REQ_FLUSH_SEQ;
 326         rq->end_io = flush_data_end_io;
 327
 328         blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
 329 }
 330
 331 /**
 332  * blk_abort_flushes - @q is being aborted, abort flush requests
 333  * @q: request_queue being aborted
 334  *
 335  * To be called from elv_abort_queue().  @q is being aborted.  Prepare all
 336  * FLUSH/FUA requests for abortion.
 337  *
 338  * CONTEXT:
 339  * spin_lock_irq(q->queue_lock)
 340  */
 341 void blk_abort_flushes(struct request_queue *q)
 342 {
 343         struct request *rq, *n;
 344         int i;
 345
 346         /*
 347          * Requests in flight for data are already owned by the dispatch
 348          * queue or the device driver.  Just restore for normal completion.
 349          */
 350         list_for_each_entry_safe(rq, n, &q->flush_data_in_flight, flush.list) {
 351                 list_del_init(&rq->flush.list);
 352                 blk_flush_restore_request(rq);
 353         }
 354
 355         /*
 356          * We need to give away requests on flush queues.  Restore for
 357          * normal completion and put them on the dispatch queue.
 358          */
 359         for (i = 0; i < ARRAY_SIZE(q->flush_queue); i++) {
 360                 list_for_each_entry_safe(rq, n, &q->flush_queue[i],
 361                                          flush.list) {
 362                         list_del_init(&rq->flush.list);
 363                         blk_flush_restore_request(rq);
 364                         list_add_tail(&rq->queuelist, &q->queue_head);
 365                 }
 366         }
 367 }
 368
 369 static void bio_end_flush(struct bio *bio, int err)
 370 {
 371         if (err)
 372                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 373         if (bio->bi_private)
 374                 complete(bio->bi_private);
 375         bio_put(bio);
 376 }
 377
 378 /**
 379  * blkdev_issue_flush - queue a flush
 380  * @bdev:       blockdev to issue flush for
 381  * @gfp_mask:   memory allocation flags (for bio_alloc)
 382  * @error_sector:       error sector
 383  *
 384  * Description:
 385  *    Issue a flush for the block device in question. Caller can supply
 386  *    room for storing the error offset in case of a flush error, if they
 387  *    wish to. If WAIT flag is not passed then caller may check only what
 388  *    request was pushed in some internal queue for later handling.
 389  */
 390 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
 391                 sector_t *error_sector)
 392 {
 393         DECLARE_COMPLETION_ONSTACK(wait);
 394         struct request_queue *q;
 395         struct bio *bio;
 396         int ret = 0;
 397
 398         if (bdev->bd_disk == NULL)
 399                 return -ENXIO;
 400
 401         q = bdev_get_queue(bdev);
 402         if (!q)
 403                 return -ENXIO;
 404
 405         /*
 406          * some block devices may not have their queue correctly set up here
 407          * (e.g. loop device without a backing file) and so issuing a flush
 408          * here will panic. Ensure there is a request function before issuing
 409          * the flush.
 410          */
 411         if (!q->make_request_fn)
 412                 return -ENXIO;
 413
 414         bio = bio_alloc(gfp_mask, 0);
 415         bio->bi_end_io = bio_end_flush;
 416         bio->bi_bdev = bdev;
 417         bio->bi_private = &wait;
 418
 419         bio_get(bio);
 420         submit_bio(WRITE_FLUSH, bio);
 421         wait_for_completion(&wait);
 422
 423         /*
 424          * The driver must store the error location in ->bi_sector, if
 425          * it supports it. For non-stacked drivers, this should be
 426          * copied from blk_rq_pos(rq).
 427          */
 428         if (error_sector)
 429                *error_sector = bio->bi_sector;
 430
 431         if (!bio_flagged(bio, BIO_UPTODATE))
 432                 ret = -EIO;
 433
 434         bio_put(bio);
 435         return ret;
 436 }
 437 EXPORT_SYMBOL(blkdev_issue_flush);