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_PREFLUSH makes sense, which
  14  * indicates a simple flush request.  If there is data, REQ_PREFLUSH 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_PREFLUSH 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_PREFLUSH
  27  * is 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  * fq->flush_queue[fq->flush_pending_idx].  Once certain criteria are met, a
  32  * REQ_OP_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 RQF_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 #include "blk-mq-tag.h"
  77 #include "blk-mq-sched.h"
  78
  79 /* FLUSH/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 bool blk_kick_flush(struct request_queue *q,
  97                            struct blk_flush_queue *fq);
  98
  99 static unsigned int blk_flush_policy(unsigned long fflags, struct request *rq)
 100 {
 101         unsigned int policy = 0;
 102
 103         if (blk_rq_sectors(rq))
 104                 policy |= REQ_FSEQ_DATA;
 105
 106         if (fflags & (1UL << QUEUE_FLAG_WC)) {
 107                 if (rq->cmd_flags & REQ_PREFLUSH)
 108                         policy |= REQ_FSEQ_PREFLUSH;
 109                 if (!(fflags & (1UL << QUEUE_FLAG_FUA)) &&
 110                     (rq->cmd_flags & REQ_FUA))
 111                         policy |= REQ_FSEQ_POSTFLUSH;
 112         }
 113         return policy;
 114 }
 115
 116 static unsigned int blk_flush_cur_seq(struct request *rq)
 117 {
 118         return 1 << ffz(rq->flush.seq);
 119 }
 120
 121 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         rq->bio = rq->biotail;
 129
 130         /* make @rq a normal request */
 131         rq->rq_flags &= ~RQF_FLUSH_SEQ;
 132         rq->end_io = rq->flush.saved_end_io;
 133 }
 134
 135 static bool blk_flush_queue_rq(struct request *rq, bool add_front)
 136 {
 137         if (rq->q->mq_ops) {
 138                 blk_mq_add_to_requeue_list(rq, add_front, true);
 139                 return false;
 140         } else {
 141                 if (add_front)
 142                         list_add(&rq->queuelist, &rq->q->queue_head);
 143                 else
 144                         list_add_tail(&rq->queuelist, &rq->q->queue_head);
 145                 return true;
 146         }
 147 }
 148
 149 /**
 150  * blk_flush_complete_seq - complete flush sequence
 151  * @rq: FLUSH/FUA request being sequenced
 152  * @fq: flush queue
 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 fq->mq_flush_lock)
 161  *
 162  * RETURNS:
 163  * %true if requests were added to the dispatch queue, %false otherwise.
 164  */
 165 static bool blk_flush_complete_seq(struct request *rq,
 166                                    struct blk_flush_queue *fq,
 167                                    unsigned int seq, int error)
 168 {
 169         struct request_queue *q = rq->q;
 170         struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
 171         bool queued = false, kicked;
 172
 173         BUG_ON(rq->flush.seq & seq);
 174         rq->flush.seq |= seq;
 175
 176         if (likely(!error))
 177                 seq = blk_flush_cur_seq(rq);
 178         else
 179                 seq = REQ_FSEQ_DONE;
 180
 181         switch (seq) {
 182         case REQ_FSEQ_PREFLUSH:
 183         case REQ_FSEQ_POSTFLUSH:
 184                 /* queue for flush */
 185                 if (list_empty(pending))
 186                         fq->flush_pending_since = jiffies;
 187                 list_move_tail(&rq->flush.list, pending);
 188                 break;
 189
 190         case REQ_FSEQ_DATA:
 191                 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight);
 192                 queued = blk_flush_queue_rq(rq, true);
 193                 break;
 194
 195         case REQ_FSEQ_DONE:
 196                 /*
 197                  * @rq was previously adjusted by blk_flush_issue() for
 198                  * flush sequencing and may already have gone through the
 199                  * flush data request completion path.  Restore @rq for
 200                  * normal completion and end it.
 201                  */
 202                 BUG_ON(!list_empty(&rq->queuelist));
 203                 list_del_init(&rq->flush.list);
 204                 blk_flush_restore_request(rq);
 205                 if (q->mq_ops)
 206                         blk_mq_end_request(rq, error);
 207                 else
 208                         __blk_end_request_all(rq, error);
 209                 break;
 210
 211         default:
 212                 BUG();
 213         }
 214
 215         kicked = blk_kick_flush(q, fq);
 216         return kicked | queued;
 217 }
 218
 219 static void flush_end_io(struct request *flush_rq, int error)
 220 {
 221         struct request_queue *q = flush_rq->q;
 222         struct list_head *running;
 223         bool queued = false;
 224         struct request *rq, *n;
 225         unsigned long flags = 0;
 226         struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
 227
 228         if (q->mq_ops) {
 229                 struct blk_mq_hw_ctx *hctx;
 230
 231                 /* release the tag's ownership to the req cloned from */
 232                 spin_lock_irqsave(&fq->mq_flush_lock, flags);
 233                 hctx = blk_mq_map_queue(q, flush_rq->mq_ctx->cpu);
 234                 blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq);
 235                 flush_rq->tag = -1;
 236         }
 237
 238         running = &fq->flush_queue[fq->flush_running_idx];
 239         BUG_ON(fq->flush_pending_idx == fq->flush_running_idx);
 240
 241         /* account completion of the flush request */
 242         fq->flush_running_idx ^= 1;
 243
 244         if (!q->mq_ops)
 245                 elv_completed_request(q, flush_rq);
 246
 247         /* and push the waiting requests to the next stage */
 248         list_for_each_entry_safe(rq, n, running, flush.list) {
 249                 unsigned int seq = blk_flush_cur_seq(rq);
 250
 251                 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
 252                 queued |= blk_flush_complete_seq(rq, fq, seq, error);
 253         }
 254
 255         /*
 256          * Kick the queue to avoid stall for two cases:
 257          * 1. Moving a request silently to empty queue_head may stall the
 258          * queue.
 259          * 2. When flush request is running in non-queueable queue, the
 260          * queue is hold. Restart the queue after flush request is finished
 261          * to avoid stall.
 262          * This function is called from request completion path and calling
 263          * directly into request_fn may confuse the driver.  Always use
 264          * kblockd.
 265          */
 266         if (queued || fq->flush_queue_delayed) {
 267                 WARN_ON(q->mq_ops);
 268                 blk_run_queue_async(q);
 269         }
 270         fq->flush_queue_delayed = 0;
 271         if (q->mq_ops)
 272                 spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
 273 }
 274
 275 /**
 276  * blk_kick_flush - consider issuing flush request
 277  * @q: request_queue being kicked
 278  * @fq: flush queue
 279  *
 280  * Flush related states of @q have changed, consider issuing flush request.
 281  * Please read the comment at the top of this file for more info.
 282  *
 283  * CONTEXT:
 284  * spin_lock_irq(q->queue_lock or fq->mq_flush_lock)
 285  *
 286  * RETURNS:
 287  * %true if flush was issued, %false otherwise.
 288  */
 289 static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq)
 290 {
 291         struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx];
 292         struct request *first_rq =
 293                 list_first_entry(pending, struct request, flush.list);
 294         struct request *flush_rq = fq->flush_rq;
 295
 296         /* C1 described at the top of this file */
 297         if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
 298                 return false;
 299
 300         /* C2 and C3
 301          *
 302          * For blk-mq + scheduling, we can risk having all driver tags
 303          * assigned to empty flushes, and we deadlock if we are expecting
 304          * other requests to make progress. Don't defer for that case.
 305          */
 306         if (!list_empty(&fq->flush_data_in_flight) &&
 307             !(q->mq_ops && q->elevator) &&
 308             time_before(jiffies,
 309                         fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
 310                 return false;
 311
 312         /*
 313          * Issue flush and toggle pending_idx.  This makes pending_idx
 314          * different from running_idx, which means flush is in flight.
 315          */
 316         fq->flush_pending_idx ^= 1;
 317
 318         blk_rq_init(q, flush_rq);
 319
 320         /*
 321          * Borrow tag from the first request since they can't
 322          * be in flight at the same time. And acquire the tag's
 323          * ownership for flush req.
 324          */
 325         if (q->mq_ops) {
 326                 struct blk_mq_hw_ctx *hctx;
 327
 328                 flush_rq->mq_ctx = first_rq->mq_ctx;
 329                 flush_rq->tag = first_rq->tag;
 330                 fq->orig_rq = first_rq;
 331
 332                 hctx = blk_mq_map_queue(q, first_rq->mq_ctx->cpu);
 333                 blk_mq_tag_set_rq(hctx, first_rq->tag, flush_rq);
 334         }
 335
 336         flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
 337         flush_rq->rq_flags |= RQF_FLUSH_SEQ;
 338         flush_rq->rq_disk = first_rq->rq_disk;
 339         flush_rq->end_io = flush_end_io;
 340
 341         return blk_flush_queue_rq(flush_rq, false);
 342 }
 343
 344 static void flush_data_end_io(struct request *rq, int error)
 345 {
 346         struct request_queue *q = rq->q;
 347         struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL);
 348
 349         /*
 350          * Updating q->in_flight[] here for making this tag usable
 351          * early. Because in blk_queue_start_tag(),
 352          * q->in_flight[BLK_RW_ASYNC] is used to limit async I/O and
 353          * reserve tags for sync I/O.
 354          *
 355          * More importantly this way can avoid the following I/O
 356          * deadlock:
 357          *
 358          * - suppose there are 40 fua requests comming to flush queue
 359          *   and queue depth is 31
 360          * - 30 rqs are scheduled then blk_queue_start_tag() can't alloc
 361          *   tag for async I/O any more
 362          * - all the 30 rqs are completed before FLUSH_PENDING_TIMEOUT
 363          *   and flush_data_end_io() is called
 364          * - the other rqs still can't go ahead if not updating
 365          *   q->in_flight[BLK_RW_ASYNC] here, meantime these rqs
 366          *   are held in flush data queue and make no progress of
 367          *   handling post flush rq
 368          * - only after the post flush rq is handled, all these rqs
 369          *   can be completed
 370          */
 371
 372         elv_completed_request(q, rq);
 373
 374         /* for avoiding double accounting */
 375         rq->rq_flags &= ~RQF_STARTED;
 376
 377         /*
 378          * After populating an empty queue, kick it to avoid stall.  Read
 379          * the comment in flush_end_io().
 380          */
 381         if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error))
 382                 blk_run_queue_async(q);
 383 }
 384
 385 static void mq_flush_data_end_io(struct request *rq, int error)
 386 {
 387         struct request_queue *q = rq->q;
 388         struct blk_mq_hw_ctx *hctx;
 389         struct blk_mq_ctx *ctx = rq->mq_ctx;
 390         unsigned long flags;
 391         struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx);
 392
 393         hctx = blk_mq_map_queue(q, ctx->cpu);
 394
 395         /*
 396          * After populating an empty queue, kick it to avoid stall.  Read
 397          * the comment in flush_end_io().
 398          */
 399         spin_lock_irqsave(&fq->mq_flush_lock, flags);
 400         blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error);
 401         spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
 402
 403         blk_mq_run_hw_queue(hctx, true);
 404 }
 405
 406 /**
 407  * blk_insert_flush - insert a new FLUSH/FUA request
 408  * @rq: request to insert
 409  *
 410  * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
 411  * or __blk_mq_run_hw_queue() to dispatch request.
 412  * @rq is being submitted.  Analyze what needs to be done and put it on the
 413  * right queue.
 414  *
 415  * CONTEXT:
 416  * spin_lock_irq(q->queue_lock) in !mq case
 417  */
 418 void blk_insert_flush(struct request *rq)
 419 {
 420         struct request_queue *q = rq->q;
 421         unsigned long fflags = q->queue_flags;  /* may change, cache */
 422         unsigned int policy = blk_flush_policy(fflags, rq);
 423         struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx);
 424
 425         /*
 426          * @policy now records what operations need to be done.  Adjust
 427          * REQ_PREFLUSH and FUA for the driver.
 428          */
 429         rq->cmd_flags &= ~REQ_PREFLUSH;
 430         if (!(fflags & (1UL << QUEUE_FLAG_FUA)))
 431                 rq->cmd_flags &= ~REQ_FUA;
 432
 433         /*
 434          * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any
 435          * of those flags, we have to set REQ_SYNC to avoid skewing
 436          * the request accounting.
 437          */
 438         rq->cmd_flags |= REQ_SYNC;
 439
 440         /*
 441          * An empty flush handed down from a stacking driver may
 442          * translate into nothing if the underlying device does not
 443          * advertise a write-back cache.  In this case, simply
 444          * complete the request.
 445          */
 446         if (!policy) {
 447                 if (q->mq_ops)
 448                         blk_mq_end_request(rq, 0);
 449                 else
 450                         __blk_end_bidi_request(rq, 0, 0, 0);
 451                 return;
 452         }
 453
 454         BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
 455
 456         /*
 457          * If there's data but flush is not necessary, the request can be
 458          * processed directly without going through flush machinery.  Queue
 459          * for normal execution.
 460          */
 461         if ((policy & REQ_FSEQ_DATA) &&
 462             !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
 463                 if (q->mq_ops)
 464                         blk_mq_sched_insert_request(rq, false, true, false, false);
 465                 else
 466                         list_add_tail(&rq->queuelist, &q->queue_head);
 467                 return;
 468         }
 469
 470         /*
 471          * @rq should go through flush machinery.  Mark it part of flush
 472          * sequence and submit for further processing.
 473          */
 474         memset(&rq->flush, 0, sizeof(rq->flush));
 475         INIT_LIST_HEAD(&rq->flush.list);
 476         rq->rq_flags |= RQF_FLUSH_SEQ;
 477         rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
 478         if (q->mq_ops) {
 479                 rq->end_io = mq_flush_data_end_io;
 480
 481                 spin_lock_irq(&fq->mq_flush_lock);
 482                 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
 483                 spin_unlock_irq(&fq->mq_flush_lock);
 484                 return;
 485         }
 486         rq->end_io = flush_data_end_io;
 487
 488         blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0);
 489 }
 490
 491 /**
 492  * blkdev_issue_flush - queue a flush
 493  * @bdev:       blockdev to issue flush for
 494  * @gfp_mask:   memory allocation flags (for bio_alloc)
 495  * @error_sector:       error sector
 496  *
 497  * Description:
 498  *    Issue a flush for the block device in question. Caller can supply
 499  *    room for storing the error offset in case of a flush error, if they
 500  *    wish to. If WAIT flag is not passed then caller may check only what
 501  *    request was pushed in some internal queue for later handling.
 502  */
 503 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
 504                 sector_t *error_sector)
 505 {
 506         struct request_queue *q;
 507         struct bio *bio;
 508         int ret = 0;
 509
 510         if (bdev->bd_disk == NULL)
 511                 return -ENXIO;
 512
 513         q = bdev_get_queue(bdev);
 514         if (!q)
 515                 return -ENXIO;
 516
 517         /*
 518          * some block devices may not have their queue correctly set up here
 519          * (e.g. loop device without a backing file) and so issuing a flush
 520          * here will panic. Ensure there is a request function before issuing
 521          * the flush.
 522          */
 523         if (!q->make_request_fn)
 524                 return -ENXIO;
 525
 526         bio = bio_alloc(gfp_mask, 0);
 527         bio->bi_bdev = bdev;
 528         bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
 529
 530         ret = submit_bio_wait(bio);
 531
 532         /*
 533          * The driver must store the error location in ->bi_sector, if
 534          * it supports it. For non-stacked drivers, this should be
 535          * copied from blk_rq_pos(rq).
 536          */
 537         if (error_sector)
 538                 *error_sector = bio->bi_iter.bi_sector;
 539
 540         bio_put(bio);
 541         return ret;
 542 }
 543 EXPORT_SYMBOL(blkdev_issue_flush);
 544
 545 struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
 546                 int node, int cmd_size)
 547 {
 548         struct blk_flush_queue *fq;
 549         int rq_sz = sizeof(struct request);
 550
 551         fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node);
 552         if (!fq)
 553                 goto fail;
 554
 555         if (q->mq_ops)
 556                 spin_lock_init(&fq->mq_flush_lock);
 557
 558         rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
 559         fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node);
 560         if (!fq->flush_rq)
 561                 goto fail_rq;
 562
 563         INIT_LIST_HEAD(&fq->flush_queue[0]);
 564         INIT_LIST_HEAD(&fq->flush_queue[1]);
 565         INIT_LIST_HEAD(&fq->flush_data_in_flight);
 566
 567         return fq;
 568
 569  fail_rq:
 570         kfree(fq);
 571  fail:
 572         return NULL;
 573 }
 574
 575 void blk_free_flush_queue(struct blk_flush_queue *fq)
 576 {
 577         /* bio based request queue hasn't flush queue */
 578         if (!fq)
 579                 return;
 580
 581         kfree(fq->flush_rq);
 582         kfree(fq);
 583 }