block/blk-barrier.c

   1 /*
   2  * Functions related to barrier IO handling
   3  */
   4 #include <linux/kernel.h>
   5 #include <linux/module.h>
   6 #include <linux/bio.h>
   7 #include <linux/blkdev.h>
   8 #include <linux/gfp.h>
   9
  10 #include "blk.h"
  11
  12 /**
  13  * blk_queue_ordered - does this queue support ordered writes
  14  * @q:        the request queue
  15  * @ordered:  one of QUEUE_ORDERED_*
  16  *
  17  * Description:
  18  *   For journalled file systems, doing ordered writes on a commit
  19  *   block instead of explicitly doing wait_on_buffer (which is bad
  20  *   for performance) can be a big win. Block drivers supporting this
  21  *   feature should call this function and indicate so.
  22  *
  23  **/
  24 int blk_queue_ordered(struct request_queue *q, unsigned ordered)
  25 {
  26         if (ordered != QUEUE_ORDERED_NONE &&
  27             ordered != QUEUE_ORDERED_DRAIN &&
  28             ordered != QUEUE_ORDERED_DRAIN_FLUSH &&
  29             ordered != QUEUE_ORDERED_DRAIN_FUA) {
  30                 printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered);
  31                 return -EINVAL;
  32         }
  33
  34         q->ordered = ordered;
  35         q->next_ordered = ordered;
  36
  37         return 0;
  38 }
  39 EXPORT_SYMBOL(blk_queue_ordered);
  40
  41 /*
  42  * Cache flushing for ordered writes handling
  43  */
  44 unsigned blk_ordered_cur_seq(struct request_queue *q)
  45 {
  46         if (!q->ordseq)
  47                 return 0;
  48         return 1 << ffz(q->ordseq);
  49 }
  50
  51 unsigned blk_ordered_req_seq(struct request *rq)
  52 {
  53         struct request_queue *q = rq->q;
  54
  55         BUG_ON(q->ordseq == 0);
  56
  57         if (rq == &q->pre_flush_rq)
  58                 return QUEUE_ORDSEQ_PREFLUSH;
  59         if (rq == &q->bar_rq)
  60                 return QUEUE_ORDSEQ_BAR;
  61         if (rq == &q->post_flush_rq)
  62                 return QUEUE_ORDSEQ_POSTFLUSH;
  63
  64         /*
  65          * !fs requests don't need to follow barrier ordering.  Always
  66          * put them at the front.  This fixes the following deadlock.
  67          *
  68          * http://thread.gmane.org/gmane.linux.kernel/537473
  69          */
  70         if (rq->cmd_type != REQ_TYPE_FS)
  71                 return QUEUE_ORDSEQ_DRAIN;
  72
  73         if ((rq->cmd_flags & REQ_ORDERED_COLOR) ==
  74             (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR))
  75                 return QUEUE_ORDSEQ_DRAIN;
  76         else
  77                 return QUEUE_ORDSEQ_DONE;
  78 }
  79
  80 bool blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error)
  81 {
  82         struct request *rq;
  83
  84         if (error && !q->orderr)
  85                 q->orderr = error;
  86
  87         BUG_ON(q->ordseq & seq);
  88         q->ordseq |= seq;
  89
  90         if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE)
  91                 return false;
  92
  93         /*
  94          * Okay, sequence complete.
  95          */
  96         q->ordseq = 0;
  97         rq = q->orig_bar_rq;
  98         __blk_end_request_all(rq, q->orderr);
  99         return true;
 100 }
 101
 102 static void pre_flush_end_io(struct request *rq, int error)
 103 {
 104         elv_completed_request(rq->q, rq);
 105         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error);
 106 }
 107
 108 static void bar_end_io(struct request *rq, int error)
 109 {
 110         elv_completed_request(rq->q, rq);
 111         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error);
 112 }
 113
 114 static void post_flush_end_io(struct request *rq, int error)
 115 {
 116         elv_completed_request(rq->q, rq);
 117         blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error);
 118 }
 119
 120 static void queue_flush(struct request_queue *q, unsigned which)
 121 {
 122         struct request *rq;
 123         rq_end_io_fn *end_io;
 124
 125         if (which == QUEUE_ORDERED_DO_PREFLUSH) {
 126                 rq = &q->pre_flush_rq;
 127                 end_io = pre_flush_end_io;
 128         } else {
 129                 rq = &q->post_flush_rq;
 130                 end_io = post_flush_end_io;
 131         }
 132
 133         blk_rq_init(q, rq);
 134         rq->cmd_type = REQ_TYPE_FS;
 135         rq->cmd_flags = REQ_HARDBARRIER | REQ_FLUSH;
 136         rq->rq_disk = q->orig_bar_rq->rq_disk;
 137         rq->end_io = end_io;
 138
 139         elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
 140 }
 141
 142 static inline bool start_ordered(struct request_queue *q, struct request **rqp)
 143 {
 144         struct request *rq = *rqp;
 145         unsigned skip = 0;
 146
 147         q->orderr = 0;
 148         q->ordered = q->next_ordered;
 149         q->ordseq |= QUEUE_ORDSEQ_STARTED;
 150
 151         /*
 152          * For an empty barrier, there's no actual BAR request, which
 153          * in turn makes POSTFLUSH unnecessary.  Mask them off.
 154          */
 155         if (!blk_rq_sectors(rq))
 156                 q->ordered &= ~(QUEUE_ORDERED_DO_BAR |
 157                                 QUEUE_ORDERED_DO_POSTFLUSH);
 158
 159         /* stash away the original request */
 160         blk_dequeue_request(rq);
 161         q->orig_bar_rq = rq;
 162         rq = NULL;
 163
 164         /*
 165          * Queue ordered sequence.  As we stack them at the head, we
 166          * need to queue in reverse order.  Note that we rely on that
 167          * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs
 168          * request gets inbetween ordered sequence.
 169          */
 170         if (q->ordered & QUEUE_ORDERED_DO_POSTFLUSH) {
 171                 queue_flush(q, QUEUE_ORDERED_DO_POSTFLUSH);
 172                 rq = &q->post_flush_rq;
 173         } else
 174                 skip |= QUEUE_ORDSEQ_POSTFLUSH;
 175
 176         if (q->ordered & QUEUE_ORDERED_DO_BAR) {
 177                 rq = &q->bar_rq;
 178
 179                 /* initialize proxy request and queue it */
 180                 blk_rq_init(q, rq);
 181                 if (bio_data_dir(q->orig_bar_rq->bio) == WRITE)
 182                         rq->cmd_flags |= REQ_WRITE;
 183                 if (q->ordered & QUEUE_ORDERED_DO_FUA)
 184                         rq->cmd_flags |= REQ_FUA;
 185                 init_request_from_bio(rq, q->orig_bar_rq->bio);
 186                 rq->end_io = bar_end_io;
 187
 188                 elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
 189         } else
 190                 skip |= QUEUE_ORDSEQ_BAR;
 191
 192         if (q->ordered & QUEUE_ORDERED_DO_PREFLUSH) {
 193                 queue_flush(q, QUEUE_ORDERED_DO_PREFLUSH);
 194                 rq = &q->pre_flush_rq;
 195         } else
 196                 skip |= QUEUE_ORDSEQ_PREFLUSH;
 197
 198         if (queue_in_flight(q))
 199                 rq = NULL;
 200         else
 201                 skip |= QUEUE_ORDSEQ_DRAIN;
 202
 203         *rqp = rq;
 204
 205         /*
 206          * Complete skipped sequences.  If whole sequence is complete,
 207          * return false to tell elevator that this request is gone.
 208          */
 209         return !blk_ordered_complete_seq(q, skip, 0);
 210 }
 211
 212 bool blk_do_ordered(struct request_queue *q, struct request **rqp)
 213 {
 214         struct request *rq = *rqp;
 215         const int is_barrier = rq->cmd_type == REQ_TYPE_FS &&
 216                                 (rq->cmd_flags & REQ_HARDBARRIER);
 217
 218         if (!q->ordseq) {
 219                 if (!is_barrier)
 220                         return true;
 221
 222                 if (q->next_ordered != QUEUE_ORDERED_NONE)
 223                         return start_ordered(q, rqp);
 224                 else {
 225                         /*
 226                          * Queue ordering not supported.  Terminate
 227                          * with prejudice.
 228                          */
 229                         blk_dequeue_request(rq);
 230                         __blk_end_request_all(rq, -EOPNOTSUPP);
 231                         *rqp = NULL;
 232                         return false;
 233                 }
 234         }
 235
 236         /*
 237          * Ordered sequence in progress
 238          */
 239
 240         /* Special requests are not subject to ordering rules. */
 241         if (rq->cmd_type != REQ_TYPE_FS &&
 242             rq != &q->pre_flush_rq && rq != &q->post_flush_rq)
 243                 return true;
 244
 245         /* Ordered by draining.  Wait for turn. */
 246         WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q));
 247         if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q))
 248                 *rqp = NULL;
 249
 250         return true;
 251 }
 252
 253 static void bio_end_empty_barrier(struct bio *bio, int err)
 254 {
 255         if (err) {
 256                 if (err == -EOPNOTSUPP)
 257                         set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
 258                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 259         }
 260         if (bio->bi_private)
 261                 complete(bio->bi_private);
 262         bio_put(bio);
 263 }
 264
 265 /**
 266  * blkdev_issue_flush - queue a flush
 267  * @bdev:       blockdev to issue flush for
 268  * @gfp_mask:   memory allocation flags (for bio_alloc)
 269  * @error_sector:       error sector
 270  * @flags:      BLKDEV_IFL_* flags to control behaviour
 271  *
 272  * Description:
 273  *    Issue a flush for the block device in question. Caller can supply
 274  *    room for storing the error offset in case of a flush error, if they
 275  *    wish to. If WAIT flag is not passed then caller may check only what
 276  *    request was pushed in some internal queue for later handling.
 277  */
 278 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
 279                 sector_t *error_sector, unsigned long flags)
 280 {
 281         DECLARE_COMPLETION_ONSTACK(wait);
 282         struct request_queue *q;
 283         struct bio *bio;
 284         int ret = 0;
 285
 286         if (bdev->bd_disk == NULL)
 287                 return -ENXIO;
 288
 289         q = bdev_get_queue(bdev);
 290         if (!q)
 291                 return -ENXIO;
 292
 293         /*
 294          * some block devices may not have their queue correctly set up here
 295          * (e.g. loop device without a backing file) and so issuing a flush
 296          * here will panic. Ensure there is a request function before issuing
 297          * the barrier.
 298          */
 299         if (!q->make_request_fn)
 300                 return -ENXIO;
 301
 302         bio = bio_alloc(gfp_mask, 0);
 303         bio->bi_end_io = bio_end_empty_barrier;
 304         bio->bi_bdev = bdev;
 305         if (test_bit(BLKDEV_WAIT, &flags))
 306                 bio->bi_private = &wait;
 307
 308         bio_get(bio);
 309         submit_bio(WRITE_BARRIER, bio);
 310         if (test_bit(BLKDEV_WAIT, &flags)) {
 311                 wait_for_completion(&wait);
 312                 /*
 313                  * The driver must store the error location in ->bi_sector, if
 314                  * it supports it. For non-stacked drivers, this should be
 315                  * copied from blk_rq_pos(rq).
 316                  */
 317                 if (error_sector)
 318                         *error_sector = bio->bi_sector;
 319         }
 320
 321         if (bio_flagged(bio, BIO_EOPNOTSUPP))
 322                 ret = -EOPNOTSUPP;
 323         else if (!bio_flagged(bio, BIO_UPTODATE))
 324                 ret = -EIO;
 325
 326         bio_put(bio);
 327         return ret;
 328 }
 329 EXPORT_SYMBOL(blkdev_issue_flush);