block/blk-flush.c

   1 /*
   2  * Functions to sequence FLUSH and FUA writes.
   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 /* FLUSH/FUA sequences */
  13 enum {
  14         QUEUE_FSEQ_STARTED      = (1 << 0), /* flushing in progress */
  15         QUEUE_FSEQ_PREFLUSH     = (1 << 1), /* pre-flushing in progress */
  16         QUEUE_FSEQ_DATA         = (1 << 2), /* data write in progress */
  17         QUEUE_FSEQ_POSTFLUSH    = (1 << 3), /* post-flushing in progress */
  18         QUEUE_FSEQ_DONE         = (1 << 4),
  19 };
  20
  21 static struct request *queue_next_fseq(struct request_queue *q);
  22
  23 unsigned blk_flush_cur_seq(struct request_queue *q)
  24 {
  25         if (!q->flush_seq)
  26                 return 0;
  27         return 1 << ffz(q->flush_seq);
  28 }
  29
  30 static struct request *blk_flush_complete_seq(struct request_queue *q,
  31                                               unsigned seq, int error)
  32 {
  33         struct request *next_rq = NULL;
  34
  35         if (error && !q->flush_err)
  36                 q->flush_err = error;
  37
  38         BUG_ON(q->flush_seq & seq);
  39         q->flush_seq |= seq;
  40
  41         if (blk_flush_cur_seq(q) != QUEUE_FSEQ_DONE) {
  42                 /* not complete yet, queue the next flush sequence */
  43                 next_rq = queue_next_fseq(q);
  44         } else {
  45                 /* complete this flush request */
  46                 __blk_end_request_all(q->orig_flush_rq, q->flush_err);
  47                 q->orig_flush_rq = NULL;
  48                 q->flush_seq = 0;
  49
  50                 /* dispatch the next flush if there's one */
  51                 if (!list_empty(&q->pending_flushes)) {
  52                         next_rq = list_entry_rq(q->pending_flushes.next);
  53                         list_move(&next_rq->queuelist, &q->queue_head);
  54                 }
  55         }
  56         return next_rq;
  57 }
  58
  59 static void blk_flush_complete_seq_end_io(struct request_queue *q,
  60                                           unsigned seq, int error)
  61 {
  62         bool was_empty = elv_queue_empty(q);
  63         struct request *next_rq;
  64
  65         next_rq = blk_flush_complete_seq(q, seq, error);
  66
  67         /*
  68          * Moving a request silently to empty queue_head may stall the
  69          * queue.  Kick the queue in those cases.
  70          */
  71         if (was_empty && next_rq)
  72                 __blk_run_queue(q);
  73 }
  74
  75 static void pre_flush_end_io(struct request *rq, int error)
  76 {
  77         elv_completed_request(rq->q, rq);
  78         blk_flush_complete_seq_end_io(rq->q, QUEUE_FSEQ_PREFLUSH, error);
  79 }
  80
  81 static void flush_data_end_io(struct request *rq, int error)
  82 {
  83         elv_completed_request(rq->q, rq);
  84         blk_flush_complete_seq_end_io(rq->q, QUEUE_FSEQ_DATA, error);
  85 }
  86
  87 static void post_flush_end_io(struct request *rq, int error)
  88 {
  89         elv_completed_request(rq->q, rq);
  90         blk_flush_complete_seq_end_io(rq->q, QUEUE_FSEQ_POSTFLUSH, error);
  91 }
  92
  93 static void init_flush_request(struct request *rq, struct gendisk *disk)
  94 {
  95         rq->cmd_type = REQ_TYPE_FS;
  96         rq->cmd_flags = WRITE_FLUSH;
  97         rq->rq_disk = disk;
  98 }
  99
 100 static struct request *queue_next_fseq(struct request_queue *q)
 101 {
 102         struct request *orig_rq = q->orig_flush_rq;
 103         struct request *rq = &q->flush_rq;
 104
 105         blk_rq_init(q, rq);
 106
 107         switch (blk_flush_cur_seq(q)) {
 108         case QUEUE_FSEQ_PREFLUSH:
 109                 init_flush_request(rq, orig_rq->rq_disk);
 110                 rq->end_io = pre_flush_end_io;
 111                 break;
 112         case QUEUE_FSEQ_DATA:
 113                 init_request_from_bio(rq, orig_rq->bio);
 114                 /*
 115                  * orig_rq->rq_disk may be different from
 116                  * bio->bi_bdev->bd_disk if orig_rq got here through
 117                  * remapping drivers.  Make sure rq->rq_disk points
 118                  * to the same one as orig_rq.
 119                  */
 120                 rq->rq_disk = orig_rq->rq_disk;
 121                 rq->cmd_flags &= ~(REQ_FLUSH | REQ_FUA);
 122                 rq->cmd_flags |= orig_rq->cmd_flags & (REQ_FLUSH | REQ_FUA);
 123                 rq->end_io = flush_data_end_io;
 124                 break;
 125         case QUEUE_FSEQ_POSTFLUSH:
 126                 init_flush_request(rq, orig_rq->rq_disk);
 127                 rq->end_io = post_flush_end_io;
 128                 break;
 129         default:
 130                 BUG();
 131         }
 132
 133         elv_insert(q, rq, ELEVATOR_INSERT_FRONT);
 134         return rq;
 135 }
 136
 137 struct request *blk_do_flush(struct request_queue *q, struct request *rq)
 138 {
 139         unsigned int fflags = q->flush_flags; /* may change, cache it */
 140         bool has_flush = fflags & REQ_FLUSH, has_fua = fflags & REQ_FUA;
 141         bool do_preflush = has_flush && (rq->cmd_flags & REQ_FLUSH);
 142         bool do_postflush = has_flush && !has_fua && (rq->cmd_flags & REQ_FUA);
 143         unsigned skip = 0;
 144
 145         /*
 146          * Special case.  If there's data but flush is not necessary,
 147          * the request can be issued directly.
 148          *
 149          * Flush w/o data should be able to be issued directly too but
 150          * currently some drivers assume that rq->bio contains
 151          * non-zero data if it isn't NULL and empty FLUSH requests
 152          * getting here usually have bio's without data.
 153          */
 154         if (blk_rq_sectors(rq) && !do_preflush && !do_postflush) {
 155                 rq->cmd_flags &= ~REQ_FLUSH;
 156                 if (!has_fua)
 157                         rq->cmd_flags &= ~REQ_FUA;
 158                 return rq;
 159         }
 160
 161         /*
 162          * Sequenced flushes can't be processed in parallel.  If
 163          * another one is already in progress, queue for later
 164          * processing.
 165          */
 166         if (q->flush_seq) {
 167                 list_move_tail(&rq->queuelist, &q->pending_flushes);
 168                 return NULL;
 169         }
 170
 171         /*
 172          * Start a new flush sequence
 173          */
 174         q->flush_err = 0;
 175         q->flush_seq |= QUEUE_FSEQ_STARTED;
 176
 177         /* adjust FLUSH/FUA of the original request and stash it away */
 178         rq->cmd_flags &= ~REQ_FLUSH;
 179         if (!has_fua)
 180                 rq->cmd_flags &= ~REQ_FUA;
 181         blk_dequeue_request(rq);
 182         q->orig_flush_rq = rq;
 183
 184         /* skip unneded sequences and return the first one */
 185         if (!do_preflush)
 186                 skip |= QUEUE_FSEQ_PREFLUSH;
 187         if (!blk_rq_sectors(rq))
 188                 skip |= QUEUE_FSEQ_DATA;
 189         if (!do_postflush)
 190                 skip |= QUEUE_FSEQ_POSTFLUSH;
 191         return blk_flush_complete_seq(q, skip, 0);
 192 }
 193
 194 static void bio_end_flush(struct bio *bio, int err)
 195 {
 196         if (err)
 197                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 198         if (bio->bi_private)
 199                 complete(bio->bi_private);
 200         bio_put(bio);
 201 }
 202
 203 /**
 204  * blkdev_issue_flush - queue a flush
 205  * @bdev:       blockdev to issue flush for
 206  * @gfp_mask:   memory allocation flags (for bio_alloc)
 207  * @error_sector:       error sector
 208  *
 209  * Description:
 210  *    Issue a flush for the block device in question. Caller can supply
 211  *    room for storing the error offset in case of a flush error, if they
 212  *    wish to. If WAIT flag is not passed then caller may check only what
 213  *    request was pushed in some internal queue for later handling.
 214  */
 215 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
 216                 sector_t *error_sector)
 217 {
 218         DECLARE_COMPLETION_ONSTACK(wait);
 219         struct request_queue *q;
 220         struct bio *bio;
 221         int ret = 0;
 222
 223         if (bdev->bd_disk == NULL)
 224                 return -ENXIO;
 225
 226         q = bdev_get_queue(bdev);
 227         if (!q)
 228                 return -ENXIO;
 229
 230         /*
 231          * some block devices may not have their queue correctly set up here
 232          * (e.g. loop device without a backing file) and so issuing a flush
 233          * here will panic. Ensure there is a request function before issuing
 234          * the flush.
 235          */
 236         if (!q->make_request_fn)
 237                 return -ENXIO;
 238
 239         bio = bio_alloc(gfp_mask, 0);
 240         bio->bi_end_io = bio_end_flush;
 241         bio->bi_bdev = bdev;
 242         bio->bi_private = &wait;
 243
 244         bio_get(bio);
 245         submit_bio(WRITE_FLUSH, bio);
 246         wait_for_completion(&wait);
 247
 248         /*
 249          * The driver must store the error location in ->bi_sector, if
 250          * it supports it. For non-stacked drivers, this should be
 251          * copied from blk_rq_pos(rq).
 252          */
 253         if (error_sector)
 254                *error_sector = bio->bi_sector;
 255
 256         if (!bio_flagged(bio, BIO_UPTODATE))
 257                 ret = -EIO;
 258
 259         bio_put(bio);
 260         return ret;
 261 }
 262 EXPORT_SYMBOL(blkdev_issue_flush);