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.  This function is called
  70          * from request completion path and calling directly into
  71          * request_fn may confuse the driver.  Always use kblockd.
  72          */
  73         if (was_empty && next_rq)
  74                 __blk_run_queue(q, true);
  75 }
  76
  77 static void pre_flush_end_io(struct request *rq, int error)
  78 {
  79         elv_completed_request(rq->q, rq);
  80         blk_flush_complete_seq_end_io(rq->q, QUEUE_FSEQ_PREFLUSH, error);
  81 }
  82
  83 static void flush_data_end_io(struct request *rq, int error)
  84 {
  85         elv_completed_request(rq->q, rq);
  86         blk_flush_complete_seq_end_io(rq->q, QUEUE_FSEQ_DATA, error);
  87 }
  88
  89 static void post_flush_end_io(struct request *rq, int error)
  90 {
  91         elv_completed_request(rq->q, rq);
  92         blk_flush_complete_seq_end_io(rq->q, QUEUE_FSEQ_POSTFLUSH, error);
  93 }
  94
  95 static void init_flush_request(struct request *rq, struct gendisk *disk)
  96 {
  97         rq->cmd_type = REQ_TYPE_FS;
  98         rq->cmd_flags = WRITE_FLUSH;
  99         rq->rq_disk = disk;
 100 }
 101
 102 static struct request *queue_next_fseq(struct request_queue *q)
 103 {
 104         struct request *orig_rq = q->orig_flush_rq;
 105         struct request *rq = &q->flush_rq;
 106
 107         blk_rq_init(q, rq);
 108
 109         switch (blk_flush_cur_seq(q)) {
 110         case QUEUE_FSEQ_PREFLUSH:
 111                 init_flush_request(rq, orig_rq->rq_disk);
 112                 rq->end_io = pre_flush_end_io;
 113                 break;
 114         case QUEUE_FSEQ_DATA:
 115                 init_request_from_bio(rq, orig_rq->bio);
 116                 /*
 117                  * orig_rq->rq_disk may be different from
 118                  * bio->bi_bdev->bd_disk if orig_rq got here through
 119                  * remapping drivers.  Make sure rq->rq_disk points
 120                  * to the same one as orig_rq.
 121                  */
 122                 rq->rq_disk = orig_rq->rq_disk;
 123                 rq->cmd_flags &= ~(REQ_FLUSH | REQ_FUA);
 124                 rq->cmd_flags |= orig_rq->cmd_flags & (REQ_FLUSH | REQ_FUA);
 125                 rq->end_io = flush_data_end_io;
 126                 break;
 127         case QUEUE_FSEQ_POSTFLUSH:
 128                 init_flush_request(rq, orig_rq->rq_disk);
 129                 rq->end_io = post_flush_end_io;
 130                 break;
 131         default:
 132                 BUG();
 133         }
 134
 135         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 136         return rq;
 137 }
 138
 139 struct request *blk_do_flush(struct request_queue *q, struct request *rq)
 140 {
 141         unsigned int fflags = q->flush_flags; /* may change, cache it */
 142         bool has_flush = fflags & REQ_FLUSH, has_fua = fflags & REQ_FUA;
 143         bool do_preflush = has_flush && (rq->cmd_flags & REQ_FLUSH);
 144         bool do_postflush = has_flush && !has_fua && (rq->cmd_flags & REQ_FUA);
 145         unsigned skip = 0;
 146
 147         /*
 148          * Special case.  If there's data but flush is not necessary,
 149          * the request can be issued directly.
 150          *
 151          * Flush w/o data should be able to be issued directly too but
 152          * currently some drivers assume that rq->bio contains
 153          * non-zero data if it isn't NULL and empty FLUSH requests
 154          * getting here usually have bio's without data.
 155          */
 156         if (blk_rq_sectors(rq) && !do_preflush && !do_postflush) {
 157                 rq->cmd_flags &= ~REQ_FLUSH;
 158                 if (!has_fua)
 159                         rq->cmd_flags &= ~REQ_FUA;
 160                 return rq;
 161         }
 162
 163         /*
 164          * Sequenced flushes can't be processed in parallel.  If
 165          * another one is already in progress, queue for later
 166          * processing.
 167          */
 168         if (q->flush_seq) {
 169                 list_move_tail(&rq->queuelist, &q->pending_flushes);
 170                 return NULL;
 171         }
 172
 173         /*
 174          * Start a new flush sequence
 175          */
 176         q->flush_err = 0;
 177         q->flush_seq |= QUEUE_FSEQ_STARTED;
 178
 179         /* adjust FLUSH/FUA of the original request and stash it away */
 180         rq->cmd_flags &= ~REQ_FLUSH;
 181         if (!has_fua)
 182                 rq->cmd_flags &= ~REQ_FUA;
 183         blk_dequeue_request(rq);
 184         q->orig_flush_rq = rq;
 185
 186         /* skip unneded sequences and return the first one */
 187         if (!do_preflush)
 188                 skip |= QUEUE_FSEQ_PREFLUSH;
 189         if (!blk_rq_sectors(rq))
 190                 skip |= QUEUE_FSEQ_DATA;
 191         if (!do_postflush)
 192                 skip |= QUEUE_FSEQ_POSTFLUSH;
 193         return blk_flush_complete_seq(q, skip, 0);
 194 }
 195
 196 static void bio_end_flush(struct bio *bio, int err)
 197 {
 198         if (err)
 199                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 200         if (bio->bi_private)
 201                 complete(bio->bi_private);
 202         bio_put(bio);
 203 }
 204
 205 /**
 206  * blkdev_issue_flush - queue a flush
 207  * @bdev:       blockdev to issue flush for
 208  * @gfp_mask:   memory allocation flags (for bio_alloc)
 209  * @error_sector:       error sector
 210  *
 211  * Description:
 212  *    Issue a flush for the block device in question. Caller can supply
 213  *    room for storing the error offset in case of a flush error, if they
 214  *    wish to. If WAIT flag is not passed then caller may check only what
 215  *    request was pushed in some internal queue for later handling.
 216  */
 217 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
 218                 sector_t *error_sector)
 219 {
 220         DECLARE_COMPLETION_ONSTACK(wait);
 221         struct request_queue *q;
 222         struct bio *bio;
 223         int ret = 0;
 224
 225         if (bdev->bd_disk == NULL)
 226                 return -ENXIO;
 227
 228         q = bdev_get_queue(bdev);
 229         if (!q)
 230                 return -ENXIO;
 231
 232         /*
 233          * some block devices may not have their queue correctly set up here
 234          * (e.g. loop device without a backing file) and so issuing a flush
 235          * here will panic. Ensure there is a request function before issuing
 236          * the flush.
 237          */
 238         if (!q->make_request_fn)
 239                 return -ENXIO;
 240
 241         bio = bio_alloc(gfp_mask, 0);
 242         bio->bi_end_io = bio_end_flush;
 243         bio->bi_bdev = bdev;
 244         bio->bi_private = &wait;
 245
 246         bio_get(bio);
 247         submit_bio(WRITE_FLUSH, bio);
 248         wait_for_completion(&wait);
 249
 250         /*
 251          * The driver must store the error location in ->bi_sector, if
 252          * it supports it. For non-stacked drivers, this should be
 253          * copied from blk_rq_pos(rq).
 254          */
 255         if (error_sector)
 256                *error_sector = bio->bi_sector;
 257
 258         if (!bio_flagged(bio, BIO_UPTODATE))
 259                 ret = -EIO;
 260
 261         bio_put(bio);
 262         return ret;
 263 }
 264 EXPORT_SYMBOL(blkdev_issue_flush);