dm: properly fix redundant bio-based IO accounting
authorMike Snitzer <snitzer@redhat.com>
Fri, 28 Jan 2022 15:58:41 +0000 (10:58 -0500)
committerJens Axboe <axboe@kernel.dk>
Fri, 28 Jan 2022 19:28:15 +0000 (12:28 -0700)
Record the start_time for a bio but defer the starting block core's IO
accounting until after IO is submitted using bio_start_io_acct_time().

This approach avoids the need to mess around with any of the
individual IO stats in response to a bio_split() that follows bio
submission.

Reported-by: Bud Brown <bubrown@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: stable@vger.kernel.org
Depends-on: e45c47d1f94e ("block: add bio_start_io_acct_time() to control start_time")
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Link: https://lore.kernel.org/r/20220128155841.39644-4-snitzer@redhat.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
drivers/md/dm.c

index 9849114b3c08fa73e56070b92c46f01f93e9603c..dcbd6d201619d0fd62931af3de1e073648852d1f 100644 (file)
@@ -489,7 +489,7 @@ static void start_io_acct(struct dm_io *io)
        struct mapped_device *md = io->md;
        struct bio *bio = io->orig_bio;
 
-       io->start_time = bio_start_io_acct(bio);
+       bio_start_io_acct_time(bio, io->start_time);
        if (unlikely(dm_stats_used(&md->stats)))
                dm_stats_account_io(&md->stats, bio_data_dir(bio),
                                    bio->bi_iter.bi_sector, bio_sectors(bio),
@@ -535,7 +535,7 @@ static struct dm_io *alloc_io(struct mapped_device *md, struct bio *bio)
        io->md = md;
        spin_lock_init(&io->endio_lock);
 
-       start_io_acct(io);
+       io->start_time = jiffies;
 
        return io;
 }
@@ -1482,6 +1482,7 @@ static void __split_and_process_bio(struct mapped_device *md,
                        submit_bio_noacct(bio);
                }
        }
+       start_io_acct(ci.io);
 
        /* drop the extra reference count */
        dm_io_dec_pending(ci.io, errno_to_blk_status(error));