block: Introduce new bio_split()
[linux-2.6-block.git] / drivers / md / raid10.c
index 69c1bc8da88f3e93a1c147a5a08091398a46f8d2..6d43d88657aa81982e1f31540878e822ffd44a39 100644 (file)
@@ -1152,14 +1152,12 @@ static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
        kfree(plug);
 }
 
-static void make_request(struct mddev *mddev, struct bio * bio)
+static void __make_request(struct mddev *mddev, struct bio *bio)
 {
        struct r10conf *conf = mddev->private;
        struct r10bio *r10_bio;
        struct bio *read_bio;
        int i;
-       sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
-       int chunk_sects = chunk_mask + 1;
        const int rw = bio_data_dir(bio);
        const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
        const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
@@ -1174,69 +1172,6 @@ static void make_request(struct mddev *mddev, struct bio * bio)
        int max_sectors;
        int sectors;
 
-       if (unlikely(bio->bi_rw & REQ_FLUSH)) {
-               md_flush_request(mddev, bio);
-               return;
-       }
-
-       /* If this request crosses a chunk boundary, we need to
-        * split it.  This will only happen for 1 PAGE (or less) requests.
-        */
-       if (unlikely((bio->bi_iter.bi_sector & chunk_mask) + bio_sectors(bio)
-                    > chunk_sects
-                    && (conf->geo.near_copies < conf->geo.raid_disks
-                        || conf->prev.near_copies < conf->prev.raid_disks))) {
-               struct bio_pair *bp;
-               /* Sanity check -- queue functions should prevent this happening */
-               if (bio_multiple_segments(bio))
-                       goto bad_map;
-               /* This is a one page bio that upper layers
-                * refuse to split for us, so we need to split it.
-                */
-               bp = bio_pair_split(bio, chunk_sects -
-                              (bio->bi_iter.bi_sector & (chunk_sects - 1)));
-
-               /* Each of these 'make_request' calls will call 'wait_barrier'.
-                * If the first succeeds but the second blocks due to the resync
-                * thread raising the barrier, we will deadlock because the
-                * IO to the underlying device will be queued in generic_make_request
-                * and will never complete, so will never reduce nr_pending.
-                * So increment nr_waiting here so no new raise_barriers will
-                * succeed, and so the second wait_barrier cannot block.
-                */
-               spin_lock_irq(&conf->resync_lock);
-               conf->nr_waiting++;
-               spin_unlock_irq(&conf->resync_lock);
-
-               make_request(mddev, &bp->bio1);
-               make_request(mddev, &bp->bio2);
-
-               spin_lock_irq(&conf->resync_lock);
-               conf->nr_waiting--;
-               wake_up(&conf->wait_barrier);
-               spin_unlock_irq(&conf->resync_lock);
-
-               bio_pair_release(bp);
-               return;
-       bad_map:
-               printk("md/raid10:%s: make_request bug: can't convert block across chunks"
-                      " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
-                      (unsigned long long)bio->bi_iter.bi_sector,
-                      bio_sectors(bio) / 2);
-
-               bio_io_error(bio);
-               return;
-       }
-
-       md_write_start(mddev, bio);
-
-       /*
-        * Register the new request and wait if the reconstruction
-        * thread has put up a bar for new requests.
-        * Continue immediately if no resync is active currently.
-        */
-       wait_barrier(conf);
-
        sectors = bio_sectors(bio);
        while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
            bio->bi_iter.bi_sector < conf->reshape_progress &&
@@ -1600,6 +1535,52 @@ retry_write:
                goto retry_write;
        }
        one_write_done(r10_bio);
+}
+
+static void make_request(struct mddev *mddev, struct bio *bio)
+{
+       struct r10conf *conf = mddev->private;
+       sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
+       int chunk_sects = chunk_mask + 1;
+
+       struct bio *split;
+
+       if (unlikely(bio->bi_rw & REQ_FLUSH)) {
+               md_flush_request(mddev, bio);
+               return;
+       }
+
+       md_write_start(mddev, bio);
+
+       /*
+        * Register the new request and wait if the reconstruction
+        * thread has put up a bar for new requests.
+        * Continue immediately if no resync is active currently.
+        */
+       wait_barrier(conf);
+
+       do {
+
+               /*
+                * If this request crosses a chunk boundary, we need to split
+                * it.
+                */
+               if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
+                            bio_sectors(bio) > chunk_sects
+                            && (conf->geo.near_copies < conf->geo.raid_disks
+                                || conf->prev.near_copies <
+                                conf->prev.raid_disks))) {
+                       split = bio_split(bio, chunk_sects -
+                                         (bio->bi_iter.bi_sector &
+                                          (chunk_sects - 1)),
+                                         GFP_NOIO, fs_bio_set);
+                       bio_chain(split, bio);
+               } else {
+                       split = bio;
+               }
+
+               __make_request(mddev, split);
+       } while (split != bio);
 
        /* In case raid10d snuck in to freeze_array */
        wake_up(&conf->wait_barrier);