md/raid1: submit IO from originating thread instead of md thread.
NeilBrown [Wed, 1 Aug 2012 22:33:20 +0000 (08:33 +1000)]
queuing writes to the md thread means that all requests go through the
one processor which may not be able to keep up with very high request
rates.

So use the plugging infrastructure to submit all requests on unplug.
If a 'schedule' is needed, we fall back on the old approach of handing
the requests to the thread for it to handle.

Signed-off-by: NeilBrown <neilb@suse.de>

drivers/md/bitmap.c
drivers/md/raid1.c

index 15dbe03..94e7f6b 100644 (file)
@@ -1305,7 +1305,7 @@ int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sect
                        prepare_to_wait(&bitmap->overflow_wait, &__wait,
                                        TASK_UNINTERRUPTIBLE);
                        spin_unlock_irq(&bitmap->counts.lock);
-                       io_schedule();
+                       schedule();
                        finish_wait(&bitmap->overflow_wait, &__wait);
                        continue;
                }
index 36a8fc0..9f01870 100644 (file)
@@ -870,6 +870,44 @@ do_sync_io:
        pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 }
 
+struct raid1_plug_cb {
+       struct blk_plug_cb      cb;
+       struct bio_list         pending;
+       int                     pending_cnt;
+};
+
+static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
+{
+       struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
+                                                 cb);
+       struct mddev *mddev = plug->cb.data;
+       struct r1conf *conf = mddev->private;
+       struct bio *bio;
+
+       if (from_schedule) {
+               spin_lock_irq(&conf->device_lock);
+               bio_list_merge(&conf->pending_bio_list, &plug->pending);
+               conf->pending_count += plug->pending_cnt;
+               spin_unlock_irq(&conf->device_lock);
+               md_wakeup_thread(mddev->thread);
+               kfree(plug);
+               return;
+       }
+
+       /* we aren't scheduling, so we can do the write-out directly. */
+       bio = bio_list_get(&plug->pending);
+       bitmap_unplug(mddev->bitmap);
+       wake_up(&conf->wait_barrier);
+
+       while (bio) { /* submit pending writes */
+               struct bio *next = bio->bi_next;
+               bio->bi_next = NULL;
+               generic_make_request(bio);
+               bio = next;
+       }
+       kfree(plug);
+}
+
 static void make_request(struct mddev *mddev, struct bio * bio)
 {
        struct r1conf *conf = mddev->private;
@@ -883,6 +921,8 @@ static void make_request(struct mddev *mddev, struct bio * bio)
        const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
        const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
        struct md_rdev *blocked_rdev;
+       struct blk_plug_cb *cb;
+       struct raid1_plug_cb *plug = NULL;
        int first_clone;
        int sectors_handled;
        int max_sectors;
@@ -1185,11 +1225,22 @@ read_again:
                mbio->bi_private = r1_bio;
 
                atomic_inc(&r1_bio->remaining);
+
+               cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
+               if (cb)
+                       plug = container_of(cb, struct raid1_plug_cb, cb);
+               else
+                       plug = NULL;
                spin_lock_irqsave(&conf->device_lock, flags);
-               bio_list_add(&conf->pending_bio_list, mbio);
-               conf->pending_count++;
+               if (plug) {
+                       bio_list_add(&plug->pending, mbio);
+                       plug->pending_cnt++;
+               } else {
+                       bio_list_add(&conf->pending_bio_list, mbio);
+                       conf->pending_count++;
+               }
                spin_unlock_irqrestore(&conf->device_lock, flags);
-               if (!mddev_check_plugged(mddev))
+               if (!plug)
                        md_wakeup_thread(mddev->thread);
        }
        /* Mustn't call r1_bio_write_done before this next test,