quoted

On Thu, 24 Jan 2013 14:02:52 +0800 majianpeng [off-list ref] wrote:

quoted

There is iobarrier in raid1 because two reason resync/recovery  or reconfigure the array.
At present,it suspend all nornal IO when reysync/recovey.
But if nornal IO is outrange the resync/recovery windwos,it  don't need to iobarrier.
So I rewrite the iobarrier.
Because the reasons of barrier are two,so i use two different methods.
First for resync/recovery, there is a reysnc window.The end position is 'next_resync'.Because the resync depth is  RESYNC_DEPTH(32),
so the start is 'next_resync - RESYNC_SECTOR * RESYNC_DEPTH'
The nornal IO Will be divided into three categories by the location.
a: before the start of resync window
b: between the resync window
c: after the end of resync window
For a, it don't barrier.
For b, it need barrier and used the original method
For c, it don't barrier but it need record the minimum position.If next resync is larger this, resync action will suspend.Otherwise versa.
I used rbtree to order those io.

For the reason of reconfigure of the arrary,I proposed a concept "force_barrier".When there is force_barrier, all Nornam IO must be suspended.

NOTE:
Because this problem is also for raid10, but i only do it for raid1. It is post out mainly to make sure it is 
going in the correct direction and hope to get some helpful comments from other guys.
If the methods is accepted,i will send the patch for raid10.

Signed-off-by: Jianpeng Ma <redacted>

Hi,
thanks for this and sorry for the delay in replying.

Hi, sorroy for delay in replying.Thanks very much for your suggestion.

quoted

The patch is reasonably good, but there is room for improvement.
I would break it up into several patches which are easier to review.

- Firstly, don't worry about the barrier for 'read' requests - it really
 isn't relevant for them (your patch didn't do this).

diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c
index bd6a188..2e5bf75 100644
--- a/drivers/md/raid1.c
+++ b/drivers/md/raid1.c

@@ -235,6 +235,7 @@ static void call_bio_endio(struct r1bio *r1_bio)
        struct bio *bio = r1_bio->master_bio;
        int done;
        struct r1conf *conf = r1_bio->mddev->private;
+       int rw = bio_data_dir(bio);
 
        if (bio->bi_phys_segments) {
                unsigned long flags;

@@ -253,7 +254,8 @@ static void call_bio_endio(struct r1bio *r1_bio)
                 * Wake up any possible resync thread that waits for the device
                 * to go idle.
                 */
-               allow_barrier(conf);
+               if (rw == WRITE)
+                       allow_barrier(conf);
        }
 }
 

@@ -1035,7 +1037,8 @@ static void make_request(struct mddev *mddev, struct bio * bio)
                finish_wait(&conf->wait_barrier, &w);
        }
 
-       wait_barrier(conf);
+       if (rw == WRITE)
+               wait_barrier(conf);
 
        bitmap = mddev->bitmap;

Above code is what's you said.But it met read-error,raid1d will blocked for ever.
The reason is freeze_array:

quoted

      wait_event_lock_irq_cmd(conf->wait_barrier,
                               conf->nr_pending == conf->nr_queued+1,

For read operation, it can't add nr_pending.
Are you have good method to do this?