On Tue, 13 Dec 2011 15:45:30 +0000 "Kwolek, Adam" [off-list ref]
wrote:

Hi Neil,

On the latest md neil_for-linus branch I've found raid0 migration problem.
During OLCE in user space everything goes fine, but in kernel process is not moved forward.
/older md works fine/

It is stopped in md in reshape_request() in line (near raid5.c:3957)
    wait_event(conf->wait_for_overlap, atomic_read(&conf->reshape_stripes)==0);

I've found that this problem is a side effect of patch:
    md/raid5: abort any pending parity operations when array fails.
and added line in this patch:
     sh->reconstruct_state = 0;

During OLCE we are going inside because condition
    if (s.failed > conf->max_degraded)
with values:
     locked=1 uptodate=5 to_read=0 to_write=0 failed=2 failed_num=4,1

and sh->reconstruct_state is set to 0 (reconstruct_state_idle) from 6 (reconstruct_state_result)
When sh->reconstruct_state is not reset raid0 migration is executed without problem.
Problem is probably in not executed code for finishing reconstruction (around raid5.c:3300)

In our case field s.failed should not reach value 2 but we've got it for failed_num = 4,1. 
It seems that '1' is failed disk for stripe in old array geometry and 4 is failed disk for stripe in new array geometry.
This means that degradation during reshape is counted two times /final stripe degradation is sum of old and new geometry degradation/.
When we reading (from old array) and writing (to new geometry) a degraded stripe  and degradation is on different positions (raid0 OLCE case) analyse_stripe() gives
us false failure information. Possible that we should have old_failed and new_failed counters to know in what geometry (old/new) failure occurs.


Here is reproduction script:

export IMSM_NO_PLATFORM=1
#create container
mdadm -C /dev/md/imsm0 -amd -e imsm -n 4 /dev/sdb /dev/sdc /dev/sde /dev/sdd -R
#create array
mdadm -C /dev/md/raid0vol_0 -amd -l 0 --chunk 64 --size  1048 -n 1 /dev/sdb  -R --force
#start reshape
mdadm --grow /dev/md/imsm0 --raid-devices 4


Please let me know your opinion.

Thanks for the excellent problem report.

I think it is best fixed by the following patch.
I also need to fixed up the calculate of 'degraded' so it doesn't say '2' in
this case, which is confusing.  Then I'll commit the fixes.


Thanks,
NeilBrown

diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
index 31670f8..858fdbb 100644
--- a/drivers/md/raid5.c
+++ b/drivers/md/raid5.c

@@ -3065,11 +3065,17 @@ static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
 			}
 		} else if (test_bit(In_sync, &rdev->flags))
 			set_bit(R5_Insync, &dev->flags);
-		else {
+		else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
 			/* in sync if before recovery_offset */
-			if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
-				set_bit(R5_Insync, &dev->flags);
-		}
+			set_bit(R5_Insync, &dev->flags);
+		else if (test_bit(R5_UPTODATE, &dev->flags) &&
+			 test_bit(R5_Expanded, &dev->flags))
+			/* If we've reshaped into here, we assume it is Insync.
+			 * We will shortly update recovery_offset to make
+			 * it official.
+			 */
+			set_bit(R5_Insync, &dev->flags);
+
 		if (rdev && test_bit(R5_WriteError, &dev->flags)) {
 			clear_bit(R5_Insync, &dev->flags);
 			if (!test_bit(Faulty, &rdev->flags)) {