Bug 225337

This is reproducable. Caused it to happen while after rebooting the cluster due
to bz 199433

[root@salem ~]# lvs -a -o +devices
  LV               VG   Attr   LSize  Origin Snap%  Move Log        Copy%  Devices
  kool             new  -wi-a- 10.00G                                     
/dev/sdc(2560)
  [kool_mimage_0]  new  vwi-a- 10.00G
  kool_mimage_1    new  -wi-a- 10.00G                                     
/dev/sde(0)
  kool_mlog        new  -wi-a-  4.00M                                     
/dev/sdd(0)
  salem            new  mwi-a- 10.00G                    salem_mlog   7.70
salem_mimage_0(0),salem_mimage_1(0)
  [salem_mimage_0] new  iwi-ao 10.00G                                     
/dev/sdc(0)
  [salem_mimage_1] new  iwi-ao 10.00G                                     
/dev/sde(2560)
  [salem_mlog]     new  lwi-ao  4.00M                                     
/dev/sdd(1)
[root@salem ~]#
[root@salem ~]#
[root@salem ~]# lvconvert -m 0 /dev/new/kool
  Logical volume kool is already not mirrored.
[root@salem ~]# lvconvert -m 1 /dev/new/kool
  Internal error: Duplicate LV name kool_mlog detected in new.
  Failed to create mirror log.

This can apparently happen by just doing mirror up and down converts.

[root@link-02 ~]# lvs -a -o +devices
  LV                 VG   Attr   LSize Origin Snap%  Move Log          Copy% 
Devices                   
  mirror1            vg   Mwi-so 5.00G                    mirror1_mlog  99.92
mirror1_mimage_0(0),mirror1_mimage_1(0)
  [mirror1_mimage_0] vg   iwi-so 5.00G                                       
/dev/sdh1(1280)           
  [mirror1_mimage_1] vg   iwi-so 5.00G                                       
/dev/sdb1(0)              
  [mirror1_mlog]     vg   lwi-so 4.00M                                       
/dev/sda1(1283)           
  mirror2            vg   Mwi-so 5.00G                    mirror2_mlog  99.92
mirror2_mimage_0(0),mirror2_mimage_1(0)
  [mirror2_mimage_0] vg   iwi-so 5.00G                                       
/dev/sdh1(2560)           
  [mirror2_mimage_1] vg   iwi-so 5.00G                                       
/dev/sdc1(0)              
  [mirror2_mlog]     vg   lwi-so 4.00M                                       
/dev/sda1(1284)           
  mirror3            vg   Mwi-so 5.00G                    mirror3_mlog 100.00
mirror3_mimage_0(0),mirror3_mimage_1(0)
  [mirror3_mimage_0] vg   iwi-so 5.00G                                       
/dev/sdh1(3840)           
  [mirror3_mimage_1] vg   iwi-so 5.00G                                       
/dev/sdd1(0)              
  [mirror3_mlog]     vg   lwi-so 4.00M                                       
/dev/sda1(1282)           

I suspect that in the quick failure case, only a recovery write failed.

It seems there is currently no mechanism to raise an event (what causes dmeventd
to reconfigure the mirror) when an error happens durring recovery.  This is
probably by design and probably the right thing to do.  This is certainly what
is causing your results.

If you lvchange -an; lvchange -ay, the problem will resolve itself.

This is less than ideal though...

We keep track of where we are in the recovery process through a variable called
'sync_search'.  This never gets set back to zero unless the mirror table is
reloaded (lvchange -an; lvchange -ay).  It might be nice to reset 'sync_search'
if it is  >= 'region_count' and we receive a 'get_resync_work' request, since we
won't get those requests if the client thinks the mirror is in-sync.

I'll have to give the above more thought, and this would definitly involve
kernel changes...  I am certain that this problem affects single machine
mirroring too.

Careful consideration will need to be made for mirrors that wish to ignore
failures (i.e. pvmove).

Wait... 'sync_search' is in the logging code.  This allows me to make the change
without affecting the kernel.  The change is:

static int _core_get_resync_work(struct log_c *lc, region_t *region)
{
	if (lc->sync_search >= lc->region_count) {
		/*
		 * FIXME: pvmove is not supported yet, but when it is,
		 * an audit of sync_count changes will need to be made
		 */
		if (lc->sync_count < lc->region_count) {
			lc->sync_search = 0;
		} else {
			return 0;
		}
	}
...

This has not been reproduced after running many up and down cmirror convert
operations. Marking VERIFIED.

Assuming this VERIFIED fix got released.  Closing.
Reopen if it's not yet resolved.