On Tue, Feb 28, 2012 at 09:14:41AM +0000, Brian Candler wrote:
> On Tue, Feb 28, 2012 at 12:48:13PM +1100, Dave Chinner wrote:
> > > Write cache: enabled
> > There's your problem. You should read all the relevant sections on
> > write caches in the XFS FAQ. e.g:
> > http://xfs.org/index.php/XFS_FAQ#Q._Which_settings_does_my_RAID_controller_need_.3F
> As a side question to this: if I am using Linux's md software RAID with an
> HBA, and XFS is using write barriers (as I believe it does by default), is
> it still necessary to disable write caching on the drives?
No. The MD layers pass the correct flushes to the drives now.
> My reading of the FAQ suggests that the write barrier by itself should be
> sufficient - but this requires that the md RAID driver implements write
> barriers correctly. I seem to remember reading a long time ago that LVM
> and/or dmraid didn't, although that may be different now.
It has since about 2.6.35.
> Incidentally I'm using 'mdadm ... -b internal' to ensure that the RAID array
> marks dirty areas, so that if a write part-completes (e.g. a block is
> written to only one half of a RAID1) then at next power-up this should be
> made consistent again.
Assuming that the dirty area tracking uses write cache flushes