On Fri, 2003-06-20 at 05:29, Gordon Henderson wrote:
> I've searched all I can, but find no real solution to the problem of the
> linux raid driver constantly outputting a stream of :
>
> Jun 20 10:15:38 red kernel: raid5: switching cache buffer size, 512 --> 4096
> Jun 20 10:15:38 red kernel: raid5: switching cache buffer size, 4096 --> 512
> Jun 20 10:15:38 red kernel: raid5: switching cache buffer size, 512 --> 4096
> Jun 20 10:15:38 red kernel: raid5: switching cache buffer size, 4096 --> 512
> Jun 20 10:15:38 red kernel: raid5: switching cache buffer size, 512 --> 4096
> Jun 20 10:15:39 red kernel: raid5: switching cache buffer size, 0 --> 4096
> Jun 20 10:15:39 red kernel: raid5: switching cache buffer size, 4096 --> 512
>
> etc. messages when doing stuff with an xfs mounted filesystem. I've tried
> putting the log on a separate raid1, and on a separate raw ide partition
> too and I still see these. Sure, I can turn them off, but it doesn't help
> the performance much. Is there a majike rune to the mkfs.xfs command that
> I've overlooked?
>
> Reading archives, etc. seems to suggest that making the logfile have a 4K
> block size would fix it, and/or putting it on a separate device, but from
> what I've already tried, it would seem to suggest that it already has a 4K
> block size anyway.
>
> Last one I tried was:
>
> # mkfs.xfs -f -b size=4096 -l logdev=/dev/hdm5,version=2 /dev/md5
>
> That puts the logfile on a separate partition on the hd device, another
> partition on that drive is part of the raid5 array, but I don't think that
> would affect it (should it?)
>
> Obviously I'd ideally like the logfile to be on a mirror or raid
> partition, but I thought I'd try to simplest solution to try to get to the
> bottom of this.
>
> meta-data=/dev/md5 isize=256 agcount=38, agsize=1048568
> blks
> data = bsize=4096 blocks=39688448, imaxpct=25
> = sunit=8 swidth=32 blks, unwritten=0
> naming =version 2 bsize=4096
> log =/dev/hdm5 bsize=4096 blocks=8024, version=2
> = sunit=8 blks
> realtime =none extsz=131072 blocks=0, rtextents=0
>
> # mount -ologdev=/dev/hdm5 /dev/md5 /mnt
>
> The system is
>
> Linux red 2.4.21-ac1 #3 SMP Wed Jun 18 20:31:28 BST 2003 i686 unknown
>
> with xfsprogs-2.3.5
>
> Any clues would be appreciated, even if it's just to comment out the
> printk in the raid5 code - performance isn't a real issue, but being up
> and runinng immediately rather than waiting an hour for an ext2 fsck (this
> box will have 4 x 150Gb partitions) is an issue should it ever crash.
>
> Gordon
With the most recent xfs user space (xfsprogs-2.5.0-0) and a kernel from
cvs, try this and see if it helps:
mkfs -t xfs -f -s size=4096 -l size=16384b /dev/mdxxxx
This will make the xfs sector size 4K and change the on disk
layout, you get a version 2 log aligned on 4K boundaries too.
I have not tried it, but this could make your problem go away.
Note that large sector support is brand new, so could probably
do with some exposure in the wild.
Steve
--
Steve Lord voice: +1-651-683-3511
Principal Engineer, Filesystem Software email: lord@xxxxxxx
|