Steve Lord wrote:
Also Note: all of XFS's meta data writes when using LVM or MD devices
should be in 512byte chunks.
The switching sizes messeages is a bit weird as no size should be
anything
but 512 (meta data) or 4096 (file data)
>
> > Hi Steve;
> >
> > Several issues are coming up in debugging the RAID5
> > and XFS issue(s). We will try this patch, but there are
> > larger issues as well one of our engineers flushed out
> > with Martin and Neils' help. However, this will solve
> > part of them.
> >
> > There is still the issue of doing 512-byte writes and
> > RAID 5 only does 1024. We are working on it with Neil
> > and Martin with two of our engineers: Danny Cox (who sent
> > this email), dcox@xxxxxxxxxx; and Robert Lasirona,
> > rlasirona@xxxxxxxxxxx See the email below.
> >
> > Danny sent a patch to allow mkfs.xfs and mounting after
> > play well while it is resyncing that we are trying now
> > and will send on to the list.
> >
> > BTW, why are 512-byte writes done?
> >
> > thanks, Scott
> >
>
> 512 byte writes are done by XFS because we have some metadata which is always
> 512 bytes long - there are 4 fields at the head of each allocation group which
> are fixed at this size. The superblock and some headers for free block and
> inode trees.
>
> Because of the transactional nature of xfs, there are conceivably cases
> where one chunk of 512 bytes must be written to disk (because it is at
> the tail of the log and we need more log space), and the other 512 bytes
> in the same 1K chunk must not be (because it is modified in memory and
> not written to disk in the log yet).
>
> Having said all that, I am not totally sure how it fits in with this:
>
> SSmyth@xxxxxxxxxx said:
> >> Raid 5 resync is done in "multiples of 1K", but XFS appears to need
> >> 512-byte resyncs as the base unit. Sorry for the terse explanation
> >> that probably seemed unclear.
>
> I have not looked at the raid5 code, so I don't know at what level it
> is doing the resync.
>
> Steve
--
Russell Cattelan
--
Digital Elves inc. -- Currently on loan to SGI
Linux XFS core developer.
|