[Cluster-devel] [RFC] gfs2: Add sb and rgrp fields to aid fsck and grow

[Andrew Price]

[Didn't CC cluster-devel - re-sending]

Hi,

So this conversation went dormant for a while but now that the resource 
group size question is out of the way (sb_rgsize definitely isn't 
needed) and the usage of rg_next is pretty straightforward, we still 
have the semantics of rg_uuid to pin down. To recap:

diff --git a/include/uapi/linux/gfs2_ondisk.h 
b/include/uapi/linux/gfs2_ondisk.h
index db3fdd0..e425413 100644
--- a/include/uapi/linux/gfs2_ondisk.h
+++ b/include/uapi/linux/gfs2_ondisk.h
@@ -186,10 +186,11 @@ struct gfs2_rgrp {
         __be32 rg_flags;
         __be32 rg_free;
         __be32 rg_dinodes;
-       __be32 __pad;
+       __be32 rg_next;
         __be64 rg_igeneration;
+       __u8 rg_uuid[16];

-       __u8 rg_reserved[80]; /* Several fields from gfs1 now reserved */
+       __u8 rg_reserved[64]; /* Several fields from gfs1 now reserved */
  };

  /*

On 30/01/14 18:50, Steven Whitehouse wrote:
> On Thu, 2014-01-30 at 17:31 +0000, Andrew Price wrote:
>> On 29/01/14 15:32, Steven Whitehouse wrote:
>>> On Wed, 2014-01-29 at 14:47 +0000, Andrew Price wrote:
>>>> rg_uuid is intended to be the same as sb_uuid for the file system. It
>>>> can be used by fsck.gfs2, when searching for resource group headers, in
>>>> order to distinguish resource groups created as part of a previous file
>>>> system on the device from resource groups in the current file system.
>>>>
>>> Again, this could be updated by writing the rgrps back to deal with
>>> older filesystems which need to be upgraded.
>>
>> Yes.
>>
>>> That could be done as a one
>>> off sweep, or as and when we write each rgrp. Also I wonder - if the
>>> field is zero, we know that the rgrp is an old one that doesn't have it
>>> set, but if someone changes the uuid at a later date, then what?
>>
>> That's a good point.
>>
>>> Maybe
>>> we can use the uuid as a way to set it to start with (in mkfs), but
>>> after that we'd use the value from the first rgrp to fill in later
>>> rgrps. If the first rgrp was zero then we'd not update the other rgrps
>>> until the first rgrp had a value in it. Or something like that... I just
>>> want to be certain that we understand what this field will mean in all
>>> possible cases,
>>
>> Yes, unless you'd prefer a separate sb_rg_uuid field in the superblock
>> we should treat it differently to the sb_uuid after mkfs and only expect
>> the rg_uuids to be consistent with themselves. There's a corner case
>> where the first rgrp might have its uuid while the others still have
>> zero, though, which will need some more though.
>>
> I don't think we need a sb field really. We can look at the first rgrp
> and if its zero, we know that we can set it to a random value, and then
> update all the other rgrps to the same value. Or something along those
> lines, anyway,

I agree that the sb field probably isn't required. From fsck's 
perspective I'm a little worried about the lack of atomicity. If we're 
using the first rgrp's uuid as an indicator of what the others' should 
be, it might be safer for gfs2 to update all of the rgrps' uuids except 
the first one and then update the first one, once we know the rest have 
been updated. Then, if the first one is still zero we know all bets are 
off and we shouldn't even look at rg_uuid any more because an update 
wasn't completed. Is that feasible/sensible?

Andy