svn commit: r212439 - head/sys/fs/nfs

[mdf at FreeBSD.org]

On Sun, Sep 12, 2010 at 12:41 AM, Rick Macklem <rmacklem at uoguelph.ca> wrote:
>> Then, fid_reserved is no more reserved ? Should we rename it ?
>>
>> Comment for fid_reserved about longword alignment is wrong.
>
> Well, it's actually more broken than that.
> fid_len - Most file systems set it to the size of their variant
>          of the entire structure, including the Xfid_len field.
>          ZFS sets it to the size of the structure - sizeof(uint16_t)
>          { presumably subtracting out the size if Xfid_len? }.
>          And xfs, well, it does weird stuff with it I can't figure
>          out, but it is definitely not the size of the entire struct.
>
> As such, exposing fid_len above the VOP_xxx() doesn't make much sense.
> (After my commit yesterday, nothing above the VOP_VPTOFH() uses it.)
>
> Personally, I'd lean towards a generic struct fid like...
> struct fid {
>       uint8_t fid_data[MAXFIDSZ];
> };

Isilon would love a generic struct like this, as to fit our filesystem
we had to make such a change locally anyways. :-)

Cheers,
matthew

> with MAXFIDSZ increased appropriately, but this will require changes
> to xfs and zfs, since they both set the generic fid_len.
>
> If you go with...
> struct fid {
>       uint16_t fid_len;
>       uint8_t fid_data[MAXFIDSZ];
> };
> then the hash functions in the two NFS servers need to be changed
> (they assume 32bit alignment of fid_data), but they should be fixed
> anyhow, since they mostly hash to 0 for ZFS at this time. (From what
> I see ZFS file handles looking like.)
>
> Or, you could just rename fid_reserved to fid_pad and not worry about it.
>
> Maybe the ZFS folks could decide what they would prefer? rick
>