There are two ways of doing this - the first is to add a CRC to the
remote attribute entry in the attribute block. The second is to
treat them similar to the remote symlink, where each fragment has
it's own header and identifies fragment location in the attribute.
The problem with the CRC in the remote attr entry is that we cannot
identify the owner of the metadata from the metadata blocks
themselves, or where the blocks fit into the remote attribute. The
down side to this approach is that we never know when the attribute
has been read from disk or not and so we have to verify it every
time it is read, and we must calculate it during the create
transaction and log it. We do not log CRCs for any other metadata,
and so this creates a unique set of coherency problems that, in
general, are best avoided.
Adding an identifying header to each allocated block allows us to
identify each fragment and where in the attribute it is located. It
enables us to rebuild the remote attribute from just the raw blocks
containing the attribute. It also provides us to do per-block CRCs
verification at IO time rather than during the transaction context
that creates it or every time it is read into a user buffer. Hence
it avoids all the problems that an external, logged CRC has, and
provides all the benefits of self identifying metadata.
The only complexity is that we have to add a header per fragment,
and we don't know how many fragments will be needed prior to
allocations. If we take the symlink example, the header is 56 bytes
and hence for a 4k block size filesystem, in the worst case 16
headers requires 1 extra block for the 64k attribute data. For 512
byte filesystems the worst case is an extra block for every 9
fragments (i.e. 16 extra blocks in the worse case). This will be
very rare and so it's not really a major concern.
Because allocation is done in two steps - the first finds a hole
large enough in the attribute file, the second does the allocation -
we only need to find a hole big enough for a worst case allocation.
We only need to allocate enough extra blocks for number of headers
required by the fragments, and we can calculate that as we go....
Hence it really only makes sense to use the same model as for
symlinks - it doesn't add that much complexity, does not require an
attribute tree format change, and does not require logging
calculated CRC values.
Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Ben Myers <bpm@sgi.com> Signed-off-by: Ben Myers <bpm@sgi.com>