On Fri, 2002-08-09 at 15:50, Yee Man Chan wrote:
> Recently when I was hacking squid, I find a need to
> update the metadata. Then I come up with a thought
> that might allow us to update the revalidated
> metadata.
What sort of updates are you looking to do? Updates to metadata already
occur, and get saved to swap.state. We lose those if the swap gets
corrupted though :[.
I think that the store interface should (if not now it can be changed)
allow the updates you need, as this makes the on-disk behaviour
orthogonal, allowing COSS to handle it differently to (say) DISKD.
We need to finish rearranging the IO code so that object directory code
(COSS, UFS/AUFS/DISKD, NULL) and IO code (AsyncIO, Kqueue, DiskD, UFS,
Win32) are separate - and then metadata updates that need to go to disk
will be an object directory responsibility.
disabling O_APPEND is not good, as Terence mentioned. If we need
arbitrary writes we should open a second file handle within the
directory code, but see above for why this shouldn't matter short term
anyway.
Rob
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:16:02 MST