On Mon, 06 Nov 2000 11:25:02 -0200, Marcelo Tosatti <marcelo@xxxxxxxxxxxxxxxx>
wrote:
>On Mon, 6 Nov 2000, Nathan Scott wrote:
>
><snip>
>
>> Its quite out of date, unfortunately, but could still be
>> used as a very rough guide to items which remain to be
>> done. The last three items for example, I'm pretty sure
>> noone is currently looking at. Some of the other items
>> listed as unassigned, are actually being pursued by SGI
>> or external folk already. The list is also no longer
>> comprehensive - Steve has another version of this doc
>> which has a number of additional items which need to be
>> done, including O_SYNC, making XFS work with more recent
>> gcc versions, and several other things.
>
>Hi,
>
>One of the things which must done is kiobuf-aware block device drivers, as
>listed on the TODO list.
In principle, agreed. There is work that needs to be done in the DMA-API
layer (calls to bus_to_virt() and virt_to_bus() which need to be redone
for "kiobuf" and "struct page *" awareness), as well.
>In current XFS CVS tree, nothing except SCSI layer is aware.
>
>When the majority of the significant block drivers are kiobuf-aware, we
>can change significant places in the VM code which currently are
>buffer-head based to be kiobuf based.
Jens Axboe is working on getting the IDE driver aware of kiobufs. Martin
Peterson is looking into the LVM side of things. At one point in time, he
was going to be looking into the `md' driver....perhaps thats one thing
that you folks could pick up.
>Looking at latest XFS block layer, function ll_rw_kio:
>
> /*
> * Only support SCSI disk for now.
> *
> * ENOSYS to indicate caller
> * should try ll_rw_block()
> * for non-SCSI (e.g. IDE) disks.
> */
> if (!SCSI_DISK_MAJOR(MAJOR(dev))) {
> *error = -ENOSYS;
> goto end_io;
> }
>
>
>I dont see any reason why we can't split the kiobuf request into buffer
>heads and send via ll_rw_block here (brw_kiovec() can do that), making
>compatibility issues transparent to ll_rw_kio callers.
>
>Comments?
Again, in principle yes, you are right. However, the only major users of
kiobuf I/O to date (raw I/O and XFS) already have their own compat.
codepaths. So, its not a burning issue to be addressed...as much as say
someone stepping forward to work on the `md' driver :^)
Cheers,
-Chait.
|