Re: [PATCH] [14/20] Don't disable direct_io for unchecked_isa_dma in st.c

[Boaz Harrosh <bharrosh@panasas.com>]

On Mon, Mar 17 2008 at 12:55 +0200, FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp> wrote:
> On Fri, 14 Mar 2008 10:24:12 -0400
> Christoph Hellwig <hch@infradead.org> wrote:
> 
>> On Fri, Mar 14, 2008 at 02:51:35PM +0100, Jens Axboe wrote:
>>> Are you sure? Seems to me that st builds its own list with
>>> get_user_pages() for direct io, where will that get bounced? The block
>>> layer will only bounce things that are mapped directly, so if st used
>>> blk_rq_map_kern() and inserted that request in the queue, we could
>>> proceed with killing that check in st.
>> Which shouldn't be all too difficult and would be the right thing to
>> do.  We really need someone to sit down and convert st/osst/sg to use
>> the proper block layer helpers.
> 
> I've been working on this (hopefully, I'll submit an updated patchset
> this week).
> 
> I think that the main problem was a lack of motivation for these
> changes. I bet that nobody reviewed the patchset that Mike posted the
> last time. Hopefully, the new patchset will be reviewed and merged.
> 

I have ran and tested these patches for a long time, in my personal
workstation (long story). I had no problems with them. Burn CDs and all.

> With sg chaining, drain, padding, etc, building a scatterlist is much
> complicated than in the past. I believe that we should kill home-made
> scatterlist code and convert the users to use the block layer to avoid
> potential scatterlist bugs.
> --

Definitely, as I remember, Mike's code just prepares BIO's and requests
and completely avoids any use of sg-lists. I will be waiting for your 
patches and will test them here.

Thanks
Boaz