Re: [PATCHSET 0/5] Peaceful co-existence of scsi_sgtable and Large IO sg-chaining

[Boaz Harrosh <bharrosh@panasas.com>]

FUJITA Tomonori wrote:
> From: Boaz Harrosh <bharrosh@panasas.com>
> Subject: [PATCHSET 0/5] Peaceful co-existence of scsi_sgtable and Large IO sg-chaining
> Date: Tue, 24 Jul 2007 11:47:50 +0300
> 
>> As Jens said, there is nothing common to scsi_sgtable and 
>> sglists. Save the fact that it is a massive conflict at 
>> scsi-ml. They touch all the same places.
>>
>> Proposed is a simple way out. Two patchsets That produce the
>> same output at the end.
>>
>> One: scsi_sgtable_than_sg-chaining
>> Two: sg-chaining_than_scsi_sgtable
> 
> Hmm, I thought that I've already posted a scsi_sgtable patch working
> with sg-chaining together.
> 
> http://marc.info/?l=linux-scsi&m=118519987632758&w=2
> 
> I quoted from my mail:
> 
> ---
> I think that the main issue of integrating sgtable and sglist is how
> to put scatterlist to scsi_sgtable structure.
> 
> If we allocate a scsi_sgtable structure and sglists separately, the
> code is pretty simple. But probably it's not the best way from the
> perspective of performance.
> 
I was just answering your other mail when this came in so I'll answer
here. 
This Approach is exactly the scsi_data_buffer approach we both
had solutions for. At the time I was all for that approach because it
is safer and could be kept compatible to old drivers. (Less work for
me) But it was decided against it. So suggesting it again is plain
going back.

Now that We have done the work, I must say that James was right. The 
sgtable work is nice and elegant. And makes scsi-ml more simple not
more complicated. And the bidi looks beautifully trivial on-top of
sgtables.

> If we put sglists into the scsi_sgtable structure (like Boaz's patch
> does) and allocate and chain sglists only for large I/Os, we would
> have the better performance (especially for small I/Os). But we will
> have more complicated code.

Only that my proposed solution is more simple more elegant and more
robust than even Jens's solution without any sgtable at all. Actually 
the sgtable does good to chaining. scsi_lib with sglist+sgtable is 
smaller than Jens's sglist by itself.

> ---
> 
> From a quick look over your patchset, you chose the latter. And my
> patch uses the former.

This patchset is not new. I have sent that solution before.
(http://marc.info/?l=linux-scsi&m=117933686313822&w=2)
In fact when first programing for sgtable I had in mind
Jens's work to make sure it will fit and converge.

Please lets not go back to old solutions again

Boaz