FUJITA Tomonori wrote:
quoted text
>>> I can't see what changing the meaning of rq->data_len (and
>>> investigating all the block drivers) gives us.
>> No matter which way you go, you change the meaning of rq->data_len and
>> you MUST inspect rq->data_len usage whichever way you go.
> 
> The patch doens't change that rq->data_len means the true data
> length. But yeah, it breaks rq->data_len == sum(sg). So it might break
> some drivers.

Yeah, that's what I was saying.  You end up breaking one of the two
assumptions.  As sglist is getting modified for any driver if it has DMA
alignment set, whether rq->data_len is adjusted together or not, sglist
and data_len usages have to be audited.

quoted text
>> Apply your patch and try to do sg IO on IDE cdrom w/ various
>> transfer lengths.
> 
> I've just tried the patch with both ata and libata and it seems to
> work.

Right, I missed you added extra_len in libata and IDE isn't using block
layer stuff yet.

quoted text
> For anyone hitting this problem, please try the following patch:
> 
> http://lkml.org/lkml/2008/3/2/218

Whether rq->data_len stays with requested data buffer size or sum(sg), I
think we need to separate out padding from address alignment; otherwise,
we'll have to audit every block driver to make sure they can deal with
extended sglist no matter which value rq->data_len ends up indicating.

If padding is applied iff explicitly requested, rq->data_len indicates
matters only to the drivers which want to see the data length adjusted,
so most of the problems go away.

-- 
tejun
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/