On Mon, 2019-06-10 at 15:36 -0400, Ewan D. Milne wrote:
> On Mon, 2019-06-10 at 11:37 -0700, James Bottomley wrote:
> > On Mon, 2019-06-10 at 23:03 +0800, Ming Lei wrote:
> > > The current way isn't safe for chained sgl, so use sgl helper to
> > > operate sgl.
> >
> > The advansys driver doesn't currently use a chained
> > scatterlist. In
> > theory it could; the
> >
> > if (shost->sg_tablesize > SG_ALL) {
> > shost->sg_tablesize = SG_ALL;
> > }
> >
> > At around line 11226 is what prevents it and that could be
> > eliminated provided someone actually has the hardware to test.
> >
> > However, provided drivers make the correct SG_ALL or less
> > declaration, they're entitled to treat scatterlists as fully
> > contiguous, so there's no real justification (beyond uniformity)
> > for making it use the chain helpers.
> >
> > James
> >
>
> I thought the whole issue came about because Ming's earlier changes
> to scsi_lib.c made the previously SG_CHUNK_SIZE scatterlist allocated
> with the struct request much smaller, (SCSI_INLINE_SG_CNT is 2) so
> everything needs to support it?
Um, well, I'm just going by the commit log. If the problem is someone
broke the SG_ALL no chaining assumption in an earlier commit, finger
that as the buggy commit you're fixing rather than implying the drivers
had a longstanding bug. In fact, preferably do this work as a
precursor to the original instead of leaving the drivers broken for a
bisect.
James
