Re: [PATCH 09/10] Change table chaining layout

[Jens Axboe <jens.axboe@oracle.com>]

On Fri, Oct 26 2007, Linus Torvalds wrote:
> 
> 
> On Fri, 26 Oct 2007, Paul Mackerras wrote:
> >
> > Linus Torvalds writes:
> > 
> > > Nobody should *ever* walk the list to find the length. Does anybody really 
> > > do that? Yes, we pass the thing down, but do people *need* it?
> > 
> > Yes, I need it for devices that use the macintosh DBDMA
> > (descriptor-based DMA) hardware.  The DBDMA hardware reads an array of
> > descriptors from system RAM, so I need to allocate an array and fill
> > it in with DBDMA command blocks (and then dma-map it and point the
> > device at it).
> 
> Yes, for allocation purposes you'd need the size ahead of time, agreed. 
> Otherwise you have to walk the list twice.

Do you really allocate a fresh table for every command, or just a max
sized one at init?

> > Maybe the drivers for devices that use DBDMA are now buggy.  Certainly
> > filling in the array of DBDMA command blocks involves walking the
> > list, but it would extremely useful to know how much to allocate
> > before we start filling them in.  So we at least need an upper bound
> > on the number of "real" entries, even if we don't have the exact
> > number.
> 
> Hmm. Depending on where you do this, and if this is some block-layer 
> specific driver/code (rather than necessarily a generic SG thing), you do 
> have the req->nr_phys_segments thing which should be that for you (ie the 
> SG list may have _fewer_ requests in it in case some of those entries got 
> squashed together due to be contiguous).
> 
> But yeah, I don't think it would be wrong at all to have a
> 
> 	struct scatterlist_head {
> 		unsigned int entries;
> 		unsigned int flags;	/* ? */
> 		struct scatterlist *sg;
> 	};
> 
> which would be passed down at higher levels.

That'd be fine with me as well, but I really don't think that a lot of
people really do need the sg count when you can just loop over the table
until it returns NULL.

-- 
Jens Axboe