Re: pci-sysfs resource mmap broken (and PATCH)

["David S. Miller" <davem@davemloft.net>]

On Thu, 28 Apr 2005 16:39:13 +1000
Benjamin Herrenschmidt <benh@kernel.crashing.org> wrote:

> On Wed, 2005-04-27 at 22:37 -0700, David S. Miller wrote:
> > On Wed, 27 Apr 2005 23:33:11 -0600
> 
> > The 'offset' argument is defined to be page aligned
> > when passed to mmap().
> 
> mmap API in general is defined to only ever deal with page aligned
> parameters & return values no ?

Yes.

> Except from that page alignment thing ... which is the root of the
> problem.

You can hide all of those problems in libpci.a or whatever.
You page align the offset, but pass back to the user a pointer
with his sub-page offset applied to it.

The kernel wants pages, so just give it pages :-)

> > I hate to say this, but the largest consumer of this stuff is the
> > X server, so we really need to force ourselves to work in parallel
> > on clean X server support.
> 
> Cleaning X.org is my goal, this is why I'm trying to clean the kernel
> side first :) I'm also working separately on the problem of VGA access
> arbitration (We'll probably do a joint session with the desktop summit
> an the kernel summit about those issue).

Yeah, that one is all about enabling VGA forwarding in the bridges.

Taking out all of the resource garbage in the X server, and replacing
it with properly synchronized calls in the kernel for mapping ROMs
and changing the current VGA forwarding seems to be the way to go.

> > On many platforms some kind of "side effect" bit in the PTE
> > determine if store buffer compression can happen in the processor.
> > We'd want to not set such a bit for things like frame-buffers and
> > the like.
> 
> Yes, and I think that pretty much match with PCI devices exposing a
> "prefetchable" BAR, don't you agree ?

Some scsi controllers have prefetchable set in their normal
register BARs.  The sym53c8xx does if I remember correctly.

Anyways, what I'm trying to say is that blinding turning prefetchable
BAR into "don't set side effect bit in PTE" might not be so wise.

I really think it's a userlevel decision.  That's where all the ioctl()
garbage came from for the /proc/bus/pci mmap() stuff.  It was for chossing
IO vs MEM space, and also for setting these kinds of mapping attributes.