Re: acpi based pci gap calculation - v3

[Alok Kataria <akataria@vmware.com>]

On Wed, 2008-07-23 at 10:52 -0700, Jesse Barnes wrote:
> On Wednesday, July 23, 2008 10:10 am Alok Kataria wrote:
> > Hi Jesse,
> >
> > On Wed, 2008-07-23 at 09:58 -0700, Jesse Barnes wrote:
> > > On Tuesday, July 22, 2008 3:52 pm Alok Kataria wrote:
> > > > So the gap that we had calculated first i.e. from e820_setup_gap did
> > > > contain a collision i.e. though a resource was reserved from
> > > > [0xf0000000 - 0xf7ffffff] our gap calculation doesn't take that into
> > > > account.  My patch fixes this issue.
> > > >
> > > > So, IMHO this is a BUG and should be fixed. Please let me know your
> > > > views.
> > >
> > > Yes, there's a conflict there, but on many machines it's probably a
> > > harmless one.  My main concerns are these:
> > >   1) it changes long standing behavior and doesn't fix any real reported
> > > bugs I'm aware of (feel free to point me at some)
> >
> > The problem that I see is with Memory Hotadd, though my BIOS exports the
> > correct SRAT table and tells the OS that some regions are hot pluggable,
> > PCI gap calculation ignores that info and assigns some "option ROMS" in
> > hot pluggable memory region.
> >
> > Because of this when i try to hot add memory, the OS sees a resource
> > allocation conflict and bails out. Which shouldn't have happened,
> > right ?
> 
> Ah ok, so you've got a real problem here.  And I don't deny that there's a
> conflict that we should fix.

Thanks :-)

> 
> > >   2) it looks like it will dramatically reduce the available PCI resource
> > >      space on at least some platforms, and that space is already scarce
> > > in our current scheme
> >
> > IMO, these gaps are used only for the option ROMS or hot pluggable
> > devices which in itself are rare. So its not like we are reducing the
> > whole available PCI resource space, but only the space that is needed by
> > these optional ROMS.
> > And i think its inevitable if we have to remove these conflicts with any
> > other subsystem.
> 
> Actually we have other allocations.  In many cases the BIOS won't assign MMIO
> resources, so we do it at boot time.  Most of the open PCI bugs we have today
> are resource allocation failures, and not just for ROMs, so hopefully you can
> understand my worry.  Using all available gaps rather than just the largest
> seems like the obvious first step to increasing our available space, and
> would give us more leeway to avoid stomping on other reserved space.
> 

Yeah i agree that this should be the approach. 
I went through TJ's wiki and in the solutions section the first point
talks about 
"keeping a list of all available PCI iomem regions which will be derived
from the e820 map (and the CRS object of the root PCI device). 

TJ/Jesse, do we have any patches/work-in-progress which actually does
this ?
I can then add the CRS object stuff to that then. 

Thanks,
Alok