[SeaBIOS] A question on SeaBIOS PCI Enumeration

[Isaku Yamahata]

On Thu, Feb 10, 2011 at 11:11:41AM -0500, Kevin O'Connor wrote:
> On Thu, Feb 10, 2011 at 03:51:04PM +0000, Prasad Joshi wrote:
> > Thanks a lot for your reply.
> > 
> > On Thu, Feb 10, 2011 at 3:31 PM, Kevin O'Connor <kevin at koconnor.net> wrote:
> > > On Mon, Feb 07, 2011 at 07:17:52PM +0000, Prasad Joshi wrote:
> > >> Hello,
> > >>
> > >>
> > >> As I notice on my host machine the same case happens, but host BIOS seems to
> > >> be allocating 128MB memory from some other region (c0000000-c7ffffff)
> > >>
> > > [...]
> > >> Can SeaBIOS do the same thing? or to put in other words Can SeaBIOS allocate
> > >> memory from some other region instead of fixed f0000000 to (0xfec00000-1)?
> > >
> > > It should be possible to alter BUILD_PCIMEM_START in src/config.h.
> > > The reason it is at 0xf0000000 is because regular memory can go up to
> > > 0xe0000000, and the bochs vga bios had soe dependency on the address
> > > space at 0xe0000000 - though this may no longer be the case.
> > 
> > I modified the memory range according the lspci -vv output.
> > Atleast on my machine, the memory range 0xf0000000 to 0xfec00000 is
> > reserved for non-prefectable memory.
> > 
> > Here is a small modification. As long as it keeps the behavior for old
> > pci assignments, I guess this would not be a problem.
> > 
> > diff --git a/src/config.h b/src/config.h
> > index 5b40488..19729cb 100644
> > --- a/src/config.h
> > +++ b/src/config.h
> > @@ -49,20 +49,18 @@
> > 
> >  // Support old pci mem assignment behaviour
> >  //#define CONFIG_OLD_PCIMEM_ASSIGNMENT    1
> > -#if CONFIG_OLD_PCIMEM_ASSIGNMENT
> >  #define BUILD_PCIMEM_START        0xf0000000
> >  #define BUILD_PCIMEM_SIZE         (BUILD_PCIMEM_END - BUILD_PCIMEM_START)
> >  #define BUILD_PCIMEM_END          0xfec00000    /* IOAPIC is mapped at */
> > +
> > +#if CONFIG_OLD_PCIMEM_ASSIGNMENT
> >  #define BUILD_PCIPREFMEM_START    0
> >  #define BUILD_PCIPREFMEM_SIZE     0
> >  #define BUILD_PCIPREFMEM_END      0
> >  #else
> > -#define BUILD_PCIMEM_START        0xf0000000
> > -#define BUILD_PCIMEM_SIZE         0x08000000    /* half- of pci window */
> > -#define BUILD_PCIMEM_END          (BUILD_PCIMEM_START + BUILD_PCIMEM_SIZE)
> > -#define BUILD_PCIPREFMEM_START    BUILD_PCIMEM_END
> > +#define BUILD_PCIPREFMEM_START    0xc0000000
> 
> As before, I believe real memory could be at 0xc0000000-0xe0000000, so
> this range can't be used for PCI bars.

Qemu may assign RAM on [0xc0000000, 0xe0000000), so the area can't be
used blindly for PCI bars.
To enlarge PCI bar area < 4G, I suppose the followings are necessary.
- enhance qemu RAM assignment logic not to allocate the region.
  i.e. customize top of low memory(TOLM).
  Maybe qemu property or machine specific option.
- make the initialization for pci bar area,
  pci_bios_mem_region and pci_bios_prefmem_region, in seabios
  aware of RAM region.
  Maybe calculate those region dynamically based on TOLM

To allocate BARs more efficiently, they can be sorted in its size.
I haven't implemented it because it requires one more pass to scan pci bus.
One for sorting BARs and another for allocating BARs.

> >  #define BUILD_PCIPREFMEM_SIZE     (BUILD_PCIPREFMEM_END -
> > BUILD_PCIPREFMEM_START)
> > -#define BUILD_PCIPREFMEM_END      0xfec00000    /* IOAPIC is mapped at */
> > +#define BUILD_PCIPREFMEM_END      0xdfffffff
> >  #endif
> > 
> >  #define BUILD_APIC_ADDR           0xfee00000
> > 
> > Also, in the function pci_bios_allocate_region(), discards the
> > prefectable memory region and places it in the non-prefectable memory
> > region, if the bus is 0.
> > 
> >          } else if ((val & PCI_BASE_ADDRESS_MEM_PREFETCH) &&
> >                     /* keep behaviour on bus = 0 */
> >                     pci_bdf_to_bus(bdf) != 0 &&
> >                     /* If pci_bios_prefmem_addr == 0, keep old behaviour */
> >                     pci_region_addr(&pci_bios_prefmem_region) != 0) {
> > 
> > I am not sure if this can be changed, can you please shade more light on this?

Seabios wasn't aware of prefetchable memory before my patches.
I kept the old behavior for compatibility with the compile time option
because I was afraid that there might be something which depends
on the logic. If we don't care of it, we can drop the compat code.

So far no one has complained about the new behavior, it might be safe to
drop it.
-- 
yamahata