Re: [Qemu-devel] [PATCH 00/28] Memory API for 1.6: fix I/O port endianness mess

[Paolo Bonzini <pbonzini@redhat.com>]

Il 22/07/2013 22:16, Hervé Poussineau ha scritto:
>> PReP is an exception, but
>> I think it could be rewritten to use an IOMMU memory region.
> 
> PReP PCI I/O area is located at 0x80000000, up to 0xbf7fffff (in main
> memory space region), while ISA I/O area is at 0x80000000, up to
> 0x8000ffff (size=64KB)

or up to 0x807ffffff for non-contiguous mode.  The IOMMU memory region
would let you implement non-contiguous mode without calling
cpu_inb/cpu_outb.

> However, as they are overlapped, some strange things can happen.
> For example, IBM 40p firmware configures the PCI SCSI bar at 0x20000000
> (ie 0xa0000000 in main memory), while Linux sets bar to 0x1000 (ie
> 0x80001000 in main memory), ie also in ISA I/O space.

If BARs have a lower priority than assigned ISA I/O space, this should
"just work" if you use an alias memory region for ISA I/O space.  Gaps
in the ISA I/O space will let you see through the ISA I/O space and
access BARs.

If BARs have a higher priority, you need to set the priority accordingly
for the ISA I/O space alias (using memory_region_add_subregion_overlap),
but that's it.

By the way, i82378.c can also use memory_region_init_alias to initialize
the memory regions that are passed to pci_register_bar.  I didn't do
this in this series.

> I don't know exactly what you mean by an "IOMMU memory region", but how
> would you modelize it, so that 0x80001000 and 0xa0000000 accesses are
> redirected to PCI SCSI card, while 0x800003f8 redirects (for example) to
> an ISA serial port?

See above.

Paolo

> If you create a new memory region for ISA I/O space, and you redirect
> all accesses from 0x80000000-0x8000ffff to this new address space,
> 0x80001000 won't work to access the SCSI I/O bar (located in the PCI I/O
> address space).