OvmfPkg/PlatformPei: reorder the 32-bit PCI window vs. the PCIEXBAR on q35

OvmfPkg/PlatformPei: reorder the 32-bit PCI window vs. the PCIEXBAR on q35

Commit 7b8fe63561b4 ("OvmfPkg: PlatformPei: enable PCIEXBAR (aka MMCONFIG
/ ECAM) on Q35", 2016-03-10) claimed that,

  On Q35 machine types that QEMU intends to support in the long term, QEMU
  never lets the RAM below 4 GB exceed 2 GB.

Alas, this statement came from a misunderstanding that occurred while we
worked out the interface contract. In fact QEMU does allow the 32-bit RAM
extend up to 0xB000_0000 (exclusive), in case the RAM size falls in the
range (0x8000_0000, 0xB000_0000) (i.e., the RAM size is greater than
2048MB and smaller than 2816MB).

In turn, such a RAM size (justifiedly) triggers

  ASSERT (TopOfLowRam <= PciExBarBase);

in MemMapInitialization(), because we placed the 256MB PCIEXBAR at
0x8000_0000 (2GB) exactly, relying on the interface contract. (And, the
32-bit PCI window would follow the PCIEXBAR, covering the [0x9000_0000,
0xFC00_0000) range.)

In order to fix this, reorder the 32-bit PCI window against the PCIEXBAR,
as follows:

- start the 32-bit PCI window where it starts on i440fx as well, that is,
  at 2GB or TopOfLowRam, whichever is higher;

- unlike on i440fx, where the 32-bit PCI window extends up to 0xFC00_0000,
  stop it at 0xE000_0000 on q35,

- place the PCIEXBAR at 0xE000_0000.

(We cannot place the PCIEXBAR at 0xF000_0000 because the 256MB MMIO area
that starts there is not entirely free.)

Before this patch, the 32-bit PCI window used to only *end* at the same
spot (namely, 0xFC00_0000) between i440fx and q35; now it will only
*start* at the same spot (namely, 2GB or TopOfLowRam, whichever is higher)
between both boards.

On q35, the maximal window shrinks from

  0xFC00_0000 - 0x9000_0000 = 0x6C00_0000 == 1728 MB

to

  0xE000_0000 - 0x8000_0000 == 1536 MB.

We lose 192 MB of the aperture; however, the aperture is now aligned at
1GB, rather than 256 MB, and so it could fit a 1GB BAR even.

Regarding the minimal window (triggered by RAM size 2815MB), its size is

  0xE000_0000 - 0xAFF0_0000 = 769 MB

which is not great, but probably better than a failed ASSERT.

Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Gerd Hoffmann <kraxel@redhat.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Ref: https://bugzilla.tianocore.org/show_bug.cgi?id=1814
Ref: https://bugzilla.redhat.com/show_bug.cgi?id=1666941
Ref: https://bugzilla.redhat.com/show_bug.cgi?id=1701710
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Philippe Mathieu-Daude <philmd@redhat.com>