soc/amd/picasso/chip: use common data fabric domain resource code

Use amd_pci_domain_read_resources function that gets the configured MMIO
regions for the PCI root domain from the data fabric's MMIO decode
registers instead of using pci_domain_read_resources. This results in
the same IO port range being used by the allocator, but makes sure that
the allocator will only allocate non-fixed MMIO resources in the address
ranges that get decoded to the PCI root complex. In order for the PCI0
_CRS ACPI resource template to match the decoded PCI root domain MMIO
windows, use amd_pci_domain_fill_ssdt to generate the _CRS ACPI code
instead of having a mostly hard-coded _CRS method in the DSDT. This
makes sure that the OS will know about the MMIO regions it is allowed to
used.

Before this patch, only the region from TOM1 to right below
CONFIG_ECAM_MMCONF_BASE_ADDRESS was advertised as usable PCI MMIO in the
PCI0 _CRS method. Also the resource allocator didn't get any constraint
on which address ranges it can use to put the non-fixed MMIO resources.
This approach worked until now, since all address range from 0 up to
right below TOM1 was filled with either usable or reserved memory and
the allocator was allocating beginning right from TOM1, since it was
using the bottom-up allocation approach and everything below TOM1 was
already in use. The MMIO region from TOM1 to right below
CONFIG_ECAM_MMCONF_BASE_ADDRESS also matched the MMIO decode window
configured in the data fabric's MMIO decode registers, so everything
seemed to work fine. However, when either selecting
RESOURCE_ALLOCATION_TOP_DOWN or enabling above 4GB MMIO, things broke
badly. This was partially due to the allocator putting non-fixed MMIO
resources in regions that weren't decoded to the PCI root, since AMD
family 17h and 19h silicon doesn't subtractively decode PCI MMIO and the
wrong ranges the allocator used also weren't advertised in ACPI.

TEST=Even when selecting RESOURCE_ALLOCATION_TOP_DOWN that usually ends
up with a non-working system when the MMIO ranges aren't reported
correctly to the resource allocator due to the reasons descried above,
Ubuntu 22.04 LTS still boots on Mandolin both with SeaBIOS and EDK2
payload and Windows 10 boots with EDK payload. There's however an EDK2
bug that results the MMCONFIG region not being advertised in the e820
table, which causes Linux to not use the MMCONFIG and fall back to the
legacy PCI config access method. This only happens with EDK2 payload and
everything works fine when using SeaBIOS as payload. That e820 issue is
unaffected by this patch.

At the end of the data_fabric_set_mmio_np call, this is the data fabric
MMIO register configuration:

=== Data Fabric MMIO configuration registers ===
idx             base            limit  control R W NP F-ID
  0         fc000000         febfffff       93 x x       9
  1      10000000000     ffffffffffff       93 x x       9
  2         d0000000         f7ffffff       93 x x       9
  3         fed00000         fedfffff     1093 x x  x    9
  4                0             ffff       90           9
  5                0             ffff       90           9
  6                0             ffff       90           9
  7                0             ffff       90           9

The limit of the data fabric MMIO decode register 1 is configured as
0xffffffffffff although this is way beyond the addressable memory space.
add_data_fabric_mmio_regions fixes this up, so the range that gets
passed to the allocator in that case is 0x7fcffffffff which takes both
the reserved most significant address bits used for the memory
encryption and the 12GB reserved data fabric MMIO at the top of the
usable address space into account.

This results in the following domain ranges passed to the resource
allocator:

DOMAIN: 0000 io: base: 0 size: 0 align: 0 gran: 0 limit: ffff done
DOMAIN: 0000 mem: base: fc000000 size: 0 align: 0 gran: 0 limit: febfffff
DOMAIN: 0000 mem: base: 10000000000 size: 0 align: 0 gran: 0 limit: 7fcffffffff
DOMAIN: 0000 mem: base: d0000000 size: 0 align: 0 gran: 0 limit: f7ffffff

The IO resource producer region is split into two parts to not cover the
PCI config IO region resource consumer. This results in these resources
being added to the PCI0 _CRS resource template:

amd_pci_domain_fill_ssdt ACPI scope: '\_SB.PCI0'
PCI0 _CRS: adding busses [0-3f]
PCI0 _CRS: adding IO range [0-cf7]
PCI0 _CRS: adding IO range [d00-ffff]
PCI0 _CRS: adding MMIO range [fc000000-febfffff]
PCI0 _CRS: adding MMIO range [10000000000-7fcffffffff]
PCI0 _CRS: adding MMIO range [d0000000-f7ffffff]
PCI0 _CRS: adding VGA resource

Kernel version 5.15.0-43 from Ubuntu 2022.4 LTS prints this in dmesg:

PCI host bridge to bus 0000:00
pci_bus 0000:00: root bus resource [bus 00-3f]
pci_bus 0000:00: root bus resource [io  0x0000-0x0cf7 window]
pci_bus 0000:00: root bus resource [io  0x0d00-0xffff window]
pci_bus 0000:00: root bus resource [mem 0x000a0000-0x000bffff window]
pci_bus 0000:00: root bus resource [mem 0xd0000000-0xf7ffffff window]
pci_bus 0000:00: root bus resource [mem 0xfc000000-0xfebfffff window]
pci_bus 0000:00: root bus resource [mem 0x10000000000-0x7fcffffffff window]

Another noteworthy thing I wasn't aware of at first when testing ACPI
changes on Windows 10 is that a normal Windows shutdown and boot cycle
won't result in it processing the changed ACPI tables; you have to tell
it to reboot to do a proper full boot where it will process the updated
ACPI tables (and fail if it dislikes something about the ACPI tables and
bytecode).

Signed-off-by: Felix Held <felix-coreboot@felixheld.de>
Change-Id: Ia24930ec2a9962dd15e874e9defea441cffae9f2
Reviewed-on: https://review.coreboot.org/c/coreboot/+/74712
Reviewed-by: Raul Rangel <rrangel@chromium.org>
Tested-by: build bot (Jenkins) <no-reply@coreboot.org>
Reviewed-by: Arthur Heymans <arthur@aheymans.xyz>

0 files changed, 0 insertions, 0 deletions