When there are no devices connected to the root bridge, no resources are
needed. GetProposedResources() currently considers this an invalid
condition (the PI spec doesn't regulate it).
Emitting an empty set of EFI_ACPI_ADDRESS_SPACE_DESCRIPTORs, followed by
the required EFI_ACPI_END_TAG_DESCRIPTOR, allows
PciHostBridgeResourceAllocator() [MdeModulePkg/Bus/Pci/PciBusDxe/PciLib.c]
to advance.
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Olivier Martin <Olivier.martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16908 6f19259b-4bc3-4df7-8a09-765794883524
According to Volume 5 of the PI spec, 10.8.2 PCI Host Bridge Resource
Allocation Protocol, SubmitResources(),
It is considered an error if no resource requests are submitted for a
PCI root bridge. If a PCI root bridge does not require any resources, a
zero-length resource request must explicitly be submitted.
Under MdeModulePkg/Bus/Pci/PciBusDxe/, we have:
PciHostBridgeResourceAllocator() [PciLib.c]
ConstructAcpiResourceRequestor(..., &AcpiConfig) [PciEnumerator.c]
PciResAlloc->SubmitResources(..., &AcpiConfig)
ASSERT_EFI_ERROR ()
If ConstructAcpiResourceRequestor() finds no resources to request (for
example because no PCI devices are on the root bridge), it places a
zero-length EFI_ACPI_ADDRESS_SPACE_DESCRIPTOR followed by an
EFI_ACPI_END_TAG_DESCRIPTOR in "AcpiConfig"; satisfying the PI spec.
However, PciHostBridgeDxe's SubmitResources() function does not expect
such input; the following part of the code rejects it:
switch (Ptr->ResType) {
case 0:
//
// Check invalid Address Sapce Granularity
//
if (Ptr->AddrSpaceGranularity != 32) {
return EFI_INVALID_PARAMETER;
}
Skip EFI_ACPI_ADDRESS_SPACE_DESCRIPTORs with zero AddrLen early. Also,
allow PciHostBridgeResourceAllocator() to proceed to the AllocateResources
phase by setting "ResourceSubmited" to TRUE.
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Olivier Martin <Olivier.martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16907 6f19259b-4bc3-4df7-8a09-765794883524
This is our MMIO space map:
> GCD:AddMemorySpace(Base=0000000010000000,Length=000000002EFF0000)
> GcdMemoryType = MMIO
> Capabilities = 0000000000000001
> Status = Success
> GCDMemType Range Capabilities Attributes
> ========== ================================= ================ ================
> NonExist 0000000000000000-0000000003FFFFFF 0000000000000000 0000000000000000
> MMIO 0000000004000000-0000000007FFFFFF C000000000000001 8000000000000001
NorFlashDxe adds this, but does not allocate it.
> NonExist 0000000008000000-000000000900FFFF 0000000000000000 0000000000000000
> MMIO 0000000009010000-0000000009010FFF C000000000000001 8000000000000001
Added by RealTimeClockRuntimeDxe, but also not allocated.
> NonExist 0000000009011000-000000000FFFFFFF 0000000000000000 0000000000000000
> MMIO 0000000010000000-000000003EFEFFFF C000000000000001 0000000000000000
Added by ourselves.
> NonExist 000000003EFF0000-000000003FFFFFFF 0000000000000000 0000000000000000
> SystemMem 0000000040000000-00000000BFFFFFFF 800000000000000F 0000000000000008*
> NonExist 00000000C0000000-0000FFFFFFFFFFFF 0000000000000000 0000000000000000
In the EfiPciHostBridgeAllocateResources phase, we allocate memory BARs
bottom up, from whichever MMIO range comes first and has room left.
Unfortunately, this places memory BARs into MMIO ranges that belong to
other devices. (Arguably, their respective drivers should not just add,
but immediately allocate those ranges as well.)
(
This problem is not seen in OVMF / PcAtChipsetPkg, because there we
allocate bottom-up from the range
[max(2GB, top-of-low-RAM), 0xFC000000).
(See the MMIO resource descriptor HOB created in MemMapInitialization()
[OvmfPkg/PlatformPei/Platform.c].)
That MMIO range fits in the static [2GB, 4GB) aperture given in
"mResAperture" in PcAtChipsetPkg/PciHostBridgeDxe; plus other MMIO
ranges (IO-APIC, HPET, LAPIC, flash chip) are higher than 0xFC000000.
Hence the bottom-up BAR allocation in OvmfPkg always finds the right
MMIO range first.
)
In ArmVirtualizationPkg/PciHostBridgeDxe we can solve the problem by
working our way downwards from the top of our own aperture.
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Olivier Martin <Olivier.martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16906 6f19259b-4bc3-4df7-8a09-765794883524
Currently we allocate IO BARs bottom-up in the
EfiPciHostBridgeAllocateResources phase of the enumeration.
> GCD:AddIoSpace(Base=0000000000000000,Length=0000000000010000)
> GcdIoType = I/O
> Status = Success
> GCDIoType Range
> ========== =================================
> I/O 0000000000000000-000000000000FFFF
Because the IO aperture is based at zero, the first allocation happens to
get the zero address. However, a zero address for a PCI BAR is considered
unmapped; see eg.:
- <http://www.pcisig.com/reflector/msg00459.html>,
- the (new_addr == 0) part in QEMU, pci_bar_address() [hw/pci/pci.c]:
new_addr = pci_get_long(d->config + bar) & ~(size - 1);
last_addr = new_addr + size - 1;
/* Check if 32 bit BAR wraps around explicitly.
* TODO: make priorities correct and remove this work around.
*/
if (last_addr <= new_addr || new_addr == 0 || last_addr >= UINT32_MAX)
{
return PCI_BAR_UNMAPPED;
}
We can avoid this problem by allocating top-down in the IO aperture.
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Olivier Martin <Olivier.martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16905 6f19259b-4bc3-4df7-8a09-765794883524
Quite non-intuitively, we must allow guest-side writes to emulated PCI
MMIO regions to go through the CPU cache, otherwise QEMU, whose accesses
always go through the cache, may see stale data in the region.
This change makes no difference for QEMU/TCG, but it is important for
QEMU/KVM, at the moment.
Because gDS->SetMemorySpaceAttributes() is ultimately implemented by
EFI_CPU_ARCH_PROTOCOL.SetMemoryAttributes() -- see
"MdeModulePkg/Core/Dxe/Gcd/Gcd.c" and "ArmPkg/Drivers/CpuDxe/" -- we add
the CPU architectural protocol to the module's DepEx.
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Olivier Martin <olivier.martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16904 6f19259b-4bc3-4df7-8a09-765794883524
VirtFdtDxe parses the following address space properties from the DTB (and
saves them in PCDs) :
ProcessPciHost: Config[0x3F000000+0x1000000)
Bus[0x0..0xF]
Io[0x0+0x10000)@0x3EFF0000
Mem[0x10000000+0x2EFF0000)@0x0
In order to allow PCI enumeration to allocate IO and MMIO resources from
the above ranges for devices, we must add the ranges to the Global
Coherency Domain.
There are two ways for that:
- building resource descriptor HOBs in the HOB producer phase (basically,
PEI), and letting the DXE core process them,
- calling gDS->AddIoSpace() and gDS->AddMemorySpace() during DXE.
We opt for the second method for simplicity.
In the address space maps, the corresponding ranges change from
"nonexistent" to "IO" and "MMIO", from which the gDS->AllocateIoSpace()
and gDS->AllocateMemorySpace() services can later allocate PCI BARs.
GCD:AddIoSpace(Base=0000000000000000,Length=0000000000010000)
GcdIoType = I/O
Status = Success
GCDIoType Range
========== =================================
-> I/O 0000000000000000-000000000000FFFF
GCD:AddMemorySpace(Base=0000000010000000,Length=000000002EFF0000)
GcdMemoryType = MMIO
Capabilities = 0000000000000001
Status = Success
GCDMemType Range Capabilities Attributes
========== ================================= ================ ================
NonExist 0000000000000000-0000000003FFFFFF 0000000000000000 0000000000000000
MMIO 0000000004000000-0000000007FFFFFF C000000000000001 8000000000000001
NonExist 0000000008000000-000000000900FFFF 0000000000000000 0000000000000000
MMIO 0000000009010000-0000000009010FFF C000000000000001 8000000000000001
NonExist 0000000009011000-000000000FFFFFFF 0000000000000000 0000000000000000
-> MMIO 0000000010000000-000000003EFEFFFF C000000000000001 0000000000000000
NonExist 000000003EFF0000-000000003FFFFFFF 0000000000000000 0000000000000000
SystemMem 0000000040000000-00000000BFFFFFFF 800000000000000F 0000000000000008*
NonExist 00000000C0000000-0000FFFFFFFFFFFF 0000000000000000 0000000000000000
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Olivier Martin <Olivier.martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16903 6f19259b-4bc3-4df7-8a09-765794883524
The RootBridgeIoCheckParameter() function currently relies on the range
limit being of the form (2^n - 1). This assumption is not necessarily
true; handle the general case.
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Olivier Martin <Olivier.martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16901 6f19259b-4bc3-4df7-8a09-765794883524
There is no IO space on ARM, and there are no special instructions that
access it. QEMU emulates the IO space for PCI devices with a special MMIO
range. We're ready to use it at this point, we just have to switch the
Io(Read|Write)(8|16|32) primitives to their MMIO counterparts, because in
"MdePkg/Library/BaseIoLibIntrinsic/IoLibArm.c", the IO primitives
correctly ASSERT (FALSE).
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Olivier Martin <Olivier.martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16900 6f19259b-4bc3-4df7-8a09-765794883524
Unlike the one in PcAtChipsetPkg, our PciHostBridgeDxe module must handle
address space translation. IO addresses expressed in the respective
aperture are mapped to a different base in CPU address space.
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Olivier Martin <olivier.martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16899 6f19259b-4bc3-4df7-8a09-765794883524
If VirtFdtDxe found no PCI host in the DTB, then the config space base
address will be left at zero -- the default is set in the DSC --, and we
should exit PciHostBridgeDxe immediately.
This causes gEfiPciRootBridgeIoProtocolGuid not to be installed, which in
turn prevents MdeModulePkg/Bus/Pci/PciBusDxe from binding (see
PciBusDriverBindingSupported()).
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Olivier Martin <Olivier.martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16898 6f19259b-4bc3-4df7-8a09-765794883524
Our PciHostBridgeDxe module creates one root bridge on the one and only
host bridge. The resource apertures of the root bridge (bus range, IO
space, MMIO space) are configured with the "mResAperture" array, which at
the moment carries static values inherited from PcAtChipsetPkg.
Set the array as first thing from the PCDs that we parsed from the device
tree.
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Olivier Martin <Olivier.martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16897 6f19259b-4bc3-4df7-8a09-765794883524
The Enhanced Configuration Access Mechanism provides access to 4096
register bytes per PCIe B/D/F. The MAX_PCI_REG_ADDRESS macro that we're
changing here is used by RootBridgeIoCheckParameter() for verifying config
space boundaries in EFI_PCI_ROOT_BRIDGE_IO_PROTOCOL.Pci.Read() and
.Write().
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Olivier Martin <Olivier.Martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16896 6f19259b-4bc3-4df7-8a09-765794883524
MdeModulePkg/Bus/Pci/PciBusDxe depends on
EFI_PCI_HOST_BRIDGE_RESOURCE_ALLOCATION_PROTOCOL and
EFI_PCI_ROOT_BRIDGE_IO_PROTOCOL. Here we clone the driver that produces
these from PcAtChipsetPkg, with the following immediate changes:
- a new FILE_GUID is generated;
- the assembly-language Ia32 / X64 specific IoFifo "accelerators" are not
copied, and their client code (which would be dead code anyway) is
removed;
- UNI files are not copied: they are used in conjunction with the UEFI
Packaging Tool (UPT), but the driver under ArmVirtualizationPkg will not
be part of UDK.
Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Olivier Martin <Olivier.martin@arm.com>
git-svn-id: https://svn.code.sf.net/p/edk2/code/trunk/edk2@16895 6f19259b-4bc3-4df7-8a09-765794883524
Laszlo Ersek