Note that whilst guest OSs should be relying on residual data rather than the
OF device tree to determine the hardware, we still update the interrupt
properties to match the new code in QEMU for consistency.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Since support for QEMU's 40p machine has been added there is now the possibility
of having a PPC architecture where !is_newworld() and !is_oldworld() are both
true.
Fix up a couple of instances in this logic that were preventing the generation
of interrupt properties in the device tree for 40p machines.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Reviewed-by: Hervé Poussineau <hpoussin@reactos.org>
Note that as part of this commit we rename the legacy virtio-blk device node
from "virtio-blk" to "scsi" to match up with the QEMU fw path generator.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
This provides an alternative mechanism for supporting boot device order
information from QEMU compared with the legacy FW_CFG_BOOT_DEVICE
functionality specified via -boot.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
This provides an alternative mechanism for supporting boot device order
information from QEMU compared with the legacy FW_CFG_BOOT_DEVICE
functionality specified via -boot.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
An error in the logic related to FREE_BASE meant that instead of mapping the
RAM for the loader at load-base, a small section was being mapped at the
bottom of RAM instead.
Fix this by deferring the mapping of the load-base RAM to arch_init() when
we can access the load-base variable and map 8MB RAM with a 1:1 phys to
virt mapping.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
For reasons lost in the history of time, the SPARC64 endian accessors were
given the opposite names to their function i.e. out_le32() would use a
standard load whilst out_be32() would use a little-endian load. Switch
them around so that their implementation matches their name (and also that
of all other architectures).
Note that since the references in the inb/inw/inl and outb/outw/outl wrappers
are also swapped around then these accessors for legacy ioports (i.e. little
endian) will still behave exactly the same as before.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Commit 8584d42 "SPARC64: implement dma-* words" introduced the dma-* words
for SPARC64 but accidentally introduced an older prototype implementation
(probably introduced during a rebase) rather than using the correct defer
words.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
The default context used by init-program is allocated from the OpenBIOS
heap which uses a different virt to phys mapping to avoid having to
physically relocate the PROM on startup for PPC machines.
It seems that the OpenBSD bootloader isn't happy with this change
introduced by the context rework, most likely because there is an
assumption that the client stack is mapped 1:1 virt to phys.
Fix this by overriding the default context stack during init-program
to a location just under the MMU hash table (SDR1) which was its
previous hard-coded location in 1:1 virt to phys mapped memory, and
allow OpenBSD to boot.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
The upper IO bridge limit was being calculated relative to the lower IO
bridge limit, rather than as an absolute address.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Instead push a valid address with zero length otherwise the Forth find-method
word can reference a NULL pointer.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
The problem with using the ioport as part of the controller node
address is that the address cannot be determined until after the PCI
BARs have been programmed.
This causes a problem when trying to generate fw bootpaths because by
definition they must be passed to the firmware before PCI initialisation.
Instead of using the controller ioport address, use the controller
index for the node address to provide a consistent device node
regardless of how the PCI BARs are programmed.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
A USB keyboard is not an ADB keyboard, so make sure that we set the keyboard
alias rather than the adb-keyboard alias.
This also enables us to simplify the alias logic in the PPC arch_of_init()
function.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
The PMU hardware supercedes the CUDA hardware on more modern Macs providing
additional power management functionality which is required for more modern
Mac OSs.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
PMU Mac models expose GPIOs via a separate address range at the start
of the via-pmu device.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
This is in preparation for allowing OpenBIOS to detect the difference
between CUDA/PMU hardware and build the device tree accordingly.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
PReP payloads such as Linux expect to be able to set standard VGA modes
on startup, but VBE prevents certain modes such as text-only.
Disable VBE extensions when executing the client program to allow this
behaviour which also matches that of Open HackWare.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
The PReP boot partition has its own separate layout, so if we detect a PReP
partition then assume success without attempting the filesystem probe (which
is always going to fail).
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Before the introduction of proper init-program contexts, OpenBIOS
used a simple assembler call_elf() function to execute the guest
kernel directly.
Switch over to using proper contexts via arch_init_program() which
enables us to remove the legacy call_elf() completely.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
PReP machines have a separate iova address space mapped into the CPU address
space at 0x80000000, so add a new dma-map-in implementation that supports
this behaviour for the PReP machine.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Otherwise we try unsucessfully to load the bootcode and fail rather than continuing
on to the next loader.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
More recent versions of QEMU disable scanning until the controller and
keyboard device have been explicitly reset.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Instead of using a C initialize function to create the esp device, switch over
to using the Forth "new-device" word so that all of the initialisation can be
done in place in ob_esp_init().
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
This enables disk drivers to call the dma-* words as required by the
specification, and in particular fixes the failing dma-alloc warnings
emitted by NetBSD PPC on boot.
Note that we also add these methods to any partition/file packages via a
new is-call-parent helper word to ensure that the dma-* words are
available at all levels of interposition.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Implement the dma-* words in the root node using Forth, except for dma-alloc
and dma-sync which require C to use the in-built aligned memory allocator and
MMU flushing.
Note that as OpenBIOS doesn't enable the IOMMU for SPARC64 it is possible to
use the aligned memory allocator directly without any extra support code.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Implement the dma-* words in the root node using Forth, except for dma-alloc
and dma-sync which require C to use the in-built aligned memory allocator and
I-cache and D-cache flushing.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Implement flush_dcache_range() similar to flush_icache_range() in preparation
for adding dma command support.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Use Forth wrappers around the existing C functions to implement all the dma-* words
in the root package.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
This is required for ensuring that CPU memory is coherent after DMA operations
have completed.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Rather than have dvma_alloc() store the result of the physical translation
in the specified location, use the fact that we now have contiguous physical
DMA memory to implement dvma_map_in() to perform the translation separately.
This allows us to remove the pointer to the physical address parameter to
dvma_alloc() and instead perform the translation in esp.c by calling
dvma_map_in() separately.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Currently we dynamically allocate physical memory for DVMA as required, however
that makes mapping between IOVA and physical addresses almost impossible since the
resulting regions are not contiguous.
Resolve this by allocating a contiguous physical memory block that we can use to
easily map between IOVA and physical addresses.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Whilst there is an existing implementation of dma-alloc, it is currently unused
so we can make this change now.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Commit a1c2e4f "Add USB OHCI + HID driver" hardcoded the high order bit of
the PCI address to that used on PPC machines instead of using the correct
PCI_ADDR() macro.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu>
The system bus frequency has been available from the QEMU fw_cfg interface for
some time, so let's use it rather than hard-coding the relevant values.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
This fixes additional warnings found when upgrading my compiler toolset from
GCC 5.x to 7.x.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
As OFMEM memory ranges don't provide an explicit upper limit, we must manually
check that CIF claim calls lie within a valid maxmimum range if attempting to
allocate a specific address. Since the claim method provides a 1:1 mapping between
physical and virtual addresses, the valid maximum address is therefore equivalent
to the machine RAM size.
This fixes booting yaboot 1.3.17 which implements a top-down search using claim to
find the next free memory region to locate its heap. Without this fix we blindly
accept a claim for physical addresses far above the available RAM which fails
silently when we try to use the mapping later.
Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
