Reuse and export low_drive_freq_update() function. This way global imx9
board_fix_fdt() doesn't duplicate code. While low_drive_freq_update()
can be reused on boards such as phyCORE-i.MX93 (TARGET_PHYCORE_IMX93)
which is not using the global imx9 board_fix_fdt() implementation.
While at it, make printout logic less verbose by only outputting on the
error condition and not on each successful clock fixup. Also drop now
invalid comment (low_drive_freq_update() now does fixup for internal and
kernel device-tree).
Signed-off-by: Primoz Fiser <primoz.fiser@norik.com>
Previously software based AES encryption was used with previously known
device specific keys (SBK), now that we have AES driver we can simply
delegate this to the engine without prior knowledge of the key (assuming
it is still loaded).
Signed-off-by: Ion Agorria <ion@agorria.com>
Now that we have working AES engine driver we can request the warmboot code
to be encrypted and signed with SBK if the device requires so. This
unlocks LP0 support for most devices in the wild as they use ODM Production
Secure.
We are not aware of any "ODM Production Open" device nor have access to
thus this has not been tested on one, merely added for completeness.
Signed-off-by: Ion Agorria <ion@agorria.com>
This driver allows using Tegra AES engines within BSEV and BSEA blocks to
encrypt and decrypt data using different AES algorithms.
One use case is allowing u-boot to self update by using the already loaded
AES key in the engine's SBK slot by the bootrom.
Particular care must be taken as chainloaded u-boot's may not have the SBK
slot loaded as the vendor bootloader erases it before leaving it.
Signed-off-by: Ion Agorria <ion@agorria.com>
DRAM init code, as per reverse engineering and matching against
previous SoCs. As usual no real documentation, and the DRAM controller
is the usual mixture of close-to-previous IP and new inventions.
This version supports LPDDR4 for now only, as seen on the early boards.
This needs improvements, but it can be done later.
Signed-off-by: Jernej Skrabec <jernej.skrabec@gmail.com>
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
This adds the early A523 clock setup code, for the basic peripheral PLL
and the basic bus clocks (APB/AHB). This is quite close to the existing
H6 and H616 clock code, so this shares the same file. A few bits and bobs
are different, though, so filter for the A523 in a few occasions.
Signed-off-by: Jernej Skrabec <jernej.skrabec@gmail.com>
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The cpu_sunxi_ncat2.h header file contains addresses of some peripherals
that are needed for the SPL, for chips that belong to the "NCAT2"
generation.
The Allwinner A523 is a member of this group, but a few addresses
differ, and we need a few more addresses, for playing with the core
reset, for instance.
Add the new addresses needed for the A523 and guard existing definitions
that conflict with that new chip.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The watchdog in the Allwinner A523 SoC differs a bit from the one in the
previous SoCs: it lives in a separate register frame, so no longer
inside some timer device, and it manages to shuffle around some
registers a bit. But it also conveniently adds a direct reset
functionality, so we don't need to use a dummy timeout period.
Avoid introducing a new MMIO register frame C struct, but just define
the one needed register offset as a macro. Then just trigger this new
direct reset functionality in the A523 specific reset_cpu()
implementation.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The Allwinner A523 features 8 CPU cores, organised in two clusters, both
driven by separate PLLs. Also there is the DSU PLL, which clocks the
hardware that connects the cores to the rest of the system.
And while the PLL registers itself are very similar, they are located in
a separate register frame, outside the main CCU, and also the register
controlling the CPU clock source (mux) is different.
Provide a separate function that reparents the two clusters and the DSU,
while their PLLs are programmed. For the actual PLL programming, we rely
on the existing shared routine.
The selection between the new A523 routine and the existing code is made
with C if statements, but since the choice is effectively made at compile
time already, the compiler optimises away the other code paths, leaving
just the one required function in.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The SPL initial clock setup code for the Allwinner H6 and H616 SoCs uses
a simple CPU PLL setup routine, which programs all register bits at once,
then waits for the LOCK bit to clear.
The manual suggests to follow a certain procedure for bringing up any
PLLs, which involves several register writes, one at a time, and some
delays. Also the H616 and the new A523 require some tiny changes in this
sequence, and the different SoCs also feature some extra bits here and
there, which we should not just clear.
So factor out the PLL setup routine, and make it follow the manual's
suggestion. This will read the PLL register at the beginning, then tweak
the bits we need to manipulate, and writes the register several times on
the way. This allows to cover the specific bits for different SoCs.
Besides improving the reliability of the PLL setup, this helps with the
A523, which requires *three* CPU PLLs to be programmed.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The Allwinner PLLs share most of their control bits, they differ mostly
in the factors and dividers.
Drop the PLL specific definition of those common bits, and use one
shared macro, for all PLLs.
This requires changing the users in the SPL clock and DRAM code.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
Since i.MX94, the ELE get_info structure is updated to add
OEM PQC SRK hash, so update it.
Signed-off-by: Ye Li <ye.li@nxp.com>
Signed-off-by: Jacky Bai <ping.bai@nxp.com>
Signed-off-by: Alice Guo <alice.guo@nxp.com>
Acked-by: Peng Fan <peng.fan@nxp.com>
With the removal of the last i.MX31 platform we can remove the rest of
the underlying architecture code as well.
Fixes: f247354708 ("arm: Remove mx31pdk board")
Signed-off-by: Tom Rini <trini@konsulko.com>
Acked-by: Peng Fan <peng.fan@nxp.com>
The current file defines a struct gpio_regs identical to the one in
<asm/mach-imx/gpio.h>. To eliminate code duplication and align with
the approach used for i.MX8M, include the common header instead of
redefining the struct.
Signed-off-by: Giulio Benetti <giulio.benetti@benettiengineering.com>
Ilias Apalodimas <ilias.apalodimas@linaro.org> says:
Instructions that lead ito an exception in the hypervisor can't modify two
CPU registers at once for the ARM ISA.
These instructions cannot be emulated by KVM as they do not produce
syndrome information data that KVM can use to infer the destination
register, the faulting address, whether it was a load or store, or
if it's a 32 or 64 bit general-purpose register.
As a result an external abort is injected from QEMU, via ext_dabt_pending.
Link: https://lore.kernel.org/r/20250618065828.1312146-1-ilias.apalodimas@linaro.org
commit 2e2c2a5e72 ("arm: qemu: override flash accessors to use virtualizable instructions")
explains why we can't have instructions with multiple output registers
when running under QEMU + KVM and the instruction leads to an exception
to the hypervisor.
USB XHCI is such a case (MMIO) where a ldr w1, [x0], #4 is emitted for
xhci_start() which works fine with QEMU but crashes for QEMU + KVM.
These instructions cannot be emulated by KVM as they do not produce
syndrome information data that KVM can use to infer the destination
register, the faulting address, whether it was a load or store, or
if it's a 32 or 64 bit general-purpose register.
As a result an external abort is injected from QEMU, via ext_dabt_pending
to KVM and we end up throwing an exception that looks like
U-Boot 2025.07-rc4 (Jun 10 2025 - 12:00:15 +0000)
[...]
Register 8001040 NbrPorts 8
Starting the controller
"Synchronous Abort" handler, esr 0x96000010, far 0x10100040
elr: 000000000005b1c8 lr : 000000000005b1ac (reloc)
elr: 00000000476fc1c8 lr : 00000000476fc1ac
x0 : 0000000010100040 x1 : 0000000000000001
x2 : 0000000000000000 x3 : 0000000000003e80
x4 : 0000000000000000 x5 : 00000000477a5694
x6 : 0000000000000038 x7 : 000000004666f360
x8 : 0000000000000000 x9 : 00000000ffffffd8
x10: 000000000000000d x11: 0000000000000006
x12: 0000000046560a78 x13: 0000000046560dd0
x14: 00000000ffffffff x15: 000000004666eed2
x16: 00000000476ee2f0 x17: 0000000000000000
x18: 0000000046660dd0 x19: 000000004666f480
x20: 0000000000000000 x21: 0000000010100040
x22: 0000000010100000 x23: 0000000000000000
x24: 0000000000000000 x25: 0000000000000000
x26: 0000000000000000 x27: 0000000000000000
x28: 0000000000000000 x29: 000000004666f360
Code: d5033fbf aa1503e0 5287d003 52800002 (b8004401)
Resetting CPU ...
There are two problems making this the default.
- It will emit ldr + add or str + add instead of ldr/str(post increment)
in somne cases
- Some platforms that depend on TPL/SPL grow in size enough so that the
binary doesn't fit anymore.
So let's add proper I/O accessors add a Kconfig option
to turn it off by default apart from our QEMU builds.
Reported-by: Mikko Rapeli <mikko.rapeli@linaro.org>
Tested-by: Mikko Rapeli <mikko.rapeli@linaro.org>
Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
The Allwinner A100 SoC has been around for a while, mostly on cheap
tablets, but didn't generate much interest in the community so far.
There were some efforts by two Allwinner employees in 2020, which led
to basic upstream Linux support for that SoC, although this momentum
dried up pretty quickly, leaving a lot of peripherals unsupported.
The A100 was silently replaced with the seemingly identical Allwinner
A133, which is reportedly a better bin of the A100. So far we assume
that both are compatible from a software perspective. There are some
more devices with the A133 out there now, so people are working on
filling the gaps, and adding U-Boot (and TF-A) support.
Based on the just added pinctrl, clock and DRAM support, this adds the
missing bits, mostly addresses and values for the SPL.
The A133 seems to be an predecessor to the H6, so we can share a lot of
code with that (and the H616 code), and just need to adjust some details.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
This adds preliminary support for the DRAM controller in the Allwinner
A100/A133 SoCs.
This is work in progress, and has rough edges, but works on at least
three different boards. It contains support for DDR4 and LPDDR4.
Signed-off-by: Cody Eksal <masterr3c0rd@epochal.quest>
[Andre: formatting fixes, adapt to mainline, drop unused parameters,
remove struct struct sunxi_mctl_com_reg, hardcode MR registers,
switch to mctl_check_pattern(), remove simple DRAM check]
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
E Shattow <e@freeshell.de> says:
Make consistent use of lowercase hexadecimal prefix '0x' throughout U-Boot.
There are a few remaining uses of uppercase 'X' to denote hexadecimal prefix
or placeholder in documentation and error messages.
External devicetree-rebasing dts/upstream and the generated code of
xilinx/zynq are ignored for the series.
Link: https://lore.kernel.org/r/20250606224558.1117422-1-e@freeshell.de
The global gd pointer is no longer volatile-qualified. Callers of this
helper function have been updated to no longer use volatile-qualifed
temporary variables, so update the prototype accordingly.
Signed-off-by: Rasmus Villemoes <ravi@prevas.dk>
There's a bunch of other places where this qualifier should be
dropped, e.g. in the set_gd() prototype and for various variables used
for stashing the value in the mach-imx/ directory and elsewhere. But
that will be done in follow-up patches.
Signed-off-by: Rasmus Villemoes <ravi@prevas.dk>
Tested-by: Anshul Dalal <anshuld@ti.com>
Add support for 8-bit CPU driven (primary and secondary) display signal
interface found in Tegra 2 and Tegra 3 SoC.
Tested-by: Ion Agorria <ion@agorria.com>
Signed-off-by: Svyatoslav Ryhel <clamor95@gmail.com>
Update GMAC speed and flow control fields in GRF_SOC_CON1 to use
RK3288_GMAC_* prefix, ensuring a consistent naming convention. It also
shifts each mask/bit definition to match the actual hardware bits, which
makes future usage easier.
Signed-off-by: Christoph Fritz <chf.fritz@googlemail.com>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
To be used for special-case oslog support in rtkit-helper.
Signed-off-by: Hector Martin <marcan@marcan.st>
Signed-off-by: Mark Kettenis <kettenis@openbsd.org>
This driver handles the MTP ASC coprocessor, which does not need any
special handling on the RTKit side and communicates out-of-band.
Signed-off-by: Hector Martin <marcan@marcan.st>
Signed-off-by: Mark Kettenis <kettenis@openbsd.org>
With the SPL clock code and the DRAM init routine we converted all users
of the H6 class "struct sunxi_prcm_reg" over to use #define'd register
offsets now.
Drop the whole definition of this struct now, since it's not needed
anymore, for all H6 and H616 boards.
This removes the entire fragile and questionable definition, and allows
new SoCs to share the code more easily.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The Allwinner H6 and H616 DRAM initialisation code uses a complex C
struct, modelling the PRCM clock register frame. For those SoCs, this
struct contains 20 registers, but the DRAM code only uses two of them.
Since we want to get rid of this struct, drop the usage of the struct in
the H6 and H616 DRAM code, by using #define'd register names and their
offset, and then adding those names to the base pointer.
This removes one more user of the PRCM clock register struct.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
U-Boot drivers often revert to using C structures for modelling hardware
register frames. This creates some problems:
- A "struct" is a C language construct to group several variables
together. The details of the layout of this struct are partly subject
to the compiler's discretion (padding and alignment).
- The "packed" attribute would force a certain layout, but we are not
using it.
- The actual source of information from the data sheet is the register
offset. Here we create an artificial struct, carefully tuning the
layout (with a lot of reserved members) to match that offset. To help
with correctness, we put the desired information as a *comment*,
though this is purely for the human reader, and has no effect on the
generated layout. This sounds all very backwards.
- Using a struct suggests we can assign a pointer and then access the
register content via the members. But this is not the case, instead
every MMIO register access must go through specific accessor functions,
to meet the ordering and access size guarantees the hardware requires.
- We share those structs in code shared across multiple SoC families,
though most SoCs define their own version of the struct. Members must
match in their name, across every SoC, otherwise compilation will fail.
We work around this with even more #ifdefs in the shared code.
- Some SoCs have an *almost* identical layout, but differ in a few
registers. This requires hard to maintain #ifdef's in the struct
definition.
- Some of the register frames are huge: the H6 CCU device defines 127
registers. We use 15 of them. Still the whole frame would need to be
described, which is very tedious, but for no reason.
- Adding a new SoC often forces people to decide whether to share an
existing struct, or to create a new copy. For some cases (say like 80%
similarity) this works out badly either way.
The Linux kernel heavily frowns upon those register structs, and instead
uses a much simpler solution: #define REG_NAME <offset>
This easily maps to the actual information from the data sheet, and can
much simpler be shared across multiple SoCs, as it allows to have all
SoC versions visible, so we can use C "if" statements instead of #ifdef's.
Also it requires to just define the registers we need, and we can use
alternative locations for some registers much more easily.
Drop the usage of "struct sunxi_prcm_reg" in the H6 SPL clock code, by
defining the respective register names and their offsets, then adding
them to the base pointer.
We cannot drop the struct definition quite yet, as it's also used in
other drivers, still.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
With the SPL clock code, the MMC driver, and the DRAM init routine we
converted all users of the H6 class "struct sunxi_ccm_reg" over to use
#define'd register offsets now.
Drop the whole definition of this struct now, since it's not needed
anymore, for all H6 and H616 boards.
This removes the entire fragile and questionable definition, and allows
new SoCs to share the code more easily.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The Allwinner H616 DRAM initialisation code uses a complex C struct,
modelling the clock device's register frame. For this SoC, the struct
contains 127 registers, but the DRAM code only uses four of them.
Since we want to get rid of this struct, drop the usage of the struct in
the H616 DRAM code, by using #define'd register names and their offset,
and then adding those names to the base pointer.
This removes one more user of the clock register struct.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The Allwinner MMC code uses a complex C struct, modelling the clock
device's register frame. We rely on sharing the member names across all
Allwinner SoCs, which is fragile.
Drop the usage of the struct in the MMC code, by using #define'd
register names and their offset, and then adding those names to the base
pointer. This requires to define those offsets for all SoCs, but since we
only use between four and six clock registers in the MMC code, this is
easily done.
This removes one common user of the clock register struct.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
U-Boot drivers often revert to using C structures for modelling hardware
register frames. This creates some problems:
- A "struct" is a C language construct to group several variables
together. The details of the layout of this struct are partly subject
to the compiler's discretion (padding and alignment).
- The "packed" attribute would force a certain layout, but we are not
using it.
- The actual source of information from the data sheet is the register
offset. Here we create an artificial struct, carefully tuning the
layout (with a lot of reserved members) to match that offset. To help
with correctness, we put the desired information as a *comment*,
though this is purely for the human reader, and has no effect on the
generated layout. This sounds all very backwards.
- Using a struct suggests we can assign a pointer and then access the
register content via the members. But this is not the case, instead
every MMIO register access must go through specific accessor functions,
to meet the ordering and access size guarantees the hardware requires.
- We share those structs in code shared across multiple SoC families,
though most SoCs define their own version of the struct. Members must
match in their name, across every SoC, otherwise compilation will fail.
We work around this with even more #ifdefs in the shared code.
- Some SoCs have an *almost* identical layout, but differ in a few
registers. This requires hard to maintain #ifdef's in the struct
definition.
- Some of the register frames are huge: the H6 CCU device defines 127
registers. We use 15 of them. Still the whole frame would need to be
described, which is very tedious, but for no reason.
- Adding a new SoC often forces people to decide whether to share an
existing struct, or to create a new copy. For some cases (say like 80%
similarity) this works out badly either way.
The Linux kernel heavily frowns upon those register structs, and instead
uses a much simpler solution: #define REG_NAME <offset>
This easily maps to the actual information from the data sheet, and can
much simpler be shared across multiple SoCs, as it allows to have all
SoC versions visible, so we can use C "if" statements instead of #ifdef's.
Also it requires to just define the registers we need, and we can use
alternative locations for some registers much more easily.
Drop the usage of "struct sunxi_ccm_reg" in the H6 SPL clock code, by
defining the respective register names and their offsets, then adding
them to the base pointer.
We cannot drop the struct definition quite yet, as it's also used in
other drivers, still.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
Thanks for Jernej's JTAG debugging effort, it turns out that the BROM
expects SP_IRQ to be saved and restored, when we want to enter back into
FEL after the SPL's AArch64 stint.
Save and restore SP_IRQ as part of the FEL state handling. The banked
MRS/MSR access to SP_IRQ, without actually being in IRQ mode, was
introduced with the ARMv7 virtualisation extensions. The Arm Cortex-A8
cores used in the A10/A13s or older F1C100s SoCs would not support that,
but this code here is purely in the ARMv8/AArch64 code path, so it's
safe to use unconditionally.
Reported-by: Jernej Skrabec <jernej.skrabec@gmail.com>
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
Reviewed-by: Jernej Skrabec <jernej.skrabec@gmail.com>
To be able to return to the BootROM FEL USB debug code, we must restore
the core's state as accurately as possible after the SPL has been run.
Since the BootROM runs in AArch32, but the SPL uses AArch64, this requires
a core reset, which clears the core's state.
So far we were saving and restoring the required registers like SCTLR
and VBAR, but could ignore the interrupt controller's state (GICC), since
that lives in MMIO registers, unaffected by a core reset.
Newer Allwinner SoCs now feature a GICv3 interrupt controller, which keeps
some GIC state in architected system registers, and those are cleared
when we switch back to AArch32.
To enable FEL operation on the Allwinner A523 SoC,
Add AArch32 assembly code to save and restore the ICC_PMR and ICC_IGRPEN1
system registers. The other GICv3 sysregs are either not relevant for the
BROM operation, or haven't been changed from their reset defaults by the
BROM anyway.
This enables FEL operation on the Allwinner A523 family of SoCs.
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
Reviewed-by: Jernej Skrabec <jernej.skrabec@gmail.com>
H616 rank and size detection code is superior to the H6. Nevertheless,
they are structurally the same. Split functions from H616 into new file
and reuse them in H6 DRAM driver too. This should also fix some bugs for
H6 too, like incorrect DRAM size detection.
Signed-off-by: Jernej Skrabec <jernej.skrabec@gmail.com>
[Andre: back out panic if test fails to allow 2^11 columns]
Reviewed-by: Andre Przywara <andre.przywara@arm.com>
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
This change is same as in commit 78aa00c38e ("sunxi: H616: dram: split
struct dram_para"), but for H6. This is needed in order to extract
common code between H6 and H616 later.
Signed-off-by: Jernej Skrabec <jernej.skrabec@gmail.com>
Reviewed-by: Andre Przywara <andre.przywara@arm.com>
