* changes:
rpi4: Add initial documentation file
rpi4: Add stdout-path to device tree
rpi4: Add GIC maintenance interrupt to GIC DT node
rpi4: Cleanup memory regions, move pens to first page
rpi4: Reserve resident BL31 region from non-secure world
rpi4: Amend DTB to advertise PSCI
rpi4: Determine BL33 entry point at runtime
rpi4: Accommodate "armstub8.bin" header at the beginning of BL31 image
Add basic support for Raspberry Pi 4
rpi3: Allow runtime determination of UART base clock rate
FDT helper functions: Respect architecture in PSCI function IDs
FDT helper functions: Add function documentation
* changes:
Migrate ARM platforms to use the new GICv3 API
Adding new optional PSCI hook pwr_domain_on_finish_late
GICv3: Enable multi socket GIC redistributor frame discovery
This patch invokes the new function gicv3_rdistif_probe() in the
ARM platform specific gicv3 driver. Since this API modifies the
shared GIC related data structure, it must be invoked coherently
by using the platform specific pwr_domain_on_finish_late hook.
Change-Id: I6efb17d5da61545a1c5a6641b8f58472b31e62a8
Signed-off-by: Madhukar Pappireddy <madhukar.pappireddy@arm.com>
Some device tree users like to find a pointer to the standard serial
console in the device tree, in the "stdout-path" property of the /chosen
node.
Add the location of the Mini UART in that property, so that DT users are
happy, for instance Linux' earlycon detection.
Change-Id: I178e55016e5640de5ab0bc6e061944bd3583ea96
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
For being able to use the virtualisation support the GIC offers, we need
to know the interrupt number of the maintenance interrupt. This
information is missing from the official RPi4 device tree.
Use libfdt to add the "interrupts" property to the GIC node, which
allows hypervisors like KVM or Xen to be able to use the GIC's help on
virtualising interrupts.
Change-Id: Iab84f0885a5bf29fb84ca8f385e8a39d27700c75
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
Now that we have the SMP pens in the first page of DRAM, we can get rid
of all the fancy RPi3 memory regions that our RPi4 port does not really
need. This avoids using up memory all over the place, restricting ATF
to just run in the first 512KB of DRAM.
Remove the now unused regions. This also moves the SMP pens into our
first memory page (holding the firmware magic), where the original
firmware put them, but where there is also enough space for them.
Since the pens will require code execution privileges, we amend the
memory attributes used for that page to include write and execution
rights.
Change-Id: I131633abeb4a4d7b9057e737b9b0d163b73e47c6
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The GPU firmware loads the armstub8.bin (BL31) image at address 0, the
beginning of DRAM. As this holds the resident PSCI code and the SMP
pens, the non-secure world should better know about this, to avoid
accessing memory owned by TF-A. This is particularly criticial as the
Raspberry Pi 4 does not feature a secure memory controller, so
overwriting code is a very real danger.
Use the newly introduced function to add a node into reserved-memory
node, where non-secure world can check for regions to be excluded from
its mappings.
Reserve the first 512KB of memory for now. We can refine this later if
need be.
Change-Id: I00e55e70c5c02615320d79ff35bc32b805d30770
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The device tree provided by the official Raspberry Pi firmware uses
spin tables for SMP bringup.
One of the benefit of having TF-A is that it provides PSCI services, so
let's rewrite the DTB to advertise PSCI instead of spin tables.
This uses the (newly exported) routine from the QEMU platform port.
Change-Id: Ifddcb14041ca253a333f8c2d5e97a42db152470c
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
Now that we have the armstub magic value in place, the GPU firmware will
write the kernel load address (and DTB address) into our special page,
so we can always easily access the actual location without hardcoding
any addresses into the BL31 image.
Make the compile-time defined PRELOADED_BL33_BASE macro optional, and
read the BL33 entry point from the magic location, if the macro was not
defined. We do the same for the DTB address.
This also splits the currently "common" definition of
plat_get_ns_image_entrypoint() to be separate between RPi3 and RPi4.
Change-Id: I6f26c0adc6fce2df47786b271c490928b4529abb
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The Raspberry Pi GPU firmware checks for a magic value at offset 240
(0xf0) of the armstub8.bin image it loads. If that value matches,
it writes the kernel load address and the DTB address into subsequent
memory locations.
We can use these addresses to avoid hardcoding these values into the BL31
image, to make it more flexible and a drop-in replacement for the
official armstub8.bin.
Reserving just 16 bytes at offset 240 of the final image file is not easily
possible, though, as this location is in the middle of the generic BL31
entry point code.
However we can prepend an extra section before the actual BL31 image, to
contain the magic and addresses. This needs to be 4KB, because the
actual BL31 entry point needs to be page aligned.
Use the platform linker script hook that the generic code provides, to
add an almost empty 4KB code block before the entry point code. The very
first word contains a branch instruction to jump over this page, into
the actual entry code.
This also gives us plenty of room for the SMP pens later.
Change-Id: I38caa5e7195fa39cbef8600933a03d86f09263d6
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The Raspberry Pi 4 is a single board computer with four Cortex-A72
cores. From a TF-A perspective it is quite similar to the Raspberry Pi
3, although it comes with more memory (up to 4GB) and has a GIC.
This initial port though differs quite a lot from the existing rpi3
platform port, mainly due to taking a much simpler and more robust
approach to loading the non-secure payload:
The GPU firmware of the SoC, which is responsible for initial platform
setup (including DRAM initialisation), already loads the kernel, device
tree and the "armstub" into DRAM. We take advantage of this, by placing
just a BL31 component into the armstub8.bin component, which will be
executed first, in AArch64 EL3.
The non-secure payload can be a kernel or a boot loader (U-Boot or
EDK-2), disguised as the "kernel" image and loaded by the GPU firmware.
So this is just a BL31-only port, which directly drops into EL2
and executes whatever has been loaded as the "kernel" image, handing
over the DTB address in x0.
Change-Id: I636f4d1f661821566ad9e341d69ba36f6bbfb546
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
At the moment the UART input clock rate is hard coded at compile time.
This works as long as the GPU firmware always sets up the same rate,
which does not seem to be true for the Raspberry Pi 4.
In preparation for being able to change this at runtime, add a base
clock parameter to the console setup function. This is still hardcoded
for the Raspberry Pi 3.
Change-Id: I398bc2f1e9b46f7af9a84cb0b33cbe8e78f2d900
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The Fast Models provide a non-volatile counter component, which is used
in the Trusted Board Boot implementation to protect against rollback
attacks.
This component comes in 2 versions (see [1]).
- Version 0 is the default and models a locked non-volatile counter,
whose value is fixed.
- Version 1 of the counter may be incremented in a monotonic fashion.
plat_set_nv_ctr() must cope with both versions. This is achieved by:
1) Attempting to write the new value in the counter.
2) Reading the value back.
3) If there is a mismatch, we know the counter upgrade failed.
When using version 0 of the counter, no upgrade is possible so the
function is expected to fail all the time. However, the code is
missing a compiler barrier between the write operation and the next
read. Thus, the compiler may optimize and remove the read operation on
the basis that the counter value has not changed. With the default
optimization level used in TF-A (-Os), this is what's happening.
The fix introduced in this patch marks the write and subsequent read
accesses to the counter as volatile, such that the compiler makes no
assumption about the value of the counter.
Note that the comment above plat_set_nv_ctr() was clearly stating
that when using the read-only version of the non-volatile counter,
"we expect the values in the certificates to always match the RO
values so that this function is never called". However, the fact that
the counter value was read back seems to contradict this comment, as
it is implementing a counter-measure against misuse of the
function. The comment has been reworded to avoid any confusion.
Without this patch, this bug may be demonstrated on the Base AEM FVP:
- Using version 0 of the non-volatile counter (default version).
- With certificates embedding a revision number value of 32
(compiling TF-A with TFW_NVCTR_VAL=32).
In this configuration, the non-volatile counter is tied to value 31 by
default. When BL1 loads the Trusted Boot Firmware certificate, it
notices that the two values do not match and tries to upgrade the
non-volatile counter. This write operation is expected to fail
(because the counter is locked) and the function is expected to return
an error but it succeeds instead.
As a result, the trusted boot does not abort as soon as it should and
incorrectly boots BL2. The boot is finally aborted when BL2 verifies
the BL31 image and figures out that the version of the SoC Firmware
Key Certificate does not match. On Arm platforms, only certificates
signed with the Root-of-Trust Key may trigger an upgrade of the
non-volatile Trusted counter.
[1] https://developer.arm.com/docs/100964/1160/fast-models-components/peripheral-components/nonvolatilecounter
Change-Id: I9979f29c23b47b338b9b484013d1fb86c59db92f
Signed-off-by: Sandrine Bailleux <sandrine.bailleux@arm.com>
Rockchip platform is using the first 1MB of DRAM as secure ram space,
and there is a vendor loader who loads and runs the BL31/BL32/BL33,
this loader is usually load by SoC BootRom to the start addres of DRAM,
we need to reserve enough space for this loader so that it doesn't need
to do the relocate when loading the BL31. eg.
We use U-Boot SPL to load ATF BL31 and U-Boot proper as BL33, the SPL
TEXT BASE is offset 0 of DRAM which is decide by Bootrom; if we update
the BL31_BASE to offset 0x40000(256KB), then the 0~0x40000 should be
enough for SPL and no need to do the relocate while the space size
0x10000(64KB) may not enough for SPL.
After this update, the BL31 can use the rest 768KB of the first 1MB,
which is also enough, and the loader who is using BL31 elf file can
support this update without any change.
Change-Id: I66dc685594d77f10f9a49c3be015fd6729250ece
Signed-off-by: Kever Yang <kever.yang@rock-chips.com>
Patch introduce the macro NS_IMAGE_MAX_SIZE to simplify the image size
calculation. Use of additional parenthesis removes the possibility of
improper calculations due nested macro expansion for subtraction.
In case of platforms with DRAM window over 32bits, patch also removes
potential problems with type casting, as meminfo.image_size is uint32_t
but macro calculations were done in 64bit space.
Signed-off-by: Radoslaw Biernacki <radoslaw.biernacki@linaro.org>
Change-Id: I2d05a2d9dd6000dba6114df53262995cf85af018
This commit change the plat/qemu directory structure into:
`-- plat
`-- qemu
|-- common (files shared with all qemu subplatforms)
|-- qemu (original qemu platform)
|-- qemu_sbsa (new sqemu_sbsa platform)
|-- subplat1
`-- subplat2
This opens the possibility of adding new qemu sub-platforms which reuse
existing common platform code. The first platform which will leverage new
structure will be SBSA platform.
Signed-off-by: Radoslaw Biernacki <radoslaw.biernacki@linaro.org>
Signed-off-by: Sandrine Bailleux <sandrine.bailleux@arm.com>
Change-Id: Id0d8133e1fffc1b574b69aa2770ebc02bb837a9b
Avoid to load FIP by hacking address. Load it by partition table instead.
Signed-off-by: Haojian Zhuang <haojian.zhuang@linaro.org>
Change-Id: I0283fc2e6e459bff14de19d92db4158e05106ee4
Avoid to load FIP by hacking address. Load it by partition table instead.
Signed-off-by: Haojian Zhuang <haojian.zhuang@linaro.org>
Change-Id: Ib476d024a51e4b9705441a0007d78f9fdf0ca078
Both kernel and U-Boot use a SMC call to the secure monitor to get the
chip ID. This call is translated by BL31 to a call to the SCP to
retrieve the ID. Add a new SiP call and the backing SCPI command.
Signed-off-by: Carlo Caione <ccaione@baylibre.com>
Change-Id: Ib128f5645ee92866e7ebbcd550dacd33f573524b
* changes:
rpi3: Do prescaler and control setup in C
rpi3: Prepare for supporting a GIC (in RPi4)
rpi3: Make SHARED_RAM optional
rpi3: Rename RPI3_IO_BASE to RPI_IO_BASE
rpi3: Move shared rpi3 files into common directory
To initialise the arch timer configuration and some clock prescaler, we
need to do two MMIO access *once*, early during boot.
As tempting as it may sound, plat_reset_handler() is not the right place
to do this, as it will be called on every CPU coming up, both for
secondary cores as well as during warmboots. So this access will be done
multiple times, and even during a rich OS' runtime. Whether doing so anyway
is actually harmful is hard to say, but we should definitely avoid this if
possible.
Move the initialisation of these registers to C code in
bl1_early_platform_setup(), where it will still be executed early enough
(before enabling the console), but only once during the whole boot
process.
Change-Id: I081c41a5476d424411411488ff8f633e87d3bcc5
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
To allow sharing the driver between the RPi3 and RPi4, move the random
number generator driver into the generic driver directory.
Change-Id: Iae94d7cb22c6bce3af9bff709d76d4caf87b14d1
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
With the incoming support for the Raspberry Pi 4 boards, one directory
to serve both versions will not end up well.
Create an additional layer by inserting a "rpi" directory betweeen /plat
and rpi3, so that we can more easily share or separate files between the
two later.
Change-Id: I75adbb054fe7902f34db0fd5e579a55612dd8a5f
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
As the PSCI "power" management functions for the Raspberry Pi 3 port
will be shared with the upcoming RPi4 support, we need to prepare them
for dealing with the GIC interrupt controller.
Splitting this code just for those simple calls to the generic GIC
routines does not seem worthwhile, so just use a #define the protect the
GIC code from being included by the existing RPi3 code.
Change-Id: Iaca6b0214563852b28ad4a088ec45348ae8be40d
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The QEMU platform port scans its device tree to advertise PSCI as the
CPU enable method. It does this by scanning *every* node in the DT and
check whether its compatible string starts with "arm,cortex-a". Then it
sets the enable-method to PSCI, if it doesn't already have one.
Other platforms might want to use this functionality as well, so let's
move it out of the QEMU platform directory and make it more robust by
fixing some shortcomings:
- A compatible string starting with a certain prefix is not a good way
to find the CPU nodes. For instance a "arm,cortex-a72-pmu" node will
match as well and is in turn favoured with an enable-method.
- If the DT already has an enable-method, we won't change this to PSCI.
Those two issues will for instance fail on the Raspberry Pi 4 DT.
To fix those problems, we adjust the scanning method:
The DT spec says that all CPU nodes are subnodes of the mandatory
/cpus node, which is a subnode of the root node. Also each CPU node has
to have a device_type = "cpu" property. So we find the /cpus node, then
scan for a subnode with the proper device_type, forcing the
enable-method to "psci".
We have to restart this search after a property has been patched, as the
node offsets might have changed meanwhile.
This allows this routine to be reused for the Raspberry Pi 4 later.
Change-Id: I00cae16cc923d9f8bb96a9b2a2933b9a79b06139
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
To allow sharing the driver between the RPi3 and RPi4, move the mailbox
driver into the generic driver directory.
Change-Id: I463e49acf82b02bf004f3d56482b7791f3020bc0
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The existing Raspberry Pi 3 port sports a number of memory regions,
which are used for several purposes. The upcoming RPi4 port will not use
all of those, so make the SHARED_RAM region optional, by only mapping it
if it has actually been defined. This helps to get a cleaner RPi4 port.
Change-Id: Id69677b7fb6ed48d9f238854b610896785db8cab
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
With the advent of Raspberry Pi 4 support, we need to separate some
board specific headers between the RPi3 and RPi4.
Rename and move the "rpi3_hw.h" header, so that .c files just include
rpi_hw.h, and automatically get the correct version.
Change-Id: I03b39063028d2bee1429bffccde71dddfe2dcde8
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
The location of the MMIO window is different between a Raspberry Pi 3
and 4: the former has it just below 1GB, the latter below 4GB.
The relative location of the peripherals is mostly compatible though.
To allow sharing code between the two models, let's rename the symbol
used for the MMIO base to the more generic RPI_IO_BASE name.
Change-Id: I3c2762fb30fd56cca743348e79d72ef8c60ddb03
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
To be able to share code more easily between the existing Raspberry Pi 3
and the upcoming Raspberry Pi 4 platform, move some code which is not
board specific into a "common" directory.
Change-Id: I9211ab2d754b040128fac13c2f0a30a5cc8c7f2c
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
This patch provides the following features and makes modifications
listed below:
- Individual APIAKey key generation for each CPU.
- New key generation on every BL31 warm boot and TSP CPU On event.
- Per-CPU storage of APIAKey added in percpu_data[]
of cpu_data structure.
- `plat_init_apiakey()` function replaced with `plat_init_apkey()`
which returns 128-bit value and uses Generic timer physical counter
value to increase the randomness of the generated key.
The new function can be used for generation of all ARMv8.3-PAuth keys
- ARMv8.3-PAuth specific code placed in `lib\extensions\pauth`.
- New `pauth_init_enable_el1()` and `pauth_init_enable_el3()` functions
generate, program and enable APIAKey_EL1 for EL1 and EL3 respectively;
pauth_disable_el1()` and `pauth_disable_el3()` functions disable
PAuth for EL1 and EL3 respectively;
`pauth_load_bl31_apiakey()` loads saved per-CPU APIAKey_EL1 from
cpu-data structure.
- Combined `save_gp_pauth_registers()` function replaces calls to
`save_gp_registers()` and `pauth_context_save()`;
`restore_gp_pauth_registers()` replaces `pauth_context_restore()`
and `restore_gp_registers()` calls.
- `restore_gp_registers_eret()` function removed with corresponding
code placed in `el3_exit()`.
- Fixed the issue when `pauth_t pauth_ctx` structure allocated space
for 12 uint64_t PAuth registers instead of 10 by removal of macro
CTX_PACGAKEY_END from `include/lib/el3_runtime/aarch64/context.h`
and assigning its value to CTX_PAUTH_REGS_END.
- Use of MODE_SP_ELX and MODE_SP_EL0 macro definitions
in `msr spsel` instruction instead of hard-coded values.
- Changes in documentation related to ARMv8.3-PAuth and ARMv8.5-BTI.
Change-Id: Id18b81cc46f52a783a7e6a09b9f149b6ce803211
Signed-off-by: Alexei Fedorov <Alexei.Fedorov@arm.com>