EFI applications need to be relocatable. Ordinarily, this is achieved
through a PE-format .reloc section, but since that requires toolchain
tricks to achieve, U-Boot's EFI applications instead embed ELF-flavored
relocation information and use it for self-relocation; thus, the
.dynamic section needs to be preserved.
Before this patch, it was tacked on to the end of .text, but this was
not proper: A .text section is SHT_PROGBITS, while the .dynamic section
is SHT_DYNAMIC. Attempting to combine them like this creates a section
type mismatch. While GNU ld doesn't seem to complain, LLVM's lld
considers this a fatal linking error.
This patch moves .dynamic out to its own section, so that the output ELF
has the correct types. (They're all mashed together when converting to
binary anyway, so this patch causes no change in the final .efi output.)
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
Cc: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Reviewed-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
While the _start label is only intended for use locally to populate the
(hand-written) PE header, the linker script includes ENTRY(_start) which
designates it as the entry point in the output ELF, resulting in linker
warnings under some linkers (e.g. LLVM's lld) due to _start not being a
globally-visible symbol. Since ELF is only an intermediary build
format, and the aforementioned PE header correctly points to _start, the
ENTRY(_start) directive could easily be removed to silence this warning.
However, since some developers who are debugging EFI by analyzing the
intermediary ELF may appreciate having correct entry-point information,
this patch instead promotes the _start labels to global symbols,
silencing the linker warning and making the intermediary ELF reflect the
true entry point.
This patch doesn't affect the final output binaries in any way.
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
Reviewed-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
This allows setting READELF=llvm-readelf in order to use the LLVM
version of the readelf utility. It also aligns with the practice of not
using $(CROSS_COMPILE) in any build recipes directly, reducing the
number of places where $(CROSS_COMPILE) is used.
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
Reviewed-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
RELRO is an instruction to a dynamic loader to make a memory range
read-only after relocations are applied, for added security. Some
linkers (e.g. LLD) require that all sections covered by the RELRO are
contiguous, so that only a single RELRO is needed. U-Boot at present
neither satisfies this requirement (e.g. x86_64 linker script currently
puts .dynamic too far from .got) nor preserves the RELRO when converting
away from ELF, therefore add `-z norelro` to global linker options.
This can be brought back in the future when the linker scripts are
cleaned up and U-Boot understands RELROs.
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
This flag only makes sense for `binary` output, because .hex/.srec are
sparse formats and represent gaps without filler. While the GNU binutils
version of objcopy does not seem to mind the extra flag being passed,
llvm-objcopy considers this a fatal error.
There is already a version of the objcopy command template in the
Makefile that doesn't use --gap-fill, which is provided for EFI. So use
this other version for all .hex/.srec outputs as well.
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
Acked-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
When a section is not flagged with SHF_ALLOC, LLD's --gc-sections
algorithm fails to visit the sections that it references. As a result of
this, LLD was dropping the call64.o(.data) section, which is itself only
referenced by .text_call64.
This appears to be a bug in LLD, but the .section directive for
.text_call64 should really have the correct flags either way.
Add `"ax"` to mark the section as ALLOC ("supposed to be loaded") and
CODE ("supposed to be executed").
Fixes: 7dc82591d6 ("x86: Move call64 into its own section")
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
LLVM's IAS does not (and cannot easily) support the 'adrl'
pseudoinstruction, and ARM developers generally do not consider it
portable across assembler implementations either.
Instead, expand it into the two subtract instructions it would emit
anyway. An explanation of the math follows:
The .+8 and .+4 refer to the same memory location; this is because the
.+4 expression occurs in a subsequent instruction, 4 bytes after the
first. This memory location is the value of the PC register when it is
read by the first sub instruction. Thus, both inner parenthesized
expressions evaluate to the same result: PC's offset relative to
image_base. The subtract instructions then remove one byte each
(low, then high) of the total offset, thereby getting the absolute
address of image_base loaded in r0.
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
There are a handful of sections that are not useful in the U-Boot output
binary. At present, the linker script moves these to the end of the
binary, after the _image_binary_end marker symbol, so that they don't
get loaded.
The linker script syntax supports discarding sections that shouldn't be
included in the output. Switch to this instead, to make the intention
clearer and reduce the ELF sections that have to be handled later in the
build. This is also consistent with the other architectures' linker
scripts.
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
These are sometimes used by LLVM's code-generator, when it can guarantee that
the memory buffer being passed is aligned on a (4- or 8-byte) boundary. They
can safely be aliased to the unaligned versions.
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
LLVM's code generator will sometimes emit calls to __aeabi_memclr. Add an
implementation of this for LLVM compatibility.
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
These symbols need to survive the IR-level dead function elimination pass,
since nothing at the IR level is referencing them (calls to these are inserted
later, at codegen time).
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
Acked-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
The .mmutable section was deprecated in 2012 [1] and finally removed
entirely from U-Boot in 2022 [2], so this special handling is no longer
necessary. Remove it to tidy up the linker script.
[1]: dde3b70dcf ("arm: add a common .lds link script")
[2]: 3135ba642f ("arm: pxa: Remove CONFIG_CPU_PXA25X")
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
On STM32MP15xx with 1 GiB of DRAM, the gd->ram_top becomes 0,
because DRAM base 0xc0000000 + DRAM size 0x40000000 leads to
gd->ram_top overflow which resets it to 0. Handle this special
case simply by checking for gd->ram_top being zero, and if it
is, assume there is no addr >= gd->ram_top .
This fixes boot hang on STM32MP15xx with 1 GiB of DRAM.
Fixes: 25fb58e88a ("ARM: stm32mp: Fix dram_bank_mmu_setup() for LMB located above ram_top")
Signed-off-by: Marek Vasut <marex@denx.de>
Reviewed-by: Patrice Chotard <patrice.chotard@foss.st.com>
A problem we have today is that some instances of IS_ENABLED(FOO) have
crept in to the code. This is in turn because with checkpatch.pl this is
only a warning and not an error, so they were overlooked. And looking
deeper, in the Linux kernel this pattern is allowed because
IS_ENABLED(DEFINED_FLAG) is allowed if discouraged and a quick skim of
the instances I saw in the current kernel follow this pattern. In U-Boot
however, this is not allowed, so bump to an error.
Signed-off-by: Tom Rini <trini@konsulko.com>
In case MAX_SYMLINK_NEST is reached while determining the size
on a symlink node, the function returns immediately.
This would not free the resources after the free_strings: label
causing a memory leak.
Set the ret value and just break out of the switch to fix this.
Signed-off-by: Andrea della Porta <andrea.porta@suse.com>
Reviewed-by: Miquel Raynal <miquel.raynal@bootlin.com>
The GPL-2.1+ SPDX License Identifier doesn't exist, but luckily the full
license text was available which shows that GPL-2.1+ was a typo and it
should have been LGPL-2.1-or-later.
As the '+' in LGPL-2.1+ is deprecated in SPDX 3.0, use the preferred
identifier, which is '-or-later'.
Normally the full license header is removed when switching to SPDX
License Identifiers, so do that now.
Signed-off-by: Diederik de Haas <didi.debian@cknow.org>
AMD/Xilinx changes for v2025.04-rc4
Zynq:
- Guard code around SPL_FS_LOAD_PAYLOAD_NAME
Versal*:
- Remove tftp block size 4096
Versal:
- Use clocks per DT binding
- Store driver data in data section
Versal Gen 2:
- Fix major/minor version decoding
Currently, Kconfig allows building CONFIG_DM_74X164 without
CONFIG_DM_SPI, which results in linker errors because this driver
actually uses dm_spi_* functions:
drivers/gpio/74x164_gpio.o: in function `gen_74x164_write_conf':
undefined reference to `dm_spi_claim_bus'
undefined reference to `dm_spi_xfer'
undefined reference to `dm_spi_release_bus'
Signed-off-by: J. Neuschäfer <j.ne@posteo.net>
With a recent Binman change, the skip-at-start property is now honoured,
meaning that all image-pos values in the affected section start from
the skip-at-start value.
The x86 code works around the old behaviour at present, so update it.
Signed-off-by: Simon Glass <sjg@chromium.org>
A discussion on the mailing list about dealing with block offsets and
binman symbols made me think that something is wrong with how Binman
deals with the skip-at-start property.
The feature was originally designed to handle x86 ROMs, which are mapped
at the top of the address space. That seemed too specific, whereas
skipping some space at the start seemed more generally useful.
It has proved useful. For example, rockchip images start at block 64,
so a skip-at-start of 0x8000 deals with this.
But it doesn't actually work correctly, since the image_pos value does
not give the actual position on the media.
Fix this and update the documentation, moving it into the 'section'
section.
Signed-off-by: Simon Glass <sjg@chromium.org>
Snow requires a fixed bloblist to operate, so re-enable this option.
Signed-off-by: Simon Glass <sjg@chromium.org>
Fixes: 864106f3c4 ("bloblist: Make BLOBLIST_ALLOC the default")
Unfortunately this change was not safe as some devices are bound before
relocation, but we don't want to probe them.
It causes 'raise: Signal # 8 caught' on jerry.
Move the bootstage timer to after autoprobe in initf_dm() since the
trace test does not tolerate any variance.
This reverts commit 21dd873572.
Signed-off-by: Simon Glass <sjg@chromium.org>
The commit bc07851897 ("board: ti: Pull redundant DDR functions to a
common location and Fixup DDR size when ECC is enabled") broke DRAM
support for the Verdin AM62. This was partially fixed with commit
3f866c47b5 ("board: verdin-am62: add dram_init_banksize"). However,
because fixup_memory_node was not called, the Linux kernel was started
with the wrong memory size on modules with less memory available. This
resulted in boot failures. Fix this issue by calling fixup_memory_node
in the board file.
spl_perform_fixups will be called in the SPL and now sets the correct
memory size in the device tree of U-Boot by calling fixup_memory_node.
U-Boot will then adjust the memory sizes of Linux during bootm/booti in
fdt_fixup_memory_banks. This chain ensures that U-Boot and Linux only
use RAM that is actually available.
Fixes: 3f866c47b5 ("board: verdin-am62: add dram_init_banksize")
Fixes: bc07851897 ("board: ti: Pull redundant DDR functions to a common location and Fixup DDR size when ECC is enabled")
Signed-off-by: Stefan Eichenberger <stefan.eichenberger@toradex.com>
Acked-by: Francesco Dolcini <francesco.dolcini@toradex.com>
When enabling net console and console multiplexing, a boot crash was
observed using mtk_eth driver with stdin/stdout set to "serial,nc"
in persistent environment:
> CPU: MediaTek MT7981
> Model: OpenWrt One
> DRAM: 1 GiB
> Core: 35 devices, 15 uclasses, devicetree: separate
> spi-nand: spi_nand spi_nand@0: Winbond SPI NAND was found.
> spi-nand: spi_nand spi_nand@0: 128 MiB, block size: 128 KiB, page size: 2048, OOB size: 64
> Loading Environment from UBI... SF: Detected w25q128 with page size 256 Bytes, erase size 4 KiB, total 16 MiB
> mtd: partition "ubi" extends beyond the end of device "spi-nand0" -- size truncated to 0x7f00000
> Read 126976 bytes from volume ubootenv to 000000007f7bf0c0
> Read 126976 bytes from volume ubootenv2 to 000000007f7de100
> OK
> "Synchronous Abort" handler, esr 0x96000004, far 0xeafffffeea000018
> elr: 0000000041e63cd4 lr : 0000000041e1b844 (reloc)
> elr: 000000007ff9ecd4 lr : 000000007ff56844
> x0 : eafffffeea000018 x1 : 000000007fb552e0
> x2 : 00000000000000fe x3 : 0000000000000000
The cause is that "serial,nc" forced the console subsystem to
initialize the ethernet driver before ethernet subsystem
initialization (console_init_r() is called before initr_net()).
During the mtk_eth driver initialization, mdio_register() will be
called, and miiphy_get_dev_by_name() will then be called.
The miiphy_get_dev_by_name() will check the list "mii_devs" to see
if the passed device name exists. However the mii_devs is defined
without initialization:
> static struct list_head mii_devs;
and the actual initialization is done in the following chain:
initr_net -> eth_initialize -> eth_common_init -> miiphy_init
Since initr_net() hasn't be called, iterating over the mii_devs
will access to physical address 0 (mii_devs.next == NULL) and will
cause the crash.
The fix is to define mii_devs using:
> static LIST_HEAD(mii_devs);
As the "current_mii" is defined as a static variable, it will
always be NULL in board_r stage and initializing it will NULL is
unnecessary. So the entire miiphy_init() can be remove.
Signed-off-by: Weijie Gao <hackpascal@gmail.com>
My previous address is no longer accessible, but I will continue to be
involved in maintaining the ADI sc5xx platforms. This updates my contact
information and hopefully avoids bouncing emails from other developers.
Signed-off-by: Greg Malysa <malysagreg@gmail.com>
Line 171 in README is describing that before relocation
no code should use global variable because global variables
are placed to BSS section which is initialized to 0 after
relocation.
On Versal platforms clock driver is initialized before
relocation (via using dm,bootph-all flag in DT) and global
variables are initialized which works if this is used only
before relocation. But the variables are used after
relocation too but values are zeroed which is ending up
incorrect behavior.
That's why place variables to data section to ensure that
values are not cleared which is for now the quickest
temporary solution. The correct way to do it is to move
all global variables to private data to avoid it.
Signed-off-by: Padmarao Begari <padmarao.begari@amd.com>
Link: https://lore.kernel.org/r/20250218052419.1141139-1-padmarao.begari@amd.com
Signed-off-by: Michal Simek <michal.simek@amd.com>
Fix build error "undefined reference to `is_warm_boot'" when
ls1021atsn and ls1021atwr boards are built with CONFIG_SPL=y and
CONFIG_DEEP_SLEEP=n.
Signed-off-by: Benjamin Lemouzy <blemouzy@centralp.fr>
Signed-off-by: Peng Fan <peng.fan@nxp.com>
When an error is detected in the TSEC receive path, the driver currently
prints an error message, but leaves the corresponding packet descriptor
in its old state (i.e. owned by the CPU side). As a result, the packet
queue can be starved of available buffers if enough errors happen.
To recover from errors, re-arm the packet buffer descriptor after an
error has been detected.
Errors can be provoked by changing a PHY with phy-mode = "rgmii-id" to
phy-mode = "rgmii".
Signed-off-by: J. Neuschäfer <j.ne@posteo.net>
Signed-off-by: Peng Fan <peng.fan@nxp.com>
This is necessary for the following patch. No functional change.
Signed-off-by: J. Neuschäfer <j.ne@posteo.net>
Signed-off-by: Peng Fan <peng.fan@nxp.com>
Check wget_info->buffer_size for overflow and do not clean the wget_info struct
on failure, let the owner of the struct handle the error. The latter is necesary
, e.g., for when a request fails because the provided buffer was too small.
Signed-off-by: Adriano Cordova <adriano.cordova@canonical.com>
This commit fixes an use after free introduced in Commit e55a4acb54
(" efi_loader: net: set EFI bootdevice device path to HTTP when loaded
from wget"). The logic in efi_net_set_dp is reworked so that when the
function is invoked it not only changes the value of the static variable
net_dp (this is how the function was implemented in e55a4acb54) but also
updates the protocol interface of the device path protocol in case efi
has started.
Fixes: e55a4acb54 ("efi_loader: net: set EFI bootdevice device path to HTTP when loaded from wget")
Signed-off-by: Adriano Cordova <adriano.cordova@canonical.com>
This board does not have any NVMe slot, so disable unneeded
CONFIG_NVME_PCI option.
Fixes: bf4a33e725 ("board: rockchip: add FriendlyElec NanoPi R3S")
Signed-off-by: Tianling Shen <cnsztl@gmail.com>
Reviewed-by: Quentin Schulz <quentin.schulz@cherry.de>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
The patch enables the second USB3.0 Type-A USB port on the NanoPi R4S
board, which prevents a crash when initializing the usb system in U-Boot
and allows both Type-A USB ports to be used for booting.
=> usb start
starting USB...
Bus usb@fe380000: USB EHCI 1.00
Bus usb@fe3c0000: "Synchronous Abort" handler, esr 0x96000010, far 0x0
Signed-off-by: Justin Klaassen <justin@tidylabs.net>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
These two files were using IS_ENABLED() to test for CONFIG flags but omitted
the CONFIG_ prefix and so did not work as expected.
Signed-off-by: Tom Rini <trini@konsulko.com>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
BootROM leave GPIO4_D6 configured as SDMMC_PWREN function and DW MMC
driver set PWREN high in dwmci_init().
However, HW revision prior to v1.2 must pull GPIO4_D6 low to access
sdmmc. For HW revision v1.2 the state of GPIO4_D6 has no impact.
Upstream Linux commit 26c100232b09 "arm64: dts: rockchip: Fix sdmmc
access on rk3308-rock-s0 v1.1 boards" fixed this issue by adding a
vcc_sd regulator.
Include the new vcc_sd regulator in SPL and enable required Kconfig
options to set GPIO4_D6 low to fix reading sdmmc on v1.1 hw revision.
Fixes: 25438c40a0 ("board: rockchip: Add Radxa ROCK S0")
Signed-off-by: Jonas Karlman <jonas@kwiboo.se>
Reviewed-by: Quentin Schulz <quentin.schulz@cherry.de>
Reviewed-by: Kever Yang <kever.yang@rock-chips.com>
