The error message
Adding disk for usb_mass_storage.lun0 failed (err=-9223372036854775788/0x8000000000000014)
provides a decimal and a hexadecimal notation of the EFI status code
EFI_ALREADY_STARTED which is defined as (EFI_ERROR_MASK | 20).
The decimal output does not convey the value 20 clearly.
With the patch we write
Adding block device usb_mass_storage.lun0 failed, r = 20
similar to other EFI error messages.
Fixes: 952018117a ("dm: sandbox: Switch over to using the new host uclass")
Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
In the SMBIOS 3 entry point the Structure Table Maximum Size field was
incorrectly named max_struct_size. A Maximum Structure Size field only
exists in the SMBIOS 2.1 entry point and has a different meaning.
Call the Structure Table Length field table_maximum_size.
Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
efi_get_variable_int() may fail, the buffer should be
cleared before using it.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Addresses-Coverity-ID: 478333 ("Error handling issues")
The code in find_smbios_table() is redundant to
efi_get_configuration_table(). Replace it.
Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
In multiple places we need a function to find an EFI configuration table.
Rename get_config_table() to efi_get_configuration_table() and export it.
Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Current U-Boot only supports the SMBIOS 3.0 entry point structure.
TCG2 measurement code should migrate to SMBIOS 3.0 entry
point structure.
efi_selftest tcg2 test also needs to be updated, and expected
PCR[1] result is changed since guid for SMBIOS EFI system table
uses different guid SMBIOS3_TABLE_GUID instead of SMBIOS_TABLE_GUID.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
In case of UCLASS_EFI_LOADER, EFI handles are managed by
EFI application/driver, we must not delete EFI handles.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Current error handling of creating raw disk/partition has
following issues.
- duplicate free for EFI handle, EFI handle is already freed
in efi_delete_handle()
- missing free for struct efi_device_path and
struct efi_simple_file_system_protocol in some error paths
To address those issues, this commit creates the common function
to free the struct efi_disk_obj resources and calls it in case
of error.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
efi_delete_handle() calls efi_purge_handle(), then it finally
frees the EFI handle.
Both diskobj and handle variables in efi_disk_remove() have
the same pointer, we can not access diskobj->dp after calling
efi_delete_handle().
This commit saves the struct efi_device_path pointer before
calling efi_delete_handle(). This commit also fixes the
missing free for volume member in struct efi_disk_obj.
This commit also removes the container_of() calls, and
adds the TODO comment of missing efi_close_protocol() call
for the parent EFI_BLOCK_IO_PROTOCOL.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
This commit auto-generates the boot option for removable
block io devices followed by fixed block io devices.
This is what EDK II reference implementation does.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Current efibootmgr auto-generates the boot option for all
disks and partitions installing EFI_SIMPLE_FILE_SYSTEM_PROTOCOL,
while EDK II reference implementation auto-generates the boot option
for all devices installing EFI_BLOCK_IO_PROTOCOL with
eliminating logical partitions.
This commit modifies the efibootmgr to get aligned to EDK II.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
check_disk_has_default_file() function checks if the
architecture-specific default file exists on the block
device, and fills the default file device path if it exists.
Rename the function name to fill_default_file_path().
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
* add generated index to table of contents
* create index entries for commands
* update Python packages used to build the documentation
* fix typos in dfu documentation
UEFI:
* split unrelated code from efi_bootmgr.c
* rename CONFIG_BOOTEFI_BOOTMGR to CONFIG_EFI_BOOTMGR
* net: tftp: remove explicit EFI configuration dependency
* fs: remove explicit EFI configuration dependency
Other:
* Add Goldfish RTC driver and make it available on RISC-V QEMU
-----BEGIN PGP SIGNATURE-----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=q+5u
-----END PGP SIGNATURE-----
Merge tag 'efi-2024-04-rc1-2' of https://source.denx.de/u-boot/custodians/u-boot-efi
Documentation:
* add generated index to table of contents
* create index entries for commands
* update Python packages used to build the documentation
* fix typos in dfu documentation
UEFI:
* split unrelated code from efi_bootmgr.c
* rename CONFIG_BOOTEFI_BOOTMGR to CONFIG_EFI_BOOTMGR
* net: tftp: remove explicit EFI configuration dependency
* fs: remove explicit EFI configuration dependency
Other:
* Add Goldfish RTC driver and make it available on RISC-V QEMU
At this point, EFI boot manager interfaces is fully independent from
bootefi command. So just rename the configuration parameter.
Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
Reviewed-by: Simon Glass <sjg@chromium.org>
Some code moved from cmd/bootefi.c is actually necessary only for "bootefi
<addr>" command (starting an image manually loaded by a user using U-Boot
load commands or other methods (like JTAG debugger).
The code will never been opted out as unused code by a compiler which
doesn't know how EFI boot manager is implemented. So introduce a new
configuration, CONFIG_EFI_BINARY_EXEC, to enforce them opted out
explicitly.
Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
QEMU provides SMBIOS tables with detailed information. We should not try to
replicate them in U-Boot.
If we want to inform about U-Boot, we can add a Firmware Inventory
Information (type 45) table in future.
Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Reviewed-by: Simon Glass <sjg@chromium.org>
This commit stores the firmware version into the array
of fmp_state structure to support the fmp versioning
for multi bank update. The index of the array is identified
by the bank index.
This modification keeps the backward compatibility with
the existing versioning feature.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
The capsule update uses the DFU framework for updating
storage. fwu_get_image_index() currently returns the
image_index calculated by (dfu_alt_num + 1), but this is
different from the image_index in UEFI terminology.
Since capsule update implementation calls dfu_write_by_alt
function, it is better that FWU returns the dfu_alt_num.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Looking back at the initrd storing functionality, we introduced three
functions, efi_dp_append_or_concatenate(), efi_dp_append/concat(). In
hindsight we could have simplified that by a lot. First of all none of
the functions append anything. They all allocate a new device path and
concatenate the contents of two device paths in one. A boolean parameter
controls the final device path -- if that's true an end node is injected
between the two device paths.
So let's rewrite this and make it a bit easier to read. Get rid of
efi_dp_append(), efi_dp_concat() and rename
efi_dp_append_or_concatenate() to efi_dp_concat(). This is far more
intuitive and the only adjustment that is needed is an extra boolean
argument on all callsites.
Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
part and dev_index in struct efi_disk_obj are not used,
let's remove it.
This commit also removes the invalid structure comment
for @dev, it does not exist.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
If one of the FMP protocols fails when calling GetImageInfo(),
populating the ESRT ends up with failure and other FMP protocols
are not added to the ESRT. We should still add all other FMP
protocols to the ESRT.
With this commit, iteration of all FMP protocols continues
even though one of the FMP protocols fails.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Tested-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
An EFI binary smbiosdump.efi is provided that can be used to check the
SMBIOS table for consistency and to dump it as a file.
The tool provides the following commands:
check
Check the SMBIOS table for consistency.
exit
Leave the tool.
help
Show available commands.
save
Save the SMBIOS table to a file on the EFI system partition. The file
can be further analyzed with the dmidecode command line tool::
dmidecode --from-dump <filename>
Specifying 'nocolor' as load option data suppresses colored output and
clearing of the screen.
Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Acked-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
The smbios.h include does not use any definitions from ofnode.h.
So don't include it.
As DECLARE_GLOBAL_DATA_PTR is no longer defined via dm/of.h we need to
add it to efi_smbios.c.
Add now missing includes to smbios-parser.c.
Remove a superfluous check comparing the sizes of the SMBIOS 2.1 and SMBIOS
3.0 anchors.
Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Reviewed-by: Simon Glass <sjg@chromium.org>
At present this code allocates memory when writing the tables and
then unnecessarily adds another memory map when installing it.
Adjust the code to allocate the tables using the normal U-Boot
mechanism. This avoids doing an EFI memory allocation early in
U-Boot, which may use memory that would be overwritten by a
'load' command, for example.
Signed-off-by: Simon Glass <sjg@chromium.org>
Only the v3 table is supported now, so always use this when installing
the EFI table.
Signed-off-by: Simon Glass <sjg@chromium.org>
Reviewed-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
EFI does not use the 'anchor string' to determine the SMBIOS table
version, instead preferring to have two separate GUIDs. Use the correct
one, depending on the table version.
Call unmap_system() to balance to the use of map_sysmem()
Signed-off-by: Simon Glass <sjg@chromium.org>
Reviewed-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
This should access arch-specific properties. Fix it and update the
existing usage.
Signed-off-by: Simon Glass <sjg@chromium.org>
Reviewed-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
We largely do not need <common.h> in these files, so drop it. The only
exception here is that efi_freestanding.c needs <linux/types.h> and had
been getting that via <common.h>.
Reviewed-by: Simon Glass <sjg@chromium.org>
Signed-off-by: Tom Rini <trini@konsulko.com>
In the prior commits, interfaces for executing EFI binary and boot manager
were carved out. Move them under efi_loader directory so that they can
be called from other places without depending on bootefi command.
Only efi_selftest-related code will be left in bootefi.c.
Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
commit 239d59a65e ("efi_loader: reconnect drivers on failure")
tried to fix the UninstallProtocol interface which must reconnect
any controllers it disconnected by calling ConnectController()
in case of failure. However, the reconnect functionality was wired in
efi_disconnect_all_drivers() instead of efi_uninstall_protocol().
As a result some SCT tests started failing.
Specifically, BBTestOpenProtocolInterfaceTest333CheckPoint3() test
- Calls ConnectController for DriverImageHandle1
- Calls DisconnectController for DriverImageHandle1 which will
disconnect everything apart from TestProtocol4. That will remain
open on purpose.
- Calls ConnectController for DriverImageHandle2. TestProtocol4
which was explicitly preserved was installed wth BY_DRIVER attributes.
The new protocol will call DisconnectController since its attributes
are BY_DRIVER|EXCLUSIVE, but TestProtocol4 will not be removed. The
test expects EFI_ACCESS_DENIED which works fine.
The problem is that DisconnectController, will eventually call
EFI_DRIVER_BINDING_PROTOCOL.Stop(). But on the aforementioned test
this will call CloseProtocol -- the binding protocol is defined in
'DBindingDriver3.c' and the .Stop function uses CloseProtocol.
If that close protocol call fails with EFI_NOT_FOUND, the current code
will try to mistakenly reconnect all drivers and the subsequent tests
that rely on the device being disconnected will fail.
Move the reconnection in efi_uninstall_protocol() were it belongs.
Fixes: commit 239d59a65e ("efi_loader: reconnect drivers on failure")
Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Reviewed-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
ACPI tables cannot convey memory reservations for ARM and RISC-V.
x86 uses the BIOS E820 table for this purpose. We cannot simply ignore the
device-tree when booting via ACPI. We have to assign EfiReservedMemory
according to the prior stage device-tree ($fdtaddr) or as fallback the
control device-tree ($fdtcontroladdr).
Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Reviewed-by: Simon Glass <sjg@chromium.org>
We currently use an outdated format 32-bit format for SMBIOS tables.
So we must allocate SMBIOS tables below 4 GiB.
Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Tested-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
It is unclear to developers why efi_var_from_file() returns EFI_SUCCESS if
file ubootefi.var is missing or corrupted. Improve the description.
Reported-by: Weizhao Ouyang <o451686892@gmail.com>
Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Reviewed-by: Weizhao Ouyang <o451686892@gmail.com>
The result of efi_append_scrtm_version() is overwritten before anyone
checks its result. Check it and exit the function on failures
Addresses-Coverity-ID: 467399 Code maintainability issues (UNUSED_VALUE)
Fixes: commit 97707f12fd ("tpm: Support boot measurements")
Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Reviewed-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
In commit 9fd3f881c6 ("efi_loader: Increase default variable store size to 64KiB")
Alper has a detailed explanation of why the size needs to be bumped to at
least 64K. However enabling Secure boot, writing db, KEK, PK etc keys
will further increase the size so bump it to 128K.
It's worth noting that when U-Boot stores the EFI variables in an RPMB the
available storage is defined statically in StandAloneMM at build time.
The U-Boot code is detecting the available true size on the fly during
writes. When StandAloneMM is present this size defines the reserved
memory U-Boot can use to copy any runtime variables, before booting an
OS.
Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Reviewed-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
This supports to boot from the URI device path.
When user selects the URI device path, bootmgr downloads
the file using wget into the address specified by loadaddr
env variable.
If the file is .iso or .img file, mount the image with blkmap
then try to boot with the default file(e.g. EFI/BOOT/BOOTAA64.EFI).
Since boot option indicating the default file is automatically
created when new disk is detected, system can boot by selecting
the automatically created blkmap boot option.
If the file is PE-COFF file, load and start the downloaded file.
The buffer used to download the ISO image file must be
reserved to avoid the unintended access to the image and
expose the ramdisk to the OS.
For PE-COFF file case, this memory reservation is done
in LoadImage Boot Service.
[Ilias fix a few memory leaks by replacing returns with gotos]
Lore: https://lore.kernel.org/u-boot/20231110042542.3797301-1-masahisa.kojima@linaro.org/T/#mbac31da301ff465b60894b38f3a587b2868cf817
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
When the image loaded by efibootmgr returns, efibootmgr
needs to clean the resources. Adding the event of returning
to efibootmgr is useful to simplify the implementation.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
const classifier is missing in EventGroup parameter of
CreateEventEx(). Fix it to remove the compiler warning.
NotifyContext parameter of CreateEventEx() is also defined
with const in UEFI specification, but NotifyContext parameter
of CreateEvent() is defined without const.
Since current implementation calls the common efi_create_event()
function from both CreateEventEx() and CreateEvent() services,
NotifyContext parameter leaves as is.
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Changes for complying to EFI spec §3.5.1.1
'Removable Media Boot Behavior'.
Boot variables can be automatically generated during a removable
media is probed. At the same time, unused boot variables will be
detected and removed.
Please note that currently the function 'efi_disk_remove' has no
ability to distinguish below two scenarios
a) Unplugging of a removable media under U-Boot
b) U-Boot exiting and booting an OS
Thus currently the boot variables management is not added into
'efi_disk_remove' to avoid boot options being added/erased
repeatedly under scenario b) during power cycles
See TODO comments under function 'efi_disk_remove' for more details
The original efi_secboot tests expect that BootOrder EFI variable
is not defined. With this commit, the BootOrder EFI variable is
automatically added when the disk is detected. The original
efi_secboot tests end up with unexpected failure.
The efi_secboot tests need to be modified to explicitly set
the BootOrder EFI variable.
squashfs and erofs ls tests are also affected by this modification,
need to clear the previous state before squashfs ls test starts.
Co-developed-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Signed-off-by: Masahisa Kojima <masahisa.kojima@linaro.org>
Signed-off-by: Raymond Mao <raymond.mao@linaro.org>
Reviewed-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Reviewed-by: Joao Marcos Costa <jmcosta944@gmail.com>
Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
The efi_gop driver uses private fields from the video uclass to obtain a
pointer to the frame buffer. Use the platform data instead.
Check the VIDEO_COPY setting to determine which frame buffer to use. Once
the next stage is running (and making use of U-Boot's EFI boot services)
U-Boot does not handle copying from priv->fb to the hardware framebuffer,
so we must allow EFI to write directly to the hardware framebuffer.
We could provide a function to read this, but it seems better to just
document how it works. The original change ignored an explicit comment
in the video.h file ("Things that are private to the uclass: don't use
these in the driver") which is why this was missed when the VIDEO_COPY
feature was added.
Signed-off-by: Simon Glass <sjg@chromium.org>
Fixes: 8f661a5b66 ("efi_loader: gop: Expose fb when 32bpp")
Reviewed-by: Bin Meng <bmeng.cn@gmail.com>
The command should not be used to enable library functionality. Add a
new BOOTEFI_BOOTMGR Kconfig for that. Adjust the conditions so that the
same code is built.
Signed-off-by: Simon Glass <sjg@chromium.org>
Suggested-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
efi_tcg2_get_active_pcr_banks doesn't immediately call the
EFI_ENTRY() wrapper once it enters the function. Move the call a
few lines above to cover the error cases properly as well.
Reviewed-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Add TPM2 functions to support boot measurement. This includes
starting up the TPM, initializing/appending the event log, and
measuring the U-Boot version. Much of the code was used in the
EFI subsystem, so remove it there and use the common functions.
Signed-off-by: Eddie James <eajames@linux.ibm.com>
For the API moving around from EFI -> u-boot core
Reviewed-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
For EFI testing
Tested-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
The driver needs to support getting the PCRs in the capabilities
command. Fix various other things and support the max number
of PCRs for TPM2.
Remove the !SANDBOX dependency for EFI TCG2 as well.
Signed-off-by: Eddie James <eajames@linux.ibm.com>
Reviewed-by: Simon Glass <sjg@chromium.org>
Acked-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
U-Boot should set up the SMBIOS tables during startup, as it does on x86.
Ensure that it does this correctly on non-x86 machines too, by creating
an event spy for last-stage init.
Tidy up the installation-condition code while we are here.
Signed-off-by: Simon Glass <sjg@chromium.org>
Reviewed-by: Bin Meng <bmeng.cn@gmail.com>
