The inner-kernel section loader checks VirtualAddress + SizeOfRawData
against kernel_size_in_memory (for the memcpy), but the memzero right
after it clears up to VirtualAddress + VirtualSize, and VirtualSize is
only constrained to be >= SizeOfRawData.
Reject a VirtualAddress + VirtualSize that overflows or exceeds
kernel_size_in_memory, mirroring the existing SizeOfRawData checks.
Follow-up for cab9c7b5a4
The stub copies the embedded kernel PE into an xmalloc_pages()
allocation with AllocateAnyPages, which only guarantees EFI_PAGE_SIZE
(4 KiB) alignment.
The arm64 kernel declares SectionAlignment = SZ_64K. When _text is not
64K-aligned the kernel's EFI stub reallocates and copies the image,
which can fail with EFI_OUT_OF_RESOURCES on memory-constrained firmware
(e.g. U-Boot, bounded by CONFIG_SYS_MALLOC_LEN rather than full DRAM).
The failure is non-deterministic since AllocateAnyPages returns whatever
page the allocator finds first.
Plumb SectionAlignment through pe_kernel_info() and allocate via
xmalloc_aligned_pages(). For alignment <= EFI_PAGE_SIZE (e.g. x86_64) it
reduces to a plain AllocatePages(), so the change is free there.
SectionAlignment comes from the PE header, which the spec requires to be
a power of 2. Sanitize it in pe_kernel_info() right where it is parsed:
a non-conforming value falls back to plain page alignment (matching the
behavior of older systemd-stub versions that ignored the field) rather
than propagate something that would break xmalloc_aligned_pages()'s
over-alignment maths, with a log_warning() so the fallback stays
diagnosable.
Co-developed-by: Claude <claude@anthropic.com>
Closes: #42443
Reported-By: Agathe Porte <agathe.porte@oss.qualcomm.com>
Tested-By: Agathe Porte <agathe.porte@oss.qualcomm.com>
The current code incorrectly subtracts ImageBase from section
VirtualAddress values when loading sections into memory. This is based
on a misunderstanding of the PE specification.
VirtualAddress in section headers is the address of the first byte of
the section relative to the image base when the section is loaded into
memory. In other words, VirtualAddress is already an RVA measured from
the image base, it is definitely NOT an absolute address that needs to
be adjusted.
So when loading a PE image into a newly allocated buffer, sections
should be copied to buffer + VirtualAddress, regardless of what
ImageBase says. The ImageBase field merely indicates the *preferred*
load address, it does not affect how section RVAs are interpreted.
This happens to not cause issues when ImageBase was 0 (since
VirtualAddress - 0 = VirtualAddress), which is why this bug went
undetected on modern kernels. However, it fails with kernels that have
non-zero ImageBase values.
So let's remove the nonsensical VirtualAddress < ImageBase check, and
remove the ImageBase subtractions from section loading offsets. This
lets all kernel UKIs work properly again.
Fixes: #40342
- Actually pass the new memory file path to parent_loaded_image->FilePath
- Restore old parent_loaded_image if Linux returns
- Pass the same kernel_file_path in load_via_boot_services path
- s/Re-use/Patch in comment explaining what we are doing
Fixes#38566
Commit 70b7e03 introduced 3 calls to log_debug() about the presence or
absence of NX_COMPAT support. Since sd-boot does not yet have the
ability to only print messages above a certain loglevel, these will
always be printed, even on top of the configured splash screen. This
commit removes the log_debug() call after a success and only prints
those for missing firmware support if the UEFI should support them in
the first place (i.e. starting with version 2.10).
When NX_COMPAT gets enabled, firmwares will enforce that executable
memory is either writable or executable.
This needs kernel compatibility, when it will happen the kernel will
have the NX_COMPAT bit set. If it is, set the memory buffer to RO.
Note that this must be undone on failure, as EDK2 in some configurations
overwrites memory ranges that are returned with FreePages() with a
fixed pattern, so if the pages are RO it will crash.
This is only an issue with the new custom PE loader, as LoadImage()
and StartImage() will always do the right thing automatically.
https://microsoft.github.io/mu/WhatAndWhy/enhancedmemoryprotection/https://www.kraxel.org/blog/2023/12/uefi-nx-linux-boot/
Follow-up for cab9c7b5a4
Fixes https://github.com/systemd/systemd/issues/38545
- Reuse parent_image instead of allocating new ones. Firmware might cast
EFI_LOADED_IMAGE_PROTOCOL * to a larger struct causing issues
- Remove loaded image protocol installation and uninstallation which are no
longer required
Fixes a bug introduced by cab9c7b5a4.
Fixes#38567.
Co-authored-by: Tobias Heider <tobias.heider@canonical.com>
It is possible to boot directly a UKI from shim, and Fedora Cloud Base UKI
does exactly this.
This used to work fine until https://github.com/systemd/systemd/pull/37372
which broke compatibility when shim < 16.0 (no loader protocol override) is
used. Shim 15.8 is still in use in several distributions, and will be for
a long time.
Restore a part of the previous implementation, and if running with secure
boot enabled, and with shim but < 16.0, apply a security override.
Follow-up for cab9c7b5a4
Fixes https://github.com/systemd/systemd/issues/38104
The comment in linux_exec() was based on Lennart's comment in
https://github.com/systemd/systemd/pull/37372#discussion_r2142340582, but
shortened. The original wording is more direct and at least for me easier to
grok, so adjust the comment to be more verbose again.
Also, move the comment from shim_loader_available() to the place where it used.
This function is for checking if the new thing is available, no need to describe
the old thing there.
Since shim 16, it is not possible anymore to override the security arch
protocol to avoid signature check and measurements.
Also on Dell servers, EFI_SECURITY2_ARCH_PROTOCOL is not implemented, so
unexpected measurements on PCR 4 were still happening.
This just loads and run the kernel pe. We verify that there is no relocation
needed. Also for simplification, we assume and verify that the base address is
expected to be 0.
I very much dislike the approach in which we were mixing Linux and UEFI C code
in the same subdirectory. No code was shared between two environments. This
layout was created in e7dd673d1e, with the
justification of "being more consistent with the rest of systemd", but I don't
see how it's supposed to be so.
Originally, when the C code was just a single bootctl.c file, this wasn't so
bad. But over time the userspace code grew quite a bit. With the moves done in
previuos commits, the intermediate subdirectory is now empty except for the
efi/ subdir, and this additional subdirectory level doesn't have a good
justification. The components is called "systemd-boot", not "systemd-efi", and
we can remove one level of indentation.
Luca Boccassi
Josh Hoffer
Zbigniew Jędrzejewski-Szmek
cdown
Lennart Poettering
Yu Watanabe
Tobias Heider
Felix Pehla
Mate KukriandTobias Heider
Zbigniew Jędrzejewski-Szmek
Valentin David
Daan De Meyer