When preparing LLEXT ELF sections only invalidate instruction cache
of executable sections. Also skip the step on platforms, managing
cache at the application level.
Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>
Not sure why this vestigal struct member was here at this point (18
months ago...) but its unused and was causing problems for certain
parts.
Signed-off-by: Tom Burdick <thomas.burdick@intel.com>
This removes 'return' statements affecting code flow from debug code in
llext_link(). Also, all valid relocations are tried before returning an
error, so that the user can see all the errors (most likely missing
symbols) at once.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
With experience it becomes clear, that failing to resolve symbols
during the linking process is likely fatal for the module loading and
a simple warning isn't enough. Fail loading instead.
Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>
Some applications need to save LLEXT context, e.g. when suspending,
to later restore it quickly without a full relinking. Add 2 functions
for context saving and restoring. Since these functions are likely to
change in the future, put them in llext_experimental.c, which depends
on CONFIG_LLEXT_EXPERIMENTAL and is disabled by default.
Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>
Beginning symbols or structure field names with an underscore isn't a
widely used style in Zephyr, remove the leading underscore from
_llext_list.
Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>
Some applications require loading extensions into the memory which does
not exist during the boot time and cannot be allocated statically. Make
the application responsible for LLEXT heap allocation. Do not allocate
LLEXT heap statically.
Signed-off-by: Alex Ivanov <alexivanov@google.com>
Make it so LLEXT allocates a memory region equal or greater
to the minimum size supported by the MPU, to ensure no region
overlap.
Signed-off-by: Lauren Murphy <lauren.murphy@intel.com>
Kconfig option LLEXT_EXPORT_DEVICES can be enabled to make all device
objects of an image available to LLEXTs, but it is disabled by default.
If an extension tries to import devices when the base image was not built
with this option, the dynamic linking equivalent of the much beloved error
"Undefined symbol __device_dts_ord_XXX" is logged, but this can be quite
cryptic to understand since user may be unaware of LLEXT_EXPORT_DEVICES.
Detect such invalid imports and print a special message that directs users
towards the appropriate Kconfig option, which should reduce the confusion.
Signed-off-by: Mathieu Choplain <mathieu.choplain@st.com>
When loading an extension, read-only sections that do not have
relocations and whose data is accessible from the ELF buffer can be
directly mapped as-is in the extension memory. This avoids the need to
allocate and copy unmodified data from the ELF buffer to the LLEXT heap.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
This patch moves the alignment checks for MPU and MMU regions to the
beginning of the llext_copy_region() function. This is done to ensure
that the correct region alignment and size are used even in the case
where the region is reused from the ELF file buffer, avoiding MMU/MPU
configuration issues.
This also relaxes the same checks for regions that are accessed by the
kernel only (e.g. symbol and string tables), which do not need special
MMU/MPU treatment. This exposed an inconsistency in the MMU code, which
was setting the permission on the correct regions, but later restoring
the default permissions on every region, including the now-unaligned
ones.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
This patch introduces an 'enum llext_storage_type' that defines the
storage type of the ELF data that will be loaded:
- LLEXT_STORAGE_TEMPORARY: ELF data is only available during
llext_load(); even when the loader supports directly accessing the
memory via llext_peek(), the buffer contents will be discarded
afterwards.
LLEXT will allocate copies of all required data into its heap.
- LLEXT_STORAGE_PERSISTENT: ELF data is stored in a *read-only* buffer
that is guaranteed to be always accessible for as long as the
extension is loaded.
LLEXT may directly reuse constant data from the buffer, but may still
allocate copies if relocations need to be applied.
- LLEXT_STORAGE_WRITABLE: ELF data is stored in a *writable* memory
buffer that is guaranteed to be always accessible for as long as the
extension is loaded.
LLEXT may freely modify and reuse data in the buffer; so, after the
extension is unloaded, the contents should be re-initialized before a
subsequent llext_load() call.
To keep full compatibility with existing code, the storage type of
LLEXT_BUF_LOADER is determined by the CONFIG_LLEXT_STORAGE_WRITABLE Kconfig
option between LLEXT_STORAGE_PERSISTENT and LLEXT_STORAGE_WRITABLE.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
Adding this information allows easy cross-reference of the effects of
memory optimization changes between subsequent commits.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
This was inspired by the detection of 2 instances of the warning:
warning: 'strncpy' specified bound 16 equals destination size
[-Wstringop-truncation]
The current code is already safe with regards to overflows, because
fixed-length string functions are used in the call tree. However, when
given a name 16 chars or larger, the current compare in llext_by_name()
will not work as expected because the stored extension name is truncated
to a max of 15.
Define a global LLEXT_MAX_NAME_LEN constant to simplify all this logic
and also implement name checks in the shell before calling llext_load().
Finally, using strlen() instead of strnlen() gets the real length of the
hex string passed as a parameter, which is important for the next safety
check.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
This patch adds support for alignment requirements for sections in the
loader. The alignment requirements are taken from the ELF file and
checked when the section is mapped to a memory region and when the
region is copied to memory (or directly used from the ELF buffer).
Supporting a larger alignment requirement than the region's current
alignment is done by adjusting the region's start address. This needs
care to avoid detecting false overlaps with other regions and to ensure
that the region address and size reported by the inspection APIs are
correct. The extra offset needed to re-adjust these values is stored in
the 'sh_info' field of the region descriptor.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
Since commit 0aa6b1c9de, the 'ldr_parm' pointer is guaranteed to be
valid inside all the functions of the llext_load() call tree.
This commit fixes the only exception of llext_copy_strings(), which was
not passed the 'ldr_parm' pointer, and remove the redundant checks.
No functional change is intended by this commit.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
This commit adjusts the logic of the LLEXT memory allocation routine by:
* defining 'align' and 'alloc' variables once and updating them in the
following code;
* swapping the condition, converting it to an 'if(IS_ENABLED(...))' and
using a direct calculation for the size to make the code easier to
read;
* updating and extending a few comments.
No functional change is intended.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
Since 3466dab804 the generic llext_symbol_name() function abstracts
the use of llext_string() for (non-section) symbols. Define a similar
llext_section_name() function and replace current occurrences of
llext_string() with the proper abstraction.
By extending llext_symbol_name(), this commit also allows to print the
correct name of sections that are referred to by a symbol.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
Rewrite the overlap checks in llext_map_sections() to use a pair of
macros to calculate the top and bottom of a region. This makes the
code more readable and less error-prone.
No functional change is introduced in this commit apart from a log
message format change.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
Use the more compact %#x format for printing out hexadecimal values in
the LLEXT subsystem. This also uses a clearer wording for the
llext_link() log message that avoids repeating information.
No functional change is introduced by this commit.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
llext_copy_section() is a misnomer, as it is copying a region of memory,
not a section. Rename it to llext_copy_region() and use 'region' in the
internal variables to better reflect their purpose.
No functional change is introduced in this commit.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
Add a 'continue' statement when the first section of a region is
detected, to avoid unnecessary indentation.
No functional change is introduced in this commit.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
Extensions with the pre_located flag set, manage memory themselves,
including MMU permissions. Skip any such adjustments in LLEXT core.
Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>
The RISC-V port of llext requires additional parameters for
handling non-adjacent HI20/LO12 relocations in arch_elf_relocate():
the current extension (struct llext), the current extension loader
(struct llext_loader), the current section header (elf_shdr_t) and
the current symbol (elf_sym_t).
This changes the signature of arch_elf_relocate accordingly.
Signed-off-by: Eric Ackermann <eric.ackermann@cispa.de>
This commit refactors logic in llext_link.c that resolves
a symbol to be imported by an llext into a utility function.
Thereby, the logic can be re-used in other functions if
needed, e.g., when resolving symbols for indirect relocations.
Signed-off-by: Eric Ackermann <eric.ackermann@cispa.de>
This patch adds support for Zstd and Zip formats to the EDK generation
process. The user can now choose between XZ, Zstd, and Zip compression
and output formats for the EDK file.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
Add a new Kconfig option to make the generation of an Extension
Development Kit (EDK) for the LLEXT subsystem optional. This
allows to cleanly separate EDK-related configuration and build
steps from the rest of the Zeprhyr build system.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
Add APIs to inspect the contents of an ELF file loaded as an extension.
This is useful for applications that need to access the contents of the
extension in a more fine-grained way than the existing LLEXT APIs.
Use of these APIs requires the 'keep_elf_data' option to be provided via
struct llext_load_param to the 'llext_load()' call.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
This is required for the functions that inspect the ELF file to work
properly. The user must then call llext_free_inspection_data() before
calling llext_unload() to free the memory left allocated in the loader
and extension memory.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
When building relocatable extensions, exported symbol names might end
up in a separate section (in my case ".rodata.str1.1") for which
storage address don't match section address. To access those names
addresses have to be relalculated.
Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>
The new CONFIG_LLEXT_IMPORT_ALL_GLOBALS option allows all global symbols
from extensions to be used by the main application. This is useful to
load basic extensions that have been compiled without the full Zephyr
EDK.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
Make sure the buffer used to load the hex file is aligned to a sensible
value; at least word size is assumed by the LLEXT APIs.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
The shell command 'llext list_symbols' currently iterates over the
'sym_tab' table. This is valid only inside 'llext_load' and is freed
before returning, so it will always be empty in shell context.
Fix it to display 'exp_tab', which is the one used at runtime by LLEXT
to provide symbols to the main application.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
The ELF format allows for multiple string and symbol tables with
complex references between them. This is especially evident when
debugging information is included.
This patch fixes the issues that have been identified with multiple
string tables to allow LLEXT to properly parse those files:
* The symbol table used by LLEXT (LLEXT_MEM_SYMTAB) is now chosen
depending on the loaded file type, and other tables are ignored.
This change is also applied to the SLID injection script.
* The LLEXT string table (LLEXT_MEM_SYMTAB) is now correctly identified
by the symbol table reference, instead of picking the first one.
* VMA range checks only make sense for allocated sections.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
This patch adds a debug log message to display the addresses of the
sections in a format that can be used with the gdb 'add-symbol-file'
command to properly inform the debugger about the final addresses for
the loaded sections.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
When building an ELF file with debugging information, the compiler
places this data in sections that do not need to be loaded in memory for
the program to run. These sections are marked as "not allocated" in the
ELF section header, and can be safely ignored when loading the file.
This patch adds a check to the relocation processing code to skip
relocations acting on those non-allocated sections, allowing to load ELF
files with debugging information.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
This new option allows to export devices using identifiers generated
from the hash of the devicetree node path, instead of the device's
ordinal number. Identifiers generated this way are stable across
rebuilds.
Add new test cases to test this new option.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
Ensure all pending log messages are processed by the log processing
thread when log_flush is called, blocking the caller until done.
Signed-off-by: Tom Burdick <thomas.burdick@intel.com>
The check combination of "is zero" and then "less than zero" leaves open
the theoretical possibility of a positive return value, which would
continue on with an uinitialized 'sym'.
Checking for "not zero" has the same effect and covers all situations.
Commit 56fd85442 did the same thing for 'rela' a few lines above, but
missed this instance. No other places in the llext code have this issue.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
Currently llext_link_plt() calculates offsets to relocation addresses
relative to the .text section and passes them to
arch_elf_relocate_global() and arch_elf_relocate_local(), where then
.text memory address is added to them. Instead it's more logical to
concentrate the entire calculation in the caller, which then also
removes the assumption, that all sections have the same VMA - LMA
offset from those functions.
Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>
Some ARC processor configurations have separate memory for code (ICCM)
and for data (DCCM). Such configurations are unsuitable for LLEXT,
except for some quite special cases. For now, disable LLEXT and its
tests for these devices completely.
Signed-off-by: Ilya Tagunov <Ilya.Tagunov@synopsys.com>
The weak syscall symbols generated by gen_syscalls.py are currently
compiled in the LLEXT subsystem library, which is then linked among all
other Zephyr libraries in an unspecified order. This can cause the weak
symbols to override the actual syscall implementations, leading to
undefined behaviour.
To fix this, the currently generated file is split in two elements:
- syscall_exports_llext.c contains the EXPORT_SYMBOL directives for all
syscalls. This part can be compiled with the LLEXT library and linked
among all other Zephyr libraries, and ensures all syscalls symbols
are preserved by the linker.
- syscall_weakdefs_llext.c contains the weak definitions for all syscalls.
This file is compiled in a separate library that is linked last, so
that the weak symbols are only used if no other implementation is
available.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
The check combination of "is zero" and then "less than zero" leaves open
the theoretical possibility of a positive return value, which would
continue on with an uinitialized 'rela'.
Checking for "not zero" has the same effect and covers all situations.
Signed-off-by: Luca Burelli <l.burelli@arduino.cc>
When linking and relocations are performed on the ELF object itself
with no copying, also global binding linking can break references.
Disable linking globally for such cases.
Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>
Now that section header cache is persistent, it can be used for
finding sections. Add a new function, similar to llext_find_section()
to use it and deprecate llext_find_section().
Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>
Move cached section headers to struct llext from struct llext_loader
to preserve them after llext_load() returns.
Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>
Currently when building LLEXT for Xtensa we use the -fPIC compiler
option, but this cannot be used when using detached sections in
extensions. Add a Kconfig option to switch between the two
compilation modes and switch -fPIC off when building relocatable
(partially linked) ELF binaries.
Signed-off-by: Guennadi Liakhovetski <guennadi.liakhovetski@linux.intel.com>
