Hi Rasmus,
On 2026-05-22T21:27:48, Rasmus Villemoes <[email protected]> wrote:
> linker_lists.h: emit lots of meta-data for debugging and sanity checking
>
> This is an attempt at to make it possible to catch various problems
> with linker lists at build time, without affecting the size of the
> final binary. Also, I did not want to have to modify all the linker
> scripts in order to not get all this meta-data thrown away by garbage
> collection.
>
> The basic trick is that one can use the .size assembly directive so
> set the st_size of an ELF symbol, without that symbol actually
> occupying that amount of space in any section or the final binary. And
> all of the .size, .type, .{push,pop}section directives are completely
> generic, so this should work for all architecture.
>
> So whenever we declare a start or end symbol for use by C code, also
> emit information about the size and alignment of the type that the
> list is supposed to be made of.
>
> A relatively straight-forward script (next patch) can then parse the
> ELF file and report on any inconsistencies. For example, suppose two
> [...]
>
> include/linker_lists.h | 43 +++++++++++++++++++++++++++++++++++++++++++
> 1 file changed, 43 insertions(+)
I see this:
test/cmd_ut.c:67:1: note: in expansion of macro ‘SUITE_DECL’
67 | SUITE_DECL(mbr);
| ^~~~~~~~~~
include/linker_lists.h:45:17: error: expected ‘)’ before ‘:’ token
45 | : : "i"(sizeof(_type)), "i"(__alignof__(_type)))
| ^
include/linker_lists.h:48:9: note: in expansion of macro ‘ll_emit_type_info’
48 | ll_emit_type_info(_list, _type);
\
| ^~~~~~~~~~~~~~~~~
include/linker_lists.h:297:9: note: in expansion of macro ‘ll_emit_start_symbol’
297 | ll_emit_start_symbol(_list, _type);
\
| ^~~~~~~~~~~~~~~~~~~~
See below
> diff --git a/include/linker_lists.h b/include/linker_lists.h
> @@ -23,7 +24,45 @@
> +#define ll_emit_type_info(_list, _type) \
> + __asm__( \
> + ".pushsection "ll_info_section_name(_list)',\'aw\'\n' \
> + ".type "ll_size_symbol_name(_list)", STT_OBJECT\n" \
> + ".size "ll_size_symbol_name(_list)", %c0\n" \
> + ".type "ll_align_symbol_name(_list)", STT_OBJECT\n" \
> + ".size "ll_align_symbol_name(_list)", %c1\n" \
> + ll_size_symbol_name(_list)':\n' \
> + ll_align_symbol_name(_list)':\n' \
> + '.popsection\n' \
> + : : 'i'(sizeof(_type)), 'i'(__alignof__(_type)))
This is extended asm (it has operand constraints), so GCC requires it
to be inside a function. ll_start_decl()/ll_end_decl() pull this into
ll_emit_start_symbol/ll_emit_end_symbol at file scope - see
SUITE_DECL() in test/cmd_ut.c
One way out is to drop the operands and have the C side emit a
zero-initialised marker object whose array dimensions encode
sizeof/__alignof__ - e.g. a static struct in a dedicated section whose
two members are sized sizeof(_type) and __alignof__(_type). That costs
a few bytes per list per TU but keeps the trick working at file scope.
> diff --git a/include/linker_lists.h b/include/linker_lists.h
> @@ -23,7 +24,45 @@
> +#define ll_emit_start_symbol(_list, _type) \
> + ll_emit_type_info(_list, _type); \
> + __asm__( \
> + ".pushsection "ll_start_section_name(_list)',\'aw\'\n' \
> + ".type "__stringify(ll_start_symbol_name(_list))",
> STT_OBJECT\n" \
> + ".size "__stringify(ll_start_symbol_name(_list))", 0\n" \
> + __stringify(ll_start_symbol_name(_list))':\n' \
> + '.popsection\n' \
> + )
These multi-statement macros are not wrapped, so they only work in
contexts that accept two declarations/statements separated by ';'.
ll_start_decl() then chains another declaration after them, leaving a
stray ';' at file scope - accepted as a GCC extension but rejected
under -Wpedantic. Please either fold the two asm blocks into one (a
single .pushsection ... .popsection can contain all the
.type/.size/label directives), or wrap them in a way that gives a
single declaration/statement. The latter is awkward at file scope,
which is another argument for the single-asm approach.
> diff --git a/include/linker_lists.h b/include/linker_lists.h
> @@ -23,7 +24,45 @@
> +#define ll_info_section_name(_list) '__u_boot_list_'#_list'_0'
Worth mentioning in the commit message that this is matched by the
existing KEEP(*(SORT(__u_boot_list*))) globs in all the linker
scripts, so the new _0 sections end up in the linker_list region of
the final image. Since each section is zero bytes that does not affect
binary size, but it does mean the metadata is reachable from the ELF
symtab without changing any board's linker script - which is exactly
what you say you wanted.
> diff --git a/include/linker_lists.h b/include/linker_lists.h
> @@ -23,7 +24,45 @@
> +#define ll_info_section_name(_list) '__u_boot_list_'#_list'_0'
> +#define ll_size_symbol_name(_list)
> '_u_boot_list_'#_list'_0_item_size'
> +#define ll_align_symbol_name(_list)
> '_u_boot_list_'#_list'_0_item_align'
The size/align symbols are emitted as local labels, so every TU that
touches a given list contributes its own local copy. That is fine for
the parser in patch 9, but it relies on strip/objcopy not throwing
away local symbols on the way to the final image. A note about this -
or about pointing the script at u-boot (the unstripped ELF) rather
than u-boot.bin - would help anyone trying this themselves.
> diff --git a/include/linker_lists.h b/include/linker_lists.h
> @@ -128,6 +167,7 @@
> #define ll_entry_start(_type, _list) \
> ({ \
> + ll_emit_start_symbol(_list, _type); \
> static char start[0] __aligned(CONFIG_LINKER_LIST_ALIGN) \
> __section(ll_start_section_name(_list)); \
Just to check: is there any reason ll_entry_start() emits the metadata
at every call site rather than only where the start/end markers are
first defined? Within a given TU every call expands to the same
.size/.type for the same symbol, which the assembler tolerates but is
wasted work. If you can move it into the declarations themselves (or
use a once-per-TU guard) the diff for existing call sites stays the
same and the assembler output gets much smaller.
Regards,
Simon
