On Tue, Mar 24, 2026 at 09:04:03AM -0700, Alexei Starovoitov wrote:
On Tue, Mar 24, 2026 at 9:00 AM Sasha Levin <[email protected]> wrote:
On Tue, Mar 24, 2026 at 08:03:30AM -0700, Alexei Starovoitov wrote:
>On Mon, Mar 23, 2026 at 9:49 AM Sasha Levin <[email protected]> wrote:
>>
>> Embed DWARF-derived function parameter name and type information in the
>> kernel image so that oops and WARN dumps display the crashing function's
>> register-passed arguments with their names, types, and values.
>>
>> A new build-time tool (scripts/gen_paraminfo.c) parses DW_TAG_subprogram
>> and DW_TAG_formal_parameter entries from DWARF .debug_info, extracting
>> parameter names and human-readable type strings. The resulting tables are
>> stored in .rodata using the same two-phase link approach as lineinfo.
>>
>> At runtime, kallsyms_show_paraminfo() performs a binary search on the
>> paraminfo tables, maps parameters to x86-64 calling convention registers
>> (RDI, RSI, RDX, RCX, R8, R9), and prints each parameter's name, type,
>> and value from pt_regs. If a parameter value matches the page fault
>> address (CR2), it is highlighted with "<-- fault address".
>>
>> Integration at show_regs() means this works for both oops and WARN()
>> automatically, since both paths provide full pt_regs at the exception
>> point.
>>
>> Example output:
>>
>> Function parameters (ext4_readdir):
>> file (struct file *) = 0xffff888123456000
>> ctx (struct dir_context *) = 0x0000000000001234 <-- fault address
>>
>> Gated behind CONFIG_KALLSYMS_PARAMINFO (depends on CONFIG_KALLSYMS_LINEINFO).
>> Adds approximately 1-2 MB to the kernel image for ~58K functions.
>>
>> Assisted-by: Claude:claude-opus-4-6
>> Signed-off-by: Sasha Levin <[email protected]>
>
>Nack.
>
>You asked claude to reinvent pahole and BTF and it did it
>completely missing years of fine tuning that pahole does.
Let's keep this on the technical side please.
>dwarf cannot be trusted as-is. pahole converts it carefully
>by analyzing optimized out arguments and dropping signatures
Fair point about pahole and optimized-out args. The problem is that BTF depends
on BPF_SYSCALL, and the environments I care about can't enable either.
This is trivially fixable without reinventing pahole.
Hmm...
Looking at the code, I'd need to:
- Split BTF parsing from kernel/bpf/btf.c to somewhere outside of kernel/bpf/.
- New init path for btf_vmlinux outside BPF verifier.
- Refactor btf_parse_vmlinux() BPF-specific bits.
- Remove BPF_SYSCALL dependency from DEBUG_INFO_BTF.
- Somehow make BTF work with DEBUG_INFO_REDUCED.
I suppose that the first 4 are straightforward, but I don't have an idea about
DEBUG_INFO_REDUCED. Though we can probably tackle it later.
Does that make sense? Did I miss anything?
--
Thanks,
Sasha
