Re: [PATCH v6 11/30] objtool: Trace instruction state changes during function validation

[Alexandre Chartre <alexandre.chartre@oracle.com>]

On 12/2/25 02:30, Josh Poimboeuf wrote:
> On Mon, Dec 01, 2025 at 01:23:29PM -0700, Nathan Chancellor wrote:
>> Hi Alexandre,
>>
>> On Fri, Nov 21, 2025 at 10:53:21AM +0100, Alexandre Chartre wrote:
>>> During function validation, objtool maintains a per-instruction state,
>>> in particular to track call frame information. When tracing validation,
>>> print any instruction state changes.
>>>
>>> Signed-off-by: Alexandre Chartre <alexandre.chartre@oracle.com>
>>
>> I am seeing a segfault after this change in -next as commit fcb268b47a2f
>> ("objtool: Trace instruction state changes during function validation")
>> when building allmodconfig with clang 21.1.6 [1] (I did not check
>> earlier versions).
>>
>>    $ clang --version | head -1
>>    ClangBuiltLinux clang version 21.1.6 (https://github.com/llvm/llvm-project.git a832a5222e489298337fbb5876f8dcaf072c5cca)
>>
>>    $ make -skj"$(nproc)" ARCH=x86_64 LLVM=1 clean allmodconfig drivers/scsi/qla2xxx/qla2xxx.o
>>    make[7]: *** [scripts/Makefile.build:503: drivers/scsi/qla2xxx/qla2xxx.o] Error 139
>>    ...
>>
>>    $ ld.lld -m elf_x86_64 --fatal-warnings -z noexecstack -r -o drivers/scsi/qla2xxx/qla2xxx.o @drivers/scsi/qla2xxx/qla2xxx.mod
>>
>>    $ tools/objtool/objtool --hacks=jump_label --hacks=noinstr --hacks=skylake --ibt --cfi --mcount --mnop --orc --retpoline --rethunk --sls --static-call --uaccess --no-unreachable --link --module drivers/scsi/qla2xxx/qla2xxx.o
>>    fish: Job 1, 'tools/objtool/objtool --hacks=j…' terminated by signal SIGSEGV (Address boundary error)
>>
>> If there is any other information I can provide or patches I can test, I
>> am more than happy to do so.
>>
>> [1]: https://mirrors.edge.kernel.org/pub/tools/llvm/files/llvm-21.1.6-x86_64.tar.xz
> 
> Objtool is overflowing the stack due to the large number of jumps it has
> to follow in that code, thanks to kasan.  The above mentioned patch
> 
>    fcb268b47a2f ("objtool: Trace instruction state changes during function validation")
> 
> added a 328-byte struct to the stack in validate_insn() which
> drastically increased the amount of stack size needed.
> 
> I suppose we could hack a fix by making it a static local variable, like
> below.
> 
> Or, objtool could setrlimit(RLIMIT_STACK) to 16MB?
> 
> diff --git a/tools/objtool/check.c b/tools/objtool/check.c
> index a02f8db75827..206b8589d82b 100644
> --- a/tools/objtool/check.c
> +++ b/tools/objtool/check.c
> @@ -3678,7 +3678,7 @@ static int validate_insn(struct objtool_file *file, struct symbol *func,
>   			 bool *dead_end)
>   {
>   	/* prev_state is not used if there is no disassembly support */
> -	struct insn_state prev_state __maybe_unused;
> +	static struct insn_state prev_state __maybe_unused;
>   	struct alternative *alt;
>   	u8 visited;
>   	int ret;

static looks good enough to me.

Reviewed-by: Alexandre Chartre <alexandre.chartre@oracle.com>

Thanks,

alex.