Re: [PATCH] kconfig: reject malformed KCONFIG_PROBABILITY values

[Julian Braha]

Hi Dmitrii,

I see that this is your first patch submission. Welcome!
And thank you for reporting this.

On 9/7/26 20:45, Dmitrii Tulnov wrote:
> randconfig checks the numeric range of each probability but does not
> validate where strtol() stops. For example, KCONFIG_PROBABILITY=50% is
> accepted as 50:0:0: the '%' is parsed twice as zero. This silently sets
> both tristate y/m probabilities to zero, reducing the coverage of random
> configuration builds. 

It would be good to note here that 'KCONFIG_PROBABILITY=50' (without the
'%') should be 25:25:50 instead.

> Empty fields and extra fields are also accepted.

Right, in these cases, I think it's reasonable to assume the user made a
typo or other oversight.

> 
> Require each field to contain an integer followed by the end of the
> string or a colon introducing another field, with at most three fields.
> Preserve the leading whitespace and optional sign accepted by strtol().

I disagree with allowing the leading whitespace and sign. The fact that
it's allowed input for strtol() is incidental to us. The intended input
format for Kconfig is:

XX:XX:XX

not:

  XX: +XX:    -XX

> Keep the strtol() result as long until the range check so that narrowing
> to int cannot turn an out-of-range value into a valid probability.
> 
> Add regression tests for malformed and out-of-range values, the supported
> probability formats, and the documented empty-value default.
> 
> Fixes: e43956e60769 ("kconfig: implement KCONFIG_PROBABILITY for randconfig")
> Assisted-by: LLM
> Signed-off-by: Dmitrii Tulnov <tulnov.dl@xxxxxxxxx>
> ---
> Validation on kbuild-next, x86_64, GCC 13.3.0:
> - make testconfig with HOSTCFLAGS=-Werror: 58 passed. With the original
>   conf binary: 18 failed, 40 passed; all failures are new regression tests.
> - ASan/UBSan at -O1, with leak detection disabled: the same 58 tests passed.
> - Both builds passed 1,635 input cases and 440 comparisons with the original
>   conf using valid inputs and fixed seeds. Malformed inputs preserved an
>   existing .config, including when KCONFIG_ALLCONFIG was set.
> - make defconfig, allnoconfig, allmodconfig and valid randconfig passed.
>   KCONFIG_PROBABILITY=50% changed from success to the expected error.
> - No vmlinux build or boot test; this changes the host configuration tool.
>   A 32-bit host build was unavailable because multilib headers were missing.
> 
> An AI coding assistant helped find the issue, prepare the fix, description
> and tests, and run validation. The requested task was to find and fix a
> useful, reproducible Linux bug suitable for a first contribution.
> 
>  scripts/kconfig/conf.c                        |  9 +++-
>  .../tests/randconfig_probability/Kconfig      | 12 +++++
>  .../tests/randconfig_probability/__init__.py  | 54 +++++++++++++++++++
>  3 files changed, 74 insertions(+), 1 deletion(-)
>  create mode 100644 scripts/kconfig/tests/randconfig_probability/Kconfig
>  create mode 100644 scripts/kconfig/tests/randconfig_probability/__init__.py
> 
> diff --git a/scripts/kconfig/conf.c b/scripts/kconfig/conf.c
> index fe8ba09b0..fa5dae74e 100644
> --- a/scripts/kconfig/conf.c
> +++ b/scripts/kconfig/conf.c
> @@ -191,7 +191,14 @@ static void conf_set_all_new_symbols(enum conf_def_mode mode)
>  		n = 0;
>  		while (env && *env) {
>  			char *endp;
> -			int tmp = strtol(env, &endp, 10);
> +			long tmp = strtol(env, &endp, 10);
> +
> +			if (endp == env || (*endp && *endp != ':') ||
> +			    (*endp == ':' && (!endp[1] || n == 2))) {
> +				errno = EINVAL;
> +				perror("KCONFIG_PROBABILITY");
> +				exit(1);

Unfortunately, sometimes users depend on unintended behavior. So I would
prefer to make this a warning for now, and then promote this to an error
later assuming nobody complains.

- Julian Braha