On 7/22/25 10:25 AM, Petr Pavlu wrote:
> On 7/22/25 5:08 AM, Wang Jinchao wrote:
>> On 7/21/25 22:40, Petr Pavlu wrote:
>>> On 7/21/25 6:52 AM, Wang Jinchao wrote:
>>>> When there is no version information, modprobe and insmod only
>>>> report "invalid format".
>>>> Print the actual cause to make it easier to diagnose the issue.
>>>> This helps developers quickly identify version-related module
>>>> loading failures.
>>>> Signed-off-by: Wang Jinchao <wangjinchao...@gmail.com>
>>>> ---
>>>> kernel/module/version.c | 4 +++-
>>>> 1 file changed, 3 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/kernel/module/version.c b/kernel/module/version.c
>>>> index 2beefeba82d9..bc28c697ff3a 100644
>>>> --- a/kernel/module/version.c
>>>> +++ b/kernel/module/version.c
>>>> @@ -42,8 +42,10 @@ int check_version(const struct load_info *info,
>>>> }
>>>> /* No versions at all? modprobe --force does this. */
>>>> - if (versindex == 0)
>>>> + if (versindex == 0) {
>>>> + pr_debug("No version info for module %s\n", info->name);
>>>> return try_to_force_load(mod, symname) == 0;
>>>> + }
>>>> versions = (void *)sechdrs[versindex].sh_addr;
>>>> num_versions = sechdrs[versindex].sh_size
>>>
>>> I think it would be better to instead improve the behavior of
>>> try_to_force_load(). The function should print the error reason prior to
>>> returning -ENOEXEC. This would also help its two other callers,
>>> check_modinfo() and check_export_symbol_versions().
>>>
>>> Additionally, I suggest moving the check to ensure version information
>>> is available for imported symbols earlier in the loading process.
>>> A suitable place might be check_modstruct_version(). This way the check
>>> is performed only once per module.
>>>
>>> The following is a prototype patch:
>>>
>>> diff --git a/kernel/module/main.c b/kernel/module/main.c
>>> index c2c08007029d..c1ccd343e8c3 100644
>>> --- a/kernel/module/main.c
>>> +++ b/kernel/module/main.c
>>> @@ -1053,6 +1053,7 @@ int try_to_force_load(struct module *mod, const char
>>> *reason)
>>> add_taint_module(mod, TAINT_FORCED_MODULE, LOCKDEP_NOW_UNRELIABLE);
>>> return 0;
>>> #else
>>> + pr_err("%s: %s\n", mod->name, reason);
>>> return -ENOEXEC;
>>> #endif
>>> }
>>> diff --git a/kernel/module/version.c b/kernel/module/version.c
>>> index 2beefeba82d9..4d9ebf0834de 100644
>>> --- a/kernel/module/version.c
>>> +++ b/kernel/module/version.c
>>> @@ -41,9 +41,9 @@ int check_version(const struct load_info *info,
>>> return 1;
>>> }
>>> - /* No versions at all? modprobe --force does this. */
>>> + /* No versions? Ok, already tainted in check_modstruct_version(). */
>>> if (versindex == 0)
>>> - return try_to_force_load(mod, symname) == 0;
>>> + return 1;
>>> versions = (void *)sechdrs[versindex].sh_addr;
>>> num_versions = sechdrs[versindex].sh_size
>>> @@ -90,6 +90,11 @@ int check_modstruct_version(const struct load_info *info,
>>> have_symbol = find_symbol(&fsa);
>>> BUG_ON(!have_symbol);
>>> + /* No versions at all? modprobe --force does this. */
>>> + if (!info->index.vers && !info->index.vers_ext_crc)
>>> + return try_to_force_load(
>>> + mod, "no versions for imported symbols") == 0;
>>> +
>>> return check_version(info, "module_layout", mod, fsa.crc);
>>> }
>>>
>>> As a side note, I'm confused why with CONFIG_MODULE_FORCE_LOAD=y, the
>>> code treats missing modversions for imported symbols as ok, even without
>>> MODULE_INIT_IGNORE_MODVERSIONS. This is at least consistent with the
>>> handling of missing vermagic, but it seems this behavior should be
>>> stricter.
>>>
>> When debugging syzkaller, I noticed that the insmod command always reports
>> errors. However, to get the exact information, I need to trace the kernel,
>> so I casually submitted this patch.
>>
>> Based on your response, I also feel that the meaning of force_load here is
>> somewhat unclear. It would be better to create a mask or a clear list to
>> indicate which fields can be forced and which cannot. Once this is clear, we
>> can create a function named may_force_check().
>
> I cannot find an explicit reason in the Git history why a missing
> vermagic is treated as if the module was loaded with
> MODULE_INIT_IGNORE_VERMAGIC, and similarly why missing modversions data
> is treated as if the module was loaded with
> MODULE_INIT_IGNORE_MODVERSIONS.
>
> I would argue that a more sensible behavior would be to consider
> a missing vermagic as an error and allow loading the module only if
> MODULE_INIT_IGNORE_VERMAGIC is explicitly specified. And analogously for
> missing modversions and MODULE_INIT_IGNORE_MODVERSIONS.
>
> Nonetheless, if I understand correctly, this should be mostly separate
> from your issue.
To answer my own confusion, the thing that I missed is that the
MODULE_INIT_IGNORE_* flags are available only for the finit_module
syscall, not for init_module. In the case of init_module, the force
logic is handled by kmod in userspace by stripping the relevant
modversions and vermagic data. This means that when using init_module,
the module loader cannot distinguish between a module that lacks this
data and one that has it deliberately removed. When finit_module was
introduced in 2012, commit 2f3238aebedb ("module: add flags arg to
sys_finit_module()") added the MODULE_INIT_IGNORE_* flags, and they were
simply implemented to mirror the kmod behavior.
-- Petr
