Re: [PATCH] module: Taint the kernel when write-protecting ro_after_init fails

[Vlastimil Babka]

On 3/6/25 17:57, Luis Chamberlain wrote:
> + linux-mm since we're adding TAINT_BAD_PAGE
> 
> On Thu, Mar 06, 2025 at 11:36:55AM +0100, Petr Pavlu wrote:
>> In the unlikely case that setting ro_after_init data to read-only fails, it
>> is too late to cancel loading of the module. The loader then issues only
>> a warning about the situation. Given that this reduces the kernel's
>> protection, it was suggested to make the failure more visible by tainting
>> the kernel.
>> 
>> Allow TAINT_BAD_PAGE to be set per-module and use it in this case. The flag
>> is set in similar situations and has the following description in
>> Documentation/admin-guide/tainted-kernels.rst: "bad page referenced or some
>> unexpected page flags".
>> 
>> Adjust the warning that reports the failure to avoid references to internal
>> functions and to add information about the kernel being tainted, both to
>> match the style of other messages in the file. Additionally, merge the
>> message on a single line because checkpatch.pl recommends that for the
>> ability to grep for the string.
>> 
>> Suggested-by: Kees Cook <kees@xxxxxxxxxx>
>> Signed-off-by: Petr Pavlu <petr.pavlu@xxxxxxxx>
>> ---
>> I opted to use TAINT_BAD_PAGE for now because it seemed unnecessary to me
>> to introduce a new flag only for this specific case. However, if we end up
>> similarly checking set_memory_*() in the boot context, a separate flag
>> would be probably better.
>> ---
>>  kernel/module/main.c | 7 ++++---
>>  kernel/panic.c       | 2 +-
>>  2 files changed, 5 insertions(+), 4 deletions(-)
>> 
>> diff --git a/kernel/module/main.c b/kernel/module/main.c
>> index 1fb9ad289a6f..8f424a107b92 100644
>> --- a/kernel/module/main.c
>> +++ b/kernel/module/main.c
>> @@ -3030,10 +3030,11 @@ static noinline int do_init_module(struct module *mod)
>>  	rcu_assign_pointer(mod->kallsyms, &mod->core_kallsyms);
>>  #endif
>>  	ret = module_enable_rodata_ro_after_init(mod);
>> -	if (ret)
>> -		pr_warn("%s: module_enable_rodata_ro_after_init() returned %d, "
>> -			"ro_after_init data might still be writable\n",
>> +	if (ret) {
>> +		pr_warn("%s: write-protecting ro_after_init data failed with %d, the data might still be writable - tainting kernel\n",
>>  			mod->name, ret);
>> +		add_taint_module(mod, TAINT_BAD_PAGE, LOCKDEP_STILL_OK);
>> +	}
>>  
>>  	mod_tree_remove_init(mod);
>>  	module_arch_freeing_init(mod);
>> diff --git a/kernel/panic.c b/kernel/panic.c
>> index d8635d5cecb2..794c443bfb5c 100644
>> --- a/kernel/panic.c
>> +++ b/kernel/panic.c
>> @@ -497,7 +497,7 @@ const struct taint_flag taint_flags[TAINT_FLAGS_COUNT] = {
>>  	TAINT_FLAG(CPU_OUT_OF_SPEC,		'S', ' ', false),
>>  	TAINT_FLAG(FORCED_RMMOD,		'R', ' ', false),
>>  	TAINT_FLAG(MACHINE_CHECK,		'M', ' ', false),
>> -	TAINT_FLAG(BAD_PAGE,			'B', ' ', false),
>> +	TAINT_FLAG(BAD_PAGE,			'B', ' ', true),
>>  	TAINT_FLAG(USER,			'U', ' ', false),
>>  	TAINT_FLAG(DIE,				'D', ' ', false),
>>  	TAINT_FLAG(OVERRIDDEN_ACPI_TABLE,	'A', ' ', false),
> 
> Reviewed-by: Luis Chamberlain <mcgrof@xxxxxxxxxx>
> 
> For our needs this makes sense, however I am curious if TAINT_BAD_PAGE
> is too broadly generic, and also if we're going to add it, if there are
> other mm uses for such a thing.

I'm not sure BAD_PAGE is a good fit. If there was a new flag that meant "a
hardening measure failed", would that have other possible uses? The
semantics would be that the kernel self-protection was weakened wrt
expectations, even if not yet a corruption due to attack would be detected.
Some admins could opt-in to panic in such case anyway, etc. Any other
hardening features where such "failure to harden" is possible and could use
this too? Kees?

>   Luis
>