Re: [PATCH] [RFC] xfrm: work around a clang-19 fortifiy-string false-positive

["Arnd Bergmann" <arnd@arndb.de>]

On Fri, Feb 16, 2024, at 22:19, Kees Cook wrote:
> On Fri, Feb 16, 2024 at 09:26:40PM +0100, Arnd Bergmann wrote:
>> From: Arnd Bergmann <arnd@arndb.de>
>> 
>> clang-19 recently got branched from clang-18 and is not yet released.
>> The current version introduces exactly one new warning that I came
>> across in randconfig testing, in the copy_to_user_tmpl() function:
>> 
>> include/linux/fortify-string.h:420:4: error: call to '__write_overflow_field' declared with 'warning' attribute: detected write beyond size of field (1st parameter); maybe use struct_group()? [-Werror,-Wattribute-warning]
>>   420 |                         __write_overflow_field(p_size_field, size);
>> 
>> I have not yet produced a minimized test case for it, but I have a
>> local workaround, which avoids the memset() here by replacing it with
>> an initializer.
>> 
>> The memset is required to avoid leaking stack data to user space
>> and was added in commit 1f86840f8977 ("xfrm_user: fix info leak in
>> copy_to_user_tmpl()"). Simply changing the initializer to set all fields
>> still risks leaking data in the padding between them, which the compiler
>> is free to do here. To work around that problem, explicit padding fields
>> have to get added as well.
>
> Per C11, padding bits are zero initialized if there is an initializer,
> so "= { }" here should be sufficient -- no need to add the struct
> members.
>
>> Since this is a false positive, a better fix would likely be to
>> fix the compiler.
>
> As Nathan has found, this appears to be a loop unrolling bug in Clang.
> https://github.com/ClangBuiltLinux/linux/issues/1985
>
> The shorter fix (in the issue) is to explicitly range-check before
> the loop:
>
>        if (xp->xfrm_nr > XFRM_MAX_DEPTH)
>                return -ENOBUFS;

I ran into this issue again and I see that Nathan's fix has
made it into mainline and backports, but it's apparently
not sufficient.

I don't see the warning with my patch from this thread, but
there may still be a better fix.

      Arnd