Re: [PATCH bpf-next] bpf: Disable -Wmissing-declarations for globally-linked kfuncs

[Anton Protopopov]

On Thu, Aug 17, 2023 at 04:35:26PM +0200, Daniel Borkmann wrote:
> On 8/17/23 6:01 AM, David Vernet wrote:
> > On Wed, Aug 16, 2023 at 08:48:16PM -0700, Alexei Starovoitov wrote:
> > > On Wed, Aug 16, 2023 at 8:38 PM Yonghong Song <yonghong.song@xxxxxxxxx> wrote:
> > > > On 8/16/23 8:06 AM, David Vernet wrote:
> > > > > We recently got an lkp warning about missing declarations, as in e.g.
> > > > > [0]. This warning is largely redundant with -Wmissing-prototypes, which
> > > > > we already disable for kfuncs that have global linkage and are meant to
> > > > > be exported in BTF, and called from BPF programs. Let's also disable
> > > > > -Wmissing-declarations for kfuncs. For what it's worth, I wasn't able to
> > > > > reproduce the warning even on W <= 3, so I can't actually be 100% sure
> > > > > this fixes the issue.
> > > > > 
> > > > > [0]: https://lore.kernel.org/all/202308162115.Hn23vv3n-lkp@xxxxxxxxx/
> > > > 
> > > > Okay, I just got a similar email to [0] which complains
> > > >     bpf_obj_new_impl, ..., bpf_cast_to_kern_ctx
> > > > missing declarations.
> > > > 
> > > > In the email, the used compiler is
> > > > compiler: gcc-7 (Ubuntu 7.5.0-6ubuntu2) 7.5.0
> > > > 
> > > > Unfortunately, I did not have gcc-7 to verify this.
> > > > Also, what is the minimum gcc version kernel supports? 5.1?
> > > 
> > > pahole and BTF might be broken in such old GCC too.
> > > Maybe we should add:
> > > config BPF_SYSCALL
> > >          depends on GCC_VERSION >= 90000 || CLANG_VERSION >= 130000
> > 
> > It seems prudent to formally declare minimum compiler versions. Though
> > modern gcc and clang also support -Wmissing-declarations, so maybe we
> > should merge this patch regardless? Just unfortunate to have to add even
> > more boilerplate just to get the compiler off our backs.
> 
> Urgh, to restrict BPF syscall with such `depends on` would be super ugly. Why
> can't we just move this boilerplate behind a macro instead of copying this
> everywhere? For example the below on top of your patch builds just fine on my
> side:
> 
> diff --git a/include/linux/btf.h b/include/linux/btf.h
> index df64cc642074..6a873a652001 100644
> --- a/include/linux/btf.h
> +++ b/include/linux/btf.h
> @@ -83,6 +83,16 @@
>   */
>  #define __bpf_kfunc __used noinline
> 
> +#define __bpf_kfunc_start	\
> +	__diag_push();	\
> +	__diag_ignore_all("-Wmissing-prototypes",	\
> +			  "Global functions as their definitions will be in vmlinux BTF");	\
> +	__diag_ignore_all("-Wmissing-declarations",	\
> +			  "Global functions as their definitions will be in vmlinux BTF");
> +

This will not solve the robot's compain, as it fails on gcc7. The
__diag_ignore_all for gcc is defined as

    #if GCC_VERSION >= 80000
    #define __diag_GCC_8(s)         __diag(s)
    #else
    #define __diag_GCC_8(s)
    #endif

    #define __diag_ignore_all(option, comment) \
            __diag_GCC(8, ignore, option)

so adding more __diag_ignore_all's will not do anything.  This is better to
patch __diag_ignore_all to include older gcc versions if anybody needs them.

> +#define __bpf_kfunc_end	\
> +	__diag_pop();
> +
>  /*
>   * Return the name of the passed struct, if exists, or halt the build if for
>   * example the structure gets renamed. In this way, developers have to revisit
> diff --git a/net/core/filter.c b/net/core/filter.c
> index c2b32b94c6bd..08dd0dd710dd 100644
> --- a/net/core/filter.c
> +++ b/net/core/filter.c
> @@ -11724,11 +11724,7 @@ bpf_sk_base_func_proto(enum bpf_func_id func_id)
>  	return func;
>  }
> 
> -__diag_push();
> -__diag_ignore_all("-Wmissing-prototypes",
> -		  "Global functions as their definitions will be in vmlinux BTF");
> -__diag_ignore_all("-Wmissing-declarations",
> -		  "Global functions as their definitions will be in vmlinux BTF");
> +__bpf_kfunc_start
>  __bpf_kfunc int bpf_dynptr_from_skb(struct sk_buff *skb, u64 flags,
>  				    struct bpf_dynptr_kern *ptr__uninit)
>  {
> @@ -11754,7 +11750,7 @@ __bpf_kfunc int bpf_dynptr_from_xdp(struct xdp_buff *xdp, u64 flags,
> 
>  	return 0;
>  }
> -__diag_pop();
> +__bpf_kfunc_end
> 
>  int bpf_dynptr_from_skb_rdonly(struct sk_buff *skb, u64 flags,
>  			       struct bpf_dynptr_kern *ptr__uninit)
> 
> Thanks,
> Daniel