Re: [net-next PATCHv2] bpf: Output error message to logbuf when loading failure

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Mon, Oct 26, 2015 at 07:13:08AM +0000, Wang Nan wrote:
> Many reasons can make bpf_prog_load() return EINVAL. This patch utilizes
> logbuf to deliver the actual reason of the failure.
> 
> Without this patch, it is very easy for user to pass an object with
> "version" section not match the kernel version code, and the problem
> is hard to determine from return code (EINVAL).
> 
> Signed-off-by: Wang Nan <wangnan0@huawei.com>
> Cc: Alexei Starovoitov <ast@kernel.org>
> Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
> Cc: David S. Miller <davem@davemloft.net>
> Cc: Wu Fengguang <fengguang.wu@intel.com>

NACK
for both implementation and design.

> +	/* Don't need care the copying result too much */
> +	WARN(copy_to_user(log_ubuf, log_buf, log_size),
> +	     KERN_WARNING "Failed to copy BPF error note '%s' to log buffer\n",
> +	     log_buf);

unprivilged user will be spamming kernel logs?!

> -	if (attr->insn_cnt >= BPF_MAXINSNS)
> +	if (attr->insn_cnt >= BPF_MAXINSNS) {
> +		bpf_prog_load_note(attr, "Too many instructions: %d > %d\n",
> +				   attr->insn_cnt, BPF_MAXINSNS);
>  		return -EINVAL;

if user space did that, it's wrong and can detect it
on its own.

>  	if (type == BPF_PROG_TYPE_KPROBE &&
> -	    attr->kern_version != LINUX_VERSION_CODE)
> +	    attr->kern_version != LINUX_VERSION_CODE) {
> +		bpf_prog_load_note(attr,
> +				   "Kernel version mismatch: 0x%x != 0x%x\n",
> +				   attr->kern_version, LINUX_VERSION_CODE);
>  		return -EINVAL;

user space (perf) could have checked that on its own
without kernel changes.

>  	/* find program type: socket_filter vs tracing_filter */
>  	err = find_prog_type(type, prog);
> -	if (err < 0)
> +	if (err < 0) {
> +		bpf_prog_load_note(attr, "Invalid program type: %d\n", type);
>  		goto free_prog;

same here.

In general syscalls muxing different error conditions into EINVAL
is a kernel wide problem and should be solved for all.