[PATCHv2] net: bpf: reject invalid shifts

Eric Dumazet eric.dumazet at gmail.com
Tue Jan 12 11:48:38 PST 2016


On Tue, 2016-01-12 at 20:17 +0100, Rabin Vincent wrote:
> On ARM64, a BUG() is triggered in the eBPF JIT if a filter with a
> constant shift that can't be encoded in the immediate field of the
> UBFM/SBFM instructions is passed to the JIT.  Since these shifts
> amounts, which are negative or >= regsize, are invalid, reject them in
> the eBPF verifier and the classic BPF filter checker, for all
> architectures.
> 

Hmm...

> diff --git a/net/core/filter.c b/net/core/filter.c
> index 672eefbfbe99..37157c4c1a78 100644
> --- a/net/core/filter.c
> +++ b/net/core/filter.c
> @@ -777,6 +777,11 @@ static int bpf_check_classic(const struct sock_filter *filter,
>  			if (ftest->k == 0)
>  				return -EINVAL;
>  			break;
> +		case BPF_ALU | BPF_LSH | BPF_K:
> +		case BPF_ALU | BPF_RSH | BPF_K:
> +			if (ftest->k >= 32)
> +				return -EINVAL;
> +			break;
>  		case BPF_LD | BPF_MEM:
>  		case BPF_LDX | BPF_MEM:
>  		case BPF_ST:

These weak filters used to have undefined behavior, maybe in a never
taken branch, and will now fail hard, possibly breaking old
applications.

I believe we should add a one time warning to give a clue to poor users
hitting this problem.

Not everybody has perfect BPF filters, since most of the time they were
hand coded.






More information about the linux-arm-kernel mailing list