Re: [PATCH bpf] xdp: use flags field to disambiguate broadcast redirect

[Martin KaFai Lau <martin.lau@linux.dev>]

On 4/18/24 12:18 AM, Toke Høiland-Jørgensen wrote:
> When redirecting a packet using XDP, the bpf_redirect_map() helper will set
> up the redirect destination information in struct bpf_redirect_info (using
> the __bpf_xdp_redirect_map() helper function), and the xdp_do_redirect()
> function will read this information after the XDP program returns and pass
> the frame on to the right redirect destination.
> 
> When using the BPF_F_BROADCAST flag to do multicast redirect to a whole
> map, __bpf_xdp_redirect_map() sets the 'map' pointer in struct
> bpf_redirect_info to point to the destination map to be broadcast. And
> xdp_do_redirect() reacts to the value of this map pointer to decide whether
> it's dealing with a broadcast or a single-value redirect. However, if the
> destination map is being destroyed before xdp_do_redirect() is called, the
> map pointer will be cleared out (by bpf_clear_redirect_map()) without
> waiting for any XDP programs to stop running. This causes xdp_do_redirect()
> to think that the redirect was to a single target, but the target pointer
> is also NULL (since broadcast redirects don't have a single target), so
> this causes a crash when a NULL pointer is passed to dev_map_enqueue().
> 
> To fix this, change xdp_do_redirect() to react directly to the presence of
> the BPF_F_BROADCAST flag in the 'flags' value in struct bpf_redirect_info
> to disambiguate between a single-target and a broadcast redirect. And only
> read the 'map' pointer if the broadcast flag is set, aborting if that has
> been cleared out in the meantime. This prevents the crash, while keeping
> the atomic (cmpxchg-based) clearing of the map pointer itself, and without
> adding any more checks in the non-broadcast fast path.
> 
> Fixes: e624d4ed4aa8 ("xdp: Extend xdp_redirect_map with broadcast support")
> Reported-and-tested-by: syzbot+af9492708df9797198d6@syzkaller.appspotmail.com
> Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
> ---
>   net/core/filter.c | 42 ++++++++++++++++++++++++++++++++----------
>   1 file changed, 32 insertions(+), 10 deletions(-)
> 
> diff --git a/net/core/filter.c b/net/core/filter.c
> index 786d792ac816..8120c3dddf5e 100644
> --- a/net/core/filter.c
> +++ b/net/core/filter.c
> @@ -4363,10 +4363,12 @@ static __always_inline int __xdp_do_redirect_frame(struct bpf_redirect_info *ri,
>   	enum bpf_map_type map_type = ri->map_type;
>   	void *fwd = ri->tgt_value;
>   	u32 map_id = ri->map_id;
> +	u32 flags = ri->flags;
>   	struct bpf_map *map;
>   	int err;
>   
>   	ri->map_id = 0; /* Valid map id idr range: [1,INT_MAX[ */
> +	ri->flags = 0;
>   	ri->map_type = BPF_MAP_TYPE_UNSPEC;
>   
>   	if (unlikely(!xdpf)) {
> @@ -4378,11 +4380,20 @@ static __always_inline int __xdp_do_redirect_frame(struct bpf_redirect_info *ri,
>   	case BPF_MAP_TYPE_DEVMAP:
>   		fallthrough;
>   	case BPF_MAP_TYPE_DEVMAP_HASH:
> -		map = READ_ONCE(ri->map);
> -		if (unlikely(map)) {
> +		if (unlikely(flags & BPF_F_BROADCAST)) {
> +			map = READ_ONCE(ri->map);
> +
> +			/* The map pointer is cleared when the map is being torn
> +			 * down by bpf_clear_redirect_map()

Thanks for the details explanation in the commit message. All make sense.

It could be a dumb question.

 From reading the "waits for...NAPI being the relevant context here..." comment 
in dev_map_free(), I wonder if moving synchronize_rcu() before 
bpf_clear_redirect_map() would also work? Actually, does it need to call 
bpf_clear_redirect_map(). The on-going xdp_do_redirect() should be the last one 
using the map in ri->map anyway and no xdp prog can set it again to ri->map.

> +			 */
> +			if (unlikely(!map)) {
> +				err = -ENOENT;
> +				break;
> +			}
> +
>   			WRITE_ONCE(ri->map, NULL);
>   			err = dev_map_enqueue_multi(xdpf, dev, map,
> -						    ri->flags & BPF_F_EXCLUDE_INGRESS);
> +						    flags & BPF_F_EXCLUDE_INGRESS);
>   		} else {
>   			err = dev_map_enqueue(fwd, xdpf, dev);
>   		}