Re: [PATCH net-next v2 3/3] net: ethtool: prevent user from breaking devmem single-binding rule

[Stanislav Fomichev <stfomichev@gmail.com>]

On 09/11, Bobby Eshleman wrote:
> From: Bobby Eshleman <bobbyeshleman@meta.com>
> 
> Prevent the user from breaking devmem's single-binding rule by rejecting
> ethtool TCP/IP requests to modify or delete rules that will redirect a
> devmem socket to a queue with a different dmabuf binding. This is done
> in a "best effort" approach because not all steering rule types are
> validated.
> 
> If an ethtool_rxnfc flow steering rule evaluates true for:
> 
> 1) matching a devmem socket's ip addr
> 2) selecting a queue with a different dmabuf binding
> 3) is TCP/IP (v4 or v6)
> 
> ... then reject the ethtool_rxnfc request with -EBUSY to indicate a
> devmem socket is using the current rules that steer it to its dmabuf
> binding.
> 
> Non-TCP/IP rules are completely ignored, and if they do match a devmem
> flow then they can still break devmem sockets. For example, bytes 0 and
> 1 of L2 headers, etc... it is still unknown to me if these are possible
> to evaluate at the time of the ethtool call, and so are left to future
> work (or never, if not possible).
> 
> FLOW_RSS rules which guide flows to an RSS context are also not
> evaluated yet. This seems feasible, but the correct path towards
> retrieving the RSS context and scanning the queues for dmabuf bindings
> seems unclear and maybe overkill (re-use parts of ethtool_get_rxnfc?).
> 
> Signed-off-by: Bobby Eshleman <bobbyeshleman@meta.com>
> ---
>  include/net/sock.h  |   1 +
>  net/ethtool/ioctl.c | 144 ++++++++++++++++++++++++++++++++++++++++++++++++++++
>  net/ipv4/tcp.c      |   9 ++++
>  net/ipv4/tcp_ipv4.c |   6 +++
>  4 files changed, 160 insertions(+)
> 
> diff --git a/include/net/sock.h b/include/net/sock.h
> index 304aad494764..73a1ff59dcde 100644
> --- a/include/net/sock.h
> +++ b/include/net/sock.h
> @@ -579,6 +579,7 @@ struct sock {
>  		struct net_devmem_dmabuf_binding	*binding;
>  		atomic_t				*urefs;
>  	} sk_user_frags;
> +	struct list_head	sk_devmem_list;
>  
>  #if IS_ENABLED(CONFIG_PROVE_LOCKING) && IS_ENABLED(CONFIG_MODULES)
>  	struct module		*sk_owner;
> diff --git a/net/ethtool/ioctl.c b/net/ethtool/ioctl.c
> index 0b2a4d0573b3..99676ac9bbaa 100644
> --- a/net/ethtool/ioctl.c
> +++ b/net/ethtool/ioctl.c
> @@ -29,11 +29,16 @@
>  #include <linux/utsname.h>
>  #include <net/devlink.h>
>  #include <net/ipv6.h>
> +#include <net/netdev_rx_queue.h>
>  #include <net/xdp_sock_drv.h>
>  #include <net/flow_offload.h>
>  #include <net/netdev_lock.h>
>  #include <linux/ethtool_netlink.h>
>  #include "common.h"
> +#include "../core/devmem.h"
> +
> +extern struct list_head devmem_sockets_list;
> +extern spinlock_t devmem_sockets_lock;
>  
>  /* State held across locks and calls for commands which have devlink fallback */
>  struct ethtool_devlink_compat {
> @@ -1169,6 +1174,142 @@ ethtool_get_rxfh_fields(struct net_device *dev, u32 cmd, void __user *useraddr)
>  	return ethtool_rxnfc_copy_to_user(useraddr, &info, info_size, NULL);
>  }
>  
> +static bool
> +__ethtool_rx_flow_spec_breaks_devmem_sk(struct ethtool_rx_flow_spec *fs,
> +					struct net_device *dev,
> +					struct sock *sk)
> +{
> +	struct in6_addr saddr6, smask6, daddr6, dmask6;
> +	struct sockaddr_storage saddr, daddr;
> +	struct sockaddr_in6 *src6, *dst6;
> +	struct sockaddr_in *src4, *dst4;
> +	struct netdev_rx_queue *rxq;
> +	__u32 flow_type;
> +
> +	if (dev != __sk_dst_get(sk)->dev)
> +		return false;
> +
> +	src6 = (struct sockaddr_in6 *)&saddr;
> +	dst6 = (struct sockaddr_in6 *)&daddr;
> +	src4 = (struct sockaddr_in *)&saddr;
> +	dst4 = (struct sockaddr_in *)&daddr;
> +
> +	if (sk->sk_family == AF_INET6) {
> +		src6->sin6_port = inet_sk(sk)->inet_sport;
> +		src6->sin6_addr = inet6_sk(sk)->saddr;
> +		dst6->sin6_port = inet_sk(sk)->inet_dport;
> +		dst6->sin6_addr = sk->sk_v6_daddr;
> +	} else {
> +		src4->sin_port = inet_sk(sk)->inet_sport;
> +		src4->sin_addr.s_addr = inet_sk(sk)->inet_saddr;
> +		dst4->sin_port = inet_sk(sk)->inet_dport;
> +		dst4->sin_addr.s_addr = inet_sk(sk)->inet_daddr;
> +	}
> +
> +	flow_type = fs->flow_type & ~(FLOW_EXT | FLOW_MAC_EXT | FLOW_RSS);
> +
> +	rxq = __netif_get_rx_queue(dev, fs->ring_cookie);
> +	if (!rxq)
> +		return false;
> +
> +	/* If the requested binding and the sk binding is equal then we know
> +	 * this rule can't redirect to a different binding.
> +	 */
> +	if (rxq->mp_params.mp_priv == sk->sk_user_frags.binding)
> +		return false;
> +
> +	/* Reject rules that redirect RX devmem sockets to a queue with a
> +	 * different dmabuf binding. Because these sockets are on the RX side
> +	 * (registered in the recvmsg() path), we compare the opposite
> +	 * endpoints: the socket source with the rule destination, and the
> +	 * socket destination with the rule source.
> +	 *
> +	 * Only perform checks on the simplest rules to check, that is, IP/TCP
> +	 * rules. Flow hash options are not verified, so may still break TCP
> +	 * devmem flows in theory (VLAN tag, bytes 0 and 1 of L4 header,
> +	 * etc...). The author of this function was simply not sure how
> +	 * to validate these at the time of the ethtool call.
> +	 */
> +	switch (flow_type) {
> +	case IPV4_USER_FLOW: {
> +		const struct ethtool_usrip4_spec *v4_usr_spec, *v4_usr_m_spec;
> +
> +		v4_usr_spec = &fs->h_u.usr_ip4_spec;
> +		v4_usr_m_spec = &fs->m_u.usr_ip4_spec;
> +
> +		if (((v4_usr_spec->ip4src ^ dst4->sin_addr.s_addr) & v4_usr_m_spec->ip4src) ||
> +		    (v4_usr_spec->ip4dst ^ src4->sin_addr.s_addr) & v4_usr_m_spec->ip4dst) {
> +			return true;
> +		}
> +
> +		return false;
> +	}
> +	case TCP_V4_FLOW: {
> +		const struct ethtool_tcpip4_spec *v4_spec, *v4_m_spec;
> +
> +		v4_spec = &fs->h_u.tcp_ip4_spec;
> +		v4_m_spec = &fs->m_u.tcp_ip4_spec;
> +
> +		if (((v4_spec->ip4src ^ dst4->sin_addr.s_addr) & v4_m_spec->ip4src) ||
> +		    ((v4_spec->ip4dst ^ src4->sin_addr.s_addr) & v4_m_spec->ip4dst))
> +			return true;
> +

The ports need to be checked as well? But my preference overall would
be to go back to checking this condition during recvmsg. We can pick
some new obscure errno number to clearly explain to the user what
happened. EPIPE or something similar, to mean that the socket is cooked.
But let's see if Mina has a different opinion..