Re: [PATCH v5 bpf-next 1/3] netfilter: nf_tables: add flowtable map for xdp offload

[Daniel Borkmann <daniel@xxxxxxxxxxxxx>]

On 6/14/24 5:40 PM, Lorenzo Bianconi wrote:
[...]
> diff --git a/net/netfilter/nf_flow_table_offload.c b/net/netfilter/nf_flow_table_offload.c
> index a010b25076ca0..d9b019c98694b 100644
> --- a/net/netfilter/nf_flow_table_offload.c
> +++ b/net/netfilter/nf_flow_table_offload.c
> @@ -1192,7 +1192,7 @@ int nf_flow_table_offload_setup(struct nf_flowtable *flowtable,
>   	int err;
>   
>   	if (!nf_flowtable_hw_offload(flowtable))
> -		return 0;
> +		return nf_flow_offload_xdp_setup(flowtable, dev, cmd);
>   
>   	if (dev->netdev_ops->ndo_setup_tc)
>   		err = nf_flow_table_offload_cmd(&bo, flowtable, dev, cmd,
> @@ -1200,8 +1200,10 @@ int nf_flow_table_offload_setup(struct nf_flowtable *flowtable,
>   	else
>   		err = nf_flow_table_indr_offload_cmd(&bo, flowtable, dev, cmd,
>   						     &extack);
> -	if (err < 0)
> +	if (err < 0) {
> +		nf_flow_offload_xdp_cancel(flowtable, dev, cmd);
>   		return err;
> +	}
>   
>   	return nf_flow_table_block_setup(flowtable, &bo, cmd);
>   }
> diff --git a/net/netfilter/nf_flow_table_xdp.c b/net/netfilter/nf_flow_table_xdp.c
> new file mode 100644
> index 0000000000000..b9bdf27ba9bd3
> --- /dev/null
> +++ b/net/netfilter/nf_flow_table_xdp.c
> @@ -0,0 +1,163 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +#include <linux/kernel.h>
> +#include <linux/module.h>
> +#include <linux/netfilter.h>
> +#include <linux/rhashtable.h>
> +#include <linux/netdevice.h>
> +#include <net/flow_offload.h>
> +#include <net/netfilter/nf_flow_table.h>
> +
> +struct flow_offload_xdp_ft {
> +	struct list_head head;
> +	struct nf_flowtable *ft;
> +	struct rcu_head rcuhead;
> +};
> +
> +struct flow_offload_xdp {
> +	struct hlist_node hnode;
> +	unsigned long net_device_addr;
> +	struct list_head head;
> +};
> +
> +#define NF_XDP_HT_BITS	4
> +static DEFINE_HASHTABLE(nf_xdp_hashtable, NF_XDP_HT_BITS);
> +static DEFINE_MUTEX(nf_xdp_hashtable_lock);
> +
> +/* caller must hold rcu read lock */
> +struct nf_flowtable *nf_flowtable_by_dev(const struct net_device *dev)
> +{
> +	unsigned long key = (unsigned long)dev;
> +	struct flow_offload_xdp *iter;
> +
> +	hash_for_each_possible_rcu(nf_xdp_hashtable, iter, hnode, key) {
> +		if (key == iter->net_device_addr) {
> +			struct flow_offload_xdp_ft *ft_elem;
> +
> +			/* The user is supposed to insert a given net_device
> +			 * just into a single nf_flowtable so we always return
> +			 * the first element here.
> +			 */
> +			ft_elem = list_first_or_null_rcu(&iter->head,
> +							 struct flow_offload_xdp_ft,
> +							 head);
> +			return ft_elem ? ft_elem->ft : NULL;
> +		}
> +	}
> +
> +	return NULL;
> +}
> +
> +static int nf_flowtable_by_dev_insert(struct nf_flowtable *ft,
> +				      const struct net_device *dev)
> +{
> +	struct flow_offload_xdp *iter, *elem = NULL;
> +	unsigned long key = (unsigned long)dev;
> +	struct flow_offload_xdp_ft *ft_elem;
> +
> +	ft_elem = kzalloc(sizeof(*ft_elem), GFP_KERNEL_ACCOUNT);
> +	if (!ft_elem)
> +		return -ENOMEM;
> +
> +	ft_elem->ft = ft;
> +
> +	mutex_lock(&nf_xdp_hashtable_lock);
> +
> +	hash_for_each_possible(nf_xdp_hashtable, iter, hnode, key) {
> +		if (key == iter->net_device_addr) {
> +			elem = iter;
> +			break;
> +		}
> +	}
> +
> +	if (!elem) {
> +		elem = kzalloc(sizeof(*elem), GFP_KERNEL_ACCOUNT);
> +		if (!elem)
> +			goto err_unlock;
> +
> +		elem->net_device_addr = key;

Looks good, as I understand (but just to double check) if a device goes away then
upper layers in the nf flowtable code will trigger the nf_flowtable_by_dev_remove()
based on the device pointer to clean this up again from nf_xdp_hashtable.

> +		INIT_LIST_HEAD(&elem->head);
> +		hash_add_rcu(nf_xdp_hashtable, &elem->hnode, key);
> +	}
> +	list_add_tail_rcu(&ft_elem->head, &elem->head);
> +
> +	mutex_unlock(&nf_xdp_hashtable_lock);
> +
> +	return 0;
> +
> +err_unlock:
> +	mutex_unlock(&nf_xdp_hashtable_lock);
> +	kfree(ft_elem);
> +
> +	return -ENOMEM;
> +}