Is it safe for a NIC driver to use all the 48 bytes of skb->cb?

From: Dexuan Cui
Date: Sat Feb 15 2020 - 00:24:02 EST


Hi,
It looks all the layers of drivers among the network stack can use the 48-byte
skb->cb array. Is there any rule how they should coordinate with each other?

I noticed the last 16 bytes are used by struct skb_gso_cb:

include/linux/skbuff.h:
struct skb_gso_cb {
union {
int mac_offset;
int data_offset;
};
int encap_level;
__wsum csum;
__u16 csum_start;
};
#define SKB_SGO_CB_OFFSET 32
#define SKB_GSO_CB(skb) ((struct skb_gso_cb *)((skb)->cb + SKB_SGO_CB_OFFSET))

Does this mean a low level NIC driver (e.g. hv_netvsc) should only use
the first 32 bytes? What if the upper layer network stack starts to take up
more space in the future?

Now hv_netvsc assumes it can use all of the 48-bytes, though it uses only
20 bytes, but just in case the struct hv_netvsc_packet grows to >32 bytes in the
future, should we change the BUILD_BUG_ON() in netvsc_start_xmit() to
BUILD_BUG_ON(sizeof(struct hv_netvsc_packet) > SKB_SGO_CB_OFFSET);
?

struct hv_netvsc_packet {
/* Bookkeeping stuff */
u8 cp_partial; /* partial copy into send buffer */

u8 rmsg_size; /* RNDIS header and PPI size */
u8 rmsg_pgcnt; /* page count of RNDIS header and PPI */
u8 page_buf_cnt;

u16 q_idx;
u16 total_packets;

u32 total_bytes;
u32 send_buf_index;
u32 total_data_buflen;
};

static int netvsc_start_xmit(struct sk_buff *skb, struct net_device *net)
{
...
/*
* Place the rndis header in the skb head room and
* the skb->cb will be used for hv_netvsc_packet
* structure.
*/
ret = skb_cow_head(skb, RNDIS_AND_PPI_SIZE);
if (ret)
goto no_memory;

/* Use the skb control buffer for building up the packet */
BUILD_BUG_ON(sizeof(struct hv_netvsc_packet) >
FIELD_SIZEOF(struct sk_buff, cb));
packet = (struct hv_netvsc_packet *)skb->cb;

Thanks,
-- Dexuan