dev->needed_headroom += EXTRA_HEADER_LENGTH;

The basic logic is, when networking stack is about to allocate a SKB to transmit via a interface, it take into consideration the value of "needed_headroom" ( in fact, a macro LL_RESERVED_SPACE (dev ) ), and the SKB would have that large headroom reserved. This way, your NIC driver could assume that most packet reach into "do_start_xmit()" should have enough headroom --- most, not all, a headroom check is still needed.


http://elixir.free-electrons.com/linux/latest/source/include/linux/netdevice.h#L260

/* Reserve HH_DATA_MOD byte-aligned hard_header_len, but at least that much.
 * Alternative is:
 *   dev->hard_header_len ? (dev->hard_header_len +
 *                           (HH_DATA_MOD - 1)) & ~(HH_DATA_MOD - 1) : 0
 *
 * We could use other alignment values, but we must maintain the
 * relationship HH alignment <= LL alignment.
 */
#define LL_RESERVED_SPACE(dev) \
	((((dev)->hard_header_len+(dev)->needed_headroom)&~(HH_DATA_MOD - 1)) + HH_DATA_MOD)
#define LL_RESERVED_SPACE_EXTRA(dev,extra) \
	((((dev)->hard_header_len+(dev)->needed_headroom+(extra))&~(HH_DATA_MOD - 1)) + HH_DATA_MOD)


skb = __netdev_alloc_skb_ip_align(netdev, adapter->rx_ps_bsize0,
    gfp);


/**
 *	__netdev_alloc_skb - allocate an skbuff for rx on a specific device
 *	@dev: network device to receive on
 *	@len: length to allocate
 *	@gfp_mask: get_free_pages mask, passed to alloc_skb
 *
 *	Allocate a new &sk_buff and assign it a usage count of one. The
 *	buffer has NET_SKB_PAD headroom built in. Users should allocate
 *	the headroom they think they need without accounting for the
 *	built in space. The built in space is used for optimisations.
 *
 *	%NULL is returned if there is no free memory.
 */
struct sk_buff *__netdev_alloc_skb(struct net_device *dev, unsigned int len,
				   gfp_t gfp_mask)



/*
 * The networking layer reserves some headroom in skb data (via
 * dev_alloc_skb). This is used to avoid having to reallocate skb data when
 * the header has to grow. In the default case, if the header has to grow
 * 32 bytes or less we avoid the reallocation.
 *
 * Unfortunately this headroom changes the DMA alignment of the resulting
 * network packet. As for NET_IP_ALIGN, this unaligned DMA is expensive
 * on some architectures. An architecture can override this value,
 * perhaps setting it to a cacheline in size (since that will maintain
 * cacheline alignment of the DMA). It must be a power of 2.
 *
 * Various parts of the networking layer expect at least 32 bytes of
 * headroom, you should not reduce this.
 *
 * Using max(32, L1_CACHE_BYTES) makes sense (especially with RPS)
 * to reduce average number of cache lines per packet.
 * get_rps_cpus() for example only access one 64 bytes aligned block :
 * NET_IP_ALIGN(2) + ethernet_header(14) + IP_header(20/40) + ports(8)
 */
#ifndef NET_SKB_PAD
#define NET_SKB_PAD	max(32, L1_CACHE_BYTES)
#endif



/* NOTE: netdev_alloc_skb reserves 16 bytes, and typically NET_IP_ALIGN
 * means we reserve 2 more, this pushes us to allocate from the next
 * larger slab size.
 * i.e. RXBUFFER_2048 --> size-4096 slab
 * However with the new *_jumbo_rx* routines, jumbo receives will use
 * fragmented skbs
 */