On Mon, 26 Oct 2015 18:23:07 -0700 (PDT), David Miller wrote:

From: Jakub Kicinski <redacted>
Date: Fri, 23 Oct 2015 19:58:11 +0100

quoted

+struct nfp_net_tx_buf {
+	struct sk_buff *skb;
+	dma_addr_t dma_addr;
+	short int fidx;
+	u16 pkt_cnt;
+	u32 real_len;
+};

This packs very poorly, and has a lot of padding holes.  Better ordering
would be:

struct nfp_net_tx_buf {
	struct sk_buff *skb;
	dma_addr_t dma_addr;
	u32 real_len;
	short int fidx;
	u16 pkt_cnt;
};

Seems to pack fine on x86:

struct nfp_net_tx_buf {
        struct sk_buff *           skb;                  /*     0     8 */
        dma_addr_t                 dma_addr;             /*     8     8 */
        short int                  fidx;                 /*    16     2 */
        u16                        pkt_cnt;              /*    18     2 */
        u32                        real_len;             /*    20     4 */

        /* size: 24, cachelines: 1, members: 5 */
        /* last cacheline: 24 bytes */
};

Are my packing skills deceiving me?  Maybe I'll try to build the driver
for more esoteric architectures.