Re: [PATCH net-next 6/6] tsnep: Add XDP RX support

[Paolo Abeni <pabeni@redhat.com>]

On Sat, 2022-12-03 at 22:54 +0100, Gerhard Engleder wrote:
> If BPF program is set up, then run BPF program for every received frame
> and execute the selected action.
> 
> Test results with A53 1.2GHz:
> 
> XDP_DROP (samples/bpf/xdp1)
> proto 17:     865683 pkt/s
> 
> XDP_TX (samples/bpf/xdp2)
> proto 17:     253594 pkt/s
> 
> XDP_REDIRECT (samples/bpf/xdpsock)
>  sock0@eth2:0 rxdrop xdp-drv
>                    pps            pkts           1.00
> rx                 862,258        4,514,166
> tx                 0              0
> 
> XDP_REDIRECT (samples/bpf/xdp_redirect)
> eth2->eth1         608,895 rx/s   0 err,drop/s   608,895 xmit/s
> 
> Signed-off-by: Gerhard Engleder <gerhard@engleder-embedded.com>
> ---
>  drivers/net/ethernet/engleder/tsnep_main.c | 100 +++++++++++++++++++++
>  1 file changed, 100 insertions(+)
> 
> diff --git a/drivers/net/ethernet/engleder/tsnep_main.c b/drivers/net/ethernet/engleder/tsnep_main.c
> index 725b2a1e7be4..4e3c6bd3dc9f 100644
> --- a/drivers/net/ethernet/engleder/tsnep_main.c
> +++ b/drivers/net/ethernet/engleder/tsnep_main.c
> @@ -27,6 +27,7 @@
>  #include <linux/phy.h>
>  #include <linux/iopoll.h>
>  #include <linux/bpf.h>
> +#include <linux/bpf_trace.h>
>  
>  #define TSNEP_SKB_PAD (NET_SKB_PAD + NET_IP_ALIGN)
>  #define TSNEP_HEADROOM ALIGN(max(TSNEP_SKB_PAD, XDP_PACKET_HEADROOM), 4)
> @@ -44,6 +45,11 @@
>  #define TSNEP_COALESCE_USECS_MAX     ((ECM_INT_DELAY_MASK >> ECM_INT_DELAY_SHIFT) * \
>  				      ECM_INT_DELAY_BASE_US + ECM_INT_DELAY_BASE_US - 1)
>  
> +#define TSNEP_XDP_PASS		0
> +#define TSNEP_XDP_CONSUMED	BIT(0)
> +#define TSNEP_XDP_TX		BIT(1)
> +#define TSNEP_XDP_REDIRECT	BIT(2)
> +
>  enum {
>  	__TSNEP_DOWN,
>  };
> @@ -819,6 +825,11 @@ static inline unsigned int tsnep_rx_offset(struct tsnep_rx *rx)
>  	return TSNEP_SKB_PAD;
>  }
>  
> +static inline unsigned int tsnep_rx_offset_xdp(void)

Please, no 'inline' in c files, the complier will do a better job
without.

> +{
> +	return XDP_PACKET_HEADROOM;
> +}
> +
>  static void tsnep_rx_ring_cleanup(struct tsnep_rx *rx)
>  {
>  	struct device *dmadev = rx->adapter->dmadev;
> @@ -1024,6 +1035,65 @@ static int tsnep_rx_refill(struct tsnep_rx *rx, int count, bool reuse)
>  	return i;
>  }
>  
> +static int tsnep_xdp_run_prog(struct tsnep_rx *rx, struct bpf_prog *prog,
> +			      struct xdp_buff *xdp)
> +{
> +	unsigned int length;
> +	unsigned int sync;
> +	u32 act;
> +
> +	length = xdp->data_end - xdp->data_hard_start - tsnep_rx_offset_xdp();
> +
> +	act = bpf_prog_run_xdp(prog, xdp);
> +
> +	/* Due xdp_adjust_tail: DMA sync for_device cover max len CPU touch */
> +	sync = xdp->data_end - xdp->data_hard_start - tsnep_rx_offset_xdp();
> +	sync = max(sync, length);
> +
> +	switch (act) {
> +	case XDP_PASS:
> +		return TSNEP_XDP_PASS;
> +	case XDP_TX:
> +		if (tsnep_xdp_xmit_back(rx->adapter, xdp) < 0)
> +			goto out_failure;
> +		return TSNEP_XDP_TX;
> +	case XDP_REDIRECT:
> +		if (xdp_do_redirect(rx->adapter->netdev, xdp, prog) < 0)
> +			goto out_failure;
> +		return TSNEP_XDP_REDIRECT;
> +	default:
> +		bpf_warn_invalid_xdp_action(rx->adapter->netdev, prog, act);
> +		fallthrough;
> +	case XDP_ABORTED:
> +out_failure:
> +		trace_xdp_exception(rx->adapter->netdev, prog, act);
> +		fallthrough;
> +	case XDP_DROP:
> +		page_pool_put_page(rx->page_pool, virt_to_head_page(xdp->data),
> +				   sync, true);
> +		return TSNEP_XDP_CONSUMED;
> +	}
> +}
> +
> +static void tsnep_finalize_xdp(struct tsnep_adapter *adapter, int status)
> +{
> +	int cpu = smp_processor_id();
> +	int queue;
> +	struct netdev_queue *nq;
> +
> +	if (status & TSNEP_XDP_TX) {
> +		queue = cpu % adapter->num_tx_queues;
> +		nq = netdev_get_tx_queue(adapter->netdev, queue);
> +
> +		__netif_tx_lock(nq, cpu);
> +		tsnep_xdp_xmit_flush(&adapter->tx[queue]);
> +		__netif_tx_unlock(nq);
> +	}
> +
> +	if (status & TSNEP_XDP_REDIRECT)
> +		xdp_do_flush();
> +}
> +
>  static struct sk_buff *tsnep_build_skb(struct tsnep_rx *rx, struct page *page,
>  				       int length)
>  {
> @@ -1062,12 +1132,17 @@ static int tsnep_rx_poll(struct tsnep_rx *rx, struct napi_struct *napi,
>  	int desc_available;
>  	int done = 0;
>  	enum dma_data_direction dma_dir;
> +	struct bpf_prog *prog;
>  	struct tsnep_rx_entry *entry;
> +	struct xdp_buff xdp;
> +	int xdp_status = 0;
>  	struct sk_buff *skb;
>  	int length;
> +	int retval;
>  
>  	desc_available = tsnep_rx_desc_available(rx);
>  	dma_dir = page_pool_get_dma_dir(rx->page_pool);
> +	prog = READ_ONCE(rx->adapter->xdp_prog);
>  
>  	while (likely(done < budget) && (rx->read != rx->write)) {
>  		entry = &rx->entry[rx->read];
> @@ -1111,6 +1186,28 @@ static int tsnep_rx_poll(struct tsnep_rx *rx, struct napi_struct *napi,
>  		rx->read = (rx->read + 1) % TSNEP_RING_SIZE;
>  		desc_available++;
>  
> +		if (prog) {
> +			xdp_init_buff(&xdp, PAGE_SIZE, &rx->xdp_rxq);
> +			xdp_prepare_buff(&xdp, page_address(entry->page),
> +					 tsnep_rx_offset_xdp() + TSNEP_RX_INLINE_METADATA_SIZE,
> +					 length - TSNEP_RX_INLINE_METADATA_SIZE,
> +					 false);
> +			retval = tsnep_xdp_run_prog(rx, prog, &xdp);
> +		} else {
> +			retval = TSNEP_XDP_PASS;
> +		}
> +		if (retval) {
> +			if (retval & (TSNEP_XDP_TX | TSNEP_XDP_REDIRECT))
> +				xdp_status |= retval;

Here you could avoid a couple of conditionals passing xdp_status as an
additional tsnep_xdp_run_prog() argument, let the latter update it
under the existing switch, returning a single consumed/pass up bool and
testing such value just after tsnep_xdp_run_prog().

Mostly a matter of personal tasted I guess.


Cheers,

Paolo