Date:   Tue, 14 Feb 2023 01:46:51 +0100
From:   Maciej Fijalkowski <maciej.fijalkowski@...el.com>
To:     Praveen Kaligineedi <pkaligineedi@...gle.com>
CC:     <netdev@...r.kernel.org>, <davem@...emloft.net>, <kuba@...nel.org>,
        Jeroen de Borst <jeroendb@...gle.com>
Subject: Re: [PATCH net-next 2/4] gve: Add XDP DROP and TX support for
 GQI-QPL format

On Tue, Feb 07, 2023 at 01:00:56PM -0800, Praveen Kaligineedi wrote:
> Add support for XDP PASS, DROP and TX actions.
> 

Hi, some questions below:

> This patch contains the following changes:
> 1) Support installing/uninstalling XDP program
> 2) Add dedicated XDP TX queues
> 3) Add support for XDP DROP action
> 4) Add support for XDP TX action
> 
> Signed-off-by: Praveen Kaligineedi <pkaligineedi@...gle.com>
> Reviewed-by: Jeroen de Borst <jeroendb@...gle.com>
> ---
>  drivers/net/ethernet/google/gve/gve.h         |  30 +++-
>  drivers/net/ethernet/google/gve/gve_ethtool.c |  37 ++++-
>  drivers/net/ethernet/google/gve/gve_main.c    | 145 +++++++++++++++++-
>  drivers/net/ethernet/google/gve/gve_rx.c      |  73 ++++++++-
>  drivers/net/ethernet/google/gve/gve_tx.c      | 130 +++++++++++++++-
>  5 files changed, 399 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/net/ethernet/google/gve/gve.h b/drivers/net/ethernet/google/gve/gve.h
> index 770e87e3ec5e..8352f4c0e8d1 100644
> --- a/drivers/net/ethernet/google/gve/gve.h
> +++ b/drivers/net/ethernet/google/gve/gve.h
> @@ -47,6 +47,10 @@
>  
>  #define GVE_RX_BUFFER_SIZE_DQO 2048
>  
> +#define GVE_XDP_ACTIONS 5
> +
> +#define GVE_XDP_PADDING 64
> +
>  /* Each slot in the desc ring has a 1:1 mapping to a slot in the data ring */
>  struct gve_rx_desc_queue {
>  	struct gve_rx_desc *desc_ring; /* the descriptor ring */
> @@ -230,7 +234,9 @@ struct gve_rx_ring {
>  	u64 rx_frag_flip_cnt; /* free-running count of rx segments where page_flip was used */
>  	u64 rx_frag_copy_cnt; /* free-running count of rx segments copied */
>  	u64 rx_frag_alloc_cnt; /* free-running count of rx page allocations */
> -
> +	u64 xdp_tx_errors;
> +	u64 xdp_redirect_errors;
> +	u64 xdp_actions[GVE_XDP_ACTIONS];
>  	u32 q_num; /* queue index */
>  	u32 ntfy_id; /* notification block index */
>  	struct gve_queue_resources *q_resources; /* head and tail pointer idx */
> @@ -238,6 +244,9 @@ struct gve_rx_ring {
>  	struct u64_stats_sync statss; /* sync stats for 32bit archs */
>  
>  	struct gve_rx_ctx ctx; /* Info for packet currently being processed in this ring. */
> +
> +	/* XDP stuff */
> +	struct xdp_rxq_info xdp_rxq;
>  };
>  
>  /* A TX desc ring entry */
> @@ -259,6 +268,9 @@ struct gve_tx_iovec {
>   */
>  struct gve_tx_buffer_state {
>  	struct sk_buff *skb; /* skb for this pkt */
> +	struct {
> +		u16 size; /* size of xmitted xdp pkt */
> +	} xdp;
>  	union {
>  		struct gve_tx_iovec iov[GVE_TX_MAX_IOVEC]; /* segments of this pkt */
>  		struct {
> @@ -373,6 +385,8 @@ struct gve_tx_ring {
>  		struct {
>  			/* Spinlock for when cleanup in progress */
>  			spinlock_t clean_lock;
> +			/* Spinlock for XDP tx traffic */
> +			spinlock_t xdp_lock;

could you explain why locking is mandatory for you when in commit message
you're mentioning that these queue will be dedicated for XDP_TX handling?

>  		};
>  
>  		/* DQO fields. */
> @@ -526,9 +540,11 @@ struct gve_priv {
>  	u16 rx_data_slot_cnt; /* rx buffer length */
>  	u64 max_registered_pages;
>  	u64 num_registered_pages; /* num pages registered with NIC */
> +	struct bpf_prog __rcu *xdp_prog; /* XDP BPF program */
>  	u32 rx_copybreak; /* copy packets smaller than this */
>  	u16 default_num_queues; /* default num queues to set up */
>  
> +	u16 num_xdp_queues;
>  	struct gve_queue_config tx_cfg;
>  	struct gve_queue_config rx_cfg;
>  	struct gve_qpl_config qpl_cfg; /* map used QPL ids */
> @@ -785,7 +801,7 @@ static inline u32 gve_num_tx_qpls(struct gve_priv *priv)
>  	if (priv->queue_format != GVE_GQI_QPL_FORMAT)
>  		return 0;
>  
> -	return priv->tx_cfg.num_queues;
> +	return priv->tx_cfg.num_queues + priv->num_xdp_queues;
>  }
>  
>  /* Returns the number of rx queue page lists
> @@ -857,7 +873,12 @@ static inline bool gve_is_gqi(struct gve_priv *priv)
>  
>  static inline int gve_num_tx_queues(struct gve_priv *priv)
>  {
> -	return priv->tx_cfg.num_queues;
> +	return priv->tx_cfg.num_queues + priv->num_xdp_queues;
> +}
> +
> +static inline int gve_xdp_tx_queue_id(struct gve_priv *priv, u32 queue_id)
> +{
> +	return priv->tx_cfg.num_queues + queue_id;
>  }
>  
>  /* buffers */
> @@ -868,7 +889,10 @@ void gve_free_page(struct device *dev, struct page *page, dma_addr_t dma,
>  		   enum dma_data_direction);
>  /* tx handling */
>  netdev_tx_t gve_tx(struct sk_buff *skb, struct net_device *dev);
> +int gve_xdp_xmit_one(struct gve_priv *priv, struct gve_tx_ring *tx,
> +		     void *data, int len, u32 flags);
>  bool gve_tx_poll(struct gve_notify_block *block, int budget);
> +bool gve_xdp_poll(struct gve_notify_block *block, int budget);
>  int gve_tx_alloc_rings(struct gve_priv *priv);
>  void gve_tx_free_rings_gqi(struct gve_priv *priv);
>  u32 gve_tx_load_event_counter(struct gve_priv *priv,
> diff --git a/drivers/net/ethernet/google/gve/gve_ethtool.c b/drivers/net/ethernet/google/gve/gve_ethtool.c
> index a27cc314e799..d2f0b53eacbb 100644
> --- a/drivers/net/ethernet/google/gve/gve_ethtool.c
> +++ b/drivers/net/ethernet/google/gve/gve_ethtool.c
> @@ -34,6 +34,11 @@ static u32 gve_get_msglevel(struct net_device *netdev)
>  	return priv->msg_enable;
>  }
>  
> +/* For the following stats column string names, make sure the order
> + * matches how it is filled in the code. For xdp_aborted, xdp_drop,
> + * xdp_pass, xdp_tx, xdp_redirect, make sure it also matches the order
> + * as declared in enum xdp_action inside file uapi/linux/bpf.h .
> + */
>  static const char gve_gstrings_main_stats[][ETH_GSTRING_LEN] = {
>  	"rx_packets", "tx_packets", "rx_bytes", "tx_bytes",
>  	"rx_dropped", "tx_dropped", "tx_timeouts",
> @@ -49,6 +54,9 @@ static const char gve_gstrings_rx_stats[][ETH_GSTRING_LEN] = {
>  	"rx_dropped_pkt[%u]", "rx_copybreak_pkt[%u]", "rx_copied_pkt[%u]",
>  	"rx_queue_drop_cnt[%u]", "rx_no_buffers_posted[%u]",
>  	"rx_drops_packet_over_mru[%u]", "rx_drops_invalid_checksum[%u]",
> +	"rx_xdp_aborted[%u]", "rx_xdp_drop[%u]", "rx_xdp_pass[%u]",
> +	"rx_xdp_tx[%u]", "rx_xdp_redirect[%u]",
> +	"rx_xdp_tx_errors[%u]", "rx_xdp_redirect_errors[%u]",
>  };
>  
>  static const char gve_gstrings_tx_stats[][ETH_GSTRING_LEN] = {
> @@ -289,14 +297,24 @@ gve_get_ethtool_stats(struct net_device *netdev,
>  			if (skip_nic_stats) {
>  				/* skip NIC rx stats */
>  				i += NIC_RX_STATS_REPORT_NUM;
> -				continue;
> -			}
> -			for (j = 0; j < NIC_RX_STATS_REPORT_NUM; j++) {
> -				u64 value =
> -				be64_to_cpu(report_stats[rx_qid_to_stats_idx[ring] + j].value);
> -
> -				data[i++] = value;
> +			} else {
> +				stats_idx = rx_qid_to_stats_idx[ring];
> +				for (j = 0; j < NIC_RX_STATS_REPORT_NUM; j++) {
> +					u64 value =
> +						be64_to_cpu(report_stats[stats_idx + j].value);
> +					data[i++] = value;
> +				}
>  			}
> +			/* XDP rx counters */
> +			do {
> +				start =	u64_stats_fetch_begin(&priv->rx[ring].statss);
> +				for (j = 0; j < GVE_XDP_ACTIONS; j++)
> +					data[i + j] = rx->xdp_actions[j];
> +				data[i + j++] = rx->xdp_tx_errors;
> +				data[i + j++] = rx->xdp_redirect_errors;
> +			} while (u64_stats_fetch_retry(&priv->rx[ring].statss,
> +						       start));
> +			i += GVE_XDP_ACTIONS + 2; /* XDP rx counters */
>  		}
>  	} else {
>  		i += priv->rx_cfg.num_queues * NUM_GVE_RX_CNTS;
> @@ -418,6 +436,11 @@ static int gve_set_channels(struct net_device *netdev,
>  	if (!new_rx || !new_tx)
>  		return -EINVAL;
>  
> +	if (priv->xdp_prog &&
> +	    (new_tx != new_rx || (2 * new_tx > priv->tx_cfg.max_queues))) {

How do I know from user POV that here i have to have the same amount of rx
and tx when xdp prog is running? I thought about adding some netdev_err()
here at least but you seem to not be using that throughout whole
gve_ethtoool.c...

> +		return -EINVAL;
> +	}
> +
>  	if (!netif_carrier_ok(netdev)) {
>  		priv->tx_cfg.num_queues = new_tx;
>  		priv->rx_cfg.num_queues = new_rx;
> diff --git a/drivers/net/ethernet/google/gve/gve_main.c b/drivers/net/ethernet/google/gve/gve_main.c
> index e69d21facd29..5050aa3aa1c3 100644
> --- a/drivers/net/ethernet/google/gve/gve_main.c
> +++ b/drivers/net/ethernet/google/gve/gve_main.c
> @@ -4,8 +4,10 @@
>   * Copyright (C) 2015-2021 Google, Inc.
>   */
>  
> +#include <linux/bpf.h>
>  #include <linux/cpumask.h>
>  #include <linux/etherdevice.h>
> +#include <linux/filter.h>
>  #include <linux/interrupt.h>
>  #include <linux/module.h>
>  #include <linux/pci.h>
> @@ -247,8 +249,13 @@ static int gve_napi_poll(struct napi_struct *napi, int budget)
>  	block = container_of(napi, struct gve_notify_block, napi);
>  	priv = block->priv;
>  
> -	if (block->tx)
> -		reschedule |= gve_tx_poll(block, budget);
> +	if (block->tx) {
> +		if (block->tx->q_num < priv->tx_cfg.num_queues)
> +			reschedule |= gve_tx_poll(block, budget);
> +		else
> +			reschedule |= gve_xdp_poll(block, budget);
> +	}
> +
>  	if (block->rx) {
>  		work_done = gve_rx_poll(block, budget);
>  		reschedule |= work_done == budget;
> @@ -974,11 +981,50 @@ static int gve_reset_recovery(struct gve_priv *priv, bool was_up);
>  static void gve_turndown(struct gve_priv *priv);
>  static void gve_turnup(struct gve_priv *priv);
>  
> +static int gve_reg_xdp_info(struct gve_priv *priv, struct net_device *dev)
> +{
> +	int err = 0, i;
> +
> +	for (i = 0; i < priv->rx_cfg.num_queues; i++) {
> +		struct gve_rx_ring *rx = &priv->rx[i];
> +		struct napi_struct *napi =
> +			&priv->ntfy_blocks[rx->ntfy_id].napi;
> +
> +		err = xdp_rxq_info_reg(&rx->xdp_rxq, dev, i,
> +				       napi->napi_id);
> +		if (err)
> +			goto out;
> +		err = xdp_rxq_info_reg_mem_model(&rx->xdp_rxq,
> +						 MEM_TYPE_PAGE_SHARED, NULL);
> +		if (err)
> +			goto out;
> +	}
> +
> +out:

i'd unwind what i registered above

> +	return err;
> +}
> +
> +static void gve_unreg_xdp_info(struct gve_priv *priv)
> +{
> +	int i;
> +
> +	for (i = 0; i < priv->rx_cfg.num_queues; i++) {
> +		struct gve_rx_ring *rx = &priv->rx[i];
> +
> +		xdp_rxq_info_unreg(&rx->xdp_rxq);
> +	}
> +}
> +
>  static int gve_open(struct net_device *dev)
>  {
>  	struct gve_priv *priv = netdev_priv(dev);
>  	int err;
>  
> +	if (priv->xdp_prog)

rcu_access_pointer() ? you mark xdp_prog with __rcu annotation, isn't
sparse yelling at you?

> +		priv->num_xdp_queues = priv->tx_cfg.num_queues;
> +	else
> +		priv->num_xdp_queues = 0;
> +
>  	err = gve_alloc_qpls(priv);
>  	if (err)
>  		return err;
> @@ -994,6 +1040,10 @@ static int gve_open(struct net_device *dev)
>  	if (err)
>  		goto free_rings;
>  
> +	err = gve_reg_xdp_info(priv, dev);
> +	if (err)
> +		goto free_rings;
> +
>  	err = gve_register_qpls(priv);
>  	if (err)
>  		goto reset;
> @@ -1058,6 +1108,7 @@ static int gve_close(struct net_device *dev)
>  	}
>  	del_timer_sync(&priv->stats_report_timer);
>  
> +	gve_unreg_xdp_info(priv);
>  	gve_free_rings(priv);
>  	gve_free_qpls(priv);
>  	priv->interface_down_cnt++;
> @@ -1074,6 +1125,95 @@ static int gve_close(struct net_device *dev)
>  	return gve_reset_recovery(priv, false);
>  }
>  
> +static int gve_set_xdp(struct gve_priv *priv, struct bpf_prog *prog,
> +		       struct netlink_ext_ack *extack)
> +{
> +	struct bpf_prog *old_prog;
> +	struct napi_struct *napi;
> +	int q;
> +	int err;
> +
> +	old_prog = rtnl_dereference(priv->xdp_prog);
> +	if (!!prog != !!old_prog) {
> +		// Adding/removing a program, need to recreate the queues.
> +		if (!netif_carrier_ok(priv->dev)) {
> +			rcu_assign_pointer(priv->xdp_prog, prog);
> +			goto out;
> +		}
> +		err = gve_close(priv->dev);
> +		if (err)
> +			return err;
> +
> +		rcu_assign_pointer(priv->xdp_prog, prog);
> +		gve_open(priv->dev);
> +	} else {
> +		// Changing the XDP program. Flush the queues.
> +		rcu_assign_pointer(priv->xdp_prog, prog);
> +		if (netif_carrier_ok(priv->dev)) {
> +			for (q = 0; q < priv->rx_cfg.num_queues; q++) {
> +				napi = &priv->ntfy_blocks[priv->rx[q].ntfy_id].napi;
> +				napi_disable(napi);
> +				napi_enable(napi);
> +				if (gve_rx_work_pending(&priv->rx[q]))
> +					napi_schedule(napi);
> +			}
> +		}
> +	}
> +
> +out:
> +	if (old_prog)
> +		bpf_prog_put(old_prog);
> +	return 0;
> +}
> +
> +static int verify_xdp_configuration(struct net_device *dev)
> +{
> +	struct gve_priv *priv = netdev_priv(dev);
> +
> +	if (dev->features & NETIF_F_LRO) {
> +		netdev_warn(dev, "XDP is not supported when LRO is on.\n");
> +		return -EOPNOTSUPP;
> +	}
> +
> +	if (priv->queue_format != GVE_GQI_QPL_FORMAT) {
> +		netdev_warn(dev, "XDP is not supported in mode %d.\n",
> +			    priv->queue_format);
> +		return -EOPNOTSUPP;
> +	}
> +
> +	if (dev->mtu > (PAGE_SIZE / 2) - sizeof(struct ethhdr) - GVE_RX_PAD) {
> +		netdev_warn(dev, "XDP is not working with mtu %d.\n",
> +			    dev->mtu);
> +		return -EOPNOTSUPP;
> +	}
> +
> +	if (priv->rx_cfg.num_queues != priv->tx_cfg.num_queues ||
> +	    (2 * priv->tx_cfg.num_queues > priv->tx_cfg.max_queues)) {
> +		netdev_warn(dev, "XDP load failed: The number of configured RX queues %d should be equal to the number of configured TX queues %d and the number of configured RX/TX queues should be less than or equal to half the maximum number of RX/TX queues %d",
ah here it is. i would add something to set_channels callback anyway.

> +			    priv->rx_cfg.num_queues,
> +			    priv->tx_cfg.num_queues,
> +			    priv->tx_cfg.max_queues);
> +		return -EINVAL;
> +	}
> +	return 0;
> +}
> +
> +static int gve_xdp(struct net_device *dev, struct netdev_bpf *xdp)
> +{
> +	struct gve_priv *priv = netdev_priv(dev);
> +	int err;
> +
> +	err = verify_xdp_configuration(dev);
> +	if (err)
> +		return err;
> +	switch (xdp->command) {
> +	case XDP_SETUP_PROG:
> +		return gve_set_xdp(priv, xdp->prog, xdp->extack);
> +	default:
> +		return -EINVAL;
> +	}
> +}
> +
>  int gve_adjust_queues(struct gve_priv *priv,
>  		      struct gve_queue_config new_rx_config,
>  		      struct gve_queue_config new_tx_config)
> @@ -1268,6 +1408,7 @@ static const struct net_device_ops gve_netdev_ops = {
>  	.ndo_get_stats64	=	gve_get_stats,
>  	.ndo_tx_timeout         =       gve_tx_timeout,
>  	.ndo_set_features	=	gve_set_features,
> +	.ndo_bpf		=	gve_xdp,
>  };
>  
>  static void gve_handle_status(struct gve_priv *priv, u32 status)
> diff --git a/drivers/net/ethernet/google/gve/gve_rx.c b/drivers/net/ethernet/google/gve/gve_rx.c
> index db1c74b1d7d3..3785bc150546 100644
> --- a/drivers/net/ethernet/google/gve/gve_rx.c
> +++ b/drivers/net/ethernet/google/gve/gve_rx.c
> @@ -8,6 +8,8 @@
>  #include "gve_adminq.h"
>  #include "gve_utils.h"
>  #include <linux/etherdevice.h>
> +#include <linux/filter.h>
> +#include <net/xdp.h>
>  
>  static void gve_rx_free_buffer(struct device *dev,
>  			       struct gve_rx_slot_page_info *page_info,
> @@ -591,6 +593,46 @@ static struct sk_buff *gve_rx_skb(struct gve_priv *priv, struct gve_rx_ring *rx,
>  	return skb;
>  }
>  
> +static void gve_xdp_done(struct gve_priv *priv, struct gve_rx_ring *rx,
> +			 struct xdp_buff *xdp, struct bpf_prog *xprog,
> +			 int xdp_act)
> +{
> +	struct gve_tx_ring *tx;
> +	int tx_qid;
> +	int err;
> +
> +	switch (xdp_act) {
> +	case XDP_ABORTED:
> +	case XDP_DROP:
> +	default:
> +		break;
> +	case XDP_TX:
> +		tx_qid = gve_xdp_tx_queue_id(priv, rx->q_num);
> +		tx = &priv->tx[tx_qid];
> +		spin_lock(&tx->xdp_lock);
> +		err = gve_xdp_xmit_one(priv, tx, xdp->data,
> +				       xdp->data_end - xdp->data,
> +				       XDP_XMIT_FLUSH);
> +		spin_unlock(&tx->xdp_lock);

there's no other critical section that this lock protects. if it is for
something introduced later in the patchset then could move this where it
starts to make sense?

> +
> +		if (unlikely(err)) {
> +			u64_stats_update_begin(&rx->statss);
> +			rx->xdp_tx_errors++;
> +			u64_stats_update_end(&rx->statss);
> +		}
> +		break;
> +	case XDP_REDIRECT:
> +		u64_stats_update_begin(&rx->statss);
> +		rx->xdp_redirect_errors++;
> +		u64_stats_update_end(&rx->statss);
> +		break;
> +	}
> +	u64_stats_update_begin(&rx->statss);
> +	if ((u32)xdp_act < GVE_XDP_ACTIONS)
> +		rx->xdp_actions[xdp_act]++;
> +	u64_stats_update_end(&rx->statss);
> +}
> +
>  #define GVE_PKTCONT_BIT_IS_SET(x) (GVE_RXF_PKT_CONT & (x))
>  static void gve_rx(struct gve_rx_ring *rx, netdev_features_t feat,
>  		   struct gve_rx_desc *desc, u32 idx,
> @@ -603,9 +645,12 @@ static void gve_rx(struct gve_rx_ring *rx, netdev_features_t feat,
>  	union gve_rx_data_slot *data_slot;
>  	struct gve_priv *priv = rx->gve;
>  	struct sk_buff *skb = NULL;
> +	struct bpf_prog *xprog;
> +	struct xdp_buff xdp;
>  	dma_addr_t page_bus;
>  	void *va;
>  
> +	u16 len = frag_size;
>  	struct napi_struct *napi = &priv->ntfy_blocks[rx->ntfy_id].napi;
>  	bool is_first_frag = ctx->frag_cnt == 0;
>  
> @@ -645,9 +690,35 @@ static void gve_rx(struct gve_rx_ring *rx, netdev_features_t feat,
>  	dma_sync_single_for_cpu(&priv->pdev->dev, page_bus,
>  				PAGE_SIZE, DMA_FROM_DEVICE);
>  	page_info->pad = is_first_frag ? GVE_RX_PAD : 0;
> +	len -= page_info->pad;
>  	frag_size -= page_info->pad;
>  
> -	skb = gve_rx_skb(priv, rx, page_info, napi, frag_size,
> +	xprog = rcu_dereference(priv->xdp_prog);
> +	if (xprog && is_only_frag) {
> +		void *old_data;
> +		int xdp_act;
> +
> +		xdp_init_buff(&xdp, rx->packet_buffer_size, &rx->xdp_rxq);
> +		xdp_prepare_buff(&xdp, page_info->page_address +
> +				 page_info->page_offset, GVE_RX_PAD,
> +				 len, false);
> +		old_data = xdp.data;
> +		xdp_act = bpf_prog_run_xdp(xprog, &xdp);
> +		if (xdp_act != XDP_PASS) {
> +			gve_xdp_done(priv, rx, &xdp, xprog, xdp_act);
> +			ctx->total_size += frag_size;
> +			goto finish_ok_pkt;
> +		}
> +
> +		page_info->pad += xdp.data - old_data;
> +		len = xdp.data_end - xdp.data;
> +
> +		u64_stats_update_begin(&rx->statss);
> +		rx->xdp_actions[XDP_PASS]++;
> +		u64_stats_update_end(&rx->statss);
> +	}
> +
> +	skb = gve_rx_skb(priv, rx, page_info, napi, len,
>  			 data_slot, is_only_frag);
>  	if (!skb) {
>  		u64_stats_update_begin(&rx->statss);
> diff --git a/drivers/net/ethernet/google/gve/gve_tx.c b/drivers/net/ethernet/google/gve/gve_tx.c
> index 429b159ac61a..a2fc4b074f52 100644
> --- a/drivers/net/ethernet/google/gve/gve_tx.c
> +++ b/drivers/net/ethernet/google/gve/gve_tx.c
> @@ -135,6 +135,9 @@ static void gve_tx_free_fifo(struct gve_tx_fifo *fifo, size_t bytes)
>  static int gve_clean_tx_done(struct gve_priv *priv, struct gve_tx_ring *tx,
>  			     u32 to_do, bool try_to_wake);
>  
> +static int gve_clean_xdp_done(struct gve_priv *priv, struct gve_tx_ring *tx,
> +			      u32 to_do);

couldn't this be avoided? especially since it's static.

> +
>  static void gve_tx_free_ring(struct gve_priv *priv, int idx)
>  {
>  	struct gve_tx_ring *tx = &priv->tx[idx];
> @@ -144,8 +147,12 @@ static void gve_tx_free_ring(struct gve_priv *priv, int idx)
>  
>  	gve_tx_remove_from_block(priv, idx);
>  	slots = tx->mask + 1;
> -	gve_clean_tx_done(priv, tx, priv->tx_desc_cnt, false);
> -	netdev_tx_reset_queue(tx->netdev_txq);
> +	if (tx->q_num < priv->tx_cfg.num_queues) {
> +		gve_clean_tx_done(priv, tx, priv->tx_desc_cnt, false);
> +		netdev_tx_reset_queue(tx->netdev_txq);
> +	} else {
> +		gve_clean_xdp_done(priv, tx, priv->tx_desc_cnt);
> +	}
>  
>  	dma_free_coherent(hdev, sizeof(*tx->q_resources),
>  			  tx->q_resources, tx->q_resources_bus);
> @@ -177,6 +184,7 @@ static int gve_tx_alloc_ring(struct gve_priv *priv, int idx)
>  	/* Make sure everything is zeroed to start */
>  	memset(tx, 0, sizeof(*tx));
>  	spin_lock_init(&tx->clean_lock);
> +	spin_lock_init(&tx->xdp_lock);
>  	tx->q_num = idx;
>  
>  	tx->mask = slots - 1;
> @@ -657,8 +665,106 @@ netdev_tx_t gve_tx(struct sk_buff *skb, struct net_device *dev)
>  	return NETDEV_TX_OK;
>  }
>  
> -#define GVE_TX_START_THRESH	PAGE_SIZE

why this move?

> +static int gve_tx_fill_xdp(struct gve_priv *priv, struct gve_tx_ring *tx,
> +			   void *data, int len)
> +{
> +	int pad, nfrags, ndescs, iovi, offset;
> +	struct gve_tx_buffer_state *info;
> +	u32 reqi = tx->req;
> +
> +	pad = gve_tx_fifo_pad_alloc_one_frag(&tx->tx_fifo, len);
> +	if (pad >= GVE_XDP_PADDING)
> +		pad = 0;
> +	info = &tx->info[reqi & tx->mask];
> +	info->xdp.size = len;
> +
> +	nfrags = gve_tx_alloc_fifo(&tx->tx_fifo, pad + len,
> +				   &info->iov[0]);
> +	iovi = pad > 0;
> +	ndescs = nfrags - iovi;
> +	offset = 0;
> +
> +	while (iovi < nfrags) {
> +		if (!offset)
> +			gve_tx_fill_pkt_desc(&tx->desc[reqi & tx->mask], 0,
> +					     CHECKSUM_NONE, false, 0, ndescs,
> +					     info->iov[iovi].iov_len,
> +					     info->iov[iovi].iov_offset, len);
> +		else
> +			gve_tx_fill_seg_desc(&tx->desc[reqi & tx->mask],
> +					     0, 0, false, false,
> +					     info->iov[iovi].iov_len,
> +					     info->iov[iovi].iov_offset);
> +
> +		memcpy(tx->tx_fifo.base + info->iov[iovi].iov_offset,
> +		       data + offset, info->iov[iovi].iov_len);
> +		gve_dma_sync_for_device(&priv->pdev->dev,
> +					tx->tx_fifo.qpl->page_buses,
> +					info->iov[iovi].iov_offset,
> +					info->iov[iovi].iov_len);
> +		offset += info->iov[iovi].iov_len;
> +		iovi++;
> +		reqi++;
> +	}
> +
> +	return ndescs;
> +}
> +
> +int gve_xdp_xmit_one(struct gve_priv *priv, struct gve_tx_ring *tx,
> +		     void *data, int len, u32 flags)
> +{
> +	int nsegs;
> +
> +	if (!gve_can_tx(tx, len))
> +		return -EBUSY;
> +
> +	nsegs = gve_tx_fill_xdp(priv, tx, data, len);
> +	tx->req += nsegs;
> +
> +	if (flags & XDP_XMIT_FLUSH)
> +		gve_tx_put_doorbell(priv, tx->q_resources, tx->req);

we usually ring a tx doorbell once per napi, doing so per every xmitted
frame might affect your performance.

> +
> +	return 0;
> +}
>  
> +static int gve_clean_xdp_done(struct gve_priv *priv, struct gve_tx_ring *tx,
> +			      u32 to_do)
> +{
> +	struct gve_tx_buffer_state *info;
> +	u32 clean_end = tx->done + to_do;
> +	u64 pkts = 0, bytes = 0;
> +	size_t space_freed = 0;
> +	u32 idx;
> +	int i;
> +
> +	for (; tx->done < clean_end; tx->done++) {
> +		idx = tx->done & tx->mask;
> +		info = &tx->info[idx];
> +
> +		if (unlikely(!info->xdp.size))
> +			continue;

shouldn't you rather break in such rare case? how would a particular tx
buf be 0 in the middle of other bufs that were processed by hw?

> +
> +		bytes += info->xdp.size;
> +		pkts++;

pkts == to_do ?

> +
> +		info->xdp.size = 0;
> +		/* FIFO free */
> +		for (i = 0; i < ARRAY_SIZE(info->iov); i++) {
> +			space_freed += info->iov[i].iov_len + info->iov[i].iov_padding;
> +			info->iov[i].iov_len = 0;
> +			info->iov[i].iov_padding = 0;
> +		}

looks like this FIFO free could be a helper of some sort as you have exact
same code in gve_clean_tx_done()

> +	}
> +
> +	gve_tx_free_fifo(&tx->tx_fifo, space_freed);
> +	u64_stats_update_begin(&tx->statss);
> +	tx->bytes_done += bytes;
> +	tx->pkt_done += pkts;
> +	u64_stats_update_end(&tx->statss);
> +	return pkts;
> +}
> +
> +#define GVE_TX_START_THRESH	PAGE_SIZE
>  static int gve_clean_tx_done(struct gve_priv *priv, struct gve_tx_ring *tx,
>  			     u32 to_do, bool try_to_wake)
>  {
> @@ -729,6 +835,24 @@ u32 gve_tx_load_event_counter(struct gve_priv *priv,
>  	return be32_to_cpu(counter);
>  }
>  
> +bool gve_xdp_poll(struct gve_notify_block *block, int budget)
> +{
> +	struct gve_priv *priv = block->priv;
> +	struct gve_tx_ring *tx = block->tx;
> +	u32 nic_done;
> +	u32 to_do;
> +
> +	/* If budget is 0, do all the work */
> +	if (budget == 0)
> +		budget = INT_MAX;
> +
> +	/* Find out how much work there is to be done */
> +	nic_done = gve_tx_load_event_counter(priv, tx);
> +	to_do = min_t(u32, (nic_done - tx->done), budget);

nit: in theory nic_done - tx->done could be bigger than INT_MAX, probably
better to play with same types here

> +	gve_clean_xdp_done(priv, tx, to_do);
> +	return nic_done != tx->done;
> +}
> +
>  bool gve_tx_poll(struct gve_notify_block *block, int budget)
>  {
>  	struct gve_priv *priv = block->priv;
> -- 
> 2.39.1.581.gbfd45094c4-goog
> 

Powered by blists - more mailing lists