2023-05-04 15:40:57

by Shenwei Wang

[permalink] [raw]
Subject: [PATCH v3 net 1/1] net: fec: correct the counting of XDP sent frames

In the current xdp_xmit implementation, if any single frame fails to
transmit due to insufficient buffer descriptors, the function nevertheless
reports success in sending all frames. This results in erroneously
indicating that frames were transmitted when in fact they were dropped.

This patch fixes the issue by ensureing the return value properly
indicates the actual number of frames successfully transmitted, rather than
potentially reporting success for all frames when some could not transmit.

Fixes: 6d6b39f180b8 ("net: fec: add initial XDP support")
Signed-off-by: Gagandeep Singh <[email protected]>
Signed-off-by: Shenwei Wang <[email protected]>
---
v3:
- resend the v2 fix for "net" as the standalone patch.

v2:
- only keep the bug fix part of codes according to Horatiu's comments.
- restructure the functions to avoid the forward declaration.

drivers/net/ethernet/freescale/fec_main.c | 13 +++++++++----
1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/drivers/net/ethernet/freescale/fec_main.c b/drivers/net/ethernet/freescale/fec_main.c
index 160c1b3525f5..42ec6ca3bf03 100644
--- a/drivers/net/ethernet/freescale/fec_main.c
+++ b/drivers/net/ethernet/freescale/fec_main.c
@@ -3798,7 +3798,8 @@ static int fec_enet_txq_xmit_frame(struct fec_enet_private *fep,
entries_free = fec_enet_get_free_txdesc_num(txq);
if (entries_free < MAX_SKB_FRAGS + 1) {
netdev_err(fep->netdev, "NOT enough BD for SG!\n");
- return NETDEV_TX_OK;
+ xdp_return_frame(frame);
+ return NETDEV_TX_BUSY;
}

/* Fill in a Tx ring entry */
@@ -3856,6 +3857,7 @@ static int fec_enet_xdp_xmit(struct net_device *dev,
struct fec_enet_private *fep = netdev_priv(dev);
struct fec_enet_priv_tx_q *txq;
int cpu = smp_processor_id();
+ unsigned int sent_frames = 0;
struct netdev_queue *nq;
unsigned int queue;
int i;
@@ -3866,8 +3868,11 @@ static int fec_enet_xdp_xmit(struct net_device *dev,

__netif_tx_lock(nq, cpu);

- for (i = 0; i < num_frames; i++)
- fec_enet_txq_xmit_frame(fep, txq, frames[i]);
+ for (i = 0; i < num_frames; i++) {
+ if (fec_enet_txq_xmit_frame(fep, txq, frames[i]) != 0)
+ break;
+ sent_frames++;
+ }

/* Make sure the update to bdp and tx_skbuff are performed. */
wmb();
@@ -3877,7 +3882,7 @@ static int fec_enet_xdp_xmit(struct net_device *dev,

__netif_tx_unlock(nq);

- return num_frames;
+ return sent_frames;
}

static const struct net_device_ops fec_netdev_ops = {
--
2.34.1


2023-05-05 10:36:26

by Leon Romanovsky

[permalink] [raw]
Subject: Re: [PATCH v3 net 1/1] net: fec: correct the counting of XDP sent frames

On Thu, May 04, 2023 at 10:35:17AM -0500, Shenwei Wang wrote:
> In the current xdp_xmit implementation, if any single frame fails to
> transmit due to insufficient buffer descriptors, the function nevertheless
> reports success in sending all frames. This results in erroneously
> indicating that frames were transmitted when in fact they were dropped.
>
> This patch fixes the issue by ensureing the return value properly
> indicates the actual number of frames successfully transmitted, rather than
> potentially reporting success for all frames when some could not transmit.
>
> Fixes: 6d6b39f180b8 ("net: fec: add initial XDP support")
> Signed-off-by: Gagandeep Singh <[email protected]>
> Signed-off-by: Shenwei Wang <[email protected]>
> ---
> v3:
> - resend the v2 fix for "net" as the standalone patch.
>
> v2:
> - only keep the bug fix part of codes according to Horatiu's comments.
> - restructure the functions to avoid the forward declaration.
>
> drivers/net/ethernet/freescale/fec_main.c | 13 +++++++++----
> 1 file changed, 9 insertions(+), 4 deletions(-)
>

Thanks,
Reviewed-by: Leon Romanovsky <[email protected]>

2023-05-05 19:42:04

by Jakub Kicinski

[permalink] [raw]
Subject: Re: [PATCH v3 net 1/1] net: fec: correct the counting of XDP sent frames

On Thu, 4 May 2023 10:35:17 -0500 Shenwei Wang wrote:
> In the current xdp_xmit implementation, if any single frame fails to
> transmit due to insufficient buffer descriptors, the function nevertheless
> reports success in sending all frames. This results in erroneously
> indicating that frames were transmitted when in fact they were dropped.
>
> This patch fixes the issue by ensureing the return value properly
> indicates the actual number of frames successfully transmitted, rather than
> potentially reporting success for all frames when some could not transmit.
>
> Fixes: 6d6b39f180b8 ("net: fec: add initial XDP support")
> Signed-off-by: Gagandeep Singh <[email protected]>
> Signed-off-by: Shenwei Wang <[email protected]>
> ---
> v3:
> - resend the v2 fix for "net" as the standalone patch.
>
> v2:
> - only keep the bug fix part of codes according to Horatiu's comments.
> - restructure the functions to avoid the forward declaration.
>
> drivers/net/ethernet/freescale/fec_main.c | 13 +++++++++----
> 1 file changed, 9 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/net/ethernet/freescale/fec_main.c b/drivers/net/ethernet/freescale/fec_main.c
> index 160c1b3525f5..42ec6ca3bf03 100644
> --- a/drivers/net/ethernet/freescale/fec_main.c
> +++ b/drivers/net/ethernet/freescale/fec_main.c
> @@ -3798,7 +3798,8 @@ static int fec_enet_txq_xmit_frame(struct fec_enet_private *fep,
> entries_free = fec_enet_get_free_txdesc_num(txq);
> if (entries_free < MAX_SKB_FRAGS + 1) {
> netdev_err(fep->netdev, "NOT enough BD for SG!\n");

This should really be rate limited :(

> - return NETDEV_TX_OK;
> + xdp_return_frame(frame);

Why return this frame? Since error is reported @sent_frames will not be
incremented, and therefore bq_xmit_all() will take care of returning it,
right?

Otherwise the other error return path (see below) needs to be changed
as well.

> + return NETDEV_TX_BUSY;

On DMA mapping error this function returns FEC_ENET_XDP_CONSUMED,
would be good if the functions return values where from the same
"enum". Are you going to clean that part up in net-next?

> }
>
> /* Fill in a Tx ring entry */
> @@ -3856,6 +3857,7 @@ static int fec_enet_xdp_xmit(struct net_device *dev,
> struct fec_enet_private *fep = netdev_priv(dev);
> struct fec_enet_priv_tx_q *txq;
> int cpu = smp_processor_id();
> + unsigned int sent_frames = 0;
> struct netdev_queue *nq;
> unsigned int queue;
> int i;
> @@ -3866,8 +3868,11 @@ static int fec_enet_xdp_xmit(struct net_device *dev,
>
> __netif_tx_lock(nq, cpu);
>
> - for (i = 0; i < num_frames; i++)
> - fec_enet_txq_xmit_frame(fep, txq, frames[i]);
> + for (i = 0; i < num_frames; i++) {
> + if (fec_enet_txq_xmit_frame(fep, txq, frames[i]) != 0)

nit: you can skip the "!= 0", but up to you

> + break;
> + sent_frames++;
> + }

--
pw-bot: cr